Lectures on Discrete and Polyhedral Geometry

Igor Pak
April 20, 2010

Contents
Introduction
Acknowledgments
Basic definitions and notations

3
7
8

Part I. Basic discrete geometry

1. The Helly theorem
2. Caratheodory and Barany theorems
3. The Borsuk conjecture
4. Fair division
5. Inscribed and circumscribed polygons
6. Dyson and Kakutani theorems
7. Geometric inequalities
8. Combinatorics of convex polytopes
9. Center of mass, billiards and the variational principle
10. Geodesics and quasi-geodesics
11. The Steinitz theorem and its extensions
12. Universality of point and line configurations
13. Universality of linkages
14. Triangulations
15. Hilberts third problem
16. Polytope algebra
17. Dissections and valuations
18. Monge problem for polytopes
19. Regular polytopes
20. Kissing numbers
1

11
20
26
32
39
54
61
72
83
96
102
109
118
125
139
148
159
170
177
185

Part II. Discrete geometry of curves and surfaces

21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.

The four vertex theorem

Relative geometry of convex polygons
Global invariants of curves
Geometry of space curves
Geometry of convex polyhedra: basic results
Cauchy theorem: the statement, the proof and the story
Cauchy theorem: extensions and generalizations
Mean curvature and Pogorelovs lemma
Senkin-Zalgallers proof of the Cauchy theorem
Flexible polyhedra
The algebraic approach
Static rigidity
Infinitesimal rigidity
Proof of the bellows conjecture
The Alexandrov curvature theorem
The Minkowski theorem
The Alexandrov existence theorem
Bendable surfaces
Volume change under bending
Foldings and unfoldings

193
207
216
229
241
249
257
263
273
279
286
292
302
307
315
322
329
336
340
348

Part III. Details, details...

41. Appendix
41.1. The area of spherical polygons
41.2. The law of cosines for spherical triangles
41.3. The irrationality of (arccos 13 )/
41.4. The Minkowski inequality
41.5. The equality part in the BrunnMinkowski inequality
41.6. The CayleyMenger determinant
41.7. The theory of places
41.8. The mapping lemma
42. Additional problems and exercises
42.1. Problems on polygons and polyhedra
42.2. Volume, area and length problems
42.3. Miscellaneous problems
Hints, solutions and references to selected exercises
References
Notation for the references
Index

360
360
360
361
361
362
364
365
368
370
370
372
375
377
410
432
433

Introduction
The subject of Discrete Geometry and Convex Polytopes has received much attention
in recent decades, with an explosion of the work in the field. This book is an introduction, covering some familiar and popular topics as well as some old, forgotten,
sometimes obscure, and at times very recent and exciting results. It is somewhat
biased by my personal likes and dislikes, and by no means is a comprehensive or
traditional introduction to the field, as we further explain below.
This book began as informal lecture notes of the course I taught at MIT in the
Spring of 2005 and again in the Fall of 2006. The richness of the material as well as
its relative inaccessibility from other sources led to making a substantial expansion.
Also, the presentation is now largely self-contained, at least as much as we could
possibly make it so. Let me emphasize that this is neither a research monograph nor
a comprehensive survey of results in the field. The exposition is at times completely
elementary and at times somewhat informal. Some additional material is included in
the appendix and spread out in a number of exercises.
The book is divided into two parts. The first part covers a number of basic results
in discrete geometry and with few exceptions the results are easily available elsewhere (to a committed reader). The sections in the first part are only loosely related
to each other. In fact, many of these sections are subjects of separate monographs,
from which we at times borrow the proof ideas (see reference subsections for the acknowledgements). However, in virtually all cases the exposition has been significantly
altered to unify and simplify the presentation. In and by itself the first part can serve
as a material for the first course in discrete geometry, with fairly large breadth and
relatively little depth (see more on this below).
The second part is more coherent and can be roughly described as the discrete differential geometry of curves and surfaces. This material is much less readily available,
often completely absent in research monographs, and, on more than one occasion, in
the English language literature. We start with discrete curves and then proceed to
discuss several versions of the Cauchy rigidity theorem, the solution of the bellows
conjecture and Alexandrovs various theorems on polyhedral surfaces.
Although we do not aim to be comprehensive, the second part is meant to be as
an introduction to polyhedral geometry, and can serve as a material for a topics class
on the subject. Although the results in the first part are sporadically used in the
second part, most results are largely independent. However, the second part requires
a certain level of maturity and should work well as the second semester continuation
of the first part.
We include a large number of exercises which serve the dual role of possible home
assignment and additional material on the subject. For most exercises, we either
include a hint or a complete solution, or the references. The appendix is a small
collection of standard technical results, which are largely available elsewhere and
included here to make the book self-contained. Let us single out a new combinatorial
proof of the uniqueness part in the BrunnMinkowski inequality and an elementary
introduction to the theory of places aimed towards the proof of the bellows conjecture.

Organization of the book. The book is organized in a fairly straightforward manner, with two parts, 40 sections, and the increasing level of material between sections
and within each section. Many sections, especially in the first part of the book, can be
skipped or their order interchanged. The exercises, historical remarks and pointers to
the literature are added at the end of each section. Theorems, propositions, lemmas,
etc. have a global numbering within each section, while the exercises are numbered
separately. Our aversion to formula numbering is also worth noting. Fortunately, due
to the nature of the subject we have very few formulas worthy of labeling and those
are labeled with AMS-TeX symbols.
The choice of material. Upon inspecting the table of contents the reader would
likely assume that the book is organized around a few of my favorite things and
has no underlying theme. In fact, the book is organized around our favorite tools,
and there are very few of them. These tools are heavily used in the second part, but
since their underlying idea is so fundamental, the first part explores them on a more
elementary level in an attempt to prepare the reader. Below is our short list, in the
order of appearance in the book.
1. Topological existence arguments. These basic non-explicit arguments are at the
heart of the Alexandrov and Minkowski theorems in Sections 3537. Sections 46
and Subsection 3.5 use (often in a delicate way) the intermediate value theorem, and
are aimed to be an introduction to the method.
2. Morse theory type arguments. This is the main tool in Section 8 and in the proof
of the FaryMilnor theorem (Section 24). We also use it in Subsection 1.3.
3. Variational principle arguments. This is our most important tool all around, giving
alternative proofs of the Alexandrov, Minkowski and Steinitz theorems (Sections 11,
35, and 36). We introduce and explore it in Subsection 2.2, Sections 9 and 10.
4. Moduli space, the approach from the point of view of algebraic geometry. The
idea of realization spaces of discrete configurations is the key to understand Glucks
rigidity theorem leading to Sabitovs proof of the bellows conjecture (Sections 31
and 34). Two universality type results in Sections 12 and 13 give a basic introduction
(as well as a counterpart to the Steinitz theorem).
5. Geometric and algebraic valuations. This is a modern and perhaps more technical
algebraic approach in the study of polyhedra. We give an introduction in Sections 16
and 17, and use it heavily in the proof of the bellows conjecture (Section 34 and
Subsection 41.7).
6. Local move connectivity arguments. This basic principle is used frequently in
combinatorics and topology to prove global results via local transformations. We introduce it in Section 14 and apply it to scissor congruence in Section 17 and geometry
of curves in Section 23.
7. Spherical geometry. This is a classical and somewhat underrated tool, despite its
wide applicability. We introduce it in Section 20 and use it throughout the second
part (Sections 24 and 25, Subsections 27.1, 29.3).
Note that some of these are broader and more involved than others. On the other
hand, some closely related material is completely omitted (e.g., we never study the

hyperbolic geometry). To quote one modern day warrior, If you try to please everybody, somebodys not going to like it [Rum].
Section implications. While most sections are independent, the following list of
implications shows which sections are not: 1 2, 3, 20 5 23 7 28, 36
9 10 25, 40 11 34 12 13 14 17 15 16 17 21 23
22 26 27 37 26 28, 29, 30 25 35 37 38 31 32 33
Suggested course content. Although our intention is to have a readable (and
teachable) textbook, the book is clearly too big for a single course. On a positive
side, the volume of book allows one to pick and choose which material to present.
Below we present several coherent course suggestions, in order of increasing difficulty.
(1) Introduction to Discrete Geometry (basic undergraduate course).
1, 2.1-2, 3, 4, 5.1-2 (+ Prop. 5.9, 5.11), 23.6, 25.1, 19, 20, 8.1-2, 8.4, 9, 10, 12, 13,
14.1-3, 15, 21.1-3, 23.1-2, 23.6, 22.1-5, 26.1-4, 30.2, 30.4, 39, 40.3-4.
(2) Modern Discrete Geometry (emphasis on geometric rather than combinatorial
aspects; advanced undergraduate or first year graduate course).
46, 9, 10, 1215, 17.5-6, 18, 2023, 25, 26, 29, 30, 33, 35.5, 36.3-4, 39, 40.

(3) Geometric Combinatorics (emphasis on combinatorial rather than geometric aspects; advanced undergraduate or first year graduate course).
25.1, 19, 20, 14, 8, 11, 12, 1417, 23, 22.1-5, 26.1-4, 32, 33, 40.3, 40.4.

(4) Discrete Differential Geometry (graduate topics course)

911, 21, 22, 2428, 3035, 7, 3638, 40.

(5) Polytopes and Algebra (intuitive graduate topics course)

Scissor congruence: 15, 16, 14.4, 17, then use [Bolt, Dup, Sah] for further results.
Realization spaces and the bellows conjecture: 1113, 31, 34, then use [Ric].
Integer points enumeration: use [Barv, 7], [Grub, 19] and [MilS, 12].
On references and the index. Our reference style is a bit idiosyncratic, but,
hopefully, is self-explanatory.1 Despite an apparently large number of references, we
made an effort to minimize their number. Given the scope of the field, to avoid an
explosion of the references, we often omit important monographs and papers in favor
of more recent surveys which contain pointers to these and many other references.
Only those sources explicitly mentioned in the remarks and exercises are included. On
occasion, we added references to classical texts, but only if we found the exposition
in them to be useful in preparing this book. Finally, we gave a certain preference
to the important foreign language works that are undeservingly overlooked in the
modern English language literature, and to the sources that are freely available on
the web, including several US patents. We made a special effort to include the arXiv
numbers and the shortened clickable web links when available. We sincerely apologize
to authors whose important works were unmentioned in favor of recent and more
accessible sources.
1See

also page 432 for the symbol notations in the references.

We use only one index, for both people and terminology. The references have
pointers to pages where we use them, so the people in the index are listed only if they
are mentioned separately.
On exercises. The exercises are placed at the end of every section. While most
exercises are related to the material in the section, the connection is sometimes not
obvious and involves the proof ideas. Although some exercises are relatively easy and
are meant to be used as home assignments, most others contain results of independent
interest. More often than not, we tried to simplify the problems, break them into
pieces, or present only their special cases, so that they can potentially be solved by
a committed reader. Our intention was to supplement the section material with a
number of examples and applications, as well as mention some additional important
results.
The exercises range from elementary to very hard. We use the following ranking:
exercises labeled [1-], [1] and [1+] are relatively simple and aimed at students, while
those labeled [2-], [2] and [2+] are the level of a research paper with the increasing
involvement of technical tools and results from other fields. We should emphasize
that these rankings are approximate at best, e.g., some of those labeled [1+] might
prove to be excessively difficult, less accessible than some of those labeled [2-] . If an
exercise has a much easier proof than the ranking suggests, please let me know and I
would be happy to downgrade it.
We mark with the exercises that are either used in the section or are mentioned
elsewhere as being important to understanding the material. Some additional, largely
assorted and ad hoc exercises are collected in Section 42. These are chosen not for
their depth, but rather because we find them appealing enough to be of interest to
the reader.
Hints, brief solutions and pointers to the literature are given at the end of the book.
While some solutions are as good as proofs in the main part, most are incomplete
and meant to give only the first idea of what to do or where to go. Open problems
and a few simple looking questions I could not answer are marked with [] . They are
likely to vary widely in difficulty.
On figures. There are over 250 color figures in the book, and they are often integral
to the proofs.

Acknowledgements. Over the years I benefitted enormously from conversations

with a number of people. I would like to thank Arseny Akopyan, Boris Aronov, Imre
Barany, Sasha Barvinok, Bob Connelly, Jes
us De Loera, Elizabeth Denne, Nikolai
Dolbilin, Maksym Fedorchuk, Alexey Glazyrin, David Jerison, Gil Kalai, Greg Kuperberg, Gilad Lerman, Ezra Miller, Oleg Musin, Janos Pach, Rom Pinchasi, Yuri
Rabinovich, Jean-Marc Schlenker, Richard Stanley, Alexey Tarasov, Csaba Toth, Uli
Wagner and Victor Zalgaller for their helpful comments on geometry and other matters.
Parts of these lectures were read by a number of people. I would like to thank
Andrew Berget, Steve Butler, Jason Cantarella, Erhard Heil, Ivan Izmestiev, Kristoffer Josefsson, Michael Kapovich, Igor Klep, Carly Klivans, Idjad Sabitov, Raman
Sanyal, Egon Schulte, Sergei Tabachnikov, Hugh Thomas, Gudlaugur Thorbergsson,

Rade Zivaljevi
c and anonymous reviewers for their corrections and suggestions on the
text, both big and small. Jason Cantarella and Andreas Gammel graciously allowed
us to use Figures 42.14 and 39.5. Special thanks to Ilya Tyomkin for his great help
with the theory of places. Andrew Odlyzkos suggestion to use the tinyurl links
helped make the references web accessible. I am also deeply grateful to all students
in the courses that I taught based on this material, at MIT, UMN, and UCLA, for
their interest in the subject on and off the lectures.
Finally, I would like to thank my parents Mark and Sofia for their help and patience
during our Moscow summers and Spring 2008 when much of the book was written.

Igor Pak
Department of Mathematics, UCLA
Los Angeles, CA 90095, USA
[email protected]

Basic definitions and notations. Let P Rd be a convex polytope, or a general

convex body. We use S = P for the surface of P . To simplify the notation we will
always use area(S), for the surface area of P . Formally, we use area() to denote the
(d 1)-dimensional volume: area(S) = vold1 (S). We say that a hyperplane H is
supporting P at x S if P lies on one side of H, and x H.
For a set X Rd we use cm(X) to denote the center of mass 2 and conv(X) to
denote the convex hull of X. We use |xy| to denote the distance between points
The vectors are
x, y Rd . Alternatively, kek denotes the length of a vector e =
xy.
always in bold, e.g., O usually denotes the origin, while 0 is the zero vector. We let
hu, w i denote the scalar product of two vectors. The geodesic distance between two
surface points x, y S is denoted by |xy|S . We use to denote the spherical angles.
All our graphs will have vertices and edges. The edges are sometimes oriented,
but only when we explicitly say so. In the beginning of Section 21, following a long
standing tradition, we study vertices smooth curves, but to avoid confusion we
never mention them in later sections.
In line with tradition, we use the word polygon to mean two different things: both
a simple closed piecewise linear closed curve and the interior of this curve. This
unfortunate lack of distinction disappears in higher dimension, when we consider space
polygons. When we do need to make a distinction, we use closed piecewise linear curve
and polygonal region, both notions being somewhat unfortunate. A simple polygon is
always a polygon with no self-intersections. We use Q = [v1 . . . vn ] to denote a closed
polygon with vertices vi in this (cyclic) order. We also use (abc) to denote a triangle,
and, more generally, (v0 v1 . . . vd ) to denote a (d + 1)-dimensional simplex. Finally, we
use (u, v) for an open interval (straight line segment) or an edge between two vertices,
[xy] and [x, y] for a closed interval, either on a line or on a curve, and (xy) for a line
through two points.
In most cases, we use polytope to mean a convex hull of a finite number of points.
Thus the word polytope is usually accompanied with an adjective convex, except in
Sections 1517, where a polytope is a finite union of convex polytopes. In addition,
we assume that it is fully dimensional, i.e., does not lie in an affine hyperplane. In
all other cases we use the word polyhedron for convex and non-convex surfaces, noncompact intersections of half-spaces, etc. Since we are only concerned with discrete
results, we do not specify whether polytopes are open or closed sets in Rd , and use
whatever is appropriate in each case.
We make a distinction between subdivisions and decompositions of a polytope,
where the former is required to be a CW complex, while the latter is not. The
notions of triangulation is so ambiguous in the literature, we use it only for simplicial
subdivisions. We call dissections the simplicial decompositions, and full triangulations
the triangulations with a given set of vertices (usually the vertices of a given convex
polytope).
Occasionally we use the standard notation for comparing functions: O(), o(),
() and (). We use various arrow-type symbols, like , , , , , , etc.,
2We

consider cm(X) only for convex of piecewise linear sets X, so it is always well dened.

for different kind of flips, local moves, and equivalence relations. We reserve for
numerical estimates.
Finally, throughout the book we employ [n] = {1, 2, . . . , n}, N = {1, 2, . . .}, Z+ =
{0, 1, 2, . . .}, R+ = {x > 0}, and Q+ = {x > 0, x Q}. The d-dimensional Euclidean
space is always Rd , a d-dimensional sphere is Sd , and a hemisphere is Sd+ , where d 1.
To simplify the notation, we use X a and X + b to denote X r {a} and X {b},
respectively.

10

Part I
Basic Discrete Geometry

11

1. The Helly theorem

We begin our investigation of discrete geometry with the Helly theorem and its
generalizations. That will occupy much of this and the next section. Although these
results are relatively elementary, they lie in the heart of discrete geometry and are
surprisingly useful (see Sections 3 and 24).
1.1. Main result in slow motion. We begin with the classical Helly Theorem in
the plane.
Theorem 1.1 (Helly). Let X1 , . . . , Xn R2 be convex sets such that Xi Xj Xk 6=
for every 1 i < j < k n, where n 3. Then there exists a point z X1 , . . . , Xn .
In other words, if all triples of convex sets intersect, then all sets intersect. The
convexity condition in the Helly theorem is necessary, as can be seen in the example
in Figure 1.1 below.

Figure 1.1. The role of convexity in the Helly theorem.

Proof. We prove the result by induction on n. For n = 3 there is nothing to prove.
For n = 4, consider points v1 X2 X3 X4 , v2 X1 X3 X4 , v3 X1 X2 X4 ,
and v4 X1 X2 X3 . There are two possibilities: either points vi are in convex
position, or one of them, say v4 , is inside the triangle on remaining the points (v1 v2 v3 ).
In the second case, z = v4 clearly works. In the first case, take z to be the intersection
of two diagonals inside the 4-gon (v1 v2 v3 v4 ). For points labeled as in Figure 1.2 we
have:
z conv{v1 , v3 } conv{v2 , v4 } (X2 X4 ) (X1 X3 ),
which implies the claim for n = 4.
Let us now make a notation which we will use throughout the section. Denote by
XI = iI Xi , where I [n] = {1, . . . , n}. For the induction step, assume that n > 4
and for every (n 1)-element subset I [n] we have XI 6= . Denote by vi any point
in X[n]i, and consider a configuration of 4 points: v1 , v2 , v3 and v4 . As in the case
n = 4, there are two possibilities: these points are either in convex position or not.
Arguing in the same manner as above we conclude that in each case there exists a
point z X[n] , as desired.

Here is a natural generalization of the Helly theorem in higher dimensions:

12

v3

v3

v4
v1

v1
v2

z
v2
v4

Figure 1.2. The Helly theorem for n = 4.

Theorem 1.2 (d-dimensional Helly theorem). Let X1 , . . . , Xn Rd be n d + 1
convex sets such that XI 6= for every subset I [n], |I| = d + 1. Then there exists
a point z X1 , . . . , Xn .
Before we present the proof of Theorem 1.2, let us note that this statement is
really about convex polytopes rather than some (at times geometrically complicated)
convex sets. This fact will prove useful later on.

Proposition 1.3. In Theorem 1.2, it suffices to prove the result for convex polytopes.
Proof. Indeed, for every I [n] such that XI 6= fix a point xI XI . Let Yi =
conv{xI | i I}. Observe that YI XI , so if XI = then YI = . On the other
hand, if XI 6= , then YI xI since each Yi xI , for all i I. Therefore, it suffices
to prove the theorem only for the convex polytopes Y1 , . . . , Yn .

Our proof of the d-dimensional Helly Theorem proceeds by induction as in the
plane. However, for d > 2 there are more cases to consider. These cases will be
handled by the following simple result.
Theorem 1.4 (Radon). Let a1 , . . . , am Rd be any m d + 2 points. Then there
exists two subsets I, J [m], such that I J = and
conv{ai | i I} conv{aj | j J} =
6 .

Proof. Write the coordinates of each point ai = (a1i , . . . , adi ) Rd , and consider a
system of d + 1 equations and m variables 1 , . . . , m :
m
X

i = 0,

i=1
m
X

i ari = 0,

i=1

for all 1 r d.

Since m d + 2, the system has a nonzero solution (1 , . . . , m ). From the first

equation, some i P
are positive and some are negative. Set I = {i : i > 0}, J = {j :
j < 0}, and c = iI i . Adding the last d equations as above and rearranging the
terms we obtain:
X
X
i ai +
j aj = O,
iI

jJ

13

X i
X j
ai =
aj .
c
c
jJ
iI

Since both sides are convex combinations of points ai from disjoint sets I and J, we
obtain the result.

Proof of Theorem 1.2. Use induction. The case n = d + 1 is clear. Suppose n d + 2
and every (n 1)-element subset of convex sets Xi has a common point vi X[n]i .
By Radons theorem, there exists two disjoint subsets I, J [n] and a point z Rd ,
such that
z conv{vi | i I} conv{vj | j J} X[n]rI X[n]rJ X[n] ,

where the second inclusion follows by definition of points vi .

1.2. Softball geometric applications. The Helly theorem does not look at all
powerful, but it has, in fact, a number of nice geometric applications. We present
here several such applications, leaving others as exercises.
Corollary 1.5. Let P1 , . . . , Pn R2 be rectangles with sides parallel to the coordinate
axes, such that every two rectangles intersect each other. Then all rectangles have a
nonempty intersection.
Proof. By the Helly theorem, it suffice to show that every three rectangles intersect
(see Figure 1.3). Indeed, project three rectangles Pi , Pj , Pk onto the x axis. By the
Helly theorem for the line, the three intervals in projection intersect at a point x0 .
Similarly, the three intervals in projection of the three rectangles onto the y axis
intersect at a point y0 . Therefore, all three rectangles contain point (x0 , y0 ), which
completes the proof.


Figure 1.3. Three pairwise intersecting rectangles as in Corollary 1.5.

Note that in the proof we used the 1-dimensional version of the Helly theorem, the
claim we took for granted when we started the section with d = 2 case. Of course,
the proof of Theorem 1.2 works fine in this case.
Corollary 1.6. Let A R2 be a fixed convex set and let X1 , . . . , Xn R2 be any
convex sets such that every three of them intersect a translation of A. Then there
exists a translation of A that intersects all sets Xi .
Note that this corollary can be viewed as an extension of the Helly theorem, which
is the case when A is a single point (its translations are all points in the plane).

14

Proof. Denote by cm(B) the center of mass of the set B. For every i, define Yi in
such a way that Xi A 6= if and only if cm(A ) Yi , for every translation A of A
(see Figure 1.4). Now apply the Helly theorem to these convex sets Yi . For a point
z Y1 . . . Yn find a translation A with cm(A ) = z. By definition of the sets Yi ,
the set A intersects all sets Xi .

cm(A)

Yi

Xi

Figure 1.4. Convex sets A, Xi , and construction of the set Yi .

Corollary 1.7. Let z1 , . . . , zn R2 be points in the plane, such that every three of
them can be covered by a circle of radius r. Then all points can be covered by a circle
of radius r.
Proof. The result follows immediately from Corollary 1.6 with xi = {zi }, and A is a
circle of radius r.

Note that here we used the other extreme of Corollary 1.6 (setting Xi to be points)
than the one which gives the Helly theorem.
Corollary 1.8. Let z1 , . . . , zn R2 be points in the plane, such that all pairwise
distances |zi zj | are at most 1. Then all points can be covered by a circle of radius 13 .
Proof. By Corollary 1.7 it suffices to show that every three points zi , zj , and zk can
be covered with a circle of radius 13 . There are three cases: triangle T = (zi zj zk )
is either acute, right or obtuse. In the last two case take a circle C centered at the
midpoint of the longest edge. Clearly, the radius of C is at most 21 < 13 .
zj
p

zi
zk

Figure 1.5. Computing the radius of the circumscribed circle C.

15

When T is acute, the proof is apparent from Figure 1.5. Let O be the center of the
circumscribed circle C. Suppose zi Ozj is the largest angle in T and let = zi Op
be as in the figure. Then = 12 ( zi Ozj ) 3 . On the other hand, |zi p| = 12 |zi zj | 12 ,
by assumption. Therefore,
radius(C) = |zi O| =
as desired.

|zi p|
1
,
sin
3


Note that Corollary 1.8 is sharp in a sense that one cannot make the radius smaller
(take an equilateral triangle). On the other hand, it is always possible to decrease
the shape of the covering set. This approach will be used in the next section.
1.3. The fractional Helly theorem. It turns out, one can modify the Helly theorem to require not all but a constant proportion of sets XI to be nonempty. While
one cannot guarantee in this case that all sets Xi are intersecting (a small fraction
of them can even be completely disjoint from the other sets), one can still show that
a constant proportion of them do intersect. We will not prove the sharpest version
of this result. Instead, we introduce the Morse function approach which will be used
repeatedly in the study of simple polytopes (see Section 8).
Theorem 1.9 (Fractional
Helly). Let > 0, and X1 , . . . , Xn Rd be convex sets
n
such that for at least d+1
of (d + 1)-element sets I [n] we have XI 6= . Then

there exists a subset J [n], such that |J| > d+1

n, and XJ 6= .
Before we prove the theorem, recall Proposition 1.3. Using the same proof verbatim,
we conclude that it suffices to prove the result only for convex polytopes.
Now, let : Rd R be a generic linear function, i.e., nonconstant on all edges
of all polytopes XI , I [n]. Then every polytope XI will have the smallest vertex
min (XI ). We will refer to as the Morse function. We need the following simple
result.
Lemma 1.10. Let I [n] be such that XI 6= , and let v = min (XI ) be the
minimum of the Morse function. Then there exists a subset J I such that v =
min (XJ ) and |J| d.
Proof. Consider a set C := {w Rd | (w) < (v)}. Clearly, C is convex and
C XI = . In other words, a family of convex sets {C} {Xi , i I} has an
empty intersection, which implies that it contains a (d + 1)-element subfamily {C}
{Xj : j J} for some J I, with an empty intersection: C XJ = . Since all
sets XJ XI contain v, this implies that (z) (v) for all z XJ (see Figure 1.6).
By construction of this implies that v = min (XJ ).

Proof of Theorem 1.9. Use a double counting argument. Fix a Morse function .
Denote by J = (I) the d-element subset J [n] given by the lemma above.

16

XJ
XI
v

Figure 1.6. Morse function in Lemma 1.10.

Now consider only (d+1)-element
subsets I [n+1] Since the number of d-element


n
subsets J is at most d , there exists a subset J0 such that J0 = (I) for at least


n
d+1
nd

=
n
d+1
d

different (d + 1)-element subsets I. Let v = min (XJ0 ). By definition of the map ,

we have I J0 for each I as above. Therefore, there exist at least nd
different i =
d+1
I J0 such that v Xi . Adding to this all sets Xj with j J0 , we conclude that v
belongs to at least
nd
n
d+
>
d+1
d+1
convex subsets Xi .


1.4. Exercises.
Exercise 1.1. (Innite Helly theorem) [1] Extend the Helly theorem (Theorem 1.2) to
innitely many (closed) convex sets.
Exercise 1.2. a) [1-] Let P R3 be a convex polytope such that of the planes spanned
by the faces, every three intersect at a point, but no four intersect. Prove that there exist
four such planes which form a tetrahedron containing P .
b) [1-] Generalize the result to polytopes in Rd .
Exercise 1.3. [1+] Prove that every polygon Q R2 of length L (the sum of lengths of all
edges) can be covered by a disk of radius L/4.
Exercise 1.4. [1+] Suppose n 3 unit cubes are inscribed into a sphere, such that every
three of them have a common vertex. Prove that all n cubes have a common vertex.
Exercise 1.5. [1] Suppose every three points of z1 , . . . , zn R2 can be covered by a triangle
of area 1. Prove that conv(z1 , . . . , zn ) can be covered by a triangle of area 4.
Exercise 1.6. [1] Suppose there are n lines in the plane such that every three of them
can be intersected with a unit circle. Prove that all of them can be intersected with a unit
circle.

17

Exercise 1.7. (Generalized Helly theorem) [2] Let X1 , . . . , Xn be convex sets in Rd . Suppose the intersections of every d + 1 k of them contain an ane k-dimensional subspace.
Prove that there exists an ane k-dimensional subspace contained in all subspaces: H Xi ,
for all 1 i n.
Exercise 1.8. a) [2-] Let Rd+1 be a simplex and let f : Rd+1 Rd be a linear map.
Denote by T1 , . . . , Td+2 the facets of . Prove that f (T1 ) . . . f (Td+2 ) 6= .
b) [2+] Show that part a) holds for all continuous maps f : Rd .
Exercise 1.9. a) [2-] Let P Rd+1 be a convex polytope and let f : Rd+1 Rd be a
linear map. Prove that there exist two disjoint faces F, G P , such that f (F ) f (G) 6= .
b) [1+] Show that part a) implies Radons theorem (Theorem 1.4).
c) [2+] Show that part a) holds for all continuous maps f : P Rd .
d) [2+] Let X Rd+1 be a convex body with nonempty interior and surface S = X. Let
f : S Rd be a continuous map. Points x, y S are called opposite if they lie on distinct
parallel hyperplanes supporting X. Prove that there exist two opposite points x, y S such
that f (x) = f (y).
Exercise 1.10. [2] Let P Rd be a convex polytope dierent from a simplex. Suppose
X1 , . . . , Xn P are convex sets (thus each lying in a facet of P ), such that every d of
them intersect. Prove that all of them intersect.
Exercise 1.11. (Spherical Helly theorem) a) [1] Suppose X1 , . . . , Xn S2+ are convex sets
on a hemisphere, such that every three of them intersect. Prove that all Xi intersect.
b) [1+] Suppose X1 , . . . , Xn Sd+ are convex sets on a hemisphere, such that every d + 1 of
them intersect. Prove that all Xi intersect.
c) [2-] Suppose X1 , . . . , Xn S2 are convex sets on a sphere, such that each of them is
inside a hemisphere, every three of them intersect and no four cover the sphere. Prove that
all Xi intersect.
Exercise 1.12. a) [1] Suppose z1 , . . . , zn S2 are points on a unit sphere, such that every
three of them can be covered by a spherical disk of radius r < /3. Prove that all points
can be covered by a spherical circle of radius r.
b) [1] Check that the claim is false if we require only that r < /2. Find the optimal
constant.
Exercise 1.13. (Kirchberger ) a) [1] Let X R2 be a set of n 4 points in general
position. Suppose points in X are colored with two colors such that for every four points
there is a line separating points of dierent color. Prove that there exists a line separating
all points in X by color.
b) [1+] Generalize the result to Rd .
Exercise 1.14. Suppose Q R2 is a simple polygon which contains a point w, such that
interval [w, x] lies in Q, for all x Q. Such Q is called a star-shaped polygon. A kernel of Q
is a set KQ of points w as above.
a) [1-] Prove that the kernel KQ is a convex polygon.
b) [1] Prove that the boundary KQ lies in the union of Q and the lines spanned by the
inection edges of Q (see Subsection 23.1).
c) [1] Dene the tangent set LQ to be the union of all lines supporting polygon Q. Prove
that LQ is the whole plane unless Q is star-shaped. Moreover, the closure of R2 r LQ is
exactly the kernel KQ .

18

d) [1+] For a 2-dimensional polyhedral surface S R3 dene the tangent set LS to be the
union of planes tangent to S. Prove that LS = R3 unless S is homeomorphic to a sphere
and encloses a star-shaped 3-dimensional polyhedron.
Exercise 1.15. (Krasnoselskii) a) [1] Suppose Q R2 is a simple polygon, such that
for every three points x, y, z Q there exists a point v such that line segments [vx], [vy],
and [vz] lie in Q. Prove that Q is a star-shaped polygon.
b) [1+] Generalize the result to Rd ; use (d + 1)-tuples of points.
Exercise 1.16. a) [2-] Let X1 , . . . , Xm Rd be convex sets such that their union is also
convex. Prove that if every m 1 of them intersect, then all of them intersect.
b) [2] Let X1 , . . . , Xm Rd be convex sets such that the union of every d + 1 or fewer of
them is a star-shaped region. Then all Xi have a nonempty intersection.
c) [2-] Deduce part a) from part b).
d) [2+] Find a fractional analogue of part b).
Exercise 1.17. a) [1] Let Q R2 be a union of axis parallel rectangles. Suppose for every
two points x, y Q there exists a point v Q, such that line segments [vx] and [vy] lie
in Q. Prove that Q is a star-shaped polygon.
b) [1+] Generalize the result to Rd ; use pairs of points.
Exercise 1.18. a) [2-] Suppose X1 , . . . , Xn R2 are disks such that every two of them
intersect. Prove that there exist four points z1 , . . . , z4 such that every Xi contains at least
one zj .
b) [2-] In condition of a), suppose Xi are unit disks. Prove that three points zj suce.
Exercise 1.19. [1+] Prove that for every nite set of n points X R2 there exist at most n
disks which cover X, such that the distance between any two disks is 1, and the total
diameter is n.
Exercise 1.20. a) [2-] Suppose X1 , . . . , Xn R2 are convex sets such that through every
four of them there exists a line intersecting them. Prove that there exist two lines 1 and 2
such that every Xi intersects 1 or 2 .
b) [2+] Let X1 , . . . , Xn Rd be axis parallel bricks. Suppose for any (d+1)2d1 of these there
exists a hyperplane intersecting them. Prove that there exists a hyperplane intersecting
all Xi , 1 i n.

Exercise 1.21. (Topological Helly theorem) a) [1] Let X1 , . . . , Xn R2 be simple polygons3 in the plane, such that all double and triple intersections of Xi are also (nonempty)
simple polygons. Prove that the intersection of all Xi is also a simple polygon.
b) [1+] Generalize part a) to higher dimensions.
Exercise 1.22. a) [1] Let X1 , . . . , Xn R2 be simple polygons in the plane, by which we
mean here polygonal regions with no holes. Suppose that all unions Xi Xj are also simple
polygons, and that all intersections Xi Xj are nonempty. Prove that the intersection of
all Xi is also nonempty.
b) [1+] Let X1 , . . . , Xn R2 be simple polygons in the plane, such that all unions Xi Xj Xk
are also simple, for all 1 i j k n. Prove that the intersection of all Xi is nonempty.
Exercise 1.23. (Weak converse Helly theorem) a) [1] Let X R2 be a simple polygon
such that for every collection of triangles T1 , . . . , Tn R2 either X T1 Tn 6= , or
X Ti Tj = for some 1 i < j < n. Prove that X is convex.
3Here

a simple polygon is a connected, simply connected nite union of convex polygons.

19

b) [1+] Generalize part a) to higher dimensions.

Exercise 1.24. (Converse Helly theorem) a) [2-] Let A be an innite family of simple
polygons in the plane which is closed under non-degenerate ane transformations. Suppose A satises the following property: for every four elements in A, if every three of them
intersect, then so does the fourth. Prove that all polygons in A are convex.
b) [2] Extend this to general compact sets in the plane.
c) [2+] Generalize this to higher dimensions.
Exercise 1.25. (Convexity criterion) a) [2-] Let X R3 be a compact set and suppose
every intersection of X by a plane is contractible. Prove that X is convex.4
b) [2+] Generalize this to higher dimensions.
Exercise 1.26. a) [2-] Let A be an innite family of simple polygons in the plane which
is closed under rigid motions and intersections. Prove that all polygons in A are convex.
b) [2] Extend this to general connected compact sets in the plane.
c) [2] Generalize this to higher dimensions.
1.5. Final remarks. For a classical survey on the Helly theorem and its applications we
refer to [DGK]. See also [Eck] for recent results and further references. The corollaries we
present in this section are selected from [HDK] where numerous other applications of Helly
theorem are presented as exercises.
Our proof of the fractional Helly theorem (Theorem 1.9) with minor modications follows [Mat1, 8.1]. The constant /(d + 1) in the theorem can be replaced with the optimal
constant
1 (1 )1/(d+1) ,
due to Gil Kalai (see [Eck] for references and discussions of this and related results).

4Already

the case when X is polyhedral (a nite union of convex polytopes) is non-trivial and is
a good starting point.

20

ra
ny theorems
2. Carath
eodory and Ba
This short section is a followup of the previous section. The main result, the Barany
theorem, is a stand-alone result simply too beautiful to be missed. Along the way
we prove the classical SylvesterGallai theorem, the first original result in the whole
subfield of point and line configurations (see Section 12).
2.1. Triangulations are fun. We start by stating the following classical theorem in
the case of a finite set of points.
Theorem 2.1 (Caratheodory). Let X Rd be a finite set of points, and let z
conv(X). Then there exist x1 , . . . , xd+1 X such that z conv{x1 , . . . , xd+1 }.
Of the many easy proofs of the theorem, the one most relevant to the subject is the
proof by triangulation: given a simplicial triangulation of a convex polytope P =
conv(X), take any simplex containing z (there may be more than one). To obtain a
triangulation of a convex polytope, choose a vertex v P and triangulate all facets
of P (use induction on the dimension). Then consider all cones from v to the simplices
in the facet triangulations.5
One can also ask how many simplices with vertices in X can contain the same
point. While this number may vary depending on the location of the point, it turns
out there always exists a point z which is contained in a constant proportion of all
simplices.
Theorem 2.2 (Barany). For every d 1 there exists a constant d > 0, such that
for every set of n points X
Rd in general position, there exists a point z conv(X)


n
contained in at least d d+1
simplices (x1 . . . xd+1 ), xi X.
Here the points are in general position if no three points lie on a line. This result
is nontrivial even for convex polygons in the plane, where the triangulations are well
understood. We suggest the reader try to prove the result in this special case before
going through the proof below.

2.2. Infinite descent as a mathematical journey. The method of infinite descent

is a basic tool in mathematics, with a number of applications in geometry. To illustrate the method and prepare for the proof of the Barany theorem we start with the
following classical result of independent interest:
Theorem 2.3 (SylvesterGallai). Let X R2 be a finite set of points, not all on the
same line. Then there exists a line containing exactly two points in X.
Proof. Suppose every line which goes through two points in X, goes also through a
third such point. Since the total number of such lines is finite, consider the shortest
distance between points and lines. Suppose the minimum is achieved at a pair:
point y X and line . Suppose also that goes through points x1 , x2 , x3 (in that
5Formally

speaking, in dimensions at least 4, this argument produces a dissection, and not necessarily a (face-to-face) triangulation (see Exercise 2.1). Of course, a dissection suces for the purposes
of the Caratheodory theorem. We formalize and extend this approach in Subsection 14.5.

21

order). Since sum of the angles x1 yx2 + x2 yx3 < , at least one of them, say
x1 yx2 < /2. Now observe that the distance from x2 to line (x1 , y) is smaller that
the distance from y to , a contradiction (see Figure 2.1).

y
v

x1

x2 w

x3

Figure 2.1. Descent step [yw] [x2 v] in the proof of the Sylvester
Gallai theorem.
One can think of the proof of the SylvesterGallai theorem as an algorithm to find
a line with exactly two points, by descending through the pairs (point, line) with
shorter and shorter distances. Of course, for infinite configurations, where every line
contains 3 points (e.g., a square grid), this descent is infinite indeed.
The following result is a colorful generalization of the Caratheodory theorem,
and has a proof from the book, again by an infinite descent.
Theorem 2.4 (Colorful Caratheodory). Let X1 , . . . , Xd+1 Rd be finite sets of points
whose convex hulls contain the origin, i.e., O conv(Xi ), 1 i d + 1. Then there
exist points x1 X1 , . . . , xd+1 Xd+1 , such that O conv{x1 , . . . , xd+1 }.
Here the name come from the idea that points in the same set Xi are colored with
the same color, these colors are different for different Xi , and the theorem shows
existence of a simplex with vertices of different color.
Proof. We call (x1 x2 . . . xd+1 ) with xi Xi , the rainbow simplex. Suppose none of the
rainbow simplices contain the origin. Let = (x1 x2 . . . xd+1 ) be the closest rainbow
simplex to O, and let z be the closets point of to O. Denote by H the
hyperplane containing z and orthogonal to (Oz). Clearly, some vertex of , say, x1 ,
does not lie on H. Since conv(X1 ) O and O, x1 lie on different sides of H, there
exists a point y X1 which lies on the same side of H as O. Now observe that
:= (yx2 . . . xd+1 ) is a rainbow simplex which contains an interval (yz) lying on the
same side of H as O (see Figure 2.2 for the case when z lies in the relative interior of
a facet). We conclude that is closer to O, a contradiction.

2.3. Generalized Radon theorem. Recall the Radon theorem (Theorem 1.4),
which states that every set of size d + 2 can be split into two subsets whose convex
hulls intersect. It is natural to ask for the smallest number t = t(d, r) of points needed
so they can be split into r sets whose convex hulls intersect. A well known theorem
of Tverberg gives t(d, r) = (r 1)(d + 1) + 1. We present here an easy proof of a
weaker version of this result.

22

x1
x2

x3

Figure 2.2. Descent step (x1 x2 x3 ) (yx2 x3 ) on rainbow simplices.

Theorem 2.5 (Tverberg). Let n (r 1)(d + 1)2 + 1 for d, r N. Then for every n
points x1 , . . . , xn Rd there exist r subsets I1 , . . . , Ir [n], such that Ii Ij = ,
i 6= j, and
r
\


conv XIi 6= .
i=1

Proof. Let k = (r 1)(d + 1), and s = n k. Observe that every d + 1 subsets Yi ,

1 i d + 1, of X = {x1 , . . . , xn } of size s have a common point. Indeed, a
simple counting gives: |Y1 | = n k, |Y1 Y2 | n 2k, . . . , |Y1 Y2 . . . Yd+1 |
n (d + 1)k 1. In other words, for every Y1 , . . . , Yd+1 as above, their convex
hulls Bi = conv(Yi ) have a point in common (which must lie in X). Therefore, by the
Helly theorem (Theorem 1.2), all convex hulls of s-tuples (and, therefore, m-tuples,
m s) of points in X have a point z conv(X) in common (not necessarily in X).
By the Caratheodory theorem (Theorem 2.1), since |X| d + 2, we can choose a
(d+1)-element subset X1 X such that z conv(X1 ). Since |X rX1 | = n(d+1)
s, from above z conv(X r X1 ). Again, by the Caratheodory theorem, there exists
X2 (X r X1 ) such that z conv(X2 ). Repeating this r times we find r disjoint
(d + 1)-element subsets X1 , . . . , Xr X whose convex hulls contain z.

2.4. Proof of Theorem 2.2. Let r = n/(d + 1)2 . By Theorem 2.5, there exist
disjoint subsets X1 , . . . , Xr X whose convex hulls intersect, i.e., contain some
point z conv(X). Now, by the colorful Caratheodory theorem (Theorem 2.4), for
every (d + 1)-element subset J = {j1 , . . . , jd+1 } [r] there exists a rainbow simplex
J = (y1 , . . . , yd+1 ), yi XJi , containing point z. In other words, if we think of [r]
as the set of different colors, then there exists a J-colored simplex containing z, for
every subset J of d + 1 colors. Therefore, the total number of such simplices is






r
n/(d + 1)2
n
=
> d
,
d+1
d+1
d+1
as desired. 

23

2.5. Exercises.
Exercise 2.1. [1-] Check that the argument below Theorem 2.1 produces a (face-to-face)
triangulation in R3 . Explain what can go wrong in R4 .
Exercise 2.2. (Dual SylvesterGallai theorem) [1] Let L be a nite set of lines in the plane,
not all going through the same point. Prove that there exists a point x R2 contained in
exactly two lines.
Exercise 2.3. a) [1] Let L be a nite set of lines in the plane, not all going through the
same point. Denote by pi the number of points which lie in exactly i lines, and by qi the
number of regions in the plane (separated by L) with i sides. Prove:
X
X
(3 i)pi +
(3 i)qi = 3
i2

i3

Conclude from here that p2 3, implying the dual SylvesterGallai theorem.

b) [1] Use combinatorial duality points lines to similarly show that for a nite set of
points, not all on the same line, there exists at least three lines, each containing exactly two
points.
Exercise 2.4. (GrahamNewman problem) [2-] Let X R2 be a nite set of points, not all
on the same line. Suppose the points are colored with two colors. A line is called monochromatic if all its points in X have the same color. Prove that there exists a monochromatic
line containing at least two points in X.
Exercise 2.5. [1] Let X = {x1 , . . . , xn } R2 be a nite set of points, not all on the same
line. Suppose real numbers a1 , . . . , an are associated with the points such that the sum
along every line is zero. Prove that all numbers are zero.
Exercise 2.6. [1] Let X1 , . . . , Xn Rd be any convex sets, and z conv(X1 . . . Xn ).
Then there exist a subset I [n], |I| = d + 1, such that z conv (iI Xi ).
Exercise 2.7. [1+] Let P Rd be a simple polytope with m facets. Fix n = m d vertices
v1 , . . . , vn . Then there exists a vertex w and facets F1 , . . . , Fn P , such that w, vi
/ Fi ,
for all 1 i n.
Exercise 2.8. (Steinitz ) a) [1+] Let X Rd be a nite set of points, let P = conv(X) and
let z P r P be a point in the interior of P . Then there exists a subset Y X, |Y | 2d,
such that z Q r Q is the point in the interior of Q = conv(Y ).6
b) [1-] Show that the upper bound |Y | 2d is tight.
Exercise 2.9. a) [2-] Let X1 , . . . , Xn be convex sets in Rd such that the intersection of
every 2d of them has volume 1. Then the intersection of all Xi has volume c, where
c = c(d) > 0 is a universal constant which depends only on d.
b) [1-] Check that the number 2d cannot be lowered.
Exercise 2.10. a) [1-] Suppose points x1 , x2 and x3 are chosen uniformly and independently at random from the unit circle centered at the origin O. Show that the probability
that O (x1 x2 x3 ) is equal to 1/4.
b) [1+] Generalize part a) to higher dimensions.
6Think

of this result as a variation on the Caratheodory theorem.

24

Exercise 2.11. [1] Let X = [x1 . . . xn ] R2 be a convex polygon. Prove that there exist
n 2 points y1 , . . . , yn2 R2 such that every triangle on X contains at least one point yi .7
Show that this is impossible to do with fewer than n 2 points.
Exercise 2.12. Let P R3 be a convex polytope of volume 1. Suppose X P is a set
of n points.
a) [1-] Prove that there exists a convex polytope Q P r X such that vol(Q) n1 .
b) [1+] For every n = 3(2k 1), prove that there exists a convex polytope Q P r X such
that vol(Q) 2k .
Exercise 2.13. (Cone triangulation) [1] Let C Rd be a convex cone, dened as the
intersection of nitely many halfspaces containing the origin O. A cone in Rd is called
simple if it has d faces. Prove that C can be subdivided into convex cones.
Exercise 2.14. a) [1+] Let Q R3 be a space polygon, and let P R3 be the convex hull
of Q. Prove that every point v P belongs to a triangle with vertices in Q.
b) [2-] Extend this result to Rd and general connected sets Q.
Exercise 2.15. [2+] We say that a polygon Q R2d is convex if every hyperplane intersects
it at most 2d times. Denote by P the convex hull of Q. Prove that every point v P belongs
to a d-simplex with vertices in Q.
Exercise 2.16. Let P Rd be a convex polytope with n facets, and let X = {x1 , . . . , xn }
be a xed subset of interior points in P . For a facet F of P and a vertex xi , dene a pyramid
i (F ) = conv(F, xi ).
a) [2] Prove that one can label the facets F1 , . . . , Fd in such a way that pyramids i (Fi ) do
not intersect except at the boundaries.
b) [2] Prove that one can label the facets F1 , . . . , Fd in such way that the pyramids i (Fi )
cover the whole P .
Exercise 2.17. (Colorful Helly theorem) [2-] Let F1 , . . . , Fd+1 be nite families of convex
sets in Rd . Suppose for every X1 F1 , . . . , Xd+1 Fd+1 we have X1 . . . Xd+1 6= .
Prove that XFi X 6= for some 1 i d + 1.
Exercise 2.18. [2] Let X1 , . . . , Xd+1 Rd be nite sets of points such that O conv(Xi
Xj ), for all 1 i < j d + 1. Then there exist points x1 X1 , . . . , xd+1 Xd+1 , such
that O conv{x1 , . . . , xd+1 }.
Exercise 2.19. a) [1+] Let P, Q Rd be two convex polytopes. Prove that P Q is convex
if and only if every interval [vw] P Q for all vertices v of P and w of Q.
b) [2] Generalize part a) to a union of n polytopes.
Exercise 2.20. a) [1] A triangle is contained in a convex, centrally symmetric polygon Q.
Let be the triangle symmetric to with respect to a point z . Prove that at least
one of the vertices of lies in Q or on its boundary.
b) [1-] Extend part a) from triangles to general convex polygons.

n
1
immediately implies that at least one of yi is covered with at least n2
3 triangles. Obviously, this is much weaker than the constant proportion of triangles given by the Barany theorem.
7This

25

2.6. Final remarks. The

method of innite descent is often attributed to Pythagoras and
his proof of irrationality of 2. In the modern era it was reintroduced by Euler in his proof
of Fermats last theorem for the powers 3 and 4. For the history of the SylvesterGallai
theorem (Theorem 2.3), let us quote Paul Erd
os [Erd]:
In 1933 while reading the beautiful book Anschauliche Geometrie of Hilbert
and CohnVossen [HilC], the following pretty conjecture occurred to me: Let
x1 , . . . , xn , be a nite set of points in the plane not all on a line. Then there
always is a line which goes through exactly two of the points. I expected this
to be easy but to my great surprise and disappointment I could not nd a
proof. I told this problem to Gallai who very soon found an ingenious proof.
L. M. Kelly noticed about 10 years later that the conjecture was not new. It
was rst stated by Sylvester in the Educational Times in 1893. The rst proof
though is due to Gallai.
For more on the history, variations and quantitative extensions of the SylvesterGallai
theorem see [ErdP, BorM], [PacA, 12] and [Mat1, 4]. We return to this result in Section 12.
The method of innite descent makes another appearance in Section 9, where it has a
physical motivation.
The B
ar
any theorem (Theorem 2.2) and the colorful Caratheodory theorem (Theorem 2.4) are due to I. B
ar
any [Bar3]. For extensions and generalizations of results in this
section see [Mat1]. Let us mention that the exact asymptotic constant for 2 in the B
ar
any
theorem (Theorem 2.2) is equal to 2/9. This result is due to Boros and F
uredi [BorF] (see
Subsection 4.2 for the lower bound and Exercise 4.10 for the upper bound).

26

3. The Borsuk conjecture

3.1. The story in brief. Let X Rd be a compact set, and let diam(X) denotes
the largest distance between two points in X. The celebrated Borsuk conjecture
claims that every convex set8 with diam(X) = 1 can be subdivided into d + 1 parts,
each of diameter < 1. This is known to be true for d = 2, 3. However for large d the
breakthrough result of Kahn and Kalai showed that this is spectacularly false [KahK].
While we do not present the (relatively elementary) disproof of the conjecture, we
will prove the classical (now out of fashion) Borsuk theorem (the d = 2 case). We
will also establish the Borsuk conjecture for smooth convex bodies. Finally, we use a
basic topological argument to prove that three parts are not enough to subdivide the
sphere S2 into parts of smaller diameter. We continue with the topological arguments
in the next three sections.
3.2. First steps. Before we show that every plane convex set can be subdivided into
three parts of smaller diameter, let us prove a weaker result, that four parts suffice
for all polygons.
Proposition 3.1. Let X R2 be a convex polygon with diam(X)
= 1. Then X can
q
be subdivided into 4 disjoint convex bodies with diameter 23 .

Proof. Let X be a convex polygon with vertices z1 , . . . , zn and diam(X) = 1. From

Corollary 1.8 we know that the points zi , and hence the whole polygon, can be covered
by a circle C of radius 13 . Subdividing the circle by two orthogonal diameters we
obtain a subdivision
q of X into four parts (see Figure 3.1). Since the diameter of each
part is at most

2
,
3

we obtain the result.

Figure 3.1. Dividing polygons into four parts of smaller diameter.

In the next subsection we show how this construction can be modified to reduce
the number of regions of smaller diameter from four to three.
8Of

course, convexity here is entirely irrelevant since taking a convex hull does not increase the
diameter. Still, we would like to keep this condition to help visualize the problem.

27

3.3. Borsuk theorem by extended meditation. We are ready to prove the Borsuk conjecture for d = 2.
Theorem 3.2 (Borsuk). Let X R2 be a convex body with diam(X) = 1. Then X
can be subdivided into 3 disjoint convex bodies with diameter (1 ), for some
fixed > 0 independent of X.
Let us try to modify the idea in the proof of Proposition 3.1 above. To obtain a
legitimate proof one has to deal with two problems:
(I) we need to extend Corollary 1.8 to infinite sets of points;
(II) we need to subdivide our disk into three parts of smaller diameter.
There are several simple ways to resolve the problem (I). First, we can go to the
proof of the corollary, starting with the Helly theorem, and prove that it holds for
infinitely many convex sets. We will leave this check as an exercise to the reader (see
Exercise 1.1).
Alternatively, one can consider a sequence of unit disks obtained by adding rational
points in the body one by one. Since they lie in a compact set, their centers have to lie
at a distance 1 from the first point, and there exists a converging subsequence. The
limiting disk of that subsequence is the desired covering unit disk (see Exercise 7.1).
In a different direction, one can also consider polygon approximations of the convex
body X (see Figure 3.2) to obtain a covering disk of radius (1 + ) diam(X). If the
proof of the second part is robust enough, as we present below, this weaker bound
suffices.

X
K
Figure 3.2. Resolving problems (I) and (II).
Now, getting around problem (II) may seem impossible since the equilateral triangle
inscribed into a circle C of radius 13 has side 1. On the other hand, such a circle C
is almost enough. Indeed, if X is covered with a truncated disk K as shown in
Figure 3.2, no matter how small the shaded segment is, the inscribed equilateral
triangle has side < 1, and the subdivision as in the figure immediately proves the
Borsuk theorem.
Finally, making two opposite sectors in disk C small enough we can make sure that
the distance between them is > (1 + ), and thus the set X does not intersect at least
one of two sectors. In other words, the set X is covered with a truncated disk, which
completes the proof of the Borsuk theorem. 

28

3.4. Making it smooth helps. As it turns out, for smooth convex bodies the Borsuk
conjecture is much easier, the reason being the simple structure of points at distance 1.
Let us present a simple and natural argument proving this.
Theorem 3.3 (Hadwiger). Let X Rd be a smooth convex body of diam(X) = 1.
Then X can be subdivided into d + 1 disjoint bodies of diameter < 1.
Proof. Let U = X be the surface of X, and let S be the surface of a unit ball. We
say that two points u, v U are opposite if the tangent hyperplanes to u and v are
orthogonal to the interval (u, v).
Observe that it suffices to subdivide X in such a way that no part contains a pair
of the opposite points. Clearly, the distance between any two points, at least one of
which is in the interior in X, is strictly smaller than 1. On the other hand, if the
tangent plane Tu is not orthogonal to (u, v), then the distance can be increased locally
as in Figure 3.3. Thus the distance |uv| < 1 unless u and v are opposite.
u

Figure 3.3. If the tangent line T is not orthogonal to (uv), then the
distance |uv| can be increased locally: |vw| > |vu|.
Now consider a continuous map f : U Sd1 defined by the condition that tangent
surfaces to u U and s = f (u) Sd1 have the same normals (see Figure 3.4).
Subdivide Sd1 into d + 1 parts such that no part contains two opposite points. One
possible way to do this is take a small cap around the North Pole and cutting along
meridians to create equal size segments.9 Use f 1 to pull the subdivision back onto U
(see Figure 3.4).

f 1

f
U

Figure 3.4. Map f : U Sd1 , the subdivision of S and U.

By construction, the opposite points of U are mapped into the opposite points
of Sd1 . Therefore, no part of the subdivision of U contains opposite points. Now
9Alternatively,

its facets.

inscribe a regular d-dimensional simplex into Sd1 and take central projections of

29

subdivide X by taking a cone from any interior point O X onto each part of the
subdivision of U. From the observation above, only the opposite points can be at
distance 1, so the diameter of each part in X is < 1.

3.5. Do we really need so many parts? It is easy to see that many convex polytopes in Rd , such as hypercubes and cross-polytopes (generalized octahedra), can be
subdivided into just two parts of smaller diameter. In fact, this holds for every centrally symmetric convex polytope (see Exercise 3.2). On the other hand, we clearly
need at least d + 1 parts to subdivide a regular simplex in Rd since no part can
contain two of the vertices of . Now, what about a unit ball? From the proof of
Hadwigers Theorem 3.3, any improvement in the d + 1 bound would imply the same
bound for all smooth bodies. The classical Borsuk-Ulam theorem implies that this is
impossible for any d. For d = 2 this is easy: take a point on the boundary between
two parts on a circle and observe that the opposite point cannot belong to either of
the parts then. We prove this result only for d = 3.
Proposition 3.4 (3-dim case). A unit ball B R3 cannot be subdivided into three
disjoint bodies of smaller diameter.
Proof. From the contrary, consider a subdivision B = X Y Z. Consider the
sphere S2 = B and the relative boundary = (X S2 ). Observe that is a union
of m closed non-intersecting curves, for some m 1. Let be the set of points
opposite to . Since = , the set consists of 2m closed non-intersecting
lines. Note (or prove by induction) that Q = S2 r ( ) consists of (2m + 1) regions
of connectivity, some of which are centrally symmetric to each other and thus come in
pairs (see Figure 3.5). Since Q is itself centrally symmetric, by the parity it contains
a centrally symmetric region U Q. Connect a point u U by a path in U to its
opposite u, and consider a closed path obtained as a union of with the opposite
path , i.e., let = . Map onto a circle S1 such that the opposite points in
are mapped onto the opposite points in S1 . Since Y Z, the problem is reduced
to the d = 2 case.

X

Q
S2

S2

Figure 3.5. Boundary = X S2 and a region Q = S2 r ( ).

30

3.6. Exercises.
Exercise 3.1. [1-] Prove that every convex polygon X of area 1 can be covered by a
rectangle of area 2.
Exercise 3.2. a) [1] Prove that centrally symmetric polytopes in Rd can be subdivided
into two subsets of smaller diameter.
b) [1] Prove the Borsuk conjecture for centrally symmetric convex sets X Rd .
Exercise 3.3. [1-] Prove that every set of n points in R3 with diameter , can be covered
by a cube with side length .
Exercise 3.4. (P
al) a) [1+] Prove that every convex set of unit diameter can be covered
by a hexagon with side 1.
b) [1-] Prove that every
convex set of unit diameter can be subdivided into three convex
sets of diameter at most 23 .
c) [2-] Prove that every planarset of n points of unit diameter can be partitioned into three
2
subsets of diameter less than 23 cos 3 n(n1)
.
Exercise 3.5. (Borsuk conjecture in R3 ) [2] Prove the Borsuk conjecture for convex sets
X R3 .
Exercise 3.6. (Convex sets of constant width) Let X Rd be a convex set. Clearly,
for every vector u there exists two supporting hyperplanes H1 , H2 with u as normals. The
distance between H1 and H2 is called the width in direction u. We say that X has constant
width width(X) if it has the same width in every direction. An example of a convex set
with constant width r is a ball of radius r/2. In the plane, another important example is
the Reuleaux triangle R obtain by adding three circular segments to an equilateral triangle
(see Figure 3.6).

Figure 3.6. Reuleaux triangle R.

a) [1] Check that width(X) = diam(X) for every set X Rd with constant width. Conclude that for every set X there exists a convex set Y X of constant width with the same
diameter: diam(Y ) = diam(X).
b) [1+] Prove that the perimeter of every region X R2 with constant width 1 is equal
to . Conclude from here that in the plane the circle has the largest area among convex
sets of given constant width.
c) [2+] Prove that in the plane the Reuleaux triangle R has the least area among convex
sets of given constant width.
d) [1-] Find two convex bodies in R3 with constant width 1 and dierent surface area.

31

Exercise 3.7. [1] Let X R2 be a body of constant width. Observe that every rectangle
circumscribed around X is a square (see Subsection 5.1). Prove that the converse result is
false even if one assumes that X is smooth.
Exercise 3.8. (Covering disks with disks) Denote by Rk the maximum radius of a disk
which can be covered with k unit disks.

a) [1] Prove that R2 = 1, R3 = 23 and R4 = 2.

b) [1] Prove that R5 > 1.64.
c) [2-] Find the exact value for R5 .
d) [1+] Prove that R7 = 2.
e) [1+] Prove that R9 = 1 + 2 cos 2
8 .
Exercise 3.9. A nite set of points A R2 is said to be coverable if there is a nite
collection of disjoint unit disks which covers A. Denote by N the size of the smallest set A
which is not coverable.
a) [1] Prove that N < 100.
b) [1] Prove that N > 10.
3.7. Final remarks. Recall that our proof of the Borsuk conjecture in dimension 2 is
robust, i.e., produces a subdivision into three parts, such that each of them have diam
(1 ), for some universal constant > 0. Although not phrased that way by Borsuk,
we think this relatively small extension makes an important distinction; we emphasize the
robustness in the statement of Theorem 3.2. Dimension 3 is the only other dimension in
which the Borsuk conjecture has been proved, by Eggleston in 1955, and that proof is also
robust (a simplied proof was published by Gr
unbaum [BolG]). However, our proof of the
Hadwiger theorem (Theorem 3.3) is inherently dierent and in fact cannot be converted
into a robust proof. We refer to [BMS, 31] for an overview of the subject, various positive
and negative results and references.
The extended meditation proof of the Borsuk theorem is due to the author, and is similar
to other known proofs. The standard proof uses P
als theorem (see Exercise 3.4) that every
such X can be covered by a regular hexagon with side 1, which gives an optimal bound on
(see [BolG, HDK]). The proof in the smooth case is a reworked proof given in [BolG, 7].
The proof in Proposition 3.4 may seem simple enough, but in essence it coincides with the
early induction step of a general theorem. For this and other applications of the topological
approach see [Mat2].
The high-dimensional counterexamples of KahnKalai [KahK] (to the Borsuk conjecture)
come from a family of polytopes spanned by subsets of vertices of a d-dimensional cube,
whose vertices are carefully chosen to have many equal distances. We refer to [AigZ, 15] for
an elegant exposition. In addition, we recommend a videotaped lecture by Gil Kalai [Kal2]
which gives a nice survey of the current state of art with some interesting conjectures and
promising research venues.

32

4. Fair division
In this section we begin our study of topological arguments, which is further continued in the next two sections. Although the results in this section are largely
elementary, they have beautiful applications.
4.1. Dividing polygons. Let Q R2 be a convex polygon in the plane. An equipartition is a subdivision of Q by lines into parts of equal area. Here is the first basic
equipartition result.
Proposition 4.1 (Equipartition with two lines). For every convex polygon Q R2
there exist two orthogonal lines which divide Q into four parts of equal area.
Proof. Fix a line R2 and observe that by continuity there exists a unique line 1 k
which divides Q into two parts of equal area. Similarly, there exists a unique line 2
which divides Q into two parts of equal area. Now lines 1 and 2 divide Q into four
polygons, which we denote by A1 , A2 , A3 and A4 (see Figure 4.1). By construction,
area(A1 ) + area(A2 ) = area(A2 ) + area(A3 ) = area(A3 ) + area(A4 )
1
= area(A4 ) + area(A1 ) = area(Q),
2
which implies that area(A1 ) = area(A3 ) and area(A2 ) = area(A4 ). Now rotate line
continuously, by an angle of /2. By the uniqueness, we obtain the same division of Q
with two lines, with labels Ai shifted cyclically. Thus, the function area(A1 )area(A2 )
changes sign and is equal to zero for some direction of . Then, the corresponding
lines 1 and 2 give the desired equipartition.

2
1

A1
A4

r1

A2

r2

r2
z

A3
Q

r2

r1

r2

Figure 4.1. Equipartitions with two and three lines.

The next result is even more impressive as it shows existence of an equipartition
of the polygon into six parts with three concurrent lines (lines that are distinct and
have a common point). Let us note that it is impossible to make an equipartition of
a convex polygon with three lines into seven parts (see Exercise 4.7).
Theorem 4.2 (Equipartition with three lines). For every convex polygon Q R2
there exist three concurrent lines which divide Q into six parts of equal area.

33

Proof. Fix a line R2 and take 1 k which divides Q into two parts of equal area.
For every point z 1 there there is a unique collections of rays r2 , r3 , r2 and r3 which
start at z and together with 1 divide Q into six parts of equal area (see Figure 4.1).
Move z continuously along 1 . Observe that the angle defined as in the figure,
decreases from to 0, while angle increases from 0 to . By continuity, there exists
a unique point z such that r2 and r2 form a line. Denote this line by 2 . As in the
previous proof, rotate continuously, by an angle of . At the end, we obtain the
same partition, with the roles of r3 and r3 interchanged. Thus, the angle between
lines spanned by r3 and r3 changes sign and at some point is equal to zero. Rays r3
and r3 then form a line which we denote by 3 . Then, lines 1 , 2 and 3 gives the
desired equipartition.

4.2. Back to points, lines and triangles. The following result is a special case of
the Barany theorem (Theorem 2.2), with an explicit constant. We obtain it as easy
application of the equipartition with three lines (Theorem 4.2).
Theorem 4.3 (BorosF
uredi). Let X R2 be a set of n = 6k points in general
position. Then there exists a point z contained in at least 8k 3 triangles with vertices
in X.


Note that 8k 3 = 29 n3 + O(n2 ), which gives an asymptotic constant 2 = 29 in the
Barany theorem. In fact, this constant cannot be improved (see Exercise 4.7). The
proof is based on the following variation on Theorem 4.2.
Lemma 4.4. Let X R2 be a set of n = 6k points in general position. Then there
exist three intersecting lines which separate X into six groups with k points each.
The proof of the lemma is similar to the proof of Theorem 4.2, and takes into
account that the points are in general position (see Exercise 4.9).
Proof of Theorem 4.3. Denote by A1 , . . . , A6 the regions in the plane divided by three
lines as in Lemma 4.4 (see Figure 4.2). Let z be the intersection point. Consider what
type of triangles must contain z. First, all triangles with one vertex in each of A1 , A3
and A4 , must contain z. The same is true for regions A2 , A4 and A6 . This gives 2k 3
triangles containing z.
A2
A1

A6

A3
A4

A5

A2

A1

z
A6

A3
A4

A5

Figure 4.2. Two types of triangles containing z.

34

Similarly, for every two vertices in the opposite regions A1 and A4 there exist at
least 2k ways to form a triangle which contains z, with either all vertices in A2 A3
or with all vertices in A4 A5 (see Figure 4.2). This gives at least 2k 3 triangles
containing z. Since the same argument holds for the other two pairs of opposite
regions, we get at least 6k 3 triangles of this type, and 8k 3 in total.

4.3. Inscribed chords. Let f : [0, 1] R be a continuous function with f (0) =
f (1) = 0. Consider two points x, y [0, 1] such that y x = and f (x) = f (y).
We call the chord [x, y] the inscribed chord of length . Now, is it true that for
every 0 < < 1 there exists an inscribed chord of length ? While the answer to
this question is easily negative (see Figure 4.3), there is a surprising connection to
topological arguments earlier in this section.
g

f
0

1
2

1
3

1
2

2
3

  

 
Figure 4.3. Functions f and g with D(f ) = 0, 12 and D(g) = 0, 13 12 , 32 .

Let us first show that there is always an inscribed chord of length 12 . Indeed, if
f ( 12 ) = 0, we found our inscribed chord. Suppose now that f ( 12 ) > 0, and consider
g(x) = f (x + 21 ) f (x), where x [0, 12 ]. Since g(x) is continuous, g(0) > 0, and
g( 21 ) < 0, we conclude that g(z) = 0 for some z [0, 12 ]. Thus, [z, z + 12 ] is the desired
inscribed chord of length 21 .
Now that we know that some distances always occur as lengths of inscribed chords,
let us restate the problem again. Denote by D(f ) the set of distances between points
with equal values:


D(f ) = y x | f (x) = f (y), 0 x y 1 .

Observe that for a given f , the set D(f ) contains all distances small enough. However,
even for very small > 0, there exists a distance < and a function f , such that

/ D(f ) (see Exercise 4.2). In fact the only distances guaranteed to be in D(f ) are
given by the following theorem.
Theorem 4.5 (Inscribed chord theorem). For every continuous function f : [0, 1] R
with f (0) = f (1) = 0, we have n1 D(f ), for all n N.
In other words, there always exist inscribed chords of length 1/n. The theorem
follows easily from the following attractive lemma.
Lemma 4.6. For every a, b
/ D(f ) and f as in the theorem, we have a + b
/ D(f ).

Proof of Theorem 4.5. Assume n1

/ D(f ). By the lemma, we have n2
/ D(f ). Simin
3
larly, n
/ D(f ), etc. Thus, n = 1
/ D(f ), which is impossible since f (0) = f (1). 

35

Proof of Lemma 4.6. Denote by G the graph of function f , and by G the graph G
shifted by , where R. Attach to G two vertical rays: at the first (global) maximum
of f pointing up and at the first (global) minimum pointing down (see Figure 4.4).
Do the same with G . Now observe that D(f ) is equivalent to having G and G
intersect. Since a, b
/ D(f ), then graph G does not intersect graphs Ga and Gb . By
construction, graph G divides the plane into two parts. We conclude that graphs Ga
and Gb lie in different parts, and thus do not intersect. By the symmetry, this implies
that G and Ga+b do not intersect, and therefore a + b
/ D(f ).

G

G
1

+1

+ 1

1
G

Figure 4.4. Graphs G, G and G , where D(f ) and G G 6= ,

while
/ D(f ) and G G = .
4.4. Helping pirates divide the loot fairly. Imagine two pirates came into possession of a pearl necklace with white and black pearls. We assume that the number
of pearls of each color is even, so each pirate wants exactly half the white pearls and
half the black pearls. Can they cut the necklace in just two places (between pearls) so
that the resulting two pieces satisfy both pirates? Surprisingly, this is always possible.
An example of such fair division is shown in Figure 4.5. The following result extends
this to k pirates.

Figure 4.5. Fair division of a necklace with two cuts.

Theorem 4.7 (Splitting necklaces). Suppose a pearl necklace has kp white pearls
and kq black pearls, for some integers k, p and q. Then one can cut the necklace in
at most 2(k 1) places so that the remaining pieces can be rearranged into k groups,
with p white and q black pearls in each group.
In fact, we show that each of the k pirates gets either one or two continuous pieces
of the necklace. Note also that the bound on the number of cuts in the theorem is
tight: if black pearls are separated from white pearls, at least k 1 cuts are needed
to divide pearls of each color.

36

Proof. Denote by n = p + q the number of pearls each pirate must receive. Fix
a starting point 0 and an orientation of the necklace. Let a(x) and b(x) denote
the number of white and black pearls among the first x pearls after 0. Consider a
discrete function f : {0, 1, . . . , k n} Z, defined as f (x) = a(x)/p b(x)/q. Extend f
linearly to the whole interval [0, k n]. Observe that f (0) = f (k n) = 0. Apply the
inscribed chord theorem (Theorem 4.5) to obtain x, y [0, k n] such that y x = n
and f (y) f (x) = 0. Since the number of pearls between x and y must be integral,
we can round x, y down and obtain a fair k1 portion of the necklace for the first pirate.
1
Repeat the procedure by cutting out a fair k1
portion of the remaining necklace,
etc. At the end, we have k 1 pirates who make two cuts each, giving the total of
2(k 1) cuts.

4.5. Exercises.
Exercise 4.1. [1-] Let f : R R be a periodic continuous piecewise linear function. Prove
that f has inscribed chords of any length.
Exercise 4.2. [1+] For every
/ {1, 12 , 13 , . . .} nd a continuous function f : [0, 1] R
with f (0) = f (1) = 0, and such that
/ D(f ). In other words, prove that Theorem 4.5
cannot be extended to other values.
Exercise 4.3. [1+] In the conditions of Theorem 4.5, prove that for every integer n there
are at least n inscribed chords whose lengths are multiples of 1/n.
Exercise 4.4. a) [1-] Consider two convex polygons in the plane. Prove that there exists
a line which divides both of them into halves of equal area.
b) [1-] Same with two halves of equal perimeter.
c) [1] Consider three convex polytopes in R3 . Prove that there always exist a plane which
divides each of the three polytopes into halves of equal volume.
d) [1] Prove or disprove: there always exist a plane which divides each of the three polytopes
into halves with equal surface area.
Exercise 4.5. a) [1+] Let Q1 , . . . , Qm R2 be convex polygons in the plane with weights
w1 , . . . , wm R (note that the weights can be negative). For a region B R2 dene the
weighted area as w1 area(B Q1 ) + . . . + wm area(B Qm ). Prove that there exist two
orthogonal lines which divides the plane into four parts of equal weighted area.
b) [1+] Generalize Theorem 4.2 to weighted areas.
Exercise 4.6. [2-] Let Q be a convex polygon in the plane. A line is called a bisector if
it divides Q into two parts of equal area. Suppose there exists a unique point z Q which
lies on at least three bisectors (from above, there is at least one such point). Prove that z
is the center of symmetry.
Exercise 4.7. [1] Prove that for every convex polygon Q R2 , no three lines can divide
it into seven parts of equal area.
Exercise 4.8. [1] Let Q R2 be a convex polygon. A penta-partition of Q is a point
z Q and ve rays starting at z, which divide R2 into ve equal cones, and divide Q into
ve polygons of equal area. Prove or disprove: every Q as above has a penta-partition.
Exercise 4.9. [1] Prove Lemma 4.4.

37

Exercise 4.10. [1+] Give an example showing that the 2 =

optimal (see Theorem 2.2 and Theorem 4.3).

2
9

in B
ar
anys theorem is

Exercise 4.11. [2] Let Q R2 be a convex polygon in the plane. Prove that there exists
a convex quadrilateral X = [x1 x2 x3 x4 ] such that X together with lines (x1 x3 ) and (x2 x4 )
divide Q into eight regions of equal area (see Figure 4.6).10 Generalize this to polygons Qi
and weighted area (see Exercise 4.5).

x2

x1

x3
x4

Figure 4.6. Cutting a polygon into eight regions of equal area.

Exercise 4.12. [2-] Let P R3 be a convex polytope. Prove that there exist three
orthogonal planes, such that every two of them divide P into four parts of equal volume.
Exercise 4.13. a) [1+] Suppose a necklace has 2p white, 2q black, and 2r red pearls. Prove
that three cuts suce to divide the necklace fairly between two pirates.
b) [1-] Check that two cuts may not be enough.
c) [2] Prove that for k pirates and s 2 types of pearls, s(k 1) cuts always suce.
Exercise 4.14. [1] Two pirates found a necklace with pearls of n dierent sizes, and each
size comes in two colors: black and white. They observed that the sizes of white pearls
increase clockwise, while the size of black pearls increase counterclockwise. Pirates want
to divide the necklace so that each gets a pearl of every size. Prove that two cuts always
suce.
Exercise 4.15. (Two integrals theorem) Let f, g : [0, 1] [0, 1] be two continuous funcR1
R1
tions such that 0 f = 0 g = 1.
R
R
a) [1] Prove that there exist 0 < 1, such that f = g = 1/2.
R
R
b) [1+] Suppose for some 0 < a < 1 and every 0 < 1, either f 6= a or g 6= a.
R
R
Prove that there exist 0 < 1, such that f = g = 1 a.
R
R
c) [1] Prove that for every n N, there exist 0 < 1, such that f = g = 1/n.
4.6. Final remarks. The equipartition with two lines goes back to [CouR, 6.6]. Theorem 4.2 is proved in [Buck] (see also Exercises 4.5 and 4.7). For the background, generalizations and references on classical equipartition results see [Gr
u2, 4.2].
10This

is called cobweb equipartition.

38

The BorosF
uredi theorem is proved in [BorF]. Our proof follows a recent paper [Bukh].
For a matching lower bound, the original proof has been shown to be false, and a complete
proof is given in [BMN] (see Exercise 4.10).
The inscribed chord theorem (Theorem 4.5) is usually attributed to Levy (1934). It was
pointed out in [Fle] that the result was rst discovered by Amp`ere in 1806. The proof we
present is due to Hopf [Hop1]. Our presentation follows [Lyu, 34].
Theorem 4.7 is due to Goldberg and West (1985), and was further generalized a number
of times. Notably, Alon showed in [Alon] that for k pirates and s 2 types of pearls,
s(k 1) cuts suce (see Exercise 4.13). We refer to [Mat2, 3.2, 6.6] for further results,
proofs and references. Our presentation is a variation on several known proofs and was
partly inuenced by [Tot].

39

5. Inscribed and circumscribed polygons

In this section we study the inscribed figures in planar piecewise linear curves. Our
main goal is the resolution of the square peg problem, but along the way we prove a
number of variations and special cases. One notable tool is the mountain climbing
lemma, an elementary result of independent interest and wide applicability. In the
next section, we explore space polygons inscribed into surfaces in R3 . The results of
this section make a reappearance in Sections 21, 23 and 24, when we start a serious
study of the geometry of plane and space curves.
5.1. Where it all begins. We start with the following easy result which gives a
flavor of results both in this and in the next section.
Let X R2 be a convex set in the plane. We say that X has a circumscribed
square W if X lies inside W and every edge of W contains at least one point x X
(see Figure 5.1).
Proposition 5.1. Every convex set X R2 has a circumscribed square.

Proof. For every unit vector u R2 denote by f (u) the distance between two lines
orthogonal to u and supporting X (see Figure 5.1). Let g(u) = f (u)f (u ), where u
is orthogonal to u. Observe that g(u) changes continuously as u changes from v to
v , for any v v . Since g(v ) = g(v ), for some u we have g(u) = 0. Now X is
inscribed into a square bounded by the four lines orthogonal to u and u .


f (u)
u

Figure 5.1. Polygons with circumscribed squares; width f (u) in direction u.

Consider now a superficially similar problem of squares inscribed into simple polygons, often called the square peg problem. We say that a closed curve C R2 has an
inscribed square if there exists four distinct points on the curve which form a square
(see Figure 5.2). Does the inscribed square always exist? If C is a general Jordan
curve, this is a classical problem, open for nearly a century. However, if the curve is
nice enough in a sense that it is either piecewise linear or smooth, or has a certain
degree of regularity, then it does have an inscribed square (see Subsection 5.8 for the
references).
Many results in this section were obtained as an attempt to resolve the inscribed
square problem and are stated for simple polygons.11 Eventually we prove that every
simple polygon has an inscribed square (see Subsection 5.6). Let us note that this
11While

some of these results easily extend to general Jordan curves, others do not, often for
delicate reasons. We hope the reader enjoys nding them.

40

Figure 5.2. Jordan curve C and an inscribed square.

implies the claim for self-intersecting closed polygons as well, since taking any simple
cycle in it suffices.
In a different direction, one can ask if Proposition 5.1 generalizes to higher dimensions. The answer is yes, but the proof is more delicate. In three dimensions this is
called the Kakutani theorem; we prove it in the next section. In fact, much of the
next section is based on various modifications and generalizations of the Kakutani
theorem.
5.2. Inscribing triangles is easy. Let X R2 be a simple polygon in the plane,
and let A R2 denote the region enclosed by X, i.e., X = A. We say that an
equilateral triangle is inscribed into X if there exist three distinct points y1 , y2, y3 X
such that |y1y2 | = |y1 y3 | = |y2y3 |.
Proposition 5.2. For every simple polygon X = [x1 . . . xn ] R2 and a point z X
in the interior of an edge in X, there exists an equilateral triangle (y1 y2 z) inscribed
into X. The same holds for every vertex z = xi with xi1 xi xi+1 /3.
First proof. Denote by X the clockwise rotation of X around z by angle /3. Clearly,
area(X) = area(X ). Thus for every z in the interior of an edge in X, polygons X
and X intersect at z and at least one other point v. Denote by u the counterclockwise
rotation of v around z by /3. Then (z, u, v) is the desired inscribed triangle.
The same argument works when z is a vertex as in the proposition. Indeed, if
z = xi and xi1 xi xi+1 /3, then the interiors of the corresponding regions X
and X intersect.

Second proof. Suppose polygon X is oriented clockwise. For every point u X,
consider an equilateral triangle (zuv) oriented clockwise. Observe that for u placed
further on the same edge as z we can make v A, while for u at maximal distance
from z we have either v X (in which case we are done), or v
/ A. The latter
case follows by continuity of v, and completes the proof of the first part. The same
argument works when z is a vertex with xi1 xi xi+1 /3.

Note that the first proof is more natural and gives an explicit construction of an
inscribed triangle. However, the second proof is more amenable to generalizations, as
can be seen in the following 3-dimensional generalizations of Proposition 5.2.

41

y2

u
X

v
y3

y1

z
X

Figure 5.3. Inscribed equilateral triangle (y1 y2 y2 ) into a polygon X

and its two constructions.
Theorem 5.3. For every simple space polygon X R3 and a point z X in the
interior of an edge in X, there exists an equilateral triangle (y1 y2 z) inscribed into X.
The same holds for every vertex z = xi with xi1 xi xi+1 /3.
Proof. For every u X, denote by C(u) the circle of all points v R3 , so that (zuv)
is an equilateral triangle (see Figure 5.4). Observe that when u is close to z the
circle C(u) is linked with X, and when u is at maximal distance from z, C(u) is not
linked with X. Therefore, C(u0 ) intersects with X, for some u0 X. The intersection
point v0 together with u0 and z form the desired equilateral triangle. The case of a
vertex z = xi is similar.

u
C(u)
u

C(u)

z
X

z
X

Figure 5.4. Space polygon X and two circles C(u) for different u.

5.3. Need to inscribe rectangles? Topology to the rescue. We say that X has
an inscribed rectangle if there exist four distinct points x1 , . . . , x4 X which form a
rectangle.
Proposition 5.4. Every simple polygon X R2 has an inscribed rectangle.
First proof. Think of X = [x1 . . . xn ] R2 as lying on a horizontal plane in R3 . For
every two points u, v X let h(u, v) be a point at height |uv| which projects onto

42

the midpoint of (u, v):

h(u, v) =

!

Ou + Ov |uv|
,
R3 .
2
2

By construction, h(u, v) forms a piecewise linear surface H which lies above X and
has X as a boundary (see Figure 5.5).
Assume first that H is self-intersecting. This means that there exist points u, v, u
and v in X, such that |uv| = |u v | and the midpoints of (u, v) and (u , v ) coincide.
Then [uvuv ] is the desired inscribed rectangle.

Figure 5.5. Triangles and parallelograms in the surface H.

Suppose now that H is embedded, i.e., not self-intersecting. Attach the interior
of X to H and denote by S the resulting polyhedral surface. Observe that S has one
n-gonal face, n triangles corresponding to edges in X, and n(n 1)/2 parallelograms
corresponding to unordered pairs of these

edges (see Figure 5.5). Therefore, the
n
surface S has the total of k = 1 + n + 2 faces, and


1
n(n 1)
m =
n+3n+4
= n2 + n edges.
2
2
On the other hand,

the vertices of S are xi and points h(xi , xj ), which gives the total
n
of N = n + 2 vertices. Therefore, the Euler characteristic of the surface S is
equal to: = N m + k = 1, which implies that S is non-orientable. Thus, S cannot
be embedded into R3 , a contradiction.

While the counting argument in the second part of the proof is elementary, it can
be substituted by the following even more straightforward topological argument.
Second proof. Let H and S be as in the first proof. Observe that H is homeomorphic
to the space of unordered pairs of points u, v X. Since the space of ordered pairs is
a torus, its quotient space H is homeomorphic to the Mobius strip with the polygon X
(corresponding to pairs with u = v) as its boundary (see Figure 5.6). Therefore, the
surface S is homeomorphic to RP2 and cannot be embedded into R3 . This shows
that H is self-intersecting, which implies the result.


43

(u, v)
(v, u)

Figure 5.6. Surface H as a quotient space of a torus; H is homeomorphic to a Mobius strip with X as a boundary.
5.4. Climbing mountains together. Suppose two climbers stand on different sides
at the foot of a two-dimensional (piecewise linear) mountain. As they move toward
the top of the mountain, they can move up and down; they are also allowed to move
forward or backtrack. The question is whether they can coordinate their movements
so they always remain at the same height and together reach the top of the mountain.
We will show that the answer is yes and give two elementary proofs. However, let us
mention here that the problem is not as simple as it might seem at first. For example,
for the mountain as in Figure 5.7 the first climber needs to move along intervals
(1, 2, 3, 4, 5, 6, 6, 5, 4, 3, 2, 2, 3, 4, 5, 6, 7) ,
while the second climbers moves along intervals
(1 , 2 , 3 , 3 , 3, 4 , 5 , 6 , 6 , 6, 7 , 8 , 9 , 9, 9 , 10 , 11) .
In other words, the first climber needs to first go almost all the way up, then almost
all the way down, and finally all the way up just to be on the same level with the
second climber at all times.

7 11
6
10
3
2
1

4 5

5 4
6
3
9
2
8 7
1

7
6
5
4
3
2
1
1 2 3 4 5 6 7 8 9 10 11

Figure 5.7. Two mountain climbers.

Theorem 5.5 (Mountain climbing lemma). Let f1 , f2 : [0, 1] [0, 1] be two continuous piecewise linear functions with f1 (0) = f2 (0) = 0 and f1 (1) = f2 (1) = 1. Then
there exist two continuous piecewise linear functions g1 , g2 : [0, 1] [0, 1], such that
g1 (0) = g2 (0) = 0, g1 (1) = g2 (1) = 1, and
f1 (g1 (t)) = f2 (g2 (t)) for every t [0, 1].

44

The mountain climbing lemma is a simple and at the same time a powerful tool.
We will use it repeatedly in the next subsection to obtain various results on inscribed
polygons.
First proof. Consider a subset A [0, 1]2 of pairs of points at the same level:
A = {(t1 , t2 ) | f1 (t1 ) = f2 (t2 )}.

By the assumptions, (0, 0), (1, 1) A. We need prove that (0, 0) and (1, 1) lie in the
same connected component of A. This follows from the fact that (0, 1) and (1, 0) lie
in different connected components of [0, 1]2 r A. Indeed, if one climber is going up
the mountain and the other is going down, there is a time t when they are at the
same level.

Second proof. Let us first assume that f1 and f2 are generic, i.e., all peaks and valleys
of the mountain are at different levels. Define the set A as in the first proof. Let us
show that A is a continuous piecewise linear curve from (0, 0) to (1, 1). Indeed, check
the transitions at peaks and valleys, where exactly four possibilities for a change in
the direction of A can occur (see Figure 5.8). We conclude that A is a continuous
curve with (0, 0) and (1, 1) its only possible endpoints. which proves the theorem
in the generic case. In a non-generic case, perturb two functions and use the limit
(i)
(i)
argument. Namely, consider a sequence of pairs of functions (f1 , f2 ), i = 1, 2, . . .

t2
f1

f2

A
t1

Figure 5.8. One of the four possible local changes of A.

Third proof. Consider the subset A [0, 1]2 defined in the first proof. Think of A
as a finite graph with straight edges. Observe that except for (0, 0) and (1, 1), every
vertex has degree 2 or 4, where the vertices of degree 4 appear when two corresponding
peaks of two valleys are at the same level. This immediately implies that (0, 0) and
(1, 1) lie in the same connected component of A.

Remark 5.6. Let us mention that the rst proof is misleadingly simple and suggests that
the mountain climbing lemma holds for all continuous functions. This is not true (see
Exercise 5.23), as we are implicitly using the fact that A is suciently nice. Formalizing
this observation is a good exercise which we leave to the reader.
For more than two climbers, climbing (two-dimensional) mountains of the same height,
one can still coordinate them to always remain at the same level. While the rst proof does
not work in this case, the second and third proof extend verbatim (see Exercise 5.24).
Consider the distance between two climbers in the mountain climbing lemma. Clearly,
this distance changes continuously from the maximal in the beginning down to zero. As in

45

Subsection 4.3, one can think of these as of inscribed chords. In other words, the lemma
implies that in contrast with the general mountains, where negative heights are allowed,
there exist inscribed intervals of all length, not just the integer fractions as in Theorem 4.5.

5.5. Rhombi in polygons. Let X R2 , be a simple polygon in the plane. An

inscribed rhombus is a rhombus with four distinct vertices in X.
Theorem 5.7. Every point of a simple polygon in R2 is a vertex of an inscribed
rhombus.
Proof. Let X R2 be a simple polygon, and let v be a point outside of X. Denote
by y, z X the points of minimal and maximal distance from v. Consider the distance
functions f1 and f2 from v to points x X on each of the [yz] portions of X. By
the mountain climbing lemma (Theorem 5.5), one can continuously move points x1
and x2 from y to z so they remain at equal distance from v.12 Denote by w the
fourth point of the rhombus with three other vertices v, x1 and x2 (see Figure 5.9).
Since z is at maximal distance from v, as x1 and x2 approach z, the point w is outside
of X. Similarly, assuming y is not a vertex of X and points v, x1 , x2 are close enough
to y, then w is inside X. Therefore, by continuity, for some x1 , x2 X there is a
point w X, so that Rv = [vx1 wx2 ] is a rhombus.
x1
w
v

y
x2

Figure 5.9. A construction of rhombi [vx1 wx2 ].

Now, let y be a fixed point in the interior of an edge e in X. Consider a sequence of
generic points v outside of X which converge to y. Let us show that the sequence of
rhombi Rv converges to an inscribed rhombus Ry , i.e., the rhombi do not degenerate
in the limit. Denote by X a polygon symmetric to X with respect to y. We assume
that X has no parallel edges. Then X X consists of an interval on the edge e
and finitely many points uk . One can choose points v to approach y in the direction
that is neither orthogonal to intervals (y, uk ), nor to the edge e. When the points v
approach y, we have the points x1 and x2 approach points uk , and by construction
this implies that Rv does not degenerate.
It remains to show that the inscribed rhombus exists when X has parallel edges,
or when y is a vertex. In both cases, use can now use the limit argument, by letting v
12Here

we implicitly use the fact that one can parameterize each path from y to z so that the
distance functions f1 and f2 are piecewise linear. This can be done by rst setting values for the
vertices of X (say, by assigning the edges of each path an equal length) and then by extending
the parametrization to the edges.

46

approach y. Indeed, since X is simple, the only way the rhombi do not converge to a
rhombus in the limit is when all four vertices converge to a point. Clearly, the latter
is impossible for the limit of polygons. This completes the proof.

Theorem 5.8. Every simple polygon in the plane has an inscribed rhombus with two
sides parallel to a given line.
Proof. Let X = [x1 . . . xn ] R2 be a simple polygon and let be a given line. We
assume that is not parallel to any edge of X. Denote by : R2 R a linear function
constant on . Let y and z be two vertices of X with the minimum and maximum
values of . We can assume for now that these vertices are uniquely defined, and use
the limit argument otherwise.
Use the mountain climbing lemma (Theorem 5.5) to continuously move points
u1 = u1 (t) and u2 = u2 (t) from y to z, so that (u1 (t)) = (u2 (t)), for all t [0, 1].
Reverse the time to continuously move points w1 = w1 (t) = u1 (1 t) and w2 =
w2 (t) = u2 (1 t) from z to y. Clearly, (w1 (t)) = (w2 (t)) for all t [0, 1]. Denote
by f1 (t) = |u1 (t)u2 (t)| and f2 (t) = |w1 (t)w2 (t)| the distances between these points.
Suppose f1 maximizes at time t0 . Consider points v1 = u1 (t0 ) and v2 = u2 (t0 ).
By construction, w1 (1 t0 ) = v1 and w2 (1 t0 ) = v2 . Rescale the time parameter t
so that points u1 , w1 leave y and z at t = 0, and reach v1 at t = 1. We have
f1 (0) = f2 (0) = 0, and f1 (1) = f2 (1) = |v1 v2 |. Use the mountain climbing lemma
(Theorem 5.5) to continuously move points u1 , w1 until they reach v1 , so that the
corresponding distances are equal: f1 (t) = f2 (t) for all t [0, 1]. We obtain four points
moving continuously toward v1 and v2 , while forming a parallelogram [u1 u2 w2 w1 ] with
two sides parallel to . Since the difference between side lengths g(t) = |u1 u2 | |u1w1 |
is initially negative and positive when they reach v1 and v2 , one of these parallelograms
is the desired rhombus.

z

w1

w2

v1

v2

u1

u2
X
y

Figure 5.10. Inscribed parallelogram with two sides parallel to a

given line.
Now let us consider the inscribed rhombi with one diagonal parallel to a given line.
We start with a special case of convex polygons.

47

Proposition 5.9. Every convex polygon in the plane has an inscribed rhombus with
a diagonal parallel to a given line.
This result is a special case of a result for general simple polygons which again uses
the mountain climbing lemma (Theorem 5.5). We present two simple proofs, both
related to far-reaching generalizations (see e.g., Exercise 5.19).
First proof. Let X = A be a convex polygon in the plane and let be a given line.
For a point x A r X, denote by a1 , a2 the intersections of a line though x and
parallel to with X. Similarly, denote by b1 , b2 the intersections of a line through x
and orthogonal to with X (see Figure 5.11). Let f : A A be a function defined
by
f (x) = cm{a1 , a2 , b1 , b2 },
where cm{} denotes the center of mass. Clearly, function f is continuous and can
be extended by continuity to X. By the Brouwer fixed point theorem, there exists a
point z A such that z = f (z). Then z is a midpoint of (a1 , a2 ) and of (b1 , b2 ), so
[a1 b1 a2 b2 ] is the desired inscribed rhombus.

b1
a1

a2

x
f (x)

A
b2

C1

C2

Figure 5.11. Function f : A A in the first proof of Proposition 5.9.

Average curves C1 and C2 in the second proof.
Second proof. Think of as a horizontal line. Denote by C1 the locus of midpoints
of vertical lines intersecting a given polygon X. Similarly, denote by C2 the locus of
midpoints of horizontal lines (see Figure 5.11). The endpoints of C1 are the leftmost
and rightmost points of X (or the midpoints of the vertical edges). Therefore, curve C2
separates them, and thus intersects C1 . The intersection point is the center of the
desired rhombus.

Perhaps surprisingly, the extension of Proposition 5.9 to all simple polygons is more
delicate. As in the proof of Theorem 5.8, we need to use the mountain climbing lemma
twice.
Theorem 5.10. Every simple polygon in the plane has an inscribed rhombus with a
diagonal parallel to a given generic line.
Proof. Let X R2 be a simple polygon, and let be a given generic line. Think of
being horizontal. Let (x) denote the height of a point x X. As in the proof of
Theorem 5.8, let y and z be the points at minimum and maximum height. Since is

48

generic, these points are unique. Without loss of generality, we assume that (y) = 0
and (z) = 1.
Use the mountain climbing lemma (Theorem 5.5) to continuously and piecewise
linearly (in time t) move points u1 = u1 (t) and u2 = u2 (t) from y to z, so that they
remain at the same height, for all t [0, 1] (see Figure 5.12). We will show that
(u1 (t), u2 (t)) is a diagonal of a rhombus, for some t [0, 1]. Let us remark here that
not all pairs of points at the same height may necessarily appear as (u1 , u2 ).
z
X

w
u2

u1

y
Figure 5.12. Line L through midpoints of an interval (u1 , u2 ) k .
For every t [0, 1], consider a vertical line L through the midpoints of (u1 (t), u2 (t))
and plot the height h = (x) of the intersection points x L X in a graph
= {(t, h), 0 t 1}.
By construction, the resulting graph is a union of non-intersecting piecewise linear
curves. There are at least two curve endpoints at t = 0 and at least two curve
endpoints at t = 1. We will prove that there are at least two continuous curves in
with one endpoint at t = 0 and one at t = 1 (see Figure 5.13).
Denote by v and w the leftmost and the rightmost points in X (see Figure 5.11).
They separate the polygon into two curves X and X . Since the intersection points
always lie between these, there is always an odd number of intersection points with X .
Therefore, there exists a continuous curve in with one endpoint in X at t = 0 and
one endpoint in X at t = 1. The same is true for X , which proves the claim.
Denote these curves by C and C , respectively. In general, curves C and C are not
functions of t. Use the mountain climbing lemma to parameterize the curves




C = { h1 ( ), t1 ( ) , [0, 1]} and C = {(h2 ( ), t2 ( ) , [0, 1]},
so that t1 ( ) = t2 ( ). Now define the average curve C of C and C as




C = h1 ( )/2 + h2 ( )/2, t1 ( ) , [0, 1] .

By construction, curve C is continuous, starts at t = 0 and ends at t = 1.

Consider a curve H given by the function h(t) = (u1 (t)). Plot H and the average
curve C on the same graph as in Figure 5.13. Since h(t) is a continuous function,

49

C
C

t
0

t
1

Figure 5.13. Graph with curves C , C , the average curve C , and curve H.
h(0) = 0 and h(1) = 1, we conclude that C intersects H at some t = T . Then
(u1 (T ), u2(T )) is the diagonal of a rhombus inscribed into X.


5.6. Squares in polygons. We are ready to prove now the main result of this section: every simple polygon in the plane has an inscribed square. We present an easy
proof for convex polygons and a more delicate proof in the general case. Another
proof is given in Subsection 23.6, and is based on a different approach.
Proposition 5.11. Every convex polygon in the plane has an inscribed square.
Proof. Assume that convex polygon X R2 has no parallel edges. By Proposition 5.9
and Exercise 5.2, for every direction u there exists a unique rhombus inscribed into X
with a diagonal parallel to u. Denote by g(u) the difference in the diagonal lengths
and check that g(u) is continuous (see Exercise 5.2). Since g(u) changes sign when
vector u rotates by /2, there exists a vector e such that g(e) = 0. This gives the
desired inscribed square.

Theorem 5.12 (Square peg theorem). Every simple polygon in the plane has an
inscribed square.
The proof of this theorem will be given in Subsection 23.6, based on other ideas.
5.7. Exercises.
Exercise 5.1. a) [1-] Prove that for every simple polygon X R2 and an interior point O,
there exist points x, y X such that O is a midpoint of (x, y).
b) [1] Let X R2 be a convex polygon with the center of mass cm(X) = O. Prove that X
has an inscribed parallelogram with the center of mass at O.
c) [1-] Let X = [x1 . . . xn ] R2 be a non-convex simple polygon. Prove that there exists a
vertex xi and points y, z X such that xi is a midpoint of (y, z). Check that this does not
necessarily hold for all vertices xi in the interior of the convex hull of X.

50

Exercise 5.2. (Inscribed rhombi) a) [1] Prove that every convex polygon in the plane
with no parallel edges has at most one inscribed rhombus with a diagonal parallel to a given
line.
b) [1] In conditions of a), prove that as the line rotates, the rhombi change continuously.
c) [1] Same for inscribed rhombi with an edge parallel to a given line.
d) [1] Find examples of polygons with parallel edges which do have inscribed rhombi as in
parts a) and b). Prove that whenever a polygon has two such rhombi, it has innitely many
of them.
e) [1-] Find a convex polygon X R2 with non-parallel edges and a vertex v, such that X
has exactly two inscribed rhombi with vertices at v.
Exercise 5.3. a) [1] Let X R2 be a simple polygon. Prove that there exist a cyclic
quadrilateral (inscribed into a circle) which is inscribed into X, and whose vertices divide
the closed curve X into four arcs of equal length.
b) [1] Let X R3 be a simple space polygon. Prove that there exist a quadrilateral inscribed
into X whose vertices lie in a plane and divide X into four arcs of equal length.
Exercise 5.4. [1] Let R R2 be the Reuleaux triangle dened in Exercise 3.6. Prove
that R has no inscribed regular n-gons, for all n 5.
Exercise 5.5. [1] Let Q1 , Q2 , Q3 R2 be three (non-intersecting) piecewise linear curves
with the same endpoint x and other endpoints y1 , y2 , and y3 , respectively. Let Q = Q1
Q2 Q3 be the union of these curves. Prove that Q has an inscribed equilateral triangle
with a vertex at either y1 , y2 or y3 .
Exercise 5.6. [1] Let Q = A, A R2 be a simple polygon in the plane and let x1 , x2 , x3
Q be three distinct points on Q. Prove that there exists a circle inside A which contains at
least one point from the three closed arcs of Q separated by the points xi .
Exercise 5.7. [2-] Let X R2 be a simple polygon. Use Exercise 5.5 to prove that for all
but at most two points z X, there exists an equilateral triangle inscribed into X with a
vertex at z.
Exercise 5.8. [1-] Prove or disprove: for every convex polygon Q R2 containing the
origin O in its relative interior, there exists a triangle inscribed into Q, with the center of
mass at O.
Exercise 5.9. a) [1-] Prove that for every three distinct parallel lines in the plane there
exists an equilateral triangle with a vertex on each line.
b) [1] Prove that for every d + 1 distinct parallel hyperplanes in Rd there exists a regular
simplex with a vertex on each hyperplane.
c) [1] Prove or disprove: such a simplex is uniquely determined up to a rigid motion.
Exercise 5.10. [1+] Prove that for every three non-intersecting lines 1 , 2 , 3 R3 there
exists a unique triangle = (a1 , a2 , a3 ), such that ai i and the lines have equal angles
with the adjacent edges of . For example, the angle between 1 and (a1 , a2 ) must be equal
to the angle between 1 and (a1 , a3 ).
Exercise 5.11. a) [1-] Prove or disprove: for every simple convex cone C R3 there
exists a plane L, such that C L is an equilateral triangle.
b) [1-] Prove or disprove: for every simple convex cone C R3 with equal face angles, there
exists a unique plane L, such that C L is an equilateral triangle.

51

c) [1] Let C be a simple cone with three right face angles. Prove that for every triangle T
there exists a plane L, such that C L is congruent to T .
d) [1] Prove that every convex cone C R3 with four faces there exists a plane L, such
that C L is a parallelogram.

Exercise 5.12. a) [1-] Prove that for every tetrahedron R3 there exists a plane L,
such that L is a rhombus. Prove that there are exactly three such planes.
b) [1] Suppose the resulting three rhombi are homothetic. Prove that is equihedral (see
Exercise 25.12).
c) [1] Suppose the resulting three rhombi are squares. Prove that is regular.
Exercise 5.13. [1+] Prove or disprove: every space polygon Q R3 has an inscribed
rectangle.
Exercise 5.14. [1] Let X R2 be a simple polygon in the plane. Prove that there exists
a rectangle Q with exactly three vertices of Q lying in X.
Exercise 5.15. Let Q R3 be a simple space polygon.
a) [2+] Prove that for every k 3 there exists an equilateral k-gon inscribed into Q. In
other words, prove that there exist distinct points y1 , . . . , yk Q such that
|y1 y2 | = . . . = |yn1 yn | = |yn y1 | .

b) [] Prove that for every k 3 there exists an equiangular k-gon inscribed into Q. In
other words, prove that there exist distinct points y1 , . . . , yk Q such that
yn y1 y2 = y1 y2 y3 = . . . = yn1 yn y1 .

Exercise 5.16. a) [1+] Prove that for every convex plane polygon and every 6= 1 there
exists at least two inscribed rectangles with aspect ratio (ratio of its sides) equal to .
b) [2] Prove that for every simple plane polygon there exists an inscribed rectangle with a
given aspect ratio (ratio of its sides).
c) [] Let Q R2 be an isosceles trapezoid. Prove that for every simple plane polygon
there exists an inscribed polygon similar to Q (equal up to homothety).
d) [1] Prove that part c) does not extend to any other quadrilateral Q.
Exercise 5.17. Let 1 , 2 , 3 and 4 be generic lines in the plane.
a) [1] Prove that there are exactly three squares with all vertices on dierent lines i .
Formalize explicitly what it means to be generic.
b) [1] Prove that there are 12 more squares with vertices on lines i .
c) [1] Compute the number of rectangles with given aspect ratio (ratio of its sides), such
that all vertices lie on dierent lines.
d) [1] Let Q R2 be a simple quadrilateral with dierent edge lengths. Compute the
number of quadrilaterals similar to Q, with all vertices on lines i .
Exercise 5.18. a) [2+] Prove that every simple space polygon in R3 has an inscribed at
(i.e., coplanar) rhombus.
b) [1-] Prove that every at rhombus inscribed into a sphere is a square. Use part a) to
conclude that every simple spherical polygon has an inscribed square.
c) [2-] Use part b) and a limit argument to give another proof that every simple polygon
in the plane has an inscribed square.
Exercise 5.19. An equihedral octahedron Q is dened as a convex hull of three orthogonal
intervals (diagonals of Q) intersecting at midpoints. Clearly all eight triangular faces in Q
are congruent.

52

a) [1+] Let P R3 be a convex polytope which is in general position with respect to the
orthogonal axes. Prove that P has an inscribed equihedral octahedron with its diagonals
parallel to the axes.
b) [1] Show that the general position condition on P is necessary.
c) [1] Find a convex polytope P R3 which has at least two inscribed equihedral octahedra
with parallel diagonals.
Exercise 5.20. (Inscribed octahedra) a) [2+] Prove that every convex polytope P R3
has an inscribed regular octahedron.
b) [2+] Generalize the result to (non-convex) surfaces embedded in R3 and homeomorphic
to a sphere.
c) [2+] Generalize the result to higher dimensions.
Exercise 5.21. [2-] Find a convex polytope in R3 which has no inscribed bricks (rectangular parallelepipeds).
Exercise 5.22. (The table and the chair theorems) Let Q = A, A R2 be a convex
polygon, and let f : R2 R+ be a continuous piecewise linear function which is zero outside
of Q.
a) [2+] Prove that for every c > 0 there exist a square of size c, such that cm[x1 x2 x3 x4 ] A
and f (x1 ) = f (x2 ) = f (x3 ) = f (x4 ).13
b) [2+] Prove that for every triangle R2 there exist a translation = [x1 x2 x3 ] such
that cm( ) A and f (x1 ) = f (x2 ) = f (x3 ).14
c) [2-] Prove that part a) does not hold for non-convex A.
Exercise 5.23. (Mountain climbing lemma for general functions) a) [1-] Show that the
mountain climbing lemma (Theorem 5.5) does not extend to all continuous functions with
the same boundary conditions.
b) [1+] Extend the mountain climbing lemma to all piecewise algebraic functions.
Exercise 5.24. (Generalized mountain climbing lemma) [1-] Let fi : [0, 1] [0, 1]
be k continuous piecewise linear functions with fi (0) = 0 and fi (1) = 1, for all 1 i k.
Prove that there exist k continuous piecewise linear functions gi : [0, 1] [0, 1], such that
gi (0) = 0, gi (1) = 1 for all 1 i k, and
f1 (g1 (t)) = . . . = fk (gk (t))

for every t [0, 1].

Exercise 5.25. (Ladder problem) [2-] Let C be a piecewise linear curve in the plane which
begins and ends with a straight unit segment. Prove that there is a continuous family of
unit intervals which are inscribed into C containing both end segments.15
Exercise 5.26. (Ring width problem) [2] Let X R2 be a simple polygon with two rays
attached as in Figure 5.14. Think of X as if it was made out of metal.
a) [2-] A metallic ring of diameter is an interval of length whose endpoints lie on dierent
sides of X. Suppose a metallic ring of diameter can slide from one ray into another. Prove
that so can a metallic ring of diameter , for all > .
b) [2-] An elastic ring of diameter is an interval of length whose endpoints lie on dierent
sides of X, and whose length is allowed to decrease to any length smaller than . Suppose an
13In

other words, there is a way to place a table of every size on a hill; thus the table theorem.
is the chair theorem.
15Thus, two people can coordinate their movements to move along a path while carrying a ladder.
14This

53

elastic ring of diameter can slide from one ray into another. Prove that so can a metallic
ring of diameter .16
c) [2] Show that the ray condition can be removed in part a).
X

Figure 5.14. Polygon X and a ring of length .

5.8. Final remarks. The piecewise linear version of the mountain climbing lemma (Theorem 5.5) is considerably simpler than the general case [Kel]. Both our proofs follow the
proof idea in [GPY], stated there in a more general context.
Our proof of Theorem 5.3 and the second (topological) proof of Proposition 5.4 follows [Nie]. Proposition 5.9 was rst proved in [Emch] and generalized in [HLM]. Our proof
of the proposition follows the proof in [Kra], where the most general result on inscribed
parallelepipeds is established (see Exercise 5.19).
The problem of inscribed squares has a long history. It was rst proved by Emch for
convex curves (see [Emch]), by Shnirelman [Shn] (see also [Gug2]) for suciently smooth
curves, and further extended in [Gri, Jer, Stro] and other papers. The proof of the square peg
theorem (Theorem 5.12) given in Subsection 5.6. We refer to surveys [KleW, Problem 11]
and [Nie] for more on the subject. See also [Mey3] for the references and connections to the
table and chair theorems (Exercise 5.22).

16Warning:

it takes an eort to realize that both parts of the problem are in fact not obvious.

54

6. Dyson and Kakutani theorems

In this short section we continue the study of inscribed figures, concentrating on
inscribed tripods. We first prove Dysons and Kakutanis theorems and then their
common generalization, which is itself a special case of the Knaster problem (see
Section 6.7). These results will not be used later on, but the underlying continuity
arguments will be repeatedly used throughout the book.
6.1. Centrally symmetric squares in space polygons. Now we can move into
R3 and prove the following attractive result. We say that a polygon Q R3 is simple
if it is not self-intersecting.
Theorem 6.1. Let Q R3 be a simple polygon centrally symmetric with respect to
the origin O. Then there exists a square inscribed into Q and centrally symmetric
with respect to O.
Proof. For simplicity, we will refer to the opposite points v and v = v in Q as the
same point, hoping this would not lead to a confusion.
For every point v Q consider a plane Hv containing O and perpendicular to (O, v).
We say that Q is generic if the intersection Tv = Q Hv contains at most one vertex
(in fact, an identified opposite pair of vertices) of Q. Then Tv is always finite, since
otherwise it must contain the whole edge e in Q, and thus at least two vertices.
Assume for now that Q is generic. Consider a graph = {(v, w) Q2 | v Q, w
Tv } which can be viewed as a subset of a square with the opposite sides identified. We
claim that since Q is generic, is a union of a finite number of disjoint curves. This
follows from the fact that locally, around every non-vertex point v Q, the graph
is a union of disjoint curves.
v

Hv

Figure 6.1. Portion of a polygon Q, plane Hv and three points of

intersection in Tv .
Observe that for a generic point v Q the number of points in Tv is odd. This
follows from assumption that Q is simple and from the fact that a plane Hv separates
two opposite points of a polygon Q. Therefore, at least one of the curves in is
defined on the whole Q.
Finally, for every point v Q let f (v) = |vO|. In this notation, we need to find
two points v, w Q such that (v, w) and f (v) = f (w). Suppose there is no
such pair. Then all curves in split into symmetric pairs of curves Ci and Ci with
f (v) > f (w) and f (v) < f (w). Since does not intersect the diagonal {(v, v), v Q},
every two corresponding curves Ci Ci together have an even number of points in the

55

intersection Tv , for a generic v. This gives a contradiction and completes the proof
when the polygon Q is generic.
C

C
v

Figure 6.2. Curve C , the corresponding symmetric curve C ,

and four points in the intersection Tv .
To remove the assumption that Q is generic, perturb all the vertices. For vertices v1
and v2 there is a unique central hyperplane H containing them, and the normal to H
will not intersect Q when v1 , v2 are in general position. Consider now a family of
generic polygons converging to Q and take the limit of the squares inscribed into
them. Since Q is embedded, it does not contain the origin O. Therefore, the squares
do not degenerate in the limit, since otherwise they would contract to O. Thus, in
the limit we obtain the desired inscribed square.

6.2. Centrally symmetric squares in polytopes. The following result is the first
of the two main results in this section.
Theorem 6.2 (Dyson). Let P R3 be a convex polytope containing the origin O in
its relative interior. Then there exists a centrally symmetric square inscribed into P .
Proof. Let us prove that the surface S = P of the polytope P contains a centrally
symmetric polygon Q. The result then follows from Theorem 6.1.
Let P be a reflection of P in O, and let S = P . Since O P , then P P 6= ,
and thus the intersection W = S S is nonempty. Indeed, if W = , then one of
the polytopes P, P would be contained in another and thus has a greater volume.
Assume for now that W is one-dimensional, and moreover a union of non-intersecting
polygons Q1 , . . . , Qn . These polygons Qi lie on the surface S homeomorphic to a
sphere S2 . Note that for all i, the reflection Qi of a polygon Qi is also one of the
polygons in W .
Let us prove that one of the polygons Qi is centrally symmetric. Consider a radial projection : W S2 of W onto a sphere centered at O. Then (Qi ) are
non-intersecting Jordan curves on S2 . By contradiction, suppose none of the (Qi ) is
centrally symmetric. By analogy with the proof of Proposition 3.4, one of the connected components of S2 (W ) is a centrally symmetric region U S2 . For x S2 ,
denote by f (x) = | 1 (x)O| the distance from the origin to the point which projects
onto x, and let g(x) = f (x) f (x ), where x S is a reflection of x. By construction,

56

the function g is continuous and g(x) 6= 0 on U. On the other hand, g(x) = g(x )
for all x, x U, a contradiction.
Suppose now that the above assumption is false, i.e., that W S is not a union of
non-intersecting polygons. Perturb the vertices of P so that no two faces are parallel
and no vertex is opposite to a point on a face. The resulting polytope satisfies the
assumption, and the theorem follows by the limit argument.


6.3. Circumscribed cubes. It is natural to assume that Proposition 5.1 can be

extended to higher dimensions. This is indeed possible, but the proof is no longer
straightforward. We present only the 3-dimensional case here (see also Exercise 6.12).
We say that a convex set X R3 has a circumscribed cube if there exists a cube
C R3 , such that X C and every (2-dimensional) face of C contains a point of X.
Theorem 6.3 (Kakutani). Every convex set X R3 has a circumscribed cube.
Proof. Let O X be the origin, and let S2 be the unit sphere centered at O. For
a point x S2 , denote by f (x) the distance between planes orthogonal to (Ox)
and supporting X on both sides. Think of f as a (symmetric) function on a sphere
f : S2 R. It suffices to show that there exists three points x, y, z S2 such that
(Ox), (Oy) and (Oz) are orthogonal to each other and such that f (x) = f (y) = f (z).
We prove this claim by contradiction.
Think of (xyz) as a right equilateral triangle on S2 , i.e., a spherical triangle with
edge lengths /2. For every right equilateral triangle (xyz) S3 , denote by (x, y, z)
the orthogonally projection of the point (f (x), f (y), f (z)) R3 onto the plane L =
{(a, b, c), a+b+c = 0} R3 containing the origin O. Clearly, the map is continuous
and if f (x) = f (y) = f (z), then (x, y, z) = O. We assume that (x, y, z) 6= O for all
right equilateral triangles.
Fix the three points x0 = (1, 0, 0), y0 = (0, 1, 0) and z0 = (0, 0, 1) in the sphere.
Observe that (x0 , y0 , z0 ), (y0 , z0 , x0 ) and (z0 , x0 , y0 ) are equilateral triangles in L.
Denote by G = SO(3, R) the group of orthogonal rotations, whose elements can be
identified with the right equilateral triangles (x, y, z) of the same orientation. Consider
a curve 0 G which connects (x0 , y0 , z0 ) to (y0 , z0 , x0 ). Use the symmetry to connect
(y0 , z0 , x0 ) to (z0 , x0 , y0 ) and (z0 , x0 , y0 ) to (x0 , y0, z0 ) with the rotations of 0 . Now
these three copies of 0 form a closed curve 1 G. Suppose (x, y, z) 6= O for every
point (x, y, z) 1 . Denote by 0 = (0 ) and 1 = (1 ) the curves in L r O. By
construction, 1 consists of three rotated copies of 0 (see Figure 6.3). Therefore, the
argument of 1 , defined as 2 times the winding number around O, is nonzero and is
of the form (3m + 1)2.
Finally, recall that G = SO(3, R) is homeomorphic to RP3 , and hence the fundamental group (G) = Z2 . Therefore the curve 2 = 21 is contractible in G. On the
other hand, its image 2 = (2 ) = 21 is not contractible in L r O, a contradiction.


57

0
0

1
O

Figure 6.3. Two examples of curves 0 and an example of curve 1

with winding number 1.
6.4. Tripods standing on surfaces. Let S R3 be an embedded compact orientable surface (without boundary), and denote by P its interior: S = P . Let O
P r S be a fixed point in the interior of S. We say that a tetrahedron (Ouvw), where
u, v, w S is a (, )-tripod if |Ou| = |Ov| = |Ow|, uOv = , uOw = , and
vOw = .
Theorem 6.4 (The tripod theorem). Let S R3 be an embedded compact surface
without boundary, and let O be a point inside S. Suppose angles and satisfy
0 < , and 2. Then there exist points u, v, w S, such that the
tetrahedron (Ouvw) is a (, )-tripod.
There are several ways to think of this result. Visually, it is saying that on every
surface one can place a tripod (a three-legged stand with given angles , between
its legs) such that the top is placed at any given point O (see Figure 6.4).
On the other hand, the tripod theorem can be viewed as a generalization of the
Kakutani theorem (Theorem 6.3). Indeed, let f : S2 R+ be a function as in the
proof of the Kakutani theorem. Consider a surface S R3 of points at distance f (x)
in direction (Ox), x S2 . Clearly, the surface S is homeomorphic to a sphere. Now
observe that the (/2, /2)-tripod is the desired right equilateral triangle corresponding to directions of the circumscribed cube.
Similarly, the tripod theorem can be viewed as a variation on the Dyson theorem
(Theorem 6.2). Indeed, let f : S2 R+ be the distance function to a point on the
surface P . Now the (/2, )-tripod corresponds exactly to three of the four vertices
of an inscribed square.
Proof. Let x0 and x1 be points on the surface S with the largest and the smallest
distance from O, respectively. We can always assume that x0 6= x1 , since otherwise
S = S2 and the result is trivial. Let {xt , t [0, 1]} be a path on the surface S
between x0 and x1 . Let Ct be the set of points y R3 such that |Oy| = |Oxt | and
yOxt = , for all t [0, 1]. Clearly, Ct is a circle which changes continuously with t.
Observe that C0 has no points inside S, while C1 has no points outside S. Let
be the supremum of t [0, 1], such that Ct contains two outside points at , bt
/ P at
angle = at Obt . If = 0, then tetrahedron (Ox0 a0 b0 ) is the desired tripod. Now
assume that > 0. Denote by (a , b ) the limit of a sequence of intervals (at , bt ) as
t . Note that a , b S since otherwise is not a supremum defined above. We
conclude that (Ox a b ) is the tripod as in the theorem.


58

bt

at

O
u

w
v

xt
Ct

Figure 6.4. A (, )-tripod (Oxyz) standing on the surface S and an

element in its construction.
6.5. General tripods. The problem of inscribed tripods can be extended to include
tripods with three possible angles. The following result shows that this is possible
for the star-shaped surfaces S R3 defined as surfaces of regions which contain the
origin O and every interval (Ox) for all x S. We present it in the form amenable
to generalizations to higher dimensions, where it is called the Knaster problem.
Theorem 6.5 (Knasters problem in R3 ). Let f : S2 R+ be a continuous function
on a unit sphere, and let x1 , x2 , x3 S be three fixed points. Then there exists a
rotation SO(3, R), such that
f ((x1 )) = f ((x2 )) = f ((x3 )) .
In other words, the theorem says that there exists a spherical triangle (y1 y2 y3 )
congruent to (x1 x2 x3 ), similarly oriented, and such that f (y1 ) = f (y2) = f (y3 ).
When
|x1 x2 |S2 = , |x1 x3 |S2 = , |x2 x3 |S2 = ,
we obtain the tripod theorem (Theorem 6.4) for star surfaces.
6.6. Exercises.
Exercise 6.1. a) [1-] Prove that for every tetrahedron R3 there exists a unique
circumscribed parallelepiped which touches all edges of .
b) [1-] Prove that the circumscribed parallelepiped as above is a brick (has right face angles)
if and only if is equihedral, i.e., has congruent faces (see Exercise 25.12).
Exercise 6.2. [1-] Extend Theorem 6.1 to self-intersecting centrally symmetric polygons.
Exercise 6.3. [1] The aspect ratio of a rectangle is the ratio of its sides. Prove that every
centrally symmetric polygon Q R3 contains an inscribed centrally symmetric rectangle
with a given aspect ratio > 0. When = 1 this is the claim of Theorem 6.1.
Exercise 6.4. [2] A quadrilateral A in R3 is called regular if all four sides are equal and
all four angles are equal. Prove that every space polygon in R3 has an inscribed regular
quadrilateral. Generalize this to Rd , for all d 3.

59

Exercise 6.5. [1] Let A be a quadrilateral inscribed into a space polygon Q R3 . Note
that for A to be a square it has to satisfy ve equations, while there are only four degrees
of freedom to choose A. Formalize this observation. Conclude that a generic space polygon
does not have an inscribed square.
Exercise 6.6. [2+] Let Rd be a xed simplex. Prove that every convex polytope
P Rd has an inscribed simplex similar to . Generalize this to polyhedral surfaces
homeomorphic to a sphere.
Exercise 6.7. [1] Let Q(z) = [ABCD] R3 be a quadrilateral with A = (0, 0, 0), B =
(0, 1, 0), C = (0, 1, z) and D = (1, 0, 0). Check that Q(1) does not have an inscribed
square. Use convergence argument to show that Q(z) does not have inscribed squares, for
all suciently small z > 0.
Exercise 6.8. [1] Check that Dysons theorem (Theorem 6.2) holds for all (non-convex)
polyhedra P R3 whose surface S = P is a star surface with respect to O. What about
general 2-dimensional orientable polyhedral surfaces?
Exercise 6.9. [2-] Use Exercise 6.3 to prove that every convex polyhedron P R3 has an
inscribed rectangle with a given aspect ratio and given center inside P .
Exercise 6.10. [1-] Let P R3 be a convex polytope centrally symmetric at the origin O.
Prove that there exists a regular octahedron inscribed into P and centrally symmetric at O.
Exercise 6.11. [1+] Let P R3 be a convex polytope centrally symmetric at the origin O.
Prove that there exists a regular hexagon inscribed into P and centrally symmetric at O.
Exercise 6.12. (Generalized Kakutanis theorem in higher dimensions) [2] Prove that every
convex set X Rd can has a circumscribed hypercube.

Exercise 6.13. [2-] Prove that every centrally symmetric convex polytope in R3 has an
inscribed cube.

Exercise 6.14. (Knaster problem in R3 ) Recall that SO(3, R) is homeomorphic to RP3 .

a) [1+] A manifold Z is called unicoherent if for every closed, connected subsets A, B X
with X = A B, the intersection A B is connected. Prove that RP3 is unicoherent.
b) [1+] Let X be a unicoherent manifold, and let be a xed point free involution on X.
Suppose a subset A X satises the following conditions:
(i) A is a closed subset of X,
(ii) A is invariant under ,
(iii) A separates x from (x), for all x X.
Then there exists a connected subset B A which satises (i)(iii).
c) [1] Let X = SO(3, R) and f as in Theorem 6.5. Dene functions g, g : X R by
g() = f ((x1 )) f ((x3 )),

g () = f ((x2 )) f ((x3 )), where X.

Dene A and A the subsets of rotations X with g() = 0 and g () = 0, respectively.

Denote by , SO(3, R3 ) the involutions which satisfy (x1 ) = x3 , (x2 ) = x3 , and dene
the involutions and on X as multiplications by and from the right:
() = ,

() = .

Check that g = g, g = g , and that the pairs (, A) and ( , A ) satisfy conditions

of part b).

60

d) [1] Denote by B A and B A the connected subsets obtained by applying part b)

to sets A and A in part c). Assume that B B = . Use the fact that , preserve the
Haar measure on SO(3, R) to obtain a contradiction. Conclude from here Theorem 6.5.
Exercise 6.15. (General tripods on surfaces) [] Let S R3 be an embedded orientable
surface and let O be a point inside S. For a simple cone C with face angles , and ,
dene a (, , )-tripod on S to be a triple of points u, v, w S, such that the tetrahedron
(Ouvw) has uOv = , vOw = , wOv = , and |Ou| = |Ov| = |Ow|. Prove that for
every C as above, there exists a (, , )-tripod on S.
6.7. Final remarks. Kakutanis theorem (Theorem 6.3) was proved in [Kaku] and became
a source of inspiration for a number of results. It was extended in [YamY] to higher dimensions (Exercise 6.12). Our version of Dysons theorem is dierent from the original [Dys],
and can be viewed as a variation on Kakutanis theorem. Dyson noticed that in the proof
Kakutani never used the fact that f is centrally symmetric, so he showed that general functions on spheres always have two orthogonal diameters such that the function is constant
on all four endpoints. This result, in turn, was also extended in a number of ways (see e.g.
Exercise 6.8 and 6.9).
Our proof of the tripod theorem (Theorem 6.4) is based on [Gor], which in turn follows
the proof in [YamY]. Theorem 6.5 is due to Floyd who used a topological argument to
prove it [Flo]. The proof is outlined in Exercise 6.14. The generalization to all embedded
2-dimensional surfaces is stated as Exercise 6.15.
Knasters problem was formulated by B. Knaster in 1947 in an attempt to further generalize the Kakutani theorem. Namely, he conjectured that for every f : Sd Rm and
x1 , . . . , xn Sd , where n = d m + 2, there exists a rotation SO(m, R) such that
f ((x1 )) = . . . = f ((xn )). This conjecture was refuted in a number of cases, for example
for all d > m > 2 [Chen], but remained open for m = 1 until recently, when it was disproved
for d > 1012 [KasS].

61

7. Geometric inequalities
The isoperimetric problems are some of the oldest problems in geometry. There
are numerous monographs written on the subject, so we see little need to expand
on this (see Subsection 7.9). We do need, however, the important BrunnMinkowski
inequality, including the equality part of it, to prove the Minkowski theorem in Section 36. Thus, we present an elementary introduction to the subject, trying to make
it as painless as possible.
7.1. Isoperimetry in the plane. It is well known that among all regions in the
plane with the same area, the disk has the smallest perimeter. Proving this result is
actually more delicate than it seems at first sight. The following proof due to Steiner
is beautiful, but also (somewhat) incorrect. While we explain the problem in the next
subsection, it is a nice exercise to find the error.
Throughout this section we always assume that our sets are convex. This allows
us to avoid certain unpleasantness in dealing with general compact sets.17
Theorem 7.1 (Isoperimetric inequality in the plane). Among all convex sets in the
plane with a given area, the disk has the smallest perimeter.
Incorrect proof of Theorem 7.1. The proof can be split into several easy steps. Let
X R2 be a compact set in the plane. We say that the set is optimal if it minimizes the perimeter given the area. We will try to characterize the optimal sets and
eventually show that only a disk can be optimal.
1) We say that two points x and y on the boundary P = X are opposite if they
divide P into two equal halves. The interval (xy) is called a diameter in this case.
We claim that in an optimal set X the diameter divides the area into two equal parts.
If not, attach the bigger part and its reflection to obtain a set Y of equal perimeter
and bigger area (see Figure 7.1).
2) Let (xy) be a diameter in an optimal set X, and let P = X. For every
point z P we then have xzy = /2. Indeed, from 1) the diameter splits the area
into equal parts. If xzy 6= /2, take chords [xy], [yz] and denote by A, B the regions
they separate as in the figure. Now attach A and B to a triangle (x y z ) with sides
|x z | = |xz|, |z y | = |zy|, and x y z = /2. Copy the construction symmetrically
on the other side of the diameter (see Figure 7.1). Since area(xyz) < area(x y z ) and
areas of segments remain the same, we obtain a region X with the same perimeter
and bigger area.
In summary, we showed that in an optimal set the angle at all points of the boundary to a fixed diameter must be /2. Thus X is a circle. This completes the proof. 

17Otherwise,

we have to specify that our set X is measurable, has a measurable boundary, etc.
For all purposes in this section the reader can assume that X has a piecewise analytic boundary.
Then it is easy to show that the convex hull conv(X) has a smaller perimeter but a bigger area, thus
reducing the problem to the convex case.

62

z
A
x

y
X

1)

y
X

y
X

B
2)

y
X

Figure 7.1. Increasing the area of convex sets while preserving diameter.
7.2. Reaching the optimum. The problem with the argument above is that at no
point do we show that the optimum exists. Naturally, if we set upon finding a convex
subset with maximum perimeter given the area, then the optimum does not exist.
Hence, showing that all sets except for the circles are not optimal does not prove that
the circles are optimal, and thus does not (yet) resolve the problem (in either case).
There are several ways to get around the problem. Let us describe two most standard
approaches.
(I) An easy patch is by an abstract compactness argument. Recall the classical
Blaschke selection theorem (see Exercise 7.1), which states that a set C() of convex
subsets of a compact set R2 is a compact. While this result is inapplicable to a
non-compact plane = R2 , we can restrict as follows.
Consider all convex sets with area and ask which of them has the smallest perimeter. Without loss of generality we can assume that our convex sets contain the origin O R2 . Let be a circle with radius 4. If a set X contains both the origin O
and point x
/ , then perimeter of X is at least 2|Ox| 8 > 2, i.e., bigger than
perimeter of a unit circle. Therefore, the optimum, if exists, must be achieved on a
convex subset of .
Now, the arguments 1) and 2) in the proof above show that the perimeter is not
minimal among all sets with area except for unit circles. On the other hand, by
compactness of C(), at least one minimum must exist. Therefore, the unit circle is
the desired minimum.
(II) Another way to prove the result is to exhibit an explicit convergence to a
circle. That is, start with a set X, apply rule 1) and then repeatedly apply rules 2) to
points z P chosen to split the perimeter into 2r pieces of equal length (occasionally
one must also take convex hulls ). We already know that under these transformations
the perimeter stays the same while the area increases. Eventually the resulting region
converges to a circle. This implies that the area of all sets with given perimeter was
smaller than that of circles, as desired. We will skip the (easy) details.
Remark 7.2. Let us emphasize the similarities and dierences between the two approaches,
and the incorrect proof above. In all cases, we use the same variational principle, by exhibiting simple rules of how to improve the desired parameter (area/perimeter2 ). In (I) we then
used a general compactness argument to show that the optimum must exist. Alternatively,
in (II) we showed that these rules when applied innitely often, in the limit will transform
any initial object to an optimum.

63

While the rst approach is often easier, more concise and elegant as a mathematical
argument, the second is more algorithmic and often leads to a better understanding of the
problem. While, of course, there is no need to use the variational principle to construct a
circle, making a distinction between two dierent approaches will prove useful in the future.

7.3. How do you add sets? Let us define the Minkowski sum of any two sets
A, B Rd as follows:
A + B := {a + b | a A, b B} .
For example, if A and B are two intervals starting at the origin O, then A + B is a
parallelogram spanned by the intervals. Similarly, if A Rd is a convex set and B is
a closed ball of radius > 0 centered at the origin, then A + B consists of all points z
at distance at most from A (see Figure 7.2).
Let us make several observations on the Minkowski sum. First, the sum is symmetric and respects translation, i.e., if A is a translate of A, and B is a translate of B,
then A + B is a translate of A + B. On the other hand, the Minkowski sum does
not respect rotations: if A and B are (possibly different) rotations of A and B, then
A + B does not have to be a rotation of A + B, as the example with two intervals
shows.
One more definition. For every R, let A = { a | a A}. We call set A
an expansion of B if there exists 6= 0, such that A = B. Clearly, if A = B, and
6= 0, then B = (1/)A. In other words, if A is an expansion of B, then B is an
expansion of A.
Now, suppose H1 and H2 are two parallel hyperplanes. Then H = 1 H1 + 2 H2 is
yet another parallel hyperplane. For A1 H1 and A2 H2 the set B := 21 (A1 + A2 )
is the average of sets A1 and A2 , consisting of midpoints of intervals between the
sets. The set B lies in the hyperplane H3 = 12 (H1 + H2 ), in the middle between
hyperplanes H1 and H3 .

A +

A+B
A

B
A+B

B
+

B
A

A+B

A+B
2

Figure 7.2. Four examples of the Minkowski sum of convex sets.

64

Recall that by area(S) = vold1 (S) we denote the (d 1)-dimensional volume of

the surface S Rd . The following result expresses the area of S via the volume of
the Minkowski sums.
Proposition 7.3. Let X Rd be a convex set with the surface S = X, and let B
be a unit ball. Then:

d
area(S) =
vol(X + tB) t=0 .
dt
The proof is clear for polytopes and requires a straightforward convergence argument for general sets. In fact, everywhere below the reader can use this proposition
as a definition of the surface area of general convex sets.
7.4. Slicing polytopes with hyperplanes. The following theorem is the key result
in the field, marking the beginning of modern studies of geometric inequalities. We
will need it to solve the isoperimetric problem in Rd , and later on to prove the
Minkowski uniqueness theorem (see Section 36).
Theorem 7.4 (BrunnMinkowski inequality). Let A, B Rd be two convex bodies.
Then
vol(A + B)1/d vol(A)1/d + vol(B)1/d .
Moreover, the inequality becomes an equality only if A is an expansion of B.
We prove in this section only the first part of the theorem. The second part is
proved in the Appendix (see Subsection 41.5).
The following interesting result is actually a corollary from the BrunnMinkowski
inequality. It was, in fact, the original motivation for Brunns studies (the work
of Minkowski came later). We deduce this theorem from Theorem 7.4 in the next
subsection.
Theorem 7.5 (Brunn). Let P be a convex body and let H1 , H2 and H3 be three parallel
hyperplanes in Rd intersecting P in that order. Denote by Ai = P Hi , 1 i 3.
Then: area(A2 ) min{area(A1 ); area(A3 )}.
Example 7.6. Suppose d = 2 and line H2 is a midway

between H1 and H3 . Then a stronger
statement is true: area(A2 ) area(A1 )+ area(A3 ) /2. On the other hand, taking averages
is not useful in higher dimensions. Take a simplex P = {0 x1 x2 . . . xd 1} Rd ,
and let Hi be the hyperplanes dened by linear equations x1 = ai , where a1 = 0, a2 = 12 ,
and a3 = 1. Clearly, the hyperplanes are parallel and at equal distance from each other.
On the other hand, the areas of the intersections are given by
1
1
area(A1 ) = 0 , area(A2 ) = d , and area(A3 ) =
.
2 n!
n!

d
Thus, the best we can hope for is area(A2 ) area(A1 ) + area(A3 ) /2 , an unattractive
inequality formally not implying Brunns theorem. On the other hand, this does suggest
that taking d-th root is a good idea, since in this case we have area(A2 )1/d = area(A1 )1/d +


area(A3 )1/d /2.

65

7.5. The functional version. Let A, B Rd be two convex bodies and let Xt =
(1 t)A + tB, where t [0, 1]. As we mentioned in the example above, the first part
of the following proposition immediately implies Theorem 7.5.
Proposition 7.7. The function (t) = vol(Xt )1/d is convex on [0, 1].18 Moreover,
(0) vol(B) vol(A), and the inequality becomes an equality if and only if A is an
expansion of B.
Proof of Theorem 7.5 modulo Proposition 7.7. Let D = A1 + (1 )A3 , where is
given by the ratio of the distance between hyperplanes: = dist(H1 , H2 )/dist(H1 , H3 ).
By construction, set D is a convex body which lies in hyperplane H2 . On the other
hand, by definition of the Minkowski sum, we have D conv{A1 , A3 } P . Thus,
D A2 , and by Proposition 7.7 we conclude:
vol(A2 ) vol(D) min{vol(A1 ), vol(A3 )} ,

as desired.

Proof of Proposition 7.7 modulo Theorem 7.4. For every three values 0 t < t <
t 1, consider A = Xt , B = Xt and C = Xt . We have C = A + (1 )B, where
= (t t )/(t t ). By the BrunnMinkowski inequality we have:

1/d

1/d  d
1/d 
1/d
vol(C)1/d vol A
+ vol (1 )B
= vol(A)
+ (1 )d vol(B)
vol(A)1/d + (1 )vol(B)1/d ,

which proves the first part.

Observe that (0) = vol(A) and (1) = vol(B). By convexity, the right derivative (0) is well defined (see Figure 7.3). From above and by the BrunnMinkowski
inequality, ( 12 ) = 12 ((A) + (B)) if and only if A is an expansion of B. Therefore,
by convexity we have (0) (( 21 ) (0))/ 12 vol(B) vol(A), and the equality
holds if and only if A is an expansion of B.


vol(B)
(t)

vol(A)
0

1
2

Figure 7.3. Convexity of (t) = vol(Xt )1/d , where Xt = (1 t)A + tB.

18Sometimes

these functions are called concave downwards, upper convex, etc. All we can do here
to convey the meaning is to point to Figure 7.3 and hope the reader is not too confused.

66

7.6. The general isoperimetric inequality. Here is the main result in this section:
Theorem 7.8 (Isoperimetric inequality in Rd ). Among all convex sets in Rd with a
given volume, the ball has the smallest surface area.
Proof. Let A be a convex set, and let B be a unit ball in Rd . By Proposition 7.7, the
function (t) = vol (1 t)A + tB) is convex on [0, 1]. Therefore, by Proposition 7.3
we obtain:


 

d
A
tB
d

d

area(A) =
vol(A + tB) t=0 =
(1 + t) vol
+

dt
dt
1+t
1+t
t=0


d 


t
d

=
(1 + t)
 = d (0)d1 (0) + (0) .
dt
1+t
t=0

By Proposition 7.7, we have (0) vol(B) vol(A), and the inequality becomes
an equality if and only if A is an expansion of B. Since vol(A) = (0), the area(A)
minimizes only when A is an expansion of B, i.e., when A is a ball.

Remark 7.9. In the spirit of Remark 7.2, it is instructive to decide what type of argument
is used in this proof: did we use a compactness or an explicit convergence argument? From
above, the rule for decreasing (X) = area(X)/vol(X)(d1)/d turns out to be quite simple:
take Xt = (1 t)X + tB, where B is a unit ball. Letting t 1 makes Xt approach the
unit ball. Thus the proof can be viewed as an explicit convergence argument, perhaps an
extremely simple case of it.

7.7. Brick-by-brick proof of the BrunnMinkowski inequality. Perhaps surprisingly, the proof of Theorem 7.4 will proceed by induction. Formally, let us prove
the inequality for brick regions defined as disjoint unions of bricks (parallelepipeds
with edges parallel to the coordinate axes). Note that we drop the convexity condition. A union of bricks does not have to be connected or the bricks be of the same
size: any disjoint union of bricks will do (see Figure 7.4). Denote by Bd the set of all
brick regions in Rd .
Lemma 7.10. If the BrunnMinkowski inequality holds for brick regions A, B Bd ,
it also holds for general convex regions in Rd .
We continue with the proof of the inequality. The lemma will be proved later.
Proof of the BrunnMinkowski inequality for brick regions. Let A and B be two brick
regions. We use induction on the total number k of bricks in A and B. For the base of
induction, suppose k = 2, so both sets consist of single bricks: A = [x1 x2 . . . xd ]
and B = [y1 y2 . . . yd ]. The BrunnMinkowski inequality then becomes the
Minkowski inequality19 (Theorem 41.4).
For the induction step, suppose the result holds for brick regions with k or fewer
bricks (in total), k 2. Now, take two brick regions A, B Bd with k + 1 bricks.
Suppose A contains at least two bricks (otherwise we can relabel the regions). Fix
19While

we move the proof of the Minkowski inequality to the appendix as a standard result, we
do suggest the reader go over the proof so as to acquaint yourself with the proof style of inequalities.

67

these two bricks, say P, Q A. Observe that there always exists a hyperplane H Rd
with normal to one of the axes, which separates bricks P and Q. Denote by A1 and A2
portions of A lying on the two sides of H. Since H either avoids or divides each brick
into two parts, we have A1 , A2 Bd . Note also that A1 and A2 , each have strictly
fewer bricks than A.
Q

A2

B2
H

A1

B1

Figure 7.4. Brick region C on plane presented as a disjoint union of

rectangles with different colors. Two regions A, B B2 divided by a
line H with the same area ratios: = 2/5.
Let = vol(A1 )/vol(A). Recall that the volume of the Minkowski sum is independent of translation. Thus, we can assume that the hyperplane H goes through the
origin O Rd . Now translate set B so that H divides B into two sets B1 and B2
with the same volume ratio: = vol(B1 )/vol(B). We have
vol(A + B) vol(A1 + B1 ) + vol(A2 + B2 ),

since the Minkowski sums A1 + B1 and A2 + B2 lie on different sides of H and thus do
not intersect (except at the boundary). Since the total number of bricks in (A1 , B1 )
and in (A2 , B2 ) is less than k we can apply to them the inductive assumption:
vol(A + B) vol(A1 + B1 ) + vol(A2 + B2 )

d

d
vol(A1 )1/d + vol(B1 )1/d + vol(A2 )1/d + vol(B2 )1/d
h

1/d

1/d id
vol(A)
+ vol(B)
h

1/d id

1/d
+ (1 ) vol(A)
+ (1 ) vol(B)

 
d
+ (1 ) vol(A)1/d + vol(B)1/d

d
vol(A)1/d + vol(B)1/d .

This completes the induction step and proves the theorem.

Proof of Lemma 7.10. For general convex regions A, B Rd the following is a standard convergence argument. First, consider parallel hyperplanes xi = m/2n , m Z,
i [d]. They subdivide the regions into small interior cubes and partly-filled cubes
near the boundary. Denote by An and Bn the union of all cubes intersecting A
and B, respectively. Observe that A1T A2 . . . , and the same holds for {Bn }
and {An + Bn }. On the other hand, n An = A since for every point a
/ A there

68

exists an > 0, such that a cube with side containing a does not intersect A. This
gives vol(An ) vol(A) and vol(Bn ) vol(B), as n . In the topology defined
by the distance d(A, B) (see Exercise 7.1), this implies that An A and Bn B,
as n .
T
Finally, A+B n An +Bn since for every cn = an +bn c Rd , by compactness
of A, B there exist converging subsequences an a and bn b, and thus c = a + b
A + B. On the other hand, An + Bn A + B for all n N, which implies that
An + Bn A + B as n . Now we can apply the inequality to the brick
regions An and Bn :
vol(A + B)1/d = lim vol(An + Bn )1/d lim vol(An )1/d + lim vol(Bn )1/d
n

vol(A)

1/d

+ vol(B)

which completes the proof of the lemma.

1/d

,


7.8. Exercises.
Exercise 7.1. (Blaschke selection theorem) [1+] For every subset Rd , let C() be the
set of convex subsets of . Dene a distance d(X, Y ) between two convex sets X, Y C(Rd )
as follows:
d(X, Y ) := vol(X r Y ) + vol(Y r X).
Check that d(X, Y ) is a metric on C(Rd ). Prove that C() C(Rd ) is compact if and only
if set Rd is compact. Conclude that an innite family of convex subsets inside a unit
cube has a converging subsequence.
Exercise 7.2. Let X R3 be a convex polytope. Dene (X) to be the set of points
obtained by reections of one point in X across another: (X) = {2x y | x, y X }.
a) [1-] Show that (X) is also a convex polytope.
b) [1-] Consider a sequence X1 , X2 , . . ., where Xi+1 = (Xi ) and X1 is a cube. Compute
polytope Xn for all n > 1.
c) [1] Same question when X1 is a regular tetrahedron, octahedron and icosahedron.
d) [1] Same question when X1 is a standard tetrahedron and a Hill tetrahedron (see Subsection 31.6).
e) [1] Prove or disprove: for every xed X1 , the number of faces of Xn is bounded, as
n .
f ) [1+] Scale polytopes Yn to have the same unit volume: Yn = Xn /vol(Xn ). Prove that
the sequence Yn converges to a convex body.
Exercise 7.3. (Steiner symmetrization) Another one of several Steiners ingenious proofs
of the isoperimetry in the plane is based on the following construction. Take a convex set
X R2 and any line . Divide X into innitely many intervals [a, b], a, b X, orthogonal
to . Move each interval [a, b] along the line (a, b) so that it is symmetric with respect to
(See Figure 7.5). Denote by Y = S(X, ) the resulting set. The resulting map : X Y
is called the Steiner symmetrization.
a) [1] Prove that diam(Y ) diam(X), area(Y ) = area(X) and perimeter(Y ) perimeter(X).
b) [1+] Prove that C = (A) + (B) ts inside D = (A + B), for all convex A, B R2 .
c) [1] Prove that C is a translate of D if and only if A and B are homothetic.

69

Figure 7.5. The Steiner symmetrization : X Y .

d) [1+] Use the Steiner symmetrization to obtain two new proofs of Theorem 7.1, by employing both types of arguments, (I) and (II), discussed above.
Exercise 7.4. (Hammers X-ray problem) [2-] Let X, X R2 be two convex polygons
in the plane. Fix three lines 1 , 2 , 3 R2 whose normals are not rationally dependent.
Suppose S(X, i ) S(X, i ) for i = 1, 2, 3, where X Y if Y is a translate of X. Prove
that X X .
Exercise 7.5. [1] Here is yet another attempt at symmetrization. Let Q R3 be a convex
polygon containing the origin O. For every line through O, shift the interval Q to
make it centrally symmetric at O. Prove or disprove: the resulting body is convex.

Exercise 7.6. [1] Prove that of all n-gons in the plane with given perimeter the inscribed
regular n-gon has the largest area. Deduce from here the isoperimetry of a circle in the
plane.
Exercise 7.7. (Bonnesens inequality) [1+] Let Q R2 be a convex polygon, a = area(Q),
and = perimeter(Q). Let r and R denote the radius of the maximal circle inside Q and
of the minimal circle outside Q. Prove that 2 4 a 2 (R r)2 .
Exercise 7.8. (Spherical isoperimetric inequality) [2-] Let X be a simple spherical polygon
of length on a unit sphere S2 . Denote by a = a(X) the area of one of the sides. Prove the
isoperimetric inequality 2 a(4 a).
Exercise 7.9. [1] Let v1 , . . . , vn R2 be a set of vectors such that |v1 | + . . . + |vn | = 1.
Prove that there exists a subset I {1, . . . , n} such that
X 
1


vi  .


iI

Check that the 1/ constant is optimal.

Exercise 7.10. (Urysohns inequality) a [1] Let X be a convex body. Prove that area(X)

2
4 diam(X) , where the equality holds if and only if X is a disk.
b) [1+] Prove that area(X) 4 w(X)2 , where w(X) is the average width of X (the width is
dened as f (u) in the proof of Proposition 5.1).
c) [2-] Generalize parts a) and b) to higher dimensions.
Exercise 7.11. (Monotonicity of the area) a) [1-] Let Q1 , Q2 R2 be convex polygons,
such that Q1 is inside Q2 . Prove that the perimeter of Q1 is smaller than the perimeter
of Q2 .

70

b) [1-] Let P1 , P2 R3 be a convex polytopes such that P1 is inside P2 . Prove that

area(P1 ) area(P2 ).
Exercise 7.12. Let P1 P2 be two convex polytopes in R3 , and let L1 , L2 denote their
total edge lengths.
a) [1] Prove that L1 L2 when the polytopes are bricks.
b) [1] Prove that L1 L2 when the polytopes are parallelepipeds.
c) [1-] Prove or disprove that L1 L2 when the polytopes are tetrahedra.
d) [1] Prove or disprove that L1 L2 when the polytopes are combinatorially equivalent
polytopes with parallel corresponding faces.
Exercise 7.13. Let P R3 be a convex polytope which lies inside a unit sphere S2 . Denote
by L the sum of edge lengths of P , and by A the surface area.
a) [1] Prove that A L. Check that this inequality is tight.
b) [1] Let O be the center of S2 , and suppose O lies in the interior of P . For every
edge e = (v, w) of P and every face F containing e, denote by (e, F ) the interior angle
between F and triangle (Ovw). Prove that A L, where = maxe,F cos (e, F ).
Exercise 7.14. Let P R3 be a convex polytope such that a circle of radius r ts inside
each face.
a) [2] Denote by S the sum of squares of edge lengths of P . Prove that S 48 r 2 . Show
that the equality holds only for a cube.

b) [2] Denote by L the sum of edge lengths of P . Prove that L 12 3 r. Show that the
equality holds only for a tetrahedron.
Exercise 7.15. a) [2] Suppose a convex polytope P R3 contains a unit sphere. Prove
that the sum of edge lengths L 24. Show that the equality holds only for the cube.
b) [2-] Suppose P is midscribed around a unit sphere, i.e., all edges of P touch the sphere.
Compute L for all regular polytopes and for all n-prisms. Show that there exist polyhedra
with a smaller L.
Exercise 7.16. (Zonotopes) Let P R3 be a convex polytope whose faces are centrally
symmetric polygons. Such polytopes are called zonotopes.
a) [1-] Subdivide each face of P into parallelograms.
b) [1] Prove that P can be subdivided into parallelepipeds, dened as the Minkowski sum
of three independent vectors.
c) [1] Prove that P is centrally symmetric.
d) [1+] Prove that P is the Minkowski sum of the intervals.
e) [1] Suppose each face of P is a parallelogram. Prove that the number of faces of P is
two times a triangular number.
f ) [1] Use Exercise 2.3 to show that every polytope P R3 has at least 6 parallelograms.
7.9. Final remarks. The proof by Steiner is one of his ve proofs of the isoperimetry
of a disk (Theorem 7.1), published in 1841. The mistake remained unnoticed until 1882.
Another one of his proofs is based on the Steiner symmetrization (Exercise 7.3). In all
fairness to Steiner, the rigor of the work by contemporaries is questionable as well. For
more on Steiners work and other early results on isoperimetric problems see [Bla3, Kry].
The proof of the general isoperimetric inequality (Theorem 7.8) via the BrunnMinkowski
inequality goes back to Minkowski [Min]. Both results can be found in a number of textbooks
(see [BonF, Grub, Had1, H
or, Schn2]). We follow [Mat1, 12.2] in our presentation of the

71

rst part of Theorem 7.4. The second part is given in Subsection 41.5, and with all its
tediousness it is our own. The advantage of this proof is that it can be extended to a
variety of non-convex regions, e.g., connected regions with smooth boundary.
Let us also mention the AlexandrovFenchel inequality generalizing the BrunnMinkowski
inequality to mixed volumes of many dierent convex bodies [BZ3, 4.2], and the PrekopaLeindler inequality, which is a dimension-free extension [Grub, 8.5]. We refer to [Gar,
BZ3] for further generalizations, geometric inequalities, their applications and numerous
references.
Dierent proofs of the Blaschke selection theorem (Exercise 7.1) can be found in [BolY,
4.3], [Egg1, 4.2] and [Fej2, Ch. 2, 1]. A completely dierent approach to the isoperimetric inequalities via averaging is given in Exercise 24.7. The spherical analogue of the
isoperimetric inequality is given in Subsection 28.5.

72

8. Combinatorics of convex polytopes

In this section we study graphs and f -vectors of convex polytopes. Most proofs use
Morse functions, a powerful technical tool useful throughout the field and preludes
the variational principle approach in the next three sections. Among other things,
we prove the Dehn-Sommerville equations generalizing Eulers formula. This is just
about the only place in the book where we do exact counting. We also show that
the graphs of convex polytopes in R3 are 3-connected (see Balinskis theorem), an
important result leading to the Steinitz theorem in Section 11. We also introduce
the permutohedron and the associahedron (see Examples 8.4 and 8.5), which are used
later in Sections 16 and 17.
8.1. Counting faces is surprisingly easy. Define simple polytopes in Rd to be
polytopes where all vertices are adjacent to exactly d edges. These polytopes are
dual to simplicial polytope, whose faces are simplices. It is important to note that the
combinatorics of simplicial polytopes (the poset of faces of all dimensions) does not
change when we slightly perturb the vertices. Similarly, the combinatorics of simple
polytopes does not change when we slightly perturb the normals to facet hyperplanes.
Let P Rd be a simple polytope. Denote by fi = fi (P ) the number of idimensional faces, 0 i d. For example, fd = 1 and f0 is the number of vertices
in P . The sequence (f0 , f1 , . . . , fd ) is called the f -vector of P .
The generalized Eulers formula for all convex polytopes is the following equation:
() f0 f1 + f2 . . . + (1)d fd = 1 .

We will show that this is not the only linear relation on fi , but the first in a series.
Theorem 8.1 (DehnSommerville equations). The f -vector of a simple polytope P
Rd satisfies the following linear equations:
 


d
d
X
X
i
i i
di
(1)
fi =
(1)
fi , for all 0 k d.
k
dk
i=k
i=dk
Now observe that the case k = 0 corresponds to Eulers formula ().
Proof. Let : Rd R be a Morse function, defined as a linear function on Rd that is
nonconstant on edges of the polytope P . Let = (V, E) be the graph of P , where V
and E are the sets of vertices and edges, respectively. Since P is simple, the graph
is d-regular, i.e., the degree of every vertex is d.
Consider an acyclic orientation O of edges of according to (edges are oriented in the direction of increase of ). For a vertex v V let the index of v,
denoted ind (v), be the number of outgoing edges in O. Let hi be the number of
vertices v V with ind (v) = i. The sequence (h0 , h1 . . . , hd ) is called the h-vector
of P . We will show that the numbers hi depend only on and are in fact independent
of the Morse function .
For every face F of P denote by min (F ) the unique vertex v F with the smallest
value of . Looking at the acyclic orientation O this is the unique source point in
an induced subgraph |F . Let us count the number fk of k-dimensional faces in P by

73

looking at their minimum vertices. Observe that for all v V , every subset (even the
empty set) of ind (v) increasing edges spans a face. Summing over all vertices v V ,
for the number of such k-element subsets we obtain:


d  
X
X
ind v
i
()
fk =
=
hi .
k
k
i=k
vV, ind (v)k

Writing
F (t) =

d
X

fi ti ,

i=0

H(t) =

d
X

hi ti ,

i=0

we can rewrite () as F (t) = H(t + 1). Indeed,

" i   #
" d   #
d
d
d
X
X
X
X i
X i
tk =
hi
hi
tk
H(t + 1) =
hi (t + 1)i =
k
k
i=0
i=0
k=0
k=0
i=k
=

d
X
k=0

tk fk = F (t).

This implies that the h-vector of P as defined above is independent of the Morse
function . Now consider the Morse function = and observe that ind (v) =
d ind (v), for all v V . Thus, hk = hdk . Expanding H(t) = F (t 1), we obtain:
 
i
hk = (1)
(1)
fi
k
i=k
k

d
X

dk

and hdk = (1)

d
X

(1)

i=dk

which implies the result.


i
fi ,
dk

Remark 8.2. The combinatorial interpretations of the integers hi in the proof above imply
the inequalities hi 0 for all 0
i d. For example, we have h0 = 1, h1 = fd1 d 0,
and h2 = fd2 (d 1)fd1 + d2 0. In fact, much stronger inequalities are known:
h0 h1 h2 . . . hr , where r = d/2.

For example, h1 h0 gives an obvious inequality for the number of facets in P : fd1 d+1.
On the other hand, already h2 h1 gives an interesting inequality:


d+1
fd2 dfd1
.
2
For more on this inequality see Remark 27.4.

8.2. A few good examples would not hurt. The Dehn-Sommerville equations
above may seem rather pointless, but in fact the Morse function approach can be
used to compute the f -vectors. Below we present three such examples for some
combinatorial polytopes.

74

Example 8.3. (Hypercube) Let Cd Rd be a standard d-dimensional cube [0, 1]d . Consider the Morse function (x1 , . . . , xd ) = x1 + . . . + xd . Clearly, ind (x1 , . . . , xd ) = d


(x1 , . . . , xd ), for all xi {0, 1}, and we obtain hi = di . Therefore,
d  
X
d i
t = (1 + t)d ,
H(t) =
i
i=1

and

F(t) = H(t + 1) = (2 + t) =

d
X

di

i=1

 
d i
t .
i

The resulting formula fi = 2di di can, of course, be obtained directly, but it is always nice
to see something you know obtained in an unexpected albeit complicated way.
(321)
(111)

(000)

(231)

(312)

(132)

(213)
(123)

Figure 8.1. Acyclic orientations O of a cube C3 and a permutohedron P3 .

Example 8.4. (Permutohedron) Let Pn Rn be a permutohedron, dened as the convex
hull of all n! permutation vectors ((1), . . . , (n)). Note that Pn is simple and (n 1)dimensional. The graph n of Pn is the Cayley graph of the symmetric group with the
adjacent transpositions as the generating set:


n = Cayley Sn , {(1, 2), (2, 3), . . . , (n 1, n)} .
Consider the Morse function

(x1 , x2 . . . , xn ) = x1 + x2 + . . . + n1 xn .
If = (n) > 0 is suciently small, the resulting acyclic orientation O of n makes a
partial order on permutations that is a suborder of the lexicographic order. In particular,
the identity permutation (1, 2, . . . , n) is the smallest and (n, n 1, . . . , 1) is the largest
permutation. Now observe that for every permutation Sn the number of outgoing
edges ind () is equal to the number of ascents in , i.e., the number of 1 i < n such
that (i) < (i + 1). Therefore, hi is equal to the Eulerian number A(n, i), dened as
the number of permutations Sn with i ascents. We refer to Exercise 8.17 for further
properties of Eulerian numbers.
Example 8.5. (Associahedron) Let Q R2 be a xed convex n-gon (not necessarily
regular), and let Tn be the set of triangulations of Q. We denote the
vertices of Q by
2n
1
integers i [n]. Euler proved that |Tn | = Cn2 , where Cn = n+1
are the Catalan
n
numbers. For every vertex v Q in a triangulation Tn , denote by (v) the sum of

75

areas of triangles in that contain v. Let Rn Rn be a convex hull of all Cn2 functions
: [n] Rn , dened as v (v), for all Rn . The polytope Rn is called an
associahedron.

2
3

Figure 8.2. Triangulation T5 with (2) = area(124) + area(234),

and an acyclic orientation O of the associahedron R5 .
One can check that Rn is in fact (n3)-dimensional; for example R4 is an interval and R5
is a pentagon (see Exercise 8.2). The edges of Rn correspond to pairs of triangulations which
dier in exactly one edge, so Rn is a simple polytope. More generally, the k-dimensional
faces of Rn correspond to subdivisions of Q with (n 3 k) noncrossing edges, and the
containment of faces is by subdivision. As in Example 8.4, consider the Morse function
: Rn R dened by
(x1 , x2 , . . . , xn ) = x1 + x2 + . . . + n1 xn ,

where > 0 is small enough. This denes an acyclic orientation O of the graph n of the
associahedron (see Figure 8.2). For example, it is easy to see that the triangulation with
edges {(1, 3), . . . , (1, n 1)} is maximal with respect to .
Recall the standard bijection : t from polygon triangulations to binary trees,
where the triangles correspond to vertices, the triangle adjacent to (1, n) corresponds to
the root, and the adjacent triangles correspond to the left and right edges to the tree (see
Figure 8.3). Observe that the number of outgoing edges ind ( ) is equal to the number of
left edges in ( ). Therefore, hi is equal to the number Bn2,i of binary trees with n 2
vertices and i left edges (see Exercise 8.18 for a closed formula).
n

Figure 8.3. Map : t. Here n = 14, ind ( ) = 7.

76

8.3. Kalais simple way to tell a simple polytope from its graph. The proof of
the Dehn-Sommerville equations shows that given an acyclic orientation O coming
from a Morse function we can compute the f -vector of a simple polytope. It turns
out that we can obtain the whole poset of faces given just the graph of the polytope.
Theorem 8.6 (BlindMani). Let P Rd be a simple polytope and let be its graph.
Then determines the entire combinatorial structure of P .
In other words, given a graph one can decide whether a given subset of vertices
forms a k-dimensional face. The proof we present below is the celebrated proof by
Kalai which uses insight of Morse functions to resolve the problem.
Proof. For a subset of vertices B V of the graph to form a k-dimensional face
the induced subgraph |B must be connected and k-regular. However, this is clearly
not enough. For example, a graph of 3-dimensional cube contains a cycle of length 6
which is not a two-dimensional face. We will use certain acyclic orientations of to
give a complete characterization.
Consider an acyclic orientation O of . We say that O is good if for every face F
P , the corresponding graph F has exactly one source. Otherwise, O is called bad.
Note that in every face F , the acyclic orientation of F must have at least one source.
Denote by hO
i the number of vertices of index i, i.e., the number of vertices with
exactly i outgoing edges and let
O
O
d O
(O) = hO
0 + 2 h1 + 4 h2 + . . . + 2 hd .

Finally, let = (P ) = f0 +f1 +. . .+fd be the total number of faces of all dimensions.
Arguing as in the proof of the Dehn-Sommerville equations, we conclude:
(O) for all O,
(O) = if and only if O is good.
To show this, count every face of every dimension according to the source vertices.
Clearly, for a vertex v with index i, the number of such faces is at most 2i , the number
of all possible subsets of edges coming out of v. Now observe that good orientations
are exactly those where every face is counted exactly once.
This gives us a (really slow) algorithm for computing (P ): compute (O) for all
acyclic orientations O of and take the smallest value. Of course, the total number
of faces is only a tiny fraction of the information we want. What is more important is
that this approach gives a way to decide whether a given orientation is good. Having
found one good orientation O,
we can take the numbers hO
i to compute the f -vector
Pd
O k
by the formula fi = k=i hk i . Still, this approach does not show how to start with
a good orientation of , choose i edges coming out of some vertex v and compute the
whole face. Morse functions again come to the rescue.
Let B V be a subset of vertices, a candidate for a face. We say that B is final
with respect to an acyclic orientation O if no edge is coming out of B. We are now
ready to state the main claim:
Face criterion. An induced connected k-regular subgraph H = B on a set of
vertices B is a graph of a face of P if and only if there exists a good orientation O
of G, such that B is final with respect to O.

77

The only if part can be proved as follows. Take a simple polytope P and a face F
with the set of vertices B V . Consider any linear function which maximizes
on F . Now perturb to make it a Morse function but keep B final with respect to .
This is clearly possible and gives the desired construction.
The if part is more delicate. Let v be any source point of the graph H with
respect to O (for now we do not know if such a point is unique). Take all k edges
in H that are coming out of v. Since O is good, there exists a k-dimensional face F
with v as its minimum. Since B is final with respect to O, we have F B and the
graph Y = F is a subgraph of H. Finally, since P is simple and F is k-dimensional,
the graph Y of is also k-regular.
Now compare graphs H and Y . Both are connected, k-regular, and Y H. This
implies that Y = H, which proves the face criterion and the theorem.20

Remark 8.7. A couple of things are worth noting while we are on the subject. First, the
above criteria gives a really slow algorithm for deciding if a subset is a facet. A polynomial
time algorithm was recently found in [Fri]. Second, one cannot easily use this idea to decide
whether a given graph is a graph of a simple polytope; this problem is believed to be very
dicult (see [Kaib]).
8.4. Graph connectivity via Morse functions. It is well known and easy to prove
that graphs of 3-dimensional polytopes are 3-connected, but the corresponding result
for d-dimensional polytopes is less standard. The proof again is based on Morse
functions.
A graph is called d-connected if the removal of any d 1 vertices leaves a connected
graph.
Theorem 8.8 (Balinski). The graph of a d-dimensional polytope P is d-connected.
Proof. Let V be a set of vertices in and let X = {x1 , . . . , xd1 } be any subset of d1
vertices. We need to prove that |V rX is connected. Fix a vertex z V r X and let
: Rd R be a linear function which is constant on X + z:
(x1 ) = . . . = (xd1 ) = (z) = c ,

for some c R. The set of vertices v

/ (X + z) is now split into two subsets:
V+ = {v V | (v) > c} and V = {v V | (v) < c}. There are two cases:
1) One of the sets V or V+ is empty. Say, V = . Perturb to obtain a Morse
function : Rd R (i.e., a function which is nonconstant on the edges of P ). Orient
all edges in , from a vertex with smaller to a vertex with larger value of , and
denote by w+ the maximal vertex. Now observe that for every vertex v V+ w+
there is an edge coming out of v. Continue walking along the path of increasing edges
until we eventually reach the unique maximum w+ . Similarly, since |X| < d, vertex z
also has an edge connecting it to V+ . Therefore, all vertices in V r X are connected
to w+ , which finishes the proof of this case.
20While

the implication is obvious in this case, it can be viewed as a special case of a general local
global principle. In a very similar context this idea also appears in the proof of Lemmas 14.7
and 35.1.

78

w+

x1
x2
z

xd1

=c

Figure 8.4. An acyclic orientation in the proof of Balinskis theorem.

2) Suppose V+ , V 6= . Again, perturb to obtain a Morse function : Rd R.
By the argument above, all vertices in V+ are connected to the maximum w+ and
all vertices in V are connected to the minimum w . It suffices to show that z is
connected to both V+ and V , since that would make all vertices connected (see
Figure 8.4).
Suppose that z does not have an edge connecting it to V . Making a small perturbation of , we can make z the minimum in X + z. By the assumption, all edges
adjacent to z either connect it with X or V+ . Therefore, all edges adjacent to z are
increasing, and z is the minimum in V . On the other hand, we still have (v) < (z)
for all vertices v V , a contradiction. This completes the proof.

8.5. Exercises.
Exercise 8.1. [1-] Denote by Fd Rd+1 the subspace spanned by all f -vectors of convex
polytopes in Rd . Prove that the DehnSommerville equations form a basis in the dual to the
quotient space Rd+1 /Fd . In other words, prove that the equations are linearly independent
and imply all linear equations which hold for all f -vectors.
Exercise 8.2. (Associahedron) a) [1] Prove that the associahedron Rn dened in Example 8.5, is (n 3)-dimensional.
b) [1] Prove that the edges of Rn correspond to diagonal ips.
c) [1] Give a complete combinatorial description of the facets of Rn .
Exercise 8.3. [1-] Let P Rd be a convex polytope and let be its graph. Suppose x, y
are vertices of P lying on the same side of a hyperplane L. Then there exists a path from x
to y in that does not intersect L.

79

Exercise 8.4. [1+] Let P Rd be a convex polytope with n vertices. For a vertex
v P denote by Cv the innite cone with a vertex at v, which is spanned by the edges of P
containing v. Prove that P is the intersection of at most n d + 1 cones Cv . Deduce from
here the Balinskis theorem (Theorem 8.8).
Exercise 8.5. Let be a graph of a convex polytope P Rd .
a) [1-] Prove that if P R3 is simplicial, then the edges of P can be colored with two
colors so that between every two vertices there are monochromatic paths of both colors.
b) [1-] Find all simple polytopes P R3 for which such a 2-coloring exists.
c) [1] Generalize part a) to simplicial polytopes in Rd and d 1 colors.

Exercise 8.6. [1] Let P Rd be a convex polytope and let be its graph. Prove that
there is an embedding of Kd+1 into , i.e., there exist d + 1 dierent vertices in P every two
of which are connected by pairwise disjoint paths in .
Exercise 8.7. [1] Let P R3 be a convex polytope with even-sided faces. Prove or
disprove: the edges of P can be 2-colored so that every face has an equal number of edges
of each color.

Exercise 8.8. a) [1-] Prove that every convex polytope P R3 has a vertex of degree
at most 5.
b) [2-] Prove that every convex polytope P R3 has two adjacent vertices whose total
degree is at most 13.
c) [2-] Prove that every convex polytope P R3 with even-sided faces has two adjacent
vertices whose total degree is at most 8.
d) [1] Show that the upper bounds in part b) and c) are sharp.
Exercise 8.9. a) [1] Let P R3 be a convex polytope and let L be the sum of edge lengths
of P . Prove that L 3diam(P ).
b) [1+] Let be the graph of a convex polytope P R3 with n vertices. Prove that the
diameter of is at most (n 1)/3. Show that this inequality is sharp.
Exercise 8.10. Let be the graph of a simple convex polytope P R3 with triangular
and hexagonal faces.
a) [1-] Prove that P has exactly four triangular faces.
b) [2-] Prove that the number of hexagonal faces P is even.
c) [2] Let P R3 be a simple polytope, whose faces have 3, 6, 9, . . . sides. Prove that the
number of faces of P is even.
Exercise 8.11. [1-] Let Q R2 be a closed (possibly self-intersecting) polygonal curve,
such that all angles are < . Fix a Morse function, i.e., a linear function : R2 R which
is non-constant on the edges of Q. Prove that Q is a convex polygon if and only if has a
unique local maximum.
Exercise 8.12. (Neighborly
polytopes) [1] Construct explicitly a convex polytope P R4
6

with 6 vertices and 2 = 15 edges. In general, we say that a polytope is neighborly if its
graph is a complete graph. For all d 4 construct a neighborly polytope in Rd with d + 2
vertices.
Exercise 8.13. [1] Let P Rd be a convex polytope with n d + 2 vertices. Prove that
there exists a vertex v P and two facets, such that v
/ F1 , F2 . Moreover, we can make F1
and F2 adjacent, which means that their intersection must be a (d 2)-dimensional face.

80

Exercise 8.14. [1] Let P Rd be a convex polytope with n d + 2 vertices. Prove that
for every k, 1 k d 1, there exist a k-face and a (d k)-dimensional face of P which
are disjoint.
Exercise 8.15. Consider all cross sections of a convex polytope P R3 which do not
contain any vertices.
a) [1-] Can all these cross sections be triangles?
b) [1-] Can all these cross sections be quadrilaterals?
c) [1-] Can all these cross sections be odd-sided polygons?
d) [1] Suppose all vertices of P have even degree. Prove that every cross section of P is an
even-sided polygon.
Exercise 8.16. Denote by () the length of the longest cycle in the graph (cycles do
not have repeated vertices).
a) [1-] Find a convex polytope P with n vertices, such that the graph = (P ) has
() < n.
b) [1+] Construct a sequence of convex polytopes {Pk } with graphs k = (Pk ) on nk
vertices, such that ni , and (k ) = O(n1
k ) as k , for some > 0.
Exercise 8.17. (Eulerian numbers) In the notation of Example 8.4, let A(n, k) be the
number of permutations Sn with k ascents. The integers (A0 , A1 , . . . , An1 ) are called
Eulerian numbers.
a) [1-] Prove that Eulerian numbers satisfy:
A(n, k) = (n k)A(n 1, k 1) + (k + 1)A(n 1, k).
b) [1] Prove that the Eulerian numbers satisfy
Ak = n! vol(Qn,k ),
where Qn,k Rn is a convex polytope dened by the inequalities
k x1 + . . . + xn k + 1,

0 xi 1, 1 i n.

c) [1] Observe that that the Eulerian numbers satisfy

A(n, k) = n! vol(Pn,k ),
where Pn,k Rn is a union of simplices 0 x(1) . . . x(n) 1, over all Sn with
exactly k ascents. Find an explicit piecewise-linear volume-preserving map : Qn,k Pn,k .
This gives an alternative proof of b).
Exercise 8.18. (Narayana numbers) In the notation of Example 8.5, let Bn,k be the
number of binary trees with n vertices and k left edges. Numbers Bn,k are called Narayana
numbers.
a) [1-] Use induction to prove a closed formula:
 

1 n
n
Bn,k =
.
n k
k+1
b) [1] Use binomial identities to obtain a closed formula for the f -vector of associahedron Rn :



1
n3
2n d 4
fi =
.
nd2 nd3
nd3

c) [1-] Use induction to give an alternative proof of this formula.

81

Exercise 8.19. The Birkho polytope Bn is dened by the following equations and
inequalities:
n
n
X
X
ai,j =
ai,j = 1, and ai,j 0, for all 1 i, j n.
i=1

j=1

a) [1-] Prove that Bn has dimension (n 1)2 .

b) [1] Prove that Bn has n! vertices which correspond to permutation matrices.
c) [1] Prove that the edges of Bn correspond to multiplication of permutations by single


cycles. Conclude the graph of Bn has diameter two, and the vertices have degree (n1)! .
d) [1-] Check that B3 is both neighborly (see Exercise 8.12) and simplicial, and that Bn ,
n 4, are neither.
e) [1-] Compute vol(B3 ).
f ) [2-] Let G be a subgraph of Kn,n . We say that G is elementary if every edge of G
belongs to some perfect matching in G. Prove that the poset of elementary subgraphs
of Kn,n ordered by inclusion is isomorphic to the face lattice of Bn .
g) [2-] Describe the f -vector of Bn in terms of elementary graphs.
Exercise 8.20. Dene the transportation polytope Tm,n by the following equations and
inequalities:
m
n
X
X
ai,j = m,
ai,j = n, and ai,j 0, for all 1 i m, 1 j n.
i=1

j=1

a) [1-] Prove that Tm,n has dimension (m 1)(n 1).

b) [2-] Prove that Tm,n is simple for all (m, n) = 1.
c) [1] Show that Tn,n+1 has (n + 1)n1 n! vertices. To do that, give a bijective proof between
these vertices of Tn,n+1 and labeled trees on n + 1 vertices, where both vertices and edges
are labeled, and use Cayleys formula.
d) [1] Find a complete description of the edges of Tn,n+1 in terms of these labeled trees.
e) [1+] Use the Morse function

approach to compute h1 and h2 of the simple polytope Tn,n+1 .
From here, obtain f2 Tn,n+1 .

Exercise 8.21. (Klyachko lemma) [1+] Let P R3 be a convex polytope. Fix > 0.
Suppose on every facet F in P , there is a particle which moves clockwise along the edges
of F with the speed at least . Prove that at some point some two particles will collide.
Exercise 8.22. (3d -conjecture) Let P Rd be a centrally symmetric convex polytope
with f -vector (f0 , f1 , . . . , fd ).
a) [1+] For d = 3, prove that f0 + f1 + f2 + f3 27.
b) [2+] For d = 4, prove that f0 + f1 + f2 + f3 + f4 81.
c) [] Prove that f0 + f1 + f2 + . . . + fd 3d , for all d 1.
8.6. Final remarks. The generalization () of Eulers formula is due to Poincare (1897).
The DehnSommerville equations (Theorem 8.1) in dimensions 4 and 5 were discovered
by Dehn (1905) and generalized to all dimensions by Sommerville (1927). Note that in
the original form, the equations gave a dierent basis in the ane space spanned by the
f -vectors of all simple polytopes in Rd (cf. Exercise 8.1). Our proof is standard and can
be found in [Br] (see also [Barv, Vi.6]), while a traditional proof can be found in [Gr
u4,
9.2]. See also [Zie1, 8.3] for the dual treatment (called shelling) of simplicial (rather than

82

simple) polytopes, as this approach is connected to further combinatorial questions and

ideas.
The proof of Theorem 8.6 is essentially the same as Kalais original argument [Kal1]
(see also [Zie1, 3.4]). Our proof of Balinskis theorem (Theorem 8.8) is a variation on the
original proof [Bali] and the argument in [Bar3] (see also [Zie1, 3.5] and Exercise 8.4), but
the use of Morse functions allows us to avoid some unnecessary technicalities.
Eulerian numbers appear in numerous instances in enumerative combinatorics; it is one
of the two major statistics on permutations.21 We refer to [Sta3] for other appearances
of Eulerian numbers, Catalan numbers and polygon subdivisions. The associahedron is
also called the Stashe polytope; in our denition we follow [GKZ, Chapter II.7]. The
calculations in the examples are partly ours and partly folklore. For more on permutohedron
and associahedron see [EKK, Zie1] and references therein. See also Section 14 for more on
triangulations.
Finally, let us mention that Morse functions are basic tools in Morse theory, a classical
branch of topology. We refer to [For] for a friendly treatment of the discrete case and further
references.

21This

statistic is also called Eulerian. The other major statistic on Sn , by the number of
inversions in permutations, is called Mahonian, after British combinatorialist Percy MacMahon.

83

9. Center of mass, billiards and the variational principle

This is our first introduction to the variational principle approach. After several elementary results on the center of mass of convex polytopes, we explore closed billiards
trajectories in smooth convex bodies and in convex polygons. These results lead the
way to the analysis of closed geodesics on convex polytopes in the next section. The
variational principle approach will also be used throughout the book.
9.1. How to roll a polytope. Here is a basic test whether your intuition is combinatorial/algebraic or physical/geometric: try proving the following result.
Theorem 9.1. Let P Rd be a convex polytope, and let z P be any point. Prove
that there exists a facet F such that an orthogonal projection of z onto F lies inside F .
The physical intuition gives an easy solution: think of z as the center of mass
of P . If not, use inhomogeneous material in P to make it so. Drop the polytope on
a (hyper)plane. If the orthogonal projection of the center of mass is not inside the
bottom facet F of the polytope, it will start rolling. Since it cannot roll indefinitely,
the polytope eventually stabilizes, and we are done.
In order to convert the above argument into a honest proof we need to have a
basic understanding of the physical nature of rolling. The following will suffice for
the intuition: the height of center of mass drops at each turn. Now that we know
what to look for, we can create a simple infinite descent style proof:
Proof. Let F be the facet such that the distance from z to a hyperplane spanned
by F is the smallest. Suppose an orthogonal projection w of z on a hyperplane H
containing F does not lie inside F . Let A be any (d 2)-dimensional face of F such
that the subspace spanned by A separates H into two parts, one containing F and
the other point w. Denote by F 6= F the other facet containing A. Observe that the
distance from z to F is smaller than the distance from z to F . This is clear from
the 2-dimensional picture on an orthogonal projection along A (see Figure 9.1). This
contradicts the assumptions.


z
F
F

A w

Figure 9.1. A polytope that is about to roll twice; projections of z

onto facets.

84

9.2. Meditation on the rules and the role of rolling. Let us further discuss
the merits and demerits of the above physical proof vs. the formal proof of the
admittedly easy Theorem 9.1. Basically, the physical proof comes with an algorithm
for finding a desired facet F : start at any facet F0 and move repeatedly to an adjacent
facet that is closer to z, until this is no longer possible. This algorithm, of course,
can be formalized, but some questions remain.
First, on the complexity of the algorithm. Clearly, not more than fd1 (P ) (the total
number of facets) of rolls is needed. In fact this is tight up to an additive constant22.
On the other hand, the algorithm does not necessarily finds the closest facet used in
the formal proof.
Now, note that the minimum plays a special role in the proof and cannot be replaced
by the maximum. The situation is strikingly different in the smooth case. We will
restrict ourselves to the dimension two case.
Let C R2 be a smooth closed curve in the plane and let O be the origin. Throughout the section we will always assume that O
/ C. For a point x C we say that
the line segment (O, x) is normal if it is orthogonal to C at point x.
Proposition 9.2. Every smooth closed curve C has at least two normals from every
point O
/ C.
Proof. Let f (x) denote the distance from x C to the origin O. Take any (local or
global) extremal point x (minimum or maximum) of f . Observe that (O, x) is normal,
since otherwise close points on the side with angle > /2 are at distance > |Ox|, and
on the side with angle < /2 are at distance < |Ox| (see Figure 9.2).

O

Figure 9.2. Normal points on a curve.

Note that we already used this argument in the proof of Theorem 3.3 (compare
Figures 3.3 and 9.2). One can again give similar physical motivation: when C is
convex, think of it as a two-dimensional die filled to make any interior point its
center of mass. Place it on a line and let it roll. The (local) minima and maxima
correspond to stable and unstable equilibrium points, respectively.
Since we are really interested in discrete geometry, let us give a discrete version
of Proposition 9.2. Let Q R2 be a polygon in the plane, possibly with selfintersections. As before, we assume that O
/ Q. An interval (O, x), x Q, is
22Try

to nd an example in any xed dimension!

85

called quasi-normal if the function |Ox| is a local minimum or maximum at x.23

When x lies on an edge e of Q this means that (O, x) is orthogonal to e, so in fact
point x is normal. When x is a vertex of Q incident to edges e, e , this means that the
angles between (O, x) and e, e are either both acute or both obtuse. We immediately
have the following result.
Proposition 9.3. For every plane polygon Q, there exist at least two quasi-normals
to Q from every point O
/ Q.
Example 9.4. Proposition 9.3 implies that a convex polygon containing the origin has at
least one normal and at least one quasi-normal. Figure 9.3 shows that these bounds cannot
be improved.

Figure 9.3. Square A and point O

/ A with two quasi-normals; heptagon B and point O B with one normal and one quasi-normal.
9.3. Center of mass of polygons. As the example above shows, there exists a
convex polygon Q and an interior point O with a unique normal. The following result
shows that this is impossible when O in the center of mass of the region inside Q.
Theorem 9.5. Let Q = A be a convex polygon, A R2 , and let z = cm(A) be
its center of mass. Then there exist at least two normals and at least two other
quasi-normals from z to Q.
In other words, if a convex polygon Q is filled uniformly and rolled on a line, then
it can stand on two or more edges and has an unstable equilibrium at two or more
vertices. The proof is based on the following elementary lemma.
Lemma 9.6. Let X, Y R2 be two convex sets of equal area such that their centers of
mass coincide: cm(X) = cm(Y ). Then X and Y intersect in at least four points.
Proof. Clearly, since area(X) = area(Y ), there are at least two points of intersection
in X Y . Assume that a and b are the only intersection points. Let Z = X Y ,
X = X Y , Y = Y X. Observe that
cm(X) = cm(Z) + (1 ) cm(X ),

23Note

that the notion of a quasi-normal can be applied to all convex sets, not just convex
polygons. All results extend verbatim in this case. We will not use this generalization.

86

cm(Y ) = cm(Z) + (1 ) cm(Y ),

where

area(Z)
area(Z)
=
.
area(X)
area(Y )
Since cm(X) = cm(Y ), we conclude that cm(X ) = cm(Y ) which is impossible
since X and Y lie on different sides of the line through a and b (see Figure 9.4). 
=

a
X

Z
b

Figure 9.4. Two polygons of equal area with two boundary intersection points; a polygon and a circle of equal area with the same center
of mass.
In a special case, the lemma gives the following corollary.
Corollary 9.7. Let X R2 be a convex set and let C be a circle centered at the
center of mass cm(X) with area equal to the area of X. Then C and X intersect in
at least four points.
Proof of Theorem 9.5. Let Q = A be a convex polygon as in the theorem and let C
be a circle of equal area centered at z = cm(A) as in the corollary. Since there are
at least four points of intersection of Q and C, the distance function f (x) = |zx| has
at least two different local maxima and two different local minima (see Figure 9.4).
This implies the result.

Observe that by analogy with the previous section, we can now easily adapt the
results to the smooth case:
Theorem 9.8. Let A R2 be a convex region bounded by a smooth convex curve
C = A, and let z = cm(A) be its center of mass. Then there exist at least four
normals from z to C.
Example 9.9. (Monostatic polytopes) One can ask whether Theorem 9.5 remains true in
higher dimensions. As it happens, it does not. Homogeneously lled polytopes in R3 which
are stable on only one face are called monostatic. Here is the construction of one such
polytope. Take a long cylinder with symmetrically slanted ends so that the resulting body
can be static only in one position as in Figure 9.5. One can use a polyhedral approximation
of the cylinder to obtain a monostatic polytope (see Exercise 9.12). Interestingly there are
no monostatic tetrahedra in R3 , but in higher dimensions there exist monostatic simplices
(see Exercise 9.10).
In a dierent direction, let us mention that there is no natural generalization of Lemma 9.6
to higher dimensions. Namely, there exist convex polytopes P and P of equal volume, such

87

Figure 9.5. Monostatic slanted cylinder and two intersecting triangles.

that cm(P ) = cm(P ) and the intersection P P has only one connected component. For
example, take two congruent triangles in orthogonal planes and arrange them as shown in
Figure 9.5. Now replace triangles with thin triangular prisms.

9.4. Special cuts of special cakes. In the previous subsection we showed that for
every smooth convex curve C R2 , there exists a point (the center of mass of the
interior region) from which there exist at least four normals to C. On the other hand,
as the smoothened version of the heptagon in Figure 9.3 shows, not all points z in the
interior must have four normals to C. In fact, in most cases, there are infinitely many
points which have four or more normals (cf. Exercise 9.7). Below we find another
point z with four normals, special in its own way.
Let C R2 be a smooth closed curve in the plane. A cut is an interval connecting
two points on a curve. We say that a cut (x, y) is a double normal if C is orthogonal
to (x, y) at both ends. Think of the region inside C as a cake and of double normals
as of a way to cut the cake.
Theorem 9.10 (Double normals). Every smooth closed curve in the plane has at
least two double normals.
Let us discuss the result before proving it. First, finding one double normal is easy:
take the most distant two points on a curve. By Proposition 9.2, this cut must be
orthogonal to the curve at both ends. However, finding the second double normal is
less straightforward. Even though it is intuitively clear that it has to correspond to
some kind of minimum rather than maximum, one needs care setting this up.

Figure 9.6. Two double normals and a family of parallel cuts K F .

88

Proof. Denote by C R2 the curve in the theorem. Consider families F of cuts K

of C, with the slope [0, 2]. Define
() c = min max |K|,
[0,2] KF

where |K| denotes the length of the cut. By compactness, the min max condition is
achieved on some cut K = (x, y), i.e., there exists an angle and a cut K F such
that |K| = c. We claim that this cut is a double normal.
First, observe that the incidence angles of (x, y) are the same at both ends, i.e.,
the tangent lines Lx and Ly at x and at y are parallel. Indeed, otherwise consider
a parallel cut K F shifted in the direction where the sum of the angles is >
(see Figure 9.7). By the smoothness of C, for small shifts the length |K | > |K|, a
contradiction.
Ly
y

y
K

K
x

Lx

Ly
K

Lx

Figure 9.7. Checking whether cut K is a double normal.

Now suppose the tangent lines Lx and Ly are parallel. By convexity, the whole
curve C lies between these lines. If K = (x, y) is not orthogonal to the curve, take a
family of cuts F of cuts perpendicular to the lines. Clearly, every cut, in particular
the maximal cut K F , has length < (K). Therefore, cut K does not satisfy the
min max condition of (), a contradiction.

There is a simpler way to think of the proof: define width(C) to be the smallest
distance between parallel lines which touch C on different sides. Then one can consider
two such lines and proceed as in the last part of the proof above. However, the
advantage of the longer proof is the insight it gives, which will be used in the next
section. Let us mention also the following straightforward discrete analogue of the
double normals theorem.
Theorem 9.11. Every convex polygon in the plane has at least two double quasinormals.
The proof follows verbatim the proof of Theorem 9.10. Of course, unless the polygon
has parallel edges, it does not have double normals.

89

9.5. Playing pool on convex tables. One can think of the double normals as of a
billiard shot which returns back after two bounces off the curve boundary. Naturally,
one can ask for more general shots with three or more bounces. Formally, a polygon
inscribed into a smooth curve C is called a periodic billiard trajectory if the incidence
angles of the adjacent sides are equal (see Figure 9.8). Note that given one periodic
billiard trajectory, one can take a power of such trajectory, repeating it several times.
Trajectories which cannot be obtained that way are called irreducible.

Figure 9.8. A periodic billiard trajectory.

Theorem 9.12 (Birkhoff). Every smooth convex curve has infinitely many irreducible
periodic billiard trajectories.
Proof. Suppose C is a smooth convex curve and let n 3 be a prime. Let us show
that C has a periodic billiard trajectory with exactly n sides. We use maximization
over all polygons inscribed into C.
Denote by Rn the set of all (possibly self-intersecting) k-gons inscribed into C,
with k n. Let
(>) c = max (R),
RRn

where (R) is the sum of edge lengths of R. We claim that the maximum is achieved
on an irreducible periodic billiard trajectory R with exactly n sides. First, note
that Rn C n is a compact in R2n , so the maximum is well defined. Second, the
maximum is achieved on an n-gon, since otherwise one can replace one edge (x, y)
with any two edges (x, z) and (z, y), increasing the length of a polygon R. Finally,
since n is prime, the maximum is not a power of a smaller trajectory.
Lz
z
y

x
C
E

Figure 9.9. Edges (x, z) and (z, y) of the inscribed polygon R.

90

Now, for every two adjacent sides (x, z) and (z, y) of the polygon R, we will show
that the incidence angles with the curve are equal. Indeed, denote by E the ellipse
with x and y as its focal points and z on the boundary. In other words, ellipse E is
the set of points w R2 with |xw| + |wy| = |xz| + |zy|, and for all points v outside E
we have |xv| + |vy| > |xz| + |zy|. Since the polygon R is maximal, the ellipse E
must contain the curve C; otherwise we can replace (x, z) and (z, y) with some (x, v)
and (v, y). Since C is smooth, this implies that E and C must be touching each other
at z C, E with the same tangent line Lz . The result follows from the equal angle
property of ellipses (see Figure 9.9).

Theorem 9.13 (Birkhoff). For every smooth convex curve C = X and every two
points x, y X, there exist infinitely many distinct billiard trajectories from x to y.

This result is formally not a corollary from Theorem 9.12, but follows directly from
the argument in the proof above (see Exercise 9.3).
9.6. Playing pool with sharp corners. There are several ways to define billiard
trajectories in polygons, since it is unclear what happens if the trajectory enters a
vertex (the edge reflections are defined as before). One way is to simply forbid the
trajectory to go through the vertices. We call this the (usual) billiard trajectory in a
polygon.
Conjecture 9.14. Every convex polygon has at least one periodic billiard trajectory.
This conjecture is classical and holds in many special cases (see Exercises 9.13
and 9.14), but is open even for triangles.24 Here is another closely related conjecture:

Conjecture 9.15. For every convex polygon X R2 and every two points x, y X,
there exists infinitely many distinct billiard trajectories from x to y.

Figure 9.10. Angles of reflection in a quasi-billiard trajectory R.

Let us consider now another variation on the theme which will prove useful in
the next section. A quasi-billiard trajectory in a convex polygon is defined by the
following rule: when a trajectory enters a vertex of angle , it satisfies
()

| | ,

where and are the reflection angles as in Figure 9.10. The following analogues of
Birkhoff theorems follows along the same lines as the proof of Theorem 9.12.
24For

triangles with angles at most 100 degrees this was proved by R. E. Schwartz (2005).

91

Theorem 9.16. For every convex polygon Q R2 , there exists a periodic quasibilliard trajectory inscribed into Q with exactly n sides, for every n 3.
Theorem 9.17. For every convex polygon Q = A, A R2 , and every two points
x, y A, there exist infinitely many quasi-billiard trajectories from x to y.

Note that the Birkhoff maximization construction when applied to polygons can
never produce the usual billiard trajectories since the maximum will always be achieved
at vertices. In other words, both theorems can be restated as existence results of vertex sequences which satisfy angle conditions as in ().
9.7. Exercises.
Exercise 9.1. [1-] a) Let Q R2 be a closed polygon and let e be the longest edge. Prove
that there exists a vertex v Q which projects onto e.
b) [1-] Deduce part a) from the proof of Theorem 9.11.
/ F,
c) [1] Find a convex polytope P R3 such that for every face F and vertex v of P , v
vertex v projects outside of F .
Exercise 9.2. a) [1-] Prove that every smooth convex curve C R2 has an inscribed
triangle = (xyz), such that a line tangent to C at each of these points is parallel to the
opposite edge of . For example, a line tangent to C at x must be parallel to (yz).
b) [1-] Show that every convex polygon Q R2 has three vertices which form a triangle ,
such that the line through either vertex of parallel to the opposite edge is supporting Q.
c) [1] Generalize a) and b) to higher dimensions.
Exercise 9.3. a) [1-] Prove Theorem 9.13.
b) [1] Give a direct proof of Theorem 9.16.
c) [1] Give a direct proof of Theorem 9.17.
d) [1] Deduce Theorem 9.12 from Theorem 9.16 by a limit argument.
d) [1] Deduce Theorem 9.16 from Theorem 9.12 by a limit argument.
Exercise 9.4. Prove or disprove:
a) [1-] Every smooth simple closed curve in R3 has a plane which touches it at three or
more points.
b) [1-] Every simple space polygon Q R3 has a plane L which contains three or more
vertices of C, such that all edges adjacent to these vertices either lie on L or on the same
side of L.
Exercise 9.5. Let X R2 be a convex set. We say that a point z X is central if every
chord through z is divided by z inside X with a ratio at most 2 : 1.
a) [1-] Prove that the center of mass cm(X) is central.
b) [1] Let be a triangle inscribed into X with maximal area. Prove that cm() is central.
Exercise 9.6. [2-] Let Q = A be a convex polygon, A R2 , and let z A be an interior
point. A triple of quasi-normals from z onto Q is called a tripod at z if they form an angle
of 2/3 with each other (see Figure 9.11). Prove that every convex polygon has a tripod at
some interior point.
Exercise 9.7. a) [1] Let Q = A, A R2 , be a convex polygon. Denote by x = cm(Q)
the center of mass of the curve with the uniform weights. Prove that there exist at least
four quasi-normals from x onto Q.

92

Figure 9.11. Tripods in an equilateral triangle and a square.

b) [1] Let V be the set of vertices of Q with the weights proportional to exterior angles
( x). Denote by y = cm(V ) the center of mass of the vertices with the weights as
above. Prove that there exist at least four quasi-normals from y onto Q.
c) [1+] Let Br R2 a disk of radius r and let Xr = X + Br , for all r > 0. Denote by
xr = cm(Xr ), z = cm(Xr ) the centers of mass as above. Prove that
lim xr = lim zr = y .

d) [1] Prove that there exist at least four normals from xr onto Q, for all r > 0. Same for
all zr .
Exercise 9.8. Let C R2 be a smooth convex curve.
a) [1+] Denote by x and y midpoints of the largest cut and the min max cut (as in the
proof of Theorem 9.10). Prove that there exist at least four normals from x onto C. Same
for y.
b) [2-] Prove that the circle is the only convex curve with a unique point with four normals.
Exercise 9.9. Let C R2 be a smooth convex curve. Denote by (C) the set of interior
points with exactly four normals to C.
a) [1] Compute (C) for an ellipse C.
b) [1+] Prove or disprove: (C) is always connected.
Exercise 9.10. (Stability of simplices) a) [2-] Prove that for every tetrahedron R3
the center of mass cm() projects onto at least two faces of P .
b) [1-] Construct a tetrahedron which when placed on a plane may roll twice.
c) [1] Construct a tetrahedron and an interior point O such that O projects only on
one of the faces.
d) [2-] Prove that for d large enough, there exists a d-simplex Rd which can stand
only on one facet, i.e., such that cm() projects onto only one face of .
e) [2] Prove that the dimension d in the previous part can be as small as 11, and must be
at least 8.
Exercise 9.11. [1-] Find a parallelepiped where all edges have equal lengths, such that it
can stand only on two opposite faces.
Exercise 9.12. (Stability of polygons and polytopes) a) [2-] Prove that every interior
point O in a regular n-gon Q projects onto at least two faces of Q.
b) [1+] Construct explicitly a convex polytope P R3 which can stand only on one face,
i.e., such that cm(P ) projects onto only one face of P (cf. Example 9.9).
c) [2+] Construct a convex polytope P R3 such that the distance from cm(P ) to P has
only one local minimum and one local maximum.
d) [1] Show that in the plane it is impossible to construct a polygon as in c).

93

Exercise 9.13. a) [1] Let R2 be an acute triangle in the plane and let T be a
triangle inscribed into T with vertices at the feet of the altitudes. Prove that T is a periodic
billiard trajectory. Prove that of all triangles inscribed into , triangle T has the smallest
perimeter.
b) [2-] Prove that there exists a quadrilateral periodic billiard trajectory in every tetrahedron with acute dihedral angles.
Exercise 9.14. [2] Let P R2 be a simple polygon with rational angles. Prove that P
has innitely many periodic billiard trajectories.
Exercise 9.15. [1] Prove that regular n-gon in the plane has exponentially many periodic
simple quasi-billiard trajectories.
Exercise 9.16. (Dense billiard trajectories) A billiard trajectory R in a convex polygon
Q R2 is called dense if for every x Q and > 0 there is a point y R such that
|xy| < .
a) [1] Prove that every billiard trajectory in a square is either dense or periodic.
b) [1+] Find a convex polygon Q R2 and a billiard trajectory that is neither dense nor
periodic.
c) [2+] Prove or disprove: in a regular pentagon every billiard trajectory is either dense or
periodic.
Exercise 9.17. a) [1+] Prove that in every smooth convex curve C there exists at least
two 3-sided periodic billiard trajectories inscribed into C.
b) [2-] Generalize this to two periodic billiard trajectories with larger number of sides.
c) [2-] Prove or disprove: in every convex polygon Q R2 there exists at least two 3-sided
periodic quasi-billiard trajectories.
Exercise 9.18. [2+] Let S R3 be a smooth convex surface. Prove that there exists a
point inside of S with at least four normals onto S.
Exercise 9.19 (Double normals in R3 ). [2+] Let S R3 be a smooth convex surface.
Prove that it has at least three double normals.
Exercise 9.20. a) [1-] Let C R2 be a convex cone. Prove that every billiard trajectory
in C has a nite number of reections.
b) [1] Generalize this to circular cones in R3 .
c) [1+] Generalize this to convex polyhedral cones in R3 .
Exercise 9.21. (Dual billiards) Let Q R2 be a convex set and let x1 be a point outside
of Q. Let x2 be a point such that (x1 x2 ) is a line supporting Q at a point which is a
midpoint of x1 and x2 . Note that there are two such x2 ; choose the one where Q is to the
right of the line. Repeat the construction to obtain x3 , x4 , etc. A sequence [x1 x2 x3 . . .] is
called a dual billiard trajectory (see Figure 9.12).
a) [1] Prove that when Q is a square, then every dual billiard trajectory is periodic.
b) [1] Same when Q is a triangle. Find all period lengths.
c) [1] Suppose Q is smooth and strictly convex. Prove that circumscribed n-gons of minimal area are periodic dual billiard trajectories. Conclude that there exist innitely many
periodic dual billiard trajectories.
d) [1] Extend part c) to self-intersecting n-gons. Show that for prime n 3 there exist at
least n1
2 n-gonal dual billiard trajectories.

94

Figure 9.12. Periodic dual billiard trajectories.

e) [1] For a smooth Q, let T be an inscribed triangle of the maximal area. Show that the
triangle with edges tangent to Q at vertices of T is a triangular dual billiard trajectory
(cf. Exercise 9.5).
f ) [1-] Extend e) to generic convex polygons Q (with no two parallel sides or diagonals).
g) [1+] Dene the generalized area A(Y ) of a polygon Y = [y1 . . . yn ] by the following
formula:
X

A(Y ) = kw k, where w =
(1)j+i+1 y i y j , y i = Oyi .
1i<jn

For a smooth Q and odd n 3, let Y be the inscribed n-gon of maximal generalized area.
Show that the n-gon with edges tangent to Q at yi is an n-gonal dual billiard trajectory.
Extend this to generic polygons.
h) [2] Prove that for every regular polygon Q, all dual billiard trajectories are bounded (do
not go to innity).
i) [1+] Consider a quadrilateral Q with vertices (0, 1), (1, 0) and (, 0). Prove that for
Q, all dual billiard trajectories are periodic.
j) [2] In notation of i), nd a
/ Q and an unbounded dual billiard trajectory.
9.8. Final remarks. Theorem 9.5 seems to be due to Arnold who proved it via reduction
to the four vertex theorem (Theorem 21.1). See a sketch in [Arn2, 3] and a complete proof
in [VD1]. In fact, Lemma 9.6 was used in [Tab1] to prove the four vertex theorem.
For more on double normals see [CFG, A3, 4]. The number of special cuts in higher
dimensions remains an open problem. Our proof of Theorem 9.12 follows the presentation
in [KleW, 1.4]. The existence of monostatic polytopes in Example 9.9 was discovered
in [CGG]. Monostatic simplices in higher dimensions we studied in [Daw2].
The theory of billiards plays an important role in ergodic theory. We refer to [Tab6] for
a historical overview and an accessible introduction, and to [Tab2] for a good survey (see
also [Schw]). We return to billiard trajectories in the next section (see also some exercises
above).
Finally, let us mention that it is often dicult to recognize and compare the direct (variational) arguments vs. the indirect (existence) arguments. While the indirect arguments
often have the advantage of being concise and amenable to generalizations, the direct arguments are often more elegant and can be implemented. We spend a great deal of the book

95

discussing both methods. For now, let us leave you with a few words of wisdom by Sun
Tzu [Sun]:
iV.5. In all ghting, the direct method may be used for joining battle, but
indirect methods will be needed in order to secure victory.
V.10. In battle, there are not more than two methods of attack the direct
and the indirect; yet these two in combination give rise to an endless series of
maneuvers.
V.11. The direct and the indirect lead on to each other in turn. It is like moving in a circle you never come to an end. Who can exhaust the possibilities
of their combination?
These quotes are especially relevant to the proofs in Sections 3537.

96

10. Geodesics and quasi-geodesics

In this section we study closed geodesics and quasi-geodesics on convex polytopes.
In the first part, we derive a baby version of the general Cohn-Vossens theorem which
states that simple closed geodesics enclose half the curvature (see Subsection 25.6). In
the second part, consider closed quasi-geodesics on convex polytopes. We use results
on periodic billiard trajectories in the previous section to obtain a simple special case
of the LyusternikShnirelman theorem.
10.1. Wise men never go to the top of the hill. 25 Let S R3 be a closed
2-dimensional polyhedral surface, e.g., the surface of a convex polytope. For two
distinct points x, y S consider all shortest paths between them. Of course, there
can be many such paths, e.g., when S is a regular n-prism and x, y are centers of
the n-gon faces. Note, however, that determining whether a path between x and y is
shortest is a non-trivial global problem. Here is the corresponding local notion.
A piecewise linear path S is called a geodesic if for every point z a sufficiently small neighborhood of around z is a shortest path. Clearly, every shortest
path is a geodesic, but not vice versa (see Figure 10.1).

a
1
a

1
b

2
b

Figure 10.1. A geodesic that is not a shortest path and an unfolding

along a geodesic.
Proposition 10.1. No geodesic on a convex 2-dimensional polyhedral surface S
R3 has a vertex in its relative interior. In addition, whenever a geodesic crosses an
edge e in S, it has equal angles with e in the faces adjacent to e.
Proof. For the first part, suppose a geodesic between points x and y passes though
vertex v. Choose points a, b close enough to v, so that the [ab] portion of is
a shortest path which consists of two straight intervals 1 and 2 as in Figure 10.1
Since the surface S is convex, the angle avb < on at least one of the sides of .
Unfold S along the edges in this angle and note that |ab| < |av|+|vb| in the unfolding,
a contradiction.
For the second part, similarly unfold along edge e and observe that must be
straight around the intersection point z = e. The details are straightforward. 
In other words, if you are on one side of the hill and you need to go to another
side, it is always faster to go around than go all the way up to the top.
25This

is a play on a Russian proverb Umny

i v goru ne po
id
et, umny
i goru obo
id
et.

97

10.2. Walking around and coming back. Suppose now S is a smooth convex
surface in R3 . Does it have closed geodesics? Does it have a closed geodesics without
self-intersections? (such geodesics are called simple).
Theorem 10.2 (Poincare). Every smooth convex body has at least one simple closed
geodesic.
Poincare proved this result in 1905 and conjectured that every convex body has at
least three simple closed geodesics (see Figure 10.2). Birkhoff later proved existence
of two such geodesics, but the conjecture was resolved by Lyusternik and Shnirelman
in 1929.
Theorem 10.3 (LyusternikShnirelman). Every smooth convex body has at least
three simple closed geodesics.
This result is highly technical and beyond the scope of this book.

Figure 10.2. Three geodesics on an ellipsoid.

Remark 10.4. There are two ways to approach the Poincare theorem. First, one can take
the shortest of the curves which divide the surface area into two equal halves. This turns
out to be a correct construction of a closed geodesic, but does not lead to generalizations.
Alternatively, imagine a solid convex body with surface S and no friction. Take a closed
rubber band and place it on a solid. The idea is that the rubber band will slide out in one
end or another unless it is in an equilibrium, in which case it is a geodesic. This basic idea
lies at the heart of the LyusternikShnirelmans approach.

10.3. Back to billiard trajectories. Let us give a connection between the double
normals theorem (Theorem 9.10) and the Lyusternik-Shnirelman theorem. Consider
a nearly flat convex body B with smooth boundary and an orthogonal projection C
(think of B as a cake obtained by fattening C). The geodesics now look like
billiard trajectories, as they alternate between the top and the bottom face of B
(see Figure 10.3). In this sense, special cuts defined in the previous section (see
Subsection 9.4) correspond to closed geodesics on B that alternate exactly once. The
double normals theorem now implies that B has at least two such closed geodesics.
One can ask what happened to the third closed geodesic as in the LyusternikShnirelman theorem. Well, it simply goes around the boundary of B.
Suppose now that S = P is the surface of a convex polytope P R3 . One can
ask what is known about closed geodesics on S. Given the poor state of art for the
periodic billiard trajectories on convex polygons, the following two conjectures are
hardly surprising and unlikely to be easy.

98

Figure 10.3. A geodesic on a convex body shaped as a pool table.

Conjecture 10.5. The surface of every convex polytope has at least one closed
geodesics.
Conjecture 10.6. Let S be the surface of a convex polytope P R3 . For every two
points x, y S, there exist infinitely many geodesics from x to y.
10.4. Simple closed geodesics on convex polyhedra. As before, let S = P be
the surface of a convex polytope P R3 . One can ask if there is always at least
one simple closed geodesic on S? The answer turns out to be negative. In fact, in
a certain precise sense, almost no tetrahedra has a simple closed geodesics. On the
other hand, a number of natural polytopes have many closed geodesics, including
simple closed geodesics (see e.g., Figure 10.4 and Exercise 10.3).

Figure 10.4. Closed geodesics on a cube and a regular tetrahedron.

Claim 10.7. The surface of a random tetrahedron does not have simple closed geodesics.
Of course, there are many different ways to define a random tetrahedron. For
example, one can choose its vertices at random from a unit cube, a unit sphere,
or some other natural distribution, as in Sylvesters problem (see Exercise 42.3).
Alternatively, one can choose a random collection of face angles of a tetrahedron,
conditioned that in each triangle they add up to and the angle sums in vertices
are < 2 (see Exercise 37.1). Similarly, one can choose a tetrahedron uniformly at
random from the compact set of all possible collections of edge lengths of tetrahedra
with a unit total edge length (cf. Exercise 31.6). The claim holds in each case, and
easily follows from the following simple observation.
Lemma 10.8. Let be a tetrahedron, and let i , 1 i 4, denote the sums of the
face angles at the i-th vertex. If i + j 6= 2 for all 1 i < j 4, then does not
have simple closed geodesics.

99

Proof. First, observe that the total sum of all face angles in a tetrahedron is equal to
1 + 2 + 3 + 4 = 4. Clearly, a simple closed geodesic on must separate either
one or two vertices. Suppose = (abc) goes around one vertex v1 (see Figure 10.5).
By Proposition 10.1, the angles at a, b and c are equal to on each side. Therefore
the sum of the face angles in a tetrahedron (v1 abc) is equal to 1 +4, a contradiction.
v1

v1
c

b
b

v2

Figure 10.5. Potential closed geodesics on a tetrahedron.

Similarly, suppose = [abcd] goes around two vertices v1 , v2 as in Figure 10.5.
Denote by the total sum of angles in two 4-gons (v1 abv2 ), (v2 cdv1 ) and two triangles
(dav1 ), (bcv2 ), i.e., all polygons on one side of . Clearly, = 6. On the other hand,
since is a geodesic, the sum of angles at points a, b, c, and d is equal to . Therefore,
= 1 + 2 + 4, and we obtain 1 + 2 = 2, a contradiction.

10.5. Quasi-geodesics. Let S = P be the surface of a convex polytope in R3 . For
a point x S, denote by (x) the sum of face angles at x, if x is a vertex of P , and
let (x) = 2 otherwise.
For a piecewise linear path S, at each vertex x of , consider the intrinsic angles
+ (x) and (x) in S, one for each side of around x. By definition, + (x)+ (x) =
(x). Path is called a quasi-geodesic if + (x), (x) for every interior x .
In other words, when crosses an edge in X, it behaves as a geodesics. On the other
hand, path is now allowed to pass the vertices as long as the angles on both sides
are . Examples of closed quasi-geodesics are shown in Figure 10.6.

Figure 10.6. Seven closed quasi-geodesics on a cube.

The notion of a quasi-geodesic is an extension of the quasi-billiard trajectories from
doubly covered polygons to general convex polyhedra. Observe that every even-sided
periodic quasi-billiard trajectory in a convex polygon Q corresponds to a closed geodesic on doubly covered Q. Similarly, an even-sided periodic quasi-billiard trajectory
can be doubled to give a closed geodesic. Theorem 9.16 then gives:

100

Proposition 10.9. Every doubly covered convex polygon has infinitely many closed
quasi-geodesics.
Similarly, Theorem 9.11 and the discussion in Subsection 10.3, implies the following
variation on the Lyusternik-Shnirelman theorem (Theorem 10.3).
Proposition 10.10. Every doubly covered convex polygon has at least three simple
closed quasi-geodesics.
The most general result in this direction in the following theorem, also beyond our
scope.
Theorem 10.11 (Pogorelov). Every convex polytope has at least three simple closed
quasi-geodesics.

10.6. Exercises.
Exercise 10.1. (Explicit constructions) a) [1-] Give an explicit construction of a polytope with no simple closed geodesics.
b) [1] Give an explicit construction of a polytope with no simple closed geodesics which has
a (non-simple) closed geodesic.
Exercise 10.2. [1-] Find a polytope P R3 with innitely many non-periodic closed
geodesics of distinct lengths. Further, nd P such that the number of self-intersections of
closed non-periodic geodesics on P is unbounded.
Exercise 10.3. a) [1+] Prove that the seven shortest geodesics on the unit cube have
squared lengths 16, 18, 20, 90, 148, 208 and 212.
b) [1+] Prove that all simple closed geodesics on the regular octahedron with edge lengths 1
have squared lengths 8 and 9.
c) [1+] Prove that all simple closed geodesics on the regular icosahedron with edge lengths 1
have squared lengths 25, 27 and 28.
Exercise 10.4. (Geodesics on equihedral tetrahedra) a) [1-] Prove that a tetrahedron has
three pairwise intersecting simple closed geodesics if an only if it is equihedral, i.e., has
congruent faces (cf. Exercise 25.12).
b) [2-] Two geodesics are called isotopic if they intersect edges of the polyhedron in the
same order. Prove that a tetrahedron has innitely many non-isotopic geodesics if and only
if it is equihedral.
c) [1] Show that one can perturb the equihedral tetrahedron so that the resulting tetrahedron has a nite number of non-isotopic geodesics, and this number can be arbitrary
large.
d) [2-] Prove that a tetrahedron has innite simple (non-self-intersecting) geodesics if and
only if it is equihedral.
e) [2] Prove that a convex polytope P R3 has an innite simple geodesic if and only if P
is the equihedral tetrahedron.

101

10.7. Final remarks. For the details, history and references on closed geodesics and the
LyusternikShnirelman theory (including a simple proof of their main result) see [Kli]. We
refer to [Alb] for a readable account of the ideas. Unfortunately, both the Lyusternik
Shnirelman theorem and the proof in [Kli] are non-combinatorial in nature (and in some
sense not even geometric) since they employ parameterized curves, and use the parameter in
an essential way. Interestingly, in higher dimensions the
number of simple closed geodesics
on convex surfaces in Rd has yet to be shown to match d2 , the smallest number of geodesics
the ellipsoids can have.
For Claim 10.7 and the related results see e.g., [Gal]. For more on quasi-geodesics see [A1]
and [AZa]. Pogorelovs Theorem 10.11 is proved in [Pog1] in the generality of all convex
surfaces (one can extend quasi-geodesics to this case). He uses approximation of general convex surfaces with smooth surfaces, which reduces the problem to the LyusternikShnirelman
theorem. We return to the study of closed geodesics on convex polytopes in Sections 25
and 40.

102

11. The Steinitz theorem and its extensions

The Steinitz theorem is one of the fundamental results in polyhedral combinatorics.
It can be viewed as the converse of the result that all graphs of polytopes in R3 are
planar and 3-connected. We present it in the first part of the book not so much
because the proof is simple, but because it uses the variational principle approach, and
the result plays a central role in the study of 3-dimensional convex polytopes. Among
other things, it proves that all 3-dimensional polytopes are rational (cf. Section 12),
and gives a basis for the algebraic approach, easily implying that almost all polyhedra
in R3 are rigid (see Section 31).
11.1. (Almost) every graph is a graph of a polytope. Recall Balinskis theorem,
which in the case of 3-dimensional convex polytopes states their graphs must be 3connected. As it turns out, this and the planarity are the only conditions on graphs
of polytopes.
Theorem 11.1 (Steinitz). Every 3-connected planar graph is a graph of a convex
polytope in R3 .
This is an important result, and like many such results it has several proofs and
interesting generalizations. We will mention three proof ideas, all somewhat relevant
to the rest of the course. Unfortunately we will not include either of the proofs for
two very different reasons: the first proof (graph theoretic) and the second proof (via
circle packing) are so well presented in Zieglers lectures that we see no room for
an improvement. On the other hand, the third proof (via Tuttes theorem) is quite
technical as its graph-theoretic part is yet to be sufficiently simplified for a leisurely
style we favor in this book.
11.2. Graph-theoretic proof. The idea of this proof goes to the original proof by
Steinitz, and is based on the following (difficult) lemma in graph theory:
Lemma 11.2. Every 3-connected planar graph can be reduced to a complete graph K4
by a sequence of simple Y transformations.
The Y transformations are shown in Figure 11.1 and can be used in either direction. The simple Y transformations are the Y transformations followed by
removal of parallel and sequential edges (the P, S reductions shown in Figure 11.1).
P
Y
S

Figure 11.1. The Y transformations , and P, S reductions.

103

An example of transforming graph C3 (graph of 3-dimensional cube) into a graph K4

(graph of a tetrahedron) is given in Figure 11.2. Here we indicate all Y transformations and P reductions (S reductions are obvious and omitted).

C3
K4

Figure 11.2. A sequence of simple Y transformations from C3 to K4 .

We will not prove Lemma 11.2, but refer to [Zie1, 4.3] for a simple and beautiful
proof. Below we sketch the rest of the proof of the Steinitz theorem.
Given a 3-connected planar graph , construct a sequence of simple Y reductions
as in the lemma:
G 1 G 2 G n = K4 .
We will work backwards, given a polytope with graph Gi we will construct a polytope
with graph Gi1 . Clearly, every tetrahedron has graph Gn = K4 , our starting point.
Now observe that each Y transformation corresponds to simply cutting a cone
with a hyperplane (the Y direction) or extending the three facets adjacent
to a triangular face to meet at a vertex (the Y direction). One can think of
Figure 11.1 as depicting a view from the top of this. The details are straightforward
and left to the reader. 
One important consequence form this proof of the Steinitz theorem is the ability
to make all vertex coordinates to be rational. This property will prove useful in the
next section.
Corollary 11.3. Every 3-connected planar graph is a graph of a convex polytope in R3
with rational coordinates.
For the proof simply observe that in the proof above we could have started with
a rational tetrahedron and continued to cut it only with rational planes, thereby
creating only rational vertices. Clearly, when we are doing extensions of rational
planes we are also creating only rational vertices, which completes the argument. In
fact, one can even blow up the resulting polytope and make all vertex coordinates
integral.

104

11.3. Circle packings are totally cool! Consider the following circle packing theorem often attributed to Koebe, Andreev and Thurston:
Theorem 11.4 (Circle packing theorem). Every planar graph G can be represented
in the plane by a collection of disjoint circles corresponding to vertices of G, with
edges of G corresponding to touching circles.
On the surface, this theorem has nothing to do with the Steinitz theorem. Here
is the connection. Start by making a stereographic projection of the circles onto a
sphere. This gives a corresponding collection of circles on a sphere. Now extend
planes through each of the circles. This gives a polytope whose edges correspond to
pairs of touching circles and in fact they touch a sphere at those points. In fact, this
polytope has a graph G , dual to the graph G in the theorem. This leads to the
following extension of the Steinitz theorem:
Theorem 11.5 (Schramm). Every 3-connected planar graph is a graph of a convex
polytope in R3 whose edges are all tangent to a unit sphere. Moreover, if we require
in addition that the origin is the barycenter of the contact points, such polytope is
unique up to rotations and reflections.
While the original proofs of both theorems are rather difficult, a recent proof by
Bobenko and Springborn of Schramms theorem via the variational principle is quite
elegant and relatively simple. Again, we do not include the proof and refer to [Zie2,
1.3] for a beautiful exposition.
11.4. Graph drawing is more than a game! When we talk about planar graphs,
one needs a little care as the edges of the graph can be drawn on a plane with lines,
polygonal arcs, curves, etc., and one has to give an argument that the notion of
planarity is in fact independent of the presentation. The following result is the best
one can hope for:
Theorem 11.6 (Fary). Every planar graph can be drawn on a plane with straight
non-intersecting edges.
It is not hard to prove the theorem directly, by starting with an outside edge and
the adjacent face, making it a convex polygon and adding points one face at a time
in such a way that the intermediate boundary remains convex (see Figure 11.3).
Slightly more carefully: take a dual graph, fix a vertex v on the boundary and orient
all edges towards v, creating a partial order on vertices. Now extend this order to a
linear order in any way and use this order to add the faces.
Note that this method of drawing graphs produces non-convex faces. The following
result may come as an initial surprise:
Theorem 11.7 (Tuttes spring theorem). Every 3-connected planar graph can be
drawn on a plane such that all interior faces are realized by non-overlapping convex
polygons.
In fact, the result follows immediately from the Steinitz theorem. Simply take a
face F of the polytope P corresponding to the outside face, place it horizontally, and

105

Figure 11.3. A simple way of proving the Fary theorem.

set a point O just below the middle of the face. Now draw inside the face the view of
the edges of P as they are seen through F (see Figure 11.4).

P
F

v
F

Figure 11.4. Fary theorem from Steinitz theorem: polytope P as

viewed from v though face F .
In fact, one can go the other way as well: first construct nice graph drawings as
in Tuttes theorem, and then lift it up to a polytope.
11.5. Using stresses to draw graphs. Imagine we physically build a graph with
edges made out of springs. Take the outside face and stretch it far enough, fixing all
the points in convex position. The remaining (interior) vertices will be forced to find
some kind of equilibrium, which we show to be unique and satisfy some additional
properties. In fact, this construction gives the drawing as in Tuttes theorem. Now
all we need is to formalize and prove everything.
Let G be a 3-connected graph with a set of vertices V and a set of edges E. A
realization of G is a map f : V R2 . We say that vertex v V is in an equilibrium
if it is in a barycenter of its neighbors:


1
f (v) =
f (w1 ) + . . . + f (wm ) ,
m
where N(v) = {w1 , . . . , wm } is the set of neighbors of v.

Lemma 11.8. Let G be a connected graph with a set of vertices V , and let f : A R2
be fixed realization in the plane of nonempty subset A V of vertices. Then there
exists a unique realization of the whole graph such that all vertices in (V r A) are in
equilibrium.
Note that the lemma does not require the graph to 3-connected or planar. In
fact, Figure 11.5 shows a (non-planar) Petersens graph with all 5 interior vertices

106

in equilibrium. Note that the resulting realization can he highly degenerate. For
example, when A consists of one or two vertices, the whole graph is realized in that
one point or on an interval between these two points, respectively.

Figure 11.5. Petersen graph with five interior vertices in equilibrium.

The proof is straightforward and is suggested by the natural notion of the energy
function coming from physics. Here is how we adapt it to our situation.
Proof. Let A = {a1 , . . . , ak } V be the fixed set of vertices. Without loss of generality we can assume that f (a1 ) = O is the origin. Let V = (V r A) denote the set
of the remaining vertices, and let E denote the set of edges with at least one vertex
in V . Define

1 X 
1 X
kf (v) f (w)k2 =
(xv xw )2 + (yv yw )2 ,
E(f ) :=
2
2

(v,w)E

(v,w)E

where f (v) = (xv , yw ) denotes the coordinates of realization points. Note that E =
E(f ) is a convex quadratic everywhere nonnegative function. Let us show that in
fact E is strictly convex. Indeed, suppose one of the points, say f (w), w V , is far
from the origin: kf (w)Ok > C. Connect w to a1 by a path with at most n = |V |1
edges. Clearly, at least one of the edge lengths is longer than C/n, so E C/n.
Therefore, for sufficiently large C > 0 we have E(f ) > E(f0 ) for any fixed f0 and
kf f0 k > C.
We conclude that E is strictly convex with a positive definite Hessian matrix. Therefore, E has a unique minimum. Writing the minimality condition for a critical realization f : V R2 , we obtain:
X
X
E
E
=
(xv xw ) = 0,
=
(yv yw ) = 0.
xv
yv
wN (v)

wN (v)

Adding these two equations, for every v V , we obtain:

X
|N(v)| f (v)
f (w) = O,
wN (v)

which is exactly the equilibrium condition.

107

Note that the variational principle approach quickly reduced the problem to a graph
theoretic question. In fact, we do not even need the uniqueness part in the lemma
just the existence of the equilibrium configuration. Formally, the rest of the Tutte
theorem follows from the following (difficult) lemma.
Lemma 11.9. Let G = (V, E) be a 3-connected planar graph and let f : V R2 be
its realization. Suppose the outside face F forms a convex polygon and the remaining
points in the realization are in equilibrium. Then all faces of G are disjoint convex
polygons.
The lemma is proved by a delicate and technical graph theoretic argument which
we skip. Let us note, however, that Lemma 11.8 holds for all graphs, e.g., for a
complete graph Kn . On the other hand, the result of Lemma 11.9 obviously cannot
hold for non-planar graphs. Thus, planarity of G must be critical in the proof.
11.6. Exercises.
Exercise 11.1. Let P R3 be a simple convex polytope whose faces are quadrilaterals.
a) [1-] Prove that it has eight vertices.
b) [1] Prove that P is combinatorially equivalent to a cube.
c) [1] Prove that if seven vertices of P lie on a sphere, then so does the eighth vertex.
Exercise 11.2. [1] Let P R3 be a simple convex polytope such that every face can be
inscribed into a circle. Prove that P can be inscribed into a sphere.
Exercise 11.3. [1-] Let C R3 be a convex cone with four faces. Prove that C has an
inscribed sphere if and only if the opposite face angles have equal sums.
Exercise 11.4. a) [1+] Let P R3 be a convex polytope whose faces are colored in black
and white such that no two black faces are adjacent. Suppose the total area of black faces
is larger than the total area of white faces. Prove that P does not have an inscribed sphere.
b) [1-] Find an example of such polytope P .
c) [1-] Show that condition in part a) is necessary but not sucient, even if there is a
unique proper coloring of the faces.
Exercise 11.5. a) [1+] As above, let P R3 be a convex polytope whose faces are colored
in black and white such that no two black faces are adjacent. Suppose the number of white
faces is smaller than the number of black faces. Prove that P does not have an inscribed
sphere.
b) [1-] Let Q be a parallelepiped and let P be obtained from Q by cutting o the vertices.
Use part a) to show that P does not have an inscribed sphere.
Exercise 11.6. a) [2] Let P be a convex polytope with graph , and let C be a simple
cycle in G. Prove that there exists a combinatorially equivalent polytope P , such that the
edges of the boundary of a projection of P correspond to C.
b) [1] Prove that the shape of the polygon in the projection cannot be prescribed in advance.
Exercise 11.7. a) [1] Prove that for all k 3, a plane can be partitioned into k-gons
(edge-to-edge).
b) [2] For every simplicial convex polytope P R3 , prove that the space R3 can be partitioned into polytopes combinatorially equivalent to P (face-to-face).

108

c) [2] For every convex polytope P R3 , prove that the space R3 can be dissected into
polytopes combinatorially equivalent to P (not necessarily face-to-face).
Exercise 11.8. (Innite F
ary theorem) [2-] Let G be an innite planar graph of bounded
degree. Prove that G can be drawn in the plane with straight edges. Check that this
generalizes part a) of the previous exercise.
11.7. Final remarks. Steinitz theorem was discovered by Ernst Steinitz in the late 1920s
and his manuscript was published posthumously in 1934 by Hans Rademacher. For a
traditional graph theoretic proof of the Steinitz theorem see e.g., [Gr
u4, 13.1], and for the
history of the Steinitz theorem see [Gr
u5]. An attractive presentation of the graph theoretic
proof is given in [Zie1, 4.3]. For the BobenkoSpringborn proof see [Zie2, 1] and [Spr]
(see also [Grub, 34]). An advanced generalization was obtained by Schramm, who showed
that the unit sphere plays no special role in Theorem 11.5, but in fact one can realize a
polytope with edges tangent to any smooth convex body [Schra]. For a complete proof of
Tuttes spring theorem, the proof of the Steinitz theorem, references and details see [Ric]
(see also [Tho2]).
The reason Theorem 11.4 is attributed to the three authors is because Koebe claimed it
in 1936, but proved it only for triangulations. Almost half a century later, in 1985, Thurston
gave a talk where he presented the theorem. He also noted in his book (which remained
unpublished for years) that the result follows easily from Andreevs results of the 1970.
Only then, with papers by Colin de Verdi`ere (1988), Schramm (1991), and others, the
result was brought into prominence. We refer to [Ber2] for an elementary introduction,
to [PacA, Ch. 8] for a nicely written standard xed-point type proof, and to [Spr, Zie2] for
the references.
It is interesting to note a fundamental dierence between the two variational proofs (of
Theorems 11.5 and 11.7). While in the proof of Tuttes theorem proving existence of an
equilibrium is easy, it is showing that the equilibrium is what we want that is hard. The
opposite is true in the Bobenko-Springborn proof (as well as in Schramms proof) it is
proving the existence of an equilibrium that is hard; showing that the equilibrium produces
a desired conguration is quite easy.

109

12. Universality of point and line configurations

This is the first of two sections on universality. In essence, it shows that the
face combinatorics studied in Section 8 and the geometry of polyhedra are largely
incompatible. To put it another way, we show that by making some restrictions on
the face lattice one can ensure that the polytopes are irrational, as complicated as
needed, or even non-existent. This is an extremely negative result showing that there
is no hope for the Steinitz type theorem in higher dimensions (see Section 11).
12.1. Universality is not for everyone. In the following two sections we present
several results with one unifying goal: to show that various constructions and configurations in discrete geometry can encode an extremely wide (universal!) range
of objects. While this may seem an empowering proposition, from the postmodern
point of view this is very disappointing, since we are basically saying that nothing
interesting can be proved about all these general geometric objects. Indeed, what can
possibly be interesting about everything?
12.2. Ruler and compass constructions are so pass
e. Recall that it is impossible
to construct a regular 7-gon using a compass and a ruler. This classical result of
Gauss is now considered iconic, but two hundred years ago the result took awhile to
understand and appreciate. What was more impressive to mathematicians of the day
was the construction of a regular 17-gon, unknown until Gauss. This constructive
part of his work is considered elementary now, but his approach, in fact, can be
viewed as the earliest universality type result.
Consider an increasing sequence of fields
K0 K1 K2 . . . Kn ,

where K0 = Q, Ki = Ki1 [ ai ], and ai Ki1 , for all 1 i n. We call a number

z C geometric if there exist a chain of fields as above, such that z Kn , for
some n N.
Theorem 12.1 (Gauss). Every geometric number can be constructed by the ruler and
the compass.
We need to elaborate on what we mean by the ruler and compass construction (in
the plane). We identify the plane with the field C, and assume that in the beginning
are given the set of two points A0 = {0, 1}. Our algorithm increases the set of points
by making the ruler and compass drawing, and adding new intersection points at each
step.
Formally, suppose at the n-th step we have the set An C of points in the plane.
At the (n + 1)-th step we can do two of the following operations:
draw a line (ai , aj ) through any two points ai , aj An , or
draw a circle with a center at ak and radius |ai aj |, for some ai , aj , ak An .
Then we can add the set of intersection point(s) B = {b1 , b2 } in the case of two circles,
or a line and a circle, and B = {b} in case two lines, to the set of points:
let An+1 := An B.

110

We say that the above algorithm constructs all points in An . The Gauss theorem
says that all geometric numbers (viewed as points in the plane) can be constructed
this way.
Now, the reason why the 17-gon can be constructed by ruler and compass lies in

(equivalently, cos 17
) is what we call a geometric number. We will
the fact that sin 17
skip the explanation of this and move to the proof of Theorem 12.1. The proof is,
in fact, completely straightforward and is included here as a motivation for further
results in this section.
Proof of Theorem 12.1. First, let us emulate the field operations. Starting with A0 =
{0, 1} we can obtain all of K0 = Q by using a finite number of additions, multiplications, subtractions and divisions. Let us show how to compute x + y, x, x y
and 1/x for every x, y C. The constructions of x + y, x can be done easily (see
Figure 12.1). To compute 1/x for complex x, first compute the argument x of 1/x
(line on which x lies), and then the norm 1/a, a = |x|. Similarly, for xy first compute
the argument, and then the norm a b, where a = |x| and b = |y|.
x

x+y

O
x
xy
i

x
bi
y

i
a

i
x

ab

Figure 12.1. Computing x + y, x, x y and 1/x, for x, y C.

Now, by induction, given ai we can now construct

any element in Ki = Ki1 [ ai ].
Thus, it remains to show how to construct the roots z given any z C (one of the
two complex roots
will suffice, of course). As before, first construct the argument and
then the norm c, where c = |z|. The latter requires some care. If c > 1, construct
the right triangle
with sides (c + 1) and (c 1), as in Figure 12.2. The length of the

third p
side, 2 c, divided by 2 gives the desired norm. If c < 1, compute the inverse 1/c,

then 1/c by the method described above, and, finally, the inverse c.


111

2 c

c+1

Figure 12.2. Computing

c1

z, where c = |z| and z C.

12.3. Rationality is a popular, but not universal, virtue. Consider a finite set
of points A = {a1 , . . . , an } R2 and lines L = {1 , . . . , m } in the plane. A point
and line arrangement is a pair (A, L). From each point and line arrangement we can
record the set of incidence pairs
C = {(ai , j ) | ai j , 1 i n, 1 j m}
We call C a (plane) configuration and think of pair (A, L) as a realization of C over R.
It is important to emphasize that in a realization no additional incidences can be
created. So, for example, two different points in a configuration cannot be realized
by a single point in the plane.
Now, we can also consider realization of C over other fields, such as Q, C, or finite
fields Fq . Here is a natural question: Is it true that configurations realizable over C
are also realizable over R? What about R versus Fq ? What about different finite
fields? The answers to all these questions turns out to be a NO, as we show below.
Furthermore, in the next subsection we prove a universality result which shows that
the situation is much worse than it seems.
Formally, write a configuration C as a pair (V, E) where V = {v1 , . . . , vn } is a set of
points, and E = {e1 , . . . , em } 2V is a set of lines which can contain any subsets
of at least two points. There are natural combinatorial conditions on E such as, e.g.,
two lines containing the same two points must coincide, but we will not be concerned
with these. Also, we consider all projective realizations rather than affine realizations
as they are easier to work with. Of course, this does not affect realizability as we can
always take a projective linear transformation to move points away from the infinite
line.
For a planar configuration C = (V, E) and a field K consider a map f : V KP2 .
Denote by f (ej ) a line spanned by some of its pairs points: f (ej ) = hf (vi ), f (vr )i
K2 , where vi , vr ej . We say that f is a (projective) realization over K if the following
condition is satisfied:
f (vi ) f (ej ) if and only if vi ej , for all vi V, ej E.
The following two examples illustrate how existence of realizations depends on the
field K.

112

Example 12.2. Consider the Fano conguration given in Figure 12.3. It can be easily
realized over F2 by taking all points on the 2-dimensional projective plane. By the Gallai
Sylvester theorem (Theorem 2.3), for every nite set of point in R2 not all on the same line,
there must be a line containing only two points. This immediately implies that the Fano
congurations cannot be realized over R. In other words, there is no way to straighten
the red line (456) circled in the gure.
2

Figure 12.3. The Fano configuration.

Example 12.3. Consider Perles conguration P as shown in Figure 12.4. Note that it
is realizable over R (when the pentagon in the middle is regular), but cannot be realized
over Q. In other words, we claim that it is impossible to make P rational. To prove this, rst
send the (1234) line to the line at innity by a projective linear transformation. We obtain a
pentagon [56798] whose four diagonals, all except for (57), are parallel to the opposite sides.
Now use another projective linear transformation to send 5 to the origin (0, 0), 6 (0, 1),
8 (1, 0). Since (78)k(56) and (69)k(58), we have 7 (1, a), 9 (b, 1). Also, since
(79)k(68), we have a = b = 1 + t for some t. Finally, since (59)k(67), we get the following
equation for the slopes:
t
1
=
,
1
1
+
t

and t {, 1}, where = 5+1

is the golden ratio.
2
To summarize, we showed that if Perles conguration P had a rational realization, there
would be a rational projective linear map into a nonrational arrangement, a contradiction.
Thus P is not realizable over Q.
2

6
5

1
8
8

Figure 12.4. Three versions of the Perles configuration P.

113

12.4. Points and lines are really all you need. We will show that there exist configurations whose realizations are not nonrational, but in fact can be as complicated
as one wishes. To make this formal, consider all complex realizations f : V CP2
of a configuration C = (V, E)
We consider finite products of ratios of the following type:
()

:=

|vi1 vi2 | |vi5 vi6 |

... ,
|vi3 vi4 | |vi7 vi8 |

where the points in the same ratio must belong to the same line. For a polynomial
P Q[t], we say that configuration C satisfies the law of P if there exists a product ()
as above, such that P () = 0 for all complex realizations f of C.
For example, in the Fano configuration without line (456) (see Figure 12.3) consider
the ratio
|16| |24| |35|
:=

.
|62| |43| |51|
The Ceva theorem states that = 1 in this case, the nicest law of all.
A different example is given by the Perles configuration. Consider the cross-ratio
on a projective line (1234) defined as follows:
:=

|12| |14|

|32| |34|

Recall that cross-ratio is invariant under projective linear transformations. In Example 12.3, we showed that a projective linear transformation maps line (1234) into the
infinite line, with points 1, 2, 3, 4 mapped into t, 0, 1, , respectively. Hence = t,
where t is as in Figure 12.4. Thus, the calculations in the example above imply that
the cross-ratio satisfies 2 1 = 0, the golden law.
Theorem 12.4 (Configuration universality). For every polynomial P Q[t], there
exists a configuration C realizable over C which satisfies the law of P .
Proof. We will draw the pictures over R for simplicity, but the whole construction
can be done over C. Fix a coordinate system by adding points and lines as in
Figure 12.5. Any complex realization of this configuration can be moved into this
one, with lines x = (w1 w2 w3 ) and y = (w1 w4 w6 ) being coordinate axes, an infinite
line (w3 w6 ), and |w1w2 | = |w1w4 | = 1. The main idea of the proof is to construct a
configuration CP starting with this coordinate system which would have points t and
have P (t) to be identically 0 (see Figure 12.5). Just like in the proof of Theorem 12.1,
we will show that it is possible to add, subtract, multiply numbers and divide by an
integer. At each step we add lines and points (including those at infinity if the lines
are parallel) so that in every realization, the points of the configuration model the
arithmetic operations. See Figure 12.6 for a step by step addition and multiplication.
For the addition of a and b make a line parallel to the y-axis (that is going through
point a and an infinite point w6 ) and take the intersection v with the y = 1 line, i.e.,
with the line (w4 w5 ). Now, adding lines (w4 a) and a parallel line through v, i.e., a
line though v and an intersection with the infinite line (w3 w6 ), gives the desired point
(0, a+b) on the x-axis. Similarly, for multiplication, obtain a point (b, 0) on the y-axis

114

w6

w4

w5

w1

w2

P (t) = 0

w3

Figure 12.5. Coordinate system for configurations.

by adding a line through point (b, 0) parallel to x + y = 1 line, i.e. line (w2 w4 ). Then
make a line through point (b, 0) parallel to a line through the points (0, 1) and (a, 0),
obtaining an intersection point (0, ab). Subtraction and division by an integer are
completely analogous and will be skipped.
b
O

a b

b
a b

O 1a b

O 1 a b

a b

a+b

a b

ab

Figure 12.6. Addition and multiplication of a, b.

Now, suppose P (t) = cN tN + . . . + c1 t + c0 , ci Q, is the given polynomial.
Start with a point t and construct all terms of the polynomial, eventually obtaining a
configuration with t, P (t) on the x axis, and one degree of freedom modulo coordinate
system. Now consider a configuration where points P (t) and w1 = O are identified,
and so are some lines by implication. Naturally, such a configuration is realizable
over C since every polynomial P (t) as above has complex roots. On the other hand,
the cross-ratio of the points w1 , w2, w3 and t on the x-axis line will satisfy the law
of P .

Now that we have the result we can take the polynomial P to be as complicated
as required. For example, if P = t2 2, then the resulting configuration of points
must be irrational. Perhaps even more surprising is the case P = t2 + 1, when a
configuration has complex, but no real, realizations.
Before we conclude, let us note that in place of one variable we could have chosen
several variables and several polynomial equations. The same approach goes through,
and at the end one obtains the configuration whose space of realizations satisfies any
system of polynomial equations. A technical checking done by Mnev shows that it

115

is always possible to proceed with this without ever creating any extra equations.
We leave all this aside. The main point of this subsection is to emphasize that in a
certain sense (which can be made completely formal), all algebraic varieties over Q
are encoded by finite configurations.
12.5. Polytopes can also be irrational! Let P Rd be a convex polytope, and
let L be the poset (lattice) of faces of P . We say that P Rd is combinatorially
equivalent to P if they have the same poset L. We say that P is realizable over a
subfield K R if there is a polytope P combinatorially equivalent to P , such that
all coordinates of P lie in K. As it turns out, just like in case of realizations of points
and lines, there exist polytopes which satisfy any prescribed law.
Theorem 12.5 (Perles, Mnev). There exists a polytope P Rd which cannot be
realized over Q. Moreover, there exists a polytope P Rd which cannot be realized
over any finite extension of Q.
Perles configuration (see Example 12.3) appeared in his original construction of
an irrational polytope. This result may seem harder and unintuitive since in three
dimensions this cannot happen: Corollary 11.3 of the Steinitz theorem asserts that
all convex 3-dimensional polytopes are rational. Let us show first how to construct
a polytope from a point and line configuration, and establish a direct map between
their realization spaces.
The Lawrence construction. Think of a plane configuration C as a collection of vectors
V = {v1 , . . . , vn } R3 , obtained by mapping points (x, y) into (x, y, 1). Clearly, three
points lie on a line if they are mapped into linearly dependent vectors. Now consider
a vector space R3 Rn with the configuration C in R3 , and an auxiliary space Rn
spanned by vectors w1 , . . . , wn . Now consider a polytope PC given by a convex hull
of points
xi = (vi , wi), yi = (vi , 2wi), where xi , yi R3 Rn , 1 i n.
The polytope PC has a special combinatorial structure so that the linear relations (on
vertices) inherit all linear relations of C. To see this, notice that points {xi , yi | i I}
are always in convex position, and thus they lie on a facet if and only if
X
X
(2xi yi ) =
(vi , O) = (O, O),
iI

iI

i.e., when the points {vi | i I} V lie on the same line in C.

From the construction, we immediately conclude that if C is irrational (cannot be

realized over Q), as in Example 12.3, then so is the polytope PC . More generally, if
a field K is a finite extension of Q, then there exists a polynomial P with rational
coefficients and roots not in K. By the configuration universality theorem there exists
a configuration C which satisfies the law of P . Then the corresponding polytope PC
cannot be realized over K.

116

12.6. Exercises.
Exercise 12.1. [1] Prove that the center of a circle can be constructed by using only a
compass (i.e., without a ruler).
Exercise 12.2. a) [1+] Prove that using the ruler, compass and angle-trisector one can
construct a regular heptagon.
b) [2] Prove that a real cubic equation can be solved geometrically using a ruler, compass
and angle-trisector if and only if its roots are all real.
Exercise 12.3. a) [1] Suppose a1 , . . . , an > 0. Find the necessary and sucient conditions
for the existence of an n-gon with these edge lengths to be inscribed into a circle.
b) [2-] Prove that for n 5, the radius of this circle cannot be constructed with a ruler
and a compass (given edge lengths ai ).
Exercise 12.4. [1+] A geometry problem asks to use a ruler and compass to nd a square
with two vertices on each of the two given non-intersecting circles. Show how to solve it or
prove that this is impossible in general.
Exercise 12.5. [1] The y = x2 parabola is given, but the axes are erased. Reconstruct the
axes using ruler and compass.
Exercise 12.6. [1] Prove that a regular pentagon cannot be realized in Q3 . Conclude that
the regular icosahedron cannot be realized in Q3 .
Exercise 12.7. [1] The Pappus theorem states that for every two triples of points on two
lines: a, b, c L and a , b , c L the points of intersection of x = (a, b )(a , b), y = (a, c )
(a , c), z = (b, c ) (b , c) must lie on a line. Remove this extra line to obtain a conguration
with nine points and eight lines. Convert the Pappus theorem into a realizability result of
this conguration over R. Can this conguration be realized over C or over a nite eld?
Exercise 12.8. [2] Find a (self-intersecting) polyhedral surface in R3 which cannot be
realized over Q.
Exercise 12.9. [2-] Find the analogue of Theorem 12.4 for polytopes. In other words,
prove a universality type theorem which extends Theorem 12.5.

12.7. Final remarks. Thinking of Theorem 12.1 as a universality result is a classical idea
(see [CouR, Ch. III 2]). Our writing was inuenced by a terric essay [Man]. Let us
mention a well-known result by Mohr and Mascheroni, which asserts that every construction
with a ruler and a compass is possible using a compass alone (see [CouR, Ch. III 5]).
In a dierent direction, Steiner showed that given a xed circle and its center, all the
constructions in the plane can be carried out by the ruler alone. We refer to classical
textbooks [Bieb, Leb2] for these and other results on geometric constructions (see also [Adl,
Schr1]), and to [Hun, Kos, Mar, Smo] for an introduction (see also nice exercises in [CouR,
Ch. IV 7]). Finally, an interesting complexity result comparing capabilities of two tools is
given in [ABe].
Realizability of point and line congurations over elds is a part of the much more general
subject of matroid theory. In a dierent direction, connections between various geometries
go beyond realizability over elds, as there exist geometries without Desargues theorem,
etc. [Hil, Kag2].

117

What we call Perles conguration is the point and line conguration which appears in
Perless construction of the rst non-rational polytope [Gr
u4, 3.5]. Theorem 12.4 is a weak
version of the Mnev Universality Theorem. For the introduction to the Mnev theorem and
the Lawrence construction see [Zie1, 6.6] and the original paper [Mnev]. The full proof of
the Mnev theorem, further extensions and references can be found in [Ric].

118

13. Universality of linkages

This is the second of two sections on universality, where we investigate realizations
of graphs in the plane with given edge lengths. The type of realization spaces we
consider here will prove crucial in Section 31 where we study realization spaces of
polyhedral surfaces. The universality of linkages, the main result in this section, also
removes much of the mystery behind various constructions of flexible polyhedra which
we discuss in Section 30.
13.1. A machine can sign your name (as long as it is algebraic). Imagine a
configuration of bars and joints in the plane where some joints are fixed to a plane.
Such configuration is called a linkage. If the linkage is not rigid, one can place a pen
in one of the joints and let the configuration move. The result is a curve in the plane,
and one can ask what curves can be obtained in this fashion.26 For example, in a
linkage shown in Figure 13.1, a pen placed in non-fixed joints 1, 2, 3 give (parts of)
the circles, while a pen in joint 4 gives a more complicated curve.
1

2
1

4
3

Figure 13.1. Two realizations of the same linkage with two fixed
point a and b.
Let us make few remarks before we formalize the setup and state the main theorem.
First, we always assume that the linkage is connected and has at least one fixed joint.
Because of the metric constraints one cannot hope to obtain an unbounded curve,
such as a parabola. Similarly, the resulting curve can be disconnected (this is in fact
the case for the linkage in Figure 13.1, where joint 1 cannot be continuously pushed
down); while the pen analogy no longer makes sense, we will consider all connected
components together as parts of the curve. Finally, in contrast with the point and line
configurations, we will allow all possible degeneracies such as intersections of bars,
overlap of joints, etc.
Formally, define a linkage L to be a graph G = (V, E) and a length function
L : E R+ , where we denote the length of the edge e = (v, w) by e = L(e). Fixed
points are the vertices H V and a function h : H R2 . A realization of linkage L
is a function f : V R2 such that f (v) = h(v) for all v H, and the distance of the
image of an edge |f (v) f (w)| = e for all e = (v, w) E. By ML denote the space of
realizations of L.
26One

can also talk about general semi-algebraic regions which can be drawn in the plane this
way. We will restrict ourselves to curves for simplicity.

119

Suppose we are given an algebraic curve C R2 , defined by

F (x, y) =

n X
n
X
i=0 j=0

ci,j xi y j = 0, where ci,j R, n N.

We say that linkage L draws curve C if the space of realizations of some joint v V ,
defined as ML (v) = {f (v) | f ML } lies on C, i.e. F (x, y) = 0 for all (x, y)
ML (v). Finally, we say that curve C is mechanical if every compact subset X C
lies in the realization space of a joint of a linkage which draws C: X ML (v) C.
Theorem 13.1 (Linkage universality). Every algebraic curve C R2 is mechanical.
The idea of the proof is roughly similar to that of the configuration universality
theorem (Theorem 12.4). We will construct a linkage with two degrees of freedom
along axes x and y. We place the pen into a joint which will always be located
at (x, y). This linkage will also have a joint z located at (F (x, y), 0) on axis x. Once
we fix the point z at the origin O, we obtain the desired linkage.
Outline of the proof. Let us start with some basic remarks. First, in the definition of
linkages, the bars are allowed to connect only in the joints. On the other hand, by
rigidifying the bars we can also connect them to the middle of bars as in Figure 13.2.
Thus, from now on we can always allow joints in the middle of bars.

Figure 13.2. Rigidifying the bars with joints in the middle.

Second, we will ignore the connected components issue: there is always a way to
add few extra bars to remove the undesired realizations. For example, in Figure 13.3
we first show how to ensure that the rigid triangle has all realizations oriented the
same way, and then how to avoid the self-intersecting realizations in a parallelogram.

Figure 13.3. Avoiding extra realizations of linkages.

120

Finally, in our linkages we will never specify the absolute lengths of the bars, only
their relative lengths to ensure the realizations lie on a given curve C. It is easy to see
that the lengths can be taken large enough to satisfy the conditions of the theorem.
Let us start constructing a linkage by making a frame. Fix the joint at the origin O,
and let x and y be the point at the end of the rhombi (Oaxb) and (Ocyd), as in
Figure 13.4, with sides 1 and 2 . Later we will ensure that these points are moved
only along the corresponding axes. We will also need to obtain point y along the x
axis, and at distance |yO| from the origin. We do this as in Figure 13.4,
by adding
an equal rhombus (Ocy d ) and adding bars (cc ) and (dd ) of length 22 , to form
right equilateral triangles (cOc) and (dOd).

y
c

y
d

d
c

2
O

O
d

Figure 13.4. Obtaining axis joints x, y and the joint y .

To obtain the desired joint F (x, y) as above we need to be able to add and multiply
distances of joints along the x axis, add constants, and multiply by constants.
The translator construction in Figure 13.5 is an easy way to make addition: z =
x + c. For c > 0, simply take two fixed point a, b at distance c, parallel to x axis.
Then connect them two points x, z through two parallelograms. Note that subtraction
follows similarly, by switching roles of x and z. The case c < 0 can be made by a
similar construction.
Similarly, the pantograph construction in is an easy way to make multiplication:
z = x c, for any real c R. Simply make homothetic triangles (Oux) and (Owz),
the second with sides c times the first. To ensure they are homothetic add a parallelogram (xuwv). Note that the pantograph construction uses joints in the middle of
bars, which we can use by the observation above.
Adding two points by linkages is in fact easy to do in general, not only when
the points are restricted to a line. The adder in Figure 13.6 models vector addition
z = x + y. On the other hand, modelling multiplication by linkages is a delicate task
which requires some preparation.
First, observe that squaring is enough to model multiplication since
xy =

1
1
(x + y)2 (x y)2 ,
4
4

121

O
u
v
d

Figure 13.5. Translator models addition z = x + d, and pantograph

models multiplication by a constant z = x c, where c = |Ow|/|Ou|.
and we already know how to add, subtract, and multiply by a constant. Similarly,
squaring can be obtained from the inversion, since
1
1
2

= 2
.
z1
z+1
z 1
Finally, the celebrated Peaucellier inversor shown in Figure 13.6, models the inversion
z = (a2 b2 )/x. Indeed, for the height of the triangle we have:

2

2
x+z
xz
2
2
2
= b
.
h = a
2
2
Therefore,
xz =

x+z
2

2

xz
2

2

= a2 b2 ,

as claimed.27 Let us summarize what we just did: through a sequence of above

reductions, we obtained the desired multiplication by linkage constructions.
y
z

a
O

b
z

O
x
Figure 13.6. Adder z = x + y and inversor z = (a2 b2 )/x.
27There

is an often overlooked problem with the Peaucellier inversor which has to do with the
central rhombus collapsing. Even if the rhombus is rigidied as above, one can still rotate the sides
around the center. Attaching two extra edges to the opposite vertices in the rhombus prevents that;
we leave the details to the reader.

122

It remains to show that we can have joints x and y move along the axis, an assumption we used in the beginning. In fact, drawing a straight line is historically the
hardest part of the proof. This can be done by a linkage as in Figure 13.7. To see
why the linkage draws a straight line, note that the main part of it is the Peaucellier
inversor, which maps the dotted circle into a dotted line (i.e., a circle of infinite
radius).

x
v

Figure 13.7. Drawing a straight line.

Now, after all these constructions we need to put everything together. Draw two
orthogonal lines, representing the axes of the coordinate system as in Figure 13.4
above. This is our frame and the desired joint will be in position v = (x, y). Rotate
the joint y onto x-axis and obtain joint F (x, y) via a series of operations (additions,
multiplications, etc.) as above. Fix this joint F (x, y) at the origin O. This restricts
the space of realizations of v to the curve F (x, y) = 0, as desired.
Finally, the relative restrictions on the sizes of our bars will give limits on the
compact subset of curve C. It remains to check that for every compact subset X
of C there exist an appropriate collection of bar lengths so that the realization space
would contain X. We omit this part.

13.2. Spherical linkages. It is natural to ask whether the configuration universality
theorem (Theorem 12.4) is related to the linkage universality theorem (Theorem 13.1).
As it turns out, the former is a special case of the spherical version of the latter. In
other words, the world of linkages is much richer and contains planar configurations
as a very special case. The following simple construction starts with a planar configuration and produces a spherical linkage with (essentially) the same realization
space.
Spherical linkage construction. Start with a planar configuration C = (V, E) and
make one joint on a sphere for every point ai V and every line j E. Whenever a
point ai lies on a line j , attach between them a (spherical) bar of length /2. Note
that for every spherical realization of the resulting linkage L, the points which lie on
the same line j E now lie on the same great circle in S2 and vice versa.

123

In order to make this construction completely rigorous and straightforward one

needs to substitute S2 with RP2 and consider all projective realizations of C. We
leave the details to the reader.
13.3. Exercises.
Exercise 13.1. [1+] Give an explicit construction of a linkage which draws a regular pentagon.
Exercise 13.2. [1+] Give an explicit construction of a linkage which draws an ellipse with
distinct axes.
Exercise 13.3. [2-] Prove that 3-dimensional algebraic curves and surfaces can be drawn
by 3-dimensional linkages.
Exercise 13.4. Dene a spherical polygon [x1 . . . xn ] S2 to be the union of geodesic
arcs [xi xi+1 ] on a unit sphere. Denote by L the set of sequences of lengths of these arcs:
(1 , . . . , n ),

where i = |xi xi+1 |S2 .

a) [1-] Prove that if a sequence (1 , . . . , n ) L then so does every permutation ((1) , . . . , (n) ),
Sn .
b) [2-] Prove that (1 , . . . , n ) L if and only if
X
iI

0 i , for all 1 i n, and

X


i
j + |I| 1 , for all I [n], |I| odd.
j I
/

c) [1] Prove that the inequalities in part b) are non-redundant.

13.4. Final remarks. The study of linkages is long, classical and exciting. It became
popular with Watts discovery in 1784 of a linkage which approximately translates a forward
into a circular motion. This mechanical linkage has been used in trains and is still in use
in many car models. Chebyshev studied Watts linkage for thirty years and introduced his
own linkages. Interestingly, he believed that a linkage which draws a line is impossible to
construct and introduced the Chebyshev polynomials in an eort to approximate the line
with low degree polynomials. Somewhat unusually, Peaucellier announced his discovery in
a letter in 1871 without a hint to the actual construction, so in 1873 it was independently
rediscovered by Lipkin. We refer to [CouR, Ch. III 5.4] and [HilC, 40] for the introductory
treatment, and to a celebrated lecture by Kempe [Kem] who essentially proved our linkage
universality theorem (Theorem 13.1). For the story of interesting polyhedral linkages in
three dimensions see [Gol1].
The universality of linkages idea was rst proposed by Kempe and Lebesgue (see above),
and was revived by Thurston who popularized it with a saying there exists a linkage which
signs your name[King]. These results were recently extended and formalized in a powerful
paper by Kapovich and Millson [KM4] (see also [King]). Our Theorem 13.1 is a toy version
of their results, which are much more technical and precise. On the other hand, the basic
idea can already be seen in the proof above. Finally, the connection between spherical
linkages and planar congurations was found in [KM2].
Let us mention that throughout history kinematics has been an honorable eld of study,
and some of the linkages described above (such as the pantograph) have been used in

124

practice. One can also use the linkages to make various computations and until very recently
such linkage based machines were still in use in navigation and other elds [Svo].28 From
a theoretical point of view, the universality theorem shows that mechanical computers in a
certain sense can nd roots of all algebraic equations over R. Fortunately, digital computers
do a much better job.

28In

1948 the author of the Math. Review on [Svo] praised them as the most economical, reliable
and sturdy machines.

125

14. Triangulations
In this short section we introduce and study Voronoi diagrams and Delaunay triangulations, some of the most basic constructions in computational geometry. They will
be used later in Section 21 in the proof of the four vertex theorem, and in Section 40
to study non-overlapping unfoldings of convex polyhedra. We also study the flips
(local moves) on the planar and regular triangulations. Flips are an important tool
which will prove useful later in Section 17.
14.1. Flips on planar triangulations. Let X R2 be a set of n points in the
plane. Denote by T (X) be the set of triangulations T of conv(X), such that vertices
of triangles in T are all in X, and every point in X appears as a vertex of at least
one triangle in T . We call these full triangulations. Define a flip (also called 2-move)
to be a transformation T1 T2 on triangulations T1 , T2 T (X) which replaces one
diagonal of a convex quadrilateral with another one (see Figure 14.1).

Figure 14.1. An example of a flip.

We say that points in X are in general position if no three points lie on a line and
no four points lie on a circle.
Theorem 14.1. Let X R2 be a set of n points in the plane in general position.
Then every two full triangulations of conv(X) are connected by a finite sequence of
flips. Moreover, n2 flips will always suffice.
In other words, for every T, T T (X) we have T T1 T2 . . . T T ,
for some T1 , T2 , . . . , T T (X) and < n2 (see Figure 14.2). We prove the theorem
later in this section.

Figure 14.2. A sequence of moves connecting all triangulations with

vertices at given six points.
Remark 14.2. (Triangulations of a convex polygon). When n points in X R2 are
in convex position, the number of triangulations |T (X)| = Cn2 is a Catalan number (see
Example 8.5). The connectivity of the ip graph Gn on all triangulations of a convex polygon
is an elementary result which can be seen directly, by connecting any triangulation to a star
triangulation (a triangulation where all diagonals meet at a vertex, see Subsection 17.5), or

126

by induction. Not only is this a special case of Theorem 14.1, but also of Theorem 14.10
presented below. Recall that the ips in this case correspond to edges of the associahedron,
a simple (n 3)-dimensional convex polytope (see Example 8.5 and Exercise 8.2). By
Balinskis theorem (Theorem 8.8), this implies that the ip graph Gn is (n 3)-connected.

14.2. Voronoi diagrams and Delaunay triangulations. Let X = {v1 , . . . , vn } be

a finite set of points in the plane. Define the Voronoi diagram VD(X) as a collection
of Voronoi cells Di , defined as the set of points z R2 which are closer to vi than to
any other point in X:


Di = w such that |wvi | < |wvj | for all i 6= j .

By construction, cells Di are open, possibly unbounded convex polygons (see Figure 14.3). Clearly, for X in general position no four cells of the Voronoi diagram
VD(X) meet at a point.

Figure 14.3. Voronoi diagram and the corresponding Delaunay triangulation.

Connect by a line segment every two points vi 6= vj , such that cells Di and Dj
are adjacent (see Figure 14.3). A Delaunay triangulation DT(X) is defined by the
resulting edges on X and can be viewed as a subdivision dual to VD(X). The name
is justified by the following result.
Proposition 14.3. Let X = {v1 , . . . , vn } R2 be a finite set of points in general
position. Then the Delaunay triangulation DT(X) defined above is a triangulation
in T (X).
Proof. First, let us show that for every vertex vi of conv(X) the cell Di is unbounded.
Take a bisector ray R of the outside angle at vi and observe that all points w R lie
in the Voronoi cell Di . Similarly, let us show that for every edge (vi , vj ) of conv(X),
the cells Di and Dj are adjacent. Take a ray R bisecting and perpendicular to (vi , vj ).
Observe that points w R far enough from (vi , vj ) have vi and vj as their closest
points. This implies that DT(X) contains all edges of conv(X). On the other hand,
since X is in general position, no four cells of the Voronoi diagram meet at a point,
which implies that DT(X) is a triangulation.


127

Let T T (X) be a triangulation of conv(X). We say that a circle circumscribed

around triangle (vi vj vk ) in T is empty, if it contains no other vertices of X in its
interior.
Proposition 14.4 (Empty circle condition). Let DT(X) be the Delaunay triangulation of a polygon conv(X), where X R2 is a finite set of points in general position.
Then the circles circumscribed around triangles in DT(X) are empty.
The conclusion of the proposition is called the empty circle condition. Later in this
section we prove that this condition is not only necessary, but also sufficient (for a
triangulation to be the Delaunay triangulation).
Proof. In the Voronoi diagram, consider the circumscribed circle C around triangle
(vi vj vk ). By construction, the center O of C is a meeting point of cells Di , Dj and Dk .
Since X is generic, no other point can lie on C. If some vr lies inside C, that means
that O must be strictly closer to vr than to vi , a contradiction.

Define a map : R2 R3 by (x, y) = (x, y, x2 + y 2 ). Consider a convex hull P
of points wi = (vi ). A face F of P is said to be in the lower convex hull if P lies
above the plane spanned by F . Projections of all faces in the lower convex hull onto
the plane (x, y, 0) give a subdivision of conv(X). The following result shows that this
subdivision is in fact a Delaunay triangulation.
Theorem 14.5 (Paraboloid construction). Let X R2 be a finite set of points in
general position. Then, the polytope P defined above is simplicial and the projections
of the triangular faces in its lower convex hull is the Delaunay triangulation DT(X).
Note that we have some freedom of choice in the paraboloid construction, by making
any parallel translation of it. The theorem implies that the construction is invariant
under these transformations.
Proof. First, let us show that every intersection of a non-vertical plane z = ax+by +c
and a paraboloid (x, y, x2 + y 2) project onto a circle (see Figure 14.4). Indeed, the

2

2
2
2
intersection satisfies x2 + y 2 = ax + by + c, i.e. x a2 + y 2b = c + a4 + b4 , as
desired. Now, if four points in (X) lie in the same plane, then their projections are
four points in X which lie in the same circle. This is impossible since X is in general
position. This implies that polytope P is simplicial.
Let F be a triangular face in the lower convex hull of P , let H be the plane
spanned by F , let be the projection of F onto the plane (x, y, 0), and let C be
the circumscribed circle around . From above, if a point vi lies inside C, then (vi )
lies below H, which contradicts the assumption that F is in the lower convex hull.
Therefore, every circumscribed circle C as above is empty, i.e., contains no other
vertices of X in its interior.
In the opposite direction, suppose T T (X) is a triangulation of conv(X), such
that all circles circumscribed around triangles = (vi vj vk ) in T , are empty. Let H be
the plane spanned by the triangle F = (wi wj wk ), where ws = (vs ). By the argument
above, points ws lie above H, for all s
/ {i, j, k}, which implies that F is a face of P .
By Proposition 14.4, in the Delaunay triangulation T = DT(X) all circumscribed

128

Figure 14.4. Intersection of a paraboloid with a plane H projects

onto a circle C.
circles as above are empty. Therefore, DT(X) is equal to the projection of the lower
convex hull of P , as desired.

Corollary 14.6 (Empty circle criterion). Triangulation T T (X) is the Delaunay
triangulation DT(X) if and only if it satisfies the empty circle condition in Proposition 14.4.
This follows immediately from the last part in the proof of the theorem.
14.3. Local convexity is just as good as global one. To prove the main theorem
(Theorem 14.1) we will need the following technical and intuitively obvious result.
Let X = {v1 , . . . , vn } R2 be a set of points in general position, and let T T (X)
be a triangulation of A = conv(X). Fix a function f : X R and extend function f
to the whole A by linearity on each triangle in T . The resulting function : A R
can be viewed as a surface S = Sf R3 which projects onto A.

Lemma 14.7 (Local convexity criterion). Suppose : conv(X) R and the surface S defined above is convex at every diagonal in T . Then S is a lower convex
hull of points {(x, (x)), x A}.
In other words, the lemma is saying that the local convexity condition on S (at
every diagonal) implies the global convexity of S. One can think of this result as a
3-dimensional generalization of Exercise 24.2, and the proof follows similar lines29
Proof. Let us first show that the cones Cx at every interior vertex x = (vi , (vi )) of
S are convex. Take x and intersect the neighborhood of x in S with a small sphere.
The intersection is a simple spherical polygon with all angles < . By the spherical
analogue in Exercise 24.2, this implies that the cone Cx is convex.
Let P = {(x, z), x A, z (x)} be the set of points above S . Proving that P
is convex implies the result. Suppose P is not convex. Let us use a version of the
argument as in the proof of Exercise 24.2. Then there exist two interior points x, y P
such that the shortest path between x and y inside P is not straight. Denote by y
a point on the surface S where is locally not straight. Now note that neither y can
29The

reader might want to solve the exercise rst, or at least read the hint.

129

lie in the vertices of S because the cones are convex, nor on the edges of S , because
all dihedral angles are < , nor on faces of S , a contradiction.


14.4. Increasing flips and the proof of Theorem 14.1. Suppose (vi vj vk ) and
(vj vk vr ) are triangles in a triangulation T T (X), such that vr lies inside a circle circumscribed around (vi vj vk ). In other words, more symmetrically, suppose
vj vi vr + vi vr vk < vi vj vk + vj vk vr . A flip replacing (vi , vk ) with (vj , vr ) is
called increasing (see Figure 14.5). Observe that we can make an increasing flip if an
only if xk lies in the circle circumscribed around (vi vj vr ).
vk

vk

vj

vj
vr
vi

vr
vi

Figure 14.5. An increasing flip.

Suppose T T (X) is a triangulation different from the Delaunay triangulation DT(X). From the empty circle criterion (Corollary 14.6), there must be a
non-empty circumscribed circle. The following result is a stronger version of this
observation.
Lemma 14.8 (Increasing flip condition). Let X = {v1 , . . . , vn } R2 be a set of
points in general position, and let DT(X) be the Delaunay triangulation. Then every
triangulation T 6= DT(X) allows at least one increasing flip.
We say that vertex vk is adjacent to triangle = (vi vj vr ) if it forms a triangle
with one of the edges of . The lemma then says that in order to check whether T is
a Delaunay triangulation it suffices to check the empty circle condition only for the
vertices adjacent to the triangles.
Proof. Consider a triangulation T of A = conv(X), and let : R2 R be as in
the paraboloid construction. Denote by S R3 the surface as in the proof of
Lemma 14.7. If T has no increasing flips, then every circumscribed circle around
triangle in T does not have have vertices adjacent to in its interior. Then, by
the same argument as in the proof of Theorem 14.5, surface S is convex upward
at every diagonal in T . Now the local convexity criterion (Lemma 14.7) implies
that S is a lower convex hull. From Theorem 14.5, we conclude that T = DT(X), a
contradiction.


130

Proof of Theorem 14.1.
Fix a triangulation T T (X). We will show that starting
with T , after at most n2 increasing flips, we can always obtain
This implies

DT(X).
n
2
that every two triangulations are connected by at most 2 2 < n flips.
Let S = S be the surface defined as above. Suppose there exists an increasing flip
T T , replacing some (vj , vk ) with some (vi vr ). Observe that the resulting surface
S lies directly below S, and, in particular, below the edge (vj , vk ). Continue making
increasing

flips. From above, the edges of the flips are never repeated, and after at
most n2 flips we obtain a triangulation without increasing flips. By the lemma, this
must be the Delaunay triangulation DT(X).

Remark 14.9. (Delaunay triangulations in higher dimensions). For a set of points X
Rd , d 3, the denitions of Voronoi diagram and Delaunay triangulations extend nearly
verbatim (see Exercise 14.2). The same goes for the empty sphere criterion (Corollary 14.6
and the Delaunay triangulation construction (Theorem 14.5).
Unfortunately, the increasing ips argument (Lemma 14.8) does not hold in full generality.
Figuring out what goes wrong is dicult to see at rst. That is because the intuitively
obvious local convexity criterion no longer holds (see Exercise 14.3). In fact, the analogue
of Theorem 14.1 is an open problem for triangulations of convex polytopes in R3 . It fails
for general sets X R3 (that are not necessarily in convex position), and for convex X in
higher dimensions (see Subsection 14.7).

14.5. Regular triangulations. While a direct generalization of Theorem 14.1 fails

for triangulations in higher dimensions (see Remark 14.9 and Subsection 14.7), choosing a nice subset of triangulations is still connected by flips. The following variation
on Theorem14.1 will prove useful in Section 17.
Let P Rd be a simplicial convex polytope and let V be the set of its vertices. We
say that P is generic if no d + 1 vertices lie in the same hyperplane. A triangulation
of P is a full triangulation with vertices in V . By a slight change of notation, denote
T (P ) the set of all triangulations of P ; these triangulations are exactly full triangulations on the set of vertices V . For the rest of this section, we will consider only full
triangulations of polytopes.
For every
function : V R, let S be the lower convex hull surface of points


d+1
v, (v) inR
defined as above. The last coordinate of a point (x, ) S is called
the height of x P , and the function is called the height function.
Observe that the projection of S onto P gives a polyhedral subdivision Q of P
(see Figure 14.6). We call Q the subdivision associated with the function . We say
that is generic if no d + 2 vertices of S lie in the same hyperplane in Rd+1 . Note
that when is generic, Q is a full triangulation: Q T (P ).
A full triangulation T T (P ) is called regular if there exists a height function ,
such that T = Q is a triangulation associated with . Denote by T (P ) the set of
all regular triangulations T T (P ). For example, by the paraboloid construction
(Theorem 14.5), the Delaunay triangulation of a convex polygon is always regular.
In fact, part a) of Exercise 14.10 says that every triangulation of a convex polygon is
regular.

131

For the rest of this section we take d = 3; already in R3 , the flips on triangulations
are no longer obvious. Take a bipyramid with two triangulations T1 and T2 as in
Figure 14.6, with two and three tetrahedra, respectively. Define the 23 move T1 T2
to be a local transformation between them.
S

P
Q

Figure 14.6. Pentagon P , surface S and a triangulation Q associated with ; a 23 move.

Theorem 14.10 (Connectivity of regular triangulations). Let P R3 be a generic
convex polytope. Then every two regular triangulations T0 , T1 T (P ) are connected
by a finite sequence of 23 moves.
Proof. Before we proceed with the proof, let us make some preliminary observations.
Denote by (P ) the set of all generic height functions on P . Because P is also generic,
the condition that (P ) is equivalent to an inequality, for every 5-tuple of vertices
of P . Geometrically, in the space W = RV of functions on the vertices, this means
that (P ) W is a complement to a finite union of hyperplanes.
Suppose T0 and T1 are two regular triangulations of P , associated with the height
functions 0 and 1 , respectively. We assume that 0 and 1 are generic, since otherwise
we can perturb them without changing triangulations.
Consider now a family of height functions t = (1 t)0 + t1 , where t [0, 1].
Again, we can assume that all t are generic, at all but a finite set of values Z [0, 1].
Indeed, if we view 0 and 1 as points in W , we can always perturb them in such a
way that the straight interval between them crosses hyperplanes discussed above one
at a time. In other words, the interval [0 , 1 ] will lie in (P ) everywhere, except for
finitely many points corresponding to these intersections.
Denote by St = St R4 the lower convex hull surface defined above, and by Qt
the associated subdivision of P , for all t [0, 1]. From above, the surfaces St are
simplicial for all t
/ Z, and Qt are regular triangulations. Note also that between
different values in Z, the triangulations are unchanged, and the change can happen
only at Z.
Consider now what happens at z Z. By construction, Sz is simplicial everywhere
except at one facet F with five vertices. Therefore, Qz is a triangulation except
for a 5-vertex polytope. Recall that the only generic 5-vertex polytope in R3 is a
bipyramid, and that the latter have only two triangulations corresponding to a 23

132

move. Therefore, by continuity, we conclude that Qt where t z, and Qt where

t z+, correspond to two triangulations which differ by a 23 move. Since Z is
finite, this implies that there is a finite sequence of 23 moves corresponding to all
z Z, which connects T0 and T1 .

14.6. Exercises.
Exercise 14.1. a) [1] State precisely what generality assumption need to be used on 0
and 1 in the proof of Theorem 14.10.
b) [1] Suppose polytope P R3 has n vertices. Give a polynomial upper bound on the
number of 23 moves between any two full triangulations of P .
c) [1] Generalize Theorem 14.10 to higher dimensions.
Exercise 14.2. (Delaunay triangulations in higher dimensions) Dene the Voronoi
diagram and Delaunay triangulation of a nite sets of points X Rd in general position,
d 1.
a) [1] Prove the empty sphere criterion, a generalization of Proposition 14.4 to all d 1.
b) [1+] Generalize Theorem 14.5 to all d 1.

Exercise 14.3. (Local convexity criterion in R4 ) [1+] Find a set X R3 , a triangulation

of A = conv(X), and a function f : R3 R, such that the local convexity criterion fails on
Sf R4 , i.e., show that Sf can be convex at all 2-faces, but not globally convex.

Exercise 14.4. (Inverse Voronoi diagrams and inverse Delaunay triangulations) Let
X R2 be a nite set of points in general position. Subdivide the plane into cells according
to the furthest point in X. For this inverse Voronoi diagram dene the corresponding
inverse Delaunay triangulation IDT(X).
a) [1] Prove the full circle condition, the analogue of Proposition 14.4.
b) [1] Give a paraboloid style construction of IDT(X), the analogue of Theorem 14.5.
c) [1] Dene decreasing ips to be the inverse of increasing

ips. Prove that every triangulation T T (X) is connected to IDT(X) by at most n2 decreasing ips.
Exercise 14.5. (Double chain conguration) a) [1] Let Xn R2 be the set of n = 2k
points as in Figure 14.7. Prove that the edges as in the gure must be present in every full
triangulation of conv(Xn ). Conclude that the number of full triangulations is equal to


2k 2
2
|T (Xn )| =
Ck2
,
k1
2n

1
where Cn = n+1
n is the Catalan number (see Example 8.5).
b) [1+] Prove that the diameter of the ip graph on T (Xn ) is (n2 ).
Exercise 14.6. [1] Let T be a full triangulation of a convex n-gon with k interior points
x1 , . . . , xi . Suppose every interior point xi is adjacent to at least 6 triangles in T . Prove
that k = O(n2 ).
Exercise 14.7. (Number of planar triangulations) a) [2-] Prove that for every X R2 in
general position, the number of full triangulations is at least exponential: |T (X)| = e(n) .
b) [2] Prove that for every X R2 , the number of full triangulations is at most exponential:
|T (X)| = eO(n) .

133

Figure 14.7. Double chain configuration of points.

Exercise 14.8. Let X R2 be the set of n points in general position.
a) [2-] Prove that every triangulation T T (X) allows at least (n 4)/2 ips.
b) [2-] Prove that every triangulation T T (X) allows at least (n 4)/6 simultaneous ips
(ips along pairwise non-adjacent diagonals).
c) [1+] Show that part b) cannot be improved to n/5 simultaneous ips.
Exercise 14.9. [1] Find a set X R2 of n points and a triangulation T , such that
(i) there is an increasing ip sequence from T to DT(X) of length (n), and
(ii) there is an increasing ip sequence from T to DT(X) of length (n2 ).
Exercise 14.10. (Flips in convex polygons) a) [1-] Prove that every two triangulations
of a convex n-gon are connected by atmost 2n ips.
b) [2] Prove the lower bound 2n O( n) in the convex case.

Exercise 14.11. Let X R2 be a nite set of points in general position.

a) [1] Consider a complete graph on X with weights on edges given by Euclidean distances.
Prove that the Delaunay triangulation DT(X) contains the minimum spanning tree in .
b) [1-] Prove or disprove: DT(X) minimizes the sum of edge lengths over all triangulations
T T (X).

Exercise 14.12. (Delaunay triangulations) Let X R2 be a nite set of points in

general position, let T (X) be the set of full triangulations, and let DT = DT(X) be the
Delaunay triangulation.
a) [1-] Suppose a triangulation T T (X) has no obtuse triangles. Prove that T = DT.
b) [1] Denote by (T ) the maximum angle in all triangles of a triangulation T T (X).
Prove that DT minimizes over all T T (X).
c) [1-] Denote by (T ) the minimum angle in all triangles of a triangulation T T (X).
Check that DT does not necessarily maximize over all T T (X).
d) [1] Prove that the inverse Delaunay triangulation IDT (see Exercise 14.4) maximizes
over all T T (X).
e) [1] Denote by (T ) the largest circumradius of all triangles in T . Prove that DT minimizes (T ).
f ) [1+] Denote by (T ) the mean circumradius of all triangles in T . Prove that DT minimizes (T ).
g) [1+] Denote by (T ) the mean inradius of all triangles in T . Prove that IDT maximizes (T ).
Exercise 14.13. (Tuttes formula) [2] Denote by an the number of rooted planar triangulations with 2n triangles (and thus with n + 2 vertices, 3n edges). Prove that
an =

2(4n 3)!
.
n!(3n 1)!

134

Exercise 14.14. (Acute triangulations) Let Q R2 be a convex polygon. An acute (nonobtuse) triangulation of a polygon Q R2 is a triangulation of Q into acute (acute or right)
triangles (see Figure 14.8). Dene the acute dissections similarly.

Figure 14.8. An acute triangulation and an acute dissection of a square.

a) [1-] Prove that every triangle has an acute triangulation. Conclude that every polygon
in the plane has an acute dissection.
b) [1] Prove that every circumscribed convex polygon has an acute triangulation.
c) [1+] Prove that every convex polygon has a non-obtuse triangulation.
d) [1+] Prove that every convex polygon Q R2 which has a non-obtuse triangulation also
has an acute triangulation.
e) [2-] Prove that the surface of every convex polytope P R3 has an acute triangulation.
f ) [] Construct a triangulation of a cube into tetrahedra with acute dihedral angles.
g) [2] A simplex is called acute if all angles between the facets (generalized dihedral angles)
are acute. Let Cd Rd be a hypercube. Prove that Cd has no triangulations into acute
simplices, for all d 5.
Exercise 14.15. A space polygon X = [x1 . . . xn ] R3 is called triangulable if there
exists an embedded triangulated surface homeomorphic to a disk, with the boundary X
and vertices at xi , 1 i n.
a) [1-] Suppose X has a simple projection onto a plane. Prove that X is triangulable.
b) [1] Prove that every space pentagon is triangulable.
c) [1-] Find a space hexagon which is not triangulable.
d) [1+] Find an unknotted space polygon which is not triangulable.
e) [2] Show that deciding whether a space polygon is triangulable is NP-hard.
Exercise 14.16. (Convex polygons) Let X = [w1 . . . wn ] R3 be a space polygon
which projects onto a convex n-gon Q in the plane. Denote by P the convex hull of X and
suppose P is simplicial. Two full triangulations T1 and T2 are associated with the top and
the bottom part of the surface of P , respectively.
a) [1-] Prove that for all Q and a triangulation T1 of Q, there exists a polygon X which
projects onto Q, and and has a shadow T1 . In other words, prove that every triangulation
of a convex polygon is regular: T (Q) = T (Q).
b) [1] Note that the shadows T1 , T2 Q must have distinct diagonals. Prove or disprove:
for every convex polygon Q and every pair of full triangulations T1 and T2 of Q with distinct
diagonals, there exists a polygon X with shadows T1 and T2 , respectively.
c) [2-] A combinatorial triangulation of an n-gon is dened by pairs of vertices corresponding
to its diagonals. Prove that for every two combinatorial triangulations T1 , T2 of an n-gon
with distinct diagonals, there exists a polygon Q L and a space polygon X R3 which
projects onto Q and has shadows T1 and T2 . In other words, prove that T1 , T2 can be
realized as shadows by a choice of Q and X.

135

Exercise 14.17. (Six bricks problem) A brick is a rectangular parallelepiped. A collection

of bricks is called disjoint if no two bricks intersect (even along the boundary). We assume
that bricks are non-transparent.
a) [1-] Can you have six disjoint bricks not containing the origin O, so no brick vertex is
visible from O?
b) [1-] Can you have six disjoint bricks not containing the origin, so that no point far enough
is visible?
Exercise 14.18. (Triangulations of non-convex polytopes) a) [1] Let Q be a simple
polygon in R3 . Prove that Q can be subdivided into triangles without adding new vertices.
b) [1] Show that part a does not extend to non-convex polyhedra in R2 . Formally, nd
an embedded (not self-intersecting) polyhedron P R3 homeomorphic to a sphere which
cannot be subdivided into simplices without adding new vertices.
c) [1+] Construct an embedded simplicial polyhedron P R3 homeomorphic to a sphere,
such that no tetrahedron spanned by vertices of P lies inside P .
d) [1+] Construct an embedded simplicial polyhedron P R3 homeomorphic to a sphere,
such that for some point O inside P , none of the intervals (Ov) lie in P , for all vertices v
of P .
e) [1+] Find an embedded simplicial polyhedron P R3 homeomorphic to a sphere, such
that for some point O outside P , all intervals (Ov) have points inside P , for all vertices v
of P .
Exercise 14.19. [2] Show that the problem whether a given non-convex polyhedron can
be subdivided into tetrahedra without adding vertices is NP-hard.
Exercise 14.20. (Regular triangulations) Let P Rd be a convex polytope. A triangulation D of P is called regular if there exists a height function such that D is associated
with . For example, part a) of Exercise 14.10 says that every full triangulation of a convex polygon is regular. Similarly, from the proof of Lemma 17.15 we know that every star
triangulation of P is regular.
a) [1] Find a (general) triangulation of a convex polygon Q R2 which is not regular.
b) [1+] Find a non-regular full triangulation of a convex polytope in R3 .
c) [1+] Let P be the 3-dimensional cube 1 x, y, z 1. Find a non-regular triangulation
of P , such the vertices of tetrahedra in the triangulation are at the vertices of P or at the
origin O.
d) [1+] Find a non-regular full triangulation of a 4-dimensional cube.
Exercise 14.21. a) [2-] Prove that every combinatorial triangulation (see Exercise 14.16)
is a Delaunay triangulation of some convex polygon.
Exercise 14.22. a) [1+] In a full triangulation of a d-dimensional cube, split the simplices
into d groups according to the number of vertices at level 0 and 1 of the rst coordinate.
Prove that the sum of the volumes of simplices in each group is the same.
b) [1+] Find a full triangulation of a d-dimensional cube with o(n!) simplices.
Exercise 14.23. (Dominoes) Let Z2 be a nite simply connected region on a square
grid. A domino tiling T of the region , write T , is a tiling of by the copies of 12
and 2 1 rectangles. Dene a 2-move to be a ip of 2 vertical dominoes forming a 2 2
square to 2 horizontal dominoes, or vice versa.
a) [1] Prove directly that every two domino tilings of are connected by a nite sequence
of 2-moves.

136

b) [1-] Show that part b) is false for non-simply connected regions. Dene by analogy the
domino tilings in R3 and construct a counterexample to 2-move connectivity for simply
connected 3-dimensional regions. Moreover, show that for every k there exist a simply
connected region in Z3 with exactly two domino tilings, and which requires the moves of at
least k dominoes.
c) [1] Color the squares in in a checkerboard fashion and orient counterclockwise.
Dene a height function h : Z by the following rule. Start at any xed point a
and let h(a) = 0. Now, when going around a black square add 1, when going around a
white square subtract 1 (see Figure 14.9). Prove that for every tileable region function h
is well dened. Show that for every tiling it extends to all (integer) points in . Check what
happens to the height function when a 2-move is applied.

-1 -2 -1 -2 -1
0
0
1
2
-1
1
0 1 0 1
2 3 2

-1 -2 -1 -2 -1
0
0
1
2
-1
1
0 1 0 1
2 3 2

Figure 14.9. Domino tilings: 2-moves, region , the height function

and the minimal tiling of .
d) [1] For two domino tilings T1 , T2 of we say that T1 T2 if for their height function
we have h1 (x) h2 (x) for all x . Suppose T satises T T for all T . Prove
that the maximal value of h (x) of the height function corresponding to T appears on the
boundary.
e) [1+] Suppose is tileable by dominoes. Construct a tiling Tmin of as follows. Compute
the height function of and nd x , such that h(x) is maximal on . Place a domino
inside so that x is adjacent to both squares. Repeat. Prove that this algorithm always
works, i.e., produces the same domino tiling for all choices of x.
f ) [1+] Start with tiling T and repeatedly apply 2-moves until a local minimum T is reached.
Prove that T = Tmin . Conclude that Tmin is global minimum: Tmin T for all T .
Conclude that every two domino tilings of are connected by 2-moves, as in part a).
g) [1] Show that when is simply connected and not tileable by dominoes, the algorithm
in d) fails. Design a linear time algorithm (in the area) for testing the tileability of by
dominoes.
h) [1+] Show that for all T1 , T2 there exist unique T and T such that




h (x) = max h1 (h), h2 (x) , h (x) = min h1 (h), h2 (x) , for all x .
Prove that this denes a lattice structure on all domino tilings of .
i) [1-] When n is a 2n2n square, nd the only two tilings which have admit exactly one
2-move. Check that these are the max and min tilings dened above.
j) [1] When n is a 2n2n square, conclude that the number of 2-moves required to go
from one tiling to another is (n3 ).
k) [2+] Generalize these results to tilings by k1 and 1 rectangles, for every xed k, 2.
Exercise 14.24. (Ribbon trominoes) Consider all tilings of a nite simply connected
region Z2 by copies of the four ribbon trominoes shown in Figure 14.10 (only translations
are allowed).

137

a) [2] Dene 2-moves on the tilings to be exchanges of two trominoes by another two. Prove
that all ribbon tromino tilings of a rectangle are connected by 2-moves.
b) [2] Extend a) to all simply connected regions .
b) [1] Conclude from a) that in any tiling of , the number of ribbon trominoes of the
second type minus the number of ribbon trominoes of the third type is a constant c()
independent of the tiling.
c) [1] Denote by n the staircase shaped region as in Figure 14.10. Show that whenever
area(n ) is divisible by 3, the region n is tileable by ribbon trominoes.
d) [1] Find all n for which n is tileable by ribbon trominoes of the second and the third
type.

4
3
7

Figure 14.10. Ribbon trominoes, their local moves and a tiling of 6 .

Exercise 14.25. (Mosaics) Let P be a centrally symmetric convex polygon with 2n
sides. A mosaic M is a subdivision of P into parallelograms with sides parallel to the sides
of P .30 Denote by R = R(P ) the set of mosaics in
the polygon P .
a) [1] Prove that every such mosaic has exactly n2 parallelograms, each corresponding to
a pair of non-parallel sides.


b) [1+] Prove that the number of mosaics |R(P )| depends only on n and is at most n2 !.

Figure 14.11. Two mosaics of a decagon and a flip.

c) [2-] Dene a ip on R to be a transformation of triples of parallelograms forming a
hexagon into another (see Figure 14.11). Prove that every two mosaics are connected to
each other by a nite sequence of ips. Prove that every mosaic of P allows at least n 2
ips.
d) [2] Dene the height function of a mosaic M to be a piecewise linear convex function
: P R whose ats give M when projected. Mosaic M is called regular in this case.
Denote by R (P ) R(P ) the set of all regular mosaics of P . Prove that when P is generic,
the number of regular mosaics depends only on n. Furthermore, prove that as n a
random mosaic is not regular.
30A

more standard name for these subdivisions is zonotopal tilings, since they extend to general
zonotopes (see Exercise 7.16).

138

e) [2-] When P is deformed, one can also deform a mosaic. Prove that there exists a mosaic
which is not regular for all deformations.
Exercise 14.26. [1] Let P be a centrally symmetric convex polygon. Consider a subdivision
of P into nitely many centrally symmetric polygons. Use Exercise 2.3 to prove that this
tiling has at least 3 parallelograms.
14.7. Final remarks. Voronoi diagrams are named after the Georgy Voronoy who introduced them in 1908. In a dierent context they were studied much earlier, most notably by
Descartes (1644) and Dirichlet (1850). Similarly, Delaunay triangulations are named after
Boris Delone (note another dierence in spelling), who proved the empty sphere criterion
(Exercise 14.2) in 1934. According to Dolbilin, this result was repeatedly rediscovered in the
West, but after Coxeter received a letter from Delone, he read and popularized the original
papers of Voronoy and Delone. For more on Voronoi diagrams and Delaunay triangulations,
their history, applications and references see [Aur, AurK].
The inverse Voronoi diagrams (and dual to them inverse Delaunay triangulations) dened
in Exercise 14.4, go back to the foundational paper [ShaH]. They are usually called the
furthest site or the farthest point Voronoi diagrams [AurK].
The local convexity criterion (Lemma 14.7) is a discrete version of a classical result by
Hadamard (1897), further generalized by Tietze (1928), Nakajima (1928) and others. We
refer to [KarB] for further references.
The ip (local move) connectivity on full triangulations has been also extensively studied
in the literature [San2]. In particular, in dimension d 5 the ip graph can be disconnected,
and there is an evidence that this might be true even for d = 3. For more on various classes of
triangulations, in particular regular triangulations, and connections to other elds see [DRS]
(see also [San2]).
The full triangulations (triangulations with a xed set of vertices) have also been studied
at length in connection with A-discriminants and variations on the theme. The height functions and regular triangulations (see Exercise 14.20) play important roles in the eld. We
refer the reader to [GKZ], where many of these fundamental results were rst summarized
(a number of connections and applications have appeared since [GKZ]).
Let us mention here the mosaics (zonotopal tilings) dened in Exercise 14.25. In a certain
precise sense, the regular mosaics are projections of the top 2-dimensional faces of a
hypercube. In a dierent direction, the mosaics correspond to the pseudoline arrangements,
while regular mosaics are dual to stretchable pseudoline arrangements. We refer to [Bj
o+]
for further results and references.
Finally, the local move connectivity plays an important role in the modern study of nite
tilings. In some special cases one can use the height functions and combinatorial group
theory to establish that certain local moves connect all tilings of a given region by a xed
set of tiles. Some examples are given in Exercises 14.23 and 14.24. See [Pak3] for these and
other tiling results, and the references.

139

15. Hilberts third problem

The name scissor congruence comes from cutting polygons on a plane (hopefully,
with scissors) and rearranging the pieces. Can one always obtain a square this way?
What happens in three dimensions? Can one take a regular tetrahedron, cut it into
pieces and rearrange them to form a cube? This problem has a glorious history and
a distinction of being the subject of one of Hilberts problems. Perhaps surprisingly,
the solutions are elementary, but use a bit of algebra and a bit of number theory.
In this section we give an introduction to scissor congruence in the plane and in
space. Our main result is the proof of Bricards condition which resolves (negatively)
Hilberts original problem. In the next section we present an algebraic approach and
prove further results on scissor congruence. Among other things, we show that if a
polytope is scissor congruent to a union of two or more similar polytopes, then it must
be scissor congruent to a cube. In Section 17 we use a different combinatorial approach
for proving general scissor congruence results. The idea of valuations employed there
will reappear later on, in the proof of the bellows conjecture (Section 34). Finally, in
Section 18 we study a related problem on Monge equivalence. Despite the similarities,
that problem turns out to be much simpler and we completely resolve it.
15.1. Scissor congruence. Let P, Q Rd be two convex polytopes.31 We say that
they are scissor congruent, write P Q, if P can be cut into finitely many smaller
polytopes which can be rearranged and assembled into Q. Formally, polytopes A, B
Rd are congruent, write A B, if B can be obtained from A by a rigid motion:
combination of translations and rotations.32 We say that P Q if each polytope is
a disjoint union33 of congruent polytopes:
P = m
i=1 Pi ,

Q = m
i=1 Qi ,

where Pi Qi . Of course, if P and Q are scissor congruent, then vol(P ) = vol(Q).

Let us first note that scissor congruence is an equivalence relation: if P Q
and Q R, then P R. Indeed, simply take the intersection of two corresponding
decompositions of Q. These smaller polytopes can be reassembled into both P and R,
proving the claim.
Now, the study of scissor congruence starts with the following intuitively clear
result, which was rediscovered on several occasions.
Theorem 15.1 (Bolyai, Gerwien). Two convex polygons in the plane are scissor
congruent if and only if they have equal area.
31Throughout

this and next section, convexity plays no role and can be weakened to any nite
union of convex polytopes. The results and proofs extend verbatim. Still, we decided to keep this
condition for the clarity of exposition.
32Throughout the section we assume that congruent polytopes must have the same orientation.
In fact, every two mirror symmetric polytopes are scissor congruent, so the reader can ignore the
dierence (see Exercise 15.3).
33Here and throughout the next three sections, when we speak of a disjoint union of polytopes
we ignore the boundary.

140

Motivated by his study of foundations of geometry Hilbert noted that this result
can be used to define the area of polygons in a combinatorial way, by setting the area
of a square and axiomatizing the scissor congruence. On the other hand, he recalled
that the Archimedes method of exhaustion for computing the volume of a pyramid
involved essentially taking the limits (cf. Exercise 15.1). In his famous list of open
problems, the third problem asked whether one can avoid taking limits and use scissor
congruence once again. The same year Hilbert posed the problem, it was resolved by
Dehn in the negative:
Theorem 15.2 (Dehn). A cube and a regular tetrahedron of the same volume are
not scissor congruent.
In this section we give elementary proofs of both theorems.
15.2. Even seamsters could use some group theory. While one can view the
BolyaiGerwien theorem as a reasonably hard exercise in elementary geometry, and
find one of the numerous ad hoc proofs, we decided to employ a more enlightening
(even if a bit postmodern) approach to obtain this classical result. Let us start with
the following easy observation.
Lemma 15.3. Let be a discrete group acting on Rd , and suppose convex polytopes P
and Q are fundamental regions of the action of . Then the polytopes are scissor
congruent: P Q.

Here by a fundamental region X Rd we mean the closure of a region X which

satisfies the following property: for every z Rd there exists a unique point x X
such that x lies in the orbit of z under action of . The proof of Lemma 15.3 is
straightforward from definition.
Proof. Consider the tiling of Rd obtained by the action of on Q. This tiling subdivides P into smaller convex polytopes Pi = P (gi Q), for some gi . Since Q
is a fundamental region, these polytopes cover the whole P , and since is discrete
there is only a finite number of gi giving nonempty intersections Pi . Now, reassemble
polytopes Pi into Q as follows: Q = i (gi1 Pi ). Since P is a fundamental region,
this indeed covers the whole polytope Q, so that polytopes Qi = gi1 Pi intersect
only at the boundary. This proves that P Q.


15.3. Cutting polygons into pieces. Now we are ready to prove Theorem 15.1.
We need the following steps:
0) triangulate every convex polygon;
1) show that every triangle with side a and height h is scissor congruent to a
parallelogram with side a and height h/2;
2) show that every two parallelograms with the same side and height are scissor
congruent;
3) show that a union
of two squares with sides a and b is scissor congruent to a
square with side a2 + b2 .

141

Let us show that these steps suffice. Indeed, after step 0) we can start with a
triangle T and convert it into a parallelogram B with
sides a b. Denote by h the
height of B, so that area(B) = ha, and let t = ha. Clearly, h b t. By 2),
parallelogram B is scissor congruent to a parallelogram D with sides a and t, and
height h. Similarly, D is scissor congruent to a square with side t. Finally,
p using 3)
repeatedly one can assemble all the squares into one big square S with sides area(P ).
Therefore, P S Q, which implies the result.

The proof of 1), 2), 3) is easy to understand from the next three figures. In Figure 15.1 let 1 Z2 be a group of translations preserving the colored triangles (or
parallelograms), and let 2 be a Z2 -extension of 1 obtained by adding a central reflection with respect to the origin, which switches colors. Now both the white triangle
and the white parallelogram are fundamental regions of 2 , and the lemma implies
that they are scissor congruent. The last of the three pictures in Figure 15.1 shows
the working of the lemma in this case.

2
1

Figure 15.1. Converting a triangle into a parallelogram.

Similarly, any two parallelograms with the same side and height are fundamental
regions of the action of same group Z2 on R2 by translations. Now the lemma
implies their scissor congruence (see Figure 15.2).
1
2
3
4
4
3
2
1

Figure 15.2. Converting a parallelogram into another parallelogram.

Finally, any two squares can be attached to each other to form a fundamental
regions of the action of the group
of translation Z2 along vectors (a, b) and

(b, a). The square with side a2 + b2 is another fundamental region of . Now the
lemma implies their scissor congruence (see Figure 15.3).
To summarize, we just showed all three congruences 1), 2), and 3) as above can be
obtained from Lemma 15.3. This completes the proof of Theorem 15.1.

142

1
2
3 4
3
4
2
1

Figure 15.3. Converting two squares into a bigger one.

15.4. Chain saw is really not as powerful as scissors. Our proof of Dehns
theorem is somewhat weaker than the original version of Dehn, but more elementary
and sufficient for our purposes.
Let E = {e1 , . . . , el } be the set of edges of a convex polytope P R3 . Denote by i
the dihedral angle at edge ei , where 1 i N. We say that polytope P is fortunate
if can be written as a rational combination of angles i :
c1 i + . . . + cN N = , where ci Q+ , for all 1 i N .

Otherwise, we say that P is unfortunate.

Lemma 15.4 (Bricards condition). An unfortunate polytope in R3 is not scissor

congruent to a cube.
Dehns theorem (Theorem 15.2) follows immediately form Bricards condition. Indeed, all dihedral angles of a regular tetrahedron are equal to = 2 arcsin 13 =
arccos 31 (see the calculation below Figure 20.4). It is not hard to show that (/)
/Q
(see Appendix 41.3), which implies that P is unfortunate. Thus, from the lemma we
conclude that is not scissor congruent to a cube of the same volume, as desired.
Before we begin the proof, let us start consider some special cases of increasing
generality. Start with polytopes P Q, and take the corresponding polytope decompositions P = i Pi and Q = i Qi , where Pi Qi . We say that these are polyhedral
subdivisions if no vertex of Pi (Qi ) lies in the interior of an edge or a face of Pj (Qj ).
Sublemma 15.5. Suppose P = i Pi and Q = i Qi are polyhedral subdivisions into
congruent polytopes Pi Qi , and Q is a cube. Then P is fortunate.
Proof of Sublemma 15.5. Denote by (X) the sum of the dihedral angles of polytope X R3 , and let
X
X
(Qi ),
=
(Pi ) =
i

where the second equality follows from the congruences Pi Qi . We will calculate
in two different ways and compare the results.
Denote by E the set of all edges in the subdivisions of P andPQ. Let i (e) be the
dihedral angle in polytope Pi , at edge e E. Define by (e) = i i (e). We have:
m X
X
X
(e).
=
i (e) =
i=1 eE

eE

143

Now observe that (e) = 2 when e is an interior edge, (e) = when e is on the
boundary of P , and (e) = s when e es is the (usual) edge of P . We conclude:
=

N
X
s=1

ks s + n , for some k1 , . . . , kN N, and n Z+ .

By the same argument for the decomposition of Q, since all dihedral angles are equal
to /2 in this case, we obtain that = l/2 for some l N. This implies that P is
fortunate, as desired.

In the general case, suppose we have vertices which subdivide the edges of Pi into
intervals. By analogy with the previous case, denote these intervals by e, and the
set of intervals by E. Let e = |e| denote the lengths of the intervals. Similarly, for
the decomposition of Q, denote by E the resulting set of intervals e , and by e their
lengths.
Sublemma 15.6. Let P = i Pi and Q = i Qi are polytope decompositions into
congruent polytopes Pi Qi , and Q is a cube. Suppose the lengths of all intervals are
rational: e , e Q for all e E, e E . Then P is fortunate.
Proof of Sublemma 15.6. Taking the previous proof as guidance, we now modify the
definition of to be as follows:
m X
m X
X
X

=
e i (e), =
e i (e ),
i=1 eE

i=1 e E

where i (e) is the dihedral angle in polytope Pi at interval e E, and i (e ) is the

dihedral angle in polytope Qi at interval e E .

Let us first
Pshow that = . For a polytope X with the set of edges U = E(X),
let (X) = uU u (u) be the sum of the dihedral angles in X weighted by the
length u of the edges u. We have
m
m
m
X
X
X
X
X
X
=
e i (e) =
u i (u) =
(Pi ) ,
i=1 uE(Pi ) eE , eu

i=1 uE(Pi )

i=1

where the second equality follows by additivity of the interval length along the edge,
and because the dihedral angles are equal along the same edge. We conclude:
m
m
X
X
=
(Pi ) =
(Qi ) = ,
i=1

i=1

as desired. By analogy with the proof above, we can redefine (e) =

can then rewrite as
X
=
(e).

i e i (e).

We

eE

Since all lengths e are rational, by the same argument as above we conclude that
=

N
X
s=1

cs s + r , for some c1 , . . . , cN , r Q+ .

144

The same argument for the cube Q gives / Q, as desired.

Proof of Lemma 15.4. In full generality, suppose now the interval lengths e are not
necessarily rational. We now change the definition of once again, substituting
values of a general function f : E R and g : E R in place of the interval lengths:
m X
m X
X
X

f =
f (e) i (e) , g =
g(e) i (e ) .
i=1 eE

i=1 e E

If we can find positive functions f, g as above so that f (e), g(e) Q+ for all e E
and e E , and such that = , the result of the lemma then follows by the same
argument as in the proof above.34
Let us write the conditions the functions f, g need to satisfy to guarantee f = g .
Again, by the argument in the proof above, it suffices to check that the sum along
the edge u E(Pi ) is equal to that along the corresponding edge u E(Qi ):
X
X
()
f (e) =
g(e)
eE , eu

e E , e u

Think of () as a set of linear equations for the values f (e), g(e) which must hold
for all polytopes Pi and all edges u E(Pi ). Because e is an obvious positive real
solution of (), there is also a positive rational solution of (). From above, this
finishes the proof of the lemma.


15.5. Exercises.
Exercise 15.1. (Method of exhaustion) [1-] Let P, Q R3 be two convex polytopes
such that vol(P ) < vol(Q). Prove that there exists a polytope P Q, such that P P .

Exercise 15.2. (Extended Bricards condition) [1-] Let P, Q R3 be scissor congruent

convex polytopes. Denote by 1 , . . . , m and 1 , . . . , n their dihedral angles. Prove that
r1 1 + . . . + rm m = s1 1 + . . . + sn n + k ,
for some ri , sj N and k Z.

Exercise 15.3. (Mirror symmetry) [1] Let P, P R3 be two polyhedra which are
congruent up to a mirror symmetry. Prove that P P .

Exercise 15.4. (Parallelepipeds) a) [1] Let R be a parallelepiped in R3 . Prove that R

is rectiable (i.e., scissor congruent to a cube of equal volume).
b) [1-] Prove that the union of two rectiable polytopes is also rectiable.
c) [1] Extends parts a) and b) to higher dimensions.
Exercise 15.5. (Prisms) Dene a prism in R3 to be the Minkowski sum of a convex
polygon in a plane and a line segment.
a) [1-] Prove that every prism is rectiable (scissor congruent to a cube of equal volume).
b) [1] Prove that polytopes Bi in the proof of Lemma 16.6 can be decomposed into two
prisms. This gives an alternative proof that Bi is rectiable.
34We

need positivity to ensure that all coecients ci > 0 in the denition of a fortunate polytope.

145

Exercise 15.6. (Zonotopes) [1-] By Exercise 7.16, recall that every zonotope can be subdivided into parallelepipeds. Use Exercise 15.4 to conclude that every zonotope is scissor
congruent to a cube.
Exercise 15.7. (HadwigerGlur ) Two polytopes P, Q Rd are called T -congruent if
they are unions of a nite number of simplices which are equal up to translations. They are
called S-congruent if central symmetries are also allowed.
a) [1] Prove the following extension of Theorem 15.1: every two polygons P, Q R2 of the
same area are S-equivalent.
b) [1+] Prove that a convex polygon Q R2 is T -congruent to a square of equal area if and
only if Q is centrally symmetric.
c) [1+] Prove that every parallelepiped in R3 is T -congruent to a cube of equal volume.
d) [2-] Prove that a convex polytope P R3 is T -congruent to a cube of equal volume is
and only if P is a zonotope.
Exercise 15.8. Two polytopes P, Q Rd are called D-congruent if P can be decomposed
into smaller polytopes which can then rearranged to Q, such that the boundary points
remain on the boundary. Think of the boundary P as painted, so the goal is to keep the
paint outside (see Figure 15.4).35

Figure 15.4. Two examples of D-congruent polygons.

a) [1] In the plane, prove that every two polygons with the same area and perimeter are
D-congruent.
b) [1+] In R3 , prove that two scissor congruent polytopes with the same volume and surface
area are D-congruent.
c) [1-] In R4 , nd two scissor congruent polytopes with the same volume and surface area,
which are not D-congruent.
Exercise 15.9. Two unbounded convex polyhedra are called scissor congruent if one can
be decomposed to a nite number of (bounded or unbounded) convex polyhedra which can
be assembled into the other.
a) [1-] Prove that this is an equivalence relation.
b) [1-] In the plane, show that innite cones with dierent angles are not scissor congruent.
c) [1+] In the plane, let C C be two unbounded convex polygons such that C r C is
bounded. Prove that C and C are scissor congruent.
d) [] Prove or disprove the 3-dimensional analogue of part c).
Exercise 15.10. (Tiling with crosses) [1] A cross in Rd is a union of 2d + 1 unit cubes,
where one cube in the center is attached to others by a facet. Prove that crosses can tile
the whole space Rd .
35Just as in the

denition of scissor congruence, the boundary condition requirement only concerns

the (d 1)-dimensional faces of P ; we ignore the lower dimensional boundary.

146

Exercise 15.11. (Tiling with notched cubes)

0 < a1 , . . . , ad < 1, is dened as follows:

A notched cube R(a1 , . . . , ad ) Rd , where

R(a1 , . . . , ad ) = { (x1 , . . . , xd ) Rd | 0 xi 1, and xi ai for some i }.

a) [1+] Prove that notched cube R(a1 , . . . , ad ) can tile the whole space Rd , for all ai as
above.
b) [2-] Find all such tilings.
c) [1] Prove that no notched cube can tile a brick.
Exercise 15.12. (Reptiles) A (non-convex) polytope P Rd is called a reptile if it can
tile a copy cP of the same tile, for some c > 1. The order of a reptile is the smallest number
N = cd needed for such a tiling.
a) [1-] Prove that every triangle is a reptile of order 4.
b) [1-] Prove that a trapezoid with sides 2, 1, 1 and 1 is a reptile of order 4. Similarly, prove
that a trapezoid with sides 3, 2, 1 and 2 is a reptile of order 9.
c) [1] Find all rectangular reptiles of order at most 5.
d) [1] Show that notched cube R( 12 , . . . , 12 ) is a reptile of order 2d 1.
e) [1] Suppose polytope P Rd is a reptile of order N . Iterate the tiling construction by
tiling bigger and bigger regions with more and more copies of P . Show that this construction
denes a limit tiling of the whole space, called a substitution tiling.
f ) [1-] For reptiles in parts a) and b), check whether the resulting tiling of the plane is
periodic or aperiodic.36
g) [1] Check that the substitution tiling of Rd with notched cube is periodic (cf. Exercise 15.11).
Exercise 15.13. (Knotted tiles) [2] For every knot K R3 , prove that there exists a
(non-convex) polytope isotopic to K, which can tile the whole space R3 .
Exercise 15.14. Let Q R2 be a convex set of unit area, area(Q) = 1, and let Q1 , . . . , Qn
be its translations. Suppose n is odd. Denote by A the set of points covered by an odd
number of Qi s.
a) [1-] Suppose Q is a square. Prove that area(A) 1.
b) [1-] Same for the regular hexagon.
c) [1] Show that convexity is necessary: if Q is non-convex, we can have area(A) < 1.
d) [1+] Show that if Q is a regular triangle, we can have area(A) < 1. For n = 3, nd the
smallest possible such area.
e) [] What happens for a regular octagon? What about when Q is a circle? How about
general centrally symmetric convex sets?
Exercise 15.15. (-congruence) Let : Rd R be a linear function non-constant on the
axis. Two polytopes P, Q Rd are called -congruent if they are unions of a nite number
of simplices which are equal up to -invariant translations.
a) [1] Suppose (x, y, z) = ax + by + cz, is a linear function on R3 , where a, b, c N. Prove
that the following two bricks are -congruent:
B1 = {(x, y, z) | 0 x b, 0 y c, 0 z a},
B2 = {(x, y, z) | 0 x c, 0 y a, 0 z b}.
36There

is certain degree of ambiguity in the denition of the substitution tiling. Thus, in each
case, check whether one can obtain an aperiodic tiling.

147

b) [1] For an axis brick B = {(x1 , . . . , xd ) | 0 xi ai , 1 i d} denote by [B] the

multiset {(a1 ), . . . , (ad )}. Prove that two axis bricks B1 and B2 are -congruent if and
only if [B1 ] and [B2 ] are equal multisets of rational numbers.
15.6. Final remarks. Although Hilbert motivated his third problem by the foundations
of geometry, the basic ideas of scissor congruence go back to Ancient Greece (see e.g.,
Archimedes Stomachion manuscript). For the comprehensive history of the subject and
an elementary exposition of the scissor congruence see [Bolt, Sah]. The group theoretic
approach is a variation on the classical tessellation technique and is formalized in [M
ul].
The proof we present in Subsection 15.2 is somewhat similar to the proof given in [AFF].
Of course, the congruences resulted in steps 1), 2) and 3) are completely standard.
Dehns original paper (1902) is based on an earlier paper by Bricard [Bri1] who proved
Bricards condition for subdivisions rather than all decompositions of polytopes (see [Ben]).
Boltyansky speculates that although Hilbert never mentioned Bricards paper, he was undoubtedly inuenced by it. Our proof starts with Bricards original approach and then
follows the argument originally due to V. F. Kagan [Kag1] (see also [Hop2, 4.3] for a similar
proof and the references). The proof of scissor congruence of mirror symmetric polyhedra
(Exercise 15.3) also follows Bricards paper.
Bricards condition (Lemma 15.4) is a corollary of Dehns general result on Dehn invariants, which are global invariants involving edge lengths and dihedral angles (see Subsection 17.3). These Dehn invariants generalize to a series of Hadwiger invariants in higher
dimensions (see [Sah]). While in R3 and R4 these invariants are proved to be both necessary
and sucient conditions for scissor congruence (these are results of Sydler and Jessen), in
higher dimensions this remains an open problem [Bolt]. For a modern treatment, extensions
to higher dimensions, spherical and hyperbolic space, and further references, see [Car, Dup].

148

16. Polytope algebra

This is the second of the three sections on scissor congruence. Here we develop
an algebraic approach, by introducing certain algebraic operations on the polytopes
which preserve scissor congruence. This approach allows us to prove that certain
families of polytopes are scissor congruent, while others are not, all without actually
exhibiting the decompositions or computing the dihedral angles. Curiously, the algebraic approach is almost completely independent of the previous section. In this
section we do not fully develop the theory; instead, we present only the most basic
tools and concentrate on attractive examples. Among other things, we prove that all
polytopes which tile the space periodically are scissor congruent to a cube, a result
of particular interest, related to classical problems on tilings.
16.1. The power of negative thinking. While the Bricard condition in Lemma 15.4
(see also a straightforward extension in Exercise 15.2) resolves Hilberts original problem by giving necessary conditions for the scissor congruence, it is very restrictive and
does not give a sufficient condition. Nor is it easy to use, as one is asked to compute
the dihedral angles and verify their rational independence. In fact, even the statement
of Bricards condition is somewhat misleading, as it suggests that scissor congruence
is a property of dihedral angles.37 The following result is an easy corollary of the
main results in this section, and gives a flavor of things to come.
Theorem 16.1 (Sydler). A regular tetrahedron is not scissor congruent to a disjoint
union of k 2 regular tetrahedra, not necessarily of the same size.
Let us emphasize that the dihedral angles are all equal in this case, so one needs
to find other tools to prove this result. Let us note that Dehns theorem combined
with the above Sydlers theorem implies that the cube, the regular tetrahedron and
the regular triangular bipyramid (see Figure 19.1) of the same volume are not scissor
congruent. In other words, there are at least three equivalence classes of polytopes
of the same volume. In Exercise 16.1 we will show that there is a continuum of such
equivalence classes.
We first develop and prove several technical algebraic tools and only then return
to further examples and applications.
16.2. Rectifiable polytopes and operations on polytopes. We say that a polytope P R3 is rectifiable if it is scissor congruent to a cube of equal volume. For
example, all parallelepipeds are rectifiable by Exercise 15.4. Denote by P Q the
disjoint union of polytopes P, Q R3 . One can show that the disjoint union of rectifiable polytopes is also rectifiable (see Exercise 15.4). Let n!!P = P . . . P (n
times) denotes the disjoint union of n copies of P .
Polytopes cP obtained by an expansion of P by a factor c > 0, are called similar
to P . Obviously, polytopes similar to a rectifiable polytope are also rectifiable. Denote
by R the set of rectifiable polytopes.
37The

proof of Bricards condition given in the previous section is somewhat more revealing, as
it is crucially based on the edge lengths.

149

We say that a polytope P R3 is self-similar if it is scissor congruent to a disjoint

union of two or more polytopes similar to P . For the rest of this section we use to
denote the disjoint union of polytopes, and write this as
P c1 P . . . ck P , where ci > 0, k 2.

The main result of this section is the following theorem.

Theorem 16.2 (Sydlers criteria). A polytope P R3 is rectifiable if and only if P

is self-similar. Alternatively, P is rectifiable if and only if P cP R, for some
R R and c < 1.
Sydlers theorem (Theorem 16.1) follows immediately from here. Indeed, if a regular
tetrahedron is self-similar, then by the above theorem it must be rectifiable, which
contradicts Dehns theorem (Theorem 15.2).
Before we move to the proof of Sydlers criteria, let us present two other useful
properties of scissor congruence. They can be viewed as saying that rectifiability of
polytopes is invariant under subtraction and division. As we show later in this
section, these properties allow a remarkable degree of flexibility in deciding whether
a given polytope is rectifiable.
Theorem 16.3 (Complementarity lemma). Suppose polytopes A, B, C, D R3 satisfy: A B C D and B D. Then A C.
In particular, the complement of a rectifiable polytope inside a brick is also rectifiable.
Theorem 16.4 (Tiling lemma). Let P R3 be a polytope such that P1 . . . Pm is
rectifiable, where every Pi is either congruent to P or a mirror image of P , 1 i m.
Then P is also rectifiable.
In other words, if we can tile a parallelepiped with copies of a given polytope, then
this polytope is rectifiable. We say that a tiling of the space R3 is periodic if there
exist three independent vectors such that the tiling is invariant under translations by
these vectors. Obviously, if a polytope tiles a parallelepiped, it has a periodic tiling
of the space.
Corollary 16.5. Let P R3 be a polytope which tiles the space periodically. Then P
is rectifiable.
Proof. Fix a periodic tiling of R3 by P and let Q be a fundamental parallelepiped of
the translation group which preserves the tiling. By Lemma 15.3, parallelepiped Q
is scissor congruent to a disjoint union of copies of P and mirror symmetric copies
of P . Since Q R, by the tiling lemma (Theorem 16.4) so is P .

16.3. Proofs of the theorems. We start with an important technical result.
Lemma 16.6. Let P R3 be a polytope and let 1 , . . . , k > 0 be fixed positive real
numbers, such that 1 + . . . + k = 1. Then
P 1 P . . . k P R,

150

for some R R.
Proof. First, let us show that it suffices to prove the result for the tetrahedra. Indeed,
suppose P = m
i=1 i and
i 1 i . . . k i Ri for all 1 i m.

Then
P

" m
k
M
M
j=1

i=1

j i

" m
M

Ri

i=1

1 P . . . k P R,

where we are using the fact R1 . . . Rm R (see Exercise 15.4).

Let us prove the result in a special case. Let R3 be a standard tetrahedron
defined by the equations x, y, z 0 and x + y + z 1. Cut the tetrahedron into layers
0 z 1 , 1 z 1 + 2 , etc. Remove the tetrahedra i i , 1 i k, from
the corners of the corresponding layers, as shown in Figure 16.1. Denote by Bi the
remaining hexahedron (polytope with 6 faces) in the i-th layer, for all 1 i k 1
(the k-th layer is equal to k ). Observe that each Bi is scissor congruent to a brick
(see Figure 16.1). By Exercise 15.4, this implies that B1 B2 . . .Bk1 is rectifiable,
and finishes the proof.
4

B3

B1

2
1

Bi

B2

Qi

Figure 16.1. Layers Bi i and scissor congruence Bi Qi .

For the general tetrahedra, use a construction obtained via a linear transformation.
The image of a brick is a parallelepiped which are rectifiable by Exercise 15.4.

Proof of Theorem 16.4. Since the mirror image of polytope P is scissor congruent to P
(see Exercise 15.3), it suffices to prove the result for congruent polytopes Pi P . The
theorem and Lemma 16.6 can be written as:
nP R1 ,

for some R1 , R2 R. We conclude:

as desired.

nP (nP ) R2 ,

nP (nP ) R2 R1 R2 R,

Proof of Theorem 16.3. First, observe that

vol(A) = vol(A B) vol(A) = vol(C D) vol(D) = vol(C).

151

Fix a large enough integer n and define

1
1
1
1
A, B = B, C = C, D = D .
n
n
n
n
By Lemma 16.6, we have:
A =

A (nA ) R1 , C (nC ) R2 ,
for some R1 , R2 R. From above,


n
n
vol(R1 ) = 1 3 vol(A) = 1 3 vol(C) = vol(R2 ).
n
n
This implies that R1 R2 and that vol(R1 ) = (1 1/n2 )vol(A) vol(A) as n .
Therefore, vol(n B )/vol(A) 0 as n , and for sufficiently large n we can
arrange n copies of B inside R1 .38 In other words, for large enough n, we have
R (nB ) S ,

for some polytope S R3 . We then have:

A (nA ) R1 (nA ) (nB ) S n(A B ) S

n(C D ) S (nC ) (nD ) S

(nC ) (nB ) S (nC ) R1 (nC ) R2 C ,

as desired.

Proof of Theorem 16.2. The only if part is trivial in both parts. For the second
criterion, by Lemma 16.6, we have:
cP R P cP (1 c)P R ,

for some R, R R. By the complementarity lemma (Theorem 16.3), we get (1

c)P R R. Write R R R for some R R. After another application of the
complementarity lemma, we obtain (1 c)P R, as desired.
For the first part, P c1 P . . . ck P and let c = c1 + . . . + ck . By Lemma 16.6,
we have:
cP c1 P . . . ck P R P R,
for some R R. Now the second criterion we just proved implies that P R.

16.4. Examples and special cases. Let O = (0, 0, 0), a = (1, 0, 0), b = (0, 1, 0), c =
(0, 0, 1), and d = (1, 1, 0). Consider the standard tetrahedron 1 = conv{O, a, b, c}
and the Hill tetrahedron 2 = conv{O, b, c, d}. Clearly, vol(1 ) = vol(2 ) = 1/6.
Proposition 16.7. The standard tetrahedron, Hill tetrahedron and the regular tetrahedron of equal volume are not scissor congruent.
38Formally

speaking, this still requires an argument (see Exercise 15.1).

152

Proof. First, let us show that 2 is rectifiable. To see this, observe that six copies
of 2 tile the unit cube C = [0, 1]3 . Indeed, take each copy to be a convex hull of a
path from O to (1, 1, 1) as shown in Figure 16.2 (there are six such paths), and use
the tiling lemma. By Dehns theorem (Theorem 15.2), this implies that 0 2 ,
where 0 is the regular tetrahedron of volume 1/6.

c0

Figure 16.2. Hill tetrahedron, standard tetrahedron and the regular tetrahedron.
To see that 1 is not rectifiable, remove four copies of 1 from the
cube C as
shown in Figure 16.2. We are left with a regular tetrahedron c0 , c = 3 2, which by
Dehns theorem is not rectifiable. By the complementarity lemma, we conclude that
the above union of four copies of 1 is not rectifiable, and by the tiling lemma so
is 1 . This implies that 1 2 .
Finally, the same construction shows that 0 1 . Otherwise, four copies of 0
and one copy of c1 are scissor congruent to a cube, which contradicts the first
Sydlers criterion (Theorem 16.2). This completes the proof of Proposition 16.7. 
By analogy with the case of a regular tetrahedron, one can use Lemma 15.4 to
prove directly that the regular octahedron is not rectifiable. In fact, one can even use
the same proof of the irrationality of dihedral angles (see Subsection 41.2). However,
if one wants to avoid calculating dihedral angles this time, here is a neat geometric
round about argument.
Proposition 16.8. The regular octahedron is not scissor congruent to the standard,
Hill, or regular tetrahedra of the same volume.
Proof. First, observe that the regular octahedron Q can be tiled with eight copies of
a standard tetrahedron. By the tiling lemma this immediately immediately implies
that Q is not rectifiable, and by the first Sydlers criterion this implies that Q is not
scissor congruent to a standard tetrahedron. We need a separate argument to show
that Q c0 .
Take a regular tetrahedron 0 and remove four corner tetrahedra 0 = 21 0 . We
are left with a regular octahedron Q (see Figure 16.3). Take a disjoint union of Q
with any two of tetrahedra 0 . Now, if Q c0 , we have:
0 Q 40 c0 40 .

This again contradicts the first Sydlers criterion (Theorem 16.2).

153

Figure 16.3. Octahedron Q and four tetrahedra 0 tile tetrahedron 0 .

16.5. What if you had a magic wand to make things bigger and smaller?
To conclude this section, we consider the following question: suppose in addition to
translation and rotations in R3 we are also allowed to inflate and deflate the polytope.
Perhaps surprisingly, one can transform every polytope into every other polytope in
this case, in stark contrast with Exercise 16.1.
Formally, we say that polytopes P and Q are -congruent, write P Q, if each
polytope is a disjoint union of similar polytopes:
P = m
i=1 Pi ,

Q = m
i=1 Qi ,

where Pi cQi , for some c > 0. Of course, all scissor congruent polytopes are also
-congruent, but the inverse is not true even if the polytopes have the same volume.
Theorem 16.9 (Zylev). Every two polytopes P, Q R3 are -congruent: P Q.
Proof. First, note that -congruence is an equivalence relation, so it suffices to prove
that every polytope P is -congruent to a unit cube C. Second, it is easy to see
that the complementarity lemma (Theorem 16.3) holds for as well, and the proof
extends verbatim. Now use Lemma 16.6 to obtain
3 R1 2 R2 ,

for some R1 , R2 R. Now the complementarity lemma implies the result.

16.6. Exercises.
Exercise

 16.1. (Continuum of scissor congruence equivalence classes) [1] For every
0, 12 dene a -truncated cube Q() R3 by the inequalities |x|, |y|, |z| 1, |x| + |y| +
|z| 3 . For example, Q(0) is the unit cube, Q( 18 ) is the truncated cube, and Q( 12 )

is the cuboctahedron (see Figure 16.4). Prove that Q() cQ(), for all 0 <
and c > 0.

1
2

Exercise 16.2. a) [1-] Prove that the plane can be tiled with copies of a convex pentagon.
b) [1-] Prove that the space R3 can be tiled with copies of a convex heptahedron (polytope
P R3 with exactly 7 faces).
Exercise 16.3. a) [1+] Prove that the plane cannot be tiled with copies of a convex
octagon.
b) [1-] Show that this is possible when the octagon is non-convex.

154

Figure 16.4. Cube, truncated cube and cuboctahedron.

Exercise 16.4. [2-] Prove that an unfortunate convex polytope cannot tile R3 .
Exercise 16.5. a) [1+] Find a convex polytope in R3 which tiles the space, but only
aperiodically.
b) [2-] Find a polygon in the plane which tiles the plane, but does not have a tiling with a
transitive symmetry group.
c) [] Find a polygon in the plane which tiles the plane, but only aperiodically.
Exercise 16.6. (Permutohedron) Let Pn be a permutohedron dened in Example 8.4.
a) [1-] Prove that Pn is vertex-transitive and all its 2-dimensional faces are squares or regular
hexagons. Conclude that P4 is an Archimedean solid called the truncated octahedron.
b) [1-] Prove that Pn is inscribed into a sphere.
c) [1] Prove that Pn is a fundamental region of a discrete group acting on Rn1 , and
generated by reections.
d) [1-] Prove that P4 tiles the space R3 periodically. Conclude from here that P4 is rectiable.
e) [1] Prove that Pn is a zonotope (see Exercises 7.16 and 15.6). Conclude from here that Pn
is rectiable.
f ) [1+] Subdivide Pn into nn2 parallelepipeds of equal volume, whose edges are parallel to
the edges of Pn (use Cayleys formula). Use this to compute the volume of Pn .
Exercise 16.7. Let Q R3 be a (non-regular) octahedron dened as a convex hull of
points (1, 0, 0), (0, 1, 0), and (0, 0, 12 ).
a) [1-] Prove that Q tiles R3 periodically. Deduce from here that Q is rectiable.
b) [1-] Prove that Q can be tiled with sixteen copies of Hill tetrahedron. Deduce from here
that Q is rectiable.
Exercise 16.8. (F
oppl polytope) Consider the truncated tetrahedron P , dened as an
Archimedean solid with four triangular and four hexagonal faces (see Figure 39.3).
a) [1-] Prove that two copies of P and two regular tetrahedra tile a parallelepiped.
b) [1] Cut into four triangular pyramids and attach them to triangular faces of P . Prove
that the resulting polytope Q tiles the space.
c) [1-] Conclude that Q is rectiable.
Exercise 16.9. Consider the truncated cuboctahedron P , dened as an Archimedean solid
with a square, hexagonal and octagonal faces meeting at each vertex.
a) [1] Prove that copies of P and 8-prisms tile the space.
b) [1-] Use a) to prove that P is rectiable. Alternatively, check that P is a zonotope and
conclude that P is rectiable (see Exercise 15.6).
c) [1] Find a periodic tiling of R3 with copies of P , cubes and permutohedra (see Example 8.4).

155

d) [1] Find a periodic tiling of the space with copies of P , truncated cube (see Subsection 16.4) and truncated tetrahedron (see Exercise 16.8).
Exercise 16.10. [1] Give an explicit construction of scissor congruence of the Hill tetrahedron 2 and a prism.
Exercise 16.11. (Rectiable tetrahedra) Prove that the following tetrahedra (xyzw) R3
are rectiable:
a) [1-] x = O, y = (1, 0, 0), z = (1, 1, 1), w = (1, 1, 1);
b) [1-] x = O, y = (1, 1, 1), z = (1, 1, 1), w = (1, 1, 1);
c) [1-] x = O, y = (2, 0, 0), z = (1, 1, 1), w = (1, 1, 1);
d) [1-] x = O, y = (1, 0, 1/2), z = (1, 1, 1), w = (1,
1, 1);
e) [1] |xz| = |yw| = 2, |xy| = |xw| = |yz| = |zw| = 3;
f ) [1+] |xy| = |zw| = |yz| = a, |xz| = |yw| = b, |zw| = 3c, where b2 = a2 + 3c2 .

Exercise 16.12. (Hill tetrahedra) For every (0, 2/3), let v 1 , v 2 , v 3 R3 be three
unit vectors with angle between every two of them. Dene the Hill tetrahedron Q() as
follows:
Q() = {c1 v + c2 v 2 + c3 v 3 | 0 c1 c2 c3 1}.
For example, Q(/2) is the Hill tetrahedron 2 dened earlier.
a) [1] Decompose Q() into three pieces which assemble into a prism. Conclude that all
Hill tetrahedra are rectiable.
b) [1+] Find a periodic tiling of space with Q(). Again, conclude that all Hill tetrahedra
are rectiable.
Exercise 16.13. (Golden tetrahedron)
by

Denote by T R3 the golden tetrahedron dened

1+ 5
z
1, where =
.
x, y, z 0, x + y +

2
a) [1] Prove that all dihedral angles of T are rational multiples of . Does T tile the space?
b) [1+] Prove that T is rectiable by an explicit construction.
c) [1+] Prove that the Hill tetrahedron Q(2/5) can be decomposed into four golden tetrahedra (possibly, of dierent size).
d) [1+] Generalize the tiling lemma (Theorem 16.4) to similar polytopes. Use this and
part c) to give another proof that T is rectiable.

Exercise 16.14. (Schl

ai simplices) For a simplex = (v1 v2 . . . vn+1 ) Rn denote
by Fi the facet containing all vertices except vi . Let G be a graph with the vertices
{1, 2, . . . , n + 1} and edges (i, j) corresponding to acute dihedral angles between facets Fi
and Fj .
a) [1+] Prove that G contains a spanning tree. Conclude that has at least n acute
dihedral angles.
b) [1] Show that
for every spanning tree T in a complete graph Kn+1 , there exists a simplex
Rn with n2 right dihedral angles and n acute dihedral angles corresponding to this
spanning tree: G = T . Such are called Schl
ai simplices.39
c) [1] Show that every Schl
ai simplex corresponding to tree T is a convex hull of mutually
orthogonal intervals which form a tree isomorphic to T .
d) [1] Denote by ij the length of the edge (vi vj ), and by ai the length of the altitude in
from vertex vi . Prove that cos ij = ai aj /ij , where ij is the dihedral angle at (vi vj ).
39H.

S. M. Coxeter calls these simplices orthogonal trees.

156

e) [1] Prove that

X 1
1
=
.
a2i
2ij
(i,j)T

Deduce from here a formula for

cos2 ij

as a polynomial in the squared edge lengths 2ij .

Exercise 16.15. a) [1-] Prove that there exists a unique circumscribed brick B around
every Schl
ai simplex , so that all vertices of are also vertices of B.
b) [1-] Prove that the midpoint of the longest edge of a Schl
ai simplex is the center of a
circumscribed sphere.
c) [1] Prove that all faces of Schl
ai simplices (of all dimensions) are also Schl
ai simplices.
d) [1+] Prove that every simplex all of whose 2-dimensional faces are all right triangles must
be a Schl
ai simplex. Deduce from here part c).
Exercise 16.16. (Orthoschemes) [1] Consider Schl
ai simplices, corresponding to paths.
These simplices are called orthoschemes.40 These can be dened as convex hulls of pairwise
orthogonal intervals in Rd forming a path. Prove that every orthoscheme in Rd can be
dissected into d + 1 orthoschemes.
Exercise 16.17. (Coxeter simplices) Consider Schl
ai simplices corresponding to trees in
Figure 16.5 with interval lengths written next to the edges.
a) [1-] Use parts c) and d) of Exercise 16.14 to compute all dihedral angles of these Schl
ai
simplices.
b) [1+] Prove that these simplices are fundamental regions of the natural action by the
corresponding ane Coxeter groups.41
c) [1-] Conclude that these simplices are rectiable.
q

1
2

e6
E

1
3

q
1

1
3

1
3

1
2

1
e7
E

e8
E

1
2

1
3

1
3

1
2

1
1

1
6

1
6

en
D

1
6

1
2

1
10

1
3

1
15

1
12

1
3

Figure 16.5. Schlafli simplices corresponding to affine Coxeter groups.

40In

computer science literature, these are called path-simplices.

is a subclass of discrete groups generated by reections: http://tinyurl.com/26fbqeu

41There

157

Exercise 16.18. (Sommerville simplices) [2] Classify all tetrahedra R3 which tile the
space face-to-face (mirror symmetry of is not allowed).
b) [] Same problem, but now mirror symmetry is not allowed.
Exercise 16.19. [1+] Let P be Jessens orthogonal icosahedron dened in part b) of Exercise 19.17. Prove that P is rectiable.
Exercise 16.20. a) [1] Use Bricards condition (Lemma 15.4) to show that no two Platonic
solids are scissor congruent.
b) [2-] Denote by , Q, I and D the regular tetrahedron, octahedron, icosahedron and
dodecahedron with unit edge lengths. Find all mi 0, such that (m1 ) (m2 Q)
(m3 I) (m4 D) is rectiable.
c) [2-] Denote by W the icosidodecahedron with unit edge length, an Archimedean solid
where two pentagons and two triangles meet at every vertex. Prove that I D W is
rectiable.
Exercise 16.21. [1] Prove that the regular cross-polytope Q in R4 can tile the whole space.
Deduce from here that Q is rectiable.
Exercise 16.22.
a) [1-] Let , Q R8 be a regular simplex and a cross-polytope with

edge lengths 2. Prove that

1920 vol() + 135 vol(Q) = 1.

b) [1+] Extend the above equation by showing that 1920 135Q is scissor congruent
to a unit cube.
c) [2-] Prove that copies of and Q can periodically tile the whole space R8 . Deduce from
here part b).
d) [2-] Decide in what other dimensions there are analogues of a) and b).
Exercise 16.23. a) [1+] Find an explicit decomposition of a regular tetrahedron into
polytopes such that similar polytopes can be arranged into a cube, i.e., prove 0 C
directly.
b) [2] Suppose only translations and no rotations are allowed between similar polytopes in
the denition of -congruence (cf. Exercise 15.7). Extend Theorem 16.9 to this case.
b) [2-] Generalize Theorem 16.9 to higher dimensions.
Exercise 16.24. (Sydler ) [2] Let P R3 be a tetrahedron with all dihedral angles rational
multiples of . Prove that P is rectiable.
Exercise 16.25. (Jessen) [2] Let P be a polytope in R4 . Prove that there exists a polytope
Q R3 such that P Q [0, 1].
16.7. Final remarks. Most results and many proofs in this section go back to Sydlers
original paper [Syd1]. In particular, the crucial idea to use Lemma 16.6 to prove the
complementarity lemma is also due to Sydler. Theorem 16.9 was proved by in [Zyl] (see
also [Had2]). Exercises 16.24 and 16.25 are partial results in the SydlerJessens theory.
We refer to [Bolt] for complete proofs, generalizations and references.
Let us mention here several connections of scissor congruence and tileability of the space
by copies of the same polytope. Already in the plane there are many examples of such
polygons, including some interesting pentagons, which tile plane. Now, if polytope P R3
can tile the space periodically, then P is rectiable (Corollary 16.5). Similarly, if P admits

158

a substitution tiling of the space (see Exercise 15.12), then P is also rectiable (this follows
immediately from Theorem 16.2). The most general result was obtained by Debrunner
(see [LM]), who showed that every polytope which tiles R3 is rectiable. On the other
hand, Conway has a simple construction of a polytope which has only aperiodic tilings
of the space (Exercise 16.5). Unfortunately, the proof by Debrunner is indirect and uses
Sydlers theorem on Dehn invariants (Theorem 17.7). Exercise 16.4 gives an idea of the
proof in a special case. We refer to [Sene] for an accessible introduction to the tiling of the
space by tetrahedra and general convex polyhedra (cf. Exercises 16.7, 16.11 and 16.6).

Figure 16.6. Generalized -congruences between a square and a circle.

Finally, in an eort to extend the notion of scissor congruence to general regions one has
to be careful not to over-extend. Not only do there exist various BanachTarski paradox
type results, but even with a restriction to almost nice regions, just about everything is
scissor congruent. In Figure 16.6 we show two ways of squaring the circle, both using
homotheties, one with innitely many piecewise smooth regions (or two disconnected regions), and another with just four simply connected regions. We refer to [HerR, KleR] for
more on these results and references.

159

17. Dissections and valuations

This is the last of three sections on scissor congruence, where we finally define the
Dehn invariants mentioned in Section 15. We begin by showing that all dissections
of a polytope are connected by certain local moves. These are the smallest possible
transformations between dissections, switching between one and two simplices. The
idea of local moves is to be able to prove various properties of all dissections by
checking that the properties are preserved under a local move. This is now a standard
approach in discrete geometry, which we repeatedly use in Section 23 as well as in
various other examples and exercises.
17.1. Slicing and dicing the polytopes. In this section we consider dissections of
convex polytopes, defined as simplicial decompositions. It is important to distinguish
dissections from triangulations defined as simplicial subdivisions and considered earlier in Section 2. Clearly, every triangulation is a dissection and every polytope has
infinitely many dissections.
Let P Rd be a convex polytope and let D be its dissection, write D P .
Define local moves on dissections as follows. We say D1 , D2 P are connected by an
elementary move, write D1 D2 , if they coincide everywhere except at one simplex
in D1 which is subdivided into two in D2 , or if one simplex in D2 which is subdivided
into two in D1 . In other words, one can cut one of the tetrahedra with a hyperplane
through all but two vertices (the only way to decompose a tetrahedron into two) or
do the inverse operation (see Figure 17.1).

Figure 17.1. Examples of elementary moves on dissections.

We say that two dissections D, D P are elementary move equivalent, write D
D , if there exists a sequence D1 , . . . , D of dissections of P such that
D D1 D2 . . . D D .

In other words, the elementary move equivalent dissections are those connected by a
finite sequence of elementary moves.
D

Figure 17.2. A sequence of elementary moves on dissections: D D .

160

Theorem 17.1 (Elementary move connectivity). Every two dissections D, D P of

a convex polytope P Rd are elementary move equivalent: D D .
We prove the theorem only for d = 2, 3, leaving the higher dimensional cases to
the reader (see Exercise 17.1). The case of space dissections (d = 3) is particularly
important in connection with scissor congruence.
17.2. Symmetric valuations. The connection between Theorem 17.1 and scissor
congruencemay
not be immediately apparent, so let us start with some definitions.
Let : R be a map from all tetrahedra Rd into real numbers. We
say that is a valuation if it is invariant under the elementary moves: (1 2 ) =
(1 ) + (2 ). Valuation is called symmetric if it is invariant under the rigid
motions (rotations and translations). For example, vol() is a natural symmetric
valuation. An extension of a valuation is a function : {P } R on all convex
polytopes P Rd , which satisfies
X
(P ) =
(), for all D P and P Rd .
D

Here and everywhere below we denote the extension also by . We first need to show
that the extensions always exist.

Lemma 17.2 (Valuation extension lemma). Every symmetric valuation on simplices in Rd has a unique extension to all convex polytopes P Rd . Moreover, this
extension is also symmetric, i.e., invariant under rigid motions.
Proof. For the existence, take any dissection D P (take, e.g., the triangulation constructed in Subsection 2.1). The uniqueness follows from the fact that the summation
in the definition is equal for all dissections D P . Indeed, by Theorem 17.1 every two
valuations are connected by elementary moves and the valuation is invariant under
elementary moves. Finally, the symmetry of the extension follows immediately from
the symmetry of .

Corollary 17.3. Let be a symmetric valuation on simplices in Rd and let P, Q Rd
be convex polytopes, such that (P ) 6= (Q). Then these two polytopes are not scissor
congruent: P Q.
Proof. To the contrary, suppose P Q. Then there exist a decomposition of both
polytopes into the same set of simplices:
But then

P = m
i=1 i ,

Q = m
i=1 i ,

where i i .

(P ) = (1 ) + . . . + (m ) = (1 ) + . . . + (m ) = (Q),
a contradiction.

For example, when is the volume and vol(P ) 6= vol(Q), we trivially have P Q.
In the next section we give some further examples of symmetric valuations.

161

17.3. Dehn invariants. We say that f : R R is an additive function if f (a + b) =

f (a) + f (b). Define the Dehn invariant to be the following function on tetrahedra
R3 :
X
() =
e f (e ) ,
e

where e is the dihedral angle at edge e , e is the length of e, and f : R R

is an additive function which satisfies f () = 0. The following two results show that
Dehn invariants can be defined on all convex polytopes P R3 and give an explicit
formula for (P ).
Lemma 17.4. Every Dehn invariant on tetrahedra in R3 is a symmetric valuation.

Proof. The symmetry follows immediately by the definition of Dehn invariants. Consider now a 2-move 1 2 . By definition, two new edges e, e are created with
corresponding dihedral angles , , and (see Figure 17.3). Therefore,
() (1 ) (2 ) = e f () + e f ( ) + e f () + e f ( )




= e f () + f ( ) + e f () + f ( ) = e f () + e f () = 0,

which implies that is a valuation.

e
e

2
1

Figure 17.3. Decomposition of a tetrahedron = 1 2 .

Lemma 17.5. The Dehn invariant defined as above has the following extension to
all convex polytopes:
X
(P ) =
e f (e ) ,
eP

where e is the dihedral angle at edge e P .

Proof. Fix a triangulation P = m

i=1 i . As in the proof of the lemma above or in the
proof of Bricards condition for subdivisions (Lemma 15.4), for every interior edge e,
the sum of dihedral angles at e is equal to 2. Similarly, for an edge inside a face,
the sum of dihedral angles at e is equal to . We have then
m
m X
X
X
X
(P ) =
(i ) =
e f (e ) =
e f (e ),
i=1

as desired.

i=1 ei

eP

162

We can now summarize the results as follows.

Theorem 17.6 (Dehn). Define the Dehn invariant on all convex polytopes P R3
by the following formula:
X
(P ) =
e f (e ) ,
eP

where e is the dihedral angle at edge e P , e is the length of e, and where f : R R

is an additive function such that f () = 0. Suppose now P, Q R3 are two convex
polytopes such that (P ) 6= (Q). Then P and Q are not scissor congruent: P Q.
17.4. From abstraction to applications. The Dehn invariants defined above may
not seem like much. In fact, at first glance it might seem too abstract to be useful in
actual applications. Before giving several examples, let us state the key result in the
field.
Theorem 17.7 (Sydler). Every two convex polytopes P, Q R3 are scissor congruent
if and only if vol(P ) = vol(Q) and for every Dehn invariant as above we have
(P ) = (Q).
In other words, Sydlers theorem implies that whenever P Q, we should be able
to prove this by finding an appropriate Dehn invariant. Unfortunately the proof is a
bit too technical to be included. Here are two basic examples.
Example 17.8. Let P = R3 be the regular tetrahedron with unit edge length, and let
Q R3 be a cube of the same volume. Let = arccos 13 denotes the dihedral angle in .
Consider an additive function f : R R which satises
f (s) = s

and f ( s) = 0,

for all s Q.

/ Q as shown in Subsection 41.3.

It is easy to see that such functions exist,42 since /
Now, for the Dehn invariant we have:

(P ) = 6 1 () = 6 and (Q) = 12 c
= 0,
2
where c is the edge length of Q. Since (P ) 6= (Q), by Theorem 17.6 we conclude that P
and Q are not scissor congruent. This gives another proof of Dehns Theorem 15.2.
Example 17.9. As before, let P = R3 be the regular tetrahedron with unit edge
length, and let Q = c1 . . . ck , k 2, be a disjoint union of regular tetrahedra of
equal volume. Let be the Dehn invariant as in the previous example. We have
(Q) = 6 c1 () + . . . + 6 ck () = 6 (c1 + . . . + ck ).

Since k 2 and vol(P ) = vol(Q), we have c31 + . . . + c3k = 1 and 0 < ci < 1. From here,
c1 + . . . + ck > 1, and (P ) 6= (Q). Again, by Theorem 17.6 we conclude that P and Q
are not scissor congruent, which gives another proof of Sydlers Theorem 16.1.
42Well,

no, it is not that easy to see. To formalize this, one needs to use the transnite induction.
However, for our purposes, a weaker statement suces (see Exercise 17.3).

163

17.5. Dissecting the plane. The proof of Theorem 17.1 will occupy much of the
rest of this section. We start with polygons in the plane and then extend the proof
to higher dimensions. First, let us start with a few definitions.
A (full) triangulation T P of a convex polygon P is a triangulation of P with
vertices in the vertices of P . Define a star triangulation T P to be a triangulation where all triangles have a common vertex a. We call a the center of the star
triangulation.
Lemma 17.10. Every two star triangulations of a convex polygon are elementary
move equivalent.
Proof. First, let us show that every two triangulations of a convex polygon in the
plane are elementary move equivalent. Define a 2-move by choosing two adjacent
triangles , T and switching one diagonal in a convex quadrilateral
to another. Observe that all triangulations of a convex polygon are connected by
2-moves (see Theorem 14.1 and Remark 14.2). Since every 2-move can be obtained
by four elementary moves (see Figure 17.4), we conclude that star triangulations are
elementary move equivalent.


Figure 17.4. Making a 2-move using four elementary moves.

Let L be a line intersecting a convex polygon P and dividing it into two polygons P1
and P2 . For every two star triangulations D1 P1 and D2 P2 , their union D1 D2
is a dissection of D.
Lemma 17.11. The union of star triangulations D1 and D2 of convex polygons P1
and P2 , respectively, is elementary move equivalent to a star triangulation D of P =
P1 P2 .
Proof. Denote by a, b the points of intersection of the line L and the boundary of P .
By Lemma 17.10, triangulations D1 and D2 are elementary move equivalent to star
triangulations centered at a. If b is not a vertex, remove the edge (a, b) which separates
two triangles (see Figure 17.5). If a is a vertex, we are done. If it lies on edge e of P , we
can choose a vertex v of P not on e. For each of the polygons separated by (a, v), use
Lemma 17.10 connect their triangulations to star triangulations centered at v. Now
remove the (a, v) edge. This give a star triangulation centered at v (see Figure 17.5).
Again, by Lemma 17.10, this triangulation is elementary move equivalent to every
star triangulation D of the polygon P .

We are now ready to prove Theorem 17.1 in the plane.

Proposition 17.12. Theorem 17.1 holds for d = 2, i.e., every two dissections of a
convex polygon in the plane are elementary move equivalent.

164

L
a

Figure 17.5. From two star triangulations to one.

Proof. Consider a decomposition Q of P into convex polygons P = m
i=1 Pi . A dissecm
tion D is called a star refinement of Q, write D Q, if D = i=1 Di , where Di is a
star triangulation of Pi . We call Q reducible if a star refinement D Q is elementary
move equivalent to a star triangulation of P . By Lemma 17.10, every star refinement
D Q is elementary move equivalent to every star triangulation of P .
We prove by induction on the number m of polygons that every decomposition Q
as above is reducible. There is nothing to prove when m = 1. For m > 1, take
any interior edge e P and cut P by a line spanned by e. Denote by P, P the
resulting two polygons, and let Pi = Pi P , Pi = Pi P for all 1 i m. By
Lemma 17.11, star triangulations of Pi are elementary move equivalent to a union of
star triangulations of Pi and Pi . In other words, if P = Pi and P = Pi are
reducible decompositions, then so is Q.

Figure 17.6. Cutting a decomposition by a line in the induction step.

Now observe that by convexity, the polygons on both sides of edge e are not cut
by . Thus, both decompositions have fewer than m polygons, and the inductive
assumption applies. From above, the decomposition Q is reducible, which completes
the proof.


17.6. Dissecting the space. Here we extend the argument in the previous subsection to convex polytopes in R3 . The heart of the proof (the proof of Proposition 17.12)
extends nearly verbatim, while the straightforward proof of Lemma 17.10 extends only
to regular triangulations (see Subsection 14.10).
Lemma 17.13. Let D1 , D2 be dissections of a convex polygon P and let Q be a
cone over P . Denote by D1 , D2 the dissections of Q obtained as union of cones over
triangles in D1 , D2 . Then D1 D2 .

165

Proof. Taking a cone over an elementary move in P corresponds to an elementary

move in Q.

Let P R3 be a convex polytope and let V be the set of its vertices. As before, a
(full) triangulation of P is a triangulation with vertices in V (we will skip the adjective
full throughout the section whenever possible). A star triangulation centered at a
vertex a V is a triangulation of P with all simplices containing a and giving star
triangulations of all faces of P . Recall the definition of regular triangulations in
Subsection 14.5.
Lemma 17.14. Every star triangulation of a convex polytope in R3 is regular.
Proof. We need to show that every star triangulation D of P is associated to some
height function : V R+ . Suppose D is centered at a V . Let (a) = 1 and
(v) = 0 for all v 6= a. The subdivision associated with consists of cones over the
faces of P . By definition, the restriction of D to a face F gives a triangulation of F .
Fix > 0 very small. Consider a height function (a) = 1, (b) = , if b is a center of
a triangulation in some face of P , and (v) = 0 otherwise. When is small enough,
in the surface S every cone over a face of P is then triangulated according to D, as
desired.

Lemma 17.15. Every two star triangulations of a convex polytope in R3 are elementary move equivalent.
Proof. Let us first show that every 23 move (see Subsection 14.5) can be obtained
by elementary moves. For each of the two tetrahedra in a bipyramid we need to have
a dissection into three cones. By Lemma 17.13 and Proposition 17.12, we can make
these dissections using elementary moves (see also Figure 17.2). Now attach together
the corresponding three pairs of resulting tetrahedra.
Now, by Exercise 17.6 extending Theorem 14.10 to all convex polytopes in R3 , it
suffices to check that remaining stellar flips can be obtained by elementary moves.
The proof is along similar lines and is left as (an easy) part d) of that exercise.

To continue with the proof as in the case d = 2 we need an analogue of Lemma 17.11.
Let L be a plane intersecting a convex polytope P R3 , dividing it into two polytopes P1 and P2 . For every two star triangulations D1 P1 and D2 P2 , their union
D1 D2 is a dissection of D.
Lemma 17.16. The union of star triangulations D1 and D2 of convex polytopes P1
and P2 , respectively, is elementary move equivalent to a star triangulation D of P =
P1 P2 .
Proof. The idea is to combine the approach in the proof of Lemma 17.11 with the
height function technique as above. Denote by A the set of points of intersection of L
and edges in P . Fix a point a A, and denote by D the star triangulation of P with
vertices in V A, centered at a and such that the center of triangulation of every
face lies on L. By construction, the triangulation D is a union of star triangulations
of P1 and P2 , which are elementary move equivalent to D1 and D2 by Lemma 17.15.

166

Now, suppose L intersects face F P , a

/ F , by an edge (b, c). The cone from a
over F is a dissection, which, by Lemma 17.13, is elementary move equivalent to a
star triangulation with vertices in V A r {b, c} (see Figure 17.7). Repeating this
over every such face F we obtain a star triangulation D of P centered at a and with
vertices in V {a}.
a

Figure 17.7. star triangulation of a cone over F (view from the bottom).
Finally, let us use the height function approach to show that D is elementary move
equivalent to a star triangulation of P with vertices in V . Suppose a lies on the
edge (v, w). Take the height function : V {a} R+ defined in the proof of
Lemma 17.14. Recall that (a) = 1 and (v), (w) > 0 are small. Continue as in
the proof of Theorem 14.10. Start increasing the values of v and w generically, until
(v), (w) > 1, when a is no longer a vertex in a triangulation. Other than 23 moves,
at some point, when is linear on (v, w) a single elementary move will be used once,
when the height function is linear on (v, w). Note that the resulting triangulation
is full, but not necessarily a star triangulation. By Theorem 14.10, we can connect it
to a star triangulation by a sequence of 23 moves, which implies the result.

Proposition 17.17. Theorem 17.1 holds for d = 3, i.e., every two dissections of a
convex polytope in R3 are elementary move equivalent.
As we mentioned above, the proof follows from the lemmas above and the same
inductive argument.
Remark 17.18. As we mentioned in Remark 14.9, it is an open problem whether all
full triangulations of a convex polytope in R3 are connected by 23 moves [San2]. The
corresponding result in dimensions 5 and higher is false. On the other hand, the graph of
all regular triangulations is not only connected, but is a graph of a convex polytope (see
Example 8.5 for a special case).

17.7. Wait, there is more! Yes, indeed. Remember we only proved the elementary
move connectivity theorem (Theorem 17.1) in the plane and in 3-dimensional space.
In higher dimensions the proof follows roughly the same lines as long as one defines
star triangulation in higher dimension (see Exercise 17.1). As a consolation prize, we
prove that the following generalization follows easily from Theorem 17.1.

167

Theorem 17.19. For every (possibly non-convex) polytope P Rd , every two dissections D, D P are elementary move equivalent: D D .
Here by a polytope we mean any finite union of convex polytopes. Note that we
implicitly used convexity, for example, in the proof of the induction step in Proposition 17.12. Interestingly, the reduction to convex polytopes is now completely
straightforward.
Proof of Theorem 17.19. Take the intersection of the dissections (i.e., superimpose
them as decompositions). We obtain a decomposition into convex polytopes. Refine
this decomposition by triangulating each polytope and denote by D the resulting
dissection of P . Now observe that every simplex in D is connected by elementary
moves to its dissection in D . Therefore, D is elementary move equivalent to D , and
we conclude: D D D .

17.8. Exercises.
Exercise 17.1. [1+] Finish the proof of Theorem 17.1 by extending the proof in Subsection 17.6 to higher dimensions.
Exercise 17.2. [1-] Show that Exercise 16.24 is a simple corollary of Sydlers theorem
(Theorem 17.7). More generally, show that Sydlers theorem implies that all polytopes
with rational dihedral angles are rectiable.
Exercise 17.3. a) [1] Suppose and Q are as in Example 17.8. As in the example,
assume that Q and consider the decompositions into tetrahedra proving this. Check
that it suces dene the additive function f only on the (nitely many) dihedral angles
that appear in the decomposition and obtain a contradiction.
b) [1+] Prove existence of the additive functions f in Example 17.8.
Exercise 17.4. (Tverbergs theorem) Let P R3 be a convex polytope. Suppose we are
allowed to cut P with a plane. Two parts are then separated, and each is then allowed to
be separately cut with a new plane, etc.43
a) [1-] Prove that a regular octahedron can be cut into tetrahedral pieces with only three
cuts. Similarly, prove that for a cube four cuts suce.
b) [1-] Prove by an explicit construction that both regular icosahedron and regular dodecahedron need at most 100 cuts.
c) [1] Prove that every convex polytope in R3 can be cut into tetrahedral pieces with nitely
many cuts.
d) [1+] Generalize this to Rd .
Exercise 17.5. a) [1-] Let Q = A be a simple polygon in R2 . Prove that every
triangulation of A contains a triangle T with two sides in Q, i.e., such that A r T is
homeomorphic to a disk.
b) [1+] In the plane, prove that for every decomposition Q = i Qi of a simply connected
polygon Q into simply connected polygons Qi , there exist Qi such that Qr Qi is also simply
connected.
c) [1+] Show that part b) does not generalize to R3 .
43Think

of chopping polytopes as if they were vegetables.

168

d) [2] Find a simplicial subdivision D of a tetrahedron , such that for every tetrahedron T
in D, the closure of r T is not homeomorphic to .
Exercise 17.6. (Stellar ips) Consider the following ips on triangulations (simplicial
subdivisions) of a polytope in R3 :
the 23 move dened above,
the 14 move dened as a subdivision of a tetrahedron into four smaller tetrahedra,
the 48 move dened as a ip from one triangulations of an octahedron to another,
as shown in Figure 17.8,
the 13 move, a degenerate case of the 14 move when the new point is on the boundary,
the 35 move, a degenerate case of the 48 move when the new point is on the diagonal
separating two coplanar triangles.
a) [2] Prove that every two triangulations of a convex polytope P R3 are connected by
these ips.
b) [2] Generalize this to non-convex polytopes in R3 .
c) [2+] Generalize this to higher dimensions.
d) [1-] Check that all stellar ips can be obtained by elementary moves.

Figure 17.8. A 48 move on triangulations of an octahedron.

17.9. Final remarks. The study of valuations has a long history and a number of applications which go outside of the scissor congruence. We changed some of the denitions to
streamline the connections and at the expense of generality. Traditionally, valuations are
dened as maps : {P } G, where P Rd are convex polytopes, G is an abelian group
(say, R by addition), and such that
(P ) + (Q) = (P Q) + (P Q),

for all P, Q Rd .

Under this denition one needs to be careful with the boundary of polytopes and dene the
valuations on degenerate polytopes, which we largely ignore for simplicity. The advantage is
the wealth of other examples of valuations, such as the surface area, the Euler characteristic,
and the mean curvature (see Section 28). We refer to [Grub, 7] for these and further results,
and to [McM2] for a detailed survey and references.
The study of local moves on triangulations is also classical and in the context of combinatorial topology goes back to J. W. Alexander (1930) and M. H. A. Newman (1926, 1931).
Here one aims to obtain a topological invariant of a manifold using simplicial subdivisions.
Thus, one wants to make sure they are indeed invariant under certain local moves. The
result then follows from the connectivity of all triangulations under such local moves. We
refer to [Lic] for the survey and references.
The geometric study of dissections and triangulations is more recent, more delicate and
has a number of negative results and open problems (see Remark 17.18). Theorem 17.19
in this form is given in [LudR]. Although our proof uses several dierent ideas, such as

169

height functions, the proofs share a similar blueprint. The triangulation counterpart of the
elementary move connectivity theorem is also known, and uses the so-called stellar ips
or Pachner moves. These ips play an important role in both geometric combinatorics
and algebraic geometry. While many natural questions turn out to have negative answers,
an important positive result (see Exercise 17.6) was established by Morelli (1996) and
Wlodarczyk (1997). We refer to [San2] for a survey of these result and to [IzmS] for an
elegant presentation of the MorelliWlodarczyk theorem.

170

18. Monge problem for polytopes

This short section presents a little known variation on the scissor congruence problem. Although we do not apply the main theorem elsewhere, it has a pleasant flavor
of universality, while the proof uses a version of the local move connectivity, fitting
well with several previous sections.
18.1. Stretching the polytopes. Let P, Q Rd be two polytopes of equal volume:
vol(P ) = vol(Q). We say that map : P Q is a Monge map if it is continuous,
piecewise linear (PL), and volume-preserving. Two polytopes P and Q are called
Monge equivalent, written P Q, if there exists a Monge map : P Q. Observe
that is an equivalence relation since the composition of Monge maps, and the
inverse of a Monge map are again Monge map.
Here is another way to think of the Monge equivalence: there must exist two
triangulations (simplicial subdivisions) of polytopes P = ni=1 Pi and Q = ni=1 Qi
and a continuous map : P Q, such that (Pi ) = Qi , the map is linear on
each Pi , and vol(Pi ) = vol(Qi ), for all 1 i n. In this setting Monge equivalence is
somewhat similar to scissor congruence: we no longer require the map : Pi Qi to
be orthogonal, just volume-preserving.44 If that was all, establishing the equivalence
would be easy (see below), but the added continuity requirement makes matters more
complicated. Here is the main result of this section.
Theorem 18.1. Every two convex polytopes in Rd of equal volume are Monge equivalent.
As the reader shall see, the result is nontrivial already in the plane. The reader
might want to look for an independent proof in this case before proceeding to the
next section. Let us mention here that the theorem has a number of far-reaching
generalizations, for example to polygons in the plane with the same area and the
same number of holes (see Exercise 18.2).
18.2. Pre-proof analysis. Before we present the proof of the theorem, let us first
consider several special cases and weak versions of the theorem.
Example 18.2. (Monge equivalence in the plane) Let P, Q R2 be two convex polygons
in the plane of equal area: area(P ) = area(Q). Take an n-gon P and any vertex v in P .
Consider a triangle formed by v and its neighbors u and w. Let z be the second neighboring
vertex of w (other than v). Now transform the triangle (uvw) into (uv w) by shifting v
along the line parallel to (uw), such that v now lies on a line containing (wz). Keep the
rest of the polygon unchanged. We obtain a convex (n 1)-gon P (see Figure 18.1).
Proceed in this manner until P is mapped into a triangle: 1 : P 1 . Proceed
similarly with the polygon Q to obtain 2 : Q 2 . Since the resulting triangles 1 , 2
have the same area, there exist a volume-preserving linear map : 1 2 . We conclude:
P 1 2 Q.
44In

other words, we replace the smaller group SO(d, R) of rigid motions of simplices, with a
bigger group SL(d, R).

171

Figure 18.1. Monge map : P P .

Example 18.3. (Monge equivalence without the continuity condition) Let us weaken the
Monge restriction on a map in the theorem by removing the continuity condition. In other
words, we prove that between every two polytopes P, Q Rd of equal volume there exists
a volume-preserving PL-map : P Q.
Without loss of generality we can assume that vol(P ) = vol(Q) = 1. Consider any two
n
simplicial subdivisions P = m
i=1 Pi , Q = j=1 Qj , and let i = vol(Pi ), j = vol(Qj ). Subdivide further each of these simplices into smaller simplices: Pi = nj=1 Pij , Qj = m
i=1 Qij , such
that vol(Pij ) = vol(Qij ) = i j . Since every simplex Pij can be mapped into simplex Qij
by a volume-preserving map, this implies the claim.
Example 18.4. (Monge equivalence without the volume-preserving condition) Let us remove the volume-preserving restriction in the Monge map, and prove a weaker version of
the theorem. In other words, let us show that between every two polytopes P, Q Rd of
equal volume there exists a continuous PL-map : P Q.
We can assume that polytopes P, Q are simplicial; otherwise subdivide each facet into
simplices. Move the polytopes so that the origin O Rd lies in the interior of both polytopes: O P, Q.
d
Now, consider a simplicial fan F = m
i=1 Fi R dened as a union of innite cones Fi
which have a vertex at O and are spanned over the facets in P . Similarly, consider a
fan G = nj=1 Gi Rd over facet simplices of Q. Let C be the union fan which consists
er a subdivision of C into simple cones (see
e = r C
of cones Fi Gj , and denote by C
Exercise 2.13). Finally, dene simplicial subdivisions P = r Pr , Q = r Qr by intersecting
e with polytopes P and Q: Pr = P C
er , Qr = Q C
er (see Figure 18.2).
the fan C
P9

O
P

O
Q

Q1

P1

P2

P8

P3
P4

P7
P6 P5

Q2

Q9

Q8

Q3
Q7

Q4
Q6 Q5

e
Figure 18.2. Polygons P, Q with fans F, G, the union fan C = C,
and the continuous PL-map : P Q.

Fix r and consider the simplices Pr and Qr . Denote by v1 , . . . , vd and w1 , . . . , wd their

vertices other than O. From above, wi = i vi for some 1 , . . . , d > 0. Dene a PL-map
: P Q to be linear on each Pr and map each Pr into Qr :
X
X
:
xi vi 7
xi wi , where xi 0.

172

The map is clearly continuous and piecewise linear, which proves the claim.
Example 18.5. (Monge equivalence for bipyramids) While Theorem 18.1 is trivial for
simplices, already for bipyramids it is less obvious, so it makes sense to start by proving the
result in this case.
Formally, let P, Q Rd be two d-dimensional bipyramids of equal volume: vol(P ) =
vol(Q). Let u1 , u2 , x1 , . . . , xd and v1 , v2 , y1 , . . . , yd be the vertices of P and Q respectively,
where u1 , u2 and v1 , v2 are the simple vertices. Let us prove that there exists a continuous
volume-preserving PL-map : P Q which is linear on each facet and which sends ui
to vi and xj to yj , for all i = 1, 2 and 1 j d.
Let us start with a volume-preserving ane transformation of Rd which maps the vertices
x1 , . . . , xd into the vertices x1 , . . . , xd of a regular (d 1)-dimensional simplex S. Denote
by z the barycenter of S, and by the line going through z and orthogonal to S. Let u1
and u2 be the orthogonal projections of u1 and u2 onto .
u1

A3
x2

x1
z
x3

B3

A2

y1

y2
y3

e
A1

v1

B1

B2

v2
u2

Figure 18.3. The map : P Q , where P = A1 A2 A3 and

Q = B1 B2 B3 .
Apply the volume-preserving linear map (u1 , x1 , . . . , xd ) (u1 , x1 , . . . , xd ) which xes S
and sends u1 to u1 . Similarly, map (u2 , x1 , . . . , xd ) to (u2 , x1 , . . . , xd ). The resulting bipyramid is now symmetric with respect to the diagonal e = (u1 , u2 ). Therefore, the simplices
A1 = (u1 , u2 , x2 , . . . , xd ), A2 = (u1 , u2 , x1 , x3 . . . , xd ), . . . , Ad = (u1 , u2 , x1 , . . . , xd1 ) have
equal volumes. They each contain the edge e, and form a simplicial subdivision P = di=1 Ai .
Denote by : P P the resulting PL-map.
Now, apply the analogues piecewise linear transformations to Q, to obtain a simplicial
subdivision Q = di=1 Bi , where all simplices Bi contain an edge e = (v1 , v2 ). Denote by
: Q Q the corresponding PL-map.
There is a natural linear transformation i : Ai Bi which maps e e by sending

u1 to v1 and u2 to v2 , and maps the boundary of P into the boundary of Q .45 Since
vol(Ai ) = vol(Bi ), these maps are volume-preserving and combine into a continuous volumepreserving PL-map : P Q (see Figure 18.3). All together, 1 : P Q is the
desired PL-map.

45Strictly

speaking, i are ane transformations of Rd which are linear on Ai .

173

18.3. Monge maps by volume sharing. We are now ready to prove Theorem 18.1.
Start with a continuous PL-map : P Q constructed in Example 18.5. This map
creates simplicial subdivisions P = ni=1 Pi , Q = ni=1 Qi , which will be fixed from here
on. For every i [n] compute ai = vol(Pi )/vol(Qi ). If all ai = 1, the map is also
volume-preserving, i.e., the Monge map is constructed. If not (which is more likely),
let us correct map with a series of local PL-maps : Q Q.
Think of the numbers ai as contraction ratios of the linear maps i : Pi Qi .
Formally, ai is the determinant of the inverse map at points in Qi . The maps we
construct below will be continuous PL-maps : P Q which will have contraction
ratios bi at all points z Qi . In other words, the maps will neither be linear on Pi ,
nor take Pi to Qi . The maps we construct will be piecewise linear on all on Pi , and
the determinant of the inverse map 1 will be a constant on Qi , for all i.
We begin with the obvious equation for the volume of simplices involved:
() vol(P ) =

n
X
i=1

ai vol(Qi ) =

n
X

vol(Qi ).

i=1

Consider two adjacent simplices Qi and Qj and real numbers ai , aj such that
(U) ai vol(Qi ) + aj vol(Qj ) = ai vol(Qi ) + aj vol(Qj ).
Let us show how to construct a continuous PL-map : P Q which is going to
have contraction ratios ai and aj on Qi and Qj , and unchanged contraction ratios
elsewhere.
Suppose simplices Qi , Qj have a common face F . Expand the height of both simei and Q
ej form
plices by a factor of ai and aj , respectively. The resulting
simplices Q


a bipyramid A of volume ai vol(Qi ) + aj vol(Qj ) . Similarly, expand the height of
e and Q
e
both simplices by a factor of ai and aj , respectively, to obtain simplices Q
i
j
which form a bipyramid B of the same volume. Let Qij = Qi Qj , and denote by
1 : Qij A, 2 : Qij B the two maps defined above. Let : A B be a volumepreserving PL-map between these two bipyramids constructed in Example 18.5.
Now, let : Qij Qij be defined as the composition = 1
2 1 and extend
to a PL-map Q Q by the identity map outside of the bipyramid Qij . This is
possible since the map in the example was linear on the corresponding faces of the
bipyramids (see Figure 18.4). Finally, let = . By construction, 1 = ( )1
outside of Qij , and the contraction ratios on Qi and Qj are now ai and aj because
is volume-preserving and 1 and 2 expand by exactly these factors.
Now that we know how to change the contraction ratios locally, let us make the
global change. Loosely speaking, we claim that we can make local changes as in (U)
to obtain a change as in the second P
equality in (). Formally, consider a set A of
all sequences (a1 , . . . , an ) such that i ai vol(Qi ) = vol(Q) and ai > 0. Let E be
the set of pairs (i, j) such that simplex Qi is adjacent to Qj (the intersection is a
(d 1)-dimensional face). We are allowed to change
(. . . , ai , . . . , aj , . . . ) (. . . , ai , . . . , aj , . . . ) whenever (U) is satisfied.

174

Qj

Qi

Qj

Qi

Qij

Qij

1
2
ej
Q

ei
Q
A

e
Q
i

e
Q
j
B

Figure 18.4. Map : Qij Qij , where = 1

2 1 .
We need to show that (a1 , . . . , an ) is connected to (1, . . . , 1) through a series of
changes.
Denote by G the graph on [n] with E as its edges. Since G is a dual graph to a
simplicial subdivision of Q, we conclude that G is connected. The desired series of
changes is possible by Exercise 18.1. This completes the proof of Theorem 18.1. 
18.4. Exercises.
Exercise 18.1. (Money exchange) [1] Finish the proof of Theorem 18.1 by showing that
global change can be made via a nite number of local changes, as in the proof above.
Exercise 18.2. (Polygons with holes) While Theorem 18.1 does not cover polygons with
holes, it is in fact true that the theorem extends to this case. Formally, let A and B be
two convex polygons and let S1 , . . . , Sk A and T1 , . . . , Tk B be non-intersecting convex
polygons. Suppose
area(A)

k
X
i=1

area(Si ) = area(B)

k
X

area(Ti ).

i=1

Consider two non-convex polygons P = A i Si and Q = B i Ti with the same area

and the same number of holes.
a) [1] Prove by a direct argument that there exists a continuous piecewise linear map
: P Q.
b) [1] Prove that there exists a Monge map : P Q.
c) [1] Suppose A = B. Prove that there exists a Monge map : P Q which is identity
on the boundary A.
d) [1+] Let A B be 22 squares, k = 1, and let S T be 11 squares, such that S is in
the center of A, and T is at distance z > 0 from the boundary. Denote by f (z) the minimal
number of triangles needed to dene a Monge map : (A r S) (B r T ). Find the lower
and the upper bounds on f (z) as z 0. An example proving that f ( 14 ) 14 is given in
Figure 18.5.

175
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxxx
x
xxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxx

10

12

11

xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

12

13

14

xxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

10

11

13

14

4
3

Figure 18.5. Monge map : P Q which is linear on 14 triangles.

Exercise 18.3. (Poset polytope) Let P be a poset with the partial order on a nite
set X, where |X| = n.46 Dene poset polytope PP Rn as a set of functions f : X R,
such that 0 f (x) 1 for all x X, and f (y) f (z) for all y z, y, z X.
a) [1-] Prove that the vertices of PP correspond to order ideals of P.
b) [1-] Prove that vol(PP ) = e(P)/n!, where e(P) is the number of linear extensions of P.
c) [1] Dene the Fibonacci polytope FPn Rn to be the convex hull of Fibonacci sequences
(z1 , . . . , zn ), where zi {0, 1} and no two ones are adjacent. Show that FPn is congruent
to a certain poset polytope. Conclude that vol(FPn ) = an /n!, where an is the number of
alternating permutations (1) < (2) > (3) < . . .
Exercise 18.4. (Chain polytope) As in the previous exercise, let P be a poset on a nite
set X. Dene chain polytope QP to be the set of functions g : X R such that g(x) 0
for every x X, and g(x1 ) + . . . + g(xi ) 1 for every antichain {x1 , . . . , xn } X.
a) [1-] Prove that the vertices of QP correspond to antichains of P.
b) [1] Suppose P has no 3-element chains. In notation of the previous exercise, prove
that PP is congruent to QP .
c) [1] Let : PP QP be a map dened by


(f ) (x) = min f (x) f (y).
yx, yX

Check that is a Monge map. Conclude that vol(QP ) = e(P)/n!.

d) [1] Use the map to show that i(cPP ) = i(cQP ), for all c N, where i(Z) denotes the
number of integral points in Z.
e) [1-] Check that PP and QP have equal number of vertices. Give a direct bijective proof
of this fact. Explain what is special about , since not every Monge map preserves the
number of vertices.

Exercise 18.5. Dene the plane partitions polytope Rn by the following equations and
inequalities:

nj
ni

X
X

bi+k,k = n i,
bk,j+k = n j ,
for all 1 i, j n.
k=1
k=1

b 0, b b
i,j
i,j
i,j+1 , bi,j bi+1,j ,
a) [2-] Prove that Rn has exactly n! integral points.
46For

poset terminology and various related results, see [Sta3, 3].

176
2

b) [2-] Recall the denition of the Birkho polytope Bn Rn (see Exercise 8.19). Prove
that the number of symmetric integer matrices (aij ) Bn is equal to the number of symmetric integer matrices (bij ) Rn .
c) [2] Give an explicit Monge map : Bn Rn . Conclude that vol(Bn ) = vol(Rn ) and
i(cBn ) = e(cRn ), for all c N, where i(Z) is as in the previous exercise.
18.5. Final remarks. This section is based on papers [HenP, Kup2]. The main theorem
and the outline of the proof is due to Kuperberg, while various details in the presentation
follow [HenP]. Both papers are motivated by a theorem of J
urgen Moser, which states
that if two manifolds are dieomorphic and have equal volume, then there is a volumepreserving dieomorphism between them. In a dierent direction, the explicit Monge maps
often appear in algebraic combinatorics to encode the bijections between integral points in
certain combinatorial polytopes (see Exercises 18.5 and 18.4). We refer to [HenP] for an
extensive discussion, for connections to Mosers theorem and piecewise linear combinatorics,
and for the references.
We name the maps in this section after Gaspard Monge, a 19-th century French geometer who in fact never studied piecewise linear maps in this context. We do this in part
due to their importance and in part for lack of a better name. Also, in the case of general convex bodies, continuous volume-preserving maps do appear in connection with the
MongeKantorovich mass transportation (or optimal transportation) problems (see [Barv,
4.14]) and the MongeAmp`ere equations (see [Caf]). The uniqueness of solutions in that
case seems to have no analogue in our (polyhedral) situation, which is not surprising in
the absence of a natural minimization functional and the non-compactness of the space of
PL-maps.
We should mention that Exercise 18.2 is a special case of a general result which states
that all PL-homeomorphic PL-manifolds of the same volume have a volume-preserving
PL-homeomorphism [HenP, Kup2]. The proof is nearly unchanged from the proof of Theorem 18.1.
A small warning to the reader: not all homeomorphic PL-manifolds are PL-homeomorphic.
While this holds for 2- and 3-dimensional manifolds, this fails for 5-dimensional manifolds
(Milnor, 1961), and the problem is known under the name Hauptvermutung. Thus, one
should be careful generalizing the result in the exercise. We refer to [RouS] for a good
introduction to the subject.

177

19. Regular polytopes

It is often believed that the regular polytopes are classical, elementary, and trivial to
understand. In fact, nothing can be further from the truth. In fact, various questions
about them lead us to the Cauchy theorem (Section 26) and the Alexandrov existence
theorem (Section 37). Although this section does not have a single theorem, it is useful
as a motivation to these crucial results in the second part.
19.1. Definition is an issue. While everyone seems to know the classification of
regular polytopes (tetrahedron, cube, octahedron, dodecahedron and icosahedron),
few people tend to know what are they, i.e., how the set of regular polytopes are
defined. This leads to several confusions and misperceptions which we will try to
dispel. In fact, even the most basic incorrect definitions of regular polytopes are
sufficiently enlightening to justify their study.
Incorrect Definition 1. Regular polytopes = convex polytopes where all sides are
regular polygons.
Why not: this definition is not restrictive enough. For example, a triangular prism
and square antiprism fit into this category.
Incorrect Definition 2. Regular polytopes = convex polytopes where all sides are
regular polygons with the same number s of sides.
Why not: again not restrictive enough; For example, a triangular bipyramid fits into
this category (see Figure 19.1).
Correct yet unsatisfactory Definition 3. Regular polytopes = convex polytopes
where all sides are regular polygons with the same number s of sides, and where every
vertex has the same degree (i.e., adjacent to the same number r of edges).
Why unsatisfactory: while the set of such polytopes is exactly what we want, there
are two problems with this definition. First, this definition does not generalize well
to higher dimensions. Second, it does not explain the large group of symmetries of
regular polytopes.

Figure 19.1. Triangular prism and bipyramid.

We are now ready to give a correct and fully satisfactory definition.
Definition 19.1. A polytope P Rd is called regular if its group of symmetries acts
transitively on complete flags of P .

178

Here by complete flags we mean a sequence: vertex edge 2-dimensional face

. . . P . Now consider the numerical consequences of the definition when d = 3.
Denote by the group of symmetries of P . Let AF be the set of edges contained
in face F . Since acts transitively on complete flags, polytope P is face-transitive.
Thus, |AF | = |AF |, for all faces F and F . In other words, all faces are polygons
with the same number s of sides.
Similarly, denote by Ev the set of edges containing vertex v of the polytope P .
Again, by transitivity of the action of we conclude that P is vertex-transitive.
Thus, |Ev | = |Ev |, for all vertices v and v . In other words, all vertices are adjacent
to the same number r of edges. This gives the conditions in definition 3. Note
that P must also be edge-transitive, restricting the possibilities even further.
19.2. Classification of regular polytopes in R3 . Let P R3 be a regular convex
polytope, where the sides are regular s-gons, and all vertices are adjacent to r edges.
Denote by n = |V | the number of vertices, m = |E| the number of edges, and f = |F |
the number of faces. Counting all edges via vertices and faces we obtain:
2 m = n r = f s.

Writing n and f through m, we rewrite Eulers formula n m + f = 2 as follows:

2m
2m
m+
= 2,
r
s
and finally
1
1
1
1
+ =
+ .
r
s
m
2
From here either 1/r or 1/s is > 1/4, and since r, s, m 3, we can check all cases to
conclude that
(r, s) {(3, 3), (3, 4), (4, 3), (3, 5), (5, 3)}.
These five cases correspond to the five platonic solids in the same order as listed in
the beginning of the previous subsection. It would seem that we are finished, except
for one annoying little problem:
How do we know these regular polytopes actually exist?
We discuss this little problem at length and eventually resolve it completely in the
next subsection.
19.3. Constructing polytopes is harder than you think. Well, we all know that
these polytopes exist. Having seen them all over the place it is easy to take things for
granted. Of course, there must be a theoretical reason. After all, that was a crown
achievement of Euclid. But are you really sure? Let us go over the list to see what is
really going on.
Naturally, there is little doubt the regular tetrahedron and the cube exist. The
centers of faces of a cube give vertices of an octahedron. Same for the dodecahedron:
it is formed by taking centers of faces of the icosahedron. Now, when we get to
icosahedron we are in trouble. A priori, there is no theoretical reason why it should
exist. Can we explicitly construct it?

179

(i) Coordinate approach. The icosahedron can be constructed bycoordinates of all 12

vertices: (0, 1, ), (1, , 0), and (, 0, 1), where = 1+2 5 is the golden ratio
(see [Hada, 564]). It takes some minor checking to see that the convex hull of these
points forms an icosahedron. First, observe that the vertices lie on a sphere centered
at the origin, and thus in convex position. Cyclically permuting coordinates and
changing signs defines an action of the group Z3 Z32 (of order 24) on all 120 = 1252
complete flags. With one additional rotation around a diagonal (see Exercise 19.1),
one can obtain the full group of symmetries A5 Z2 of order 120, and prove
existence of the icosahedron.
It is important to observe that the above argument was indirect: we did not really
construct the icosahedron we just checked someone elses construction. Also, Euclid
clearly could not have conceived this construction. So here is another approach.
(ii) Continuity argument. Start with a regular pentagon Q with side 1 and a pentagon Q obtained from Q by rotation (see Figure 19.2). Start lifting Q up until at some
point the side edges connecting two levels become equal to 1. This will happen by
continuity. Finally, add a regular pentagonal cap on top and on the bottom to obtain
an icosahedron (see Figure 19.2). Although, by construction, all faces are equilateral
triangles, this does not immediately imply that the resulting polytope is regular.
Now, consider the group of symmetries of the resulting polytope P . Clearly, we have
all symmetries which preserve the main diagonal (North to South Pole). This gives
a group of symmetries with 20 elements. However, the action of the full symmetry
group on vertices is not obviously transitive: to prove that P has all 120 symmetries
we need to exhibit an orthogonal transformation which maps the main diagonal into
another diagonals.

Q
Q

Figure 19.2. Construction of the regular icosahedron, step by step.

(iii) Inscribing into a cube. Here is another approach, arguably more elegant. Draw
intervals of length x symmetrically on each face of the square as in Figure 19.3. Denote
by Px a convex hull of these intervals. Beginning with a small length x, increase x
until all edges of Px have the same length. Again, the resulting polytope P has all
faces equilateral triangles, but that in itself does not a priori prove that P is regular.
By construction, the group of symmetries of P is transitive on all vertices. In
fact, already the subgroup of joint symmetries of P and the cube is transitive on
the vertices of P and has order 24. However, the action on faces is not obviously

180

transitive: triangles containing original intervals and those that do not may a priori
lie in different orbits.

Figure 19.3. Another construction of the regular icosahedron.

Unfortunately, there seem to be no soft argument which would give a construction
of the icosahedron and prove that it has the full group of symmetries. Either an
additional calculation or an advance result is needed.
Let us the classical Cauchy theorem to prove that polytope P has the full group
of symmetries . Fix two flags v e F P and v e F P in polytope P .
Let P be a polytope obtain from P by relabeling of faces with respect to these flags.
Formally, let P be a polytope combinatorially equivalent to P , where the map in
the definition of combinatorial equivalence is given by v v , e e , and F F .
By the Cauchy theorem, polytope P can be moved into P by a rigid motion, thus
giving the desired symmetry between the flags. Therefore, the symmetry group
acts transitively on complete flags of P , and the polytope P is a regular icosahedron.
There are two more constructions of the icosahedron that we would like to mention.
First, one can start with an unfolding of 20 equilateral triangles and glue the polytope
(see Figure 19.4). The Alexandrov existence theorem (Theorem 37.1) implies that
this is possible. Still, to prove uniqueness in this case one needs to use the Alexandrov
uniqueness theorem (Theorem 27.7), an extension of the Cauchy theorem, and the
edges a priori are not guaranteed to be at the right positions.

Figure 19.4. An unfolding of the regular icosahedron.

(iv) The original construction. Finally, there is a construction that was essentially
used by Euclid. Let us construct a dodecahedron rather then an icosahedron. It is a
classical fact (but still remarkable for those seeing it the first time), that a cube can be
inscribed into a dodecahedron such that the edges become diagonals of the pentagonal

181

faces. In Figure 19.5 we show the view from the top: the diagonals are equal because
each face is a regular pentagon and the angles are straight by the symmetry.
Now, one can start with a cube and build a dodecahedron by attaching caps to
all six sides (see Figure 19.5 for a cube and two caps). Note that the Cauchy theorem
is obvious in this case as the polytope is simple, and thus three pentagons completely
determine the dihedral angles in all vertices. By the argument as above, this implies
that our polytope is regular. We leave it as an exercise to see why this construction produces a regular dodecahedron seemingly without any extra calculations (see
Exercise 19.2).

Figure 19.5. Construction a regular dodecahedron, after Euclid.

19.4. Exercises.
Exercise 19.1. [1] Find an explicit 3 3 matrix for a rotational symmetry of order 5
to show that the icosahedron constructed via coordinates is indeed a regular polytope.
Exercise 19.2. [1-] Compute the dihedral angles in Euclids construction of the dodecahedron to show that the pentagons are indeed at.

Figure 19.6. Folding six pentagons to form a half of the dodecahedron.

Exercise 19.3. a) [1-] Start with ve regular pentagons attached to one. Fold them all
along the edges continuously until they form a half of the surface of the regular dodecahedron
(see Figure 19.6). Note that there is a unique way to attach ve regular pentagons to extend
this surface. Check that by the symmetry these pentagons t together. Again, check that

182

by the symmetry, the remaining hole in the surface is a regular pentagon. Is this an honest
calculation-free construction of the regular dodecahedron?
b) [1-] Start the same way by gluing together 6 regular pentagons. Consider their projection
as in Figure 19.6. Use Euclidean geometry to show that the projection is a regular 10-gon.
Conclude that two such surfaces t together to form the surface of the dodecahedron.
c) [1-] Use the rigidity of simple polytopes, to show that the resulting polyhedron has the
full group of symmetries.
d) [1-] Use this to construct the regular icosahedron.
Exercise 19.4. a) [1] Decide on theoretical grounds (as opposed to explicit construction)
which of the three unfoldings in Figure 19.7 are unfoldings of convex polytope whose faces
are equilateral triangles and squares. Is there a unique way of gluing them together?
b) [1+] Prove that these polytopes exist, by an explicit construction.

Figure 19.7. Unfoldings of three convex polytopes.

Exercise 19.5. Let P be a convex polytope with at least one symmetry line, dened as an
axis of rotational symmetry by 180 .
a) [1-] Prove that all symmetry lines pass through the same point.
b) [1] Prove that P has an odd number of symmetry lines.
Exercise 19.6. [1] Show that a regular octahedron Q can be inscribed into a cube C R3
so that every vertex of Q lies on the edge of C.
Exercise 19.7. a) [1] Prove that there exists a 3-coloring of edges of the icosahedron such
that every face has all three colors. Moreover, show that such 3-colorings are combinatorially
equivalent (up to the symmetry of the icosahedron and relabeling of the colors).
b) [1] The same question for 5-colorings of vertices of a 600-cell (a regular simplicial polytope
in R4 with 600 facets; see Exercise 19.15).
Exercise 19.8. A polytope P R3 is called face-transitive if for every two faces F, F P
there exists a symmetry of P which maps F into F .
a) [1-] Find a face-transitive polytope with n triangular faces, for every even n 4.
b) [1] Find a face-transitive polytope with n quadrilateral faces, for every even n 4.
c) [1+] Show that there are no face-transitive polytopes in R3 with an odd number of faces.
Exercise 19.9. Let P R3 be a convex polytope. Think of P as a playing die. Denote
by O = cm(P ) the center of mass of P . For every face F in P , denote by CF a cone from O
over F . We say that P is fair if the solid angles (CF ) are equal (see Subsection 25.4). We
say that P is super fair if O projects orthogonally onto centers of mass cm(F ) of all faces,
and all faces F P are centrally symmetric.
a) [1] Prove that there exists a fair die with n faces, for all n 4.
b) [1] Prove that there exist a super fair die with n faces, for all n 4.

183

Exercise 19.10. (Gr

unbaumShephard ) [2] Let S R3 be an embedded polyhedral surface
of genus of g > 0. Prove that S cannot be face-transitive.
Exercise 19.11. [1] Suppose P R3 is a simple polytope such that all faces are inscribed
polygons. Prove that P is inscribed into a sphere.
Exercise 19.12. a) [1-] Find a convex polytope in R3 with equal edge lengths, which is
midscribed (has all edges touch the sphere) and circumscribed, but not inscribed into a
sphere.
b) [1-] Find a convex polytope in R3 with equal edge lengths, which is midscribed and
inscribed into sphere, but not circumscribed around a sphere.
Exercise 19.13. Let P R3 be a convex polytope which is midscribed and has edges of
equal length.
a) Suppose one of the faces of P is odd-sided. Prove that P is also inscribed.
b) Suppose all faces of P have the same number of sides. Prove that P is circumscribed.
Exercise 19.14. a) [1] Find all convex polytopes whose faces are equilateral triangles.
b) [1] Prove that there is only a nite number of combinatorially dierent convex polytopes
whose faces are regular convex polygons with at most 100 sides.
c) [1] Prove that for every n > 100 the number of such polytopes with n faces is two if n is
even, and one if n is odd.
Exercise 19.15. (Regular polytopes in higher dimensions) a) [2-] Show that the only
regular convex polytopes in Rd , d 5 are the simplex, the hypercube and the cross-polytope.
b) [1+] Prove that in R4 there are three additional regular convex polytopes, with 24, 120,
and 600 vertices. Give a construction of these polytopes in the same manner as in the
section.
c) [1+] Prove that in R4 , cross-polytopes tile the space in a regular fashion, i.e. the resulting
tiling is ag-transitive.
Exercise 19.16. (Catalan zonotopes) Dene the rhombic dodecahedron to be a facetransitive polytope dual to the cuboctahedron (see Figure 16.4). Similarly, dene rhombic
triacontahedron to be a face-transitive polytope dual to the icosidodecahedron (see Exercise 16.20).
a) [1] Prove that in addition to the Platonic solids and icosidodecahedron, these are the
only convex polytopes in R3 which are edge-transitive.
b) [1-] Prove that both polyhedra are zonotopes (see Exercise 7.16).
c) [1] Find an explicit decomposition of both polyhedra into parallelepipeds.
Exercise 19.17. (Jessens orthogonal icosahedron) a) [1-] Start with a regular icosahedron positioned in a cube as in Figure 19.3. For every (red) edge e of the icosahedron lying
on the cube remove the tetrahedron spanned by two triangles containing e (see Figure 19.8).
Prove that the resulting (non-convex) polyhedron is vertex-transitive and combinatorially
equivalent to an icosahedron.
b) [1] Construct a combinatorial icosahedron as in Figure 19.3 where the distance x is
equal to half the edge length of the cube. Starting with this convex polytope construct a
non-convex polyhedron as in part a). Prove that all dihedral angles are either /2 or 3/2.
Exercise 19.18. (Equifacetal simplices) a) [1] Prove that there exist > 0 small enough,
such that for every collection of lengths ij [1 , 1 + ], there exists a simplex =
[v0 v1 . . . vd ] Rd with edge lengths ij = |vi vj |.

184

Figure 19.8. Construction of Jessens orthogonal icosahedron.

b) [1-] A simplex with congruent facets is called equifacetal. Construct an equifacetal
simplex which is not regular.
c) [1+] Prove that every equifacetal simplex is vertex-transitive.
a) [1+] Find the largest possible in part a).
19.5. Final remarks. Regular polytopes in R3 are called Platonic solids, while pairs (r, s)
dened above are called their Schl
ai symbols. Note that values (r, s) {(3, 6), (4, 4), (6, 3)}
also make sense: they correspond to innite plane tessellations (honeycombs): the hexagonal, square and triangular lattices. In dimension 4 there are three special regular polytopes (see Exercise 19.15) which are even harder to construct without coordinates (their
uniqueness, however, is much easier given the 3-dimensional case; cf. Subsection 27.2).
For a comprehensive overview of regular polytopes in all dimensions see classical monographs [Cox1, Fej2], and a recent opus magnum [McS]. It is worth noting that already
(perhaps, especially) in dimension 4, the regular polytopes and tessellations are unintuitive.
The reader can take solace in the following warning by Coxeter [Cox1, 7.1]:
Only one or two people have ever attained the ability to visualize hyper-solids
as simply and naturally as we ordinary mortals visualize solids.

Although perhaps overly pessimistic, in all fairness, this quote is taken somewhat out of
context, as Coxeter dedicates the whole section to this discussion (aided, in turn, with a
nice quote by Poincare).
Let us mention here the 13 Archimedean solids, a related family of polytopes which are all
vertex-transitive and have faces regular polygons. Similarly, the face transitive polytopes
dual to Archimedean solids are called Catalan solids. Finally, all convex polytopes with
regular faces are called Johnson solids (see Exercise 19.14), and are completely classied by
Johnson and Zalgaller (see [Crom, Joh, Zal3]).
Various constructions of the icosahedron and connection to rigidity are described in [BorB,
20]. Note that our constructions (i) and (iii) essentially coincide - we chose to emphasize
the dierence in the geometric argument rather than the outcome. As for the original
construction (part (iv)), by itself it does not produce a regular polytope: one has to check
that the parts of pentagonal faces align properly (see Exercise 19.2). However, there are
other constructions which seem to require fewer computation (see Exercise 19.3), albeit
with some sort of weak rigidity argument. The connection to rigidity was rst discovered
by Legendre in his eort to rewrite Euclids Elements (see [Sab6]).

185

20. Kissing numbers

The problem of determining kissing numbers is one of the most celebrated problems
in discrete geometry. The problem is surprisingly difficult, so we prove only the most
basic results in small dimensions. While these results are completely unrelated to the
rest of the book, we found them an excellent way to give a quick and easy introduction
to spherical geometry. As the reader will see, the basic spherical geometry is used
throughout the second part.
20.1. Basic results. Denote by Kd the d-dimensional kissing number, defined as the
largest number of unit balls touching a fixed unit ball in Rd . Observe that Kd can be
also viewed as the largest number of points on a unit sphere Sd1 Rd with pairwise
distances 1.
Proposition 20.1. K1 = 2, K2 = 6.
Proof. The first part is clear. For the second part, observe that K2 6 as shown
in Figure 20.1. To see that K2 6, consider k points a1 , . . . , ak on a unit circle
with center at the origin O, and with pairwise distances 1. Take their convex hull.
Since from the edge lengths of (a1 , a2 ), (a2 , a3 ), . . . , (ak , a1 ) are 1, we conclude that
a1 Oa2 , a2 Oa3 , . . . , ak Oa1 /3. Therefore

2 = a1 Oa2 + a2 Oa3 + . . . + ak Oa1 k ,

3
and k 6.


O1

O2

O3

Figure 20.1. Six touching circles and centers of twelve touching spheres.
It is natural to use the idea of the proof of Proposition 20.1 to obtain the bounds
in dimension three.
Proposition 20.2. 12 K3 14.
Proof. The lower bound follows from a grocery style sphere packing. Consider
six unit spheres with centers on a horizontal plane as above, and start lowering the
three new unit spheres so that their centers are projected orthogonally onto centers
O1 , O2 , O3 of triangles, as shown in Figure 20.1. By the symmetry, at one point all
three spheres will be touching the spheres below. Now observe that these three new
spheres are non-overlapping (in fact touching themselves) since the distance between

186

their centers is equal to 2. Now adding in a similar fashion the three spheres from
below we obtain the desired 12 spheres.
For the upper bound, consider k unit spheres S1 , . . . , Sk kissing a fixed unit sphere
S0 R3 centered at the origin. Now let Ci be a cone over Si centered at the origin O,
and let Ri = S0 Ci be the spherical caps on a sphere. Since the centers Oi of
spheres Si are all at the same distance from O and their pairwise distances are 2,
we conclude that the cones Ci must be disjoint and thus so are caps Ri . We conclude
that area(Ri ) /k, where = area(S) = 4 denotes the area of a unit sphere.
To compute the area(Ri ), recall that area of a spherical cap is proportional to the
height of the cap. From Figure 20.2, we have yOz = wOOi = arcsin 12 = 6 .
Therefore, for the height h of the cap Ri we have:

3
h = |zx| = 1 |Oz| = 1 cos = 1
.
6
2

w
y
Ci
Ri

Oi

Figure 20.2. Computing the height of a cap Ci on a sphere.

Finally, we compute the area of a cap:

1 23

h
>
.
area(Ri ) = =
2
2
14.9283
Since from above area(Ri ) /k, we conclude that k 14.

20.2. The problem of fourteen spheres. Let us employ a somewhat sharper (but
still elementary) technique to show that K3 13. For that, we will need only two
elementary results: Eulers formula (see Section 25) and Girards formula for the area
of spherical triangles (see Appendix 41.1). In fact, it is known that K3 = 12, but all
known proofs are much more involved.
Let us first discuss the limitations of the approach in the proof above. Note that
we significantly underestimated the area of a sphere by adding the areas of the caps,
leaving all regions between the cap unaccounted for. From a combinatorial point

187

of view, it is natural to make the caps larger so as to count more of the area, and
then use the inclusion-exclusion principle to account for overlap of the caps. The
argument below is exactly of this kind: we first overestimate the area of a sphere and
then subtract the intersection areas.
Theorem 20.3. K3 13.

Proof. Suppose one can place points a1 , . . . , a14 on the unit sphere S2 so that the
pairwise distances |ai aj | 1, for all i 6= j. Denote by P = conv{a1 , . . . , a14 } the
convex polytope obtained as a convex hull of points ai . Note that all ai are vertices
of P .
We will assume that the origin O is the center of S2 , and that O P since otherwise
all ai lie in the same half-space, and one can add an extra point a15 contradicting
Proposition 20.2. Finally, we will assume that P is simplicial, i.e., each face of P is a
triangle. Otherwise, triangulate each face and proceed as follows without any change.
By Eulers Theorem, the polytope P with n = 14 vertices will have 3n 6 = 36
edges and 2n 4 = 24 triangular faces. Now, for each face (ai aj ak ) consider a circle
Cijk around the vertices; one can think of Cijk as an intersection of S2 and the plane
which goes thorough ai , aj and ak .
Let us consider the 14 caps R1 , . . . , R14 with centers atpoints ai and such that the
circles of the caps Ci = Ri have the same radius = 1/ 3. We need several simple
results on geometry of caps Ri . As before, let = area(S2 ) = 4.
Lemma 20.4. For the caps R1 , . . . , R14 defined as above, we have:
1) area(Ri ) > 0.0918 ,
for all
1 i 14;
2) area(Ri Rj ) < 0.0068 ,
for all
1 i < j 14;
3) area(Ri Rj Rk ) = 0,
for all
1 i < j < k 14.
We prove the lemma in the next subsection, after we finish the proof of the theorem.
Observe that the only pairs (i, j) for which we can have area(Ri Rj ) 6= 0 are those
corresponding to 36 edges of the polytope P . Using all three parts of the lemma we
get:
14
X
X
= area(S)
area(Ri )
area(Ri Rj )
i=1

1i<j14

> 14 (0.0918 ) 36 (0.0068 ) > 1.03 ,

a contradiction. Therefore, K3 < 14 as desired.

20.3. Proof of Lemma 20.4. For part 1), recall that radius(Ri ) = = 1/ 3. From
Figure 20.3, the height h of the cap Ri is given by47
r
r
p
1
2
h = 1 1 2 = 1 1 = 1
.
3
3
47Note

that in the proof of Proposition 20.2 we used spherical cap of radius 1/2. Loosely speaking,
we are undercounting in the proof of the proposition, and are overcounting in the proof of the
theorem.

188

O
S

Ri
Figure 20.3. Computing the area of spherical caps Ri .
Therefore, by the same reasoning as in the proof of Proposition 20.2, for the area
of Ri we have:
q
1 23
h
area(Ri ) = =
> 0.0918 .
2
2
For part 2), we can assume that the distance |ai aj | = 1 since for larger distances
the area of the link L := Ri Rj is smaller. Denote by b the remaining vertex of
the equilateral spherical triangle (ai aj b) and by u, v the corners of the link as in
Figure 20.4.

O
u

Ri

ai

/2

aj
v

Rj
S

Figure 20.4. Computing the area of lenses Ri Rj .

By the symmetry, the point u S is the center of the triangle (ai aj b), and we have
equal spherical angles = uai v = uaj v = bai aj . Now, for the area of L we have


area(L) = 2 area(Q) area(T ) ,

189

where Q denotes the sector [ai uv] Ri and T denotes the spherical triangle (ai uv)
Q. Again, by the symmetry and from Girards formula we have:
p
p
1 2/3

area(Ri ) =

(4) = 2/3 ,
area(Q) =
2
2
2

area(T ) = + + = .
3
3
3
Now, observe that is a dihedral angle in the regular tetrahedron (Oai aj b) with side 1.
Computing it as shown on Figure 20.4, we obtain = 2 arcsin 1/2
= 2 arcsin 13 . We
3/2
conclude:
p


area(L) = 2 area(Q) area(T ) = 2 (/3 2/3 )
r
2
2
1
=
4
arcsin < 0.0843 < 0.0068 .
3
3
3
For part 3), we begin with the result which follows directly from the argument used
in the proof of Corollary 1.8, with all inequalities reversed.
Lemma 20.5. For all faces (ai aj ak ) in P , the radius(Cijk ) , where =

1 .
3

Suppose now there is a point z S which lies in the interior of three caps Ri , Rj
and Rk . By definition, the distances from z to vertices ai , aj and ak are strictly
smaller than . Thus, the same is true for the projection z on a plane spanned by
ai , aj and ak . This implies that the triangle (ai aj ak ) can be enclosed into a circle C
with center z and radius r < . Expand the triangle (ai aj ak ) to a triangle (ai aj ak )
inscribed into C with bigger edge lengths, as shown in Figure 20.5. By Lemma 20.5,
the circumradius of a triangle with edge lengths 1 is , a contradiction. This
completes the proof of Theorem 20.3.
ak

ak

C
ai
aj

ai

aj

Figure 20.5. Expanding triangle (ai aj ak ) (ai aj ak ).

20.4. Exercises.
Exercise 20.1. (Asymptotics) a) [1] Prove that Kd = exp O(d).
b) [1] Prove that Kd = exp (d).
Exercise 20.2. [1] Find the largest number of lines in R3 , which pass through the origin
and have equal angles between them.

190

Exercise 20.3. [1+] Let P R3 be a convex polytope with 19 faces. Suppose P is

circumscribed around a sphere of radius 10. Prove that diam(P ) > 21.
Exercise 20.4. [1] Let x1 , . . . , x30 S2 be points on a unit sphere. Prove that there exist
1 i < j 30, such that


xi xj  2 < .
S
4

In other words, show that some two of the vectors Oxi have angle < /4 between them.
Exercise 20.5. a) [1] Suppose n points on a unit sphere have pairwise spherical distances
/2. Prove that n 6, and that for n = 6 these points must be vertices of an inscribed
regular octahedron.
b) [1+] Prove an icosahedral analogue of a).
c) [2-] Suppose x1 , . . . , xn and y1 , . . . , yn are two congurations of points in the unit
sphere S2 , such that |xi xj | |yi yj | for all i, j [n]. Suppose further that O conv{x1 , . . . , xn },
where O is the center of S2 . Prove that |xi xj | = |yi yj | for all i, j [n].
Exercise 20.6. a) [1-] Is it possible to cover the plane by a union of parabolas, i.e., regions
obtain by translations and rotations of y ax2 , where a > 0.
b) [1] Suppose in the space R3 there is a nite number of disjoint cones, i.e., regions obtain
by translations and rotations of x2 + y 2 az, where a > 0. Prove that these cones cannot
be moved to cover the whole space.

Exercise 20.7. On a unit sphere, dene an -arc to be an arc of a great circle of length .
a) [1-] For every < , prove that there are innitely many not self-intersecting -arcs on
a unit sphere.
b) [1] For = 5/3, prove that there are at most two not self-intersecting -arcs on a unit
sphere.
c) [1] Prove that there is an unbounded number of not self-intersecting -arcs on a unit
sphere.
d) [1+] For every > , prove that there is at most a bounded number n = n() of not
self-intersecting -arcs on a unit sphere.
Exercise 20.8. Prove that the following triangles (given by their angles) can tile the
sphere S2 without overlap:
a) [1-] (90 , 90 , 1 ), (90 , 60 , 60 ), (120 , 60 , 60 ),
b) [1] (120 , 45 , 45 ), (72 , 60 , 60 ), (150 , 60 , 60 ),
c) [1+] (80 , 60 , 60 ), (100 , 60 , 60 ), (100 , 80 , 60 ).
Exercise 20.9. a) [1-] Place the centers of twelve unit spheres in R3 at the vertices
of the icosahedron inscribed into a sphere of radius 2. Check that these spheres are not
overlapping and, moreover, no two of them touch each other. This gives an alternative
proof of the lower bound in Proposition 20.2.
b) [2-] In a), the outside spheres can be continuously moved while they remain nonoverlapping and touching the center sphere. Is it possible to switch any two spheres that
way?
c) [2+] Can these twelve spheres be moved into twelve spheres in the grocery style sphere
arrangement?
Exercise 20.10. Denote by KCd the number of innite cylinders of unit radius in Rd ,
which are non-overlapping and touching the unit sphere. Such conguration is called kissing
cylinders.

191

a) [1-] Prove that KCd Kd1 .

b) [1-] Find a continuous family of dierent congurations of six kissing cylinders in R3
(dierent up to rotations).
c) [2-] Prove that KC3 7.
d) [] Prove that KC3 6.

Exercise 20.11. For a nite arrangement B of balls in Rd with disjoint interior (not
necessarily of the same radii), denote by a(B) their average kissing number. Dene AKd to
be the supremum of a(B) over all nite ball arrangements.
a) [1-] Show that the AKd > a(B) for any nite B.
b) [1] Prove that AK3 > 12.
c) [1] Prove that AKd 2Kd .
d) [2] Prove that AK3 8 + 4 3.
Exercise 20.12. (Pairwise kissing spheres) Denote by P Kd the maximal number of pairwise kissing spheres in Rd (possibly, of dierent radius).
a) [1-] Use the previous exercise to show that P Kd = exp O(n). Explain the dierence with
Exercise 42.45.
b) [1] Use the algebraic approach (see Section 31) to show that P Kd = O(d).
c) [1] Prove that P Kd = d + 2.
20.5. Final remarks. Finding kissing numbers Kd is a classical problem connected to
the study of error-correcting codes, lattices, number theory and many problems in geometry [ConS, Fej2, Zon1] (see also [Bez] for a recent survey). Proving that K3 = 12 is known as
the NewtonGregory problem named after their celebrated exchange in 1694. The problem
was resolved by Sch
utte and van der Waerden only in 1953, but most known proofs are
quite technical. In fact, this section can be viewed as an introduction to Leechs proof, an
elementary presentation of which recently appeared in [Mae2].
We refer to [Mus6] for a recent interesting proof of K3 = 12. Let us also note that while
the exact value K4 = 24 was determined recently by Musin [Mus5], the kissing numbers
K8 = 240 and K24 = 196560 have been known since 1979 (see [ConS, Mus5, PfeZ]). The
proof of Theorem 20.3 presented in this section is a minor reworking of the proof in [Yag2].

192

Part II
Discrete Geometry of Curves and Surfaces

193

21. The four vertex theorem

In this section we present a number of results on the geometry of planar and space
polygons, all related to the four vertex theorem. While motivation for some of these
results lies in differential geometry, the piecewise linear results are often more powerful
and at the same time easier to prove. On the other hand, some continuous results
seem to have no discrete analogues, while others have several. To give a complete
picture, in the beginning and at the end of this section we present (without proof)
a number of continuous results as a motivation. Interestingly, the proof of the main
theorem in this section uses Voronoi diagrams (see Section 14).
Although the results in this section are similar in spirit to various results in the first
part (see Sections 5 and 9), there is a common thread of results in the second part.
Namely, in the next section we obtain several results on relative geometry of convex
polygons, further extending the four vertex theorem. One of these extensions, the
Legendre-Cauchy lemma is then used to prove the Cauchy theorem (see Section 26).
Much of the rest of the book is dedicated to various extensions and generalizations,
including Dehns theorem, the Alexandrov theorem, etc. The relative geometry of
polygons reappears several times, often in disguise, but still at the heart of many
proofs.
21.1. The flavor of things to come. Let C R2 be a smooth curve in the plane.
Throughout this section, all curves will be closed. We say that a curve is simple if it
is not self-intersecting. A point x C is called a vertex if it is an extremum (local
maximum or minimum) of the curvature . Here by a curvature (x) we mean the
inverse of the radius R(x) of the osculating circle at point x, so (x) = 0 when x is a
flat point.48
Observe that every smooth closed curve has at least two vertices. The following
classical result shows that if C is simple the number of vertices is at least four.
Theorem 21.1 (Four vertex theorem). Every smooth simple curve has at least four
vertices.
A few quick examples: an ellipse with distinct axes has exactly four vertices, while
all points on a circle are vertices. On the other hand, for non-simple curves the claim
is easily false (see Figure 21.1).
The next extension of the four vertex theorem is based on restricting the notion of
vertices. Let C be a smooth convex curve. An osculating circle R tangent to C at
point x C is called empty if it lies entirely inside C. Of course, then is a local
maximum at x, but the inverse does not necessarily hold. Similarly, an osculating
circle R tangent to C at point y C is called full if C lies entirely inside R. Then
has a local minimum at y. The empty and full osculating circles are called extremal
circles. By definition, the number of extremal circles is at most the number of vertices
of curve C.
48If

the curve is parameterized by arc length then the curvature is the length of the second
derivative vector.

194

Figure 21.1. An osculating circle in a convex curve with four vertices

and a non-simple curve with two vertices.
Theorem 21.2. Every smooth convex curve has at least four extremal circles.
Note that extremal circles have a global nature in contrast with a local nature
of vertices. Here is a quantitative extension of Theorem 21.2 due to Bose (1932). We
say that C is generic if it is not tangent to any circle at more than three points.
Theorem 21.3. Let C R2 be generic smooth convex curve in the plane. Denote
by s+ and s the number of full and empty osculating circles, respectively. Denote
by t+ and t the number of full and empty circles tangent to C at three points. Then
s+ t+ = s t = 2.
The theorem immediately implies Theorem 21.2, which in this language states that
s+ + s 4. We postpone other extensions and generalizations of the four vertex
theorem till Subsection 21.8.
21.2. Discretizing is ambiguous and harder than it looks. There are several
natural discrete analogues of the four vertex theorem (Theorem 21.1). We present a
few of them in this section, but more will appear later on. Remember that the word
vertex when applied to convex polygons has the usual meaning.
Let Q = [x1 . . . xn ] R2 be a convex polygon. We say that Q is generic if no
four vertices of Q lie on a circle. Denote by Ri a circle circumscribed around triangle
(xi1 xi xi+1 ), where all indices are taken modulo n. We say that a vertex xi is extremal
if xi2 and xi+2 lie on the same side of Ri , i.e., either both vertices lie inside or both
vertices lie outside of the circle.
Theorem 21.4 (Discrete four vertex theorem). Every generic convex polygon with
at least four vertices has at least four extremal vertices.
The following quantitative version of the four vertex theorem follows easily from
the theorem (see Exercise 21.1). We say that Q is coherent if the center of Ri lies
inside the cone of Q at xi .
Corollary 21.5. Let Q = [x1 . . . xn ] R2 be a generic coherent convex polygon,
n 4, and let ri denotes the radius of Ri , 1 i n. Then there are at least four
sign changes in the cyclic sequence (r1 r2 , r2 r3 , . . . , rn r1 ).
The idea is that circles Ri for convex polygons play the role of the osculating circles
for the curves, and their inverse radii correspond to the curvature. In this language,

195

the corollary implies that convex polygons have at least two maximal and at least
two minimal radii.
Let us note that for non-coherent polygons the corollary does not hold (see Exercise 21.1). On the other hand, all obtuse polygons (convex polygons where all angles
are right or obtuse) are coherent. Thus Corollary 21.5 for obtuse polygons can be
viewed as a direct extension of the (usual) four vertex theorem (Theorem 21.1).
Consider the following further reduction of the theorem. Let Q be an equilateral
convex polygon, i.e., a polygon with equal edge lengths a. Note that all equilateral
convex polygons are coherent. Denote by i = xi1 xi xi+1 the angle in vertex xi .
Since a = 2ri cos(i /2), the larger angles i correspond to larger radii ri , and we
obtain the following.
Corollary 21.6. Let Q = [x1 . . . xn ] R2 be a generic equilateral convex polygon,
n 4, and let i = xi1 xi xi+1 denote the angle in Q, 1 i n. Then there are at
least four sign changes in the cyclic sequence (1 2 , 2 3 , . . . , n 1 ).
Here is a natural analogue of the four extremal circles theorem. Let Q = [x1 . . . xn ]
be a convex polygon as above. A circle Rijk through a triple of vertices xi , xj , xk ,
i < j < k, is called disjoint if no two vertices are adjacent; it is called neighboring
if two of the vertices are adjacent to the third. The remaining circles, with only one
pair of adjacent vertices, are called intermediate. The circle Rijk is called empty if no
other vertices xr are inside, and it is called full if all other vertices xr are inside. We
are now ready to discretize Theorem 21.3.
Theorem 21.7. Let Q R2 be a generic convex polygon with at least four vertices.
Denote by s+ , t+ and u+ the number of full circles that are neighboring, disjoint and
intermediate, respectively. Similarly, denote by s , t and u the number of empty
circles that are neighboring, disjoint and intermediate, respectively. Then
s+ t+ = s t = 2,
s+ + t+ + u+ = s + t + u = n 2.
We prove the theorem in the next subsection. Until then, let us obtain a stronger
version of the discrete four vertex theorem (Theorem 21.4). 5 Define an extremal
circle in Q to be a neighboring full circle or a neighboring empty circle.
Corollary 21.8. Every generic convex polygon with at least four vertices has at least
four extremal circles.
Clearly, the corollary immediately implies Theorem 21.4. Therefore, all results in
this section follow from Theorem 21.7.
Example 21.9. (Non-convex polygons) Recall that the original four vertex theorem (Theorem 21.1) holds for all simple curves, while the results in this section are stated for convex
polygons. In fact, without extra conditions the discrete four vertex theorem (Theorem 21.4)
is false for non-convex polygons (see Figure 21.2). We present a non-convex version later
in this section (see Theorem 21.17).

196

Figure 21.2. A (non-convex) pentagon with two extremal vertices

and a 12-gon with no extremal circles.
21.3. Proof of Theorem 21.7 via Voronoi diagrams. Let Q = [x1 . . . xn ] be a
generic convex polygon. Consider the Voronoi diagram VD(V ) of the set of vertices
V = {x1 , . . . , xn } (see Figure 21.3 and Subsection 14.2). The cells of VD(V ) are
unbounded convex polygons which consist points which are strictly closer to one
vertex xi than to all others.

Figure 21.3. Voronoi diagram of vertices of a convex polygon, where

broken lines correspond to infinite rays. White nodes and squares correspond to centers of neighboring and disjoint circles, respectively.
Define the cut locus C(V ) to be the complement to VD(V ), i.e., the set of points y
in the plane which are equidistant to some two vertices
xi and xj of Q: |yxi| = |yxj |.
By construction, the cut locus lies in the union of a n2 lines equidistant from pairs
of points.
Let us prove that C(V ) is a plane binary tree (a tree where all vertices have degree 0
or 3) with n infinite rays. For every path from xi to xj , when a point moves from xi
along , by continuity there exists a point in C(V ). This implies that the cut locus
is connected. Observe that the cell containing a vertex xi is unbounded since it must
contain the outer bisector at xi . Moreover, C(V ) has no other cycles since every
point z VD(V ) is closest to some xi , and thus (z, xi ) does not intersect the cut
locus. Therefore, C(V ) is a tree. Note that C(V ) has rays which go to infinity and
are orthogonal at midpoints to intervals (xi , xj ). Since Q is convex, these intervals

197

must be edges of Q, so there are exactly n of them.49 Since Q is generic, no point in

the plane is at equal distance to four or more vertices, which implies that C(V ) is a
binary tree.
There are three types of vertices of the cut locus tree: vertices which are connected
to two, one, or zero infinite rays (there are no vertices connected to three infinite rays
since n 3). Observe that they correspond to empty neighboring, intermediate and
disjoint circles, respectively. The result now follows by induction.
We need to prove that s t = 2 and s + t + u = n 2. The base n = 4 is
trivial: we have s = 2 and t = u = 0. For larger n, remove one of the endpoints.
There are two cases to consider, depending on what kind of vertex the removed vertex
is adjacent to (see Figure 21.4). In both cases the relations are easily satisfied.

Figure 21.4. Inductive step of the proof: removing an endpoint of

the cut locus tree.
Finally, to prove the corresponding relations for full circles, we need to consider the
inverse Voronoi diagram, subdividing points in R2 according to the farthest vertex xi
(see Exercise 14.4). Again, by convexity, there are n infinite rays corresponding to
edges of Q and pointing in the opposite direction to edge normals. The proof extends
verbatim to this case, and we obtain equations on s+ , t+ and u+ as in the theorem.
This completes the proof of Theorem 21.7. 
21.4. A dual version. Note that from a continuous point of view there is no reason
to prefer circumscribed circles to inscribed circlesboth are the same in the limit. As
we show below, there is in fact a precise analogue of the discrete four vertex theorem
and its extensions to this case. The basic idea is to use the standard points to lines
duality on a projective plane. We will not formalize it, but rather use it implicitly
whenever needed (cf. Exercise 2.2).
Let Q = [x1 . . . xn ] R2 , be a convex polygon with edges e1 = (x1 , x2 ), . . . ,
en = (xn , x1 ). Denote by i , 1 i n the lines spanned by the edges ei . We say
that Q is (dually) generic if no point z R2 lies at equal distance to four of these

lines. A circle Rijk

is called inscribed if it is tangent to three lines i , j and k and
lies on the same side of these lines as the polygon.

Denote by Ri = R(i1)i
(i+1) a circle inscribed into three consecutive lines. We say
that edge ei is extremal if Ri either intersects both edges ei2 and ei+2 , or intersects
neither of the two. Now we are ready to state a dual analogue of Theorem 21.4.
49This

is an easy but crucial point in the proof (cf. the proof of Proposition 14.3).

198

Theorem 21.10 (Four edge theorem). Every dually generic convex polygon with at
least four edges has at least four extremal edges.
Corollary 21.11. Let Q = [x1 . . . xn ] R2 be a generic convex polygon, n 4, and
let ri denotes the radius of Ri , 1 i n. Then there are at least four sign changes
in the cyclic sequence (r1 r2 , r2 r3 , . . . , rn r1 ).
Now suppose Q is an equiangular convex polygon, i.e., a polygon with equal angle (n 2)/n. Since larger lengths |ei | correspond to larger radii ri , we obtain the
following.
Corollary 21.12. Let Q = [x1 . . . xn ] R2 be a generic equiangular convex polygon,
n 4, with edges ei = (xi , xi+1 ), 1 i n. Then there are
at least four sign changes
in the cyclic sequence |e1 | |e2 |, |e2 | |e3 |, . . . , |en | |e1 | .
By analogy with the four vertex theorem, Theorem 21.10 can be proved by a
counting all inscribed circles argument (see Exercise 21.4). We will prove the result
later in this section by a different argument.

21.5. A picture is worth a thousand words. Another way to understand the

discrete four vertex theorem (Theorem 21.4) is to interpret it by using the notion of
evolute (sometimes called caustic).
Let Q = [x1 . . . xn ] R2 be a generic convex polygon as above, and let Ri be
the circumscribed circles around the triangle (xi1 xi xi+1 ), 1 i n. Denote by ri
the radius and by Oi the center of Ri . The (discrete) evolute of Q is a polygon
= [O1 . . . On ]. Clearly, the edge (Oi1, Oi ) of lies on a line perpendicular to edge
ei = (xi , xi+1 ) of Q. Orient the edges of away from ei . The vertex Oi of the evolute
is called a cusp if both edges adjacent to Oi are oriented to Oi , or both edges oriented
from Oi .
Corollary 21.13. The evolute of a generic coherent convex polygon with at least four
vertices has at least four cusps.
The corollary follows immediately from Corollary 21.5. To see this, simply observe
that the circle radii increase in the direction of edges in the evolute. Note that from
this point of view it is obvious that the evolute of every (not necessarily convex)
generic polygon has at least two cusps.
Now, one reason to consider the evolute is the visually appealing property that the
between the cusps the evolute is concave, so the cusps are visually distinctive (see
Figure 21.5. In the continuous case this phenomenon is even more pronounced, with
sharp angles at the cusps.
The dual evolute is defined to be a polygon = [O1 . . . On ] with vertices at centers
of inscribed circles. Other details are similar: centers Oi lie on angle bisectors, which
are oriented away from the vertices and the cusps are defined as vertices of with
outdegree zero or two.
Corollary 21.14. The dual evolute of a dually generic convex polygon with at least
four vertices has at least four cusps.

199

Figure 21.5. Evolutes of two convex polygons with four and eight cusps.
In the next subsection we generalize Corollary 21.14 in a rather unexpected direction.
21.6. Relative evolutes of parallel polygons. Let Q = [x1 . . . xn ] be a convex
polygon in a plane. We say that a polygon W = [w1 . . . wn ] is parallel to Q, if
the edges (wi , wi+1 ) are parallel to (xi , xi+1 ), for all 1 i n. We say that W
surrounds Q if Q is inside W , and vectors ui = (xi wi ) turn in the same direction as
points xi as in Figure 21.6.
wi
i
xi

ui
ui+1

i
wi+1

i
i+1

xi+1
Q

zi

Figure 21.6. Polygon W surrounds Q and their relative evolute with

four cusps.
Denote by i the lines spanned by (xi , wi) and let zi be the intersection of i and i+1 ,
1 i n. Define the relative evolute as the polygon = [z1 . . . zn ]. Orient the edges
of away from the vertices xi . As before, define the cusps to be points zi with
with either two ingoing or two outgoing edges. Finally, we say that Q and W are
(relatively) generic if no three lines i , j and k intersect.
Theorem 21.15. Let Q and W be parallel convex polygons which are relatively
generic. Suppose also that W surrounds Q. Then the relative evolute defined
above has at least four cusps.
Let us first show that Theorem 21.15 is a direct generalization of Corollary 21.14.
Denote by i and i the angles of i with edges in Q. Observe that the condition

200

that (wi , wi+1 ) is parallel to (xi , xi+1 ) is equivalent to |ui+1 |/|ui| = sin i / sin i+1 .
Therefore, for a polygon Q and fixed lines i with angles (i , i ), 1 i n, a similar
polygon W exists if and only if
sin 1 . . . sin n = sin 1 . . . sin n .

Similarly, such W surrounds Q if and only if

i + i+1 < , for all 1 i n .

Thus when i are bisectors in Q, there exists a parallel polygon W which surrounds Q,
and the theorem applies. Let us note also that as a consequence of Theorem 21.15,
we obtain a new proof of Theorem 21.10 via the Corollary 21.14.
Proof of Theorem 21.15. Let i = |xi wi |/|zi xi |. Denote by Ti a triangle spanned by ui
and ui+1 and let i be a triangle (zi xi xi+1 ). Now observe that i and Ti are similar
with similarity coefficient i . Therefore, the number of relative cusps as in the theorem
is equal to the number of sign changes in the cyclic sequence (1 2 , 2 3 , . . . , n 1 ).
Since Q and W are generic, the i are distinct and there are at least two sign changes.
Suppose now that there are exactly two sign changes. We can assume that i i+1 > 0
for 1 i k and < 0 for k + 1 i n. Observe that for every point y inside Q we
have:
n
n
n
n
X
X
X
X

(i i+1 ) yxi =
i yxi+1 yxi =
i xi xi+1 =
ui+1 ui = 0.
i=1

i=1

i=1

i=1

Set O to be a point on a line L crossing (xn , x1 ) and (xk , xk+1 ). Then the sum on the
l.h.s. lies on one side of L, a contradiction.

xn

x1
y

xk
xk+1

Figure 21.7. Line L separating positive and negative coefficients (i i+1 ).

21.7. Going into space. The main result of this subsection can be viewed as a generalization of the four vertex theorem. Although simple, it is surprisingly powerful as it
implies an extension of the discrete four vertex theorem in the plane (Theorem 21.4).
Let Q = [x1 . . . xn ] R3 , be a simple space polygon. We say that Q is generic if
no four vertices of Q lie on the same plane. We say that Q is weakly convex 50 if it
50This

notion is dierent and in higher dimensions is less restrictive than the notion of convexity
given in Exercise 2.15.

201

lies on the surface of the convex polytope P = conv(Q). We say that vertex xi is a
support vertex if (xi1 , xi , xi+1 ) is a face in P . In other words, the plane spanned by
edges (xi1 , xi ) and (xi , xi+1 ) is a supporting plane (see Figure 21.8).
Theorem 21.16 (Four support vertex theorem). Every generic weakly convex simple
space polygon with at least four vertices has at least four support vertices.
Proof. Let Q R3 be the polygon as in the theorem. Since Q is generic, the polytope
P = conv(Q) is simplicial. Observe that Q is a Hamiltonian cycle in the graph
of P . Thus, it divides the surface of S = P into two triangulations T1 and T2 , neither
of which have interior vertices. If Q has n 4 vertices, triangulations T1 , T2 have
(n2) 2 triangles. Since the graph dual to T1 is a tree, it has at least two endpoints,
i.e., T1 has at least two triangles (xi1 xi xi+1 ) with two edges Q. This implies that T1
has at least two support vertices. Similarly, T2 also has two support vertices. Finally,
note that these four support vertices must be distinct since otherwise they have degree
two in .

x3

x3
x5

x2

x4

x1
x6

x4

x2

x5

x1
x6

Figure 21.8. Two weakly convex space polygons [x1 . . . x6 ] R3 , one

with four and another with six support vertices.
To see the connection to extremal circles in polygons, assume that a simplicial
convex polytope P R3 as above is inscribed into a sphere. Then a circumscribed
spherical circle around every triangular face (xi1 , xi , xi+1 ) is extremal, i.e., contains
all remaining vertices on one side. Thus, one can think of Corollary 21.8 as a limiting
case of Theorem 21.16, when the circumscribed sphere becomes a plane.
Recall that a polygon Q R2 is generic if no four points lie on a circle. We say
that Q is a Delaunay polygon if it is simple, generic, and for every edge (x, y) Q
there exists a circle R through x and y, such that the remaining vertices are either
all inside of R or all outside of R.51 For example, every generic convex polygon is a
Delaunay polygon.
Theorem 21.17. Every Delaunay polygon in the plane with at least four vertices has
at least four extremal circles.
51The

name comes from Delaunay triangulations and the empty circle condition (see Section 14).

202

Proof. Let Q = [x1 . . . xn ] R2 be a polygon as in the theorem and let Q =

[x1 . . . xn ] S2 be a stereographic projection of Q onto a sphere. Since no four
vertices in Q lie on a line or a circle, we conclude that no four vertices in Q lie on a
spherical circle, and thus on the same plane in R2 . Since Q is a Delaunay polygon,
there exists a circle around every edge which is then mapped into a circle on a sphere,
such that all vertices of Q are on one side of it. If you think of Q R3 as of a
space polygon now, this implies that Q is a generic weakly convex space polygon.
Similarly, if two edges (xi , xi+1 ) and (xj , xj+1) intersect, then all four points lie in the
plane and thus on a spherical circle, which implies that xi , xi+1 , xj , and xj+1 lie on a
circle or a line. Since this is impossible by the assumption that polygon Q is generic,
we conclude that Q is simple. The result now follows from Theorem 21.16.

While Theorem 21.17 is a direct extension of the four vertex theorem (Theorem 21.4), and Corollary 21.8, the analogue of Corollary 21.5 is less obvious (see
Exercise 21.7).
Let Q = [x1 . . . xn ] be a simple polygon in the plane and let ri , 1 i n denotes
the radius of the circumscribed circle Ri as above. Define the curvature i at xi to
be 1/ri if xi is convex and 1/ri if xi is concave.
Corollary 21.18. Let Q = [x1 . . . xn ] R2 be a coherent Delaunay polygon, n 4.
Then there are at least four sign changes in the cyclic sequence of the curvature values:
(1 2 , 2 3 , . . . , n 1 ).
Now, recall that all equilateral polygons are coherent. We conclude that Corollary 21.6 holds verbatim for all (not necessarily weakly convex) simple polygons:

Corollary 21.19. Let Q = [x1 . . . xn ] R2 be an equilateral Delaunay polygon,

n 4. Denote by i = xi1 xi xi+1 the interior angle in Q. Then there are at least
four sign changes in the cyclic sequence (1 2 , 2 3 , . . . , n 1 ).

Finally, let us give a non-convex analogue of Theorem 21.7. Note that the triangular
faces of P = conv(Q) correspond to both empty and full extremal circles. Thus it is
natural to consider circle numbers s = s+ + s , t = t+ + t and u = u+ + u in this
case.
Theorem 21.20. Let Q R2 be a coherent Delaunay polygon with at least four
vertices. Denote by s, t and u the number of neighboring, disjoint and intermediate
circles, respectively. Then
s t = 4,

s + t + u = 2n 4.

Sketch of proof. Start as in the proof of Theorems 21.16 and 21.17, by considering
triangulations T1 , T2 on two sides of Q of the surface S = P , where P = conv(Q).
In each triangulation there are no interior vertices and triangles of three types: with
zero, one and two edges in Q. These triangles correspond to disjoint, intermediate
and neighboring circles, respectively. Since the dual graph to T1 is a binary tree and
the same is true for T2 , we can proceed as in the proof of Theorem 21.7 (compare
Figure 21.3 and Figure 21.9).


203

Figure 21.9. Triangles of three types in a triangulation.

21.8. Further generalizations. The (usual) four vertex theorem has a number of
other interesting generalizations, some of them natural and some quite surprising.
Here are some of them. First, for spherical curves one can define spherical vertices in
a similar manner, by considering osculating spherical circles.
Theorem 21.21 (Spherical four vertex theorem). Every simple smooth curve C on
a sphere S2 has at least four spherical vertices.
The theorem has a straightforward discrete analogue which will be proved and
generalized in the next section. Note that the (spherical) curvature can be strictly
positive on a curve C in a hemisphere S2+ . The following result is an extension of the
spherical four vertex theorem, proving that there are at least two (spherical) vertices
with positive curvature and two vertices with negative curvature under a certain area
condition.
Theorem 21.22 (Tennis ball theorem). Every simple smooth curve C on a sphere S2
which divides the area into two equal parts has at least four inflection points.
The next result is an even stronger generalization, showing that the area condition
is too restrictive.
Theorem 21.23 (Four inflection points theorem). Every simple smooth curve C on
a sphere S2 which is not contained in a closed hemisphere has at least four inflection
points.
If one considers centrally symmetric curves, an even stronger result is known.
Theorem 21.24 (Mobius). Every centrally symmetric simple smooth curve C on a
sphere S2 has at least six inflection points.
Back to the plane, the next result improves the bound on the number of vertices.
Theorem 21.25. Every smooth convex curve C R2 intersecting a circle at least k
times has at least k vertices. In particular, if C touches the smallest circumscribed
circle at least m times, it has at least 2m vertices.

204

Clearly, a circumscribed circle touches the curve in at least two points, so this result
is stronger that the four vertex theorem. To see that the second part follows from the
first part, take a circumscribed circle and shrink it by a small > 0. In fact, a circle
in the theorem can be substituted with any smooth convex curve. Thus one can view
Theorem 21.25 as another example of a relative result, similar to Lemma 9.6 and
Theorem 21.15. Further results of this type will be given in the next section.
Our next extension is an extension to non-simple curves. Think of a simple curve as
the boundary of a 2-dimensional disk embedded into R2 . We say that a curve C R2
bounds an immersed disc if there exists a 2-dimensional disc D immersed into R2 with
boundary C = D.
Theorem 21.26. Every smooth curve C R2 which bounds an immersed disk has
at least four vertices.

Figure 21.10. A smooth curve which bounds an immersed disk.

In a different direction, the four vertex theorem generalizes to contact numbers of
higher order. For a point x on a smooth convex curve C R2 we can consider an
osculating conic, defined to have contact with C of order 4 at point x. We say that x
is a sextactic vertex if the osculating conic has contact with C of order 5 at x.
Theorem 21.27 (Six vertex theorem). Every smooth convex curve has at least six
sextactic vertices.
In conclusion, let us mention that the four vertex theorem has a natural converse.
Theorem 21.28 (Converse four vertex theorem). Let : S1 R be a continuous
function with at least two local maxima and two local minima. Then there is an
embedding : S1 R2 whose curvature at the point (t) is equal to (t).
The results of this type are called existence theorems and play a special role later
in the book. In fact, Theorems 35.4, 36.2, and 37.1 can be viewed as discrete generalizations of the convex case of Theorem 21.28.
21.9. Exercises.
Exercise 21.1. a) [1-] Deduce Corollary 21.5 from Theorem 21.4.
b) [1-] Show that the coherence condition in the corollary is necessary.
c) [1] Deduce Corollary 21.18 from Theorem 21.17.

205

Exercise 21.2. [1-] Let Q = [v1 . . . vn ] R2 be a convex polygon with equal angles.
Suppose
|v1 v2 | |v2 v3 | . . . |vn1 vn | |vn v1 |
Prove that Q is a regular polygon.
Exercise 21.3. [1] Use limit argument to deduce Theorem 21.1 in full generality from
Theorem 21.17.
Exercise 21.4. [1+] The lines i and j are called adjacent if |i j| = 1, i.e., they are
, i < j < k, is called distant if no two lines , ,
spanned by adjacent edges. A circle Rijk
i j k
are adjacent, it is called mediocre if exactly one pair of lines is adjacent, and it is called
is called clear if it does not intersect
close if two pairs of lines are adjacent. The circle Rijk
any other lines, and it is called crossing if it intersects all other lines m , m 6= i, j, k.
Let Q R2 be a dually generic convex polygon with at least four vertices. Denote by s+ ,
t+ and u+ the number of full close, distant and mediocre crossing circles, respectively.
Similarly, denote by s , t and u the number of close, distant and mediocre clear circles,
respectively. Prove the following linear relations:
s+ t+ = s t = 2,

s+ + t+ + u+ = s + t + u = n 2.
Exercise 21.5. [1+] Find a dual version to Theorem 21.15, generalizing Corollary 21.13.
Exercise 21.6. [1] Show that when the number of vertices is odd, the dual evolute and
orientation of its edges uniquely determine the convex polygon.
Exercise 21.7. [1] Deduce Corollary 21.18 from Theorem 21.17.
Exercise 21.8. [1] Use a limit argument to deduce Theorem 21.1 from Corollary 21.6
when the curve is convex. Similarly, use Corollary 21.19 to deduce Theorem 21.1 in full
generality.
Exercise 21.9. [2-] Find and prove a discrete analogue of Theorem 21.21.
Exercise 21.10. [2-] Find and prove a discrete analogue of Theorem 21.22. Show that the
result is false for self-intersecting curves.
Exercise 21.11. [2] Find and prove a discrete analogue of Theorem 21.27.
Exercise 21.12. [2] Recall the cross-ratio of an ordered 4-tuple of distinct numbers, dened
as
(a c)(b d)
[a, b, c, d] =
.
(a b)(c d)
Let (x1 , . . . , xn ) and (y1 , . . . , yn ) be distinct real numbers, and let
i = [xi , xi+1 , xi+2 , xi+3 ],

i = [yi , yi+1 , yi+2 , yi+3 ],

where, all indices are taken modulo n. Prove that a cyclic sequence (1 1 , . . . , n n )
has at least four sign changes.
Exercise 21.13. a) [2-] Prove that every smooth convex curve C in the plane has at least
three pairs of opposite points (points with parallel tangents) which have equal curvature.
b) [] Find a discrete analogue of part a).

206

21.10. Final remarks. The four vertex theorem (Theorem 21.1) is a classical result in
Dierential Geometry, available in numerous textbooks and survey articles (see e.g., [Cher,
Gug1]). It was proved by Mukhopadhyaya for convex curves (1909) and by A. Kneser in
full generality (1912). Theorem 21.2 is due to H. Kneser (1922), the son of A. Kneser, and
Theorem 21.3 is due to Bose (1932).
Interestingly, a related result of M
obius (Theorem 21.24) is much older (1852). For the
history of the four vertex theorem, various extensions and references see [DGPV, Mus4,
Ume, Weg2] and [OT2, Chapter 4].
Corollary 21.6 and Corollary 21.5 for obtuse polygons are due to S. Bilinski (1961, 1963).
The general version of Corollary 21.5 is due to Musin (see [Mus1] for an introduction and an
elementary proof). Theorem 21.7 can be viewed as the most general discrete Bose theorem,
and is a variation on several known results (see e.g., Theorem 1.7 in [Mus4]). The proof of
Theorem 21.7 is new in this form. It uses the cut locus idea in the Rene Thoms proof (1972)
of the four vertex theorem (see [Ume]), subsequently repeatedly rediscovered and adapted
to polygons (see [BanG, Weg2, Weg4]). The four edge theorem (Theorem 21.10) seems to
be new. In computational geometry the cut locus is well studied under the name medial
axis [AurK].
Theorem 21.15 and the proof in Subsection 21.5 are due to Tabachnikov [Tab4]. Theorems 21.16 and 21.17 are due to Sedykh [Sed] (a version of Corollary 21.19 was also discovered by Dahlberg). Interestingly, there is also a two vertex theorem for general Jordan
curves [Hau].
Theorem 21.23 was rst proved by Segre (1968), but a version of it was stated by Blaschke
as an exercise in [Bla1]. Its corollary, the tennis ball theorem (Theorem 21.22) was discovered, aptly named and popularized by Arnold [Arn1, 20]. There is a great deal of literature
on these results and their various generalizations (see [Arn3, pp. 99, 553] for a short survey
and further references).
The rst part of Theorem 21.25 goes back to Blaschke (1916) and Mukhopadhyaya (1931)
(see also [Bla3]). It was rediscovered several times with the denitive version due to Jackson [Jac]52. The second part is an easy corollary of the rst and was popularized in [Oss]).
Similarly, the six vertex theorem (Theorem 21.27) was proved by Mukhopadhyaya in his
original paper (1909), and was repeatedly rediscovered. For more on the history and advanced generalizations, including common generalizations with the M
obius theorem and
connections to classical Cayleys results, see [ThoU].
Theorem 21.26 is due to Pinkall [Pink] who proved it in the generality of all immersed
surfaces. The converse of the four vertex theorem (Theorem 21.28) was proved by Gluck
for convex curves (1971), and by Dahlberg in full generality (1997, published posthumously
in 2005, see [DGPV]). We should mention that much of the diculty in the proof is
analytical and we are not aware of a nontrivial discrete version.

52References to

older results, mostly in German, can be found in the AMS Math. Reviews on [Jac].

207

22. Relative geometry of convex polygons

Here we continue our study of four vertex theorems, shifting in the direction of
their relative geometry. We limit our scope and concentrate only on results which
will be used later on, in the proof of rigidity results of convex polyhedra.
22.1. The LegendreCauchy lemma. We start with the following classical result
which can be viewed as relative versions of the four vertex theorem (Theorem 21.4).
Theorem 22.1 (LegendreCauchy lemma). Let Q = [v1 . . . vn ] and Q = [v1 . . . vn ]
be two convex polygons in the plane with equal corresponding edge lengths: |vi vi+1 | =

|vi vi+1
|. Denote by i = vi1 vi vi+1 and i = vi1
vi vi+1
the angles in the polygons,
where 1 i n and the indices are taken modulo n. Then either there are at least
four sign changes in the cyclic sequence (1 1 , . . . , n n ), or the sequence is
zero.
To see the connection to the four vertex theorem, consider the case of equilateral
polygons Q = [v1 v2 . . . vn ] and Q = [v2 . . . vn v1 ]. The theorem implies in this case
that either all corresponding angles are equal, or there are at least four sign changes
in the sequence (1 2 , . . . , n 1 ). This is a minor variation on Corollary 21.6.
Let us immediately state a spherical version of the theorem, which will prove useful
in the proof of the Cauchy theorem (see Section 26).
Theorem 22.2 (Spherical LegendreCauchy lemma). Let Q = [x1 . . . xn ] and Q =
[x1 . . . xn ] be two spherical convex polygons on a hemisphere S+ with equal corresponding edge lengths: |xi xi+1 |S2 = |xi xi+1 |S2 . Denote by i = vi1 vi vi+1 and i =

vi1
vi vi+1
the spherical angles in the polygons, 1 i n. Then either there are at
least four sign changes in the cyclic sequence (1 1 , . . . , n n ), or the sequence
is zero.
One can think of this result as a generalization of Theorem 22.1 since the plane
polygons are the limits of spherical polygons, as the radius of a sphere tends to infinity.
Later in this section we prove both result by the same general argument.
Note that the condition that both polygons lie inside a hemisphere S+ is necessary,
since otherwise one can find two spherical triangles with equal corresponding sides
and all angles of the first strictly smaller than the corresponding angles of the second
triangle. For example, take any small spherical triangle and its complement =
S2 r . The angles of are < , while the angles of are > . Similarly, as can be
seen already for two quadrilaterals in Figure 22.1, the result is false for non-convex
polygons.
22.2. Making errors at all the right places. The proof of the Legendre-Cauchy
lemma is elementary, but delicate at one point. The original proof, which by now
has become standard, contained a famous error which eluded discovery for nearly
a century and has led to a number of interesting (and correct) results. Thus we
start with the original proof, then point out the famous error, and then correct it.
The reader might want to pay extra attention to the figures and think it over before
learning the answer.

208

x2

y2
-

x3

y3

+
-

y4

x1

y1

x4

Figure 22.1. Non-convex polygons with two sign changes.

We present the proof of only the spherical LegendreCauchy lemma (Theorem 22.2).
The plane version (Theorem 22.1) follows by essentially the same argument. Throughout the proof we will treat spherical geometry almost the same as plane geometry,
relying on readers intuition. This is not where the mistake lies.53
Proof of the LegendreCauchy lemma. Think of the signs as being placed in the vertices of the polygon Q. Now suppose a spherical polygon has exactly two sign changes.
Then there exists a diagonal (y, z) such that on one side of it there are only (+) and (0)
labels, with at least one (+), and on the other side only () and (0) labels, with at
least one (), as in Figure 22.2.

+
z

- 0

Figure 22.2. A spherical polygon Q with exactly two sign changes.

The idea of the proof is to show that if the edge lengths are the same and angles
on one side are increasing, the length of diagonal (y, z) is also increasing. Similarly,
if angles on the other side are decreasing, the length of (y, z) is also decreasing, a
contradiction. Formally, we prove the following result.
Lemma 22.3 (Arm lemma). Let X = [x1 x2 . . . xn ] and X = [x1 x2 . . . xn ] be two
convex spherical polygons in the hemisphere S2+ , such that:
x1 x2 x3 x1 x2 x3 , x2 x3 x4 x2 x3 x4 , . . . , xn2 xn1 xn xn2 xn1 xn ,
and
|x1 x2 | = |x1 x2 |, |x2 x3 | = |x2 x3 |, . . . , |xn1xn | = |xn1 xn |.
53In

fact, if one uses the spherical law of cosines (see Appendix 41.2) in place of the usual law of
cosines that was implicitly used in the proof, one proof easily translates into the other.

209

Then |x1 xn | |xn1 xn |, and the equality holds only if all inequalities between the
angles are equalities. Similarly, if X and X are plane convex polygons with equal
corresponding edge lengths and inequalities on the angles, the same conclusion holds.
Let us first deduce the LegendreCauchy lemma from the arm lemma. Let X be
a polygon on the side of the diagonal (y, z) in Q with positive and zero signs, and
let X be the corresponding polygon in Q , with the diagonal (y , z ) as one of its
sides. Since not all labels are zero, by the arm lemma we have |yz| < |y z |.
Similarly, reverse the role of X and X for the other side of the diagonal. Let X be
the polygon on the side of the diagonal (y, z) in Q with negative and zero signs, and
let X be the corresponding polygon in Q. By the arm lemma we have |yz| > |y z |, a
contradiction. Therefore, having exactly two sign changes is impossible.
It remains to prove that zero sign changes is impossible unless all labels are zero.
Clearly, if all labels are (+) or (0), with at least one (+), we can apply the arm lemma
to any edge to get a contradiction.54 Use the same argument for () and (0) labels.
This completes the proof of the Legendre-Cauchy lemma.

22.3. An incorrect proof of the arm lemma (be vigilant!) Use induction on the
number n of sides of the polygons. The claim is clear for n = 3, when X and X are
spherical triangles in the upper hemisphere. It follows easily from the cosine law on
a sphere (see Proposition 41.3 in the Appendix), and the observation that all angles
are in this case.
Suppose now that n > 3, and assume we have an equality between some of the
angles: xi1 xi xi+1 = xi1 xi xi+1 , for some 1 < i < n. We can simply remove
triangles = (xi1 xi xi+1 ) and = (xi1 xi xi+1 ) from X and X and consider the
remaining polygons (see Figure 22.3). Note that the side lengths and the angle
between them determine the triangles, so = . Thus, the remaining (n 1)-gons
satisfy conditions of the lemma, and the claim follows by the inductive assumption
in this case.
x1

x1

+
xi 0+

xn

xi

xn

Figure 22.3. Removing a zero label in vertex xi of a polygon.

Now suppose all inequalities between the corresponding angles are strict. Start
increasing angle x1 x2 x3 until the angle is equal to the desired value x1 x2 x3 (see
Figure 22.4). Denote by Y = [y1 x2 x3 . . . xn ] the resulting polygon. In a triangle
(x1 x2 x3 ), the lengths of the side (x1 x2 ) and the diagonal (x2 xn ) remain the same, so
54Of

course, in the plane we can use equality of the angle sums of X and X , which removes the
need for the arm lemma in this case.

210

the length |x1 xn | increases: |x1 xn | < |y1xn |. Let us compare polygons Y and X . By
construction, y1 x2 x3 = x1 x2 x3 , by the previous argument |y1 xn | |x1 xn |. This
proves the induction step and establishes the arm lemma. 
y1

x1
x2

xn

+
x3

+
X

x2

xn

0
x3

+
Y

Figure 22.4. Increasing the angle x1 x2 x3 in a spherical polygon.

22.4. An explanation, a discussion, and an idea of correction. Now, before
we show the mistake in the proof let us point out why the mistake must exist. Note
that throughout the proof we occasionally state and always implicitly assume that
the spherical polygons are convex. This, of course, is essential to the proof: without
convexity the claim in the arm lemma claim is false, as shown in Figure 22.5. However,
in the inductive proof above we never used the fact that X is convex. Since the claim
is wrong for non-convex polygons, the proof must also be incorrect.

+
0 0

+
- 0
0

Figure 22.5. Diagonal length |x1 xn | may decrease in a non-convex polygon.

To see the error we need to get into details of the inductive argument. The problem
arises after we increase the angle x1 x2 x3 (see Figure 22.4). Unfortunately, the
resulting polygon Y does not have to be convex, and the induction step fails. In
Figure 22.6 we show a convex spherical polygon which after increase of x1 x2 x3
becomes non-convex, and after subsequent increase of xn xn1 xn2 becomes convex
once again. By the symmetry, there is no order in which these two angles can be
increased without producing a non-convex polygon in between.
Now that we know the mistake, correcting it is not difficult. Here is a natural way,
somewhat involved and educational at the same time. We start to increase the angles
in a convex polygon X = [x1 x2 . . . xn ]. Observe that we can increase the angle in x2
all the way until vertex x1 lies on a line (xn1 xn ). After that, further increase may
give a non-convex polygon, so we do not do it. Now increase the angle in xn1 until
it lies on a line (x1 x2 ). Then increase the angle in x2 , etc. Keep increasing until one
of the angles reaches the desired value. Since all intermediate polygons are convex by

211

y1
x1
x2 +

xn

xn

xn
+

x3

x1

x2

x3

xn1

Figure 22.6. Increasing angle x1 x2 x3 may result in a non-convex polygon.

construction, the distance |x1 xn | will only increase by the argument as in the proof
above.
The trouble with this argument is that it requires an unbounded number of steps (as
long as this number is finite it is not really a problem), and that it takes some work to
show that the increase in the angles does not converge to a value lower than necessary
(see Figure 22.7). This latter part is trickier and uses convexity of X ; it is left to the
reader. Once all the details are set and done, one would really want a different proof
of the arm lemma, somewhat simpler even if perhaps not as straightforward (see also
Subsection 23.5).

Figure 22.7. The iterative process of increasing two angles.

22.5. Alls well that ends well. Based on the first impression one would assume
that the arm lemma is obvious. After some thinking and working out the iterative
argument above one can conclude that it is inherently complicated. Well, one would
be wrong again. Here is a simple and ingenious inductive proof of the arm lemma.
Proof of the arm lemma (for real now! ). Use induction on n. As in the argument
above, we can assume that all angles of X as in the lemma are strictly increasing.
The base n = 3 for triangles is also established. The argument in the original proof
fails if there exists a polygon Y = [y1 x2 . . . xn ] where x1 x2 x3 > y1 x2 x3 x1 x2 x3 ,
and point y1 lies on the (xn1 xn ) line. Apply the base of induction to go from X to Y :
in the triangle (x1 x2 x3 ) we have |y1 xn | |x1 xn |.
Think of Y as (n 1)-gon with one side comprised of two: |y1 xn1 | = |y1 xn | +
|xn1 xn |. Let Z = [x1 x2 . . . xn1 ] be a (n 1)-gon with sides and angles as in the
lemma. By inductive assumption, going from Y to Z, we have |x1 xn1 | |y1xn1 |.

212

Now take point zn on the edge (x1 , xn1 ), such that |zn xn1 | = |xn1 xn |, and think
of Z = [x1 x2 . . . xn1 zn ] as of an n-gon. Clearly, |x1 xn1 | = |x1 zn | + |zn xn1 |, and
|x1 zn | = |x1 xn1 | |zn xn1 | |y1 xn1 | |xn xn1 | = |y1 xn |

Since X is convex, after comparing it with Z we have xn2 xn1 xn xn2 xn1 zn .
Thus the angle in xn1 should be increased. Apply the base of induction to go from Z
to X : in the triangle (x1 xn1 zn ) we have |x1 xn | > |x1 zn |. Putting everything together
we have:
|x1 xn | > |x1 zn | |y1 xn | |x1 xn | ,
as desired.

y1
x1

x2

xn

xn

x2

xn1

xn1

+
+

x3

xn2

x3

xn2
Y

x1

xn

x1
zn

xn1

x2

x2
x3

xn2

x3

xn1

xn2

Figure 22.8. Transformation of polygons X Y Z X .

In Figure 22.8 we show the process of moving from X to X , in the plane for
simplicity. Note that the angle at xn1 here decreases when going from Y to Z; the
interesting feature of the proof is that this does not affect the conclusion.
22.6. The Alexandrov lemma. Just like one can view the LegendreCauchy lemma
(Theorem 22.1) as the relative version of the discrete four vertex theorem (Theorem 21.4), one can ask about the relative version of the dual discrete four vertex
theorem (Theorem 21.10). The following result is an unexpected generalization.
Recall that two convex polygons are called parallel if they have parallel edges. Of
course, parallel polygons have equal corresponding angles.

213

Theorem 22.4 (Alexandrov lemma). Let X = [x1 . . . xn ] and Y = [y1 . . . yn ] be two

parallel convex polygons in the plane. Denote by ai and bi the edge lengths in X and Y ,
respectively. Suppose neither X fits inside Y , nor Y fits inside X by a translation.
Then either there are at least four sign changes in the cyclic sequence (a1 b1 , . . . , an
bn ), or the sequence is zero.
The condition that neither polygon fits inside another may seem unusual, as there
is no such condition in Theorem 22.1. In fact, without this condition the claim is
false (see Figure 22.9).

+
X

+
+

Figure 22.9. Parallel hexagons with exactly two sign changes.

The connection to the dual discrete four vertex theorem is similar to the case of
the LegendreCauchy lemma. Consider an equiangular polygon X = [x1 . . . xn ] and a
polygon Y obtained from X by a clockwise 2/n rotation. If X = Y , the sequence in
the theorem is a zero sequence. Otherwise, since area(X) = area(Y ), neither polygon
fits inside another, and the cyclic sequence (a1 a2 , . . . , an a1 ) has at least four
sign changes. Thus we obtain Corollary 21.12. Here is a more general corollary from
Alexandrovs lemma.
Corollary 22.5. Let X = [x1 . . . xn ] and Y = [y1 . . . yn ] be two parallel convex polygons in the plane. Denote by ai and bi the edge lengths in X and Y , respectively.
Suppose either of the following conditions holds:
(i) area(X) = area(Y ),
(ii) perimeter(X) = perimeter(Y ).
Then either there are at least four sign changes in the cyclic sequence (a1 b1 , . . . , an
bn ), or the sequence is zero.
The follows from the Alexandrov lemma since under either of these conditions
neither polygon fits inside another (for the perimeter see part a of Exercise 7.11). Of
course, one can replace the area and perimeter in (i), (ii) with any other function f
of the polygons which satisfies: f (X) f (Y ) for all X Y , where the equality holds
only if X = Y . For example, take any polynomial of area(X) and perimeter(X) with
positive coefficients. Alternatively, one can use two asymmetric conditions, such as,
e.g., diam(X) < diam(Y ) and width(X) > width(Y ).
The proof of the Alexandrov lemma is elementary but tedious. Instead, we prove
only the second part of the corollary.
Proof of part (ii) of Corollary 22.5. Let e =
xx be an edge vector in X, and let
i

i i+1

u i = e i /ai be a unit vector in direction e i , 1 i n. Since these unit vectors are

214

the same for Y , we have

n
X

ai u i =

i=1

n
X

bi u i = 0.

i=1

Therefore, for every vector w , we obtain

n
n
n
X
X
X
(ai bi )(u i w ) =
(ai bi )u i
(ai bi )w = 0,
i=1

i=1

i=1

where the second sum is zero since X and Y have equal perimeters. The rest of the
proof follows the same argument as in the proof of Theorem 21.15. Clearly, the cyclic
sequence (a1 b1 , . . . , an bn ) is either zero or has at least two sign changes, since the
sum of its elements is zero. Suppose there are exactly two sign changes. Choose w
so that vectors v i = e i w with positive and negative coefficients (ai bi ) lie on
different sides of the line. But then the sum on the l.h.s. in the equation above cannot
be zero, a contradiction.


ui
+

vi

w
ei
xi+1

ui

xi

Figure 22.10. Edge vectors of polygon X (dotted lines separate vectors with positive and negative signs).

22.7. Exercises.
Exercise 22.1. (Alexandrovs local lemma) For a polygon P R3 , denote by i the
edge lengths and by u i the unit outer normals to edges, where i [n]. Write P as the
intersection of halfplanes hx , u i i hi , for some hi R.
a) [1-] Prove that 1 u 1 + . . . + n u n = 0.
b) [1-] Prove that 2area(P ) = 1 h1 + . . . + n hn .
c) [1] Suppose {Pt } is a continuous deformation of P which preserves the area. Prove that
all edge lengths i cannot be increasing.
d) [1] Furthermore, prove that the sequence (1 , . . . , n ) of derivatives is either zero or has
at least four sign changes. Note that this is an immediate consequence of Corollary 22.5,
part (i).
Exercise 22.2. [1+] Prove Alexandrovs lemma (Theorem 22.4).
Exercise 22.3. (Extended arm lemma) a) [2-] Let X = [x1 x2 . . . xn ] R2 be a convex
polygon, and let X = [x1 x2 . . . xn ] be a (possibly non-convex) polygon such that
|x1 x2 | = |x1 x2 |, . . . , |xn1 xn | = |xn1 xn |

215

and
| x1 x2 x3 | | x1 x2 x3 | , . . . , | xn2 xn1 xn | | xn2 xn1 xn | .

Then |x1 xn | |xn1 xn |.

b) [1+] What happens for X R3 ?

Exercise 22.4. (Line development) a) [2-] Let P R3 be a convex polytope and let
A P be a convex polygon on the surface of P , dened as a polygonal region with all
interior angles satisfying 0 i . Let Q = A. Unfold the faces of F containing Q
onto a plane, starting with any one of them, in the order of intersection with Q (note that,
generally speaking, Q might go though the same facet more than once). Prove the unfolding
of Q is not self-intersecting.55
b) [1] Let P R3 be a convex polytope and let L be a plane which does not contain vertices
of P . Consider the curve C = L P . Prove the unfolding of C is not self-intersecting.
22.8. Final remarks. While the arm lemma (Lemma 22.3) seem to be due to Legendre,
it is rarely attributed to him [Sab6]. It became famous in Cauchys original proof of the
rigidity of convex polytopes (see Section 26). In any case, the mistake was discovered and
corrected only by Steinitz in the 1920s (see Steinitzs proof in [Lyu]). The mistake is so
subtle and occurs in such an obvious claim, a number of textbooks continue to repeat
it until this day. Since Steinitz, a number of correct proofs of the arm lemma have been
proposed, some simpler and more elegant than others (compare the proofs in [A2, Ber1,
Hada, Lyu, Sab6, SchoZ, Sto] and try not to be awed). The concise (and correct) proof of
the arm lemma given in Subsection 22.5 was discovered in [SchoZ] (see also [AigZ, 11]).
By now there are a number of papers completely dedicated to the arm lemma and its
generalization (see [Sab6, Schl4] and references in [Con5]). We should mention here that
the Pogorelov lemma (Lemma 28.5) can be viewed as another variation on the arm lemma.
Let us note that the arm lemma is more transparent and easier to prove in the plane than
on a sphere (see e.g. the rst proof in [SchoZ]), thus several authors gave a correct proof of
a plane version, and then mislead the reader by saying that the proof in the spherical case
is almost the same. The reader should be careful with such claims.
The proof of the Alexandrov lemma (Theorem 22.4) is elementary and can be found
in [A2, 6.1] and [Lyu].56 The lemma is used in Section 36 in the elementary proof of the
Minkowski theorem (Theorem 36.2) in R3 (see also Exercise 36.4). The elegant proof of
the innitesimal analogue of part (i) is given in Exercise 22.1 (see [A2, 9.1] and a concise
presentation in Alexandrovs original 1937 article). The proof of part (ii) of Corollary 22.5
is new.

55If

one imagines rolling a die of shape P on a plane, the unfolding of Q is the trace left by Q.
See Section 40 for more on unfoldings.
56A simpler proof was recently found by G
unter Rote (personal communication).

216

23. Global invariants of curves

In this section we concentrate on the local move connectivity method which
we explored earlier in a very different context (see Section 17). We show how one
can deform polygons and prove certain relations based on their invariance under
transitions. Among other things, we sketch how this can be used to prove one of
the four vertex theorem extensions (see Section 21) and outline our own proof of the
square peg problem (see Section 5). We also give an unexpected application to halving
lines, an important subject, which we do not explore elsewhere in the book.
23.1. Counting the intersections. The following result is elementary and can be
proved in a number of ways. We present three different short proofs which give a
good illustration of three flavors of proofs in polyhedral geometry.
Let X = [x1 . . . xn ] R2 be a plane, possibly self-intersecting polygon, and let i
be the line spanned by the edge ei = (xi , xi+1 ), 1 i n. We say that X is generic
if no vertex xr lies on a line i , for all i 6= r 1, r. Denote by a(X) the number of
intersections of lines i and edges ej , i 6= j, j 1, where 1 i, j n. Note that
when X is not simple, the intersections of edges ei and ej are counted twice: once as
i ej and once as ei j .
Theorem 23.1. Let X = [x1 . . . xn ] R2 be a generic plane polygon. Then the
number a(X) of intersections of lines i and edges ej is even.
For convex polygons we have a(X) = 0 and the theorem is trivial. It is less trivial
in the general case. We first present three different proofs of the theorem and then
continue in the next subsections with two generalizations of the result.
First proof. Orient X counterclockwise. An edge ei of X is called convex if the
edges ei1 and ei+1 lie on the left of line i . Similarly, edge ei is concave if ei1
and ei+1 lie on the right of i . Alternatively, we say that ei is a rightward inflection
edge if ei1 lies on the left and ei+1 lies on the right of i . Finally, edge ei is a
leftward inflection edge if ei1 lies on the right and ei+1 lies on the left of i . Clearly,
the leftward and rightward inflection edges alternate. Thus the number of inflection
edges is always even.
Observe that every line i intersects edges in X an even number of times when ei
is convex or concave, and an odd number of times when ei is an inflection edge (see
Figure 23.1). Thus the total number of intersections a(X) is even.


Figure 23.1. Lines through convex, concave and inflection edges.

217

Second proof. Since the number of intersections of two edges of X is counted twice
in a(X), it suffices to count the remaining intersections. Orient X counterclockwise
and denote by i and i two rays spanned by edge ei , along and against orientation,
starting at xi and xi+1 , respectively. Observe that rays i and i1 start at the same
vertex xi and separate the plane into two parts. Therefore, for every i the total
number of intersections of i and i1 and edges in Q is always even (see Figure 23.2).
Summing these numbers, we conclude that a(X) is also even.


ei
X

ei1
xi
i

i1

Figure 23.2. Rays i and i1 intersect an even number of edges in X.

Third proof. Let us deform X = [x1 . . . xn ] into a convex polygon Y = [y1 . . . yn ], by
moving one vertex a time: first move linearly x1 into y1 , then move linearly x2 into y2 ,
etc. By choosing yi in general position the resulting continuous family of polygons is
generic except at a finite number of elementary transitions. There are four types of
these transitions as shown in Figure 23.3. In transitions of the first and fourth type
the number of intersections a(X) changes by two, and in transitions of the second
and third type a(X) is unchanged. This immediately implies the result.


Figure 23.3. Four types of transitions.

23.2. Post-proof analysis. The best things about having several proofs of the same
result is the ability to make a relative comparison of their strengths and weaknesses
when it comes to generalizations. This can often shed some light on the nature of the
result.

218

(1) The first proof is a global counting argument, suggesting that the key to the
theorem is the even number of inflection edges. In fact, this kind of argument works
in other interesting cases, and has analogues in higher dimensions as well (see below).
(2) The second proof is the shortest and the most ingenious of the three. It is
fundamentally a local counting argument, which suggests that the key to the theorem
is a local structure of lines around vertices. Of course, in other cases there is little
hope of having this kind of argument.
(3) The third proof is the most general approach of all. As we shall see later in this
section, it is widely applicable whenever one tries to prove a global enumerative
statement about general curves. On the other hand, the local moves approach gives
the least insight into the nature of the result. It is essentially a verification technique
which gives no hints to potential generalizations.
To underscore the point in (1), let us present the following simple 3-dimensional
result motivated by the proof. A generalization to higher dimensions can be obtained
in a similar way.
Let X = [x1 . . . xn ] R3 be a space polygon. We say that X is generic if no three
vertices lie on the same line and no four vertices lie on the same plane. Denote by Li
the plane spanned by edges ei1 and ei , where ei = (xi , xi+1 ). We say that Li is an
inflection plane if edges ei2 and ei+1 lie on the different sides of Li .
Proposition 23.2. The number of inflection planes of a generic space n-gon in R3
has same parity as n, for all n 4.
In particular, when n is odd, the proposition implies that there is at least one
inflection plane (see Figure 23.4).
x2

x3
x1

x4
x5

Figure 23.4. A space pentagon (drawn on the surface of a hexahedron) with a unique inflection plane L5 = (x4 x5 x1 ).

Proof. Let X = [x1 . . . xn ] R3 be a generic space polygon. Denote by e i =

x
i xi+1
the edge vectors of the polygon, 1 i n, and let i {1} be the sign of the
determinant det(e i1 , e i , e i+1 ). Let i = i1 i , and observe that Li is an inflection

219

plane if and only if i = 1. Clearly,

n
Y
i=1

i =

n
Y

i1 i =

i=1

n
Y
i=1

!2

= 1,

which implies that the number of planes Li that are not inflection planes, is even. 
23.3. Counting double supporting lines. Let X = [x1 . . . xn ] R2 be a plane
polygon. Denote by ei the edge (xi , xi+1 ) and by ij the line through xi and xj . We
say that X is generic if no three vertices lie on a line and no three edges intersect.
Recall that ei is called an inflection edge if edges ei1 and ei+1 lie on different sides
of line i,i+1 . It is easy to see that the number of inflection edges is always even.
Line R2 is called supporting at xi if xi and both edges ei and ei+1 lie on
the same sides of . Line ij through non-adjacent vertices xi and xj is called double
supporting if it supporting at xi and xj . Observe that there are two types of double
supporting lines depending on the direction of the edges adjacent to xi and xj . We
say that ij is exterior if all four edges ei1 , ei , ej1 and ej lie on the same side of ij .
Otherwise, if edges ei1 , ei and ej1 , ej lie on different sides of ij , we say that ij is
interior (see Figure 23.5). Finally, a crossing is a pair of intersecting edges ei , ej X.

Figure 23.5. An example of two exterior and two interior lines in a polygon.
Theorem 23.3 (Fabricius-Bjerres formula). Let X R2 be a generic plane polygon.
Let a(X) be the number of inflection edges, let c(X) be the number of crossings,
and let t0 (X) and t1 (X) the number of exterior and interior double supporting lines,
respectively. Then
a(X)
t0 (X) t1 (X) = c(X) +
.
2
For example, when X is convex, all numbers in the theorem are equal to zero. For
a polygon X in Figure 23.5 we have: a(X) = 6, c(X) = 2, t0 (X) = 8, t1 (X) = 3,
and the theorems states that 8 3 = 2 + 6/2. The proof below is similar to the third
proof of Theorem 23.1, based on deformation of polygons.
Proof. Deform X = [x1 . . . xn ] into a convex polygon Y = [y1 . . . yn ], by moving one
vertex a time (see Subsection 23.1). By choosing yi in general position the resulting
continuous family of polygons is generic except for a finite number of elementary

220

transitions when either three points lie on a line or three lines intersect at a point.
The number of such transition is clearly finite, even if rather large. We summarize
all transitions in Figure 23.6 and Figure 23.7. Now it suffices to check the relation in
the theorem is invariant under these transitions. Since the relation holds for a convex
polygon Y , it also holds for X. This completes the proof.


Figure 23.6. Transitions where some a(X), c(X), t1 (X) and t2 (X) change.

Figure 23.7. Transitions where a(X), c(X), t1(X) and t2 (X) do not change.
One advantage of the local move approach is its flexibility. Let us now present a
straightforward extension of Theorem 23.3 to unions of polygons.
Let Y be a finite union of polygons in the plane. We still assume that no three
vertices lie on a line (and thus no two coincide), but do allow multiple intersections
of edges. Denote by ci (Y ) the number of points where i edges intersect. Define the
number t0 (Y ) and t1 (Y ) of exterior and interior double supporting lines of Y in the
same way.
Theorem 23.4 (Extended Fabricius-Bjerres formula). For a union of polygons Y as
above, we have:
X i
a(Y )
t0 (Y ) t1 (Y ) =
ci (Y ) +
.
2
2
i>1

To prove this extension, start with a nested union of convex polygons, when all
parameters are zero and the formula trivially holds. Applying transformations as in
the proof of Theorem 23.3, we obtain the formula for generic unions of polygons.
When multiple crossings are

allowed, slightly perturb the vertices of Y . Each ii
intersection now creates 2 intersections, which are all counted as in the formula.

221

Before we move on, let us note that Theorem 23.4 is non-trivial already when Y
is a union of convex polygons. In this case there are no inflection edges, i.e., a(Y ) =
0. Note that the interior double supporting lines separate the polygons they are
supporting, while the exterior double supporting lines have them on the same side.
As before, we say that Y is generic if no three vertices lie on a line and no three edges
intersect. We obtain the following.
Corollary 23.5. For a generic union Y of convex polygons, we have:
t0 (Y ) t1 (Y ) = c(Y ).

The corollary can also be proved directly (see Exercise 23.1).

Remark 23.6. While the checking of elementary transitions in the proof above might seem
overly tedious, there are certain advantages in the proof of this type. First, it is automatic
in a sense that the number of transitions is obviously nite and they involve a nite number
of edges. Thus, given a relation one can in principle write a program which veries whether
the relation is invariant under all transitions.
In a dierent direction, one can use elementary transitions to determine (the vector space
of) all relations for any given set of parameters. For that, simply take a quotient of the
space of relations by the vectors spanned by the transition vectors.

23.4. An application: counting k-lines. Let X = {x1 , . . . , xn } R2 be a set of n

points in general position, i.e., such that no three points xi lie on a line and no three
lines through different pairs of points intersect. Consider the set of directed lines ij
through vertices xi and xj , which are directed from xi to xj . Line ij is called a k-line
if it contains exactly k vertices of X on the left. Denote by mk = mk (X) the number
of k-lines in X.
Fix k 0 and consider an oriented graph k on X, where edges (xi , xj ) corresponding to k-lines ij (see Figure 23.8). For example, 0 consists of the boundary of the
convex hull of X, oriented clockwise. Denote by di = di (k ) the number of vertices of
out-degree i. Finally, denote by c = c(k ) the number of intersections of edges in k .

Figure 23.8. Graphs 1 , 2 and 3 on the same set of nine points.

Theorem 23.7. Let X be a set of n points in general position, and let k n/2 1.
Then
X i
(z)
m0 + m1 + . . . + mk1 = c +
di
.
2
i>1

222

Proof. First, observe that

in the graph k the in-degree is equal to the out-degree at every vertex xj . To see
that, start rotating clockwise line ij around xj . The number of edges to the left of
the line will initially drop to k 1, but eventually will increase up to (n k 2) when
the line turns into ji. Clearly, when the line passes vertices on the left of ij , this
number decreases, and when those on the right it increases. Thus, the first time this
number is equal to k happens when the line passes through xj and the vertex on the
right, which gives a correspondence between ingoing and outgoing edges at xj . This
splits edges of k into a union Y of polygons (which may have common vertices and
be self-intersecting).
We need two more observations which follow by the same argument, i.e., by rotating
a line around a vertex. One can check that at every vertex the order of ingoing edges
is the same as the order of corresponding outgoing edges (see Figure 23.9). Further,
at no point xj can there be two pairs of corresponding (ingoing and outgoing) edges
which lie on one side of a line ij .
ij

xj

xi

xj

Figure 23.9. Impossible configurations of edges in k .

We are now ready to apply Theorem 23.4. Since xi are in general position, no three
different lines ij can intersect. Further, by construction of Y the edges always turn
right, so there are no inflection edges. This implies that the r.h.s. of (z) is equal to
t0 (Y ) t1 (Y ). Moreover, by the observation above, all these double supporting lines
must be different.
Now observe that a line ij is a supporting line at i if two edges of the same polygon
in Y lie on the same side of ij . This can happen only if on the other side of ij there
are fewer than k vertices. This implies that there are no interior double supported
lines: t1 (Y ) = 0. In the opposite direction, by rotating the line again one can check
that for every r-line ij , r < k, there exists two pairs of edges in Y at xi and xj , such
that ij is an exterior double supporting line. Thus, the number t0 (Y ) of exterior
double supporting lines is equal to the number of r-lines, 0 r < k.


3
Corollary 23.8. For the number of k-lines, we have mk = O(n k).
The proof of the corollary is outlined in Exercise 23.3.
23.5. Rigid deformations. Let X = [x1 . . . xn ] and Y = [y1 . . . yn ] be two polygons
in the plane with equal edge lengths:
|xi xi+1 | = |yiyi+1 | = ai , where 1 i n.

223

We say that X can be rigidly deformed into Y if there is a continuous family of

polygons Xt = [x1 (t) . . . xn (t)], t [0, 1], such that |xi (t)xi+1 (t)| = ai for all 1 i n,
and X0 = X, X1 = Y . The family {Xt , t [0, 1]} is called a rigid deformation.
Theorem 23.9. Every two polygons with edge lengths (a1 , . . . , an ) can be rigidly
deformed into each other, unless ai + aj > p/2, ai + ak > p/2, and aj + ak > p/2, for
some i < j < k, where p = a1 + . . . + an .
The inequalities in the theorem are necessary, since, for example, a polygon in
Figure 23.10 cannot be deformed into its inverse. More generally, consider a polygon
linkage L with given edges length (see Section 13), and let ML be its realization
space. The theorem says that ML is connected unless these inequalities hold. In
fact, in the latter case, it is known that the space ML has two connected components
separating polygons and their inverses.

Figure 23.10. A polygon which cannot be deformed into its inverse.

We define a simple rigid deformation to be a rigid deformation {Xt }, where all Xt
are simple. Clearly, a simple deformation preserves orientation of a polygon. The
following result is the analogue of Theorem 23.9 for simple polygons.
Theorem 23.10 (Carpenters rule problem). Every two simple polygons with the
same corresponding edge lengths and the same orientation can be rigidly deformed
into each other by a family of simple polygons.
Just like in the previous two sections, this result can be used to prove a version
of the four vertex theorem (Theorem 21.1) for all equilateral (but not necessarily
convex) polygons.
Sketch of proof of Corollary 21.19. Denote by N = s t the difference in the number
of neighboring and disjoint cycles. If we can prove that N = 4 for a specific simple
polygon and that N does not change under the rigid deformation, Theorem 23.10
implies the result. Observe that N does not change while the polygon is generic.
Check that we can always deform one simple polygon into another without creating
more than one degeneracy at a time: line with three vertices or circle with four
vertices. Call a transition through one degeneracy a local move. Going over all
possible local moves as in Figure 23.11, check that s t is indeed invariant under
rigid deformations. It remains to check that N = 4 for a specific polygon.


224

R2

R1

Figure 23.11. Rigid deformation of a polygon: a neighboring circle R1 and a disjoint circle R2 disappear after a local move.
23.6. Back to inscribed squares. We conclude this section with another proof of
Theorem 5.12, which says that every simple polygon in the plane has an inscribed
square.57 In the spirit of this section, we prove this result by deforming the polygon,
but first we need to find a global relation which holds for (almost all) simple polygons.
This relation is simple: every generic simple polygon has an odd number of inscribed
squares. Now that we have the relation, we can try to prove that it is invariant under
certain elementary transitions.
Theorem 23.11. Every generic simple polygon has an odd number of inscribed
squares.58
Theorem 5.12 now follows by a straightforward limit argument, as we repeatedly
did in Section 5. Note also that the theorem is false for all simple polygons; for
example every right triangle has exactly two inscribed squares. We begin the proof
with the following simple statement, which will prove crucial (cf. Exercise 5.17).
Lemma 23.12. Let 1 , 2 , 3 and 4 be four lines in R2 in general position. Then there
exists a unique square A = [a1 a2 a3 a4 ] such that xi i and A is oriented clockwise.
Moreover, the map (1 , 2 , 3 , 4 ) (a1 , a2 , a3 , a4 ) is continuously differentiable, where
defined.
Proof. Fix z1 1 . Rotate 4 around z1 by /2, and denote by 4 the resulting line,
and by z2 = 2 4 the intersection point. Except when 2 4 , such z2 is unique.
Denote by z4 4 the inverse rotation of z2 around z1 . We obtain the right isosceles
triangle = (z2 z1 z4 ) oriented clockwise in the plane. The fourth vertex z3 of a
square is uniquely determined. Start moving z1 along 1 and observe that the locus
of z3 is a line, which we denote by 3 . Since line 3 is in general position with respect
to 3 , these two line intersect at a unique point x3 , i.e., determines uniquely the
square [a1 a2 a3 a4 ] as in the theorem. The second part follows immediately from the
above construction.

57Of

course, the result extends to general (self-intersecting) polygons, but for technical reasons it
is easier to work with simple polygons.
58It takes some eort to clarify what we mean by a generic polygon (see the proof). For now, the
reader can read this as saying that in the space R2n of n-gons, almost all are generic.

225

Sketch of proof of Theorem 23.11. We begin with the following restatement of the
second part of the lemma. Let X = [x1 . . . xn ] be a generic simple polygon and let
{Xt , t [0, 1]} be its continuous piecewise linear deformation. Suppose A = [a1 a2 a3 a4 ]
is an inscribed square with vertices ai at different edges of X, and none at the vertices
of X, i.e., ai 6= xj . Then, for sufficiently small t, there exists a continuous deformation
{At } of inscribed squares, i.e., squares At inscribed into Xt . Moreover, for sufficiently
small t, the vertices ai of At move monotonically along the edges of Xt .
Consider what can happen to inscribed squares At as t increases. First, we may
have some non-generic polygon Xs , where such a square in non-unique or undefined.
Note that the latter case is impossible, since by compactness we can always define a
limiting square As . If the piecewise linear deformation {Xt } is chosen generically, it
is linear at time s, and we can extend the deformation of At beyond As .
The second obstacle is more delicate and occurs when the vertex ai of square As
is at a vertex v = xj of Xs . Clearly, we can no longer deform As beyond this point.
Denote by e1 the edge of X which contains vertices ai of At for t < s. Clearly,
e1 = (xj1 , xj ) or e1 = (xj , xj+1 ). Denote by e1 the other edge adjacent to v. Denote
by e2 , e3 and e4 the other three edges of X containing vertices of At (see Figure 23.12).
e1

e2

xj
e1

At

As
e3

Br
e4

Figure 23.12. Inscribed squares At , As = Bs and Br , where t < s <

r. Here e2 , e3 and e4 are fixed, while e1 and e1 move away from the
squares.
Now consider a family {Bt } of squares inscribed into lines spanned by edges e1 , e2 , e3
and e4 . By construction, As = Bs . There are two possibilities: either the corresponding vertex bi approaches xj from inside e1 or from the outside, when t s and t < s.
In the former case, we conclude that the number of inscribed squares decreases by 2
as t passes through s. In the latter case, one square appears and one disappears, so
the parity of the number of squares remains the same. In summary, the parity of the
number of squares inscribed into Xt with vertices at different edges is invariant under
the deformation.
It remains to show that one can always deform the polygon X in such a way that
at no point in the deformation do there exist inscribed squares with more than one
vertex at the same edge, and such that the resulting polygon has an odd number of
inscribed squares.
Fix a triangulation T of X. Find a triangle in T with two edges the edges of X
and one edge a diagonal in X. Subdivide the edges of X into small edges, so that
neither of the new vertices is a vertex of an inscribed square. If the edge length is now
small enough, we can guarantee that no square with two vertices at the same edge

226

is inscribed into X. Now move the edges along two sides of the triangle toward
the diagonal as shown in Figure 23.13. Repeat the procedure. At the end we obtain
a polygon Z with edges close to an interval. Observe that Z has a unique inscribed
square (see Figure 23.13). This finishes the proof.


Z
X

Figure 23.13. The first step of the polygon deformation which preserves the parity of the number of inscribed squares; the final polygon Z.
23.7. Exercises.
Exercise 23.1. [1] Prove Corollary 23.5 by a direct argument.
Exercise 23.2. [1] Suppose a closed curve C R2 has no triple intersections and a winding
number n around the origin. Prove that C has at least n 1 intersections.
Exercise 23.3. a) [1+] Let G be a graph with n vertices and m 4n edges drawn on
a plane with c crossings. For a xed probability p, compute the expected number of edges
in a random p-subgraph of G. Use Eulers formula to show that there exists on average
at least one crossing when p is suciently large. Compare this to the average number of
crossing in a p-subgraph. Optimize for p to prove that c m3 /64n2 .
b) [1+] In the notation of Theorem 23.7, prove that m0 + .
. . + mk = O(k n). Conclude from
here that c = O(k n) and use part a) to obtain mk = O(n 3 k) (see Corollary 23.8).
Exercise 23.4. [1+] Find all local moves described in Subsection 23.5 and complete the
proof of Corollary 21.19.
Exercise 23.5. a) [1-] For the case of numbers a(X), c(X), t0 (X) and t1 (X) as above,
show that the relation in Theorem 23.3 is the only possible relation with these parameters.
b) [1] Prove that every nonnegative integer 4-tuple (t0 , t1 , c, a) Z4+ which satises a 2
and the Fabricius-Bjerre formula can be realized by a polygon X R2 .
c) [2-] In notations of Theorem 23.3, prove that if a(X) = 0, then t1 (X) is even and satises
t1 (X) c(X)2 c(X).
d) [1+] Prove that every nonnegative integer 4-tuple (t0 , t1 , c, 0) Z4+ which satises the
Fabricius-Bjerre formula and the inequality in part c), can be realized by a polygon X R2 .

Exercise 23.6. [2-] Let Q = [x1 . . . xn ] R3 be a simple space polygon such that no four
vertices lie on the same plane. Denote by Lijk the plane plane containing vertices xi , xj
and xk , 1 i < j < k n. Note that L can be oriented according to the orientation of
the triangle (xi , xj , xk ) L. We say that L is tritangent if two edges adjacent to each of
the vertices xi , xj and xk lie either on L side of L. Classify all planes Lijk by the number
of edges on it and by the side (positive or negative) on which adjacent edges lie. Find all
linear relations for the resulting numbers.59
59For

the weakly convex polygons some of these numbers were studied in Subsection 21.7. However, in the full generality not all relations remain true.

227

Exercise 23.7. [2-] Use the stereographic projection to obtain the analogues of results in
the previous exercise for the numbers of neighboring and disjoint circumscribed circles (see
Subsection 21.7).
Exercise 23.8. a) [1+] Find a space polygon in R3 isotopic to the trefoil knot, and without
tritangent planes.
b) [2-] Let Q be space polygon in R3 isotopic to the trefoil knot, and such that Q projects
onto a plane with the usual diagram with three crossings. Prove that Q has a tritangent
plane.
Exercise 23.9. [2-] Let X = [x1 . . . xn ] R2 be a polygon with vertices in general position,
and let ei = (xi , xi+1 ) be the edges of X. An interval u = (xi , y), y X, is called quasinormal at xi if both angles of u and ei1 , ei are acute or both are obtuse. Dene a sign
(u) {1} of a normal u to be 1 if the angles are acute and 1 if the angles are obtuse.
A diagonal (xi , xj ) is called a double quasi-normal if it is quasi-normal at both vertices
(see Subsection 9.4). It is called positive if (xi , xj )(xj , xi ) = 1; otherwise, it is called
negative. Denote by n+ (X) and n (X) the number of positive and negative double quasinormals. Dene a perpendicular to be an interval u = (xi , y) such that y ej for some ej
and (xi , y)ej . Denote by p+ (X) and p (X) the number of perpendiculars u such that
(u) = 1 and (u) = 1, respectively. Use deformations of polygons to prove that
n+ (X) n (X) + p+ (X) p (X) + c(X) = 0,

where c(X) is the number of crossings of X.

Exercise 23.10. Let C R2 be an oriented closed piecewise linear curve with n double
intersections and no triple intersections. A loop is a closed portion of curve without double
points.
a) [1] Note that C has at most 2n loops. Prove that if C has exactly 2n loops, then between
every two points on C corresponding to a double crossing there are exactly n curve arcs.
Such curves are called maximally looped.
b) [1+] Suppose C has at least n + 1 loops. Prove that n is odd.
c) [1+] Denote by r(C) the sum of ( i ), over all angles i on the left of the curve. Prove
that r(C) = 2(n+ n 1), where n+ and n is the number of positive and negative
double points, respectively.
d) [1] Use this to prove that if C is maximally looped, then r(C) {4, 0, 4}.
d) [1] Use part c) of Exercise 23.5 to show that if C is maximally looped and has no inection
edges, then C has no interior double supporting lines.
e) [1-] Use Theorem 23.3 to conclude that every curve C as in part d) has exactly n double
supporting lines.
Exercise 23.11. [2-] Find a simple unknotted space polygon Q R3 which cannot be
rigidly deformed into a at convex polygon. In other words, prove that Theorem 23.10
does not extend to R3 .
Exercise 23.12. [1+] Prove that every two simple polygons in R4 , can be rigidly deformed
into each other. In other words, prove that Theorem 23.10 extends to R4 .
Exercise 23.13. (Erd
os ip problem) [2] Let Q R2 be a non-convex polygon and let C
be the convex hull of Q. Choose edge e of C that is not in Q and reect the portion of Q
across e, as in Figure 23.14. Repeat such ip transformations until a convex polygon is
obtained. Prove that the process stops after nitely many steps.

228

Figure 23.14. First few flip transformations of a polygon.

Exercise 23.14. [1+] Let Q R2 be a self-intersecting polygon. Dene a reection move
to be a reection of an arc [xy] of Q along the line (x, y), where x, y Q are any points
on the polygon (see Figure 23.15). Prove that every polygon can be made simple (not
self-intersecting) by a nite sequence of reection moves.

Figure 23.15. A sequence of reflection moves of a polygon.

23.8. Final remarks. Theorem 23.1 and the rst two proofs are given in [VasE], Problem 235. The original smooth analogue of Theorem 23.3 is due to Fabricius-Bjerre; this
version and the proof follows [Ban2] (see also [Fab]). A complete description of the values
in the Fabricius-Bjerre formula is given in Exercise 23.5. For a precise statement in Remark 23.6 see [Pak3] which discusses various local move connectivity arguments for nite
tilings.
The bound in Corollary 23.8 on the number of k-lines is due to Dey (1998). Theorem 23.7
implying it (see Exercise 23.3) was proved in [And+]. A dierent, more combinatorial proof
of the theorem was given in [AroW].60 The proof of Theorem 23.7 we present here is due to
Uli Wagner (unpublished), who graciously allowed us to use it here (see also a related proof
in [And+]). Let us note that all these proofs use basic properties of graphs k , discovered
by Lov
asz in [Lov].
Theorem 23.9 was proved independently in [KM1] and [LenW]. For other versions of
this results in the literature see e.g., [Whit]. Theorem 23.10 is known as carpenters rule
problem and was proved in [CDR] (see also [Stre]).
The deformation idea in the proof of Theorem 23.1 is due to Shnirelman [Shn] (see
also [Gug2]), although the details are largely dierent. Our presentation follows [Pak9] (see
Subsection 5.8 for further results on inscribed squares).

60This

proof can be viewed as a combination of the argument we give and the Fabricius-Bjerres
proof in [Fab] applied to the special case of polygons with no inection edges and interior double
supporting lines.

229

24. Geometry of space curves

In this mostly stand-alone section we finish our exploration of the geometry of space
curves. Much of this section is aimed at the proof of the FaryMilnor theorem and
related results. Although these results are standard in differential geometry, their
proofs are completely discrete and use important technical tools. Let us single out
the averaging technique in Croftons formula (see also Exercises 24.524.7), and knot
decomposition in the second hull theorem.
24.1. Curve covers. Let C Rd be a curve of length L. It is obvious that C can
be covered by a ball of radius L/2. The following result shows that this bound can
be improved to a sharp bound of L/4.
Theorem 24.1. Every space polygon Q Rd of length L can be covered by a ball of
radius L/4.
The theorem is a d-dimensional generalization of Exercise 1.3. By the Helly theorem
(Theorem 1.2) it suffices to check this for all (d + 1)-gons in Rd . Here is a simple
direct proof.
Proof. Let v, w be two points on Q which divide the length into equal halves. For the
midpoint z of the interval (v, w) and any point x Q we have:




1
1
1
L
|zx| =
vx + wx
|vx| + |wx|
L/2 = .
2
2
2
4
Thus, the ball of radius L/4 centered at z covers Q.

The following is a spherical analogue of Theorem 24.1. The proof above can also
be modified to work in this case.
Theorem 24.2. Every spherical polygon Q Sd , d 2, of length L 2 can be
covered by a spherical ball of radius L/4.
Proof. As in the proof above, let v, w be two points on Q which divide Q into paths
of equal length L/2, and let z be the midpoint of the arc of a great circle from v
to w. For every point x Q, denote by x Sd a point on a sphere such that z is the
midpoint of the arc of a great circle from x to x . By the symmetry, we have:




1
1
1
|xz|Sd = |xx |Sd
|xv|Sd + |vx |Sd =
|xv|Sd + |xw|Sd .
2
2
2
Since the geodesic distance |xv|Sd + |xw|Sd L/2, we obtain the result.61

We will need the following special case of the theorem.

Corollary 24.3. Every spherical polygon Q Sd , d 2, of length L < 2 can be

covered by an open hemisphere. Similarly, every spherical polygon Q Sd , d 2, of
length L = 2 can be covered by a closed hemisphere.
61Note

that we never used condition L 2, which is added for simplicity. The midpoints of
intervals longer than are ambiguous and the spherical balls of radius more than /4 are less
natural.

230

24.2. Total curvature. For a triangle (xyz), the exterior angle (xyz) at a vertex y
is defined as
(xyz) = xyz .

Let Q Rd be a simple (non-self-intersecting) space polygon. The total curvature

(Q) is the sum of the exterior angles of Q. This definition is robust enough so that
when polygons approximate a smooth closed curve C Rd , the total curvature of
the polygons converges to the total curvature of the curve (C).
The following beautiful result is the first step in the understanding of total curvature.
Theorem 24.4 (Fenchel). For every simple space polygon Q R3 , the total curvature
(Q) 2. The equality holds if and only Q is flat and convex.
Proof. Let Q = [v1 . . . vn ] R3 . For every edge vector e1 = (v1 v2 ), . . . , ed1 =
(vn1 vn ), en = (vn v1 ), consider a point on a unit sphere xi S2 so that (Oxi) is
a unit vector along ei . Connect points xi and xi+1 with the shortest path, so that
R = [x1 . . . xn ] is a spherical polygon. By definition, the exterior angle (vi1 vi vi+1 )
of Q is equal to the length |xi xi+1 |S2 on a sphere. Thus (Q) = |R|S2 , the length of
the spherical polygon.
Now observe that R cannot lie in an (open) hemisphere. Indeed, otherwise all edge
vectors ei are pointing into the same half-space, and since the sum of ei is equal to O,
this is impossible. On the other hand, by Corollary 24.3, any curve on a sphere of
length < 2 can be covered by a hemisphere. Therefore |R|S2 2, which proves the
first part of the theorem.
For the second part, from above and the second part of Corollary 24.3, the equality
can occur only when Q lies in a plane. Finally, for the plane convex polygons the
result follows from a direct calculation of the interior angles.


24.3. The F
aryMilnor theorem. In the previous section we proved Fenchels theorem that every curve in R3 has total curvature at least 2. The main result of this
section is a lower bound of 4 for knotted curves.
We say that a space polygon Q R3 is knotted if it is simple and knotted as a
closed curve.
Theorem 24.5 (FaryMilnor). For every knotted space polygon Q R3 , the total
curvature (Q) 4.
The proof is again based on spherical geometry. Let C S2 be a spherical polygonal
curve on a unit sphere. For every point x S2 , denote by nC (x) the number of points
of intersection of C and a plane Hx orthogonal to (Ox). The values of nC subdivide
the sphere into spherical polygons A1 , . . . , Ak with constant value i on Ai . Define
Z
k
X
NC =
i area(Ai ) =
nC (x) dx .
i=1

S2

231

Lemma 24.6 (Croftons formula). The length |C| of a spherical polygonal curve
C S2 satisfies
1
|C| = NC
4
Proof. Both sides are clearly additive, so it suffices to prove the lemma only for the
intervals on a sphere. Suppose C S2 is an interval, |C| = . Then nC (x) = 1
when x S2 lies on a great circle with normal (Oz) for some z C, and nC (x) = 0
otherwise. The set A of points x with nC (x) = 1 is a union of two sectors with
angle . We have:
2
NC = area(A) =
4 = 4 ,
2
which completes the proof.

We are now ready to prove the FaryMilnor theorem.
Proof of Theorem 24.5. For a closed knotted space polygon Q R3 , fix an orientation
and consider the spherical polygon R S2 as in the proof of Theorem 24.4. Then
(Q) = |R|. For a spherical polygon R, consider a subdivision A1 , . . . , Ak of a sphere
and the corresponding values 1 , . . . , k . If i 4 for all i, we have |R| = NQ /4 4,
as desired.
Suppose now that i < 4. Choose a point x Ai in general position, so no edge
of Q is parallel to Hx . Recall that i is equal to the number of intersections of R
and Hx , which in turn is equal to the number of changes in direction of edges of Q
with respect to Hx . Therefore, i = 2. The result now follows from the assumption
that Q is unknotted and the following Milnors lemma.

Lemma 24.7 (Milnors lemma). Let Q R3 be a simple space polygon such that the
distance function to a plane L has exactly two critical points. Then Q is unknotted.
Proof. Denote by x and y the points of Q with the minimum and maximal distance,
respectively. Then x, y divide Q into two paths where the distance from L is monotone
increasing. Consider a projection of Q onto a generic plane HL as in Figure 24.1.
Undo all crossings, from x to y to obtain a projection without crossings. This implies
that Q is an unknot.


Figure 24.1. Projection of Q onto a plane orthogonal to L.

232

24.4. Linked tubes. Consider two linked simple space curves C1 and C2 in R3 of
length 2. The question is how big a tube can be made around the curves so that
these tubes do not intersect. It is intuitively obvious that the tubes cannot have
radius bigger than 1/2. This radius is achieved when two curves are circles which lie
in the orthogonal planes, and such that each circle passes through the center of the
other (see Figure 24.2).

Figure 24.2. Two linked circles with maximal distance between them.
The following result is a discrete version of this observation. For two space polygons
Q1 , Q2 R3 denote by


d(Q1 , Q2 ) = min |xy|, x Q1 , y Q2
the distance between them.

Theorem 24.8. For every two linked space polygons Q1 , Q2 R3 the distance
d(Q1 , Q2 ) between satisfies:


1
d(Q1 , Q2 )
min |Q1 |, |Q2 | .
2
Proof. Fix a point x Q1 and a surface A spanned by intervals (x, x ), where x Q1 .
Since polygons are linked, Q2 intersects A. Fix a point y Q2 A and consider the
largest ball Br of radius r around y, such that Q1 is outside Br . Then the projection R
of Q1 onto the sphere S = Br has length |R| |Q1 |. Since y A, the projection R is
a spherical polygon which contains two opposite points. Since the spherical distance
between the opposite points is equal to , we have |R| 2 r.62 Therefore, for the
radius r we have r 21 |Q1 |, which implies the result.

24.5. Inscribed space polygons. The main goal of this subsection is to give a
combinatorial proof of the FaryMilnor theorem (Theorem 24.5). We begin with an
important technical lemma.
We say that a space polygon X is inscribed into a space polygon Q, where X, Q
3
R , if the vertices of X lie on the edges of Q.
Lemma 24.9 (Monotonicity of the total curvature). Let Q, X R3 be space polygons,
such that X is inscribed into Q. Then (X) (Q).
62This

also follows immediately from Theorem 24.2.

233

Proof. Add vertices of X to Q and observe that the curvature (Q) does not change.
One by one, remove the vertices of Q that are not in X and check that the curvature
is nonincreasing.
Formally, let Q = [v1 . . . vn ]. Remove vertex vi and replace edges (vi1 , vi ) and
(vi , vi+1 ) with (vi1 , vi+1 ). Denote by Q the resulting space polygon. Let us show
that (Q ) (Q). We have:
(Q) (Q ) = ( vi2 vi1 vi ) + ( vi1 vi vi+1 ) + ( vi vi+1 vi+2 )
( vi2 vi1 vi+1 ) ( vi1 vi+1 vi+2 )

= ( vi vi1 vi+1 + vi2 vi1 vi+1 vi2 vi1 vi )

+ ( vi vi+1 vi1 + vi1 vi+1 vi+2 vi vi+1 vi+2 ) 0,

where the last inequality is the spherical triangle inequality applied to each term. 
vi
vi1

vi+1
vi+2

vi2
Figure 24.3. Removing vertex vi from space polygon Q = [v1 . . . vn ].
Here is a simple idea of another proof of the FaryMilnor theorem. Starting with
a knotted polygon Q R3 , consider a sequence of space polygons inscribed into
each other. If we can show that the last polygon in a sequence has total curvature
4, we obtain the desired lower bound. In fact, we will prove that one can take
the last inscribed polygon to be a degenerated quadrilateral D = [xyxy] which has
(D) = 4.
Combinatorial proof of Theorem 24.5. We prove the theorem by induction on the
number n of vertices of a knotted space polygon Q = [v1 . . . vn ] R3 . When n = 3,
polygon Q is a triangle. Then Q is unknotted and the claim trivially holds.
By the continuity of the total curvature, we can always assume that the vertices
of Q are in general position. We start with the following observation. Suppose there
exists a vertex vi such that triangle (vi1 vi vi+1 ) does not intersect any other edge of Q.
We can then remove vertex vi as in the proof above. Since the resulting polygon is
still knotted, we obtain the step of induction.
We now begin to change polygon Q as follows. Starting at v1 , consider all points
z (v2 , v3 ), z 6= v2 , such that (v1 , z) intersects an edge in Q. If this set of such points
is empty, we can remove the triangle as above. Otherwise, let z2 be the closest such
point to v2 . Since Q is generic, we have z2 6= v3 . Denote by y1 the corresponding
point of intersection of (v1 , z2 ) and Q (see Figure 24.4).
Move z2 slightly towards v2 . Denote by x1 the closest to y1 point on a new
(v1 , z2 ). Now replace vertex v2 of Q with z2 , i.e., replace edge (v1 , v2 ) and interval

234

v2

z2

v2

v3

y1

x1

v1

v3

y1

v1

Figure 24.4. Transforming space polygon Q = [v1 . . . vn ].

(v2 , z2 ) with (v1 , z2 ). The resulting polygon Q remains knotted, but has points x1
and y1 very close to each other.
Repeat the procedure for the (new) vertex v2 . Again, replace vertex v3 with z3
(v3 , v4 ), and obtain two points x2 (v2 , v3 ) and y2 Q very close to each other.
There are two possibilities for y2 relative to the points x1 , y1 Q. If y2 is on the arc
[y1 . . . x1 ] of the polygon, then the inscribed quadrilateral [x1 x2 y1 y2 ] is very close to
a degenerated quadrilateral D as above, and can be made as close as desired.
Alternatively, if y2 is on the arc [x1 . . . y1 ] of the polygon, then the chords (x1 , y1 )
and (x2 , y2 ) as in Figure 24.5 do not intersect. Repeat the procedure until two intersecting chords are found. That will always happen since by construction the points
(xi , yi) cannot lie on adjacent edges. This completes the induction step and proves
the theorem.

x2

y2

v2
x1

y1

v1

y2

Figure 24.5. Two possibilities for y2 .

24.6. Second hull of a polygon. The goal of this section is to present the following
unusual combinatorial generalization of Milnors lemma (Lemma 24.7).
Let Q R3 be a simple space polygon and let P = conv(Q) be its convex hull.
Observe that every plane H through an interior point x P intersects Q in at least
two points. We say that a plane is generic if it contains no vertices of Q. Define
the second hull sh(Q) to be the set of points x P such that every generic plane H
through x intersects Q in at least four points. The second hull sh(Q) can be empty,

235

for example when Q projects onto a plane without intersections. Thus it is natural
to ask what happens for knotted polygons.
Theorem 24.10 (Second hull theorem). For every knotted space polygon Q R3 ,
the second hull sh(Q) is nonempty.

Figure 24.6. Second hull of the trefoil knot.

To see that the second hull theorem generalizes Milnors lemma, let us restate them
in the same language. Milnors lemma says that if for a space polygon Q there exists
a family of parallel planes through every point in R3 which intersects Q at most twice,
then Q is unknotted. The second hull theorem makes the same conclusion without
the restriction on all planes to be parallel.
Let us mention that sh(Q) can be disconnected, but is always a disjoint union
of convex polytopes, except at the boundary (see Exercise 24.9). Also, the second
hull theorem has no natural inverse, i.e., there exist unknots Q with sh(Q) 6= (see
Exercise 24.10).
Proof of Theorem 24.10. For two points x, y Q consider a polygon Q obtained by
replacing the portion of Q between x and y by a straight interval (x, y). Observe that
sh(Q ) sh(Q). Indeed, if a generic plane H through a point in sh(Q ) intersects Q
four times, it contains at least three of them in Q (x, y) Q. Since the number of
intersections of H and Q is even, it must be at least four.
From now on we assume that Q is geometrically prime, i.e., that there are no
planes which cut Q into two knotted arcs. Otherwise, we can consider a knotted
space polygon Q instead and reduce the problem to a polygon with fewer edges.
Let F R3 be a (closed) half-space on one side of plane H. We say that F
essentially contains Q if it either contains Q, or intersects Q at exactly two points x, y
such that the arc outside of F union (x, y) is unknotted. In this case, define the
clipping of Q by F to be clipF (Q) = (F Q) (x, y) the isotopic polygon inside F .
Let F, G be two half-spaces which essentially contain Q. We shall prove that G
essentially contains the clipping clipF (Q). Note that there are three possible relative
positions of points x, y Q F and points x , y Q G (see Figure 24.7).
In the first case, we have x , y
/ F . Then clipF (Q) G and the claim holds by
definition. In the second case, we have x , y F . Since the portion between x and y
in Q is unknotted, so is the arc between x and y in clipF (Q) and the claim holds by
definition again. Finally, in the third case, we have x
/ F and y
/ G. Then planes F

236

and G divide Q into four arcs, such that the sum of two outside of F and the sum
of two outside of G are unknotted. Since the sum of two non-trivial knots cannot
be unknotted (see Exercise 24.3), the arc A outside F G is unknotted. Therefore,
clipF (Q) consists of A inside G and an unknotted arc outside G, which proves that G
essentially contains clipF (Q) in this case as well.

y
G

x
x

y
F

Figure 24.7. Clipping of knots: three cases.

We are now ready to finish the proof. Denote by F = F (Q) the set of all closed
half-spaces which essentially contain Q, and let Z = F F F be the intersection of all
such half-spaces. We claim that Z is a nonempty convex set, such that Z sh(Q).
For the second part, consider a generic plane H through any point z Z and assume
that H intersects Q at two points. Since Q is geometrically prime, the arcs on both
sides cannot be knotted. On the other hand, if either arc is unknotted, take the
half-space F which lies in the other side of H, such that the plane F is at distance
from H. For > 0 small enough, the half-space F essentially contains Q, but z
/ F,
a contradiction with the choice of z Z. Thus H intersects Q in at least three points.
Since H is generic, it has at least four intersection points with Q then.
By definition, the set Z is the intersection of half-spaces F F . Since the halfspaces are convex and closed, by the Helly theorem (Theorem 1.2 and Exercise 1.1),
to prove that Z is nonempty it suffices to show that the intersection of every four
half-spaces F1 , . . . , F4 F is nonempty. Consider a sequence of knotted polygons
Q0 , . . . , Q4 , defined by Q0 = Q, and Qi = clipFi (Qi1 ), 1 i 4. From above,
by induction, plane F1 essentially contains Q0 , plane F2 essentially contains Q1 , etc.
Therefore, Qi are well defined knotted space polygon. Since Q4 F1 F2 F3 F4 ,
we conclude that this intersection is nonempty, which finishes the proof.

24.7. Exercises.
Exercise 24.1. a) [1-] Let X R2 be a set of n points in general position. Prove that
there exist a simple (not self-intersecting) polygon with vertices in X.
b) [1-] Let X R3 be a set of n points in general position. Prove that there exists an
unknotted polygon with vertices in X.
c) [1-] Let X R3 be a set of 3n points in general position. Prove that there exist n
disjoint triangles with vertices in X, such that no two triangles are linked.
d) [1+] Let X R3 be a set of 6 points in general position. Prove that there exist 2 linked
disjoint triangles with vertices in X.

237

Exercise 24.2. (Local convexity of polygons) [1] Prove that every simple polygon Q R2
with interior angles < is convex. Prove a spherical analogue.
Exercise 24.3. (Sum of knots) [2-] Prove that the sum of two non-trivial knots cannot be
an unknot.
Exercise 24.4. (Borsuk ) [1] Generalize Fenchels Theorem 24.4 to curves in Rd .
Exercise 24.5. (Croftons formula for surfaces) [1] Let S R3 be a non-convex 2dimensional polyhedral surface
P with the set of vertices V = {v1 , . . . , vn }. Dene the total
absolute curvature (S) = i |i | the sum of absolute values of curvatures i = (vi ) of
vertices (see Section 25). Generalize Croftons formula (Lemma 24.6) to show that (S) is
2 times the average number of critical points of linear functions on S.
Exercise 24.6. (Buons needle) Consider an innite family L of horizontal lines in the
plane which stand apart at distance 1 from each other. Let C be a rectiable curve in R2 .
Denote by N (C) the number of intersections of C and lines in L. Dene = (, b) to be a
rigid motion of a plane obtained as a rotation by angle and then a translation by vector b.
Denote by B an axis parallel unit square. Finally, let


N (C) = E N (C) ,

where = (, b), [0, 2] is a random angle, and b B is a random vector. Heuristically,

think of N (C) as the average number of intersections of C and a random rigid motion of C.
a) [1] Prove that the average N (C) is invariant under rigid motions. Prove that N (C) is
additive when a union of two curves is taken. Check that N (C) is continuous as curve C is
deformed.
b) [1-] Observe that N (C) = 2 when C is a circle of diameter 1. Conclude from here that
N (C) = 1/ when C is a unit interval.63 More generally, prove that N (C) = |C|/ for
every C.
c) [1-] Observe that N (C) = 2 when C is a curve of constant width 1 (see Exercise 3.6).
Conclude that |C| = .
Exercise 24.7. (Random intersections) Let R be the set of rigid motions (, b) as above.
Dene the kinematic density d = | sin | ddb. For two convex sets X, Y R2 , denote by
N (X, Y ) the number of points in the intersection X Y . Fix a convex set X and let mi be
the measure of the set of A such that N (X, X) = i. Finally, denote = perimeter(X)
and a = area(X).
a) [2-] Prove that
2 4 a = m4 + 2m6 + 3m8 + . . .
Conclude from here the isoperimetric inequality in the plane (Theorem 7.1).
b) [1] Consider a subset A(X) R of rigid motions (, b) such that N (X, X) 4. Use
part a) to show that if X is a convex set not equal to a circle, then the subset A(X) has
positive kinematic measure.
c) [1] Give an independent proof of part b) using Lemma 9.6.
Exercise 24.8. (Wienholtz ) a) [1+] Prove that every polygon Q R2 of length |Q| has a
projection onto a line of length at most |Q|/.
b) [2-] Prove that every space polygon Q R3 of length |Q| has a projection onto a plane
which has diameter |Q|/.
63This

result, well known under the name Buons needle, gives a Monte Carlo method for approximating the number .

238

c) [2] Prove that every space polygon Q R3 of length |Q| has a projection onto a plane
which lies in a circle or diameter |Q|/.
Exercise 24.9. a) [1] Prove that the closure of every connected component of sh(Q) is
a convex polytope.
b) [1-] Let L R3 be a disjoint union of space polygons which form a nontrivial link.
Extend the denition of the second hull to links. Prove that sh(L) 6= .
Exercise 24.10. a) [1-] Find a simple space polygon in R3 with a disjoint second hull.
b) [1-] Find an unknotted space polygon in R3 with a nonempty second hull.
c) [1] Find an unknotted space polygon in R3 such that its second hull contains a unit ball.
Exercise 24.11. [2+] In the conditions of the second hull theorem (Theorem 24.10), prove
that the closure of sh(Q) intersects Q.
Exercise 24.12. [2] Let L be a link with n components such that every two of them
are linked. Prove that there exists a point x conv(L) such that every plane through x
intersects at least n/2 components of L.
Exercise 24.13. (Alexandrov ) [1+] Let C be a polygonal curve in R2 with endpoints x, y.
Dene the total curvature = (C) to be the sum of exterior angles in the vertices of C.
Suppose (C) < . Prove that
|xy|
|C|
,
cos
and the equality holds only when C has two edges of equal length and angle ( ) between
them. Extend the result to Rd .
Exercise 24.14. a) [1+] Let C be a polygonal curve in R2 with endpoints x, y. Suppose C
does not contain the origin O. Dene the visibility angle (C) to be the length of a projection
of C onto a unit circle centered at O. Denote by , the angles between (Ox), (Oy) and C.
Prove that
()
(C) (C) + .
b) [1+] Prove the analogue of a) for paths in R3 . Formally, let C be a polygonal curve in R3
with endpoints x, y, and such that O
/ C. Let (C) be the length of a projection of C onto
a unit sphere centered at O, and let , be the angles between (Ox), (Oy) and C. Prove
that () holds. Conclude that (Q) (Q) for every (closed, simple) space polygon.
c) [1] Use the second hull theorem (Theorem 24.10) to prove that for every knotted space
polygon Q R3 there exists a translation Q of Q such that (Q ) 4. Use part b) to
obtain the F
aryMilnor theorem (Theorem 24.5).
Exercise 24.15. [1+] Let Q R3 be a knotted space polygon. Prove that there exist three
distinct points x, y, z Q such that y is a midpoint of (x, z).

Exercise 24.16. (Quadrisecants of knots) Let Q R3 be a knotted space polygon.

a) [1-] Prove that every point in Q lies on a line which intersects Q in at least two other
points.
b) [2] Prove that there exists a line which intersects Q in at least four points lying on
dierent edges. Such quadruples of points are called quadrisecants.
c) [2+] Prove that there there exists a quadrisecant with the order of points on a line
alternating with the order of points on Q. These are called alternating quadrisecants.
Conclude from here the F
aryMilnor theorem (Theorem 24.5).

239

Exercise 24.17. (F
ary theorem and generalizations) a) [1+] Let C Br be a general
(possibly non-convex and self-intersecting) polygon in the plane, where Br is a disk of
radius r > 0. Denote by |C| the length of C and by (C) the total curvature of C. Prove
that |C| < r (C).
b) [2-] Let C Br be a polygon in Rd , where Br is a ball of radius r > 0. As before,
let (C) be the total curvature of C. Prove that |C| < r (C).
c) [2] Let C, Q R2 be two polygons in the plane such that Q is convex and C is inside Q.
Prove that the average total curvature is nonincreasing:
(C)
(Q)

.
|C|
|Q|

Moreover, the equality holds only when C is a multiple of Q.

Exercise 24.18. Let S1 , S2 Rd be two PL-homeomorphic PL-surfaces. Dene the Frechet
distance
distF (C1 , C2 ) = inf max |x(x)|,
xC1

where the inmum is over all piecewise linear homeomorphisms : S1 S2 .

a) [1+] Let C1 , C2 R2 be two convex polygons in the plane. Prove that
|C1 | |C2 | 2 distF (C1 , C2 ).

b) [2-] Let C1 , C2 R2 be two simple polygons in the plane. Prove that



|C1 | |C2 | (C1 ) + (C2 ) 2 distF (C1 , C2 ).

c) [2] Let C1 , C2 R3 be two simple space polygons. Prove that



|C1 | |C2 | (C1 ) + (C2 ) 2 distF (C1 , C2 ).

d) [1-] Check that part a) of Exercise 24.17 follows from here, while part b) does not.
b the graph on
Exercise 24.19. (Curvature of graphs in R3 ) For a graph G denote by G
the same set of vertices, and with twice as many edges between every two vertices. For an
b dene
embedding : G , where R3 , and an Eulerian circuit C in G,


1
C () = (C) .
2
This notion is called the net total curvature of .
a) [2-] Find the analogue of Croftons formula (Lemma 24.6) for the net total curvature
b) [1+] Dene the Kotzig transformation on Eulerian circuits by changing the direction on
a closed segment of a circuit. Prove that every two Eulerian circuits C1 , C2 in G can be
obtained from each other by a nite sequence of Kotzig transformations.
c) [1+] Use parts a) and b) to prove that the net total curvature is independent of the
Eulerian circuit.
Exercise 24.20. Let R3 be a union of three non-intersecting space polygonal curves
between two points x, y. Let () be the sum of all exterior angles between pairs of adjacent
edges, where the exterior angles in x and y are weighted with 1/2.
a) [1] Check that () is equal to the net total curvature of .
b) [1+] Prove that () 3.
c) [2-] Let G be a graph with two vertices and three edges between them, and let be an
embedding of G in R3 . Prove that if () < 4, then the is unknotted.

240

Exercise 24.21. a) [2-] Prove that the total length of a net (of inextensible strings) which
holds a sphere is at least 3.64
b) [1] Show that the bound in a) is tight.
c) [2] Suppose a net holds a sphere even if any k links are broken. Prove that the total
length of such net is at least (2k + 3).
24.8. Final remarks. Theorem 24.1 goes back to Segre (1937) and was subsequently rediscovered several times. This proof is given in [ChaK]. Theorem 24.4 was proved by Fenchel
in 1929, and was extended by Borsuk in 1947 (see Exercise 24.4). A continuous version of
the proof above was given by Liebmann and later rediscovered by Horn (see [Horn] for a
simple proof and references).
We refer to [Ada2, Sos] for an accessible introduction to knot theory. Theorem 24.5 was
discovered independently by F
ary and Milnor. In our proof we follow the sketch in [Cher],
which is close to the original proof in [Mil1]. The linked tube problem in Subsection 24.4
was proposed by F. Gehring in 1973. The proof we present is due to M. Ortel [Can+] (see
also [EdeS]). We refer to [Can+] for various extensions and further references.
The combinatorial proof of the F
aryMilnor theorem presented in Subsection 24.5 is due
to Alexander and Bishop [ABi]. They give additional arguments extending the result to
general curves and making the inequality in the theorem strict. The result on alternating
quadrisecants in the Exercise 24.16 is due to Denne [Den] and can be viewed as yet another
proof of the F
aryMilnor theorem.
The second hull theorem (Theorem 24.10) is proved in [CKKS], where various extensions
and applications are also presented. These include the extension to nontrivial links (see
Exercise 24.9) and another connection to the F
aryMilnor theorem (see Exercise 24.14).

64The

precise denition is a non-trivial part of the exercise. For a simple example of a net which
2
holds a sphere, take vertices of an equilateral tetrahedron inscribed
q into S and connect them by

geodesics. The total length of these geodesics is equal to 12 arcsin

2
3

3.65.

241

25. Geometry of convex polyhedra: basic results

We begin the study of polyhedral surfaces with two most basic results: Eulers
formula and the GaussBonnet theorem. We use bits and pieces of spherical geometry
and apply the results to closed geodesics on convex polyhedra (see Section 10).
25.1. Eulers formula. Let P R3 be a convex polytope, and let V, E, F denote
the set of vertices, edges, and faces of P , respectively. The classical Eulers formula
in this case is the following result.
Theorem 25.1 (Eulers formula). |V | |E| + |F | = 2.
Even though there is a straightforward inductive proof, the spherical geometry
proof below will be a helpful introduction to the method. Note also that in Section 8
we already proved the much more general Dehn-Sommerville equations for all simple
polytopes (Theorem 8.1).
Proof. Move P so that it contains the origin O and consider the unit sphere S2 R3
with the center at O. Consider a radial projection of the surface P onto S2 . In other
words, for every edge (v, w) E draw an intersection of S2 and a (planar) cone over
(v, w) starting from O. This gives an arc (v , w ) S2 of a great circle. Clearly, every
k-gonal face in P becomes a spherical k-gon.
Let F F be an i-gonal face of the polytope, let F denotes the corresponding
spherical i-gon, and let (F ) be the sum of the angles of F . By Girards formula for
polygons (see Theorem 41.2) we have:
(F ) = area(F ) + (i 2) .
Denote by fi the number of i-gonal faces in F and sum the above equation over all
faces F F . Since the sum of the angles on a sphere in the vertex image v is 2,
for all v V , we have:
X
X
2 |V | = area(S2 ) +
fi (i 2) = 4 2 |F | +
i fi .
P

On the other hand, i i fi = 2 |E|, since every edge belongs to exactly two faces.
Substituting this into above equation gives Eulers formula.


Of course, Eulers formula holds in much greater generality, e.g., for all planar
maps, or in higher dimensions (see Section 8). Here is a simple corollary which will
prove useful.
Corollary 25.2. Every simplicial polytope with n vertices has 3n6 edges and 2n4
faces. More generally, this holds for every triangulated surface homeomorphic to a
sphere S2 .
Proof. Suppose P has k faces. Then the number of edges is 23 k, and by Eulers formula
k 32 k + n = 2. Solving this for k gives the result.


242

25.2. GaussBonnet theorem. Let P R3 be a convex polytope with the set of

vertices V = {v1 , . . . , vn }. Denote by i = (vi ) the sum of the face angles around vi
and let i = 2 i be the (Gaussian) curvature of vi . The sum of curvatures at all
vertices is called the total curvature of the polytope.
Theorem 25.3 (GaussBonnet). Let 1 , . . . , n be the curvatures of vertices of a
convex polytope P R3 . Then 1 + . . . + n = 4 .
Observe that i is an intrinsic rather than extrinsic parameter of vertices of a closed
convex surface S = P . In view of the theorem, one can think of i /4 as a point
concentrated measure on S. Here is a straightforward proof of the theorem based on
a counting argument. A more enlightening proof based on spherical geometry is given
in the next subsection.
Proof. Triangulate the faces of P . By Corollary 25.2, the resulting triangulation has
2n 4 triangles. Thus, the total sum of face angles A = (2n 4). We conclude
1 + . . . + n = 2 n A = 4 .


25.3. Spherical geometry proof. Now we present a different proof of the Gauss
Bonnet theorem (Theorem 25.3), based on spherical geometry. Not only this proof is
more insightful and from the book, it also leads to some useful extensions.
For every vertex vi V of polytope P , consider all planes H containing vi , such
that the whole polytope lies on one side of H. Such planes are said to be supporting
vertex vi . Taking normals uH (unit vectors H) these planes define a region Ri on a
unit sphere S2 . Clearly, these regions are disjoint (except for the boundary) and they
cover the whole sphere: S2 = i Ri . Indeed, taking any plane H and moving it from
infinity towards the polytope must hit some (perhaps more than one) vertex first (see
Figure 25.1).
uH
vi

S2

uH

S2

vi

Ri

Figure 25.1. Normal uH to plane H. Spherical k-gon Ri S2 corresponding to vertex vi of the polytope.
Suppose vi is a vertex of degree k (i.e., has k adjacent edges). Denote by 1 , . . . , k
the corresponding face angles. Observe that Ri is a spherical k-gon with sides orthogonal to projections of edges onto S2 (see Figure 25.1). Therefore, polygon Ri has

243

angles 1 , . . . , k and by Girards formula for spherical k-gons we have:

area(Ri ) =

k
X
j=1

( j ) (k 2) = 2 i = i .

From here, for the total area of a sphere we have:

2

4 = area(S ) =

n
X

area(Ri ) =

i=1

n
X

i ,

i=1

which finishes the proof of the GaussBonnet theorem.

25.4. GaussBonnet theorem in higher dimensions. One useful thing about a

good proof is its robustness and ability to work in generalizations. It is even better
if the proof leads to the right definitions and then becomes a complete triviality.
This is what happens with the generalization of the GaussBonnet theorem to higher
dimensions.
Let C Rd be a cone starting at the origin O, and let S = Sd1 Rd be a unit
sphere centered at O. Define the solid angle (C) to be (d 1)-dimensional volume
of the intersection:
(C) := area(C S) = vold1 (C S).
Now, let P Rd be a convex polytope with the set of vertices V = {v1 , . . . , vn }.
Consider cones Ci starting in the vertices vi V of the polytope and spanned by
the edges. As the proof above shows, the right measure of Ci is not (Ci ), but
rather (Ci ), the solid angle of the dual cone defined as follows.
The dual cone Ci is the cone of the normals uH to the hyperplanes H Rd
supporting vi , i.e., hyperplanes H containing vi and having P in one half-space of H
(as in Figure 25.1 above). Define the curvature of a cone (C) = (C ), and let
i = (Ci) = (Ci ). Finally, denote by d the (d 1)-dimensional volume of the
unit sphere: d = area(Sd1 ).
Theorem 25.4 (GaussBonnet in Rd ). Let 1 , . . . , n be the curvatures of vertices
of a convex polytope P Rd . Then: 1 + . . . + n = d .
The proof is completely straightforward and follows the steps in the proof above.
Consider the regions Ri = Ci S, and check that
n
X
i=1

We omit the details.

i =

n
X
i=1

area(Ri ) = area(S) = d .

244

25.5. Infinite polyhedra. Let P Rd be a (possibly unbounded) convex polyhedron

defined as an intersection of a finite number of half-spaces. Convex polytopes and
cones are examples of convex polyhedra. Below we show how to extend GaussBonnet
theorem to convex polyhedra in every dimension.
We assume that convex polyhedron P contains the origin O. Consider all rays
(half-lines) starting at O and going to infinity and contained inside the polytope:
P . Denote by CP the cone spanned by these rays. It is easy to see that CP is
unchanged under translations of P , still containing O. Cone CP is called the base
cone of P . Define the curvatures i of vertices vi of P as for the polytopes. Similarly,
let (CP ) = (CP ) be the curvature of the base cone.
Theorem 25.5 (Total curvature of convex polyhedra). Let 1 , . . . , n be the curvatures of vertices of a convex polyhedron P R3 . Then: 1 + . . . + n = (CP ).
In particular, the theorem implies that (CP ) depends only on P and is invariant
under the rigid motions. The proof is straightforward once again. Note that the
spherical region RP = CP S corresponds to the set normals of hyperplanes supporting
vertices of the polyhedron P .
Note that (CP ) is well defined even when P is degenerate, i.e., has dimension < d.
For example, suppose polyhedron P R3 has a base cone CP which consists of a
single ray. Such convex polyhedra are called convex caps. From the theorem, for the
total curvature of convex caps we have (CP ) = area(S2 )/2 = 2. More generally,
for all infinite convex polyhedra P R3 , we have (CP ) 2, where the equality
holds only for convex caps.

25.6. Back to closed geodesics. Let S be the surface of a convex polytope P

R3 . A surface polygon M = [w1 , . . . , wk ] is a collection of shortest paths (wi , wi+1 ),
i [k 1], on the surface S. By Proposition 10.1, the shortest paths may cross edges
but not vertices, except when wi V . In the intrinsic geometry of S the edges of P
are invisible, in a sense that the metric is flat along them.
As before, polygon M is simple if it is not self-intersecting, and is closed if w1 = wk .
Denote by A the region on one side of M. Denote by (A) the sum of curvatures of
vertices vi A. Recall that M is a closed geodesic on S if only if M does not contain
vertices of P and all surface angles of A at wi are equal to (see Section 10.2). The
following result formalizes and extends the results on closed geodesics obtained earlier
in Section 9.
Theorem 25.6 (Total curvature of simple closed geodesics). Let M be a simple closed
geodesic on the surface S = P of a convex polytope. Then the total curvature on
each side is equal to 2.
Proof. Suppose M S is a surface k-gon and a region A on one side has interior
vertices. A triangulation of A has


m = 2(k + ) 4 (k 2) = k + 2 2 triangles.

245

Thus, the total sum (A) of the face and boundary angles in A satisfies (A) = m.
On the other hand, summing over all vertices, we have (A) = 2 (A) k .
Together these equations imply the result.

The proposition implies that if there is no way to split the total curvature of all
vertices into two equal parts, then there are no simple closed geodesics (see Section 9).
Here is another result in this direction.
Corollary 25.7. An infinite convex polyhedron P R3 has a simple closed geodesic
if and only if P is a convex cap.
Proof. The corollary follows immediately from Theorem 25.5 and the observations
above: 1 + . . . + n = (CP ) < 2 unless P is a convex cap. To prove the second
part, suppose P is a convex cap. Consider a hyperplane H orthogonal to a ray CP ,
and such that all vertices of P lie on one side of H. Then the intersection M = H P
is a closed geodesic.

The following is a somewhat unexpected application of the proposition.
Theorem 25.8 (Cohn-Vossen). Two simple closed geodesics on a surface of a convex
polytope either intersect or have equal lengths.
For example, a hexagon and a square shown in Figure 10.4 are two closed geodesics
on a cube of different length, and they do, in fact, intersect. The theorem also allows
an easy calculation of the lengths of geodesics. For example, in a regular tetrahedron
in the same figure, the rectangular geodesics can be sled down towards an edge. Thus
its length is twice the edge length.
Proof. Let P be a convex polytope, and let M1 , M2 be two simple closed geodesics
on the surface S of P . As before, let V be the set of vertices of P . Assume that M1
and M2 do not intersect. Then polygons M1 and M2 divide the surface S into three
regions: B1 , B2 and C, such that B1 = M1 , B2 = M2 , and C = M1 M2 . By
proposition, the sum of curvatures inside B1 and B2 is equal to 2, so by the Gauss
Bonnet theorem region C does not have interior points.
In the language of the graph G = (V, E) of P , observe that the sets of interior
vertices V1 = V (B1 ) and V2 = V (B2 ) are disjoint, nonempty and contain all graph
vertices: V = V1 V2 . Thus, there exists an edge (x, y) E, such that x V1 ,
y V2 Since both M1 and M2 must intersect (x, y), denote by z1 , z2 points of intersection. Now cut the region C with an interval (z1 , z2 ), and unfold C on a plane
(see Figure 25.2). Each geodesic becomes a straight interval (z1 , z1 ) and (z2 , z2 ),
where (z1 , z2 ) used to be identified with (z1 , z2 ). Thus, the unfolding is a plane 4-gon
Z = [z1 z1 z2 z2 ] with |z1 z2 | = |z1 z2 |. Since M1 and M2 are both geodesics, we have:
z1 z1 z2 + z1 z1 z2 = z1 z2 z2 + z1 z2 z2 = .
Therefore, Z is a parallelogram and geodesics M1 and M2 have equal length.

246

z2
z1

M2
M1

z2

z2
z1

z1

Figure 25.2. Two non-intersecting closed geodesics M1 and M2 on a

cube, and an unfolding Z of the region between them.
25.7. Exercises.
Exercise 25.1. [1] Extend Eulers formula and the GaussBonnet theorem to general
closed polyhedral surfaces (of any genus).
Exercise 25.2. a [1] Let S = P be the surface of a convex polytope P , and let A S be
a polygonal region. Denote by (A) the sum of the interior surface angles at the boundary
vertices. Suppose the boundary M = A contains r surface polygons, and k vertices. Then:
X
(A) = (n + 2r 4) +
i .
vi A

b) [1] Deduce the GaussBonnet theorem from here.

Exercise 25.3. (From GaussBonnet back to Euler ) [1] Consider a polygonal region A
P obtained as an -neighborhood of all edges of the polytope (see Figure 25.3), for > 0
small enough. Apply the previous exercise to A and derive Eulers formula.

A
P

Figure 25.3. Region A obtained as an -neighborhood of edges of P .

Exercise 25.4. [1-] Let Cv Rd be a convex cone with vertex v. Check that the dual
cone satises: Cv = {w R3 | hw, vi 0 for all v C}.

Exercise 25.5. (Monotonicity of cone curvature) a) [1-] Suppose C, D Rd are convex

cones such that C D. Prove that (C) (D).
b) [1] Moreover, (C) = (D) only if the cone C is a translation of D.
c) [1-] Suppose P, Q Rd are convex polyhedra such that P Q. Prove that (P ) (Q).

Exercise 25.6. a) [1] Suppose all faces of a convex polytope P R3 are centrally symmetric. Prove that P has at least eight simple vertices (vertices adjacent to exactly three
edges). Note that this is tight since the cube has exactly eight simple vertices.
b) [1-] Prove that there is no upper bound of the number of faces.
Exercise 25.7. [1] Let P R3 be a convex polytope whose vertices have degrees 4.
Prove that it has at least eight triangular faces.

247

Exercise 25.8. Let P R3 be a convex polytope. Denote by c(P ) the sum of the number
of triangular faces and the number of vertices of degree 3.
a) [1-] Check that the graph-theoretic proof in Subsection 11.2 implies that c(P ) 1.
b) [1] Use Eulers formula to prove that c(P ) 8.
Exercise 25.9. a) [1] Suppose the surface S of a convex polytope P R3 can be decomposed into a nite number of rectangles. Prove that P has at most eight vertices.
b) [1] Suppose S can be decomposed into a nite number of rectangles. Prove that P has
at most twelve vertices.
Exercise 25.10. (de Guas formula) [1] Let P R3 be a convex polytope. Denote by A
and B the sums of all solid and dihedral angles, respectively. Prove that
2B A = 2 (m 2),
where m is the number of faces in P .
Exercise 25.11. Let P R3 be a convex polytope containing the origin O. For a facet F
of P , denote by (F ) the sum of the angles of F and by (F ) the sum of the angles of the
projection of F onto a unit sphere centered at O. Finally, let (F ) = (F ) (F ). Prove
that
X
(F ) = 4 .
F P

Exercise 25.12. (Equihedral tetrahedra) [1+] Let R3 be a tetrahedron. Prove that

the following conditions are equivalent:
(i) all faces of are congruent triangles,
(ii) all faces have equal perimeter,
(iii) all vertices have equal curvature,
(iv) the opposite edges have equal dihedral angles,
(v) all solid angles are equal.
Such are called equihedral tetrahedra.
Exercise 25.13. [1+] The edges of a convex polytope P R3 are colored with two colors.
Prove that there exists a vertex in P such that the adjacent edges either have the same
color or have two color intervals (in cyclic order).
Exercise 25.14. [1] Let be the graph of a convex polytope P R3 and let O be an
orientation of edges. As always, let V and F denote the set of vertices and faces of P . For a
vertex v V , denote by ind(v) to be 1 c(v)/2, where c(v) is the number of changes in the
orientation of edges adjacent to v (in cyclic order). Similarly, for a face F F, denote by
ind(F ) to be 1 c(F )/2, where c(F ) is the number of changes in the orientation (clockwise
vs. counterclockwise) of edges of F . Prove that
X
X
ind(v) +
ind(F ) = 2.
vV

F F

Exercise 25.15. (Discrete PoincareHopf index theorem) a) [1] Let P R3 be a convex

polytope. We say that an orientation of edges of P is balanced if every vertex has at least one
ingoing and one outgoing edge (i.e., the oriented graph of P has no sinks and no sources).
Prove that the edges of at least two faces of P form oriented cycles.
b) [1] Generalize part a) to orientable 2-dimensional polyhedral surfaces of higher genus.

248

Exercise 25.16. [1] Prove that the inverse to Theorem 25.6 does not hold, i.e., construct
a convex polytope P such that a subset of vertices has sum of curvatures equal to 2, but
which has no simple closed geodesics.
Exercise 25.17. [1+] Let P R3 be an unbounded polyhedron with (CP ) > 0. Prove
that between every two points in S = P there is only a nite number of geodesics.
Exercise 25.18. Let denote the largest dihedral angle in a convex polytope P R3
with n vertices.
a) [1] Suppose < /2. Prove that P is a tetrahedron.
b) [1] Suppose /2. Prove that n 8.
c) [1+] Suppose , for some > 0. Disprove: n is bounded as a function of .
Exercise 25.19. [1+] Suppose integers n, m, k > 0 satisfy n m + k = 2, n 2k 4, and
k 2n 4. Give a direct construction of a convex polytope in R3 with n vertices, m edges
and k faces.
Exercise 25.20. (Eberhards theorem) a) [1] Use Eulers formula to show that for every
simple polytope P R3 , we have:
X
(6 i) fi (P ) = 12,
i3

where fi (P ) is the number of faces with k sides.

b) [2] Suppose (f3 , f4 , . . .) is an integer sequence which satises the above equation. Use
the Steinitz theorem (Theorem 11.1) to prove that there exists a simple polytope P R3 ,
such that fi (P ) = fi for all i 6= 6.
25.8. Final remarks. Our presentation of Eulers and GaussBonnet theorems is a variation on a standard theme (see [SCY, 57], [Hop2] and [Ber1, 12.7]). The curvature i
is the discrete version of the classical Gaussian curvature, and is usually called the angle
defect.
The GaussBonnet theorem for 3-dimensional convex polyhedra (Theorem 25.3) is due
to Descartes [Fed], so it precedes Eulers formula by over a century. In fact, the spherical
geometry proof of Eulers theorem is due to Legendre, and is probably the oldest rigorous
proof (see [Bla2, Crom] for more on the history of the subject). The extension of the Gauss
Bonnet theorem to general polyhedra (Theorem 25.5) will be used at a technical point in
the proof of the main result in Section 35. See [A2, 1.5] for 3-dimensional treatment. The
higher dimensional generalization (Theorem 25.4) is due to Shephard [Gr
u4].
The idea of taking an -neighborhood in Exercise 25.3 is standard in geometry. In a
similar context it leads to an advanced generalization of the GaussBonnet theorem to all
even-dimensional Riemannian manifolds [Gray, 5.5]. We refer to [Gray, 5.6] for the history
of the GaussBonnet theorem, various extensions and further references. Theorems 25.6
and 25.8 are essentially due to Cohn-Vossen [CV]. Although he does not consider convex
polyhedral surfaces, his study of geodesics on Riemann surfaces have polyhedral analogues
(see [Bus] for further results and references).
The formula in Exercise 25.10 is due to de Gua (1783) and can be generalized to higher
dimensions by scaling the angles and taking the alternating sum (see [Gr
u4]). For an
interesting discussion on the nature of Eulers formula see also [Lak].

249

26. Cauchy theorem: the statement, the proof, and the story
In this section we give the classical proof of the Cauchy theorem, of Alexandrovs
extension, and of Stokers converse. Much of the rest of the book is based on this
section.
26.1. Polytopes are surprisingly rigid. Let P and P be two combinatorially
equivalent polytopes in R3 (see Section 12), with vertices v1 , . . . , vn and v1 , . . . , vn ,
and faces F1 , . . . , Fm and F1 , . . . , Fm , respectively. From here on, we will always
assume that the equivalence maps the corresponding vertices and the corresponding
faces, i.e., vi vi and Fj Fj , all i and j. Finally, suppose the corresponding
vertices belong to the corresponding faces: vi Fj if and only if vi Fj . We say that
two faces Fj and Fj are equal if they are equal as polygons, i.e., the corresponding
edges have equal lengths and the corresponding face angles are equal.
Theorem 26.1 (Cauchy theorem). Let P and P R3 be two convex polytopes as
above, with all faces equal polygons: Fj Fj , for all 1 j m. Then P and P are
equal polytopes: P P , i.e., P can be moved into P by a rigid motion.
There is a way to think of the Cauchy theorem in terms of graph realizability.
Suppose P is simplicial, i.e., all faces are triangles. Then faces are determined by the
edge lengths. Formally, let G be the graph of P , and let L : E R+ be the length
function on edges of P . The Cauchy theorem implies that if convex polytope P has
the same pair (G, L), then P P . In other words, polytope P with (G, L) is unique
up to a rigid motion. This uniqueness property is an important way of thinking of
Theorem 26.1.
More generally, let T be a graph of some triangulation of faces of P , i.e., a graph on
the set V of vertices of P with the set of edges H E being edges of the polytope and
some diagonals in T . Denote by L : H R+ be the corresponding length function.
Now the full power of the Cauchy theorem implies that if convex polytope P has the
same pair (T, L), then P P .
A more traditional way to think of the Cauchy theorem is in terms of (continuous)
rigidity. A continuous family of polytopes {Pt : t [0, 1]} is called a continuous
deformation of P0 into P1 if under deformation the faces remain equal. The Cauchy
theorem now implies the following result.
Corollary 26.2 (Rigidity of convex polytopes). Let {Pt : t [0, 1]} be a continuous
deformation of convex polytopes. Then Pt is a rigid transformation from P0 to P1 .
While the corollary follows immediately from the Cauchy theorem, it is easier to
prove, has many applications and generalizations, some of which will be mentioned
below.
Remark 26.3. Let us calculate the degrees of freedom of convex polytopes to see if the
Cauchy theorem makes sense. A simplicial polytope P R3 is determined by its 3n
coordinates. The space of rigid motions has dim O(3, R) = 6 dimensions (three rotations
around pairs of axes and three translations along the axes). Thus, up to a rigid motion,
polytope P has 3n 6 degrees of freedom. Since the number of edges is |E| = 3n 6,

250

this makes sense (see Corollary 25.2). If we had |E| < 3n 6, we would have a hard time
showing that they have a unique real solution.65 While this is not a formal argument, the
magic number (3n 6) is good to keep in mind.

26.2. What to expect when you are expecting a proof. Surprise! There are
many interesting proofs of the Cauchy theorem and its relatives. The classical proof
(essentially due to Cauchy) will occupy the rest of this section. As the reader shall
see in the next section, the Cauchy theorem can be extended and generalized in a
number of ways. We outline a few of these directions and prove some extensions.
In fact, much of the rest of the book is also connected to the Cauchy theorem, one
way or another, but the following seven sections are directly related. In Sections 28, 29
we present two other proofs of the Cauchy theorem, one of a local, while another of
global nature. In Section 30 we present several examples of non-unique and flexible
polyhedra. By the Cauchy theorem these polyhedra must be non-convex, but they
are interesting and enlightening nonetheless.
Later, in Section 31, we introduce an algebraic approach to rigidity of polyhedra.
This approach leads us to two proofs of Dehns rigidity theorem (in Sections 32 and 33)
and eventually culminates in Section 34 with the proof of the bellows conjecture
(Theorem 31.2). Let us note that the final remarks subsections in these sections
contain references not listed in other sections, so the reader is advised to read all of
them to get the big picture.
26.3. The traditional proof of the Cauchy theorem. Let P and P be two
polytopes as in the Cauchy theorem (Theorem 26.1). Denote by G = (V, E) the
graph of P (and of P ), and by e and e the dihedral angles in polytopes P and P
of the corresponding edge e E. Let us label every edge e E with (+) if e < e ,
with () if e > e , and with (0) if e = e . Now let us analyze this labeled graph G.
For a vertex v in P , consider the labels of the edges containing v, written in cyclical
order, and ignoring the zero labels. Denote by mv the number of sign changes, i.e.,
pairs of adjacent edges (cyclically, skipping all 0s) with labels of different signs (see
e.g., Figure 26.1). Clearly, mv is always even.
Lemma 26.4 (Sign changes lemma). For every vertex v V , the number of sign
changes mv among pairs of edges containing v is at least four, unless all labels of
edges containing v are zero.
Proof. Let Cv be the cone spanned by edges of the polytope P containing v. Think
of Cv as an infinite cone starting at v and containing P . Now let Qv be the spherical
convex polygon obtained by intersection of Cv with a unit sphere S2 centered at v.
The sides of Qv are equal to the face angles, while angles are equal to dihedral angles
in edges containing v. Note also that Qv lies in a hemisphere S2+ . Similarly, let Qv
be the corresponding polygon obtained from P . From above, Qv and Qv have equal
65The

system would have innitely many complex realizations of (G, L), but that make very little
sense in the context of convex polytopes over the real numbers.

251

w1

v
w6

Qv

w2

+
0

+
w5

Cv
P

Qv

w3

w4

Figure 26.1. Spherical polygon Qv corresponding to vertex v. A vertex v in graph G with mv = 2 sign changes.
edge lengths but may have different angles. The result now follows immediately from
the spherical LegendreCauchy lemma (Theorem 22.2).

We are now ready to prove the Cauchy theorem.
Proof P
of Theorem 26.1. Assume for now that the labeling has no zeroes. Denote by
M = vV mv the total number of pairs of edges with a sign change. By the lemma,
we have a lower bound on the number of sign changes: M 4 |V |.
Now let us get an upper bound on M by counting the number of sign changes
summing over all faces F F . As before (see Section 25), let fk denotes the number
of faces with k sides. We have:
|F | = f3 + f4 + f5 + f6 + f7 + f8 + . . . and
Therefore,

2 |E| = 3f3 + 4f4 + 5f5 + 6f6 + 7f7 + 8f8 + . . .

4 |E| 4 |F | = 2f3 + 4f4 + 6f5 + 8f6 + 10f7 + 12f8 + . . .

Observe that the number of sign changes in a (2r + 1)-gon face is at most 2r, the
same as in a 2r-gon (see Figure 26.2). From here we have:
M 2f3 + 4f4 + 4f5 + 6f6 + 6f7 + 8f8 + . . .
2f3 + 4f4 + 6f5 + 8f6 + 10f7 + 12f8 + . . .
4 |E| 4 |F | = 4 |V | 8 ,

where the last equality follows from Eulers formula. Since M 4 |V |, we get a
contradiction.
Now suppose some edges in G are labeled zero, but not all of them. Remove zero
labeled edges from G and let H be a connected component in the resulting graph
(see Figure 26.2). Note that for the graph H we also have Eulers formula, since
the removal of non-bridge edges (i.e., those edges whose removal does not make H
disconnected) decreases only the number of faces and edges, so it does not change
|V | |E| + |F |. Then the argument above gives a contradiction once again.
From above, we conclude that H is empty and all edges are labeled zero. In other
words, polytopes P and P have equal dihedral angles between the corresponding

252

faces. This immediately implies the polytopes are equal, i.e., can be moved by a rigid
motion.


+
-

+
-

4
+

Figure 26.2. Maximum number of sign changes in different faces, and

graph H obtained after removal of zero labeled edges in G.
The proof above is important enough to be stated as a separate result, so that we
can refer to it later on. Since the proof is essentially a graph-theoretic statement, let
us phrase it as such:
Lemma 26.5 (Sign counting lemma). Suppose the edges of a plane simple graph are
labeled with 0, (+) and () such that around each vertex either all labels are 0 or have
at least four sign changes. Then all signs are 0.
Here by plane graph we mean a graph which is already drawn on a plane, since
otherwise, for graphs with low connectivity, the faces are not well defined. Note that
the faces include the outside face as well.
26.4. Better language for the Cauchy theorem. As the reader can see, we use
a rather clumsy way of saying that two polytopes are assembled in the same way
from the same faces. Let us straighten the language and restate the result.
Theorem 26.6 (Cauchy, restated). Let P and Q R3 be two combinatorially equivalent convex polytopes whose corresponding faces are isometric. Then P and Q are
isometric.
Here by isometry between polytopes we mean that one can be mapped into another
such that the pairwise distances between the corresponding points are always equal.
This map is called the isometry map. Same for the isometry between faces, but here
we are assuming that the isometry respects combinatorial equivalence.
Note that in R3 two polytopes are isometric if and only if they are equal, i.e., can
be mapped into each other by a rigid motion. Beside sounding more scientific, the
isometry is a useful concept in several generalizations of the Cauchy theorem.
Note also that the underlying theme in this version is the local global property
of isometry, since we are saying that isometry of the faces (all lying on a surface of
polytopes) implies global isometry. As we will see in Section 30, this principle does
not hold for non-convex polytopes. However it does apply in many other convex
situations.

253

Now let us redefine rigidity of convex polytopes and restate Corollary 26.2. We
say that {Pt : t [0, 1]} is a continuous deformation of polytope P if the faces of Pt
remain isometric, for all t [0, 1]. We say that P is (continuously) rigid if in every
continuous deformation {Pt } of polytope P the polytopes Pt are isometric.
Corollary 26.7 (Rigidity of convex polytopes, revisited). Every convex polytope P
R3 is continuously rigid.
This definition of rigidity is also robust enough to allow advanced generalizations.
For simplicity, we drop continuous and use rigidity for the rest of the lectures66.
26.5. Parallel polytopes. We start a short series of extensions of the Cauchy theorem (Theorem 26.1) with the following side result. While easily equivalent to the
Cauchy theorem, it brings to light some properties of the proof given in Section 26.3.
Theorem 26.8 (Alexandrov). Let P, Q R3 be two combinatorially equivalent convex polytopes with equal corresponding face angles. Then they have equal corresponding
dihedral angles.
This result easily implies the Cauchy theorem: faces and dihedral angles determine
the whole polytope. In fact, we used this observation in the proof of the Cauchy
theorem. On the other hand, the result is applicable to distinct polytopes, such as
bricks [a b c].
The proof of Theorem 26.8 is essentially an observation. Note that in the proof of
the Cauchy theorem we never used the geometry of faces, except for the face angles.
In addition to the (intrinsic) convexity of polyhedra, we use only only one global
parameter: Eulers theorem, that is a consequence from the fact that the surfaces of
convex polytopes are homeomorphic to a sphere. Thus, basically, in the proof of the
Cauchy theorem, we first establish Theorem 26.8, and only then prove the result.
We say that two polytopes P, Q Rd are parallel if they are combinatorially equivalent and the corresponding facets are parallel. Clearly, every two combinatorially
equivalent polytopes with equal corresponding dihedral angles can be made parallel
by a rigid motion. Basically, one can view Theorem 26.8 as a local condition for being
parallel, up to a global rotation.
Let us mention that the inverse to Theorem 26.8 is also true. Given two polytopes
which are combinatorially equivalent and have equal corresponding dihedral angles
we conclude that they are parallel, i.e., the corresponding faces lie on parallel planes.
Therefore, the corresponding edges lie on parallel lines. Finally, the corresponding
face angles are equal as angles between pairwise parallel lines.
26.6. Face angles are not as good as you think they are. Before we move to
further generalizations of the Cauchy theorem, let us mention that from the point of
view of convex polytopes, face angles are not a good way to define a polytope. In
fact, they overdetermine the polytope, even up to translation of faces. Indeed, let us
66There

are other kinds of rigidity, such as static rigidity, innitesimal rigidity, k-th order rigidity,
global rigidity, etc. (see [Con5]). We will dene the rst two in Sections 31 and 33.

254

compare the degrees of freedom for a simplicial polytope P with n vertices. First,
there are
X
ifi = 2 |E| = 2 (3n 6)
i

face angles. These angles satisfy |F | = (2n 4) linear relations on the sums along
faces67. In total, angles give 2 |E| |F | = 4n 8 degrees of freedom.
On the other hand, n vertices of a polytope P R3 have (3n 6) degrees of
freedom (see Remark 26.3). There are |F | = (2n4) degrees of freedom of translation
of faces, |F | 3 = 2n 7 up to translations of the whole space R3 . Thus there are
(3n 6) (2n 7) = n + 1 remaining degrees of freedom. Comparing n + 1 and 4n 8
we obtain a number of (nonlinear) relations on face angles of polytopes, making face
angles an unattractive way to define a polytope.
Let us illustrate this in a tetrahedron: 12 face angles in this case determine a
tetrahedron up to rigid motions and expansion, a 5-dimensional space of realizations.
Taking into account 4 linear relations on face angles, we conclude that there are three
extra nonlinear equations on the angles. Let us compute one of these two equations
(the others are similar).
In a tetrahedron as in Figure 26.3, denote the vertices by i [4]. Recall the law of
sines:
|12|
sin (132)
=
.
|23|
sin (213)
Applying this equation to all four faces, we have:
1 =

|12| |23| |34| |14|

sin (132) sin (243) sin (314) sin (124)

.
|23| |34| |14| |12|
sin (213) sin (324) sin (134) sin (142)

This is a non-linear equation on 8 angles, all linearly independent otherwise.

3

4
1
2
Figure 26.3. Getting a nonlinear relation on face angles of a tetrahedron.

67The

GaussBonnet theorem (Theorem 25.3) gives another linear relation, but it follows from
the linear relations on faces (as made explicitly in the proof of the theorem).

255

26.7. The converse of the Cauchy theorem. From the discussion above, one can
think of the Cauchy theorem as a statement that in all convex polytopes in R3 the
edge length and face angles determine the dihedral angles, and with them the whole
polytope. Thus, it is reasonable to ask if the converse is true as well: is it true that
the edge lengths and dihedral angles determine the face angles as well? The following
result confirms the suspicion:
Theorem 26.9 (Stoker). Let P, Q R3 be two combinatorially equivalent convex
polytopes with equal corresponding edge lengths and dihedral angles. Then P and Q
are isometric.
Note that the theorem is obvious for simplicial polytopes: in that case the edge
lengths alone determine face angles (consider separately each triangular face), and
with them the whole polytope. Of course, for simple polytopes knowing edge lengths
is insufficient. On the other hand, dihedral angles alone can determine the face angle
(consider separately each vertex cone), and with them the whole polytope once again.
Therefore, this result is actually easy for extreme cases and the main difficulty is with
intermediate polytopes. The main idea of the proof below is to combine these two
different approaches into one argument.
Proof. Using the approach in the proof of the Cauchy theorem, compare face angles
in P and Q and label them with (+), () and 0 accordingly. Note that around
every face either all labels are zero or there are at least four sign changes. This is
the analogue of the arm lemma (Lemma 22.3) for planar polygons and the proof is
verbatim. Similarly, around every vertex all labels are zero or there are at least four
sign changes. Indeed, consider a vertex v of P and the dual cone Cv (see Section 25).
The dihedral angles i equal to i , where i are face angles in v. Now, as in
the proof of the sign changes lemma (Lemma 26.4), by the arm lemma for spherical
polygons we have the claim.
Consider the medial graph H with vertices corresponding to edges of P and edges
connecting two adjacent edges in P lying in the same face. Label the edges of H
with the same label as the corresponding face angle in P . Note that the faces of H

Figure 26.4. Medial graph H of the graph of a square prism is the

graph of a square antiprism.
correspond to vertices and faces of P , and have all zero labels, or at least four sign
changes. Applying the sign counting lemma (Lemma 26.5) to the dual graph H , we
conclude that all labels in H are zero. In other words, the corresponding face angles
in P and Q are equal. Now the isometry of faces and the equality of dihedral angles
implies the result.


256

26.8. Exercises.
Exercise 26.1. (Stokers conjecture) The conjecture states that every two combinatorially equivalent convex polytopes in R3 with equal corresponding dihedral angles have equal
corresponding face angles.
a) [1-] Check that the conjecture implies Theorem 26.9.
b) [1] Prove that the conjecture holds for simple polytopes.
c) [1+] Prove that the conjecture holds for circumscribed polytopes.
d) [] Prove the conjecture for simplicial polytopes.
Exercise 26.2. [1+] State the spherical analogue of Stokers conjecture. Find a counterexample.
26.9. Final remarks. The Cauchy theorem was proposed to Cauchy by Legendre, who
established it in several special cases (see [Bla2, Mi3]). Legendre wrote a standard geometry
course, and was led to the problem by trying to formalize (and prove) Denition 11.10 from
Euclids Elements [Euc]:
Equal and similar solid gures are those contained by
similar planes equal in multitude and in magnitude.
Here by planes Euclid means the faces of the polytope, which he required to be equal
polygons. For centuries this denition was viewed as a claim that polytopes with equal faces
are equal. However, the notion of equality as congruence of polytopes is absent in Euclid,
and the statement really concerns the volumes of polytopes. As was noted in [Mi3, 2.5],
it seems Euclid meant to apply the denition only to volumes of basic polytopes, such
as triangular prisms and pyramids. In any event, Legendre proved the Cauchy theorem in
several special cases, and gave the general problem to his student Cauchy.68
The proof of the Cauchy theorem more or less follows Cauchys original approach. The
Alexandrov theorem (Theorem 27.7) can be found in [A2]. Let us mention that similar
polytopes play an important role in the study of mixed volumes of convex bodies. We use
them in the proof of the Minkowski theorem (Theorem 36.2).
Our proof of Theorem 26.9 follows the original proof by Stoker, who further conjectured
that convex polytopes with equal corresponding dihedral angles have equal face angles [Sto,
6] (see Exercise 26.1). Let us mention that Theorem 26.8 extends verbatim to S3 and H3
(see Subsection 27.1), and so does Stokers theorem (Theorem 26.9). Although the natural
analogue of Stokers conjecture is open in full generality, the spherical analogue in false (see
Exercise 26.2), and the innitesimal version was recently established in [MazM] (see also
[Ale6, 5], [Schl2] and Exercise 32.2).
In the next several sections we present further results on rigidity of polyhedral surfaces.
In Sections 28, 29 we present two other (more involved) proofs of the Cauchy theorem, and
in Sections 32, 33 we present about four dierent (and somewhat less involved) proofs of
the rigidity of convex polytopes (Corollary 26.2).

68The

works of Legendre were studied by I.Kh. Sabitov (personal communication).

257

27. Cauchy theorem: extensions and generalizations

In this short section we discuss a number of extensions of the Cauchy theorem,
some easy and straightforward, some important (the Alexandrov uniqueness theorem
and its converse in Section 37) and some beyond the scope of this book (the Pogorelov
uniqueness theorem).
27.1. Spherical polyhedra. As it turns out, in order to extend the Cauchy theorem
to higher dimensions, we first need a spherical analogue.
Note first that the proof of the Cauchy theorem is robust enough to work in other
geometries, e.g., on a sphere S3 , Lobachevsky space H3 , as well as other metric spaces
with bounded curvature.
We begin with the definitions. Let Sd Rd+1 be a d-dimensional unit sphere.
A convex spherical d-dimensional polyhedron P Sd is an intersection of a finite
number of hemispheres Sd+ , i.e., half-spheres lying on one side of a hyperplane in R4
containing the origin O. Note that the boundaries of these hemispheres are the
(d 1)-dimensional spheres, which we denote by C1 , . . . , Cm . The vertices, edges
and k-dimensional faces of a convex spherical polytope P Sd are the points of
the boundary P which lie in the intersections of d, d 1, and d k spheres Ci ,
respectively. We can now define combinatorially equivalent spherical polyhedra the
same way as we do in the Euclidian space.
Define the spherical distance between points x, y Sd as the length of the shortest
path between them. Equivalently, this is the length along the unique great circle which
contains x, y (such circle is unique unless the points are equal or opposite). Define
now isometric spherical polyhedra, write P Q, the same way as before. Note
that spherical polyhedra P, Q Sd are (globally) isometric if and only if there is an
orthogonal transformation M O(d + 1, R) which maps P into Q, i.e. M(P ) = Q.
We will need the isometry only for d = 2 (faces), and d = 3 (polytopes).
Theorem 27.1 (Cauchy theorem for spherical polyhedra). Let P, Q S3+ be two
combinatorially equivalent convex spherical polyhedra in the upper hemisphere, whose
corresponding faces are isometric. Then P and Q are isometric.
The proof follows verbatim the proof of the Cauchy theorem, with appropriate
substitutions of spherical geometry concepts, such as dihedral angles in edges of P
and Q. We skip the details.
27.2. Cauchy theorem in higher dimensions. We are now ready to present the
following important extension.
Theorem 27.2 (Cauchy theorem in higher dimensions). Let P, Q Rd (or P, Q
Sd+ ), d 3 be two combinatorially equivalent (spherical ) convex polyhedra whose
corresponding facets are isometric. Then P and Q are isometric.
Here by (global) isometry between polyhedra we mean that one can be mapped
into another such that the pairwise distances between the corresponding points are
always equal. By Sd+ we denote a fixed (upper) hemisphere.

258

Now, to prove Theorem 27.2, we need the corresponding result for convex spherical
polytopes first. Such a result is easy to prove by a straightforward induction. For
d = 3 this is Theorem 27.1. Suppose the result is established for some d 3. Take a
(d + 1)-dimensional spherical polyhedron P and let Qv be an intersection of P and a
d-dimensional sphere centered at a vertex v of small radius. By inductive assumption
applied to the d-dimensional convex spherical polytope Qv , we conclude that Qv and
thus the spherical pyramid Pv = conv{v, Qv } Sd+1 are completely determined. This
gives the Cauchy theorem in any dimension for spherical polytopes.
Now, for a Euclidean polytope P Rd , again consider an intersection Qv with a
small (d 1)-dimensional sphere. Use the result for spherical polyhedra to conclude
that the pyramid Pv over Qv is completely determined, and thus gives all dihedral
angles between facets. We omit the details.
Before we conclude, let us state the rigidity of higher-dimensional polytopes as well.
Corollary 27.3 (Rigidity in Rd ). Every convex polytope in Rd , d 3, is rigid.

Remark 27.4. There is a case to be made that in higher dimensions the polytopes are
largely overdetermined, i.e., conditions in the corollary can be weakened in most cases.
Without making this precise, let us make a counting argument calculating the degrees of
freedom (cf. Remark 26.3).
Consider a convex simplicial polytope P Rd , in which case the facets are determined
by the edge lengths. Suppose P has n vertices
and k edges. The space of realizations of
the graph of P up to rigid motions is dn d+1
2 , since the dimension of the group of rigid
motions (translations and rotations) is d + dim O(d, R) = d(d + 1)/2. Now the rigidity of P
implies69 that this is the lower bound for the number of edges:


d+1
()
k dn
.
2
Interestingly, we already have seen this inequality. Denote by P Rd a dual polytope to P .
Clearly, P is simple, and in the notation of Section 8 we have fd1 = n and fd2 = k.
Thus, the inequality g2 g1 presented in Remark 8.2 becomes inequality () given above.
Let us show that inequality () can be quite weak. This means that a polytope can
have too many edges, more than necessary to justify the rigidity. The easiest example is a
cross-polytope Cd Rd , dened as a regular polytope dual to a cube.
For example, C3 is


a regular octahedron. Clearly, Cd has n = 2d vertices and k = 2d

d
edges, making ()
2
into strict inequality for d 4.

27.3. Unbounded convex polyhedra. Recall the notion of (unbounded) convex

polyhedra considered in Section 25. Namely, let P R3 be an intersection of a finite
number of half-spaces. The question is: can we extend the Cauchy theorem to this
case? Phrased differently, do there exist non-isometric convex polyhedra which are
combinatorially equivalent and have isometric corresponding faces?
The answer is easily no, and examples include a cone and a product of an unbounded convex polygon and a line (see Figure 27.1). Somewhat surprisingly, there
is an easy way to enforce rigidity and generalize the Cauchy theorem to this case.
69Formally

speaking, this conclusion can be made for realizations only over C, rather than over R.
The reader can think of this as a heuristic argument which can in fact be formalized.

259

Figure 27.1. Non-rigid unbounded convex polyhedra.

Theorem 27.5 (Cauchy theorem for unbounded polyhedra). Let P and Q R3
be two combinatorially equivalent unbounded convex polyhedra which do not contain
straight lines and whose corresponding faces are isometric. Suppose further that P
and Q contain the origin O and have equal base cones: CP = CQ . Then polyhedra P
and Q are isometric.
One can think of the theorem as of a limiting case of the Cauchy theorem, when
vertices of a face are sent to infinity along the rays of the base cone. Unfortunately,
we do not know a rigorous limit argument in this case. In fact, there is a global
proof along the same lines as that of the Cauchy theorem. Again, consider dihedral
angles and the sign changes. The infinite edges of P and Q have equal dihedral angles
to that of CP = CQ . Now proceed as in the proof of the Cauchy theorem, noting that
Eulers formula also applies in this case.
27.4. Non-strictly convex polytopes. Consider a sequence of convex polytopes
Pi R3 converging to a convex polytope Q. Note that the geometric convergence
does not necessarily translate into a combinatorial convergence, i.e., the limit graph
G = limi Gi will, of course, contain graph H of Q, but may be much larger than H
as many edges now can lie on the surface Q (see Figure 27.2). One can think of such
polyhedra as non-strictly convex polytopes, defined not to have vertices in the relative
interior of faces. We will use a different terminology.

Figure 27.2. Examples of non-strictly convex polyhedral surfaces.

Let S = Fi R3 be a polyhedral surface defined as a union of convex polygons Fi
embedded into R3 . We say that S is a convex polyhedral surface if S = P for some
convex polytope P R3 , where the equality is in term of sets of points. Of course,

260

different polyhedral surfaces may be equal (as sets of points) to the boundary of the
same convex polytope.
We say that convex polyhedral surfaces S and S in R3 are globally isometric,70
write S S , if there is a map between them (as sets of points in R3 ) which preserves
distances in R3 . In other words, S S if there exists a rigid motion which moves S
into S .
Theorem 27.6 (CauchyAlexandrov theorem for polyhedral surfaces). Let S =

3
m
i=1 Fi and S = i=1 Fi be two convex polyhedral surfaces in R , which are combinatorially equivalent and the corresponding faces are isometric: Fi Fi . Then
surfaces S and S are globally isometric.
Proof. This is again essentially an observation on the proof of the Cauchy theorem.
We say that a point v in S is flat if the sum of the face angles in it is equal to 2
(the point v may actually lie on an edge of P ). Since the sum of the face angles
around vertices is always 2, every flat vertex v of polygon Fi is mapped into a
flat vertex v of polygon Fi . There are three possibilities: either both v and v lie in
the interior of faces, or one of them lies
on the proof of the Cauchy theorem.
Suppose S = P and S = P . From above, the union of all polygons Fi which
form a polytope face Pj P is isometrically mapped into the union of all Fi which
form a face Pj P . Now use the same argument as in the Cauchy theorem to obtain
the result.

27.5. Alexandrovs uniqueness theorem for isometric convex surfaces. Let
S Rd be a (convex) surface embedded into a space. For two points x, y S define
the surface distance |x, y|S to be the length () of the shortest path S between x
and y. Thus we obtain the surface metric on S.

3
Let S = m
i=1 Fi and S = j=1 Fj be two (convex) polyhedral surfaces in R . We
say that S and S are intrinsically isometric, write S S , if there exists a map
: S S (between sets of points), which maps the surface metric on S into the
surface metric on S :
|x, y|S = |(x), (y)|S , for all x, y S.

Of course, global isometry implies intrinsic isometry, but the converse is not true for
non-convex surfaces. For example, two polyhedral surfaces that lie on the boundary of
the same polytope are intrinsically and globally isometric. Alternatively, the surfaces
of two realizations of the same polytope are intrinsically, but not globally, isometric
(see, e.g., Figure 30.1). By the Cauchy theorem (Theorem 26.1), one of these polytopes must be non-convex. The following result shows that one cannot obtain two
intrinsically isometric surfaces from two different convex polytopes.
Theorem 27.7 (Alexandrov uniqueness theorem). Let S, S R3 be two convex
polyhedral surfaces which are intrinsically isometric. Then they are globally isometric.
70We

use the term globally isometry in place of the usual isometry to emphasize that we are
working with surfaces rather than polytopes. The intrinsic isometry is dened below.

261

Once again, let us emphasize that one can view Theorem 27.7 as of a uniqueness
result. It says that up to a rigid motion there exists only one convex polyhedral
surface with a given metric.
Proof. Let : S S be a map in the definition of intrinsic isometry, S = P ,
and S = P . Consider the regions Gi = (Fi ). Since all interior points in Fi are flat
in S, so are the points in Gi ; otherwise shortest paths in Fi will not be mapped into
shortest paths in Gi by Proposition 10.1. Similarly, since sides of the polygon Fi are
the shortest path in S, the sides of Gi are also shortest paths. We conclude that Gi
are polygons in S not containing vertices of P in their relative interior. Similarly,
regions Hj := 1 (Fj ) are polygons in S not containing vertices of P in their relative
interior. Finally, note that all Gi and Hj are intrinsically convex in the surface metric
of S and S.
Let Gij = Gi Fj , and Hij = Hi Fj , for all 1 i m, i j m . Clearly, each
Gij and Hij is a convex polygon and the map : Hij Gij is an isometry between
them. Finally, combinatorial equivalence of convex polyhedral surfaces Se = ij Hij
and Se = ij Gij follows from adjacency of the corresponding edges as zero distance.
By Theorem 27.6, surfaces Se and Se are globally isometric. Since S = Se and S = Se
(as sets of points), we obtain the result.

An example of an isometric embedding of a polygon into a surface is given in
Figure 27.3. The edges subdivide it into smaller convex polygons, as in the proof
above.

Figure 27.3. An isometric embedding of a rectangle into a surface of a cube.

27.6. General convex surfaces. Let S = P be the surface of a compact convex

body. Define a metric on S again via the geodesic distance, i.e., let |xy|S be the length
of a shortest path (it may not be unique). As before, we say that surfaces S and S
are intrinsically isometric if there exists a map : S S , which maps the surface
metric on S into the surface metric on S (see above). The following result is an
advanced extension of the Cauchy theorem and the Alexandrov uniqueness theorem.
Theorem 27.8 (Pogorelov uniqueness theorem). Let S, S R3 be two convex surfaces which are intrinsically isometric. Then they are globally isometric.

262

The proof is a delicate and at times technical argument based on approximation of

convex surfaces by convex polytopes. We will not present it here.
27.7. Exercises.
Exercise 27.1. [1+] Let S R3 be a (non-convex) 2-dimensional polyhedral surface homeomorphic to a sphere, such that every face of S is a unit square. Prove that all dihedral
angles of S are right angles.
Exercise 27.2. a) [1-] Prove or disprove: R3 is a regular tetrahedron if and only if
it has ve equal dihedral angles.
b) [1-] The same for all six equal dihedral angles.
Exercise 27.3. (Locally convex polyhedra) [1] A polyhedron is called locally convex if
the cone at every vertex is a convex cone. Prove that every locally convex 2-dimensional
polyhedral surface homeomorphic to a sphere and embedded in R3 is convex.71
Exercise 27.4. [1+] Show that every two spherical tetrahedra in S3 with equal corresponding dihedral angles are isometric.
Exercise 27.5. [1] Complete the proof of Theorem 27.1.
Exercise 27.6. [1+] Complete the proof of Theorem 27.2.
Exercise 27.7. [1+] Complete the proof of Theorem 27.5.
Exercise 27.8. [1] Complete the proof of Theorem 27.6.
Exercise 27.9. (Olovianishniko ) [2] Let P R3 be a convex polytope and let Q R3
be a convex body. Suppose the surface P is isometric to Q. Prove that Q is a convex
polytope. Conclude from here that P and Q are globally isometric.
27.8. Final remarks. There are a number of extensions of the Cauchy theorem to polyhedra in other metric spaces, some of them going back to Alexandrov, while some are more
recent and quite advanced. In particular, extensions of the Cauchy theorem to spherical
polytopes and polytopes in a Lobachevsky space (hyperbolic polytopes) are due to Alexandrov. We refer to [A1, A2, Pog3] for classical results, to [Con5, IKS] for the surveys, and to
papers [Mi6, Mi4, Schl1] for more recent advances. Theorem 27.6 is also due to Alexandrov,
as are most of the results in this section.
The study of unbounded polyhedra and Theorem 27.5 are due to Olovianishniko [Olo1]
(sometimes spelled Olovjanisnikov). He also established an important intermediate result
between the Alexandrov and Pogorelov uniqueness theorems in [Olo2] (see Exercise 27.9).
For Pogorelovs theorem see [Pog3].
Also, a number of further extensions of the uniqueness theorems (to convex spherical
surfaces, unbounded convex surfaces, surfaces with bounded curvature, etc.) were obtained
by Alexandrov and his students. We refer to an extensive survey [Sab1] (see also [Sen2])
for rigidity of general surfaces.

71This

is unfortunate since, of course, the Cauchy rigidity theorem immediately extends to locally
convex polyhedra homeomorphic to a sphere.

263

28. Mean curvature and Pogorelovs lemma

We introduce the mean curvature of polyhedral surfaces and prove that it is invariant under isometric deformations. We then present Pogorelovs lemma, which implies
both the Cauchy theorem and the isoperimetry of spherical polygons (cf. Section 7).
28.1. Mean curvature. The proof below requires a few preliminary results on geometry of convex polyhedra. Since we need these results for other purposes, we separate
them from the main part of the proof.
Let P R3 be a convex polytope with the set of vertices V and the set of edges E.
For every e = (v, w) E denote by e = |vw| the length of edge e. Let u v,e = (v, w)/e
be a unit vector starting at v V and pointing along edge e = (v, w) E. Fix a real
function f : e fe , where fe R, e E. Define vectors a v (f ) R3 , v V , by the
following summation formula:
X
fe u v,e .
a v (f ) :=
e=(v,w)E

Finally, let

M(P, f ) :=

X
eE

fe e .

A special case of M(P, f ) is of particular importance. Denote by e the dihedral angle

in e, and let M(P ) := M(P, )/2 be the mean curvature of P :
1 X
M(P ) :=
( e ) e .
2 eE

This function comes from differential and integral geometry and has a special invariance property which we prove later in this section.

Lemma 28.1 (Edge summation lemma). For every point z R3 , vectors r =

vz
v

corresponding to vertices v V , and function f : E R as above, we have:

X


M(P, f ) =
av (f ), rv .
vV

Proof. Observe that

()

hu v,e , r v i + hu w,e , r w i = e ,

for every edge e = (v, w) E with length e . Indeed, the l.h.s. in () is the sum of
the (signed) lengths of projections pv and pw of vectors r v and r w onto line (vw) (see
Figure 28.1). Changing the order of summation from vertices to edges we obtain:
E
X

XD X

fe u v,e , r v
a v (f ), r v =
vV

e=(v,w)E

as desired.

vV

e=(v,w)E

h
i
X
fe hu v,e , r v i + hu w,e , r w i =
fe e = M(P, f ),
eE

264

rv
v
u v,e

pv

rw
pw
u w,e w

Figure 28.1. The equality pv + pw = e , where pv = hu v,e , r v i and

pw = hu w,e , r w i are projections onto edge e = (v, w).
28.2. Some things never change. Let us study the mean curvature in greater
detail, by observing what happens when we continuously change the polytope. Start
with the continuous deformations of a cone, defined as follows.
Let C R3 be a polyhedral cone, not necessarily convex, defined as follows. Let u 1 ,
. . . ,u k be unit vectors in R3 , and fix a point v R3 . Consider cone faces starting at v
and spanned by consecutive pairs of vectors (u 1 , u 2 ), . . . ,(u k1 , u k ), and (u k , u 1 ).
Denote by i , i [k], the (signed) dihedral angle of C at u i .
Now, suppose we have a continuous family of cones {Ct : t [0, 1]}, such that
C0 = C, spanned by continuously changing unit vectors u i (t), where u i = u i (0). We
call {Ct } a continuous deformation of the cone C.
Lemma 28.2 (Angular velocity equation). Let {Ct : t [0, 1]} be a continuous
deformation of the cone C preserving face angles, and let i (t) be the dihedral angles
in the cone Ct . Then
k
X
i=1

ui (t) i (t) = 0 for all t [0, 1].

Here in the lemma we are assuming that the dihedral angles e (t) on [0, 1] have
derivatives e (t) with respect to t. If not, by continuity of e (t), the claim can be
replaced with the left or right derivative at any point t = t0 with no change in the
proof. We find the former version more aesthetically pleasing, while the second more
useful.
Proof. The lemma follows immediately from the following observation. Note that
i (t)u i (t) are angular velocity vectors of the i-th face around (i 1)-st face of Ct .
Thus the sum is the rotational velocity of the 1-st face over itself, a zero.
Formally, consider a vector w R3 in the first face of C. Suppose w is the
composition of rotations around u i by an angle i (t) i (0), for all i [k]. We have:
" k
#
k
X
X
w =
i u i w =
i u i w = a w .
i=1

i=1

By the argument above, w = 0, and we conclude that a = 0.

265

Before we turn to the proof of the Cauchy theorem, let us make an interesting
corollary from the lemma. Let {Pt : t [0, 1]} be a continuous family of combinatorial equivalent polyhedra in R3 , not necessarily convex or even embedded (see
Section 26.2), such that vertex coordinates change continuously. Then, with the same
disclaimer as after Lemma 28.2, we obtain the following result.
Theorem 28.3 (Schlafli formula). For a continuous family of polyhedra {Pt : t
[0, 1]} preserving the faces, we have:
X
e (t) e (t) = 0, for all t [0, 1],
eE

where e (t) are edge lengths and e (t) are dihedral angles.
Proof. In the notation above, let fe = e (t) be derivatives of the dihedral angles.
By Lemma 28.2
we have a e ( (t)) = 0, for every edge e E. On the other hand,
P
M(Pt , (t)) = eE e (t)e (t), as in the theorem. Now the edge summation lemma
(Lemma 28.1) gives:
X


M(Pt , (t)) =
a v ( (t)), r v = 0,
vV

as desired.

The following corollary follows easily from the Schlafli formula, but is important
enough to single out. Recall that a continuous deformation {Pt : t [0, 1]} is a
continuous family of polyhedra with all edge lengths constant under deformation:
e (t) = ce for some ce > 0 and all t [0, 1].
Corollary 28.4 (Invariance of the mean curvature). Let {Pt : t [0, 1]} be a continuous isometric deformation of a (possibly non-convex) polyhedron. Then the mean
curvature M(Pt ) is a constant independent of t.
For convex polyhedra this follows immediately from Corollary 26.7. On the other
hand, for non-convex polyhedra this is a new result, a special property of flexible
polyhedra. We present their examples and further properties in Sections 30, 31.
Proof. Suppose first that all dihedral angles e (t) are differentiable on [0, 1]. Then
#
"
X
1 X
1
( e ) e =
e (t) e (t) = 0,
M (Pt ) =
2 eE
2 eE
so the mean curvature M(Pt ) is a constant independent of t. Suppose now that not
all e (t) are differentiable. By continuity, at every point t0 (0, 1) we can compute
the left and right derivatives of e (t), and then of M(Pt ). From above, the left and
right derivatives of M(Pt ) are equal to 0 at all t (0, 1). Thus, M(Pt ) is differentiable
with zero derivative in (0, 1). This again implies that M(Pt ) is independent of t. 

266

28.3. The ultimate local proof. Observe that the structure of the proof of the
Cauchy theorem splits naturally into a local and a global part: the sign changes
lemma (Lemma 26.4) which easily reduces to the arm lemma (Lemma 22.3) and the
sign counting lemma (Lemma 26.5) based on the use of a counting argument and
Eulers formula. There are several other proofs of the Cauchy theorem which follow
the same pattern, but have varying levels of complexity of local and global parts. The
idea of this proof is to make the local claim strong enough, so that the contradiction
follows easily, by a geometric and a straightforward double counting argument, rather
than by a calculation and Eulers formula. After all, Eulers formula can also be
deduced by a global argument (see Section 25), so it makes sense that the Cauchy
theorem can be proved along these lines.
Let C, D R3 be two convex cones starting at the origin O and with the same
sequence of side angles (in cyclic order). In this case we say that cones C and D have
isometric sides. The cones are called equal, write C D if there exists a rotation
mapping one into the other. Finally, define the dual cone C as in Subsection 25.4,
to be the cone of all normal vectors: C = {w R3 | (w, v) 0 for all v C}.
Lemma 28.5 (Pogorelov). Let C, D R3 be two convex cones with isometric sides
and dihedral angles 1 , . . . , k and 1 , . . . , k , respectively. Denote by u1 , . . . , uk the
unit vectors along edges of C. Let w = w(C, D) be a vector defined as follows:
w :=

k
X
i=1

(i i ) ui .

Then w C . Furthermore, w = 0 if and only if C D.

We postpone the proof of Pogorelovs lemma and start with the following original
proof of the Cauchy theorem based on the lemma.
Proof of the Cauchy theorem modulo Pogorelovs lemma. Let P and Q be combinatorially equivalent convex polytopes with isometric faces. Denote by V , E the set
of vertices and edges of P , respectively, and let e denote the dihedral angles in P .
To simplify the notation we use the same notation E for the sets of edges in Q and
let e , e E, be the dihedral angles in Q.
Denote by Cv the cone spanned by the edges of P containing the vertex v V .
Let w v be the vector as in the lemma, corresponding to the cone Cv . Fix a point z P
and let r v = (v, z), v V , be as above. Since r v Cv for all v V , by Pogorelovs
lemma we have (w v , r v ) 0. Summing this over all vertices, we get:
X
A :=
hw v , r v i 0 .
vV

In the notation of the edge summation lemma (Lemma 28.1), let fe = (e e ) and
note that w v = a v (f ). By the lemma, we have:
X
X
X
A =
(e e ) e =
( e ) e
( e ) e = 2M(P ) 2M(Q),
eE

eE

eE

267

where the last equality follows from the isometry between the corresponding edges of
polytopes P and Q. From here, M(Q) M(P ).
Now switch the roles of P and Q. We similarly get M(P ) M(Q). Therefore,
M(P ) = M(Q), A = 0, and every inequality above is an equality: hw v , r v i = 0, for
all v V . By convexity and the assumption that z P , this implies that w v = 0,
for all v V . By the second part of Pogorelovs lemma, we conclude that the
corresponding cones have isometric sides, and thus the corresponding dihedral angles
are equal: e = e , for all e E. Thus, the polytopes P and Q are also equal.

28.4. Proof of Pogorelovs lemma. While Pogorelovs lemma may seem like a
quantitative version of the sign changes lemma (Lemma 26.4), the proof we present
below reduces the former to the latter. Unfortunately, as of now, there is no independent proof of Pogorelovs lemma. So while the search for such a proof is ongoing, let
me present the proof we have.
Proof of Pogorelovs lemma modulo the sign changes lemma. Clearly, if C D, then
the vector w = 0. Now suppose C and D are not equal. Since cone C is convex, it
suffices to show that hw , u i i < 0, for all i [k]. In other words, we need to prove
that
k
k
DX
E
X
hw , u i i =
(j j )u j , u i =
(j j ) cos ij < 0, for all j [k],
j=1

j=1

where ij is the angle between u i and u j . Fix i [k] and denote by i the summation
we need to bound:
k
X
i :=
(j j ) cos ij .
j=1

We fix cone D and vary cone C. Let us show that i = i (C) maximizes only
when C D. This would imply that i < 0, and prove the result.
First, note that i is bounded, i.e., by definition it is at most k. Thus, by
compactness of the space of k-cones (equivalently, spherical k-gons), the maximum
of i is reached at some cone, which can be degenerate and not strictly convex. Either
way, let C be such a cone, and assume that i (C) > 0.
Think of the cone C and the corresponding spherical polygon. Place labels (+), ()
and 0 on the edges of C as in the proof of the Cauchy theorem, i.e., according to the
signs of (j j ), j [k]. By the assumption above, C and D are not equal, and by
the sign changes lemma (Lemma 26.4), there are at least four sign changes.
Let us assume that the i-th edge has label (+). Then, there exists r [k] such
that r-th edge also has label (+) and together the i-th and r-th edges divide all
k edges into two parts (in cyclic order) so that each of them has positive and negative
labels. Suppose edges p and q are the closest on each side with labels (). Consider a
continuous deformation {Ct } of the cone C, which increases dihedral angles at the i-th
and r-th edges, decreases dihedral angles at the p-th and q-th, and leaves unchanged
angles at the remaining edges. We claim that such a deformation is possible, with the

268

cones Ct convex on a sufficiently small interval t [0, ]. Indeed, we need to worry

only about increasing angles at the i-th and r-th edges beyond , and decreasing
angles at the i-th and r-th edges below 0. But since the corresponding angles at D
are greater or less, respectively, we conclude that i , r < , and p , q > 0, which
implies the claim.
p

+ r

-q

Figure 28.2. Change of dihedral angles in a continuous deformation {Ct } of the cone C. Here the angle ij (t) decreases.
Observe that by the arm lemma (Lemma 22.3), the angles ij (t) are decreasing
for j on an interval between p and q containing r, and are unchanged elsewhere (see
Figure 28.2). Differentiating directly, we have:72
k
X


i (t) =
(j j (t)) cos ij (t)
j=1

k
X
j=1

(j (t) j ) sin ij (t) ij

(t)

k
X

j (t) cos ij (t).

j=1

Note that the second sum is equal to 0. This follows from the angular velocity
equation (Lemma 28.2) by taking the scalar product with u i on both sides. Note
also that each term in the first sum is 0, with the r-th term > 0. We conclude
that i (t) > 0, which contradicts our assumption that i maximizes at C.
We have two more cases to consider: when the label of the i-th edge is 0 and (). In
the latter case we proceed similar fashion, by taking a deformation with the dihedral
angles increasing at p-th and q-th edges, while decreasing at i-th and r-th edges.
When the label is 0 we can treat it either as (+) or as (), since the i-th term in
the first summation is always zero. One has to be careful with the extremal cases:
if i = 0 we should increase the angle, and if i = , then decrease it. Either way we
get a contradiction with the maximality assumption on i .

72It

is easy to make the dierentiation rigorous. Let us parameterize the deformation, say, by
having i increase linearly: i (t) = i +t, for some > 0. Then the remaining angles j (t) and ij (t)
are completely determined, and can be computed by the spherical law of cosines (see Appendix 41.2).
Clearly, the resulting functions are all analytic in t.

269

28.5. Isoperimetry of spherical polygons. The following result is classical and

has no place in this section, and fits well other isoperimetric problems in Section 7, except that the proof is a straightforward application of Pogorelovs lemma (Lemma 28.5).
We say that two spherical polygons P, Q S2+ are isometric if they have equal
corresponding side lengths. By restricting them to the upper hemisphere S2+ we can
define the area of the polygons as the area of the corresponding polygonal regions on
a sphere.
Theorem 28.6. Let P, Q S2+ be two isometric convex spherical polygons, such
that P is inscribed into a circle in S2 and contains its center. Then area(P )
area(Q).
Proof. Denote by i and i the corresponding angles of polygons P and Q, i [k].
By Girards formula (see Section 41.3) we have:
area(P ) =

k
X
i=1

i (k 2) ,

area(Q) =

k
X
i=1

i (k 2) .

Denote by O1 the center of the sphere containing convex polygons P and Q. Denote
by C and D the cones over P and Q, respectively, with center at O1 . By the assumption, the polygon P = [x1 . . . xk ] is inscribed into a circle with center O2 S2 .

Finally, let r = O1 O2 and u i = O1 xi , for all i [k]. We have:

hu 1 , r i = . . . = hu k , r i = c,

for some c > 0.

Now, let w = w (C, D) be as in Pogorelovs lemma:

w=

k
X
i=1

(i i )u i .

By the lemma, w C . By the assumption in the theorem, r C, and, therefore,

hw , r i 0. We conclude:
area(Q) area(P ) =
=
as desired.

k
X
i=1

(i i ) =

k
1 DX

i=1

k
X
i=1

(i i )

(i i )u i , r

hu i , r i
c

1
hw , r i 0,
c


28.6. Exercises.
Exercise 28.1. [1-] Prove the result of Theorem 28.6 without the convexity assumption
on Q, and the assumption that P contains the center O2 of the circle it is inscribed into.
Deduce from here the isoperimetric inequality in the plane (Theorem 7.1).
Exercise 28.2. (Monotonicity of the mean curvature) a) [1+] Let P R3 be a convex
polytope inside a unit ball B. Denote by e the length of edge e, and let e be the dihedral
angle at e. Prove that the mean curvature of P satises: M (P ) 4 .
b) [2-] Let P, Q R3 be two convex polytopes and P Q. Then M (P ) M (Q).

270

c) [1] Dene M (X) = supP X M (P ) to be the mean curvature of a convex body X. Prove
that M (B) = 4.
Exercise 28.3. a) [1+] Let C R3 be a convex cone with a vertex at the origin. Denote
by u 1 , ..., u k the unit vectors along the cone edges, and let 1 , ..., k be the corresponding
dihedral angles. Prove the following inequality:

1
1 u 1 + . . . + k u k (C),
2

where (C) is the curvature of C.

b) [1-] Use the edge summation lemma (Lemma 28.1) and the GaussBonnet theorem
(Theorem 25.3) to obtain part a) of Exercise 28.2.
c) [] Generalize a) to higher dimensions.

Exercise 28.4. Let S1 , S2 R3 be two 2-dimensional polyhedral surfaces. Dene the

Frechet distance distF (S1 , S2 ) as in Exercise 24.18.
a) [2-] When S1 , S2 are convex, prove that


M (S1 ) M (S2 ) 4 distF (C1 , C2 ).

Check that this implies part a) of Exercise 28.2.

b) [2] Suppose S1 and S2 are homeomorphic to a sphere. The total absolute curvature (S)
is the sum of absolute values of curvatures of vertices in S (see Exercise 24.5). Prove that:




M (S1 ) M (S2 ) (S1 ) + (S2 ) + 4 distF (C1 , C2 ).

Exercise 28.5. (Minkowski formula) There is a classical dierential geometry approach

to the mean curvature of surfaces in Rd , but that would lead us away from the subject.
Instead, we present the classical Minkowski formula which can be viewed as an alternative
denition:
Z
Md (P ) =

H(u) d(u),

Sd1

where P Rd is a convex polytope containing the origin O, and H(u) is the support
function dened by H(u) = max{hx, ui | x P }.
a) [1] Prove that M2 (P ) = perimeter(P ).

Now, let us compute M3 (P ) by rewriting the integral as a sum over vertices. In the notation
above, using Rv = Cv S2 , we have:
E
XZ
X DZ
M3 (P ) =
hu, r v i d(u) =
u d(u), r v .
vV

Rv

vV

b) [1+] For a simple cone C R3 calculate the integral

Rv

R u d(u),

where R = C S2 .

c) [1+] Use additivity to compute the integral above for general cones. Write the answer
for each vertex:
Z
X
u d(u) =
e u v,e .
Rv

e=(v,w)E

d) [1] In the edge summation lemma, let fe = e and compute M3 (P ) explicitly.

271

Exercise 28.6. (General Schl

ai formula) a) [1+] Prove the Schl
ai formula for every
(not necessarily isometric) deformation {Tt , t [0, 1]} of a tetrahedron T0 R3 , i.e.,
X
e (t) e (t) = 0, for all t [0, 1].
eE

b) [1] Deduce from here the Schl

ai formula for deformations of all polyhedral surfaces.

Exercise 28.7. (Spherical Schl

ai formula) a) [2-] Let T S3 be a spherical tetrahedron with edge lengths e and dihedral angles e , as above. Suppose {Tt , t [0, 1]} is a
deformation of T . Prove:
1 X
e (t) e (t) = vol(Tt ) .
2
eE

b) [1] Extend this to deformations of general spherical polyhedral surfaces in S3 .

Exercise 28.8. [1+] Let R3 be a xed tetrahedron and let a, b denote its two
edges. When we change the length a of a while keeping other edge lengths xed, the
dihedral angle b at b also changes. Let Lab = b /a . Prove that Lab = Lba , for all a
and b.
Exercise 28.9. a) [1] Let x1 , x2 , x3 , x4 R2 be four points in general position, and let
ij = |xi xj | be the pairwise distances between them. Suppose all distances except for 12
and 34 are xed. Since (x1 x2 x3 x4 ) has ve degrees of freedom, change of 12 inicts a
change of 34 . Prove that
34 d34
12 d12
=
.
area(x1 x2 x3 ) area(x1 x2 x4 )
area(x1 x3 x4 ) area(x2 x3 x4 )
b) [1+] Let x1 , x2 , x3 , x4 , x5 R3 denote ve points in general position, and let ij = |xi xj |
be as above. Suppose all ij except for 12 and 34 are xed. Prove that
12 d12
34 d34
=
.
vol(x1 x2 x3 x5 ) vol(x1 x2 x4 x5 )
vol(x1 x3 x4 x5 ) vol(x2 x3 x4 x5 )
28.7. Final remarks. The angular velocity equation (Lemma 28.2) is standard, see e.g.,
[A2, 10.1]. The edge summation lemma (Lemma 28.1) is a straightforward extension of
the argument in [Pog2]. Corollary 28.4 is due to Alexander [Ale] who presented two proofs.
The rst proof is based on simplicial subdivisions of polyhedra and the additivity property
of the mean curvature. Our proof of the (special case of) Schl
ai formula (Theorem 28.3)
is in fact similar to the second proof of Alexander (see [Ale, 9]).
The local proof presented above is due to Pogorelov [Pog2], who never seems to have
published the proof of his lemma (Lemma 28.5). Pogorelovs lemma was proved in a companion paper [Vol1] by Volkov, which we follow. The combined PogorelovVolkov proof
of the Cauchy theorem is obviously harder than Cauchys proof, since it basically replaces
the straightforward sign counting lemma (Lemma 26.5) with two non-trivial proofs (make
it four, if one counts two supporting lemmas). On the other hand, this approach has the
advantage of being connected to other results, and also gives ground for further investigations.
Most recently, Schlenker found an innitesimal analogue of Pogorelovs lemma, and a
similar in spirit (double counting) proof of the innitesimal rigidity [Schl4] (see Section 32).
His proof of the innitesimal analogue of Pogorelovs lemma is independent of the sign
changes lemma, and is based on some explicit calculation for spherical 4-gons. It would

272

be natural to look for a similar argument to obtain a new proof of Pogorelovs lemma,
independent of the sign changes lemma.
As we mentioned above, the fact that the mean curvature of polyhedral surfaces are
invariant under the exing was shown in [Ale]. This result was rediscovered and extended
in [AlmR], where further connections to the Schl
ai formula were found (see also [SchlS]).
For the history of the Schl
ai formula, references and related results see [AVS, Mil2]. We
refer to [Sant, Ch. 13] for more on mean curvature in higher dimensions, its connections to
dierential and integral geometry, and the proof of the Minkowski formula (Exercise 28.5).
Our proof of Theorem 28.6 follows [Mi2] (see also [Mi3, 1.5]). Note that isoperimetric
inequalities were studied earlier in Section 7 and in fact one can view Theorem 28.6 as a
generalization of the isoperimetric inequality in the plane (see Exercise 28.1).

273

29. Senkin-Zalgallers proof of the Cauchy theorem

29.1. Avoiding counting at all cost. Below we present yet another proof of the
Cauchy theorem (Theorem 26.1) with the opposite philosophy compared to the proof
in the previous section. Previously, of the Cauchy theorem, we replaced a simple
double counting global argument with a delicate and well crafted local argument
(Pogorelovs Lemma 28.5). Now, instead, we give an elaborate geometric global
argument, reducing the problem to a simple looking local lemma, which is a weak
version of the arm lemma (Lemma 22.3). The latter will be transformed in such a way
as to allow an advanced but straightforward global proof. This makes the following
proof fundamentally global, at least as much as such proofs ever are. We begin by
stating the local result.
Lemma 29.1 (Weak arm lemma). Let X = [x1 x2 . . . xn ] and X = [x1 x2 . . . xn ] be
two convex spherical polygons in the upper hemisphere, such that
x1 x2 x3 < x1 x2 x3 , . . . , xn1 xn x1 < xn1 xn x1 , xn x1 x2 < xn x1 x2 .
Then X and X are not isometric, i.e., the following equations cannot hold:
|x1 x2 | = |x1 x2 |, . . . , |xn1 xn | = |xn1 xn | , |xn x1 | = |xn x1 | .
In the notation of Section 26, the lemma says that the number of sign changes
cannot be zero unless all labels are zero. This case appears at the very end of the
proof of the sign changes lemma (Lemma 26.4), and is an immediate consequence
of the arm lemma (Lemma 22.3). Following the pattern of the previous proofs, we
continue with the proof of the Cauchy theorem, and then derive the lemma.
29.2. Step back and take a look from the outside. The idea of this proof is
to use the global convexity of convex polytopes by comparing the distances from the
same vertex to the corresponding points on the surfaces. There are certain restrictions
on this distance function which can be utilized by the following geometric argument.
Proof of the Cauchy theorem modulo the weak arm lemma. Fix any vertex v of the
polytope P , and the corresponding vertex v of P . Denote by S = P and S = P
the polyhedral surfaces of the polytopes, and let : S S be the isometry map as
in the Cauchy theorem (Theorem 26.6). We will use primes in the notation : x x
between points of the surfaces throughout the proof.
Consider the dual cone Cv (see Section 25) and fix a point w Cv (see Figure 29.1).
Denote by f = fv : S R+ and f = fw : S R+ the distance function (in R3 )
between v, w and points of the surfaces: f (x) = |vx|, and f (x ) = |wx |. Consider a
subset G = Gw S of points where function f is greater than f (on the corresponding
points):
G := {x S | f (x) > f (x ), where x = (x)} .
Similarly, let F = Fw S be the set of points where the functions are equal:
F := {x S | f (x) = f (x ), where x = (x)} .

274

To visualize set F , consider a rigid motion of R3 which moves face A into A . Denote
by w
e the image of the point w. The set of points at equal distance from v and w
e is
the plane T which intersects face A either by a face or by an interval with endpoints
at the edges of A, or T does not intersect A at all (see Figure 29.1). In other words,
the set F is a union of faces and straight intervals between the edges.

w
v

w
e
F

Figure 29.1. Point w C and the construction of F .

In the construction above we have some flexibility in the choice of a point w, an
advantage we can exploit. When w = v , we can assume without loss of generality
that G is nonempty. From the contrary, if G = switch the role of P and P . If we
still have G = , then F = S and the polytopes are equal. Indeed, simply note that
the distances between any vertex and the corresponding points on faces determine
the polytope up to a rigid motion.
Now, start moving w away from v in direction of the cone C . Then the set G =
Gw is decreasing since all the distances between w and points in S are increasing.
Since Gv is nonempty, when w is close enough to v the set Gw will remain nonempty.
Now, in order for a vertex u of the polytope P to be in F , the point w must be on a
sphere with radius |uv| centered at u = (u). For points w in general position this
cannot happen.
To summarize, we can choose w is such a way that Gw is nonempty and Fw does
not contain any vertices of P . The set F = Fw is a union of straight intervals between
the edges (see Figure 29.2).
Denote by H a connected component of Gw . From above, the boundary H is
a union of polygons on the surface of S, with polygon vertices on the edges of P ,
but not in the vertices of P . Denote by Z = [z1 z2 . . . zn ] a polygon corresponding to
the exterior boundary of H, defined as a boundary of the only connected component
of S r H containing v. Similarly, define Z = [z1 z2 . . . zn ] = (Z), where zi = (zi ),
the corresponding polygon in S .
By construction, polygons Z and Z have equal sides and the distance from v and w
to the corresponding points is the same: |vzi | = |wzi |, for all 1 i n. Consider a
polygon vertex zi , and let y H be a point on the same edge as zi . Let a = zi1 ,

275

v
H

Figure 29.2. Connected component H of the set G (shaded) on the

surface S of the polytope P . Polygon Z is the exterior boundary of H.
and b = zi+1 be the previous and next vertices of the polygon P (from here on in
this section all indices in subscripts are taken mod n). Define a = (a), b = (b),
y = (y) to be the corresponding points on S (see Figure 29.3).
We can now consider a four-sided cone D starting at zi and spanned by the vectors
(zi , y), (zi , a), (zi , b), and (zi , v). Similarly, consider a four-sided cone D starting at zi
and spanned by the vectors (zi , y ), (zi , a ), (zi , b ), and (zi , w). Comparing D and D
we see that these are cones with isometric sides (see above) and by definition of the
set G y, the distance |vy| > |wy |. This implies that |ab| < |a b |, and that the
dihedral angle in D at (vzi ) is strictly smaller than the corresponding dihedral angle
in D at (wzi ), for every i [n] (see Figure 29.3). This crucial observation allows us
to obtain a contradiction with the weak arm lemma (Lemma 29.1).
v
a

a
zi
y

zi

b
D

b
D

Figure 29.3. Two corresponding four-sided cones D and D .

Let S and S be the unit spheres centered at v and w. Consider X S and X
S the projections of Z and Z from v and w on S and S , respectively. By the

equality of triangles (vzi zi+1 ) and (vzi zi+1

), we have xi vxi+1 = xi wxi+1 , and thus
the spherical polygons X and X have equal corresponding sides. Observe that the
spherical angles xi1 xi xi+1 and xi1 xi xi+1 in X and X are equal to the dihedral
angles in (vzi ) and (wzi ), defined as above. Thus, xi1 xi xi+1 < xi1 xi xi+1 for all
i [n].

276

Finally, note that X lies in the cone Cv and thus can be moved into an upper hemisphere. Similarly, X lies in the cone Cv . Since the cone over X with a vertex at w is
strictly smaller than Cv , the polygon X also can be moved into an upper hemisphere.
By the weak arm lemma (Lemma 29.1), we conclude that such polygons X and X
do not exist, a contradiction. This proves that P Q.

29.3. Killing the monsters. It would be a shame to have a global but rather
intricate proof as above, and a rather unexciting inductive proof of the weak arm
lemma (Lemma 29.1). Fortunately, the fact that the lemma is qualitative rather than
quantitative allows us to create from two polygons a nice imaginary monster, i.e.,
an impossible construction whose nonexistence is easy to establish.
We need a few preliminary definitions and remarks. Let M be a compact metric
space homeomorphic to a sphere. The example to keep in mind is a surface S = P
of a convex body with the distance defined as the length of the shortest path in S
between the points. The advantage of the abstract metric spaces is that it can be
studied without an explicit embedding into Euclidean space. Below we use only the
abstract spherical polyhedral surfaces (s.p.s.) which are defined by gluing along the
edges a finite number of spherical polygons (on a unit sphere S2 ), such that the
resulting manifolds are homeomorphic (but not necessarily isometric) to a sphere.
These are analogues of polyhedral surfaces for spherical polyhedra (we will add an
analogue of convexity below). An example is a spherical tetrahedron which can
be obtained by gluing together four equilateral spherical triangles with the same
sidelength.
Given a s.p.s. M as above, we can define lines, circles, and angles accordingly. At
a point x M, define the curvature (x) = 2 (x) where (x) is the total angle
around x. A point x S is called flat if (x) = 0, and non-flat otherwise. For a
k-gon Q M define
X
(Q) := area(Q) +
(x),
xNF(Q)

where NF(Q) is the set of non-flat interior points x Q. Using these definitions
one can extend the inductive proof of the GaussBonnet theorem (Theorem 25.3) to
obtain that the total curvature of every s.p.s. satisfies (M) = 4.
Finally, we will always assume that our s.p.s. is nonnegatively curved: (x) 0
for all x M. Observe that for every two points x, y M we have |xy| . Indeed,
the shortest path cannot go through points of positive curvature by Proposition 10.1.
Thus, every shortest path M between x and y must lie in a flat neighborhood R
M. Since R is isometric to a region in S2 (possibly overlapping), the path of length
|| > can be shortened. This condition will be important in the proof.
Proof of the weak arm lemma. Let X, X be two spherical polygons as in the lemma.
Remove X from a unit sphere S2 and attach X in its place. Denote by M the
resulting s.p.s. The only non-flat points in M are the vertices of X = [x1 x2 . . . xn ].
Using notation i = xi1 xi xi+1 xi1 xi xi+1 , we have (xi ) = i > 0. Since X is

277

in the upper hemisphere, the s.p.s. M contains a hemisphere B S2 , and xi M r B

for all i [n].
Consider a shortest path between x1 and x2 and cut M along . Let us prove that
the length || = |x1 x2 |M < . Indeed, from above |x1 x2 | . Now, if |x1 x2 | = , we
can consider all shortest paths from x1 to x2 , and show that they either go through
other points xi , or n = 2, both leading to a contradiction.
Construct two symmetric spherical triangles T1 = [a1 a2 a3 ] and T2 = [b1 b2 b3 ] on a
unit sphere U |a1 a2 | = |b1 b2 | = |x1 x2 |, a2 a1 a3 = b2 b1 b3 = 1 /2 and a1 a2 a3 =
b1 b2 b3 = 2 /2. From |x1 x2 |M < , for the third angle we have a1 a3 a2 = b1 b3 b2 <
. Glue T1 and T2 to each other: (a1 a3 ) with (b1 b3 ), (a2 a3 ) with (b2 b3 ). Now glue
sides (a1 a2 ) and (b1 b2 ) of the triangles to the sides of the cut . This gives the
nonnegatively curved s.p.s. M1 with only n 1 non-flat points.
Continue this process to obtain M2 , M3 , etc., until we obtain Mn1 with only one
non-flat point y. The complement Mn1 y is isometric to the complement to a point
of a unit sphere, and has area 4. Therefore, (Mn1 ) = 4 + (y) > 4, which
contradicts the GaussBonnet theorem for s.p.s. (see Exercise 29.1).

Remark 29.2. The condition that the polygons (or at least X ) are in the upper hemisphere
is critical in the weak arm lemma by the same argument as in Section 22.4. Let us see how
this aects the proof above. Consider a sector R on a unit sphere which can be viewed as
a 2-gon with vertices x, y at the North and South Pole, and some angle > 0 between two
meridians. Dene M to be a s.p.s. obtained by gluing two sides of A. Now the curvature
in each vertex is (x) = (y) = 2 , while area(R) = 2. The triangles T1 = T2 have
two angles ( /2) and the third angle is . Therefore, when we glue R and two triangles
together we obtain back the sphere S2 . As expected, no contradiction in this case.
29.4. Exercises.
Exercise 29.1. (GaussBonnet theorem for s.p.s.) [1+] State and prove the analogue
of the GaussBonnet theorem (Theorem 25.3) for spherical polyhedral surfaces.
Exercise 29.2. [1] Check that the proof above extends verbatim to the proof of the Alexandrov uniqueness theorem (Theorem 27.7).
Exercise 29.3 (Alexandrov-Senkin). [1] Let S, S R3 be two intrinsically isometric
polyhedral surfaces with the same boundary Q = S = S . Suppose the origin O is
separated from S, S by a hyperplane, and that both S, S are seen from O from inside
(outside). Then S = S .
Exercise 29.4. [2-] Modify the proof above to obtain the Pogorelov uniqueness theorem
(Theorem 27.8) in the piecewise smooth case.
29.5. Final remarks. The global proof of the Cauchy theorem is due to Senkin [Sen1]
and is not alike any other proof we know. The proof of the weak arm lemma in this section
follows a companion paper by Zalgaller [Zal1]. While not as simple as the arm lemma (even
in the corrected version), it is very insightful and brings to light some useful ideas. We refer
to [BVK] for a slightly expanded presentation of the original proof. The extension of the
Cauchy theorem in Exercise 29.3 is given in [AS].

278

Interestingly, there seems to be a bit of a tradeo: the easier the global part is, the harder
the local part is, and vice versa. Pogorelovs proof in the previous section has a strong local
part, and an easy global part. Similarly, the proof in this section has a reverse emphasis.

279

30. Flexible polyhedra

30.1. The importance of being flexible. To better understand and appreciate the
Cauchy theorem consider what happens with non-convex polyhedra. At this point
there are several alternative definitions that can be accepted, some of them more
interesting than others. We will go over these definitions in this section and see
where they lead us.
30.2. Rigidity and uniqueness of realization. Let S be a 2-dimensional polyhedral surface, defined as a closed compact metric space obtained by gluing a finite
number of triangles. Sometimes it is convenient to use polygons Fi in place of triangles. Of course, this is an equivalent definition since one can always triangulate
polygons Fi . A realization of S is an isometric immersion Q R3 given by a map
f : S R3 , which maps polygons Fi into equal polygons in Q = f (S). In other
words, polygons Fi lie on faces of Q, but they are allowed to intersect as sets in R3 .
Of course, one cannot hope to have a unique realization even when S is convex.
For example, in Figure 30.1 we present two realizations of the same bipyramid, one
convex and one non-convex. Further, for a cyclic polytope Zk as in the same figure, the
number of realizations is exponential in the number of vertices (polytope Zk has k + 3
vertices and 2k1 realizations). On the other hand, it is easy to show that the number
of different realizations of Zk is always finite, thus it does not have any continuous
deformations (cf. Corollary 26.2).
In summary, one cannot hope for uniqueness of realizations, just the lack of continuous families or realizations, corresponding to continuous deformations of polyhedra.

Z4

Figure 30.1. Two realizations of the same bipyramid and a cyclic polytope Z4 .

30.3. Tight polyhedra. A polyhedron P R3 (a 2-dimensional polyhedral surface) is called tight if every plane divides P into at most two connected components.
The definition is equivalent to the following weak convexity condition: every edge
of conv(P ) must lie in P and every vertex v P which is a local minimum of some
linear function : R3 R, must also be a vertex in conv(P ) (see Exercise 30.1).
The tightness condition is so restrictive, the reader might find it difficult to find
any non-convex examples of tight polyhedra. As it happens, all tight polyhedra
homeomorphic to a sphere are convex (see Exercise 30.1), but there are various natural
examples of tight polyhedra of higher genus (see Figure 30.2).

280

Figure 30.2. Surfaces S, S R3 , where S is tight and S is not.

Let us show that the Cauchy theorem does not extend to tight polyhedra. We
present two non-congruent embedded toric polyhedra which are both tight and have
congruent corresponding faces. In other words, we show that being embedded and
tight is not sufficient for the uniqueness of realization.
Consider a convex polygon Q = [v1 . . . vn ] in the right half-plane of the xz-plane.
Rotate Q around the z axis by /2, and 3/2 to obtain four copies of Q, two in the
xz-plane and two in the yz-plane. Connect the corresponding edges in the orthogonal
copies of Q by trapezoid faces to obtain a polyhedron PQ homeomorphic to a torus.
Let A = [a1 . . . a4 ], B = [b1 . . . b4 ] be the following two quadrilaterals:
a1 = (1, 2),

a2 = (2, 4),

a3 = (4, 3),

a4 = (3, 1),

b1 = (1, 1),

b2 = (2, 3),

b3 = (4, 4),

b4 = (3, 2),

(see Figure 30.3). Consider polyhedra PA and PB defined as above. It is easy to

check that the corresponding faces of PA and PB are congruent, while the polyhedra
themselves are not.
a2

b3
b2

a3
a1

B
b4
b1

a4

Figure 30.3. Polygons A and B corresponding to toric polyhedra with

congruent faces.

30.4. Flexible polyhedra. We say that a polyhedral surface S is flexible if there

exists a continuous family {Qt : t [0, 1]} of realizations of S which are (globally)
pairwise non-isometric. In other words, no two realizations Qt , Qt can be moved into
each by a rigid motion. Such realizations are called flexible polyhedra. Below we

281

give an example of a flexible polyhedron73 called the Bricard octahedron, which is a

non-convex realization of a (non-regular) convex octahedron in R3 .
Consider all plane self-intersecting realizations of the polygon with sides a, b, a, b,
where 0 < a < b. By symmetry, such polygon realizations Q can be inscribed into
a circle. Take a bipyramid with the main diagonal orthogonal to the plane and
side edge length c (a + b) (see Figure 30.4). By the argument above, the resulting
polyhedral surfaces are intrinsically isometric and form a continuous family of globally
non-isometric realizations.
Note that the polyhedral surface S of the above realizations is that of a convex
octahedron with four triangles with sides a, c, c, and four triangles with sides b, c, c.
Therefore, S is homeomorphic to a sphere.

Figure 30.4. Three self-intersecting realizations of the same parallelogram, and the Bricard octahedron.
Since the surface S is triangulated, one can think of realizations of S as realizations
of the corresponding graph. Formally, let G be a graph of an octahedron and let L be a
length function taking values a, b, c as in the figure. From above, there is a continuous
family of realizations of (G, L) which cannot be obtained one from another by rigid
motions.

30.5. Flexors. While the Bricard octahedra can be viewed as an obstacle to an extension of the Cauchy theorem (Theorem 26.1) to non-convex polytopes, one can
restrict the set of examples further, by requiring a realization to be an embedding
(i.e., with no self-intersections). Such polyhedra are called flexors and were first constructed by Connelly. This is the most restrictive definition of non-convex polyhedra,
suggesting that the Cauchy theorem really cannot be extended in this direction.
Unfortunately very few constructions of flexors are known, all of them somehow
related to a trick used by Connelly. Being a flexible polyhedron carries already too
many restrictions and flexors do not seem to have any additional properties separating
them from the flexible polyhedra. We skip their constructions.

73In

this section we will use the term polyhedron quite loosely, applying it to all objects at hand.

282

30.6. Flexible polyhedron of higher genus. Let us give an example of a flexible

polyhedron which is not homeomorphic to a sphere. Consider a Renault style polyhedron T shown in Figure 30.5. This polyhedron consists of four symmetric tubes
with equal parallelogram sides, glued together along rhombi which are parallel in R3 .
Observe that T is homeomorphic to a torus and can be continuously deformed by
flattening. We omit the details.

Figure 30.5. Flexible polyhedron T homeomorphic to a torus.

30.7. Flexible spherical polyhedron. As we explain in Section 27.2, if we want

to have any chance of making flexible polyhedra in higher dimensions, we should be
able to make flexible spherical polyhedra in S3 . As it turns out, flexible spherical
polyhedra are easier to construct than (the usual) flexible polyhedra in R3 .
We start with a simple construction of a convex spherical polyhedron P S3 . There
are two ways to think about it. First, one can take a spherical rhombus R = [abcd],
i.e., a polygon with all sides of length on a unit sphere S1 S2 . Think of S1 as
the equator on the unit sphere S3 and consider a bipyramid over R with additional
vertices x, y in the North and South Pole. Another way to think about it is to take
four equal sectors [xayb] with angle on a unit sphere S2 S2 . Then glue them
along the sides, as schematically illustrated in Figure 30.6. Clearly, the rhombus R is
non-rigid, and thus so is P . In fact, at the extremes the polytope becomes flat; more
precisely, we get a doubly-covered spherical sector, a polyhedron of zero volume.

x
c

b R

a
d c
a
P
b

S1

S2
y

d
b

a
b

x
c

a d
c
a
c
b

Figure 30.6. Construction of flexible spherical polyhedra P and Q

283

It is instructive to compare the construction of P with Theorem 27.1 which states

that all spherical convex polyhedra are rigid. At first glance the two may seem
contradictory until one goes into the details. Note a condition in the theorem that
the polyhedron must lie in the upper hemisphere. This was used in the proof of the
arm lemma (Lemma 22.3 and Lemma 29.1), a key ingredient in the proofs of the
Cauchy theorem and its relatives. By construction, polyhedron P is almost there: it
can be rotated so that only points x, y are not in the upper hemisphere. This shows
that the upper hemisphere condition is in fact necessary in Theorem 27.1.74
To obtain a spherical (non-convex) polyhedron in the upper hemisphere, let us
modify the above construction as follows. Note that [abcd] is really unnecessary to
define P ; in fact, these are not the real edges of the polyhedron. To simplify the
construction, choose the rhombus [abcd] now close to the North Pole. It divides the
surface of the polyhedron P into two parts. Rotate the part containing x around
diagonal (bd) to obtain a new top: pyramid with vertex x over the rhombus [a bc d].
Keep the bottom part, containing y unchanged. Now attach the top part and
the bottom part to intermediate triangles (aa b), (aa d), (cc b) and (cc d) (see Figure 30.6). The resulting polyhedron Q is flexible for the same reason as P . On the
other hand, the direction of rotation around (bd) can be chosen in such way so that Q
lies completely in a hemisphere.
30.8. Flexible polyhedra in higher dimension. Unfortunately, very little is known
about flexible polyhedra in higher dimensions. The following construction of a 4dimensional cross-polytope C4 is an interesting variation on the Bricard octahedron
theme.
Think of C4 as dual to a 4-dimensional cube. Denote by V = {a1 , a2 , b1 , . . . , d2 } the
set of 8 vertices of C4 . All pairs of vertices in the graph of C4 are connected, except for
the opposite pairs: (a1 , a2 ), (b1 , b2 ), (c1 , c2 ), and (d1 , d2 ). We think of a 4-dimensional
realization of C4 as of a map f : V R2 R2 , where f (a1 ) = (a1 , a1 ),. . . ,f (d2 ) =
(d2 , d2 ). Visualize the first coordinates on one plane R2 and the second coordinate on
another R2 . We call these plane coordinates.
a1
b1

a2

O1

c1
b2
d1

c2

O2

d2

c1 , c2 , d1 , d2
a1 , a2 , b1 , b2

Figure 30.7. Walzs flexible 4-dimensional cross-polyhedron.

74We

already mentioned in Remark 29.2 what happens to the (weak) arm lemma without this
condition.

284

Now let the plane coordinates of C4 be as shown in Figure 30.7. Here all plane
coordinates either lie on circles or in their centers O1 and O2 . Thus, the pairwise
distances between
p points are either equal to the interval lengths as in the figure, or
equal to = 21 + 22 , where 1 , 2 are the radii of the circles. Note that both
inscribed 4-gons are flexible, as the circle radii change (compare this with the Bricard
octahedron). Thus, we can keep the distance constant while changing the radii of
the circles. Therefore, the above construction gives a flexible polyhedron.
Finally, note that by the argument in Subsection 27.2 every construction of a flexible
d-dimensional polyhedron gives several constructions of flexible (d 1)-dimensional
spherical polyhedra, one per vertex. Of course, in this case all such polyhedra are the
spherical analogues of the Bricard octahedron.
30.9. Polyhedral surfaces with boundary. Now that we started expanding the
class of non-convex polyhedra, there is no reason to stop on closed surfaces. One
can (and, in fact, some people do) consider polyhedral surfaces with boundary. Unfortunately, in this case flexibility is not an exception but a rule. Basically, if one
removes two adjacent faces from a surface of a simplicial polytope one obtains a
flexible surface:
Theorem 30.1. Let P R3 be a simplicial convex polytope, let e be an edge of P ,
and let F1 , F2 be two faces containing e. Then S := P r (F1 F2 ) is a flexible
polyhedral surface.
Of course, the result cannot be extended to all polytopes. For example, if we remove
two adjacent faces of a cube, the resulting surface is still rigid. The proof of the
theorem is an easy application of the Alexandrov existence theorem (Theorem 37.1),
and is presented in Subsection 37.3.
30.10. Exercises.
Exercise 30.1. (Tight polyhedra) a) [1+] Prove that a polyhedron P R3 is tight if and
only if every edge of conv(P ) must lie in P and every vertex v P which is a (strict) local
minimum of some linear function : R3 R must be a vertex of conv(P ).
b) [1-] Show that neither of the two conditions in a) suce for the tightness.
c) [1-] Check that the two toric polyhedra constructed in Subsection 30.3 are tight.
d) [1] Prove that every tight polyhedron embedded into R3 and homeomorphic to a sphere
is convex.75
Exercise 30.2. [2-] Generalize to higher dimensions the construction of a toric polyhedron
in Subsection 30.6. Check whether they are exible or rigid.
Exercise 30.3. [1] Consider the Hooker polyhedron dened in the US Patent 3894352.76
Is it exible or rigid?
Exercise 30.4. a) [] Find examples of exible polyhedra in Rd , for all d 5.
b) [] Prove or disprove: every realization of a cross-polytope in R5 is rigid.
75One
76See

can view this result as a variation on the convexity criterion given in Exercise 1.25.
http://www.google.com/patents?vid=USPAT3894352.

285

30.11. Final remarks. The construction of tight toric polyhedra with isometric faces is due
to Bancho [Ban1]. See [Kuh] for more on tight polyhedra, generalizations and references.
We refer to [Con2, Con4, FucT] for an introduction to the subject of exible polyhedra
and easy-to-construct examples (see also [Ale4, Dol, Sab5]). Let us note also that Bricard
completely characterized all exible octahedra in 1897; there are in fact two additional
families dierent from the construction we present in this section [Bri2] (see also [Leb3]).
We refer to a well written survey [Ale6, 7] for detailed constructions and the references.
Flexible polyhedron homeomorphic to a torus in Subsection 30.6 is one of a large family
of exible polyhedra introduced by Goldberg [Gol2] (see also [Ale2]). Flexible spherical
polyhedron in Subsection 30.7 is due to V. A. Alexandrov [Ale3]. Let us mention also that
exible polyhedra exist in the hyperbolic space H3 , where the analogues of the Bricard
octahedra can be proved to be exible [St2].
The construction in Subsection 30.8 is due to A. Walz (unpublished). Our presentation
follows [St1], where the author gives a more general construction of exible cross-polytopes.
While no examples of exible polytopes in higher dimensions are known at the moment,
we do expect the multitude of examples. On the other hand, Stachel conjectured that in
dimension d 5 there are no exible cross-polytopes [St1].

286

31. The algebraic approach

The algebraization of realization spaces introduced earlier in the context of linkages
(see Section 13) will prove critical to understand the rigidity of non-convex polyhedra. In this section we prove Glucks theorem and set up a general approach for
infinitesimal rigidity in the next two sections, as well as the bellows conjecture.
31.1. Rigidity vs. flexibility: what is more likely? Let P be a convex polytope
and S = P be its polyhedral surface. The Cauchy theorem states that S is rigid.
On the other hand, as the Bricard octahedron and the Connelly polytope show, this
is no longer true for immersions and even for embeddings of S (see the previous
section). The question is what is more likely: that a surface is rigid or flexible?
Perhaps unsurprisingly, the answer is unambiguous: a polyhedral (not necessarily
convex) surface in R3 is almost surely rigid.
We start with the definitions. Let G = (V, E) be a plane triangulation, i.e., a connected planar graph with all faces (including exterior face) triangles. Let |V | = n. By
Eulers formula we have m = |E| = 3n 6 (see Corollary 25.2). Suppose we are given
a length function L : E R+ . For each triangle in the plane triangulation G, make
a metric triangle with edge lengths given by the length function L. Now glue these
triangles along the corresponding sides to obtain a simplicial surface S homeomorphic
to a sphere.
Clearly, not all length functions define a metric space as the triangle inequality
must be satisfied. Denote by L(G) the set of all length functions leading to a metric
space. The set L(G) forms a convex cone in Rm with facets corresponding to triangle
inequalities for each
P triangle in G. Denote by L1 (G) the set of all length functions L
L(G) such that eE L(e) 1. From above, L1 (G) is a convex polytope in Rm .
3
 As before,
 a realization of (G, L) is a map f : V R , such that the actual distance
f (v)f (w) = L(e) for every edge e = (v, w) E. Given L L(G), such a realization
defines an (intrinsically) isometric immersion of S into R3 . In the notation above,
realizations of (G, L) now correspond to realizations of the surface S.
Finally, rigid motions in R3 act naturally on realizations of (G, L) by acting on
sets {f (v) | v V }. We are now ready to state the main result.
Theorem 31.1 (Gluck, Poznjak). Let G be a planar triangulation and let L be a
random length function in L1 (G). Then, almost surely, (G, L) has only a finite number
of realizations, up to rigid motions.
In other words, the set of non-rigid realizations has measure zero in L(G). We
prove the result by an algebraic argument later in this section.
31.2. The reason why metallic bellows do not exist. The bellows, as any dictionary will explain, are mechanical devises which by expansion and contraction pump
air. Basically, these are surfaces which change the volume under bending. Of course,
in theory it would be nice to make these surfaces polyhedral, with all faces made out
of some kind of metal. In practice this does not work. Here is why (a theoretical
version).

287

We start with a celebrated bellows conjecture recently resolved by Sabitov. It states

that the volume of flexible polyhedra is invariant under continuous deformations:
Theorem 31.2 (Former bellows conjecture). Let {Qt : t [0, 1]} be a continuous
deformation of a closed polyhedral surface S homeomorphic to a sphere. Then the
volume vol(Qt ) is independent of t.
We prove this theorem in Section 34. In fact, the result holds for all closed orientable
surface immersed in R2 (see Subsection 34.7). To illustrate the theorem, note that
the (signed) volume of the Bricard octahedron and the toric polyhedron T defined in
Subsection 30.6, is equal to zero by symmetry. Of course, one can attach a rigid convex
polyhedron to the surface of the either of the two, to obtain a flexible polyhedron of
positive volume, still constant under flexing.
31.3. Constructing polytopes from scratch. Let P R3 be a simplicial convex
polytope with n vertices: V = {v1 , . . . , vn }. For simplicity, let us assume that (v1 v2 v3 )
is a face of P . Observe that there always exists a rigid motion which maps vertex v1
into the origin O = (0, 0, 0), v2 into (a, 0, 0), and v3 into (b, c, 0), for some a, b, c R+ .
Note also that up to reflection such rigid motion is unique. We call the resulting
polytope planted.
Now, in order to construct a planted polytope from its graph and edge length we
need to set up a system of algebraic (in fact, quadratic) equations. By the Cauchy
theorem, there exists only one convex solution, so we are done. Of course, the first
part of the plan is easy to set up as follows.
Formally, let G = (V, E) be a plane triangulation and let L : E R+ be a length
function, L L(G). Denote by f : vi (xi , yi , zi ), i [n], realizations of (G, L). For
every pair of vertices (vi , vj ) consider a polynomial
Fij := (xi xj )2 + (yi yj )2 + (zi zj )2 .
We consider only planted solutions, i.e., values (x1 , . . . , zn ) R3n where
()

x1 = y1 = z1 = y2 = z2 = z3 = 0.

Denote by R = C[x2 , x3 , . . . , xn , y3 , . . . , yn , z4 , . . . , zn ] the ring of polynomials on the

remaining variables. Now, in the ring R we have the system of equations:


()
Fij = (ij )2 , for all e = (vi , vj ) E and ij = L(e) .

As we mentioned before, () is a system of m = 3n 6 equations with 3n 6

variables. Therefore, if the equations are algebraically independent, we have a finite
set of solutions. The following result positively resolves the problem:
Theorem 31.3 (Gluck). Let G = (V, E) be a plane triangulation. Then polynomials Fij corresponding to edges (vi , vj ) E, vi , vj V , are algebraically independent.
We prove the theorem later in this section. First, let us mention the following
corollary, which follows easily from the theorem.

288

Corollary 31.4. Under conditions of Theorem 31.3, every polynomial g R is a

root of a nontrivial algebraic equation:
()

cN tN + . . . + c1 t + c0 = 0,

where cr C[Fij , (vi , vj ) E], 0 r N.

We call polynomial relations () in the corollary the Sabitov polynomial relations.
To obtain the corollary simply note that transcendency degree of R is m = 3n 6,
equal to the number |E| of algebraically independent polynomials Fij in the theorem.

Remark 31.5. (Does this give an algorithm? ) It was shown in [FedP] that the degree
of Sabitov polynomials () is at most 2m and can be exponential even in the most simple
cases (see Exercise 31.5). Starting with edge polynomials, one can use standard computer
algebra tools to determine (numerically) the coordinates of all vertices.

31.4. Why polynomials are always better than numbers. Let us continue our
construction of the polytope. For each coordinate xi , yi , and zi we need to compute
the Sabitov polynomial relations (). Evaluating polynomials Fij at 2ij makes ()
into polynomials of the desired values. Compute their roots and try one by one all
resulting combinations until a convex one is found.
Alternatively, and more invariantly, we can compute Sabitov polynomials for all
diagonals F1i , F2i and F3i . Again, after their lengths are determined to belong to a
certain finite set of solutions, we have a finite number of possibilities to consider.
Now, the above argument clearly contradicts the existence of flexible polyhedra
(say, with self-intersections) since it implies that there is always only a finite number
of solutions. The mistake in this argument is very important and may not be obvious
at first sight.
. . . . . . [Think about it for a few minutes!]

... ...

The mistake is that all polynomial coefficients ci in the Sabitov polynomial relations () may become zero when evaluated at {ij }. When this happens, we cannot
determine the corresponding diagonal lengths, thus allowing for the flexible polyhedra. On the other hand, if at least one ci 6= 0 in all Sabitov polynomials for diagonals
as above, then there exists only a finite number of planted realizations. In particular,
all realizations are (continuously) rigid.
We are now ready to prove Theorem 31.1. First, observe that the polytope L1 of
length functions has full dimension m = 3n 6 (see next subsection). On the other
hand, for every diagonal (vi , vj ), the set of roots of the equation cr = 0 (over R) has
codimension at least 1, where cr is a coefficient of () corresponding to Fij . Thus,
all relations () are nonzero almost surely, and there exists only a finite number of
realizations of (G, L), as desired.
Let us think about what is needed to prove Theorem 31.2, the former bellows
conjecture. First, one has to check that vol(P ) is a polynomial in the ring R, i.e.,
depends polynomially on the vertex coordinates. Then, by Corollary 31.4 it satisfies
a Sabitov polynomial relation (). If one can prove that the coefficient cN is nonzero,

289

then the volume takes only a finite number of values, and, therefore, remains constant
under continuous deformations. While proving polynomiality of the volume is an easy
exercise, checking that cN 6= 0 is a major task; we prove this in Section 34.
31.5. Proof by an algebraic manipulation. Let us first show that Theorem 31.3
follows easily from the Cauchy theorem and the Steinitz theorems. The proof is trivial
in essence, but requires a good absorbtion of the results and definitions involved.
Proof of Theorem 31.3. Let G = (V, E) be a plane triangulation as in the theorem.
Recall that by the Steinitz theorem (Theorem 11.1) there exists a convex polytope P
with graph G. We can assume that P is planted; otherwise use a rigid motion to
make it so. Perturbing the vertex coordinates of P by < , while keeping P planted,
does not change convexity and the graph of the polytope, for sufficiently small > 0.
By the Cauchy theorem, all perturbations give different length functions L : E R+
of these realizations of P . In other words, we obtain an open subset X Rm of
perturbations of coordinates, each giving a length function L L(G) Rm .
Now consider a characteristic map F = (. . . , Fij , . . .) : Rm Rm , where Fij will
always correspond to graph edges (vi , vj ) E. The Jacobian J() in this case is
a determinant of a m m matrix of partial derivatives of Fij by the coordinates.
Since Fij are polynomials, the Jacobian is also a polynomial. From above, J() does
not vanish on an open subset X Rm of the coordinates. Therefore, the Jacobian J()
is a nontrivial polynomial, and polynomials Fij are algebraically independent.

Remark 31.6. The proof above is based on the Cauchy theorem, but uses a local rather
than a global argument. On the other hand, the (continuous) rigidity is insucient to make
the argument work, since in general the Jacobian can vanish at a point while a polyhedral
surface remains continuously rigid. The examples are given in the previous section. This
will also lead us to a notion of static and innitesimal rigidity which are stronger than
continuous rigidity, but somewhat easier to prove than the Cauchy theorem.

31.6. Exercises.
Exercise 31.1. [1+] Let P R3 be a bipyramid with side length c, over an inscribed
n-gon in the plane, with side lengths a. Find a minimal polynomial for the volume of P
with coecients in C[a2 , c2 ].
Exercise 31.2. Denote by Qn (a, b) the radius of a convex polygon inscribed into a circle
with n 1 sides a and one side b. Let rn (a, b) be the radius of the this circle, i.e., the radius
of the circle circumscribed around Qn (a, b).
a) [1+] Compute a formula for rn (a, b) as a root of a polynomial equation in a2 and b2 in
terms of the Chebyshev polynomials.
b) [1+] Compute the minimal degree of such polynomial relation for rn (a, b).
Exercise 31.3. (Robbinss problem) Denote by A(a1 , . . . , an ) and R(a1 , . . . , an ) the area
and the circumradius of an inscribed convex polygon with sides ai (see Example 34.6 and
Exercises 34.1, 34.6).
a) [1] Prove that A2 () and R2 () are algebraic over C[a21 , . . . , a2n ].

290

b) [1+] Prove that in both cases the degrees n and n of the minimal polynomial is at
least Tk when n = 2k + 1, and 2Tk when n = 2k + 2, where Tk is the number of (possibly
self-intersecting) polygons inscribed into a circle.
c) [1+] Prove an explicit formula for Tk :
 
2k + 1 2k
22k1 .
Tk =
2
k
d) [2] Give an explicit formula for n . Check that it proves the upper bound matching the
lower bound in b).
e) [2] Prove that n = n .
Exercise 31.4. (Connellys suspensions) Dene a suspension to be a bipyramid over an
inscribed polygon (as in Bricards octahedron).
a) [2] Prove that in every exible suspension, the inscribed (self-intersecting) polygon must
have each edge length repeated twice, once in each direction.
b) [1] Check that such polygons have zero area (cf. Exercise 34.6). Conclude that every
exible suspension has a zero volume.
Exercise 31.5. (Degrees of Sabitov polynomials) a) [1] Let Qn denote the bipyramid
over an n-gon and let n be the degree of the Sabitov polynomial of the main diagonal.
Prove that n = n (see Exercise 31.3). Conclude that n is exponential.
b) [1] Let n denote the degree of Sabitov polynomial of the main diagonal in the cyclic
polytope Zn (see Figure 30.1). Prove that n is exponential.
c) [2] Use a Bezout type result to show that degrees of Sabitov polynomials () is at most 2m .
Exercise 31.6. Let = (v0 v1 v2 v3 ) R3 be a tetrahedron, and let ij be its edge lengths,
0 i < j 3.
a) [1-] Find positive values {ij , 0 i < j 3} which satisfy the triangle inequalities
ij + jk > ik ,
and such that no tetrahedra with these edge lengths exist.
b) [1] Consider the set of all possible edge lengths of tetrahedra in R3 :
L = {(01 , . . . , 34 )} R6 .
Prove that L is not convex.
c) [2-] Consider the set of all possible square edge lengths of tetrahedra in R3 :
S = {(201 , . . . , 234 )} R6 .
Prove that S is convex.
Exercise 31.7. Let = (v1 v2 v3 v4 ) R3 be a tetrahedron, and let ij denote the dihedral
angle at the edge (vi , vj ), for all 1 i < j 4.
a) [1] Prove that 12 + 23 + 34 + 14 2.
b) [1-] Prove that 2 12 + 13 + 14 + 23 + 24 + 34 3.
c) [1] Prove that 0 cos 12 + cos 13 + cos 14 + cos 23 + cos 24 + cos 34 2.
d) [1-] Prove that the second inequality in c) is an equality only if is equihedral.
e) [1+] Show that part a) holds also in the hyperbolic space H3 .

291

31.7. Final remarks. Versions of Theorem 31.1 have been proved in much greater generality, e.g., for closed surfaces of every genus. This result is usually attributed to Gluck [Glu],
but has also appeared in an earlier paper by Poznjak [Poz].
We prove the bellows conjecture (Theorem 31.2) and give references in Subsection 34.7.
To further appreciate Sabitovs theorem, compare it with the following Connellys conjecture: In the conditions of Theorem 31.2, polytopes enclosed by Qt are scissor congruent
(see [Con2, Con4]). A version of this conjecture was recently disproved in [AleC].
Finally, let us mention that the algebraic approach to the subject and the method of
places recently found unexpected applications to the Robbins conjectures (see Exercise 31.3
and 34.6). These conjectures are concerned with properties of polynomial relations on areas
of inscribed convex polygons, as functions of squares of its sides. We refer to [FedP] for
connections between two problems and an algebraic background, and to [Pak4] for a short
survey and further references (see also Exercise 12.3).

292

32. Static rigidity and Dehns theorem

This is the first of two sections where we prove Dehns theorem, an infinitesimal version of the Cauchy rigidity theorem. We include two proofs: a variation on Cauchys
original proof in Section 26 and Dehns original proof.
32.1. Who needs rigidity? In the next two sections we introduce two new concepts, the static and the infinitesimal rigidity of convex polyhedra, which turn out
to be equivalent to each other and imply continuous rigidity. These ideas are crucial
in modern rigidity theory; their natural extensions to general frameworks (of bars,
cables and struts) were born out of these considerations and have a number of related
properties. While we spend no time at all on these extensions, we find these ideas
useful in discussions on Cauchys and Glucks theorems.
To summarize the results in the next two sections, we show that Glucks Theorem 31.3 follows from Dehns lemma on the determinant of a rigidity matrix, which
we also introduce. We also show that the continuous rigidity also follows from Dehns
lemma. We then present three new proofs of Dehns lemma, all without the use of the
Cauchy theorem, as well as one extra proof of continuous rigidity of convex polyhedra.
As the reader shall see, all this is done and motivated by the two rigidity concepts.
32.2. Loading the edges. To define the static rigidity, we need to extract the key
ingredient in the proof of Glucks theorem we presented in the previous section.
Let V = {v1 , . . . , vn } be the set of vertices of a plane triangulation = (V, E), and
denote by f : V R3 its planted realization. Let E be a set of ordered pairs: if
(vi , vj ) E, then (vj , vi ) E as well. Now, for every edge e = (vi , vj ) E, denote by

e ij = f (vi )f (vj ) = (xj xi , yj yi , zj zi ) R3

the corresponding edge vector in the realization. In this notation, e ij = e ji, for
all (vi , vj ) E. The set of scalars {ij R, (vi , vj ) E} is said to be an edge load if
ij = ji , 12 = 13 = 23 = 0, and
X
ij e ij = 0, for all i [n].
j : (vi ,vj )E

We say that a planted realization f : V R3 of (V, E) defining the polytope is

statically rigid if there is no nonzero static load {ij }. Finally, a simplicial convex
polytope P with graph = (V, E) is statically rigid if so is the planted realization
of obtained by a rigid motion of P . The following result is the key result of this
section.

Theorem 32.1 (Dehns theorem; static rigidity of convex polytopes). Every simplicial convex polytope in R3 is statically rigid.
We already proved this result in a different language. To see this, consider a ma(ij)
trix R with rows R corresponding to edges (vi , vj ) E, written in lexicographical
order:
(ij)

= (. . . , xi xj , yi yj , zi zj , . . . , xj xi , yj yi , zj zi , . . .).

293

The matrix R is called the rigidity matrix. Now observe, the Jacobian J() is a
determinant of the matrix with the following rows:


Fij Fij Fij
(ij)
,
,
, . . . = 2 R .
dFij = . . . ,
xr yr zr

We showed that the Jacobian J() = 2m det R 6= 0, when evaluated at a planted

realization f : R3 defined by a convex polytope P . Thus, there is no nonzero
linear combination of the rows of the rigidity matrix, with coefficients ij as above.
Interpreting the set of coefficients {ij } as the edge load, we obtain the statement of
Dehns theorem.
In the opposite direction, given Theorem 32.1, we obtain that det R 6= 0. Therefore, the Jacobian is nonzero, which in turn implies Glucks theorem without the use
of the Cauchy theorem. To conclude this discussion, the static rigidity of convex
polytopes is equivalent to the following technical statement.
Lemma 32.2 (Dehn). Let P R3 be a simplicial convex polytope with a graph
= (V, E). Then the rigidity matrix R is nonsingular.
In the following two subsections we present three independent proofs of Dehns
lemma, all (hopefully) easier and more elegant than any of the previous proofs of the
Cauchy theorem. Until then, let us make few more comments.
First, let us show that Dehns lemma implies the (continuous) rigidity of convex
polytopes (Corollary 26.7). Indeed, consider the m-dimensional space W of planted
realizations of = (V, E). The space W is mapped onto the m-dimensional space
of all length functions, and the determinant J() = 2m det R is nonzero at convex
realizations. Therefore, in a small neighborhood of a convex realization the edge
lengths are different, and thus a simplicial convex polytope is always rigid.
Our second observation is that the Cauchy theorem is more powerful than Theorem 32.1. To see this, recall the CauchyAlexandrov theorem on uniqueness of convex
polyhedral surfaces (Theorem 27.6). This immediately implies the rigidity of these
surfaces77. On the other hand, such realizations are not necessarily statically rigid,
as the example in Figure 32.1 shows. Here we make arrows in the directions with
positive coefficients which are written next to the corresponding edges.
32.3. Proof of Dehns lemma via sign changes. This proof goes along the very
same lines as the traditional proof of the Cauchy theorem (see Section 26.3). We first
show that the edge load {ij } gives a certain assignment of signs on edges, then prove
the analogue of the sign changes lemma (Lemma 26.4), and conclude by using the
sign counting lemma (Lemma 26.5).
Proof of Dehns lemma. Consider an edge load {ij | (vi , vj ) E} on the edges in P .
To remove ubiquity, consider only coefficients ij with i < j. Let us label the edge
(vi , vj ) E, i < j, with (+) if ij > 0, with () if ij < 0, and with (0) if ij = 0.
77One

has to be careful here: this only proves rigidity in the space of convex realizations. In fact,
the continuous rigidity holds for all non-strictly convex realizations; this is a stronger result due to
Connelly (see [Con5])

294

1
2

1
2

2
1

S
Figure 32.1. Stresses on a non-strictly convex polyhedral surface S
show that it is not statically rigid.
Lemma 32.3 (Static analogue of the sign changes lemma). Unless all labels around
a vertex are zero, there are at least four sign changes.
By Lemma 26.5, we conclude that all labels must be zero. This proves Dehns
lemma modulo Lemma 32.3.

Proof of Lemma 32.3. Denote by e 1 , . . . , e k the edge vectors of edges leaving vertex w
of a convex polytope P . We assume that w is at the origin and the edge vectors are
written in cyclic order. Suppose we have a nonzero linear combination
u := 1 e 1 + . . . + k e k = 0.
Denote by H any generic hyperplane supporting P at w, i.e., a hyperplane containing w, such that all vectors e i lie in the same half-space. If there are no sign changes,
i.e., 1 , . . . , k 0 or 1 , . . . , k 0. Then their linear combination u is also in the
same half-space unless all i = 0, a contradiction.
Suppose now that there are two sign changes, for simplicity 1 , . . . , i 0 and
i+1 , . . . , k 0. Denote by H a hyperplane which contains vectors e 1 , . . . , e i in a
half-space H+ , and e i+1 , . . . , e k in the other half-space H . Then the linear combie3
e2

e4
H
e1
e5

Figure 32.2. Hyperplane H separating edges in a polytope P .

295

nation u is in H+ unless all i = 0, another contradiction (see Figure 32.2). Thus,

there are at least four sign changes unless all labels are zero.

32.4. Proof of continuous rigidity from the angular velocity equation. Before
we continue with other proofs of Dehns lemma, let us show how continuous rigidity
(Corollary 26.7) follows from the angular velocity equation (Lemma 28.2), which was
proved in Section 29 by a simple independent argument. The proof will be almost
completely the same as the above proof of Dehns lemma.
Proof of Corollary 26.7 modulo Lemma 28.2. Consider the angular velocity equation
for each vertex and assign labels to all edges according to the signs of derivatives e (t).
By the proof of Lemma 32.5, either all labels around a vertex are zero, or there are at
least four sign changes. Now use the sign counting lemma (Lemma 26.5) to conclude
that all labels bust be zero, i.e., all derivatives are zero (more precisely, all left and
right derivatives at each point are zero, which is equivalent). Thus the dihedral
angles remain constant under the continuous deformation, and the deformation itself
is a rigid motion.

32.5. Graph-theoretic proof of Dehns lemma. Let = (V, E) be a plane triangulation, and let R (. . . , xr , yr , zr , . . .) be the rigidity matrix defined above. To
prove Dehns lemma we compute D = det(R ) and show that it is 6= 0 for convex
realizations.
Let us use the fact that most entries in R are zero. Observe that every 3 3
minor of R either contains a zero row or column, or two columns which add up to
zero, or looks like

xa xb ya yb za zb
M(a | b, c, d) = xa xc ya yc za zc ,
xa xd ya yd za zd

where a, b, c, d represent distinct integers in [n]. Here we assume that b < c < d and
the ordering on rows corresponding to edges (vi , vj ) E is lexicographic. In addition
to these minors, there is one special non-degenerate 3 3 minor of R , with rows
corresponding to the edges (v1 , v2 ), (v1 , v3 ), and (v2 , v3 ), and the columns to x2 , x3 ,
and y3 :

x2
0
0
x3
y3 .
M(1, 2, 3) = 0
x2 x3 x3 x2 y3
Now, using the Laplace expansion for det R over triples of rows we conclude that
the determinant D is the product of determinants of the 3 3 minors as above, each
given up to a sign. Since we need these signs let us formalize this as follows.
We say that vertices v1 , v2 , v3 are base vertices and the edges between them are base
edges. A claw in is a subgraph H of isomorphic to K1,3 , i.e., a subgraph K(a |
b, c, d) consisting of four distinct vertices va , vb , vc , vd and three edges: (va , vb ), (va , vc ),
and (va , vd ). We call vertex va the root of the claw K(a | b, c, d). Recall that

296

contains (3n 6) 3 = 3(n 3) non-base edges, exactly three per non-base vertex
(cf. Corollary 25.2). One can ask whether non-base edges of can be partitioned into
claws, with a root at every non-base vertex. Denote such claw partitions by , and
the set of claw partitions by . We have:
X
Y
D = det R = det M(1, 2, 3)
()
det M(a | b, c, d),

H(a|b,c,d)

where () {1} and is given by the sign of the corresponding permutation ()

Sm of the set of edges E. Now, observe that each determinant in the minor as above
can be written as a volume of a parallelepiped spanned by the edge vectors:
1
det M(a | b, c, d) = volhe a,b , e a,c , e a,d i.
6
Fix an orientation of the surface S = P induced by R3 . By convexity of P ,78 the
determinant is positive if the order of edges (va , vb ), (va , vc ), (va , vd ) coincides with the
orientation of S, and negative if it is the opposite. We call such claws K(a | b, c, d)
positive and negative, respectively. For a claw partition , denote by s() the
number of negative claws H , and let () = (1)s() . We need the following two
lemmas:
Lemma 32.4 (Existence of claw partitions). For every plane triangulation and
every base triangle, the set of claw partitions is nonempty.
Lemma 32.5 (Signs of claw partitions). Every two claw partitions , have
the same products of signs: () () = ( ) ( ).
Now the claim follows easily from the lemmas: the determinant D is a nonempty
sum of products, all of the same sign. Thus, D 6= 0, which completes the proof of
Lemma 32.2 modulo the above two lemmas. 
Proof of Lemma 32.4. Let us orient the non-base edges in a claw away from the root.
Every non-base vertex must have exactly three outgoing edges then, which determine
the claw partition. Use induction on the number n of vertices in . For n = 4,
when = K4 the result is obvious.
For n > 4, let a be a vertex of degree at most 5. This vertex exists, since otherwise
the number of edges |E| = 3n6 (6 n)/2 = 3n. Remove vertex a with all adjacent
edges, and denote by any triangulation of the remaining graph. Since has a claw
partition by inductive assumption, return the removed vertex and modify the edge
orientation as in Figure 32.3. There are five cases to consider: when deg(a) = 3,
when deg(a) = 4, and three cases corresponding to different orientations of diagonals
in a pentagon, when deg(a) = 5.
All the cases are straightforward except for the last one, with two diagonals leaving
from the same pentagon vertex v. Since at most three edges can leave v, one of the
two side edges of the pentagon, say (w, v), must go into v. Then make the change
as in the figure. We should add that when triangulating the pentagon, we need to
78Incidently,

this is the only use of convexity of P in the whole proof.

297

v
w

v
w

Figure 32.3. Inductive proof of existence of claw partition.

make sure that v is not a base vertex; this is easy to do since there are only three
base vertices.

Proof of Lemma 32.5 is motivated by the following example.
Example 32.6. It is easy to see that for every base triangle in the graph of an octahedron,
there exist exactly two claw partitions. In Figure 32.4 we show both of them and how they
are connected by reversal of a cycle of length 3. Here the base vertices are the vertices of
the outside triangle. We also show a claw partition of the graph of an icosahedron and an
oriented cycle of length 6. Reversal along this cycle gives a dierent claw partition.

Figure 32.4. Two claw partitions of an octahedron and a claw partition of an icosahedron.
We say that two claw partitions s , are connected by a 3-cycle reversal if
they coincide everywhere except for three edges which belong to different claws and
form a 3-cycle as in the example. We need the following lemma.
Lemma 32.7. Every two claw partitions , can be connected by a sequence
of claw partitions:
= 0 1 . . . 1 = ,

298

such that i , i [], and every i is obtained from i1 by the reversal of edges
in a 3-cycle.
Note that the number of number s() of negative claws in remains invariant under
a 3-move reversal, so the sign () is constant on all . Similarly, because 3cycle is an even permutation, the sign () is also constant. Therefore, Lemma 32.7
immediately implies Lemma 32.5.
Proof of Lemma 32.7. Let e1 = (vi , vj ) E be an edge in which belongs to a claw
with root vi in partition and belongs to a claw with root vj in partition . If
there are no such edges, then = , and there is nothing to prove. Think of e1
as an oriented edge: it belongs to , but not to . Clearly, there also exists an
edge e2 = (vj , vr ) which belongs to , but not to , etc. Continuing in this manner,
we eventually return to the starting vertex vi , and obtain an oriented cycle of edges
C = (e1 , e2 , . . . , ek ). Note that if we reverse the direction of the edges, we obtain an
oriented cycle in (see a cycle on the right in Figure 32.4).
Note that the reversal operation of edges in an oriented cycle does not move from
to . Instead, we obtain a new claw partition 1 . Now, comparing 1 to we can
find a new oriented cycle of edges in but not in , whose reversal will create a new
partition 2 . Repeating this procedure we obtain a sequence of claw partition with
more and more edges oriented as in , and eventually reach :
()

1 2 . . . .

We can also assume that all cycles reversed at each step are not self-intersecting since
otherwise we can split each such cycle at the intersection point (see Figure 32.5), and
further refine sequence ().

Figure 32.5. Splitting a self-intersecting cycle into two cycles.

In summary, we obtain a sequence () of claw partitions, each subsequent obtained
by a reversal along some oriented not self-intersecting cycle C. Let further refine this
sequence, by writing each reversal of edges of C as a composition of reversals along
3-cycles.
Clearly, if cycle C is removed from the surface S = P one part will contain the
base triangle (v1 v2 v3 ) Denote by H an induced subgraph of obtained by removal
of all interior vertices in that part. Suppose that H = (V , E ) contains m interior
vertices. Note that H is planar, and consists of triangles and one k-gonal face. We
have:
|V | = m + k, 2 |E | = k + 3t, and |F | = t + 1 .

299

By Eulers formula |V | |E | + |F | = 2, graph H has exactly t = 2m+ k 2 triangles

and
1
|E | = (k + 3t) = 3m + 2k 3
2
edges. Observe that there are 3m edges in claws rooted at the interior m vertices,
and neither of these edges can belong to the cycle C. Thus, the remaining (2k 3)
edges consist of k edges in C, and of s = (k 3) edges rooted at vertices of C and
oriented inside of H. Therefore, every k 4 there exist at least one such edge e (see
Figure 32.6).
Consider a path starting with e, obtained as in the argument above. Clearly,
we eventually reach C again. Two sides of path will divide C into two parts, and
reversal along C is the composition or reversals along one of them, call it C1 , and
then along the other, call it C2 (see Figure 32.6). Since cycle C1 , C2 have smaller
area than C, we can continue splitting of the cycles until we have a composition of
3-cycles, refining ().


C1

C2
C
Figure 32.6. Reversal of a long cycle.

32.6. Exercises.
Exercise 32.1. (Rigidity of non-convex polyhedra with saddle vertices) a) [1] Let P be a
non-convex polyhedron in R3 homeomorphic to a sphere, but possibly with self-intersections.
We say that a vertex is a saddle, if it is adjacent to exactly 4 edges, and there exists a
hyperplane separating two opposite edges from the other two. Check that Lemma 32.3
extends to saddle vertices.
[1] We say that a vertex v in P is convex if the corresponding cone Cv is convex. Check
that the proof above extend verbatim to polyhedra homeomorphic to a sphere, with only
convex and saddle vertices. An example is given in Figure 32.7.

Figure 32.7. A barbell style polyhedron with only convex and saddle vertices.

300

Exercise 32.2. [2] Let P R3 be a simplicial convex polytope and let S = P . Suppose {St } is a polyhedral deformation which preserves combinatorial structure and dihedral
angles. Prove that {St } is a rigid motion.79

Exercise 32.3. a) [2-] Let P R3 be a convex polytope with even-sided faces (quadrilaterals, hexagons, etc.) with surface S = P . Suppose S R3 is an embedded surface
which is combinatorially equivalent to S and whose faces are isometric. Prove that S is
continuously rigid.
b) [2-] Extend part a) to polyhedra with at most 7 odd-sided faces.
Exercise 32.4. (Alexandrov ) [1+] Let S = P be a triangulated surface of a convex
polytope P R3 . Suppose there are no triangle vertices in the relative interior of faces
(they are allowed only at vertices and on natural edges of P ). Prove that S is statically
rigid.

32.7. Final remarks. Dehns lemma was used by Dehn [Dehn] to prove Theorem 32.1
and deduce from here the continuous rigidity of convex polyhedra (Corollary 26.7). This
approach was repeatedly rediscovered and is now only the rst basic step in the study of
rigidity of general frameworks. We refer to [Con5, Whi2] for the introduction and a broad
overview of general problems in rigidity theory and various applications.
The static (and thus continuous) rigidity of various families of non-convex polyhedra (see
Exercise 32.1) is an interesting part of modern rigidity theory. We refer to surveys [Ale6,
Con5] for various references on the subject.
The rst proof of Dehns lemma (Subsection 32.3) follows [FedP], but can also be found
in the early works on rigidity of frameworks (see [Roth, Whi1]). The proof of continuous
rigidity given in Subsection 32.4 is due to Alexandrov [A2, 10.3].
The rst part of the graph theoretic proof (Subsection 32.5, until Example 32.6) follows
the original paper of Dehn [Dehn] (see also [Ale6, 6]). The connectivity of all claw partitions
by a sequence of 3-cycle reversals is probably new. The argument here is a variation
on theme of several local move connectivity arguments (see Subsection 14.7) We refer
to [KorP] for further discussion on claw partitions of graphs and connections to tilings, and
to [Pak3] for a survey on local move connectivity arguments (cf. Section 23).
The extension of the Cauchy and Dehn theorems to polytopes with saddle vertices (as
in Exercise 32.1) in fact goes through with many proofs. This result was was proved by
Stoker [Sto] and then repeatedly rediscovered. Alexandrovs extension (Exercise 32.4) is
the rst step towards Connellys theorem on the second order rigidity of all triangulated
convex surfaces (Exercise 33.5).
Finally, let us note that for small enough angles the argument in the original (incorrect)
proof of the arm lemma (Lemma 22.3) is completely valid. Thus the incorrect proof in
fact implies the continuous rigidity of convex polytopes (Corollary 26.2). As we show in this
and the next section, the rigidity of convex polyhedra is somewhat easier to prove than the
uniqueness (the Cauchy theorem). Thus, it is quite tempting to ignore the full power of the
Cauchy theorem, thinking that the continuous rigidity is the most important implication of
uniqueness. From the real life applications point of view this idea is actually not without
merits (you really need rigidity, not uniqueness, to prevent a bridge from falling apart).
On the other hand, the proofs of a number of important results, such as the Alexandrov
79This

is the innitesimal version of Stokers conjecture for simplicial polytopes.

301

existence theorem (Theorem 37.1), substantially rely on uniqueness. As a consequence, the

full strength of the Cauchy theorem should be neither undervalued nor undermined.

302

33. Infinitesimal rigidity

In this short section we give another approach to the rigidity, and give our own
new proof of Dehns theorem in the previous section.
33.1. You cannot change anything if you do not know how to start. Consider
a continuous deformation {Pt : t [0, 1]} of the simplicial polytope P = P0 R3 . As

before, let V and E be the set of vertices and edges of P . Denote by v i (t) = Ovi the
vector from the origin to the vertex vi = vi (t) of Pt . Think of vectors v i (t) as velocities
of vertices vi .80 For an edge length |vi vj | to be constant under the deformation we
need kv i (t) v j (t)k = 0, where kw k = (w, w) = |w|2. Thus, in particular, at t = 0
we must have:





d
d
v i (t) v j (t)
v i (0) v j (0) + t v i (0) v j (0)
0 =
=
t=0
t=0
dt
dt

= 2 v i (0) v j (0), v i (0) v j (0) .

This leads to the following natural definition.

Let P R3 be a planted simplicial polytope. Suppose we are given a vector a i ,
for every vertex vi V . We say that the set of vectors {a i } defines an infinitesimal
rigid motion if
()

(v i v j , a i a j ) = 0, for every (vi , vj ) E.

An infinitesimal rigid motion is called planted if the velocities of base vertices are
equal to zero: a 1 = a 2 = a 3 = 0. Finally, we say that a simplicial polytope P R3
is infinitesimally rigid if every planted infinitesimal rigid motion is trivial: a i = 0 for
all vi V .
Theorem 33.1 (Dehns theorem; infinitesimal rigidity of convex polytopes). Every
simplicial convex polytope in R3 is infinitesimally rigid.
Of course, the restriction to planted infinitesimal rigid motions is necessary, as the
usual rigid motions of P in R3 can define nontrivial infinitesimal rigid motions. Also,
by definition, the infinitesimal rigidity implies the continuous rigidity.
To prove the theorem, simply observe that infinitesimal rigidity is in fact equivalent
to static rigidity, but written from a dual point of view. Indeed, if the determinant D
of the rigidity matrix is nonzero that means that locally all edge lengths of planted
realizations must be different. Therefore, the space of planted infinitesimal rigid
motions is zero, and vice versa. The details are straightforward.
To summarize, both static and infinitesimal rigidity are equivalent to Dehns lemma
(Lemma 32.2). In fact, this equivalence extends to general bar frameworks (see [Con5,
Whi2]) with no difference in the proof.
80As

in Section 28, when the derivatives do not exist we can consider both left and right
derivatives.

303

33.2. A synthetic proof of infinitesimal rigidity. Now, one may ask what is
so great about yet another way to rephrase Dehns lemma. Well, as it turns out,
the infinitesimal rigidity notion is nice enough to allow a graph theoretic approach
markedly different from the one used above.
The proof we present below yet again splits into two parts: global and local. The
local part, while different, has roughly the same level of difficulty as the static analogue
of the sign changes lemma (Lemma 32.3). The global part is a simple graph theoretic
argument similar in style to the proof of the sign counting lemma (Lemma 26.5).
Proof of Theorem 33.1. First, note that equations () above say that the difference
in the velocity of vertices is orthogonal to the edges of the polytope. Think of velocity vectors as vector functions on vertices of P which are equal to 0 on base
vertices v1 , v2 , v3 . The idea of the proof is to enlarge the set of such functions and
prove a stronger result.
Let V = {v1 , . . . , vn } be the set of vertices of a simplicial convex polytope P R3 ,
and let E be the set of edges. Consider the set of all vector sequences (a 1 , . . . , a n ),
a i R3 , such that for every edge (vi , vj ) E we have one of the following three
possibilities:
1.
2.
3.

(v i v j , a i ) = (v i v j , a j ) = 0,
(v i v j , a i ) < 0 and (v i v j , a j ) < 0,
(v i v j , a i ) > 0 and (v i v j , a j ) > 0.

In other words, we require that projections of velocity vectors a i and a j onto edge (v i , v j )
have the same signs. We say that a vertex vi is dead if a i = 0; it is live otherwise. We
need to prove that for every vector sequence (a 1 , . . . , a n ) as above, if base vertices
are dead, then all vertices vi V are dead. By definition of infinitesimal rigidity, this
would immediately imply the theorem.
Denote by = (V, E) the graph of P . Since P is simplicial, is a plane triangulation. Consider an orientation edges of in the direction of projections of the
velocity vectors. More precisely, we orient the edge vi vj in case 2, we orient the
edge vi vj in case 3, and leave it unoriented vi vj in case 1.
Consider two edges e = (vi , vj ) and e = (vi , vr ), e, e E, with a common vertex vi ,
such that (vi , vj , vr ) is a face in P . We say that edges e and e

have one inversion if one of them is oriented into vi , and the other out of vi ,
have zero inversions if both of them are oriented into vi or out of vi ,
have a half-inversion if one of the edges is oriented and the other is unoriented,
have one inversion if both of them are unoriented and vi is a live vertex,
have zero inversions if both of them are unoriented and vi is a dead vertex.

We say that a triangle is active if at least one of its vertices is live; it is inactive
otherwise. Now consider orientations of an active triangle (vi vj vr ) where vertex vi is
live (see some of them in Figure 33.1). A simple enumeration of all possible cases
gives the following result.
Lemma 33.2. Every active triangle has at least one inversion.

304

vr

vi

vr

vj

vi

vr

vj

vi

vr

vj

vi

vj

Figure 33.1. Different orientations of (vi vj vr ), where vertex vi is live.

This gives a lower bound on the number of inversions . To get an upper bound,
we use the following infinitesimal analogue of the sign changes lemma.
Lemma 33.3. There are at most two inversions around every live vertex.
We postpone the proof of the lemma until we finish the proof of the theorem.
Consider what this gives us when the only unoriented edges in are the base edges
(v1 , v2 ), (v1 , v3 ), and (v2 , v3 ). In this case we have n3 live vertices and (2n4)1 =
2n 5 triangles with at least one live vertex. By Lemma 33.3, there are at most
2(n 3) = 2n 6 inversions in , while by Lemma 33.2 there are at least 2n 5
inversions, a contradiction.
We use the same strategy in the general case. Remove from all inactive triangles
together with all edges and vertices which belong only to inactive triangles. Denote
by H = (V , E ) a connected component of the remaining graph. Since is planar, the
induced subgraph H of has a well defined boundary H. Denote by k the number
of vertices in H (all of them dead), and by the number of connected components
of H. Finally, denote by m the number of vertices in H r H (some of them live
and some possibly dead). By Lemma 33.3, there are at most 2m inversions in H.
Now let us estimate the number of inversions via the number t of triangles in H.
Observe that the total number of vertices, edges, and faces in H is given by
|V | = m + k,

2 |E | = k + 3t,

and |F | = + t .

Thus, by Eulers formula, graph H has exactly t = 2m + k + 2 4 triangles. Since

there is at least one inactive triangle (v1 v2 v3 ), we have k 3 and 1. Therefore,
by Lemma 33.2, there are at least
t = 2m + k + 2 4 2m + 3 + 2 4 = 2m + 1
inversions in H, a contradiction.

Proof of Lemma 33.3. Consider all possibilities one by one and check the claim in
each case. Suppose a vertex vi is adjacent to three or more unoriented edges. This
means that a i is orthogonal to at least three vectors spanning R3 . Therefore, a i = 0
and vi is a dead vertex with zero inversions.
Suppose now that vi is adjacent to exactly two unoriented edges e, e . This means
that a i 6= 0 is orthogonal to a plane spanned by these edges. Observe that e, e
separate the edges oriented into vi from those oriented out of vi . Thus, there are

305

Figure 33.2. The number of inversions around a vertex in different cases.

either two half-inversions and one inversion if the edges e, e are adjacent, or four
half-inversions if e, e are not adjacent (see Figure 33.2).
Next, suppose that vi is adjacent to exactly one unoriented edge e. Since a i 6= 0 in
this case, consider a plane containing vi and orthogonal to a i . This plane contains
and separates the edges oriented into vi from those oriented out of vi . Therefore,
there are either two half-inversions if all other edges are oriented into vi , or out of vi ,
or two half-inversions and one inversion otherwise.
Finally, if vi is a live vertex and not adjacent to unoriented edges, then the plane
orthogonal to a i separates the edges into two parts: those oriented into vi from those
oriented out of vi . Thus, there are exactly two inversions in this case.

33.3. Exercises.
Exercise 33.1. [1+] Prove the innitesimal analogue of Alexandrovs theorem (Exercise 32.4).
Exercise 33.2. [1] Prove that Jessens orthogonal icosahedron (see Exercise 19.17) is continuously rigid, but not innitesimally rigid.
Exercise 33.3. [1] Give a direct proof that the static and the innitesimal rigidity are
equivalent.
Exercise 33.4. [1+] Modify and prove the analogue of Dehns theorem for unbounded
polyhedra (cf. Subsection 27.3).
Exercise 33.5. (Connellys second order rigidity theorem) a) [1] In notation as in the
beginning of the section, take the second derivative of the edge length |vi vj |. Letting both
the rst and second derivatives to be zero, extend () to dene the second order rigidity in
combinatorial terms. Conclude that this is a necessary condition for the continuous rigidity.
b) [1-] Prove that the non-strictly convex polyhedron in Figure 32.1 is second order rigid.
c) [1] Give a polynomial time algorithm for testing the second order rigidity.
d) [2] Prove that every triangulated convex polyhedral surface in R3 is second order rigid.
Exercise 33.6. [2] Prove that the innitesimal rigidity is invariant under projective transformations.
Exercise 33.7. [1+] Let X = [x1 . . . xn ] R2 be a convex polygon in the plane. Suppose
the edges of X are bars and all diagonals (xi , xi+1 ) are cables, where the index i {1, . . . , n}
is taken modulo n.81 This framework is called the tensegrity polygon. Formally state and
prove that it is innitesimally rigid.
81Here

bars lengths must remain the same, while cables can be contracted.

306

Exercise 33.8. [1+] Let (x000 , x001 , . . . , x111 ) R3 be eight vertices corresponding to vertices of a cube. Suppose all cube edges, as well as the diagonals in the top and bottom face,
are connected by cables. Also, suppose the diagonals in the four side edges are connected
by bars. Prove that this framework is innitesimally rigid.
Exercise 33.9. Let = (x1 x2 x3 x4 ) be a tetrahedron in R3 , let ij = |xi xj |, and let ij be
the dihedral angle at edge (xi xj ).
a) [1] Prove that angles ij determine uniquely, up to expansion.
b) [1] Show that lengths 12 , 13 , 14 and angles 23 , 24 , 34 do not necessarily determine .
c) [1+] Show that cannot be continuously deformed so that the angles and lengths in b)
remain invariant.
Exercise 33.10. (Rivin, Luo) a) [2+] Let S be an abstract 2-dimensional triangulated
surface homeomorphic to a sphere, and dened by the edge lengths ij = |eij |. Denote by
ij = , where and are the angles opposite to edge eij . Prove that every
continuous deformation of S which preserves all ij is a homothety.
b) [2+] Same result for ij = ln(tan ) + ln(tan ).
c) [2+] Same result for ij = cot + cot .
Exercise 33.11. (Schramm) [2] Suppose two combinatorially equivalent simplicial convex
polyhedra P1 , P2 R3 have a midscribed sphere, i.e., all their edges touch a xed unit
sphere S2 . Suppose further that a face in P1 coincides with the corresponding face in P2 .
Prove that P1 = P2 .
33.4. Final remarks. The synthetic proof of innitesimal rigidity in this section follows [Pak5] and is based on the proof idea in [Tru]. We refer to [Con5, 4.6] for further
references and other proofs of Dehns theorem.
Polyhedra that are continuously but not innitesimally rigid are called shaky polyhedra [Gol4]. Jessens orthogonal icosahedron is a classical example of a shaky polyhedron
(see Exercises 19.17 and 33.2). A non-strictly convex polyhedron in Figure 32.1 is another,
degenerate example.
While continuous rigidity is often hard to establish even for very specic frameworks,
the innitesimal rigidity is equivalent to a nonzero determinant of the rigidity matrix R .
Similarly, one can dene the second and higher order rigidity, which give further necessary
conditions on continuous rigidity (see Exercise 33.5). We refer to [ConS] for the introduction
to higher order rigidity, and to survey papers [Con5, IKS] for further references.

307

34. Proof of the bellows conjecture

We continue the algebraic approach started in Section 31. Here we give a complete
proof of the bellows conjecture stated there, modulo some basic algebraic results given
in the Appendix (see Subsection 41.7).
34.1. Polynomiality and integrality of the volume. Let S Rd be a polyhedral
surface homeomorphic to a sphere. When S = P is a surface of a convex polytope,
we can define the volume enclosed by S as the volume of the polytope P . In general,
we can triangulate the surface and define
vol(S) :=

1
n!

det[v1 , . . . , vd ] ,

(v1 ...vd )S

where (v1 . . . vd ) is an oriented simplex in S and det[v1 , . . . , vd ] is a determinant of a

matrix where vi are column-vectors. This implies that the volume of convex polytopes
is polynomial in the vertex coordinates.
Now let S = P be the surface of a simplicial convex polytope P R3 . Its volume is polynomial in vertex coordinates, and by Corollary 31.4 this implies that as
a polynomial this volume is a root of an equation with coefficients in C[Fij ], where
Fij are polynomials giving the squared distances between vi and vj . In other words,
there exists a Sabitov polynomial relation () for the volume polynomial (see Corollary 31.4).
More generally, consider a simplicial surface S R3 homeomorphic to a sphere,
i.e., a polyhedral surface with only triangular faces. From above, the volume is a
root of a polynomial obtained by an evaluation of the Sabitov polynomial relation ()
at the vertex coordinates. Now, as we showed in Section 31, the bellows conjecture
(Theorem 31.2) follows immediately if we show that the leading coefficient is 1, or
any other nonzero integer, since this implies that the above polynomial does not
degenerate. Let us formalize this property as follows.
Let L be a field which contains a ring R. For our purposes one can always assume
that L is either C or R. We say that an element x L is integral over R if
xr + cr1 xr1 + . . . + c1 x + c0 = 0.
Now comes the crucial result.
Lemma 34.1 (Integrality of the volume). Let S R3 be a simplicial surface homeomorphic to a sphere and let R be a ring generated by the squares of the edge lengths 2ij .
Then 24 vol(S) is integral over R.
From above, the lemma implies the bellows conjecture (Theorem 31.2). Its proof is
based on a mixture of algebraic and geometric properties of volume and will occupy
the rest of this section.

308

34.2. Playing with infinities. To prove the integrality of the volume we need to
develop some algebraic tools. Below is a quick introduction to the theory of places,
which will help resolve the problem.
Let F be a field, and let Fb = F {}. Elements a F are called finite elements
in Fb. As before, let L be a field. A map : L Fb is called a place if (1) = 1 and
it satisfies the following conditions:
(>)

(a b) = (a) (b),

Here it is understood that for all a F

(a b) = (a) (b), for all a, b L.

a
1
= ,
= 0,
0

and b = for all b 6= 0. Furthermore, the expressions

0
,
, 0 , and
0
are not defined, and the conditions on do not have to hold in the case when the
r.h.s. of (>) is undefined. Note that (0) = (1 1) = (1) (1) = 1 1 = 0.
For example, take the field of rational Laurent polynomials L = Q(x) and define
b by setting (x) = , ( 1 ) = 0, and (a) = a, for all
a place : Q(x) Q
x
a Q. Viewing the elements f (x) Q(x) as functions in x defined on R+ , have then:
(f ) = limx f (x).
A place : L Fb is called finite on L (on R) if (a) F for all a L (for all
a R). The following standard algebraic result is a key tool in our approach.
a = =

Theorem 34.2 (Integrality criterion). Let L be a field containing a ring R. An

element x L is integral over R if and only if every place that is finite on R, is also
finite on x.

Obviously, all elements x R are integral over R. Note that by the theorem, if x
and y are integral over R, then so are x, x + y and x y. On the other hand, even
if x L, x 6= 0, is integral, then
1/x Lis not necessarily integral. For example, if
L = C and R = Z, then 2, 2 and i = 1 are integral over Z, while 1/2 is not.
34.3. Proof of the bellows conjecture. Just like in several proofs of the Cauchy
and Dehns theorems, the proof of the bellows conjecture (Theorem 31.2) neatly splits
into the local and the global part. As before, we first present the global part of the
proof, and only then prove the local part.
Proof of Lemma 34.1. Let L = R be the field of vertex coordinates. As in the lemma,
denote by R the ring generated by the squared edge lengths. Let F be any fixed field.
Consider a place : L Fb which is finite on R.
Now, let = (V, E) be a graph of the vertices and edges in S. We prove the claim
by induction on the number n = |V | of vertices and the smallest degree of a vertex
in . Let us start with the base of induction, which is summarized in the following
lemma, proved later in this section.

309

Lemma 34.3. Let R3 be a simplex and let R be a ring generated by the squares
of the edge lengths 2ij . Then 24 vol() is integral over R.
Suppose has vertex w of degree 3. In this case we show that one can make the
inductive step by reducing the number of vertices. Denote by v1 , v2 , v3 the neighbors
of w. Consider a new surface S obtained from S by removing triangles (wv1 v2 ),
(wv2 v3 ), (wv3 v1 ), and adding triangle (v1 v2 v3 ) (see Figure 34.1). Clearly, S has
|V | 1 vertices and
vol(S) = vol(S ) + vol(),
where is a simplex (wv1 v2 v3 ). By the lemma and by the inductive assumption,
24 vol() and 24 vol(S ) are both integral over R. Thus, 24 vol(S) is also integral
over R, and we proved the claim in this case.
w
v3

v1

v3

v1

v2
S

v2
S

Figure 34.1. The inductive step S S when the degree of w is 3.

Suppose now that has no vertices of degree 3. The inductive step in this case
consists of reducing the minimal degree of . Since is a plane triangulation, there
exists a vertex w of degree at most 5 (otherwise there are 6n/2 = 3n edges in ).
We will show that is also finite on certain diagonals in S. Then, by the argument
similar to the one above, we conclude that is finite on the volume.
Formally, we say that g = (v, v ) is a diagonal if v, v V and g
/ E. Let d = |g|
be the length of the diagonal g in S. The diagonal g is called supportive if (d2 ) is
finite.
Let w be a vertex in S of the smallest degree k, 4 k 5, and let v1 , . . . , vk
be its neighbors, in cyclic order. Denote by Hw the graph induced by the vertices
w, v1 , . . . , vk . Define small diagonals in Hw to be pairs of vertices g1 = (v1 , v3 ), . . . ,
gk2 = (vk2 , vk ), gk1 = (vk1 , v1 ), and gk = (vk , v2 ). The following lemma shows
that at least one of the small diagonals in is supportive.
Lemma 34.4 (Supportive small diagonals). Let w be a vertex in S of degree 4 or 5,
and let : L Fb be a place which is finite on squared edge lengths of Hw . Then one
of the small diagonals gi in Hw is supportive.
In other words, the lemma says that at least one of the place values (|g1 |2 ), . . . ,
(|gk |2 ) is finite. The lemma represents the local part of the proof of the bellows
conjecture and will be proved later in this section.

310

Suppose (|gi |2 ) is finite, for the small diagonal gi = (vi , vi+2 ) as above. Consider
a new surface S obtained from S by removing triangles (wvivi+1 ), (wvi+1 vi+2 ), and
adding triangles (wvi vi+2 ) and (vi vi+1 vi+2 ) (see Figure 34.2). Clearly, the degree of w
in S is k 1, and
vol(S) = vol(S ) vol(),
where is a simplex (wvivi+1 vi+2 ) with finite squared edge lengths. Here the orientation of the surface of can be either positive or negative, which determines the sign.
By the same argument as above, Lemma 34.3 and the inductive assumption imply
vi

vi
vi+1

vi+1

w
S

w
vi+2

vi+2

Figure 34.2. The inductive step S S when the degree of w is 5.

that 24 vol() and 24 vol(S ) are both integral over R. Thus, 24 vol(S) is also integral
over R, and we proved inductive claim. This completes the proof of Lemma 34.1. 
34.4. Computing the volume of a simplex. To prove Lemma 34.3, we present
the following classical result which implies the lemma.
Denote by = (v0 v1 . . . vd ) the simplex in Rd with vertices vi and edge lengths ij .
Define the CayleyMenger determinant CM() as follows:

0 1
1
1 ... 1
1 0 201 202 . . . 20d

1 201 0 212 . . . 21d

CM(v0 , v1 , . . . , vd ) := det
1 202 212 0 . . . 22d .
. .
..
.. . .
.
.. ..
. ..
.
.
1 20d 21d 22d . . .

The following formula computes the volume of via the CayleyMenger determinant.

Theorem 34.5 (CayleyMenger). For every simplex = (v0 v1 . . . vd ) Rd , we

have:
(1)d1
vol2 () =
CM(v0 , v1 , . . . , vd ).
2d d!2
The proof of the CayleyMenger theorem is a straightforward calculation with
determinants (see Appendix 41.6). We can now prove Lemma 34.3. When d = 3, we
have from the theorem:

2
2 CM(v0 , v1 , v2 , v3 ) = 2 288 vol2 () = 24 vol() .

311

Since the l.h.s. in this equation is a polynomial in the squared edge lengths, we
conclude that 24 vol() is integral over R, as desired.
Example 34.6. (Herons formula) Let P be a triangle in R2 with side lengths a, b, and c.
Recall the classical Herons formula for the area of a triangle:
p
1
area(P ) = p (p a)(p b)(p c) , where p = (a + b + c).
2
Expanding the product we obtain the same result as given by the CayleyMenger determinant:

0 1 1 1
1 0 a2 b2


1
1

area2 (P ) =
2a2 b2 + 2a2 c2 + 2b2 c2 a4 b4 c4 =
det
1 a2 0 c2 .
16
16
1 b2 c2 0

This implies that 4 area(P ) is integral over the ring generated by a2 , b2 and c2 . Note also
that the multiple 4 cannot be lowered here.

34.5. Proof of Lemma 34.4. We start with the following technical result. Let
w, v1 , v2 , v3 , v4 be distinct points in R3 . Consider the squared distances a1 = |v1 v2 |2 ,
a2 = |v2 v3 |2 , a3 = |v3 v4 |2 , and bi = |wvi|2 , 1 i 4. Further, let c = |v1 v4 |2 ,
s1 = |v1 v3 |2 , and s2 = |v2 v4 |2 .
Lemma 34.7. Let : L Fb be a place such that (ai ) and (bj ) are finite, where
1 i 3, and 1 j 4. Suppose also that (s1 ) = (s2 ) = . Then (c/s1 ) = .

This lemma easily implies Lemma 34.4. For k = 4, suppose that (s1 ) = (s2 ) =
(see Figure 34.3). Then the conditions of Lemma 34.7 are satisfied. In addition,
we require that (c) is finite, which implies that (c/s1 ) = 0. This contradicts
(c/s1 ) = in the lemma. Therefore, either (s1 ) or (s2 ) is finite. In other words,
at least one of the two small diagonals must be supportive.
a2

v2

v3

v3

a3

a1
s1
v1

v2

s2
c

s1
v4

v1

v4

v5
Figure 34.3. Two cases in Lemma 34.4: when k = 4 and when k = 5.
For n = 5, suppose neither of the small diagonals in a pentagon [v1 v2 v3 v4 v5 ] is
supportive. From the lemma, (c/s1 ) = . Similarly, by the symmetry, (s1 /c) =
(see Figure 34.3). But then
(1) = (c/s1 ) (s1 /c) = = ,

312

a contradiction. This implies that at least one of the small diagonals is supportive,
and completes the proof of Lemma 34.4 from Lemma 34.7.
Proof of Lemma 34.7. Think of five points v1 , . . . , v4 and w as being in R4 . Then the
CayleyMenger determinant CM(v1 , v2 , v3 , v4 , w) = 0. In the notation of the lemma,
we have:

0 1 1 1 1 1
1 0 a1 s1 c b1

1 a1 0 a2 s2 b2
det
= 0.
1 s1 a2 0 a3 b3
1 c s a 0 b
2
3
4
1 b1 b2 b3 b4 0
Divide the second row and second column by s1 . Similarly, divide the fifth row and
fifth column by s2 . We obtain:

1 1 s12
1
0 s11
b1
c
1
0 a1 1

s1
1 a1 s01 a s11s2 bs1
2
2

s1
D := det
= 0.
a3
1
1
a
0
b

2
3
s2

1
0 sb42
s2 s1cs2 1 as23
1 sb11 b2 b3 sb42
0


Suppose that s1cs2 = 0. Using the assumptions on the values of and conditions (>) in the definition of a place, we obtain:

0 0
1
1
0
1
0 0 0 0 0 1
0
1
0
0
0 0
0 0 0 1 0 0

0
(a2 ) 1 (b2 )
1 0
0 0 0 0 1 0
(D) = det
= det
= 1,
0
0 (b3 )
1 1 (a2 )
0 1 0 0 0 0
0 0
0 0 1 0 0 0
1
0
0
0
1 0 (b2 ) (b3 ) 0
0
1 0 0 0 0 0


which contradicts (D) = (0) = 0. Therefore, s1cs2 6= 0, and
 


c
c

=
(s2 ) = ,
s1
s1 s2
as desired.

34.6. Exercises.
Exercise 34.1. (Brahmaguptas formula) a) [1-] Prove an explicit formula for the
area of a convex quadrilateral Q with side lengths a, b, c, d, and which is inscribed
into a circle:
area2 (Q) = ( a)( b)( c)( d),
where = 21 (a + b + c + d) is half the perimeter of Q.
b) [1] Conclude that 4 area(Q) is integral over C[a2 , b2 , c2 , d2].

313

Exercise 34.2. a) [1-] Show that a triangle in the plane cannot have odd sides and
integral volume.
b) [1] Show that a tetrahedron in R3 cannot have odd edge lengths and integral
volume.
c) [1-] Conclude from a) that no 4 points in the plane can have odd pairwise distances.
Exercise 34.3. Two sets A and B of n points are called equivalent if their have the
same multiset of pairwise distances. Similarly, they are called congruent if one can
be obtained from the other by a rigid motion.
a) [1] Find two equivalent but non-congruent sets of 4 points in the plane.
b) [1+] Show that almost all sets of n 4 points in the plane have no sets equivalent
but not congruent to them.
Exercise 34.4. [1] Extend Lemma 34.4 to vertices of degree > 5.
Exercise 34.5. [1+] Let v0 , . . . , vd Rd1 be any d + 1 points. Denote by
CM (v0 , v1 , . . . , vd ) = det(2ij )
the principal minor of the CayleyMenger matrix, obtained by removing the first row
and the first column. Prove that CM (v0 , v1 , . . . , vd ) = 0 if and only if the points lie
on a sphere or a hyperplane.
Exercise 34.6. (Robbins problem) Denote by A(a1 , . . . , an ) the area of a convex
polygon inscribed into a circle with sides ai (cf. Example 34.6, Exercises 13.1, 31.3,
and 34.1).
a) [1-] Prove that A() is a symmetric function.
b) [2-] Prove that 4A() is integral over C[a21 , . . . , a2n ].
Exercise 34.7. (Extended bellows conjecture) a) [1+] Extend the bellows conjecture
to all orientable 2-dimensional polyhedral surfaces in R3 . In other words, extend
Lemma 34.1 to surfaces of any genus.
b) [] Extend the bellows conjecture to higher dimensions.
Exercise 34.8. (Equiareal triangulations) A dissection of a polygon into triangles of
equal area is called an equiareal triangulation.
a) [2] Prove that every equiareal triangulation of a unit square has an odd number
of triangles.
b) [2] Suppose a unit square is dissected into triangles with areas a1 , . . . , am . Prove
that there exists a polynomial f Z[x1 , . . . , xm ] such that f (a1 , . . . , am ) = 1/2.
c) [2+] Prove that every equiareal triangulation of a centrally symmetric polygon has
an odd number of triangles.
d) [2+] Extend part c) to centrally symmetric polygons of area 1.
e) [2] Find a plane quadrilateral which has no equiareal triangulations.
f ) [2] Find a convex polytope R R3 which cannot be dissected into tetrahedra of
equal volume.

314

34.7. Final remarks. The bellows conjecture is usually attributed to Connelly and
Sullivan (see [Con2]). The idea of finding and studying polynomial relations for the
volume (as in Corollary 31.4) seems to be due to Connelly and Sabitov and goes back
to the late 80s (see [Sab1]). The bellows conjecture, including the case of surfaces
of higher genus, was completely resolved by Sabitov in a series of papers of varying
difficulty [Sab2, Sab3, Sab4]. Our presentation is based on [CSW], where the idea
of using the theory of places has been introduced. This approach was later used
by Connelly to give a simple proof of one of the Robbins conjectures [Con6] (see
Exercise 34.6 and [Pak4]). We refer to [Schl3] for for more on the background of the
conjecture and the proof.
The bellows conjecture remains open in higher dimensions (see Exercise 34.7).
It is known to hold for the few available examples of flexible polyhedra (see Subsection 30.8). Interestingly, the bellows conjecture fails for spherical polyhedra, as
exhibited by a flexible spherical polyhedron given in Subsection 30.7 (see [Ale3]). A
number of counterexamples to the infinitesimal versions of the conjecture was given
in [Ale1]. Finally, the conjecture remains open for hyperbolic polyhedra (see [Ale6]
for the references).
The theory of places is often presented as part of a more general theory of valuations. In this context the integrality criterion (Theorem 34.2) is called the Chevalley
extension lemma (see [Cohn, 9.1]) and is usually stated in terms of valuation rings
(see [AtiM, Corollary 5.22] and [Mats, Theorem 10.4]).

315

35. The Alexandrov curvature theorem

This is the first of three sections giving a description of convex polyhedra in terms
of certain metric invariants. In this section we consider polytopes whose vertices
lie on given rays starting at the origin. The main result is Alexandrovs theorem
giving their complete characterization. The tools include the GaussBonnet theorem
and its variations (see Section 25) and the mapping lemma in the Appendix (see
Subsection 41.8). The proof ideas from this section are also used in the next two
sections.
35.1. First touch rule. Let u 1 , . . . , u n S2 be a fixed set of unit vectors. Denote by
R1 , . . . , Rn R3 the rays starting at the origin O R3 and spanned by these vectors.
Consider all convex polytopes P with vertices on the rays (one on each). In other

words, let P = conv{v1 , . . . , vn } where Ovi = ri u i , ri > 0, and i [n] = {1, . . . , n}.
We say that polytope P lies on rays Ri , and call ri the ray coordinates of P .
We assume that rays Ri do not lie in the same half-space, so that polytope P
contains the origin O in its relative interior. As before, denote by i = (Ci) the
curvature of vertex vi (see Section 25).
Define the expansion of polytope P with ray coordinates r1 , . . . , rn to be a polytope cP with ray coordinates cr1 , . . . , crn , for some c > 0. The following results
shows that the curvature uniquely determines the polytope up to expansion:
Lemma 35.1 (Curvature uniqueness). Let P, P R3 be polytopes which lie on
rays Ri containing the origin O, where i [n]. Suppose these polytopes have equal
curvature at all vertices: i = i , for all i [n]. Then P is an expansion of P .
Proof. Consider all expansions cP which lie inside P . Clearly, for sufficiently small
c > 0 such expansions exist, and for sufficiently large c they do not. Let Q = P be
the largest such expansion, Q P . Since no further expansion is possible, at least
one vertex of Q lies on the boundary P , and by construction of polytopes on rays
this boundary point is a vertex of P (see Figure 35.1).

Q
P

Figure 35.1. Polygons P, P and the maximal expansion Q = cP P .

Consider all edges of Q which coincide with edges of P . There are two possibilities:
either all of them do, or (by connectivity of the graph of Q there exists an edge

316

e = (vi , vj ) in Q, with one vertex vi in P and the other vertex vj in the interior
of P . In the first case, all vertices of Q coincide with those in P and polytope P
is an expansion of P . In the second case, denote by Ci the cone spanned by edges
of P starting at vi , and by Di the cone spanned by edges of Q starting at vi . Then,
by construction, Ci Di and Ci 6= Di . Therefore, by the monotonicity of the cone
curvature (Exercise 25.5), we have i = (Ci) < (Di ) = i , a contradiction.

35.2. Inequalities on curvatures. Before we can state the converse of Lemma 35.1,
we need to give the necessary conditions on the curvatures.
As before, let P be a convex polytope which lies on rays Ri , which do not lie in
the same half-space, and let i > 0 be the curvatures of the vertices, i [n]. First,
recall the GaussBonnet theorem (Theorem 25.3):
1 + . . . + n = 4 .
In addition, there is a number of inequalities given by the geometry of rays.
Lemma 35.2. For a subset I [n], denote by CIP
= conv{Ri : i I} the convex
cone spanned by the rays. Assume CI 6= R3 . Then
k I
/ k > (CI ).
Proof. Let P be a convex polyhedron obtained as a convex hull of P and rays Ri ,
for all i [n]. Since CI 6= R3 , polyhedron P is well defined. Note that the vertices
of P are given by vk
/ CI , so in particular, k
/ I. Since O is a point in the relative
interior of P , the cones of P and P at all such vertices vk satisfy Ck Ck . By
the monotonicity of the cone curvature (Exercise 25.5), we have k k , and the
inequality is strict for vertices of unbounded faces of Ck . Applying Theorem 25.5 to
polyhedron P , we obtain:
X
X
X
k >
k
k = (CI ),
k I
/

as desired.

k I
/

k[n] : vk C
/ I

35.3. Inequalities on ray coordinates. Let us now describe the set of ray coordinates of all convex polytopes which lie on given rays R1 , . . . , Rn . As it turns out, it is
convenient to use the inverse ray coordinates. Formally, let xi = r1i and let X denote


the set of all (x1 , . . . , xn ) Rn , such that xi > 0 and x11 , . . . , x1n are ray coordinates
of a polytope on rays Ri .
Lemma 35.3. Suppose R1 , . . . , Rn R3 are rays which do not lie in the same halfspace. Then set X of inverse ray coordinates is an open convex cone in Rn .
Proof. Since all polytopes P which lie on rays Ri must contain the origin, their combinatorics is completely determined as follows. For every three independent rays Ra , Rb
and Rc , consider a cone C = C{a,b,c} spanned by these rays. Whenever ray Ri lies
inside C, the vertex vi of Ri must be separated from O by a plane H spanned by va , vb
and vc .

317

As before, let u a , u b , u c and u i be the unit vectors along the rays. Suppose
u i = u a + u b + u c .
We immediately have








ra u a + ri
rb u b + ri
rc u c .
ri u i = ri (u a + u b + u c ) = ri
ra
rb
rc

Observe that the plane H consists of points a va + b vb + c vc , for all a + b + c = 1,

and let s > 0 be defined by



s
+
+
= 1.
ra
rb
rc

Therefore, vertex vi lies on H whenever ri = s and vi is separated from O by H if

ri > s. Rewriting the last inequality in the inverse ray coordinates xi = r1i we obtain
xi < xa + xb + xc . Taking the intersection of these halfspaces for all cones C as
above, with the halfspaces xi > 0, we obtain convex cone X as in the lemma.

35.4. Finding polytopes with given curvatures. We are ready now to prove the
main result of this section. For a subset I [n], denote by CI = conv{Ri : i I}
the convex cone spanned by the rays.
Theorem 35.4 (Alexandrov curvature theorem). Suppose R1 , . . . , Rn R3 are rays
which do not lie in the same half-space, and suppose 1 , . . . , n R satisfy the following conditions:
1) i > 0, for all i [n],
2)
P1 + . . . + n = 4,
3
3)
k I
/ k > (CI ), for every I [n], such that CI 6= R .
Then, up to expansion, there exists a unique convex polytope P R3 which lies on
rays Ri and has curvatures i at vertices vi Ri , for all i [n]. Conversely, all
curvatures i of such polytopes P must satisfy conditions 1) 3).
One can think of this theorem as both the extension and the converse of Lemma 35.1.
Of course, the second part follows from Lemma 35.2 and the Gauss-Bonnet theorem.
Remark 35.5. Let us rst check that this result is plausible. There are n 1 degrees

of freedom of the ray coordinates modulo expansion. This matches the n 1 degrees of
freedom of the vertex curvatures modulo the GaussBonnet formula. Therefore, by the
curvature uniqueness theorem and the inverse function theorem, for small perturbations
of i , there is always a small perturbation of the ray coordinates ri giving these curvatures.
The following proof uses a variation on the same argument.

Proof of Theorem 35.4. Let X Rn be the cone of inverse ray coordinates xi , and let
X be the intersection of X with a hyperplane defined by x1 + . . . + xn = 1. Clearly,
X is an open (n 1)-dimensional convex polytope. Also, for every polytope P with
inverse ray coordinates x X , there is a unique c > 0 and an expansion cP with
inverse ray coordinates cx X .

318

Denote by Rn the set of all w = (1 , . . . , n ) as in the theorem. Denote by

: X the map defined by (x ) = w , where x = (x1 , . . . , xn ). By Lemma 35.2,
the curvatures satisfy three conditions in the theorem, so this map is well defined.
We now deduce the result from the mapping lemma (Theorem 41.8). First, both X
and are open connected manifolds of the same dimension n 1. By Lemma 35.1,
map is injective. Obviously, map is continuous. It remains to show that is
proper.
Suppose w k w as k , where w k , w . Assume w k = (x k ) for some
x k X . Since X is a bounded convex polytope, the sequence {x k } has a converging
subsequence to some point x Rn . We need to show that x X . Denote by Pk the
polytopes corresponding to x k . By contradiction, if x
/ X , then it lies on one of the
facets, i.e., either xi = xa + xb + xc for some i, a, b, c [n], or xi = 0 for some
i [n], in notation of the proof of Lemma 35.3. In the first case, the limit of Pi is a
non-strictly convex polytope P where vertex vi lies in the plane spanned by (va vb vc ).
But then the limit of the curvatures i in w k is 0, a contradiction with w .
In the second case, if xi = r1i 0, we have ri , and the limit polyhedron
Pk P is unbounded. Let us first show that P is well defined. Denote by I the set
of all i [n] such that ri . By convexity, this implied that the same holds for
all i [n] such that Ri CI . Thus, if CI = R3 we have xi 0 for all i [n], a
contradiction with x1 + . . . + xn = 1. Therefore, polyhedron P is well defined and
contains only those rays Ri , s.t. i I. In other words, the
P base cone CP coincides
with the cone CI . Applying Theorem 25.5 to P , we have j I
/ j = (CP ) = (CI ),
a contradiction with w . This finishes the proof that is proper. By the mapping
lemma, is a homeomorphism, which proves the result.

35.5. Convex cap curvatures. Let H R3 be a plane, which we always think as
horizontal. Fix n points a1 , . . . , an H in general position, i.e., such that no three
of them lie on a line. Consider lines Li H, such that ai Li , for all i [n]. For
convenience, assume that the origin O is in the interior of conv{a1 , . . . , an }. Consider
a convex cap P R3 with vertices vi Li and the base cone CP a ray orthogonal
to H and pointing down.
One can ask if there is a way to characterize the curvatures i of vertices of convex
caps. Of course, shifting the convex cap P up by a fixed constant does not change the
curvatures. Also, by the GaussBonnet theorem for convex caps (see Section 25.5),
we have 1 + . . . + n = 2. As the following theorem shows, there are no nontrivial
inequalities on the curvatures.
Theorem 35.6 (Alexandrov curvature theorem for convex caps). Fix a plane H
in R3 and lines L1 , . . . , Ln R3 orthogonal to H. Suppose 1 , . . . , n > 0 satisfy
1 + . . . + n = 2. Then, up to translation, there exists a unique convex cap P which
lies on lines Li and has curvatures i at vertices vi Ri , for all i [n].
There is a topological proof (using the mapping lemma) of this theorem which we
leave to the reader (Exercise 35.2). Below we present a variational principle argument
proving a stronger result (see the reduction in Exercise 35.5).

319

Suppose a1 , . . . , ak are the vertices of a polygon A = conv{a1 , . . . , an }, so the

remaining points ak+1 , . . . , an lie in the relative interior of A. Consider a convex
cap P with vertices vi Li , for all i [n]. Space polygon Z = [v1 . . . vk ] is called the
border of P (see Figure 35.2).

Li

vi
hi
H

ai

Figure 35.2. Space polygon Z and a convex cap P on lines Li with border Z.
Theorem 35.7 (Pogorelov curvature theorem for convex caps). Fix a plane H in
R3 , lines L1 , . . . , Ln R3 orthogonal to H, and space polygon Z = [v1 . . . vk ] as
above. Suppose k+1 , . . . , n > 0 satisfy k+1 + . . . + n < 2. Then there exists a
unique convex cap P which lies on lines Li , has the border Z, and has curvatures i
at vertices vi Li , for all k + 1 i n.

Proof. We start with the uniqueness part. Suppose we have two convex caps P and P
as in the theorem. Shift P down until it lies below P and start moving it up until
the first time some two vertices vi and vi coincide. As in the proof of Lemma 35.1,
for the cones Ci and Ci at these vertices we have Ci Ci and (Ci) = (Ci ). By the
monotonicity of the cone curvature, we have Ci = Ci . Moving along the edges of the
graph G of P , starting at vi and that are not in Z, we conclude that vj = vj for all
vertices in G.
For a convex cap P which lies of lines Li , denote by hi the height of points vi ,
i [n]. Let P be the set of convex caps P which lie on lines Li , have border Z and
have curvatures i = (Vi) i for all k + 1 i n. Of course, for P P, the
heights h1 , . . . , hk are fixed, while hk+1 , . . . , hn can vary. Let (P ) = hk+1 + . . . + hn
be the sum of the heights.
We will show that the cap P P with maximal value of is the desired convex
cap. First, let us show that the set P is bounded, i.e., that hi c for some constant c
which depends on ai , i and Z. Let w = k+1 + . . . + n < 2. Then there exists a
vertex vi , i [k], with curvature i > (2 w)/k > 0. Therefore, the maximal slope
of the ray in the convex cone Ci starting at vi is at most ( i )/2 < /2. Since
every vertex vj lies inside Ci , the height hj is bounded from above (see Figure 35.3).
In a different direction, by convexity, the heights hj are bounded from below. This
implies that P is a compact set and the sum of heights maximizes on P.
Suppose now that function maximizes at some P P, and that for the curvature i of vertex vi we have i < i. Increase the height of vi by > 0 to obtain

320

vj
vi

hj

Figure 35.3. The heights hj of convex caps P P are bounded.

a convex cap P . By convexity, cone Ci can be shifted down by to fit inside Ci .
Thus, i = (Ci ) > (Ci ) = i . If is small enough, the curvature i < i . For the
remaining vertices, we have Cj Cj and thus j j i . This proves that P P.
By construction, (P ) = (P ) + , a contradiction. Therefore, maximizes on the
convex cap as in the theorem.

35.6. Exercises.
Exercise 35.1. Let P, P R3 be polytopes which lie on rays R, and have equal solid
angles: (Ci ) = (Ci ), for all i [n].
a) [1-] Prove that P is an expansion of P .
b) [1] Think of part a) as the uniqueness result (Lemma 35.1). Explain why there is no
obvious counterpart of the existence theorem in this case.
c) [1+] Prove the analogue of Theorem 35.7 for solid angles.
Exercise 35.2. [1] Give a topological proof of Theorem 35.7.
Exercise 35.3. [1+] Generalize Theorem 35.4 to higher dimensions.
Exercise 35.4. [] Give a variational principle proof of Theorem 35.4.
Exercise 35.5. [1] Show that Theorem 35.7 implies Theorem 35.6.
Exercise 35.6. [1-] Suppose n particles in space move at the same speed in dierent
directions. Prove that all particles will eventually move into a convex position.
Exercise 35.7. [2] In notation of Lemma 35.1, suppose polytopes P and P have curvatures
which satisfy |i i | < . Prove that there exist a constant c > 0, such that |ri cri | <
O( ) for all i [n], where c > 0 is a universal constant independent of the number n of
rays.
Exercise 35.8. a) [1] Following the approach in Subsection 9.1, convert the variational
proof of Theorem 35.7 into a greedy algorithm. Explain where the combinatorics of faces
of the convex caps appears in the proof.
b) [] For a greedy algorithm as in part a), consider triangulations of A = conv{a1 , . . . , an },
corresponding to projections of convex caps P . Either construct an example which requires
going through exponentially many such triangulations, or prove a polynomial upper bound.
Check that the same triangulation can be repeated. Give the exponential upper bound on
the number of triangulations in the greedy algorithm.

321

Exercise 35.9. Let (1 , . . . , n ), satisfy n 4, 0 < i < 2, for all 1 i n, and

1 + . . . + n = 4.
a) [2-] Prove that there exists a convex polytope P R3 with n vertices, which has
curvatures i , 1 i n.
b) [2-] Suppose n is even. Prove that there exists a simple polytope P R3 with n vertices,
which has curvatures i , 1 i n.
Exercise 35.10. a) [1] Find a collection of rays Ri as in Theorem 35.4, such that the
number of dierent cones CI is exponential.
b) [] Prove or disprove: deciding whether condition 3) is satised for given R1 , . . . , Rn R3
and 1 , . . . , n > 0, is NP-hard.

Exercise 35.11. [1] In notation of the mapping lemma (Theorem 41.8), suppose A
and B are connected, but is only locally injective: for every b B we have (a1 ) = (a2 )
implies a1 = a2 in a suciently small neighborhood of b. Prove the analogue of the mapping
lemma in this case.
35.7. Final remarks. The curvature existence theorem (Theorem 35.4) is due to Alexandrov and our proof follows [A2]. Both Theorem 35.6 and 35.7 follow from Alexandrovs
most general result. The version with the border Z and the variational principle proof in
Subsection 35.5 is due to Pogorelov (see [Pog3, 2.4]).
To explain the motivation behind the results in this section, consider the history of the
problem. Let f : B R be a convex function dened on a convex region B R2 . Think
of points sb = (b, f (b)), where b B as forming a convex surface S R3 . Following
the Cauchy problem, one can ask whether for given boundary values f : B R and
the curvature function : b (sb ), one can reconstruct the surface S? Note here the
logic in the question: the curvature is a fundamentally local parameter, so we are asking
whether such local parameter can determine a global structure of the surface. This problem
was resolved by Alexandrov in a much more general form. Theorem 35.7 is both a simple
special case and an important lemma in the proof.
Now, one ask a similar question for the surfaces of general convex bodies in R3 . Here the
curvature is dened not on a plane but on a unit sphere, via spherical projections. This case
is more technical because of the inequalities as in Lemma 35.2. Again, this was resolved by
Alexandrov in full generality. Theorem 35.4 corresponds to the case of convex polytopes,
when the curvature is concentrated in the vertices.
Recall that prior to this section, we proved a number of uniqueness and rigidity results
on convex polyhedra, including the Cauchy and Dehn rigidity theorems, as well as their
variations and generalizations. In the next two sections we give two more existence results.
One can think of this section as a gate to the next two sections.
Namely, the Minkowski theorem (Theorem 36.2) on existence and uniqueness of convex
polytopes with given normals and areas of the faces. Informally one can think of this
result as dual to Theorem 36.2. As the reader shall see, we use an unusual variational type
argument to prove this result, and a rst touch rule analogue for unbounded polyhedra.
Similarly, for the Alexandrov existence theorem (Theorem 37.1) we give a topological proof
based on the mapping lemma. In fact, the proof uses Lemma 35.3.
Let us note also that while the existence is often harder and more delicate to establish
(as in this section), the uniqueness results play an important role in classical Dierential
Geometry (see Subsection 38.5).

322

36. The Minkowski theorem

We continue the study of global geometry of convex polyhedra. The main result of
this section is the Minkowski theorem, which characterizes convex polytopes by the
facet areas and normals. The proof is based on the BrunnMinkowski inequality (see
Section 7). We also present Pogorelovs variation for the unbounded polyhedra and a
curious Alexandrovs uniqueness theorem in R3 , based on the Alexandrov lemma in
Section 22.
36.1. Normals to faces. Let P Rd be a convex polytope with facets F1 , . . . , Fn .
Denote by u i the unit normal to the facet Fi , and by ai = area(Fi ) = vold1 (Fi ) its
areas, 1 i n. The following result gives a connection between them.
Proposition 36.1. For every convex polytope P Rd as above, we have:
a1 u1 + . . . + an un = 0.
The vector equation in the proposition is obvious in the plane (see Figure 36.1).
u1

w1

w2
w3

u6
u2

u5

w6

w1
0

w4

w2
w5

u3
w5

u4

w4

w6

w3

Figure 36.1. Polygon P , normals u i and weighted normals w i = ai u i .

Proof. Let z = u H be the normal to a hyperplane H Rd . Consider an orthogonal
projection Q H of polytope P onto H. Similarly, consider projections Ai of facets Fi
onto H. Recall that facets Fi are oriented according to the (fixed) orientation of the
surface P of the polytope. Observe that these projections cover Q twice: once with
a positive and once with a negative orientation. Therefore, for the vector x defined
by x = a1 u 1 + . . . + an u n we have:
* n
+
n
n
X
X
X
ai u i , z = hx , zi.
0 =
area(Ai ) hu i , zi =
ai hu i , zi =
i=1

i=1

i=1

In other words, the scalar product hx , zi = 0 for every unit vector z Rd . This
immediately implies that x is a zero vector, as desired.

We are now ready to state the main result of this section. We prove it in the next
subsection.

323

Theorem 36.2 (Minkowski theorem). Let u1 , . . . , un Rd be unit vectors which

span Rd , and let a1 , . . . , an > 0 satisfy a1 u1 + . . . + an un = 0. Then there exists a
unique convex polytope P Rd with n facets F1 , . . . , Fn , such that area(Fi ) = ai and
the normal to Fi is ui , for all i [n].
36.2. Proof of the Minkowski theorem. Denote by P the set of all convex polytopes P Rd , with normals u i as in the theorem, up to a translation. Polytope P
can be written as the intersection of halfspaces corresponding to hyperplanes Hi at
distance zi from the origin.
n
\


()
P =
x Rd : hOx, u i i zi
i=1

To simplify the notation, we assume that z1 = . . . = zd = 0, and that u 1 , . . . , u d

are linearly independent, i.e., polytope P has a unique translation such that the
facets F1 , . . . , Fd contain the origin O. Denote by W the space of all vectors z =
(zd+1 , . . . , zn ). Let W+ W be the cone of vectors z with nonnegative coordinates,
so W+ Rnd
+ . Of course, not all z W+ correspond to a polytope in P, since some
facets may become degenerate or disappear. We denote by Pz the polytope defined
in (), and by Z the set of z W+ , such that Pz P. Observe also that
vol(Pz )
, for all d + 1 i n.
zi
The following result is a key to both parts of the Minkowski theorem. It is an easy
consequence of the BrunnMinkowski inequality (see Section 7)
area(Fi ) =

Lemma 36.3. Let K W be a set of all z Z such that vol(Pz ) 1. Then K

is a closed convex set with a unique supporting hyperplane through every boundary
point z K. Moreover, K is strictly convex, i.e., no straight interval can lie in the
boundary of K.
Proof. Since the volume vol(Pz ) is continuous in zi , we obtain that K is closed. Similarly, since the supporting hyperplanes to K are given by the partial derivatives as
above and the facet areas are continuous, we obtain uniqueness of the supporting
hyperplanes.
For the convexity, take z = (1 )z + z , where z , z P and [0, 1].
Let P = Pz + Pz be the Minkowski sum of polytopes. For every x P we have
x = x + x , where x (1 )Pz and x Pz . Therefore,

hOx, u i i = hOx, u i i + hOx , u i i (1 )zi + zi = zi ,

which implies that x Pz and P Pz . Thus, by Proposition 7.7, we have:



1/d
vol(Pz ) vol (1 )Pz + Pz (1 )vol(Pz )d + vol(Pz )d

1/d
= 1.
(1 ) +

Hence, z K, which proves that K is convex. By the second part of Proposition 7.7,
the above inequality is an equality if and only if the polytopes Pz and Pz are similar

324

(equal up to homothety). Now, if [z , z ] K is an interval on the boundary,

polytopes Pz , z [z , z ] are similar and have volume 1. Thus, all polytopes Pz are
congruent, and since they have the same cone at vertex 0, we conclude that all Pz
coincide, a contradiction. Therefore, K is strictly convex.

Proof of Theorem 36.2. We prove the existence first. Consider a hyperplane H W
defined by ad+1 zd+1 + . . . + an zn = 0. Since ai > 0, the intersection H W+ contains
only the origin O. By the lemma, set K is closed and strictly convex. Since K W+ ,
there exists a supporting hyperplane Hz kH, for some z K (see Figure 36.2). Let
us show that Pz is similar to the desired polytope.
aj
K
z
Hz

ai

Figure 36.2. Hyperplane Hz tangent to K at z .

Denote by Ai = area(Pz ) the facet areas of polytope Pz , for all i [n]. Since the
supporting hyperplane Hz is tangent to K at z , we have
(k)

Aj
1 vol(Pz )
1 vol(Pz )
vol(Pz )
Ai

=
= 0,
ai
aj
ai
zi
aj
zj
ij

where ij = ai zi aj zj lies in H, and d + 1 i < j n. We immediately have:

Ad+1 = sad+1 , . . . , An = san ,

for some s > 0.

Since vectors u 1 , . . . , u d R are linearly independent by assumption, the vector relation in the theorem implies that A1 , . . . , Ad are uniquely determined by Ad+1 , . . . , An .
We conclude that A1 = sa1 , . . . , Ad = sad , and 1s Pz is the desired polytope.
For the uniqueness, reverse the above argument. For every P P as in the
theorem, there exists a unique similar polytope Pz = cP of volume 1. The supporting
hyperplane Hz is tangent at z and thus must satisfy (k). By the lemma, Hz can
contain only one point in z , which implies that z is uniquely determined. Thus, so
is P , which completes the proof.

36.3. Orthant shape variation. Denote by Q = Rd+ the positive orthant. We say
that a normal vector u is positive if it lies in the interior of Q. We say that an
unbounded polyhedron P Rd has orthant shape if it is an intersection of Q and and
halfspaces with positive normals (see Figure 36.3).

325

Figure 36.3. Orthant shape polyhedron Q R3 with 5 boundary faces.

Theorem 36.4 (Pogorelov). For every positive unit vectors u1 , . . . , un Q, and
a1 , . . . , an > 0, there exists a unique orthant shape convex polyhedron P Rd with n
bounded facets F1 , . . . , Fn , such that the facet Fi has normal ui and area ai , for all
i [n].
Proof. Denote by P = P(u 1 , . . . , u i ) the set of all octant shape polyhedra P with
normals u i and face areas ai . Let us prove that P is nonempty. Take halfspaces
spanned by the hyperplane at unit distance from the point 1 = (1, . . . , 1) Rd and
with these normals. Denote the resulting polyhedron by P and observe that all facet
areas > 0. Then, for sufficiently small c > 0, the polyhedron cP has facet areas < ai ,
as desired.
Denote by Hi the planes spanned by Fi , and let wi 1 = Hi T be the intersection
of Hi with the diagonal T = {t 1 : t 0}. A polyhedron P P is uniquely
determined by w = (w1 , . . . , wn ) Rn . Define (w ) = w1 + . . . + wn and let w be
a maximum of . We first prove that polyhedron P corresponding to w has face
areas ai , and then that such maximum is unique.
Suppose for the area of the facet Fi in P we have area(Fi ) < ai . Increase the
value of wi in w by > 0 and let w be the resulting vector. The area of Fi will
increase, and the area of Fj , j 6= i, will decrease or remain unchanged. Thus, for
> 0 sufficiently small, we obtain a polyhedron P P with (w ) = (w ) + ,
a contradiction. Therefore, for every maximum of , the corresponding polyhedron
has facets ai .
For the uniqueness, suppose two polyhedra P, P P have facet areas ai . Consider
the corresponding vectors w , w Rn and let = maxi (wi wi ). We assume that
0; otherwise, relabel the polyhedra. Let P = P + 1 be a polyhedron
with hyperplanes Hi at least as far away from the origin as Hi . By construction,
wk = wk + for some k [n], which implies that Hk = Hk . Therefore, for the
corresponding facets we have Fk Fk , and Fk = Fk only if Hj = Hj for all facets Fj
adjacent to Fk , where by adjacent facets we mean two facets which have a common
(d 2)-dimensional face. But then wj = wj + and we can repeat the argument for
all such j. Since all facets of P are connected, we conclude that wi = wi + for all
i [n]. Applying the same argument to the unbounded facets of P , we conclude that
the unbounded facets of P intersect T at the origin. Since by construction they are
shifted by 1, we immediately have = 0 and P = P = P , as desired.


326

36.4. Alexandrovs approach. Perhaps surprisingly, in R3 the Minkowski theorem

can be proved in two steps: first the uniqueness by the argument as in the Cauchy
theorem (Theorem 26.1), and then by the topological argument as in the previous
section.
Theorem 36.5 (Alexandrov). Let P, P R3 be two convex polytopes with equal face
normals and such that neither face Fi in P fits inside the corresponding face Fi of P ,
nor vice versa. Then P and P are congruent, i.e., equal up to a translation.
In particular, when the areas of the corresponding faces are equal, we obtain the
uniqueness part of the Minkowski theorem without using the BrunnMinkowski inequality. The existence part now can be done by a topological argument and is left
to the reader (Exercise 36.3).
36.5. Exercises.
Exercise 36.1. [1-] Make a translation of all hyperplanes Hi and use equations area(Fi ) =
vol(Pz )
to obtain an alternative proof of Proposition 36.1.
zi
Exercise 36.2. [1+] Use Alexandrovs lemma (Theorem 22.4) to prove Alexandrovs
Theorem 36.5.
Exercise 36.3. [1+] Prove the Minkowski theorem in R3 using the Alexandrov theorem
(Theorem 36.5) and the topological lemma (Theorem 41.8).
Exercise 36.4. [1+] Use Exercises 22.1 and 35.11 to obtain an alternative proof of the
Minkowski theorem in R3 , without the full power of the Alexandrov lemma (Theorem 22.4).
Exercise 36.5. a) [1-] In R3 , prove the analogue of the Pogorelov theorem (Theorem 36.4) when perimeters of the faces are used in place of the areas.
b) [1-] In R4 , prove the analogue of the Pogorelov theorem (Theorem 36.4) when the surface
area of the facets, which are 3-dimensional convex polytopes, are used in place of their 3dimensional volumes.
c) [1-] In R4 , the same with mean curvatures.
Exercise 36.6. [] Find the perimeter analogue of the Minkowski theorem in R3 ?
Exercise 36.7. [1] Extend Pogorelovs theorem to general convex cones in Rd .
Exercise 36.8. (Dihedral angles in simplices) a) [1-] Use Proposition 36.1 to prove that
every facet of a simplex in Rd has an acute dihedral angle with some other facet.
b) [1] Prove that every simplex in Rd has at least d acute dihedral angles.
Exercise 36.9. (Equihedral tetrahedra) [1+] Prove that every tetrahedron R3 with
equal face areas is equihedral (see Exercise 25.12).
b) [1-] Prove that for all d 4, every simplex Rd with equal areas of 2-dimensional
faces is regular.
Exercise 36.10. (Shephard) A polytope P Rd is called decomposable if P = P1 + P2 for
some polytopes which are not similar to P . Otherwise, P is called indecomposable.
a) [1] Prove that except for a triangle, every convex polygon is decomposable.
b) [1-] Prove that every simplicial polytope P Rd is indecomposable.
c) [1-] Prove that the truncated cube and the regular dodecahedron are decomposable.

327

d) [1+] Prove that except for a tetrahedron, every simple convex polytope in R3 is decomposable.
e) [1+] Find two combinatorially equivalent polytopes in R3 , one of which is decomposable
and the other is not.
f ) [2-] Suppose every two edges of a polytope P R3 are connected by a chain of adjacent
triangles. Prove that P is indecomposable.
g) [2-] Prove that every convex polytope P R3 with n faces is the Minkowski sum of at
most n indecomposable polytopes.
h) [2-] Generalize e), f ) and g) to higher dimensions.
Exercise 36.11. (Wul ) [1] Let P be a polytope dened by (k) in the proof of Theorem 36.5. Then P maximizes the sum z1 + . . . + zn among all polytopes P with the same
volume and facet normals.
Exercise 36.12. (Lindel
of ) Let P, P Rd be polytopes with the same face normals
u 1 , . . . , u n and facet areas ai = area(Fi ) and ai = area(Fi ), where i [n]. Denote by zi
and zi be as in Subsection 36.2.
a) [1-] Dene (P, P ) = 1d (z1 a1 + . . . + zn an ). Observe that (P, P ) = dvol(P ). Check
that (P, P ) is translation invariant.
b) [1+] Fix P and consider all P as above which satisfy (P, P ) = dvol(P ). Use Lagrange
multipliers to show that vol(P ) is maximized if and only if P is an expansion of P .
c) [1+] Conclude that 3 (P, P ) vol(P )vol(P )2 and the equality holds only if P is an
expansion of P .
d) [1+] Let P be a circumscribed polytope around a unit sphere. Observe that vol(P ) =
1

d
d area(P ). Prove that of all polytopes in R with given volume and facet normals, the
circumscribed polytope has the smallest surface area.
e) [1-] Deduce from here the isoperimetric inequality.
Exercise 36.13. Let P, P Rd be convex polytopes as in the previous exercise.
a) [1+] Dene f (t) = P + tP . Prove that f (0) = (P, P ).
b) [1+] Use the BrunnMinkowski inequality (see Proposition 7.7) to deduce part c) in the
previous exercise.
c) [1] Conversely, deduce the BrunnMinkowski inequality from part c) in the previous
exercise.
Exercise 36.14. (Minkowskis symmetry criterion) [1] Let P Rd be a convex polytope
such that for every facet F of P there exist a facet with the same area and opposite normal.
Prove that P is centrally symmetric.
Exercise 36.15. (Zonotopes in Rd ) a) [1] Suppose convex polytope P Rd and all its
faces are centrally symmetric. Prove that P is also centrally symmetric.
b) [1+] Generalize Exercise 7.16 to prove that P is the Minkowski sum of intervals.
Exercise 36.16. Let P Rd be a convex polytope such that translations of P tile the
space.
a) [1] Prove that the opposite facets of P have equal areas.
b) [1] Use the Minkowski theorem to prove that P is centrally symmetric.
c) [1+] Extend this to show that all faces of P are centrally symmetric. Use the previous
exercise to conclude that P is a zonotope.
d) [1] In R3 , show that P has at most 14 faces and this bound is tight.

328

e) [1+] Classify completely the combinatorics of all such P in R3 .

f ) [1+] In Rd , show that P has at most 2d+1 2 facets.

Exercise 36.17. [] Suppose a convex polytope P R3 can tile the whole space (not
necessarily periodically or even face-to-face). Prove that P has at most 106 vertices.
Exercise 36.18. a) [2-] Let P, P R3 be convex polytopes such that for every edge e
of P there exists a parallel edge e of P such that |e| |e |. Suppose further, that for every
face F of P there exists a parallel face F of P such that a translation of F is contained
in F . Prove that there exists a translation of P which is contained inside P .
b) [1] Check that neither of the two conditions suce by themselves.

Exercise 36.19. [2+] Find two convex polytopes P, Q R3 with parallel and non-equal
faces which satisfy the following property: for every face F of P and the corresponding
parallel face G of Q, there exists a unique translation which either ts F inside of G or
ts G inside of F .82
36.6. Final remarks. Our proof of the Minkowski theorem is a reworking of Minkowskis
original proof [Min], parts of which were later claried in [McM1]. The main idea of this
proof is implicitly based on Wuls theorem in crystallography (see [Grub, 8.4]). Let
us mention that most standard proofs the Minkowski theorem are based on the Brunn
Minkowski inequality (see [BonF, Schn2]), to the extent that they can be shown essentially
equivalent [Kla].83 Minkowski extended Theorem 36.2 to general surfaces (see [Pog4, Pog3]
for further results and references). Let us mention also the algorithmic approach [GriH]
and the important work on stability of solutions in the Minkowski theorem (see [Gro] and
references therein).
The orthant shape variation given in Subsection 36.3 is due to Alexandrov, who proved
it in [A2, 6.4] in the generality of all cones in R3 (cf. Exercise 36.7). It was extended to
higher dimensions and other functionals (see Exercise 36.5) by Pogorelov [Pog3, 7.6] (see
also [A2, 6.5] and [B
arV]). The proof in Subsection 36.3 follows the original Pogorelovs
proof.
For the complete proof of the Alexandrov theorem (Theorem 36.5) see [A2, 6.2] (see
also [Lyu, 30] and [Mi3]). Let us mention that a series of Exercises 22.1, 35.11 and 36.4 outline a similar approach, which substitutes the tedious Alexandrovs lemma (Theorem 22.4)
with an elegant argument in the Alexandrovs local lemma (Exercise 22.1).

82The

corresponding faces must lie on the same side of the polytopes; otherwise any tetrahedron P
and Q = 2P will work.
83To quote [Kla], they have equiprimordial relationship.

329

37. The Alexandrov existence theorem

In this section we prove the Alexandrov existence theorem, giving a complete characterization of convex polyhedral surfaces in R3 . This is a celebrated result with has
a number of important applications, including Theorem 30.1 (see later in this section)
and Leibins theorem (Theorem 38.2). We present a topological proof using the mapping lemma (Theorem 41.8) and the Alexandrov uniqueness theorem (Theorem 27.7).
37.1. The statement. Let S be an abstract 2-dimensional polyhedral surface with n
vertices. We say that S is intrinsically convex if the curvatures 1 , . . . , n of all
vertices satisfy 0 < i < 2, for all i [n]. Of course, the surface of every convex
polytope in R3 is intrinsically convex. The following result shows that there are
essentially no other examples.
Theorem 37.1 (Alexandrov existence theorem). Every intrinsically convex 2-dimensional polyhedral surface homeomorphic to a sphere is isometric to the surface P of
a convex polytope P R3 , or to a doubly covered polygon.
By the Alexandrov uniqueness theorem (Theorem 27.7), such a polytope is uniquely
determined, up to a rigid motion. In other words, one can view Theorem 37.1 as the
converse of Theorem 27.7.
To fully appreciate the Alexandrov existence theorem, let us note that the edges
of the polytope are determined by the surface S, but there does not seem to be an
efficient algorithm to construct it directly (see Exercise 37.11). If the edges are given
and form a triangulation, one can obtain explicitly all realizations of S with these
edges (see Remark 31.5). The Cauchy theorem says that for each triangulation there
is at most one convex realization (up to rigid motions). In this language, the existence
theorem implies that such convex realization exists for one or more triangulations of S.
Example 37.2. To see how the triangulation of the surface can change under a small
deformation, consider the following two cases. Start with the surface of the cube and push
the opposite vertices towards each other by a small . Take the convex hull. Alternatively,
pull these vertices away from each other by a small . Take the convex hull. Now observe
that although these are combinatorially dierent triangulations, intrinsically the surfaces
are -close to each other.

S1

S2

Figure 37.1. Two triangulated -close surfaces S2 and S2 .

330

Example 37.3. (Surfaces from unfoldings) Following Alexandrov, let us restate the theorem in terms of unfoldings. Let X R2 be a simple polygon in the plane with an even
number of edges, and let U R2 be the non-convex region bounded by X, i.e., X = U .
Orient the edges of X clockwise. Let : X X be an involution on edges of X describing
the gluing of U . One can show that the surface U/ is well dened and homeomorphic to
a sphere if and only if the following conditions are satised:
(i) the attached edges must have equal lengths and the opposite orientation,
(ii) the involution on edges in this case can be drawn inside U (see Figure 37.2).
Note that the involution maps several dierent vertices of X into the same vertex v of the
surface S = U/. The 2-dimensional polyhedral surface S is intrinsically convex in the sum
of angles around all preimages of v is < 2. When all these conditions are satised, the
Alexandrov existence theorem says that U is the unfolding of a convex polytope. Finally, let
us mention that U is a general unfolding, not necessarily an edge unfolding (see Section 40).

U
v

Figure 37.2. The standard unfolding U of a cube and the gluing map .

37.2. Topological proof. Denote by Pn the space of all convex polytopes in R3

with n (labeled) vertices. Here the distance between polytopes P1 , P2 Pn can be
defined in a number of different ways, e.g., as the maximal distance between the
corresponding vertices. We include into Pn the doubly covered n-gons, which can be
viewed as surfaces of flat polyhedra.
Lemma 37.4. Space Pn is an open connected 3n-dimensional manifold.
It is tempting to deform a polytope so as to simplify it and then connect to a
special polytope P Pn . The main obstacle here is that the number of vertices of
polytopes must remain the same at all times. Here is how we get around this problem.
Proof. For connectivity, fix a polytope P Pn and translate it so that the origin O
lies in its relative interior. Consider rays Ri from 0 to all vertices vi of P , where
i [n]. Denote by P the polytope which lies on rays Ri and has unit ray coordinates
(see Section 35). By Lemma 35.3, polytope P is connected to P . Observe that P
is inscribed into a unit sphere S2 . Similarly, every doubly covered n-gon is connected

331

to a doubly covered n-gon inscribed into S2 . Therefore, it suffices to show that all
polytopes in Pn inscribed into S2 are connected.
Let vi S2 denote the vertices of P . By definition, P = conv{v1 , . . . , vn }. In
the other direction, for every u1, . . . , un S2 the polytope Q = conv{u1, . . . , un } lies
in Pn . Therefore, the space of polytopes Q Pn inscribed into S2 is homeomorphic
to the space of distinct n-tuples of points in S2 , and thus connected.
Now, take P Pn and perturb the vertices. When the perturbations are sufficiently
small they give a convex polytope. The space of perturbations is 3n-dimensional,
which implies that Pn is an open 3n-dimensional manifold.

Denote by Mn the space of all 2-dimensional polyhedral surfaces homeomorphic to
a sphere with n (labeled) vertices, defined as points with positive curvature. Again,
the topology on Mn can be defined via the pairwise geodesic distance between the
corresponding vertices.
Lemma 37.5. Space Mn is an open connected (3n 6)-dimensional manifold.
Proof. Take S Mn with vertices v1 , . . . , vn . Fix a shortest path ij between vi
and vj , for all 1 i < j n. If there is more than one such path, choose either one.
Since S is intrinsically convex, by Proposition 10.1 paths ij do not contain vertices
in its relative interior. The geodesic triangle = 12 13 23 subdivides S into two
regions. Vertex v4 lies in one of these regions and connected by 14 , 24 and 34 to
the vertices in its boundary. Vertex v5 lies in in one of the resulting four regions, etc.
Denote by T the resulting geodesic triangulation of S. Observe that T has 3n 6
edges (see Corollary 25.2 or use induction in the construction).
By construction, every triangle in T is flat, i.e., has no vertices in its interior.
Perturb the edge lengths of T . Since every vertex vi has, by definition, strictly
positive curvature, when the perturbations are sufficiently small they determine a
surface S Mn . Thus, Mn is open and the space of such perturbations is (3n 6)dimensional.
To prove that Mn is connected, let us show that triangulation T can be realized
in R3 with the length function ij = |ij | (see Subsection 31.1). In other words,
we claim that there exists a (possibly, self-intersecting) polyhedral surface S R3
isometric to S, and such that S is triangulated into flat triangles according to T ,
with the same edge lengths ij .
Use the construction of T . Suppose vertex vn is connected to a geodesic triangle
[vi vj vk ]. The geodesic triangles A1 = [vn vi vj ], A2 = [vn vi vk ] and A3 = [vn vj vk ] are flat
and their angles at vn are together < 2 and satisfy the triangle inequality. Thus,
there exists a tetrahedron in R3 with faces A1 , A2 and A3 . Change S into a surface S1
by substituting A1 , A2 and A3 with a flat triangle [vi vj vk ]. Since the vertex curvatures
do not increase, we conclude that S1 Mn1 .84 Repeat the construction. At the
end, we obtain a doubly covered triangle Sn3 M3 , which can be realized in R3 .
Going backwards, we obtain the desired realization S .
84Although

ij in the surface S1 may no longer be the shortest paths, they are still geodesics and
dene a geodesic triangulation of S1 .

332

Figure 37.3. Deformation of a surface into a convex surface.

Now, let us deform the surface S to a convex surface. As in the inductive construction above, let z S1 be the center of mass of the triangle (vi vi vk ). Move vn
along [vn z] towards z, until the resulting vertex vn is sufficiently close to z. Repeat
the same procedure for vn1 , etc. Here we change the surface triangles linearly, as
the vertices vi are moved (see Figure 37.3). At the end we obtain a convex surface
sufficiently close to a doubly covered triangle (v1 v2 v3 ). By Lemma 37.4, all convex
surfaces in Pn are connected, which implies that Mn is also connected.

The proof of the Alexandrov existence theorem now easily follows from the lemmas
above and the mapping lemma (Theorem 41.8).
Proof of Theorem 37.1. Denote by Pn the space of convex polytopes P Pn modulo
the rigid motions. Let : Pn Mn be the natural map giving the surface of
the polytope: (P ) = P . Map is continuous by definition and is injective by
the Alexandrov uniqueness theorem (Theorem 27.7). The group of rigid motions is
6-dimensional, so by Lemmas 37.4 and 37.5 manifolds Pn and Mn have the same
dimensions.85 Similarly, the lemmas imply that both manifolds are connected. It
remains to prove that is proper. By Theorem 41.8, this would imply that is a
homeomorphism, and prove the theorem.
Consider a sequence of polytopes P1 , P2 , . . . Pn such that Si = (Pi ) S Mn
as i . We can always assume that Pi contain the origin O. Denote by D the
geodesic diameter of the limit surface S. Then, for sufficiently large N, polytopes Pi
lie in the ball of radius D centered at O. Thus, a subsequence of {Pi } converges to
a convex polytope P . To prove that P Pn , simply observe that the curvature
of a vertex vk in P is the limit of the corresponding curvatures of i in Pi . Since i
are the curvatures in S and Si , for all i 1, by continuity we conclude that is the
curvature of vk in S. By the assumption, S Mn , which implies that > 0 and that
P is strictly convex at vk , for all k [n]. We conclude that P Pn . This implies
that is proper and finishes the proof of the theorem.


37.3. Back to flexible polyhedral surfaces. We can now prove a holdover result
on flexible polyhedral surfaces.
85Formally

speaking, one needs a separate argument to show that Pn is a (3n 6)-dimensional

manifold. For that, consider planted polytopes as in Subsection 33.1 and follow the proof of
Lemma 37.4.

333

Proof of Theorem 30.1. Let S0 = P be a polyhedral surface of the convex polytope P . Denote by St , t [0, 1], the polyhedral surface obtained from S0 by decreasing
the length of edge e by t. Clearly, the surfaces St are (intrinsically) isometric to S0
everywhere except two triangles containing the edge e. For > 0 small enough, the
surfaces St are locally convex. By the Alexandrov existence theorem (Theorem 37.1),
there exist convex polytopes Pt realizing the surfaces: St = Pt . Moreover, for small
enough the combinatorial structure of Pt is the same as of P . Thus, removing two
triangular faces containing e in the polyhedral surfaces St gives the desired flexing
of the surface S as in the theorem. It remains to show that these surfaces are not
globally isometric. This follows from the difference in the edge length |e|.

37.4. Exercises.
Exercise 37.1. [1-] Suppose S is an abstract 2-dimensional polyhedral surface obtained
by gluing four triangles as in a tetrahedron. Suppose also that the sum of angles at each
vertex of S is < 2. Prove that there exists a tetrahedron R3 with the surface isometric
to S.
Exercise 37.2. a) [1] Find the analogue of Theorem 37.1 for convex caps (cf. Subsection 25.6 and Theorem 27.5). Deduce this version from the Alexandrov existence theorem.
b) [1+] Give a topological proof of this result.
Exercise 37.3. [1-] For the unfoldings in Figures 19.4 and 19.7 draw the involutions inside
the polygons.
Exercise 37.4. a) [1-] Find a polygon in the plane which is an unfolding of two dierent
tetrahedra (corresponding to dierent gluing maps )
b) [1] Show that the square is an unfolding of innitely many dierent convex polytopes.
c) [1+] Describe all polyhedra which have a square unfolding.
Exercise 37.5. (Harer-Zagier formula) Let Rn a regular 2n-gon and denote by n the set
of xed point free involutions : [2n] [2n] on edges of Rn . Clearly, |n | = (2n 1)!! =
1 3 . . . (2n 1). Draw by straight lines inside Rn , and take the orientable surface Rn /
obtained by gluing the corresponding edges. Denote by ak (n) the number of n such
that genus(Rn /) = k.
a) [1] Prove that Rn / is homeomorphic to a sphere if the drawing of has no crossings.
Conclude that a0 (n) is a Catalan number.
b) [1] Give a combinatorial description of the genus(Rn /).
c) [2] Prove that for every N 0, we have:
n/2

X
k=0

ak (n)N

n+12 k

= (2n 1)!!

N
X

m=1

m1

 
n
N
.
m1 m

Exercise 37.6. [1] Prove that conditions (i) and (ii) in the Example 37.3 are necessary
and sucient. Compare with part a) of the previous exercise.
Exercise 37.7. [1+] Let S be a 2-dimensional polyhedral surface homeomorphic to a
sphere (not necessarily intrinsically convex). Use the construction in the proof of Lemma 37.5
to obtain a piecewise linear immersion of S into R3 .

334

Exercise 37.8. Start with unfolding the regular icosahedron given in Figure 19.4. Perturb
the edges by at most so that the corresponding edges under remain of the same length.
b) [1-] Use the Alexandrov existence theorem to prove that for suciently small this is
an unfolding of a convex polytope. Find an explicit bound on .
b) [1] Use the algebraic approach in Section 31 to show that for suciently small this is
an edge unfolding of a convex polytope combinatorially equivalent to the icosahedron.
c) [1] Use the proof of the Alexandrov existence theorem to obtain part b). Explain the
connection to the proof of b).
Exercise 37.9. (Twisting the surfaces) Let S = P be the surface of a convex polytope
P R3 , and let S be a closed geodesic. Denote by S1 and S2 the surfaces on both
sides of . A surface S obtained by gluing S1 and S2 along the boundary is called a twist
of S.
a) [1-] Use the Alexandrov existence theorem to check that all surface twists are surfaces
of convex polytopes.
b) [1+] Start with a regular tetrahedron. Check that all polytopes obtained by a sequence of
twists must be equihedral tetrahedra (see Exercise 25.12). Prove or disprove: all equihedral
tetrahedra with equal surface area can be obtained this way.
c) [] Start with a square pyramid where all vertices have equal curvatures. Describe all
polytopes resulted by a sequence of twists.
d) [] Same question for a regular octahedron.
Exercise 37.10. (Volkovs stability theorem) [2+] Let S, S two intrinsically convex
2-dimensional polyhedral surfaces homeomorphic to a sphere with geodesic diameter at
most 1. Suppose : S S is a homeomorphism such that
|x, y|S < |(x), (y)|S < |x, y|S + ,

P, P

R3

for all x, y S .

Denote by

the corresponding convex polytopes. Prove that there exists a rigid

motion such that P and P are close to each other:
|z, (z)| < C 1/24 , for all z P (so that (z) P ),

and where C is a universal constant.

Exercise 37.11. [] Prove or disprove: computing edges in the polytope corresponding to

a given intrinsically convex 2-dimensional polyhedral surface is NP-hard.
37.5. Final remarks. The Alexandrov existence theorem and its unbounded variations
(cf. Subsection 27.3) was proved by Alexandrov in a series of papers. He presented two
interrelated proofs in [A1, A2], both based on elementary but delicate work with unfoldings.
The proof in this section is largely new, although the use of the mapping lemma is motivated
by the proof in [A2].
A variational proof for the case of polyhedral caps (see Exercise 37.2) was given by
Volkov in [Vol2], and was extended to general polytopes and general unbounded polyhedra in [VP1]. This proof was motivated by Pogorelovs proof in Subsection 35.5. Most
recently, a conceptual variational proof was found in [BobI], which also proved amenable
to a computer implementation. This proof builds on top of Volkovs proof and uses several
technical innovations, including an extension of the AlexandrovFenchel inequality, which is
an advanced extension of the BrunnMinkowski inequality and other geometric inequalities.
An accessible exposition of the convex cap variation is given in [Izm].

335

Let us mention here an important Volkovs stability theorem (Exercise 37.10), which he
proved for both bounded convex surfaces and for the convex caps [Vol3]. In a dierent direction, the BuragoZalgaller theorem is a non-convex analogue of the Alexandrov existence
theorem (see Exercise 39.13 and compare with Exercise 37.7).

336

38. Bendable surfaces

In this section we discuss bendable surfaces whose properties turn out to be very
different from those of flexible surfaces in Section 30. As in the Alexandrov existence
theorem, we place the emphasis on intrinsic geometry of surfaces rather than on
geometry of faces.
38.1. Bending surfaces is easy. We say that a 2-dimensional polyhedral surface
S R3 is bendable if there exists a continuous family {Bt : t [0, 1]} of embeddings
Bt R3 , which are intrinsically isometric to S, but every Bt and Bt are not globally
isometric. The above family of surfaces {Bt } is called a bending of S. Of course, by
Pogorelovs uniqueness theorem (Theorem 27.8), at most one surface in the bending
of S can be convex. On the other hand, every convex polyhedral surface is bendable
(see Figure 38.1).

B0

B1

Figure 38.1. Surface S = B0 of the cube is bendable.

Proposition 38.1. Let S = P be a polyhedral surface of a convex polytope P R3 .
Then S is bendable.
Proof. Let v be a vertex of P and let H be a plane supporting v and in general
position, i.e., a plane such that H P = {v}. Denote by {Ht } the family of planes
parallel to H, and shifted by t in the direction of P . Let Qt = P Ht be the polygon in
the intersection. Finally, let Bt be a non-convex surface obtained by gluing along Qt
two parts of S divided by Ht (as in Figure 38.1). The details are straightforward. 
38.2. Convexly bendable polyhedral surfaces. Let us restrict out attention to
part of polyhedral surfaces of convex polytopes. Formally, we say that (non-closed)
polyhedral surface S R3 is globally convex if there exists a convex polytope P
such that S P . We say that S is convexly bendable if there exists a bending
{Bt : t [0, 1]} where all surfaces Bt are polyhedral86 and globally convex. For
example, the union of three sides of a tetrahedron is a convexly bendable surface (see
Figure 38.2).
The following result shows that in fact this example generalizes to every polytope.
86Non-closed

surfaces, of course, can have non-polyhedral bending, e.g., the surface of a cube
without two opposite faces can be easily deformed into a cylinder. We restrict ourselves to polyhedral
surfaces for simplicity.

337

Figure 38.2. Three sides of a tetrahedron form a convexly bendable surface.

Theorem 38.2 (Leibin). Let F be a face in a convex polytope P R3 . Then the
polyhedral surface S = P r F is convexly bendable.
The moral is that bendability is a general property, and without additional restrictions, such as smoothness of closed surfaces, one cannot hope to have any kind of
non-bendability rigidity. Once again, the proof is an easy application of the Alexandrov existence theorem.
Proof. The strategy is similar to that in the proof of Theorem 30.1. Fix an interior
point w F and consider a pyramid Zt over F with a vertex z which projects onto w
and has height t, for every t [0, 1]. Denote by Pt a polytope obtained from P = P0
by attaching a pyramid Zt . Clearly, for sufficiently small > 0, polytopes Pt are
convex. Furthermore, the convex polyhedral surface S is a polygonal region on the
closed surface St of Pt .
Denote by A the boundary polygon of S, i.e., A = F = S. Fix a flat point x A
on an edge e = (p, q) of A. For sufficiently small , all shortest paths on St between x
and points x F lie on the surface Zt . Since the shortest path cannot go through
the vertex z of Zt (see Proposition 10.1), points x 6= x will split into two intervals
depending on which side of z they have a shortest path to x. By continuity, there
exists a point y A with exactly two shortest paths and between x and y.87
Denote by Rt St the closed polygonal region between the paths. Consider a closed
surface Ut obtained from St by removing Rt and gluing along the corresponding points
on paths and . It is easy to check that for small enough the surfaces Ut are
all locally convex. By the Alexandrov existence theorem (Theorem 37.1), they can
be realized by convex polytopes Qt . Removing the remaining parts of the pyramid
boundary Zt from Qt , we obtain a continuous family of surfaces {Bt , t [0, 1]} of
the original surface S = B0 .
We claim that the {Bt } is the desired convex bending of S. It remains to prove
that polyhedral surfaces Bt are not globally isometric. First, observe that point x
is no longer flat and is a vertex of Qt . Moreover, since the curvature of vertex x
in Qt is increasing with t, the surface angle = pxq is decreasing (on Ut r Bt ).
Since the (usual space) angle = pxq in R3 between (new) edges (p, x) and (x, q)
87This

is a basic step in the cut locus construction (see Section 40).

338

F
p

Rt

q
P

Pt

Qt

Figure 38.3. A convex bending {Bt Qt } of the polyhedral surface

S = P F with boundary A = F .
is at most , it is also decreasing for sufficiently small > 0.88 Since the edges (p, x)
and (x, q) lie on the boundary Bt , we obtain the claim.


ir

Kef

ir

Kef

Mil

lo
Mo

ko

38.3. Packaging liquid can be a hard science. In the good old days milk (and
other liquid) used to be sold in small containers in the shape of a regular tetrahedron
(see Figure 38.4). Such containers are easy to manufacture (they have a nice unfolding
which are easy to cut out of cardboard) and easy to store.89 Unfortunately they were
terrible in use, since to open them one had to use scissors, and no matter how small
the hole was, the container would bend and some milk would inevitably spill. This
can all be explained, of course, by Theorem 38.2 which states that all such surfaces
with cuts are bendable.

Figure 38.4. Old and new milk cartons.

Now, the present day milk cartons look like bricks with a big hole is in the middle
of a face (see Figure 38.4). When closed, the hole is covered with a plastic cap, but
even when the cap is removed, such cartons are quite rigid. Naturally, this begs a
theoretical explanation. The following result does exactly that.
Theorem 38.3 (Alexandrov). Let A P be a polygon on a surface of a convex
polytope P R3 , and suppose no vertices of P lie in the polygon A or on its boundary.
Then the polyhedral surface S = P r A is not convexly bendable.
88A

priori the surface angle may be greater than the space angle if the triangle (pxq) is not
at on the surface Ut .
89There are even quite good tetrahedral packings of R3 : http://tinyurl.com/23y8v5e

339

For example, let F is the face of a convex polytope P (milk carton), and let A F
be a polygon in the relative interior of the F (the hole). The theorem says that
S = P r A is not convexly bendable (i.e., milk is harder to spill).
Proof. Consider a convex polytope P and a polyhedral surface S P intrinsically
isometric to S. By the GaussBonnet theorem (Theorem 25.3, the total curvature
(S) = (S ) = 2. Thus, all vertices of P lie in the interior of S , and points in
A := P r S are all flat. Since polygons A S and A S may not lie in one face,
denote by B and B the unfoldings of A and A on a plane, respectively.
Observe that by the intrinsic isometry of surface S and S , polygons B and B in R2
have equal side lengths and equal outside angles. Therefore, B and B are equal,
which implies that A and A are isometric. We conclude that surfaces P = S A
and P = S A are both closed convex polyhedral surfaces which are intrinsically
isometric. By the Alexandrov uniqueness theorem (Theorem 27.7) they are globally
isometric, which completes the proof.

38.4. Exercises.
Exercise 38.1. a) [1] Give a direct construction proving Leibins theorem (Theorem 38.2)
for the cube and the regular octahedron.
b) [1+] Same for the regular icosahedron and dodecahedron.
Exercise 38.2. a) [1+] Let S1 P1 , S2 P2 be isometric polyhedral surfaces homeomorphic to a disk. Assume that all (intrinsic) angles at the boundaries S1 S2 are > .
Prove that there exists a convex bending from S1 to S2 .
b) [2-] Same result under assumption that all boundary angles are .
b) [] Same result without restriction on the angles.
38.5. Final remarks. The study of exible and bendable surfaces was initiated by Liebmann who proved rigidity (in fact, uniqueness) of a sphere as an analytic closed surface of
constant curvature (see an especially pretty proof by Hilbert in [Bla1, 91]). This study of
rigidity of analytic surfaces was continued in the works of Hilbert, Blaschke, Cohn-Vossen,
Weyl, Alexandrov, Pogorelov, and others. We refer to [Pog3, Sab1, Sen2] for references and
historical overview.
Theorem 38.2 is proved in [Leib], and holds in much greater generality, for all convex
surfaces with removed regions of strictly positive curvature (see also [Pog3, 3.6]). The
terminology we use here diers from the standard, complicated by the ambiguity of Russian
phraseology. The essence of the proof remains the same, however. Theorem 38.3 is due to
Alexandrov, and has been later extended to unbounded and other surfaces with boundary,
see [Pog3, Sho2] (also, compare with [Sho1]).

340

39. Volume change under bending

We have seen two examples when a closed convex polyhedral surface has several
realizations (say, embeddings) in R3 (see Figure 30.1, 38.1). In each case the volume
of non-convex embeddings is smaller than that of convex. One can ask if this is
a general rule, i.e., whether the unique convex embedding (by Theorem 27.6) has
a greater volume than all other embeddings. Surprising or not, the answer to this
question is negative. In this section we exhibit volume-increasing bendings of the
surface of a doubly covered triangle, regular tetrahedron, and a unit cube. We then
discuss various general results in this direction.
39.1. BuragoZalgallers bending. The following construction of a bending {Bt }
has an extremal point doubly covered triangle of volume 0. Start with a rhombus
divided into two equilateral triangles (ABD) and (BCD). Take two points E, F on
the diagonal [AC] so that [BF DE] is a rhombus and |EF | = t |AC|. Now fold the
rhombus [ABCD] along the diagonal (BC), and glue (AB) to (BC), (AD) to (DC).
Push the edge (EF ) inside so that we obtain a polyhedron Bt with faces (ABE),
(ADE), (CBF ), (CDF ), (BEF ) and (DEF ), as shown in see Figure 39.1. Note
that is is concave along the edge (EF ), and convex along other edges.
B

A
E

B
E
F

F
D

Figure 39.1. BuragoZalgallers bending construction.

This shows a non-convex polyhedron can have a greater volume than a convex
polyhedron with an isometric surface (see also Exercise 39.5). If the reader is unhappy
about the doubly covered triangle as polytope, there is a simple way to modify this
construction to make it the usual polytope. Below we present somewhat different
(and more complex) examples of the same phenomenon.
39.2. Bleeckers bending. Let be a regular tetrahedron with vertices v1 , v2 , v3 , v4 ,
side lengths |vi vj | = 1, and center O. Consider a pyramid T = (Ov1 v2 v3 ), which can
be viewed as one quarter of . Subdivide each side of symmetrically, as in
Figure 39.2. Denote by S = .
Consider a non-convex polyhedron Pt with surface Bt = Pt (intrinsically) isometric
to S, as in the figure. Bend the edges perpendicular to the middle of the edges so that
the equilateral triangles in the middle of the faces stick out. In the figure we show
only one quarter Q of the resulting polytope Pt , corresponding to T . More precisely,
let (w1 , w2 , w3 , O) be a rectangular pyramid homothetic to T , and fix edges (ai , bi )

341

v3

v3

w3

a2 t
t a1
s
b1
b2
b3
v1

a3

v2

v1

v2

w1

w2

Figure 39.2. One quarter T of a regular tetrahedron and an one

quarter Q of isometric non-convex polyhedron P of bigger volume.
perpendicular to edges
(ai wj ), for all j 6= i. Denote by s = |bi bj | and t = |ai bi |. Note

that |a1 a2 | = s + 3t, so s is determined by t.

By construction, all surfaces Bt are intrinsically isometric to the surface B0 = S of
the regular tetrahedron. Thus we obtain a bending {Bt } of S, where parameter t can
take any value between 0 and = 21 3 .
Proposition 39.1. We have vol(Pt ) > 1.37 vol(), for some t [0, ].
In fact, one can show that vol(Pt ) is increasing for all t > 0 such that t 2s (see
Exercise 39.2). Thus, one can think of the resulting surfaces {Bt } as of a volumeincreasing bending of a surface of a regular tetrahedron . Rather than make this
elaborate ad hoc calculation, let us compute directly the maximal volume of Pt .
Proof. Choose parameter t such that it is equal to half the side of the equilateral
triangle (b1 b2 b3 ) in the middle: |b1 b2 | = s = 2t. Since |a1 a2 | = 21 , from above we have:
t =

|a1 a2 |
1
=
0.1340 .
(2 + 3)
4+2 3

We also have

1
+ t2 0.5176 .
4
Observe that the volume of a regular tetrahedron is equal to
r

1
1
3
2
1
vol() = area(v1 v2 v3 ) height() =

= 0.1179 .
3
3
4
3
6 2
:= |vi bj | =

On the other hand, polyhedron P is a union of four hexagonal pyramids Ri with apexes
at vertices wi , and
a polytope Y shown in Figure 39.3. The height of each pyramid Ri
is equal to h := 2 s2 0.4429. Note that Y is a truncated tetrahedron with edge
lengths equal to s. The volume of Y is equal to the volume of a regular tetrahedron
with edge length (3s) minus four times the volume of a regular tetrahedron with side

342

length s. We obtain:
(3s)3 4s3
1

vol(P ) = vol(Y ) + 4 vol(Ri ) =

+ 4
3
6 2

3 2
6
s
4

h 0.1623 .

s
s
s

Figure 39.3. Truncated tetrahedron Y and a way to compute its volume.

Thus, vol(P ) > 1.37 vol(), as desired.

For example, the volume of a tetrahedral milk carton can increase by as much as
37% under continuous (but non-convex) bending. By applying the same construction
to small pyramids in P one can further increase the volume by a bending, but it is
unclear how large the volume can get. Of course, such volume is bounded by the
isoperimetric inequality (Theorem 7.8) which gives an upper bound of about 82%. It
would be interesting to see which of these two bounds is closer to the truth.
39.3. Milkas bending. In this section we present a symmetric volume-increasing
bending of a cube. This construction requires few calculations and motivates the
general construction in the next subsection.
Consider a cube C with side length 1 and the surface S = C. Fix a parameter
(0, 12 ). Think of as being very small. On each face of the cube, from every
corner remove a square of side-length . Denote by R S the resulting surface with
boundary. On every face of R there are four boundary points which form a square.
Call these points corners. Translate each square directly away from the center of the
cube (without expanding the squares) until all of the distances between the nearest
corners on adjacent faces reach 2. Take the convex hull of the corners to obtain a
polytope Q . To each triangular face of Q attach a triangular pyramid whose base
is equilateral with side-length 2 and whose other faces are right triangles. Denote
by P the resulting (non-convex) polyhedron.
Note that the surface Q without triangular faces between the corners is isometric
to R. Similarly, three squares meeting at a vertex of the cube can be divided
into six right triangles which can be then bent into three faces of a pyramid (see
Figure 39.4). This easily implies that the surface P is isometric to the surface S.
Proposition 39.2. We have vol(P ) > 1 for > 0 sufficiently small. Moreover,
vol(P ) increases for > 0 sufficiently small.

343

Figure 39.4. Construction of Milkas volume-increasing bending.

Proof. Cut Q with six planes, each parallel to a square face and containing the nearest
edges of its four neighboring square faces. This subdivides Q into one (interior) cube,
six slabs (along the faces), twelve right triangular prisms (along the edges), and eight
pyramids
(one per cube vertex). Observe that the cutting planes are at distance

d = 2 from the square faces of Q . We have:

vol(Q ) = (1 2)3 + 6 (1 2)2 d + 12 (1 2)
Since vol(P ) = vol(Q ) + 8

d2
d3
+8 .
2
6

d3
,
6

we conclude that

vol(P ) = 1 + 6( 2 1) + O(2),

This implies that for small > 0, we have vol(P ) > 1, and the volume vol(P ) is
increasing.

39.4. General volume-increasing bendings. We are ready to state the main result of this section generalizing the previous examples.
Theorem 39.3 (Pak). Let P R3 be a convex polytope, and let S = P be its surface. Then there exists an embedded polyhedral surface S R3 which is intrinsically
isometric to S and encloses a larger volume.
By the Alexandrov uniqueness theorem (Theorem 27.7) convex surface S is unique
up to a rigid motion, which implies that the surface S in the theorem must be nonconvex. Here is a convex variation on Theorem 39.3.
We say that a surface S R3 is submetric to S, write S 4 S, if there exist a
homeomorphism : S S which does not increase the geodesic distances: |x, y|S
|(x), (y)|S for all x, y S. Of course, the (intrinsically) isometric surface are
submetric.

344

Theorem 39.4 (Pak). Let P R3 be a convex polytope, and let S = P be its

surface. Then there exists a convex polytope P R3 such that the surface S = P
is submetric to S and vol(P ) < vol(P ).
In other words, the theorem says that the surface S can be triangulated in such
a way that each triangle can be made smaller to assemble into a submetric convex
surface S 4 S which encloses a larger volume. Note that if a new triangulation is
consistent (i.e., the corresponding triangle edges have equal lengths) and is locally
convex, then the surface S always exists by the Alexandrov existence theorem (Theorem 37.1). Let us mention also that Theorem 39.4 implies Theorem 39.3 modulo the
BuragoZalgaller theorem in the Exercise 39.13.
To see an example of Theorem 39.4, let P = C and P = Q as in the previous
subsection. Define a map : C Q as in Figure 39.4, where each of the eight
attached pyramids is projected onto triangles. This shows that P 4 P . On the
other hand, since the volume of eight pyramids is O(3 ), the same argument as above
shows that vol(Q ) > vol(P ) for all > 0 small enough.
Remark 39.5. Here is a physical interpretation of Theorem 39.3. Imagine a polyhedron is
made out of bendable, but non-stretchable material. One can then blow more air inside to
make the volume larger. Of course, one can continue blowing air until a non-inatable shape
emerges (see also Exercise 39.6). The theorem says that the resulting surface cannot be a
polyhedron. Two examples of such surfaces (for a doubly covered square and a cube) are
given in Figure 39.5. Note that the surface shrinks in both cases and in the limit becomes
submetric to the original surface (this is visible in cube).

Figure 39.5. The square and cubic balloons.

39.5. Exercises.
Exercise 39.1. a) [1-] In Milkas bending, let Q and P be the limit of polyhedra Q
1
and P , as = 2
. Check that Q is an octahedron, and that vol(P ) < 0.95.
3
b) [1-] Show that the volume of Pt maximizes at about 1.19. Compare this bound with
what follows from the isoperimetric inequality (Theorem 7.8).
Exercise 39.2. [1] In Bleeckers bending, compute explicitly vol(Pt ). Check that
that vol(Pt ) increases for t s/2 as in the proof, and compute the maximum value of
the volume. Check that at the maximum, t > s/2.

345

Exercise 39.3. a) [1] Modify Bleeckers bending to prove Theorem 39.3 for a cube. Compute the maximal volume.
b) [1] Modify Bleeckers bending to prove Theorem 39.3 for a regular dodecahedron and
icosahedron. Find the optimal bounds for the volume.
c) [1+] Modify Bleeckers bending to prove Theorem 39.3 for all simplicial polytopes.
Exercise 39.4. [1] Modify Milkas bending to prove Theorem 39.3 for a regular dodecahedron and a regular icosahedron. Compare the optimal bounds with those obtained in the
previous exercise.
Exercise 39.5. [1] Prove Theorem 39.3 directly for the doubly covered polygons (see
Figure 39.1 and Figure 39.6).

Figure 39.6. Volume-increasing bending of a doubly covered square.

Exercise 39.6. (Mylar balloon) [2-] Think of a circular party balloon as a doubly covered
circle lled with gas as much as possible (see Remark 39.5). Assume the balloon retains the
cyclic symmetry. Compute the volume of the balloon. Compute the surface area. Explain
why the surface area decreases.
Exercise 39.7. (Curvilinear cube) Consider a curvilinear surface S dened in Figure 39.7.
a) [1-] Prove that S is isometric to a cube.
b) [1+] Show that surface S exists and is uniquely dened for every symmetric convex
piecewise linear curve close enough to the edges.
c) [1] By varying the curve, give an explicit bending construction of these convex polyhedra.
Prove that this bending is volume-decreasing.

Figure 39.7. Curvilinear deformation (buckling) of a cube, and its unfolding.

346

Exercise 39.8. (Chadwick surfaces) Consider the curvilinear surfaces constructed in US

Patent 449272390 We call them Chadwick surfaces, after their inventor Lee Chadwick.
a) [1-] Explain why the Chadwick surfaces are isometric to doubly covered regular polygons,
Platonic and Archimedean solids, etc.
b) [2-] Show that the Chadwick surfaces exist and are uniquely dened for every symmetric
convex piecewise linear curve close enough to the edges.
c) [1+] By varying the curve, give an explicit bending construction in each case. Prove that
all these bendings are volume-decreasing.
Exercise 39.9. We say that a surface S R3 has a at region if it contains a surface
polygon lying in the plane. For example, the curvilinear surface in Figure 39.7 has two at
squares.
a) [1] Consider the surfaces in the patent from the previous exercise. Despite the appearance, check that all of them have at regions.
b) [1+] Construct an isometric curvilinear embedding of the surface of the regular tetrahedron without at regions.
c) [2-] Extend b) to all Platonic solids.
Exercise 39.10. An isometry between two closed piecewise linear curves C1 and C2 is
a length-preserving piecewise linear map : C1 C2 . Of course, the curves must have
the same length: |C1 | = |C2 |. An (intrinsic) isometry : S1 S2 between 2-dimensional
polyhedral surfaces S1 and S2 with boundary curves C1 = S1 and C2 = S2 denes an
isometry between the curves.
a) [1-] Suppose now that C1 = S1 , where S1 R2 is a convex polygon, and S2 R3 . Prove
that every isometry : S1 S2 as above satises the distance condition: |(x)(y)| |xy|
for all x, y C1 .
b) [1] Prove that as above satises the distance condition if and only if the inequality
holds for all x, y either vertices of C1 or preimages of vertices in C2 .
c) [1] Extend part b) to non-convex simple polygons C1 in the plane.
Exercise 39.11. (Realization of polygons) Let S = P be a 2-dimensional polyhedral
surface. Suppose S is subdivided into triangles i , 1 i k. Suppose further that there
exist a collection of triangles 1 , . . . , k Rd such that:
(1) the corresponding triangles are congruent: i i ,
(2) whenever i and j share an edge, i and j share the corresponding edge.
In other words, suppose the abstract polyhedral surface S obtained as the union of triangles i is (intrinsically) isometric to S. We say that S is a realization of S, and that S can
be realized in Rd .91
a) [1+] Let C1 = S1 , S1 R2 , be a convex polygon in the plane, and let C2 R3 be
a simple space polygon of the same length. Suppose : C1 C2 is an isometry which
satises |(x)(y)|R3 |xy|R2 for all x, y C1 . Prove that there exists a 2-dimensional
polyhedral surface S2 isometric to S1 , such that the isometry map : S1 S2 coincides
with on C1 .
b) [2-] Extend a) to non-convex simple polygons C1 in the plane.
c) [2-] Extend b) to general plane polygons with holes.
90Available

at http://www.google.com/patents?vid=USPAT4492723.
that in Rd triangles i can intersect, overlap or even coincide, so S may not be a surface
in the usual sense.
91Note

347

d) [1] Use part a) to prove that for every > 0, the surface of a given convex polytope
P R3 can be realized inside the ball of radius .
e) [1+] Let S 4 S be two convex surfaces in R3 . Prove that for every > 0, the surface S
can be realized in the -neighborhood of S . Check that this is a stronger result than that
in part c).
Exercise 39.12. (Cramming surfaces) Let S = P be the surface of a convex polytope
P R3 .
a) [1-] Use Zalgallers theorem (Theorem 40.10) show that for every > 0, there exists a
realization of S inside the ball of radius .
b) [2-] Prove that for every > 0, there exists a surface S which is isometric to S and is
embedded inside the ball of radius .
c) [2] Prove there exists a bending (continuous piecewise linear isometric deformation)
{St : t [0, 1]} such that S0 = S and S1 = S .
Exercise 39.13. (BuragoZalgallers theorem) Let S be an abstract 2-dimensional
polyhedral surface homeomorphic to a sphere.
a) [2] Prove that there exists an isometric embedding of S in R3 .
b) [2] Suppose S 4 S for some 2-dimensional polyhedral surface S R3 . Prove that for
every > 0, there exists an isometric embedding of S in the -neighborhood of S .
39.6. Final remarks. The BuragoZalgallers volume-increasing bending was presented
in [BZ4, 9]. Bleeckers bending was introduced in [Ble] (see also a friendly exposition
in [Ale5]).
Milkas bending was discovered in [Mi5] where the other regular polyhedra were also
explored and their bending analyzed. Interestingly, Milka did not notice that his bendings
were volume-increasing. Our presentations follows [Pak7].
Both theorems in Subsection 39.4 are proved in [Pak8]. The proof of Theorem 39.3 uses
Theorem 39.4 and a delicate result of Burago and Zalgaller [BZ4] (see Exercise 39.13). A
weaker version of the BuragoZalgaller theorem is given in Exercise 39.11, the rst part of
which is based on Tasmuratovs results [Tas1, Tas2]
Theorem 39.4 extends to higher dimensions and its proof uses an advanced generalization
of Milkas bending construction. Similarly, Theorem 39.3 extends to non-convex (possibly
self-intersecting) polyhedral surfaces [Pak8]. The curvilinear surface in Figure 39.7 is due
to Shtogrin [Sht1]. However, the subject of curvilinear surfaces is quite old and in the case
of the cylinder has been extensively studied by Pogorelov [Pog5].92

92According

to Zalgaller, some of the work in this direction by Alexandrov, Pogorelov and others,
remains classied because of rocket science applications (personal communication, not a joke).

348

40. Foldings and unfoldings

In this section we prove that every convex polytope can be unfolded into the plane
if cuts along faces are allowed. Our tools include the shortest paths (Section 10) and
Voronoi diagrams (Section 14).
40.1. Unfoldings of convex polyhedra. Let be the graph of a convex polytope
P R3 . For a spanning tree t , cut the surface S = P along the edges of t and
unfold S r t on the plane. The resulting edge unfolding can be either overlapping
or non-overlapping (Figures 19.4, 19.7 and 40.1). We call t the cut set and say
that P has an unfolding with the cut set t. If the unfolding is non-overlapping, the
resulting polygon is called a foldout. Since non-overlapping unfoldings are visually
very appealing, the following old conjecture says that they always exist (see also
Subsection 40.6).
Conjecture 40.1 (D
urers conjecture). Every convex polytope P R3 has a nonoverlapping edge unfolding.

Figure 40.1. Two edge unfoldings of a snubbed cube.

Here is a way to weaken the conjecture. Recall that from the metric point of view,
there is nothing significant about edges of the polytope. Suppose T S is a piecewise
linear tree, defined as a contractible union of intervals on the surface. We say that T
is spanning if it contains all vertices of P . We can then consider an unfolding of P
with the cut set T (see Figure 40.2 to appreciate the difference).

Figure 40.2. Foldouts of three unfoldings of a cube.

Let us now we present two constructions of non-overlapping unfoldings.

349

Theorem 40.2 (Source unfolding). Let S be the surface of a convex polytope P R3 .

Fix a point w S in the relative interior of a face and let Tw be the set of points
z S which have two or more shortest paths to w on the surface. Then the unfolding
of P with the cut set Tw is non-overlapping.
The point w is the called the source point, thus the source unfolding. The cut set in
this case is called the cut locus. The third unfolding in Figure 40.2 is an example of a
source unfolding of the cube, where the source point w is in the middle of a face. Note
that the cut set can be complicated even in the most basic cases (see Figure 40.3). It
is by no means obvious that Tw is a spanning tree on the surface S. We will prove
this in the next section.

w
w

Figure 40.3. Source unfolding of the cube with the source on the
bottom face and a foldout of the Alexandrov unfolding.

Theorem 40.3 (Alexandrov unfolding). Let S be the surface of a convex polytope

P R3 . Fix a point w S which has a unique shortest path to every vertex of P .
Let Tw be the union of these paths. Then the unfolding of P with the cut set Tw is
non-overlapping.
We do not prove Theorem 40.3, leaving it to the reader (see Exercises 40.1 and 40.9).
Remark 40.4. Although the statements of both theorems may seem completely clear, we
never really dened what is a (general) unfolding. The denition simplies in the nonoverlapping case. A non-overlapping (general) unfolding of the surface S with the cut set T
is an isometric piecewise linear homeomorphism : S r T U where U is a polygon in
the plane. Such a homeomorphism is called the unfolding map.
The great advantage of the source unfolding over the Alexandrov unfolding is the fact that
the former gives a star-shaped polygon centered at (w). Thus the shortest path wz unfolds
into a straight interval in U . This construction can now be used to resolve the discrete
geodesic problem of computing the geodesic distances between points (see Exercise 40.8).
Curiously, the Alexandrov unfolding is not necessarily a star-shaped polygon, even though
it is usually referred in the literature as the star unfolding.

350

40.2. Source unfolding as the source of inspiration. As it turns out, the proof
of Theorem 40.2 is straightforward once we show that the source unfolding is well
defined. We prove this result in a sequence of lemmas.
Lemma 40.5. The cut set Tw is a finite union of intervals.
Proof. Suppose point w lies in the face F of the polytope P . For every face A, consider
the set X = X(w, A) H of images under various unfoldings of the source point w
onto the plane H spanned by A. Points x X are called source images. Formally,
every shortest path wz joining w to a point z in the relative interior of A ends in
a straight segment end(wz ) in A. The source image corresponding to such path wz
is the point x H, such that the straight segment [zx] has the same length as and
contains end(wx ).
The idea is that when restricted to the face A, the Voronoi diagram VD(X) is a
complement to Tw .93 More precisely, denote by Cw (e) S the geodesic cone over an
interval e in the edge of A through which source image sees the face A, i.e., a family
of shortest paths from w to points in A which cross e. Denote by Aw (e) = A Cw (e)
the region of F which can be seen from w through e, and thus the corresponding
source image x X (see Figure 40.4). Now, for every point z A, take all regions
Aw (e) containing z and decide which source image is the closest to z. Since we
consider only shortest paths, no facet can appear twice in the corresponding geodesic
cones, which implies that the number of different cones Cw (e) is finite. Thus, face A is
subdivided into a finite number of intersections of regions Aw (e) with different e, and
in each such intersection the cut set Tw is the complement of the Voronoi diagram of
the source images. This implies that in the face A, the cut set Tw is a finite union of
intervals, as desired.

e

Aw (e)

w
F

Figure 40.4. A geodesic cone G = Cw (e) and region Aw (e).

Lemma 40.6. The complement S r Tw to the cut set is contractible.
In particular, the lemma implies that the cut set Tw is connected and simply
connected.
93This

is true as stated, but requires a separate proof by a rather delicate technical argument (see
Exercise 40.6). We use the geodesic cones in the proof to avoid this technicality even if that makes
the cut set structure less transparent.

351

Proof. By Proposition 10.1, no shortest path xy between two points x, y S can

contain a vertex of P in its relative interior. Therefore, for all z S, a shortest
path wz cannot contain any points in the cut set Tw in its relative interior, since for
all flat points c Tw we can make the shortcut locally around c as in Figure 40.5.
This shows that S r Tw is contractible to w along the shortest paths.

to z

to w
to w

Figure 40.5. An intersection that is Y-shaped cannot locally minimize length in R2 (segment is a shortcut). Cut set interval R in the
neighborhood of a vertex v of P
Lemma 40.7. The cut set Tw is a tree in S which contains all vertices of P .
Proof. By Lemma 40.5, the cut set Tw is a connected union of intervals. Since Dw
is contractible by Lemma 40.6, Tw has no cycles. Therefore, Tw is a tree. To see
that it is a spanning tree, consider two cases. If a vertex v P has two or more
shortest paths, then w Tw . Otherwise, if there is a unique shortest path wv , we
can unfold the faces of P along wv . Consider a bisector R of the angle at v (see
Figure 40.5). Observe that for points z in the neighborhood of v, the points z R
have two shortest paths to w (cf. the cone unfolding argument in Subsection 10.1).
Therefore, vertex v lies in the closure of an interval in the cut set Tw .

Proof of Theorem 40.2. Suppose point w lies in face F of P . Denote by H the plane
spanned by F . Let Dw = S r Tw . For a point z Dw , z 6= w, consider the unique
shortest path wz from w to z. When restricted to F , this path begins as a straight
segment start(wz ) F . Following the construction of source images in the proof
of Lemma 40.5, define a map : Dw R2 , such that [w(z)] is an interval which
begins as start(wz ) and has the same length: |w(z)| = |wz |. Set w = (w) and
denote by U R2 the image of .
Note that is a homeomorphism since no two distinct points have the same image:
(x) 6= (y) for all x 6= y, x, y Dw . Otherwise, we have two points with shortest
paths of the same length |xw |S = |yw |S which start in the same direction from w.
Taking the last point of the initial segment where these paths coincide we obtain
again the Y-shaped intersection as in Figure 40.5.
Now, since Tw is a tree, we conclude that the homeomorphism is piecewise linear.
Similarly, since Dw is contractible, we conclude that U is a polygon. To prove that
is the unfolding map, it remains to show that is isometric (see Remark 40.4). By
definition of , the polygon U is star-shaped at w since for every y wz we have
wy wz . Consider any shortest path xy Dw . Again, by definition of , we have

352

|wx|S = |w(x)|, |wy|S = |w(y)|, and S xwy = (x)w(y), where the S is the
angle between two shortest paths in S. Since the geodesic triangle (wxy) Dw is flat,
it is isometric to (w(x)(y)) U. This implies that |xy|S = |xy | = |(x)(y)|U for
all x, y Dw , i.e., Dw and U are isometric.


Example 40.8. Consider the source unfolding of the cube shown in Figure 40.3. For the
most involved top face A there are 12 potential source images in the plane H spanned
by A. Of which only 8 have points in A corresponding to shortest paths to the source
point w (see Figure 40.6). The Voronoi diagram of these 8 points gives the part of the cut
set Tw in A. To make a distinction, we mark by the remaining four false source images
in the gure. Note that we are explicitly using Mounts lemma (Exercise 40.6) in this case.
*

A
*

w
*

Figure 40.6. Source point v on the bottom face F , 12 source images

for the top face A, and the Voronoi diagram of the source images.
Remark 40.9. (Connellys blooming conjecture) One can think about unfoldings as a
continuous process in the following sense. Take a polyhedral non-overlapping foldout made
of hinged metal, is it always possible to glue its corresponding edges together? Because
metal is rigid, we need not only a non-overlapping property on the foldout as it lies at on
the ground, but also a nonintersecting property as we continuously fold it up. Viewing this
process in reverse, can we continuously unfold the polyhedral boundary so that all dihedral
angles monotonically increase, until the whole polyhedral boundary lies at on a plane? An
example is given in Figure 40.7.

Figure 40.7. Blooming of the source unfolding of a cube.

This problem is called the blooming conjecture, and it can be stated in a number of
dierent (inequivalent) ways. It is believed that such blooming always exists for the source
unfolding, but as of now it is not known if any non-overlapping unfolding of a convex
polyhedron can be always bloomed (see [MilP] for a precise statement).

353

40.3. Folding polyhedra is easier than unfolding. Let S = P be the surface

of a convex polytope P R3 . Suppose S can be subdivided into a finite number of
triangles which can be placed on a plane R2 in such a way that adjacent triangles
in S are also adjacent along the corresponding edges.94 The resulting piecewise linear
continuous map : S R2 is called a flat folding. By construction, the map is
locally isometric everywhere except on some edges of P .
For example, the surface of a cube has a flat folding shown in Figure 40.8. Here we
subdivide the surface of a cube into 24 congruent right triangles which are mapped
onto the same triangle according to the coloring as in the figure. Since triangle edges
with equal colors are still adjacent, this is indeed a flat folding.

Figure 40.8. The surface of a cube can be folded onto a plane.

Theorem 40.10 (Zalgaller). Let S = P be the surface of a convex polytope P R3 .
Then S has a flat folding.
In other words, every convex surface can be folded onto a plane. The proof idea
is in fact more general and works for all 2-dimensional polyhedral surfaces (see Exercise 40.2).
Proof. Let U R2 be the foldout of the source unfolding constructed above. Recall
that U is a star-shaped polygon with two corresponding edges per every interval
in the cut set, i.e., polygon U has edges (a1 , b1 ) and (a2 , b2 ) for every (a, b) Tw .
The triangles (wa1 b1 ) and (wa2 b2 ) are congruent since straight intervals to point
z1 (a1 , b1 ) and the corresponding point z2 (a2 , b2 ) are the images under of the
shortest paths to the same point in the cut set.
Subdivide U into cones over the edges and consider the corresponding subdivision
of S. The surface S is now a union of quadrilaterals [wawb], for all (a, b) Tw . In
every quadrilateral, cut each triangle (wab) with a bisector (wc) as in Figure 40.9.
Place all resulting triangles in the plane so that their vertices w are at the origin and
the edges (w, a) and (w, b) are in the positive part of the x axis. By construction,
the adjacent triangles in each quadrilateral [wawb] remain adjacent. Similarly, the
adjacent triangles in different quadrilaterals can only be adjacent along the edges
(wz), where z is a vertex of U, and all such edges now lie on the x axis. This finishes
the proof.

94This

is called a realization of the surface S in R2 (see Subsection 31.1 and Exercise 39.11).

354

b
c

c
w

a
w

c
w

w
O

x
w

Figure 40.9. Zalgallers flat folding construction.

Example 40.11. For the construction in Figure 40.8, it is not at all obvious that this at
folding can be obtained continuously. Here is an alternative, continuous folding of the cube.
Start with the back side of the cube and push it forward until the back half is glued to the
front half (see Figure 40.10). Do this two more times as in the gure. The resulting surface
is 8-folded and looks like corner of a cube, consisting of three squares. This surface can now
be folded onto the plane in a number of ways, e.g., as in the gure. It is an open problem
whether this procedure is possible for every convex polyhedron (see Exercise 40.14).

Figure 40.10. A step-by-step construction of a continuous folding of

the cube.
40.4. Folding a napkin. Let Q R2 be a unit square which we view as a 2dimensional polyhedral surface with boundary. The perimeter of a flat folding :
Q R2 is the perimeter of the set of points in the plane with one or more preimage.
It was an open problem for decades whether this perimeter is always at most 4. As
it turns out, one can make this perimeter as large as desired.
Theorem 40.12 (Napkin folding problem). The perimeter of a flat folding of a unit
square is unbounded.
In other words, the square napkin can be folded into a figure whose perimeter is as
large as desired. The theorem is only one way to formalize this problem, and there
are several ways to strengthen the folding condition (see Exercise 40.16).
Proof. Subdivide the square into k 2 congruent squares. Subdivide each square into
N = 8m triangles of with equal angle at the center, as in Figure 40.11. Fold triangles
in each square on top of each other, into N layered triangular shaped booklet. Place
all k 2 booklets on top of each other. This is a flat folding since by the symmetry the
adjacent triangles in different booklets are congruent and thus adjacent in the plane.
Observe that the booklets are adjacent along edges outside the circles inscribed into
the squares. Thus the booklets can be further folded to spread them out as in the

355

k2 N

Figure 40.11. Flat folding of a square with large perimeter.

figure. To compute the perimeter of the resulting flat folding, observe that one can
make the triangles as slim as desired. Thus, the overlap of the booklet perimeters can
be made O( k12 ). Since the height of each triangle (or booklet) is at least 21k , and there
are k 2 of them, this gives the total perimeter k O(1). This implies the result. 
40.5. Exercises.
Exercise 40.1. [1-] Let w S be a generic point on the surface of a convex polytope
P R3 . Take the union of cut sets of the source and the Alexandrov unfolding with the
same point w. Show that this subdivides S into n convex polygons, called peels, where n is
the number of vertices in P (see Figure 40.12).

7
2 3

1
8

1
2

7
3

Figure 40.12. Source and Alexandrov unfoldings of the cube, subdivided into peels (cf. Figure 40.12).
Exercise 40.2. (Extended Zalgallers theorem) Let S R3 be a 2-dimensional polyhedral surface. For a nite set of points X = {x1 , . . . , xk } S dene the geodesic Voronoi
diagram VDS (X) to be a union of sets Di of points z S, which are closest to xi and have
a unique shortest path to xi , 1 i k. Dene the cut locus C(X) to be the complement
to VDS (X). Note that for convex surfaces S, the cut locus of a single point is exactly the
cut set in the source unfolding.
a) [1-] Check that the cut locus C(X) is not necessarily 1-dimensional (for non-convex
surfaces).
b) [1] Let V be the set of vertices of S. Prove that the cut locus C(V ) is 1-dimensional.
c) [1] Prove or disprove: sets Di are connected, i.e., the geodesic Voronoi diagram VDS (V )
has exactly k cells.

356

d) [1-] Extend Theorem 40.10 to all abstract 2-dimensional polyhedral surface (dened by
a collection of triangles with given edge lengths and combinatorics).
e) [1+] Extend Theorem 40.10 to spherical polyhedra.
f ) [2-] Extend Theorem 40.10 to convex polytopes in R4 .
Exercise 40.3. Consider the geodesic Voronoi diagram VD of vertices on the surface of
a convex polytope P R3 .
a) [1-] Give an example of a convex polytope P where two polygons in VD have two or
more common edges.
b) [1-] Give an example of a convex polytope P where a polygon in VD has self-adjacent
edges.
Exercise 40.4. a) [1] Prove that for all n 7, the regular tetrahedron cannot be unfolded
into a convex n-gon. Show that this possible for 3 n 6.
b) [1+] For every convex polytope P R3 nd an explicit bound on N = N (P ) such that
for all n N , polytope P cannot be unfolded into a convex n-gon.
Exercise 40.5. a) [1+] Let P R3 be a pyramid with the base a convex polygon Q. Prove
that when other faces are collapsed (rotated around the edges onto the plane spanned by Q),
they cover the whole of Q.
b) [1] Similarly, if the faces are rotated around the edges onto the outside of Q, they do not
intersect, and thus give the edge unfolding of P .
c) [2-] Generalize parts a) and b) to higher dimension and to general polytopes P whose
facets intersect a given facet Q of P by a facet of Q.
Exercise 40.6. (Mounts lemma) a) [1+] In notation of the proof of Lemma 40.7, prove
that if a point z F lies in the Voronoi cell of the source image x X(w, F ), then
|xz| = |xz |S .
b) [2-] Generalize part a) to higher dimensions.
Exercise 40.7. (Number of shortest paths) a) [2-] Let S be the surface of a convex
polytope P R3 with n vertices. Prove that the number of shortest paths between every
pair of points on S is at most polynomial in n.
b) [1-] Show that part a) fails for non-convex polyhedral surfaces.
c) [2-] A combinatorial type of a shortest path is a sequence of faces it enters. Prove that
the number of combinatorial types shortest paths on S is at most polynomial in n.
d) [] Generalize part a) to higher dimensions.
Exercise 40.8. (Discrete geodesic problem) a) [2-] Convert the source unfolding construction into a polynomial time algorithm to solve the discrete geodesic problem: compute
the geodesic distance between two points on the surface of a convex polytope in R3 .
b) [2] Find a polynomial time algorithm to compute the geodesic diameter of a polytope
in R3 .
Exercise 40.9. (Alexandrov unfolding) [2] Prove Theorem 40.3.
Exercise 40.10. [2] Generalize Theorem 40.2 to higher dimensions.
Exercise 40.11. a) [2-] The shortest path problem asks to nd a shortest path between
points x, y R3 which avoids a given set of convex polytopes Q1 , . . . , Qk . Prove that this
problem is NP-hard, if the polytopes are given by their vertices.
b) [1] Conclude from part a) that the discrete geodesic problem on 3-dimensional surfaces
in R4 is also NP-hard.

357

Exercise 40.12. Let S = P be the surface of a convex polytope, and let = xy be a

shortest path between some two points x, y S.
a) [1] For every > 0, give a construction of P , such that for some shortest path we have
for the total curvature: () > 2 .
b) [2] Give a construction of P , such that for some shortest path we have for the total
curvature: () > 2.
c) [2-] Dene the folding angle () to be the sum of exterior angles of edges of P intersected
by . Check that () (). Show that () is unbounded, i.e., for every N > 0 give a
construction of P and , such that () > N .
d) [] Prove or disprove: () is bounded for all P and .
Exercise 40.13. (Volkov ) Let S = P be the surface of a convex cap P R3 (see
Subsection 25.6). Suppose x, y S are two points in the bounded faces of S.
a) [1+] Prove that the length of all geodesics between x and y is bounded.
b) [2-] Conclude that there is at most a nite number of geodesics between x and y.
Exercise 40.14. a) [1-] Formalize the continuous folding introduced in Example 40.11.
Show that there exists a continuous folding which produces the folding in Figure 40.8.
b) [1-] Find a continuous folding of the truncated cube (see Figure 16.4).
c) [1] Find a continuous folding of the regular icosahedron and the regular dodecahedron.
d) [1] Prove that every tetrahedron has a continuous folding.
e) [1+] Show that every convex surface has a continuous folding.
Exercise 40.15. Let S = P be the surface of a convex polytope in R3 . Dene a layered
folding to be a at folding map : S R2 with the ordering map : S N dened so
that adjacent triangles in the layers can be physically glued together.
a) [1-] Formalize the denition. Check that every continuous folding of S gives a layered
folding.
b) [1+] Prove that every S as above has a layered folding.
c) [1+] Generalize b) to abstract 2-dimensional polyhedral surfaces homeomorphic to a
sphere.
Exercise 40.16. (Napkin problem) a) [1] Dene the reection folding of a polygon to
be a at folding obtained sequentially, by reecting the whole polygon along a line (ab) as in
Figure 40.13. Show that the perimeter does not increase under these reections. Conclude
that the perimeter of every reection folding of a napkin is at most 4.
b) [] Dene a natural folding of a polygon to be a at folding obtain sequentially, by
reecting layers along a line (see Figure 40.13). Show that every natural folding of a unit
square has perimeter at most 4.
c) [1+] Find a layered folding (see Example 40.15) of a unit square with perimeter as large
as desired.
d) [2-] Same for continuous foldings (see Example 40.11).
e) [2] Same for continuous piecewise linear foldings.
Exercise 40.17. a) [1-] Find a non-convex polytope P in R3 which does not have a nonoverlapping unfolding.
b) [1+] Same for P with convex faces.
urers conjecture for zonotopes (see Exercise 7.16).
Exercise 40.18. [] Prove D
Exercise 40.19. [] Prove that every star-shaped surface in R3 has a (general) nonoverlapping unfolding.

358

b
a

b
a

Figure 40.13. A reflection folding A and a natural folding B of a square.

Exercise 40.20. [1-] Find a painting by Salvador Dali which exhibits an unfolding of a
4-dimensional cube (among other things).
40.6. Final remarks. D
urers conjecture (Conjecture 40.1) is named in honor of Albrecht
D
urer, who described a number of unfoldings in his classical monograph (c. 1500). In
modern times it is usually attributed to Shephard (1975). The conjecture inspired a lot of
passion and relatively little evidence in either direction (a rare negative result was recently
obtained in [Tar3]). Personally, I would bet against it, while most experts probably would
bet on it. Either way, as it stands, the conjecture remains unassailable.
The non-overlap of the source unfolding (Theorem 40.2) was rst proved in [ShaS], although most supporting results about the cut loci on convex polyhedra were proved much
earlier in [VP2]. Mounts lemma (Exercise 40.6) is due to D. M. Mount (1985, unpublished), and was generalized to higher dimension in [MilP]. It can also be deduced from the
Toponogovs theorem on geodesics in all metric spaces with curvature bounded from below
(see [BBI, 10.3]).95
For non-convex 2-dimensional polyhedral surfaces S R3 the discrete geodesic problem
was resolved in [MMP] by a continuous Dijkstra algorithm, generalizing the classical graph
connectivity algorithm. The problem is NP-hard in higher dimensions (Exercise 40.11). For
convex surfaces an optimal time algorithm was found in [SchrS].
The Alexandrov unfolding was introduced by Alexandrov in [A1, 6.1]. It was proved
to be non-overlapping in [AroO]. The algorithmic consequences were further explored
in [Aga+]. Zalgallers theorem is a special case of a general result outlined in [Zal2] (see
also Exercise 40.2).
The napkin problem goes back to V. I. Arnold (1956) and is often called Arnolds rouble
problem or Marguliss napkin problem. The problem is stated in [Arn3, pp. 2, 158] in
ambiguous language, perhaps intentionally. Variations on the problem were resolved in
[Yas], [Lang, 9.11] and [Tar2], the latter paper giving the strongest version. As Lang
explains in his book, the construction we present in Subsection 40.4 is standard in the
origami literature, going back several centuries.

95Ezra

Miller, personal communication.

359

Part III
Details, details...

360

41. Appendix
41.1. The area of spherical polygons. In the next two subsections we present several
basic denitions and classical results in spherical geometry. A few preliminary words. We
consider a unit sphere S2 with center at the origin O. The role of lines play great circles,
dened as circles of radius 1 with centers at the origin O. Triangles, polygons, areas, etc.
are dened by analogy with the plane geometry. The angle between two lines is dened as
the dihedral angle between planes containing the corresponding great circles. Finally, recall
that area(S2 ) = 4.
Theorem 41.1 (Girards formula). Let T be a spherical triangle with angles , and .
Then area(T ) = + + .
Proof. Let A, B and C be the triangular regions attached to the triangle as in Figure 41.1.
Observe that area(A T ) = 2 area(S2 ) = 2, area(B T ) = 2, and area(C T ) = 2.
Now for the upper hemisphere H we have:


area(H) = area(A) + area(B) + area(C ) + area(T ) = area(A) + area(T )




+ area(A) + area(T ) + area(C) + area(T ) 2 area(T )


= 2 + + area(T )

On the other hand, area(H) = area(S)/2 = 2, which implies the result.

x
C
y

c z

b
a

u
S2

v
S2

Figure 41.1. The area and the law of cosines for spherical triangles.
Theorem 41.2 (The area of a spherical polygon). Let Q be a spherical n-gon with angles
1 , . . . , n . Then area(Q) = 1 + . . . + n (n 2).
We refer to the formula in the theorem as Girards formula for polygons. For the proof,
subdivide the polygon into triangles and sum the areas of all triangles according to Girards
formula. The details are straightforward.
41.2. The law of cosines for spherical triangles. In Section 21 we repeatedly use the
following claim: if the spherical triangles (xyz) and (x y z ) satisfy |xy| = |x y |, |xz| = |x z |,
and yxz > y x z , then |yz| > |y z |. In the plane this follows immediately from the
(rst) law of cosines:
xz.
|yz|2 = |xy|2 + |xz|2 2 |xy| |xz| cos yd

361

For spherical triangles there is a similar formula which also implies the result. We present
a simple proof below.
Proposition 41.3. Let (xyz) S2 be a spherical triangle with edge lengths a = |yz|,
b = |xy|, and c = |xz|. Let = yxz. Then:
cos a = cos b cos c + sin b sin c cos .
Proof. Consider the unit sphere S2 centered at the origin O. Place point x at the North
Pole of the sphere, and let u, v be the intersections of meridians (xy) and (xz) with the

equator (see Figure 41.1). Denote by s = Os, for any point s S2 . We have:
y = cos b x + sin b u , z = cos c x + sin c v .
Now observe that (y , z) = cos a. We conclude:
cos a = (y , z) = (cos b x + sin b u, cos c x + sin c v ) = cos b cos c + sin b sin c cos ,
since (u, v ) = cos and (x , u) = (x , v ) = 0.

41.3. The irrationality of (arccos 13 )/. Let = arccos 13 be the dihedral angle in a
regular tetrahedron (see Sections 20 and 15.1). In this subsection we prove that (/)
/ Q.
More precisely, by induction on n we show that cos(n)
/ Z, for all n N. Now, if = m
n
for some m, n N, then cos(n) = 0, a contradiction.
Let us make an even stronger inductive claim: for every n N we have cos n = r/3n ,
where 3 r. The base of induction is clear: cos = 13 . Now recall that
cos( + ) + cos( ) = 2 cos cos .
Substituting = n and = , we obtain:
2
cos n cos (n 1) .
3
The inductive claim follows immediately from here.
cos (n + 1) =

41.4. The Minkowski inequality. The main result of this subsection in the following
inequality used in Section 7 to prove the BrunnMinkowski inequality (Theorem 7.4). As
the reader shall see this is really a disguised form of the arithmetic mean vs. geometric mean
inequality.
Theorem 41.4 (The Minkowski inequality). For every x1 , . . . , xn , y1 , . . . , yn > 0 we have:
"n
#1/n
" n
#1/n
" n
#1/n
Y
Y
Y
(xi + yi )

xi
+
yi
.
i=1

i=1

i=1

Moreover, the inequality becomes an equality if and only if xi = cyi , for all i [n], and
some c > 0.
Proof. Recall the arithmetic mean vs. geometric mean inequality:
a1 + . . . + an
(a1 an )1/n for all a1 , . . . , an > 0,
n

362

and the inequality becomes an equality if and only if a1 = . . . = an . We have:

1/n

1/n
Qn
Qn
1/n
1/n
n 
n 
Y
Y
+
xi
yi
i=1 xi
i=1 yi
Qn
=
+
1/n
xi + yi
xi + yi
i=1 (xi + yi )
i=1
i=1

n
n
1 X xi
1 X yi
+
= 1,
n
xi + yi
n
xi + yi
i=1

i=1

and the inequality becomes an equality if and only if there exists c > 0, such that xi /yi = c,
for all i [n].

41.5. The equality part in the BrunnMinkowski inequality. Let us start by noting
that one can view the second (equality) part in Theorem 7.4 as a uniqueness result: for
every A there exists a unique B of given volume such that the BrunnMinkowski inequality
becomes an equality. Of course, the claim does not follow from the convergence argument
in Subsection 7.7: just because the equality holds for a pair of convex sets (A, B) does not
imply that it must hold for pairs of brick regions (An , Bn ). Nevertheless, one can still use
the brick-by-brick approach to prove the claim.
Proof of the equality part. Denote the unit vectors in the direction of axis coordinates of Rd
by e 1 , . . . , e d . Let C Rd be a convex set. Divide A into two parts C0 , C1 of equal volume
by a hyperplane orthogonal to e 1 . Then divide each part into two parts by (separate)
hyperplanes orthogonal to e 2 , to obtain four parts C00 , C01 , C10 , and C11 of equal volume.
Continue cutting each part into two, cyclically changing the normals e i of the hyperplanes.
After n = k d iterations (k rounds of all d directions) we obtain 2n convex regions Ci ,
where 0 i < 2n and corresponds to the binary expression of i.
We say that Ci is a boundary region if it contains points of the surface: Ci C 6= .
Clearly, non-boundary regions are bricks. Denote by Rn C the union of all non-boundary
regions (bricks) Ci after n iterations. The process is illustrated in Figure 41.2.

C1
C0

C01 C11
C00 C10

Figure 41.2. Division of the region C into 16 equal parts, 12 boundary

and 4 non-boundary (shaded).
For a convex region X Rd denote by W (X) = [w1 (X) . . . wd (X)] the smallest brick
containing X. Here wi (X) denotes the width of X in direction e i (see Exercise 3.6).
Now let A and B be any convex sets in Rd for which the BrunnMinkowski inequality
becomes an equality. We can always assume that vol(A) = vol(B), or take an appropriate
expansion of B otherwise. Rather than apply the above process to A and B, we alter
the process and make it dependent of both sets. First, divide A into two parts A0 , A1 a
hyperplane orthogonal to e 1 , such that w1 (A0 ) = w1 (A1 ) = w1 (A)/2. Similarly, divide B
into two parts B0 , B1 by another hyperplane orthogonal to e 1 , such that vol(B0 ) = vol(A0 )
and vol(B1 ) = vol(A1 ). Then, divide parts A0 , A1 by hyperplanes orthogonal to e 2 to four
parts A00 , A01 , A10 , and A11 such that w1 (A00 ) = w1 (A01 ) = w1 (A0 )/2 and w1 (A10 ) =

363

w1 (A11 ) = w1 (A1 )/2. Then divide sets B0 , B1 by hyperplanes into four parts B00 , B01 , B10 ,
and B11 , such that the corresponding sets in A and B have equal volume. Then repeat this
iteration of cuts for e 3 ,. . . ,e d , making a total of d iterations of cuts. Then switch the roles
of A and B and in the next d cuts halve the widths of regions in B. This make the total
of 2d cut, which we call a round. Now make k rounds of such cuts. In the notation above,
we obtain regions Ai and Bi , 0 i < 2n , where n = 2k d.
Denote by Xn and Yn the unions of all non-boundary regions Ai and Bi , respectively.
Clearly, there may be regions in Xn that do not have corresponding regions in Yn and vice
versa. Let Xn Xn be a union of non-boundary regions Ai such that the region Bi is also
non-boundary. Dene Yn Yn analogously. A priori Xn and Yn can be disconnected, so
let Xn , Yn be their connected components containing centers of mass cm(A) and cm(B),
respectively.
Observe that after k rounds of cuts as above we have wr (Ai ) wr (A)/2k diam(A)/2k .
Letting D = max{diam(A), diam(B)} we conclude that wr (Ai ), wr (Bi ) , where =
(D/2k ) > 0. For large enough k = n/2d we can ensure that neither center of mass lies in a
boundary region. Further,
vol(Xn ) vol(A) [ area(A)],

and thus vol(Xn ), vol(Yn ) vol(A) = vol(B) as n .

Let us further restrict our regions: denote by Xn and Yn the union of bricks which
have no points at distance from the boundary, where > 0 is as above. From the
reasoning as above, for large enough n we obtained two connected brick regions Xn , Yn
with corresponding bricks Ai A, Bi B, and such that Xn A, Yn B, as n .
Now comes the key observation. In the inductive step in the proof of the Brunn
Minkowski inequality in Section 7, we divided the regions A, B into two disjoint parts by a
hyperplane in the same ratio, and concluded that the inequality holds for sets A, B only if it
holds for the corresponding parts. Clearly, the equality also holds only if the equality holds
for the corresponding parts. Repeating the procedure we get again the equality for the four
corresponding pairs of smaller parts, etc. By induction, we conclude that the equality in
the BrunnMinkowski inequality holds for all pairs (Ai , Bi ), 0 i < 2n .
Now, if both Ai and Bi are bricks the equality can hold only if one is an expansion of
the other. Since vol(Ai ) = vol(Bi ) by construction, bricks Ai , Bi must be equal up to a
translation. By itself, this does not imply that Xn and Yn are equal. It does, however,
imply that they are converging to the same convex set, by the following argument.
We say that two bricks Ai and Ai touch each other if their boundaries intersect in points
interior to the faces. Observe that the corresponding bricks Bi and Bi do not necessarily
have to touch, but they must both touch a hyperplane separating them, a hyperplane
corresponding to a hyperplane separating Ai and Ai .
Denote by Aj and Bj , 0 j < 2d , the regions resulted after the rst round of cuts.
Consider Aj = Aj Xn and Bj = Bj Yn . Both brick regions contain a large number
of smaller regions, but none of them boundary regions. Furthermore, since point of the
boundary region can lie in an interval parallel to e r between two points in Aj or in Bj .
Now, we claim that Aj Bj for all j. In other words, these regions can be obtained
from each other by a translation. Indeed, start at a corner which must contain equal
corresponding bricks. Start adding bricks along the straight interior edges of Aj , and all of
them will sequentially touch each other. Thus, the interior edges of Bj will be lined up with
equal corresponding bricks. Now start lling all 2-dimensional interior faces in Aj and Bj ,
again by the equal corresponding bricks, etc. Eventually we obtain two equal llings of the

364

regions, implying that the regions themselves are equal. Note that we are implicitly using
the fact that by convexity of A, B and the construction, there are no holes in Aj , Bj .
While the corresponding regions Aj , Bj are equal, we can no longer use the same argument
to prove that Xn = Yn . Instead, let us put them together in the reverse order they were
obtained by separation with hyperplanes. When put together, regions Aj and Bj can be
shifted along the axes. Observe that by convexity of A and B, they cannot be shifted by
more than (2) in each direction.
Thus in fact the distance between the corresponding

points in Xn , Yn is at most d (2) 0 as n .

Make a translation of Yn so that it has the came center of mass as Xn . From above,
we have vol(Yn r Xn ) 0. Since Xn A and Yn B, this implies that A B, and
completes the proof.


41.6. The CayleyMenger determinant. Here we give an explicit computation of the

CayleyMenger determinant dened and studied in Section 34. This is an important ingredient in the proof of the bellows conjecture (Theorem 31.2).
Theorem 34.5 (CayleyMenger).

vol2 () =

(1)d1
2d d!2

simplex = (v0 v1 . . . vd ) Rd , we have:

1
1
1 ... 1
0 201 202 . . . 20d

201 0 212 . . . 21d

,
202 212 0 . . . 22d

..
..
.. . .
.
. ..
.
.
.
1 20d 21d 22d . . . 0

For every

0
1

det 1

..
.

where ij = |vi vj |, for all 0 i < j d.

Proof of Theorem 34.5. Suppose vi = (xi1 , . . . , xid ) Rd , for all 0 i d.

ces A and B as follows:

x01 x02 . . . x0d 1

x01 x02 . . . x0d
x11 x12 . . . x1d 1
x11 x12 . . . x1d

A = ..
B = ..
..
..
.. ,
..
..
..
..
.

.
.
.
.
.
.
.
.
xd1 xd2 . . . xdd 1
xd1 xd2 . . . xdd

Dene matri0
0
..
.
0

Clearly, det(B) = 0, and

det(A) = det

Note that

x01
x02
...
x0d
1
x11 x01 x12 x02 . . . x1d x0d 0

d
..
..
..
.. = (1) d! vol().
..

.
.
.
.
.
xd1 x01 xd2 x02 . . . xdd x0d 0

A AT = hvi , vj i + 1 0i,jd ,

and

B B T = hvi , vj i 0i,jd .

365

From here we have:

det(A A ) = det

hv0 , v0 i + 1 hv0 , v1 i + 1
hv1 , v0 i + 1 hv1 , v1 i + 1
..
..
.
.
hvd , v0 i + 1 hvd , v1 i + 1
1
1

hv0 , v0 i hv0 , v1 i
hv1 , v0 i hv1 , v1 i

..
..
= det
.
.

hvd , v0 i hvd , v1 i
1
1

...
...
..
.
...
...

. . . hv0 , vd i + 1
. . . hv1 , vd i + 1
..
..
.
.
. . . hvd , vd i + 1
...
1

hv0 , vd i 1
hv1 , vd i 1

..
.. .
.
.

hvd , vd i 1
1
1

0
0
..
.

0
1

Since det hvi , vj i = det(B B T ) = det2 (B) = 0, in the determinant above we can replace
the bottom right entry 1 with 0. Denote the resulting matrix by M .
Now observe that 2ij = hvi , vi i 2hvi , vj i + hvj , vj i. Denote by C the matrix as in the
theorem. Using row and column operations we obtain:

2hv0 , v0 i 2hv0 , v1 i . . . 2hv0 , vd i 1

2hv1 , v0 i 2hv1 , v1 i . . . 2hv1 , vd i 1

..
..
..
.. = (2)d det(M ).
..
det(C) = det
.
.
.
.
.

2hvd , v0 i 2hvd , v1 i . . . 2hvd , vd i 1

1
1
...
1
0


2
Since det(M ) = det(A AT ) = d! vol() , we conclude that
det(C) = (1)d1 2d d!2 vol2 (),

as desired.

41.7. The theory of places. The proof of the bellows conjecture (Theorem 31.2) is based
on the integrality criteria (Theorem 34.2), a classical result in the theory of places. In this
section we present a reworking of the presentation in [Lan], making it self-contained and,
hopefully, more accessible to the reader unfamiliar with the eld.
Let L be a eld containing ring R. We will always assume that R contains 1. To prove
the integrality criteria we need to show:
A. if x L is integral over R, then every place that is nite over R, is also nite
over x;
B. if x L is not integral over R, then there exists a place that is nite over R,
and innite over x.
While the proof of the bellows conjecture uses only part B, we include a simple proof of
part A for the sake of clarity and completeness.
Proof of part A. Since x L is integral over R, by denition we have
xn + an1 xn1 + . . . + a1 x + a0 = 0 ,

where ai R. Dividing both sides by xn we obtain:

an1
a1
a0
1+
+ . . . + n1 + n = 0 .
x
x
x

366

Recall that Fb = F {} and let : L Fb be a place which is nite on R. Suppose

that (x) = . Then ( x1 ) = 0. Applying to both sides of the above equation, we obtain
1 = 0, a contradiction.

Proof of part B. Since x is not integral over R, we have x 6= 0. Let us construct a place
: L Fb, for some (possibly very large) eld F , such that ( x1 ) = 0. This would imply
that (x) = and prove the claim.
Denote by Rh x1 i the ring generated by R and x1 . Consider the ideal I = x1 Rh x1 i generated
by x1 . Note that I 6= Rh x1 i. Indeed, otherwise 1 x1 Rh x1 i and we can write
an
a1
a2
1 =
+ 2 + ... + n ,
x
x
x
for some ai R. Multiplying both sides by xn gives a monic polynomial equation for x,
implying that x is integral over R, a contradiction.
From above, there exists a maximal ideal m Rh x1 i containing I. Since m is a maximal
ideal, we can consider a eld F1 = Rh x1 i/m. This denes a natural map : Rh x1 i F1 ,
such that ( x1 ) = 0. Denote by F the algebraic closure of F1 . Let us extend the map
from F1 to F , still calling it . We have : Rh x1 i F . 96
Now, consider a set of ring homomorphisms S : S F such that Rh x1 i S
L,
1
and S (y) = (y) for all y Rh x i. There is a natural partial order S < S on these
homomorphisms, where S S and S is an extension of S .
Consider an increasing sequence of ring homomorphisms 1 < 2 < . . . < i < . . .,
corresponding to an increasing sequence of rings Rh x1 i S1 S2 . . . Si . . ., where
Si
L. We can construct a ring S = i Si and a homomorphism : S F , where
i < for all i = 1, 2, . . . Clearly, S
L, since if S = L then x Si for some i, and
1 = (1) = (x x1 ) 6= (x) ( x1 ) = 0. Therefore, homomorphism is an upper bound
on i . Clearly, the same argument works for all innite chains of homomorphisms (not
necessarily countable), and by Zorns lemma the partial order dened above has a maximal
element, i.e., a homomorphism which cannot be extended.
The following lemma characterizing the ring of the maximal homomorphism is the key
result which allows us to nish the construction. We postpone the proof of the lemma until
the end of the proof of the theorem.
Lemma 41.5. Let : S F be a maximal homomorphism, where < and Rh x1 i
S L. Then for every y L either y S or y1 S.
b
b
We can now dene b : L Fb by letting (y)
= (s) for all y S, and (y)
= for all
y
/ S. Since S L is a subring, it is easy to see (by checking all relations in the denition
of a place) that
 
1
b
= 0 for all y
/S
()

y
implies that b is a place. Therefore, to prove part B it suce to establish ().
By the lemma, for every y
/ S we have y1 S. Now suppose that ( y1 ) 6= 0. Take a
bigger ring S = Shyi generated by S and y, with a homomorphism : S F dened
by (y) = 1/( y1 ) and (s) = (s), for all s S. To see that the map is indeed a
P
P
ring homomorphism, simply note that ni=0 si y i = 0 is equivalent to ni=0 si ( y1 )ni = 0,
96The

importance of taking the algebraic closure F F1 will come only at the end of the proof
of Lemma 41.5. Until then, the arguments work for any extension of F1 .

367

for all si S. Since y S , we conclude that < , a contradiction with being

maximal. Therefore, ( y1 ) = 0 and we proved (). This completes the proof of part B
modulo Lemma 41.5.

Proof of Lemma 41.5. Suppose y, y1
/ S for some y L. Dene S1 = Shyi and S2 = Sh y1 i.
We obtain a contradiction with maximality of : S F by showing that can be extended
to at least one of the rings S1 or S2 .
Let m = Ker( : S F ) S be the ideal of elements mapped into 0 F . Let us start
with the following two technical lemmas whose proofs we postpone.
Sublemma 41.6. The ideal m S is maximal. In addition, for every y
/ m, we have
1
97
y S.
Sublemma 41.7. Either mS1 6= S1 or mS2 6= S2 .
By Sublemma 41.6, the ideal m S is maximal and S/m is a eld. Thus, we obtain
a natural eld homomorphism : S/m F . By Sublemma 41.7, we can assume that
mS1 S1 .
Since mS1 is an ideal in S1 , we can consider a maximal ideal m1 m in S1 . Then S1 /m1
is a eld extension of S/m, obtained by adding = y, dened as the class of y. Note here
that this implies that is algebraic over S/m, since otherwise S1 /m1 = S/m[] is not a
eld.
Now, to prove the lemma it suces to construct an extension 1 : S1 /m1 F of ,
since this would give a ring homomorphism 1 : S1 F by letting 1 : m1 0. This
would imply that < 1 , i.e., : S F is not maximal, a contradiction. Clearly, we need
only to dene 1 () as both homomorphisms must coincide on S/m. Let 1 () be a root
of (f ) over F . This is possible since F F1 is algebraically closed. Thus, we obtain the
desired extension 1 . This completes the proof of the lemma.

Proof of Sublemma 41.6. Let : S F be the maximal homomorphism as in the lemma,
let : S/m F be the corresponding homomorphism modulo ideal m, and let : S S/m
be a natural projection.
Let y S, y
/ m, be an element of the ring, and let = (y) S/m be its projection.
From above, 6= 0. By the denition of m = Ker() we conclude that is an injection.
Since 6= 0, we conclude that () 6= 0 and (y) 6= 0.
1
Now take S = Sh y1 i. Let (s) = (s) for all s S, and let ( y1 ) = (y)
. It is easy

to see that uniquely extends to a homomorphism : S F . Since < , from

the maximality of we conclude that S = S , and y1 S. This proves second part of the
sublemma.
For the rst part, note that there is an element 1 = ( y1 ). Therefore, every nonzero
element in S/m is invertible, and S/m is a eld. This implies that m is maximal.

Proof of Sublemma 41.7. From the contrary, suppose mS1 = S1 and mS2 = S2 . Then
there exist polynomials
 j
n
k
X
X
1
i
()
1 =
ai y
and 1 =
bj
,
y
i=0

97This

j=0

says that S is a local ring. To make the presentation self-contained we will refrain from
using this notion.

368

for some ai , bj m, where the degrees k, n are chosen to be the smallest possible. By
symmetry, we can assume that k n. Rewriting the second polynomial, we have:
yk =

k
X

bj y kj .

Equivalently,

j=0

(1 b0 ) y k =

k
X

bj y kj .

j=1

Now observe that (1b0 ) is invertible. Indeed, since 1

/ m and b0 m we have (1b0 )
/ m,
and the second part of Sublemma 41.6 implies that (1 b0 )1 S. This gives:
k
1 X
y =
bj y kj .
1 b0
k

Equivalently,

j=1

k
1 X
y =
bj y nj .
1 b0
n

j=1

This allows us to decrease the degree of the rst polynomial in ():

1 =

n1
X

ai y i + an y n =

i=0

a contradiction with the minimality of n.

n1
X
i=0

ai y i +

n1
X
an
bni y i ,
1 b0
i=nk

41.8. The mapping lemma. In Sections 35 and 37 we use the following standard result.
Theorem 41.8 (The mapping lemma). Let A and B be two manifolds of the same dimension. Suppose a map : A B satises the following conditions:
1) every connected component of B intersects the image (A),
2) map is injective, i.e., (a1 ) = (a2 ) implies that a1 = a2 ,
3) map is continuous,
4) map is proper.
Then is a homeomorphism; in particular, is bijective.
Here by a proper map we mean that for every sequence of points {ai A} and images
{bi = (ai )}, if bi b B as i , then there exists a A, such that b = (a) B, and
a is a limit point of {ai A}.
To get better acquainted with the lemma, consider the following examples which show
that the conditions in the lemma are necessary.
Example 41.9. Suppose A = S 1 and B = S 2 , i.e., we have two manifolds of dierent
dimensions. Then any embedding : S 1 S 2 satises conditions 1) to 4), but is
obviously not bijective.
Similarly, if A = S 1 and B = S 1 S 1 , i.e. B is a union of two disjoint circles. Then the
map sending A into the rst circle in B satises conditions 2) to 4), but is not a bijection.
If in this example one takes B to be two circles attached at a point, then all conditions 1)
to 4) are satised. Clearly, is not bijective, which may be puzzling at rst, until one
realizes that B is no longer a manifold.
Further, if now A = S 1 S 1 , B = S 1 and sends both circles into one, then all conditions
are satised, except for condition 2). Finally, suppose both A and B are open unit disks
and let be a map contracting A into a disk of smaller radius. Then all conditions are
satised except for the condition 4), and is obviously not bijective.

369

Proof of Theorem 41.8. By continuity of , dierent connected components of A are mapped

into dierent connected components of B. Thus, it suces to prove the result when B is
connected.
Let B = {(a), a A} B be the image of the map . By injectivity of , the
map 1 : B A is well dened. Furthermore, condition 4) implies that 1 is also
continuous. Therefore, : A B is a homeomorphism, and B is open in B. On the other
hand, condition 4) implies that B is also closed in B. Since B is connected, this implies

that B = B and completes the proof.
41.9. Final remarks. The results in spherical geometry can be found in virtually all classical geometry textbook (see [Hada]). The presentation of Girards formula and the law of
cosines follows [Ber1, 18.3, 18.6]. Our proof of the Minkowski inequality (Theorem 41.4)
is standard. See [BecB] for a dierent proof, and numerous proofs of the arithmetic mean
vs. geometric mean inequality. For a similar computation of the CayleyMenger determinant, and several related results we refer to [Ber1, 9.7], while further references can be
found in [GriK, 3.6]. For the original expanded presentation of the mapping lemma (Theorem 41.8) see [A2, 2.2]. See also [Ale6] for a background and a survey of applications to
convex polyhedra.

370

42. Additional problems and exercises

42.1. Problems on polygons and polyhedra.
Exercise 42.1. [1-] A quadrilateral in R3 is tangent to a sphere. Prove that the tangency
points are coplanar.
Exercise 42.2. Let Q = [x0 , x1 , . . . , xn ], xi R3 be a 3-dimensional n-gon, where x0 = xn .
We say that Q is regular if |xi1 , xi | = 1 and xi1 xi xi+1 = , for all 1 i n and
some xed 0 < < . Space polygon Q is called non-degenerate if no four points of Q are
coplanar.
a) [1] Prove that every regular pentagon lies on a plane.
b) [1] Prove that for all n 6 there exist non-planar regular n-gons.
c) [1+] Prove that for n large enough there exist non-degenerate regular n-gon.
Exercise 42.3. (Sylvesters problem) [1+] Let A be a convex set and let x1 , x2 , x3 , x4 be
chosen at random from A. Denote by p(A) the probability that four points form a convex
quadrilateral. Compute p(A) for a circle, a square, and an equilateral triangle.
Exercise 42.4. [2] Let Q be a piecewise linear curve in the plane. Prove that one can
add straight shortcuts of total length at most c1 |Q| such that the distance between every
two points x, y along Q is at most c2 times the Euclidean distance |xy|, for some universal
constants c1 , c2 .
Exercise 42.5. [1-] Prove that in every simplicial polytope P R3 there exists an edge
such that all face angles adjacent to it are acute.
Exercise 42.6. [1-] Let P R3 be a convex polytope with faces Fi , 1 i n, and let
Xi Fi be convex polygons inside the faces. Prove that for every > 0 there exists a
convex polytope Q P such that all Xi are faces of Q and vol(Q) > (1 )vol(P ).
Exercise 42.7. [1] Let R3 be a tetrahedron and let x, y be the midpoints of opposite
edges. Prove that every plane passing through x, y divides into two polytopes of equal
volume.
Exercise 42.8. [1] Let P Rd+ be a convex polytope lying in the positive orthant. Prove
that for every interior point a = (a1 , . . . , ad ) P there exists a vertex v = (v1 , . . . , vd ), such
that vi d ai for all 1 i d.
Exercise 42.9. [1] Find the maximal number of sides in a polygon obtained as the intersection of a unit hypercube and a 2-dimensional plane.
Exercise 42.10. [1+] Prove that for every tetrahedron R3 there
exist two planes
L1 , L2 , such that the areas A1 , A2 of projections of satisfy A1 /A2 2. Check that this
bound is optimal.
Exercise 42.11. [1+] Let P R3 be a convex polytope with n facets, and let Q be its
projection on a plane. What is the maximal possible number of edges the polygon Q can
have?
Exercise 42.12. [1-] Let D R3 be a unit cube centered at the origin O. Suppose C is
a cone with O as a vertex and all face angles equal to /2. Prove or disprove: at least one
vertex of D lies in C.

371

Exercise 42.13. a) [2] Suppose P R3 is a convex body, such that all projections of P
are convex polygons. Prove that P is a convex polytope.
b) [2-] Suppose further, that all projections are polygons with at most n sides. Then P has
nO(n) facets.
c) [1] Give an explicit construction of polytopes P with e(n) facets.
Exercise 42.14. a) [2] Suppose P R3 is a convex body, such that all projections of P
are convex polygons. Prove that P is a convex polytope.
b) [2-] Suppose further, that all projections are polygons with at most n sides. Then P has
nO(n) facets.
Exercise 42.15. a) [1] Let P R3 be a convex body with countably many extremal
points. Prove or disprove: the set of extremal points is closed, i.e., the limit of every
converging sequence of extremal points is also extremal.
b) [2] Let P R3 be a convex body such that every projection of P on a plane is a convex
polygon. Prove that P is a convex polytope.
Exercise 42.16. (Pohlkes theorem) [1+] Prove that every convex quadrilateral Q R2 is
an oblique projection of a regular tetrahedron.
Exercise 42.17. Let Q be a 3-dimensional unit cube.

a) [1] Denote by X a projection of Q onto a plane. Prove that 1 area(X) 3.

b) [1+] Denote by X the projection of Q onto a plane H, and by the length of a projection
of Q onto a line orthogonal to H. Prove that = area(X).
c) [1+] Denote by Y a cross section of Q, i.e., the intersection of Q with a plane. Prove
that 1 area(Y ) 2.
Exercise 42.18. a) [1] Let P R3 be a convex polytope, and let be a tetrahedron with
vertices at vertices of P of largest volume. Prove that for every plane H, an orthogonal
projection of onto H has area at least 1/9 of the area of the projection of P onto H.
b) [1+] Improve the above bound to 1/7.
Exercise 42.19. Let P, P R3 be two convex polytopes. We say that P passes through P
if there exist projections Q and Q (of P and P ) such that Q ts inside Q . When P = P
we say that P passes through itself.98
a) [1] Prove that every tetrahedron passes through itself.
b) [1-] Prove that the cube passes through itself.
c) [1] Prove that the regular octahedron passes through itself.
d) [] Prove or disprove: every convex polytope passes through itself.
Exercise 42.20. (Shadows of polytopes) a) [2] Let X L be a polygon in the plane
L R3 . Suppose Y H is an orthogonal projection of X onto another plane H. Prove
that there exists a rigid motion such that (Y ) X. In other words, the shadow of any
polygon is smaller than the polygon itself.
b) [1-] Show that part a) does not generalize to 3-dimensional convex polytopes in a hyperplane in R4 .
c) [1+] Prove that when d 3 no d-dimensional polytope P can cover all its shadows. For
example, when d = 4, every convex polytope satises b).
98The

idea is that in a solid Q one can make a hole so that P can pass through that hole.

372

Exercise 42.21. a) [1-] Let Cd Rd be a unit cube, d 3. Prove that there exists a
hyperplane which intersects every facet of Cd .
b) [1+] Prove that for every convex polytope P R2 and two vertices v, w of P there exists
a plane L containing v, w, and such that at least three faces of P are not intersected by L.
c) [1+] Prove that for every simplicial convex polytope P R3 and two vertices v, w of P
there exists a hyperplane H containing v, w, and such that at least three facets of P are
not intersected by H.
d) [2-] Find a polytope P R4 and two vertices v, w, such that every hyperplane H
containing v and w intersects all but at most two facets.
Exercise 42.22. [1+] Let P R4 be a product of two triangles:

P = {(x1 , x2 , y1 , y2 ) | (x1 , x2 ) 1 , (y1 , y2 ) 2 }.

Prove that there are no 9-gon projections of P on a 2-dimensional plane.

Exercise 42.23. [2-] Let P R3 be a convex polytope with the property that for every
face F of P there is a parallel face F and a parallel plane L, such that L contains all vertices
of P that are not in F or F . Prove that K has at most 14 faces. Show that this bound is
tight.
Exercise 42.24. [1+] Find a polyhedral embedding of a torus which has 7 (non-convex)
faces, such that every two faces have a common edge.
Exercise 42.25. [1] Find a polyhedral embedding of a torus which has 7 vertices, 14 triangular faces, and such that every two vertices are connected by an edge.
42.2. Volume, area and length problems.
Exercise 42.26. [1-] Let R3 be a tetrahedron. Prove that has a vertex v, such
that the lengths of three edges of adjacent to v satisfy the triangle inequality.
Exercise 42.27. [1] For a tetrahedron R3 , denote by h() the minimal height in ,
and by w() the minimal distance between the opposite edges in . Prove that 23 h()
w() h().
Exercise 42.28. (Sarrons formula) [1] Let X = [x1 x2 . . . xn ] R2 be a simple polygon
and let 1 , . . . , n be the angles dened as in Figure 42.1. Denote by i = |xi xi+1 | the edge
lengths of X. Prove:
1 X
area(X) =
i j sin(j i ) .
2
i<j

Exercise 42.29. [1] Suppose points x1 , . . . , xn lie on a unit sphere S2 . Prove that
X
|xi xj |2 n2 .
1i<jn

Exercise 42.30. (Law of sines in S2 ) [1] Let (ABC) S2 be a triangle on a unit sphere
with edge lengths |BC|S2 = a, |AC|S2 = b, and |AB|S2 = c. Suppose also = A, = B
and = C be the spherical angles in the triangle. Then
sin(a)
sin(b)
sin(c)
=
=
.
sin()
sin()
sin()

373

2
1

x2
3

x1
x3
n
xn

Figure 42.1. Polygon X = [x1 x2 . . . xn ] and angles i in Sarrons formula.

Exercise 42.31. (Law of sines in R3 ) [1-] Let a = OA, b = OB and c = OC be three

non-collinear vectors in R3 . Denote by K the cone spanned by these vectors at the origin O.
Dene the 3-dimensional sine function as
3

[a, b, c]
,
[a, b] [a, c] [b, c]

sin(O, ABC) =

where [a, b, c] is the volume of a parallelepiped spanned by the vectors; similarly, [a, b] is
the area of a parallelogram spanned by a, b. Consider a tetrahedron (OABC) R3 . Prove
the following law of sines:
area(ABC)
3 sin(O, ABC)

area(ABO)
3 sin(C, ABO)

area(ACO)
3 sin(B, OAC)

area(BCO)
3 sin(A, OBC)

Exercise 42.32. (Polar sine in R3 ) [1] Let a, b and c be three non-collinear vectors in R3 .
Dene the polar sine function as
psin(a, b, c) =

[a, b, c]
.
kak kbk kck

For every nonzero vector u in R3 , prove the following inequality:

psin(a, b, c) psin(u , b, c) + psin(a, u, c) + psin(a, b, u) .

Exercise 42.33. (The law of cosines in R3 ) [1+] Let P R3 be a convex polyhedron with
n + 1 faces F, G1 , . . . , Gn . Denote B = area(F ), and Ai = area(Gi ), 1 i n. Finally, let
ij be an angle between between planes spanned by Gi and Gj . Prove that
B2 =

n
X
i=1

A2i

Ai Aj cos ij .

i6=j

Generalize the result to non-convex polyhedra and to higher dimensions.

Exercise 42.34. [1-] Let x1 , x2 , x3 S2 be points on a unit sphere in R3 centered at the
origin O, and let = (O, x1 , x2 , x3 ) be the simplex spanned by these points. Denote by ij
the angles between (O, xi ) and (O, xj ). Prove that

1
cos 12 cos 13
1
1
cos 23 .
vol2 () =
det cos 12
36
cos
cos
1
13

23

374

Exercise 42.35. [1] Fix a vertex v of a given tetrahedron . Let a, b, c be the lengths of
edges adjacent to v, and let , , be the face angles between these edges. Prove that
abc p
++
vol() =
( )( )( ) , where =
.
3
2

Exercise 42.36. (Kahans formula) [1+] Let = (x1 x2 x3 x4 ) be a tetrahedron with edge
lengths
|x1 x2 | = a, |x1 x3 | = b, |x2 x3 | = c,
Dene

|x3 x4 | = p, |x2 x4 | = q , |x1 x4 | = r .

u = (p + q c)(p + q + c), v = (p + r b)(p + r + b), w = (q + r a)(q + r + a),

s = (p q + c)(q p + c), t = (p r + b)(r p + b), z = (q r + a)(r q + a),

and = uvz , = uwt, = vws, = stz .

Prove that
vol() =

p
1
( + + )( + + )( + + )( + + ).
192pqr

Exercise 42.37. [1] For every six points x1 , x2 , x3 , y1 , y2 , y3 R3 dene the CayleyMenger
bideterminant CM() as follows:

0
1
1
1
1 |x1 y1 |2 |x1 y2 |2 |x1 y3 |2

CM(x1 , x2 , x3 | y1 , y2 , y3 ) = det
1 |x2 y1 |2 |x2 y2 |2 |x2 y3 |2 .
1 |x3 y1 |2 |x3 y2 |2 |x3 y3 |2

Prove that the dihedral angle between the planes X = (x1 x2 x3 ) and Y = (y1 y2 y3 ) satises:
cos2 =

CM 2 (x1 , x2 , x3 |y1 , y2 , y3 )
.
CM (x1 , x2 , x3 ) CM (x1 , x2 , x3 )

Exercise 42.38. Let P R2 be a convex polygon containing the origin O in its relative
interior. Denote by fP (r) = area(P Br ), where Br is the circle centered at O of radius
r > 0.
a) [1] Suppose fT = fT for two triangles T, T R2 . Prove that T T .
b) [1-] Prove or disprove: if fP = fP , then P P .
c) [1+] What happens in higher dimensions?
Exercise 42.39. (Eulers inequality) a) [1-] Let R and r be the circumradius and the
inradius of a triangle. Prove that R 2r.
b) [1] Let R and r be the radii of circumscribed and inscribed sphere for a simplex Rd .
Prove that R dr.
c) [1+] Show that the volume of a simplex inscribed into a unit sphere in Rd , is maximal
for a regular simplex.
d) [1+] Show that the volume of a simplex circumscribed into a unit sphere in Rd , is minimal
for a regular simplex.
e) [1-] Deduce part b) from c) and d).

375

42.3. Miscellaneous problems.

Exercise 42.40. a) [1] Let Q R3 be an unbounded convex polyhedron such that every
projection of Q onto a plane is a cone. Prove that Q is a 3-dimensional cone.
b) [1] Same result for Q Rd and projections on all 2-dimensional planes.

Exercise 42.41. a) [1-] Let Q and Q be two convex polygons in the XY and XZ planes,
respectively such that projections of Q and Q on the X line are the same. Prove that there
exists a convex polytope P with projections Q and Q .
b) [1] Generalize this to simple (not necessarily convex) polygons.

Exercise 42.42. a) [1-] Prove that every tiling of a d-dimensional cube by smaller cubes
contains at least 2d cubes.
b) [1] Prove that a 3-dimensional cube cannot be tiled by a nite number of cubes of distinct
size.
c) [2-] Prove that a square can be tiled by a nite number of squares of distinct size.
Exercise 42.43. [1+] A trapezoid is called isosceles if its non-parallel sides have equal
lengths. Prove that every convex polygon has a dissection into isosceles trapezoids.
Exercise 42.44. a) [1] Prove that an equilateral triangle cannot be tiled by a nite number
of equilateral triangles of dierent sizes.
b) [1+] Prove that no convex polygon can be tiled by a nite number of equilateral triangles
of dierent sizes.
c) [2-] Prove that the plane cannot be tiled with distinct equilateral triangles whose side
length are at least 1.
d) [1] Prove that an isosceles right triangle can be tiled by a nite number of isosceles right
triangles of dierent sizes.
e) [1+] Prove that every non-equilateral triangle T can be tiled by a nite number of triangles
similar to T and of dierent sizes.
f ) [2-] Prove that a square can be tiled by a nite number of isosceles right triangles of
dierent sizes.
Exercise 42.45. Two convex polytopes in R3 are called adjacent if they touch by a face.
a) [1-] Find eight pairwise adjacent tetrahedra.
b) [1+] Prove that there are no 100 pairwise adjacent tetrahedra.
c) [1+] For every n, nd a family of n pairwise adjacent convex polytopes.
d) [2] For every n, nd a family of n pairwise adjacent congruent convex polytopes.
Exercise 42.46. Let Q be a convex polygon in the plane. Points x1 , . . . , xn Q are
said to x Q, if for every direction on the plane an innitesimal translation of Q in that
direction contains one of the xi in its relative interior.
a) [1] Prove that one can always nd six points which x Q.
b) [1] Prove that six points is necessary only when Q is a hexagon with parallel edges.
Exercise 42.47. Let P1 , P2 , . . . , Rd be a family of non-intersecting convex bodies. We say
that Pi can be extracted if there exists a continuous vector valued function f : [0, ) Rd
such that kf (t)k = t and a polytopes P1 , . . . , Pi + f (t), . . . , Pn are non-intersecting for all
t 0.
a) [1] Prove that for d = 2 and a nite family of non-intersecting convex polygons as above,
there is always a polygon Pi which can be extracted.

376

b) [1] Prove that for d = 2 and a nite family of three or more non-intersecting convex
bodies, at least three of them can be extracted.
c) [1+] Prove that for all d and any nite family of d + 1 or more balls in Rd (not necessarily
of the same radius), then at least one can be extracted.
d) [1+] For d = 3, construct an innite family of non-intersecting convex polytopes Pi which
lie between two parallel planes, and such that none can be extracted.
e) [2-] For d = 3, construct a nite family of non-intersecting convex polytopes Pi such
that none can be extracted.
f ) [2-] For d = 3, construct a nite family of non-intersecting unit cubes, such that none
can be extracted.
g) [] Is it always possible, for a given convex polytope P R3 , to construct a nite family
of polytopes congruent to P such that none can be extracted?
h) [2+] We say that a family of polytopes Q1 , . . . , Qk R3 can be extracted from a family
of polytopes Q1 , . . . , Q R3 if for some f : [0, ) R3 , kf (t)k = t, and some xed k,
1 < k n, polytopes Q1 , . . . , Qk , Q1 + f (t), . . . , Q + f (t) are non-intersecting. A family
of polytopes P = {P1 , . . . , Pn } in R3 can be taken apart with two hands if some subset A
of P can be extracted from P r A. Find a nite family of non-intersecting convex polytopes
which cannot be taken apart with two hands.

377

Hints, solutions and references to selected exercises

1.2. For b), consider all halfspaces dening P , and suppose all normals are in general
position.
1.3. By Corollary 1.7, it suces to check that every three vertices of Q lie inside a circle
of radius L/4. For obtuse triangles the claim is trivial, and for acute triangles the claim
for the circumradius can be checked directly. For a generalization and another proof see
section 24.1.
1.4. This problem is given by V. Proizvolov in [Kvant], M1665 (1993, no. 3).
1.5. Start with the largest area triangle with vertices at zi .
1.8 and 1.9. See [BajB].
1.10. See [BolS, 19].
1.11. a), b) and c) These are special cases of results by Moln
ar and Baker. See [Bak] for
the references.
1.12. a) From part a) of the previous exercise and the logic as in the proof of Corollary 1.7,
it suces to check that no four spherical circles of radius r can cover S2 . This follows from
the area argument when r < /3.
For b), when r = /2 , the claim is not true; vertices of a regular tetrahedron give a
counterexample. The bound r < 2 arcsin 13 given by this example is optimal.
1.13. See [Leic, 1.7] and [DGK, 5].
1.14. Part b) is a discrete version of [Hal1]. Similarly, parts c) and d) are discrete versions
of special cases of [Hal2].
1.15. See [YagB], Problem 20.
1.16. Part a) is due to Berge (1959) and part b) is due to Breen (1990). We refer to [B
arM]
for references, details and extensions.
1.17. This is proved in [BolS, 21].
1.19. An ingenious solution of this problem is given by N. B. Vasiliev in [Kvant], M30 (1971,
no. 4).
1.20. See [ErdP, 6.4] for references and related results.
1.22. Both results are due to Breen [Bre].
1.23. A short proof is given in [Gug4].
1.24. This is due to Dvoretzky [Dvo] (see also [BZ2, 6]).
1.25. Part a) is due to Schreier (1933) and the generalization to higher dimension is due to
Aumann (1936). See [BZ2, 5] for further results and references.
1.26. See [Rub].
2.1. Note that in R4 the facets can intersect by a 2-dimensional face which can have dierent
triangulations on each side.
2.3. a) Denote by n, , and m the number of points, line intervals and regions in the plane.
Prove by induction that n + m = 1 (or use Eulers formula). We have:
X
X
X
X
n=
pi , =
ipi , m =
qj , 2 =
j qi ,
i2

i2

3 = (3n ) + (3m 2) =

j3

X
i2

(3 i)pi +

j3

X
j3

(3 j)qi ,

378

p2 = 3 +

X
i4

This proof follows [BorM].

(3 i)pi +

X
i4

(3 j)qi 3.

2.4. A simple proof, references and generalizations are given in [ErdP, 3.3].

2.5. Denote by mi the number of line passing through point xi . Summing the numbers
over all these lines, we obtain (mi 1)ai + s = 0, where s = a1 + . . . + an . We conclude
that (mi 1)ai = (mj 1)aj for every 1 i, j n. Since not all points line on the same
line, we conclude that all ai have the same sign, a contradiction. This proof is given by
F. V. Vainshtein in [Kvant], M451 (1978, no. 5).
2.7. Use the previous exercise. See [B
ar] for details.
2.8. For b), note that center of a cross-polytope Q Rd does not lie in the interior of any
subset of vertices of Q.
2.9. See [BKP].
2.11. For each xi , consider n 2 disjoint triangles with a vertex at xi . Now place y1
close to x1 in the rightmost of these triangles around x1 , place y2 in the second rightmost
triangle around x2 , etc., until the leftmost triangle around xn2 . Check that y1 , . . . , yn2
are as desired. This simple solution is given by N. B. Vasiliev in [Kvant], M551 (1980,
no. 2).
2.12. For a), take a Voronoi diagram of points X in P (see Subsection 14.2). Choose the
cell of largest volume. For b), prove the result by induction on k, for n = 3(2k 1). For
k = 1, we have n = 3 points. Cut P with a plane which contains X and choose the larger
part. For k > 1, take two points in X and cut P with a plane through these points which
divides P into two parts with equal volume. Now apply the inductive assumption to the
half with fewer points in X. This part of the problem is based on a solution by L. Lipov
in [Kvant], M375 (1976, no. 11).
2.14. Caratheodory theorem says that every such v belongs to a tetrahedron with vertices
in Q. Consider all triangles with two vertices the vertices of and the third vertex on Q.
Check that one of them work.
2.15. The proof is given in [Ani, 1]. See also [Scho] for classical references and applications
of convex polygons in Rd .
2.16. This result was proved in [Naz]. See also [Kara] for extensions and further references.
2.17. This is a classical result of Lov
asz (1974).
2.18. This is a recent result of Holmsen and Pach (2008).
2.19. See [B
aFu].
2.20. This result is due to Dolnikov and has appeared in the 2004 Russian Math. Olympiad.
3.1. Take the furthest pair of points of the polygon to be the median of a square.
3.2. For both parts, note that in every centrally symmetric set X, the points x, y X with
|xy| = diam(X) must be opposite. For a), intersect the polytope with a generic hyperplane
through the center of symmetry and note that opposite vertices lie on dierent sides. For b),
note that X B, where B is the ball of radius diam(X)/2. The result now follows from
the proof of the Borsuk conjecture for B.
3.4. For a) and b), see [BolG, HDK]. Part c) is proved in [BojF, 3.38].

379

3.6. Part b) is called Barbiers theorem; see [BolG, YagB] and Exercise 24.6. Part c) is due
to Blaschke and Lebesgue, and a simple proof can be found in [Egg1, 7.2] and [YagB, 7].
3.7. Take four copies of the same curve joined at the vertices of a square.
3.8. See [Hep2, Mel] and the references therein. Part b) was proved by Neville (1915) who
found the optimal disk conguration. Part c) is proved in K. Bezdek (1983).
3.9. This problem is discussed on the Math. Overow, see http://tinyurl.com/ykzgfxh
4.1. Take a function g(x) = f (x + a) f (x). Either g(x) changes sign or it does not. In
the rst case, g has a zero. In the second case, if g(x) > 0 for all x, then g(x) , a
contradiction. Alternatively, this follows from the mountain climbing lemma (Theorem 5.5),
when two climbers start climbing at two minima.
4.2. For a (0, 1] r {1, 12 , 13 , . . .}, Levy used the following function:
x

f (x) = sin2
x sin2
.
a
a
See [Lyu, 34] for an elementary proof and further references.
4.3. This is proved in [Ros].
4.5. For a), the argument in the proof of Proposition 4.1 does not work since the lines 1 , 2
are not necessarily unique. Note that for generic the number of lines 1 (or 2 ) must be
odd. Now consider the set of pairs of lines which work and use the parity argument as in
Section 5.
4.7. See [Buck].
4.8. Take a 1 t rectangle, where t > 0 is large enough.
4.10. It is easy to make a mistake in [BorF] as pointed out in [BMN]. In the correct
construction, take n points on the curve y = 9x3 very far apart from each other. See the
proof in [BMN].
4.11. See [Schu1].
4.13. For c), see [Alon].
4.14. Choose the closest two pearls of the same size and cut between them, right after the
rst one. This problem is given by V. Proizvolov in [Kvant], M1684 (1999, no. 6).
4.15. This is equivalent to the inscribed chord theorem (Theorem 4.5). See [Tot] for proofs
and several related results.
5.1. For a), consider a polygon X symmetric to X with respect to O. By the area argument, X X 6= , which gives the desired triple. For b), use the same construction and
Lemma 9.6. For c), take any concave vertex xi . Such a vertex always exists by Exercise 24.2.
5.2. An easy uniqueness argument for part a) is given in [Emch]. For b), the continuity is
immediate from the second proof, in all directions except for those orthogonal to the edges.
On the other hand, the limit of rhombi is also a rhombi. Now the uniqueness from part a)
implies the result.
5.3. For a), denote the inscribed quadrilateral by Q = [v1 v2 v3 v4 ]. Clearly, vertex v1
determines v2 , v3 and v4 . Suppose Q is not cyclic, i.e., f (v1 ) = v1 v2 + v3 v4 6= 0.
As v1 moves along X, function f changes sign: f (v1 ) = f (v2 ), and by the intermediate
value theorem we have f (z) = 0 for some z X.
For b), consider f (v) = vol(v1 v2 v3 v4 ), the signed volume of a tetrahedron and proceed
analogously. While both parts are probably well known, [Stru] is the only reference we
were able to nd.

380

5.4. See [Egg2].

5.5. Start with an inscribed equilateral triangle near the vertex x and continuously expand
it until one of the endpoints is reached. See [Mey1] for details.
5.6. The following proof by Erd
os is presented in [Jac]. Denote by A R2 the region
enclosed by Q and by Qi the arcs of Q separated by xi . Let di (z) be the distance between
z A and Qi . Finally, dene Ai to be a subset of A consisting of points z A such
that di (z) dj (x) for all j. Observe that each Ai is connected since for every z Ai we
have z is connected to Qi within Ai by a straight line to the closest point in Qi . Now, if
A1 A2 A3 = , then A1 (A2 A3 ) consists of two nonempty sets A1 A2 and A1 A3 , and
thus disconnected. But this is impossible since A is simply connected and thus unicoherent
(see Exercise 6.14). Finally, a circle at z A1 A2 A3 with radius d1 (z) = d2 (z) = d3 (z)
is the desired circle.
5.7. This was proved in [Mey1] (see also [Mey3]).
5.9. For a), consider the location of the third vertex for equilateral triangles with two
vertices on given parallel lines, and use the continuity argument. For b), use induction on d
to construct various regular simplices with d + 1 points on given hyperplanes, and use the
continuity argument again.
5.11. a) This is false. A counterexample is given by a cone with face angles /2, /2
and /3. Let O be the vertex of C and suppose (A, B, C) are the vertices of an equilateral
triangle, such that AOB = /3. Since |CA| = |CB| we have |OA| = |OB|, which in turn
implies that |AB| = |OA|. But |CA| > |OA|, a contradiction. Part b) is also false for all
cone face angles < /3.
d) Denote the faces of the cone by F1 , F2 , F3 , F4 (in cyclic order), and let = F1 F3 ,
= F2 F4 . Check that every plane parallel to , intersects the cone by a parallelogram.
5.12. For a), take L parallel to opposite edges. Continuously move it between the edges
and compare the edge lengths of the resulting parallelograms.
5.13. This is false. Consider a polygon Q = [x1 x2 x3 x4 ], where x1 = (0, 0, 0), x2 = (0, 0, 1),
x3 = (1, 0, 0), and x4 = (1, 1, 0). This construction is given in [MeyA] (see also [Mey3]).
5.14. See [Koe] and further references in [CFG, B2].

5.15. For general k and smooth curves part a) was proved in [Wu]. The piecewise linear
case follows by a limit argument.
5.16. Part a) is proved in [Kake], while part b) is proved in [Gri]. For d), take a polygon X
approximating a circle. This implies that Q must be cyclic (inscribe into a circle). Similarly,
take a triangle X with sides , , and 2 . Observe that a quadrilateral similar to Q is
inscribed into X, for all large enough, only if Q has parallel edges. Together these two
conditions imply the result.
5.17. For a), see Subsection 23.6. The same idea works in other parts as well. For d),
there are 24 ways to arrange vertices on dierent lines i , there are 4 24 ways to arrange
vertices on lines i , such that two adjacent vertices lie on the same line, and there are 2 24
ways where two diagonal vertices lie on the same line. This gives a total of 168 distinct
quadrilaterals.
5.18. This proof is outlined in [Mak2].
5.19. For a), use the proof idea of Proposition 5.9. See [Kra] for the details. Parts b) and c)
are given in [HLM].

381

5.20. Part a) was claimed by Pucci (1956), but the proof was shown to be incorrect in [HLM].
Even though the question was listed as an open problem in [KleW], the proof of a much
more general result is outlined in [Gug2].
5.21. Take the intersection of two tall pencil-like cones at a xed acute angle to each other.
See [Biel] for details.
5.22. Part a) was proved in [Fenn], part b) in [Zak1] (see also [KroK]). See [Mey2] for a
construction in c).
5.23. For a), take f1 (x) = 12 for all x [ 13 , 23 ], and let f2 have x sin( x1 ) singularity at 12 .
See [Kel] For b), show that a variation on the rst proof works.
5.24. Use the second proof of the mountain climbing lemma (Theorem 5.5). Alternatively,
prove the claim by induction, since fi (gi (t)) = fj (gj (t)) immediately implies fi (gi (h(t))) =
fj (gj (h(t))), for every h : [0, 1] [0, 1].
5.25. Consider the graph of points (x, y) such that |xy| = 1, where x, y C. Then use the
idea of the second proof of the mountain climbing lemma (Theorem 5.5). See [GPY] for
details and further references.
5.26. Parts a) and b) were proved in [GPY]. Part c) was proved in [Ger].
6.3. This was proved in [NieW].
6.6. A smooth case is outlined in [Mak1].
6.9. This was proved in [Liv, Zar].
6.10. This follows immediately from the proof of the Kakutani theorem (Theorem 6.3).
6.11. This follows immediately from the tripod theorem (Theorem 6.4), with = /3 and
= 2/3.
6.12. This was proved in [YamY].
6.13. This was proved in [HMS].
7.3. Prove the claim for polygons rst and then use the limit argument. See [Grub, 9.1]
for a concise proof of parts a) and b).
7.4. This is a special case of [GarM].
7.5. This is almost always false.
7.7. This is a classical result of Bonnesen (see [Sant, 7.5]).
7.9. Without loss of generality, we can assume that the angle of vectors vi increases clockwise. Consider a convex polygon Q = [O a1 . . . an ] where ai = v1 + . . . + vn (we may have
O = an ). The perimeter of Q is at least 1. Suppose the diameter is achieved on a diagonal
(ai , aj ). Then choose I = {i + 1, . . . , j}. Finally, use the isoperimetric inequality to show
that the diameter of a convex polygon is at least 1/ times perimeter. In the opposite
direction, the constant 1/ is obtained in the limit of polygons Qn which approach a unit
circle. This proof follows [SCY], Problem 6.
7.10. See [Grub, 8.3].
7.11. For a), project Q2 onto Q1 . Check that this map shrinks the distances. The same
approach works in any dimension.
7.12. For a), use monotonicity of the mean curvature (Exercise 28.2). For b), consider
parallel projections of the edges of P1 onto the faces of P2 . Sum the lengths of projections,
compare these to L2 and note that the sum of three projections of the same edge is at most
the length of the edge. This part is based on a solution by A. Kh. Shen and V. O. Bugaenko
in [Kvant], M1687 (1999, no. 6).

382

Part c) is false. The idea of a counterexample is given in Figure 42.2. When the edge
lengths of long edges increases, four such edges of inner tetrahedron 1 will overcome
the three edges of the outer tetrahedron 2 . This problem and solution appeared at the
Moscow Math. Olympiad in 2002; available at http://tinyurl.com/2vftjn
Part d) is also false. Consider any two convex polytopes P1 , P2 with L1 > L2 , as e.g., in c).
Now the polytopes Q1 = P1 + P2 , Q2 = P1 + P2 for suciently small > 0 is the desired
pair of polytopes (cf. Section 36).

Figure 42.2. Tetrahedra 1 2 with L1 > L2 .

7.13. For a), take a plane through a face F and observe that it ts inside a circle of
radius 1. Therefore, perimeter(F ) 2. By the isoperimetric inequality for the plane,
we obtain: area(F ) L2 /4 L/2. Summing this inequality over all faces and taking
into account that all edges will be used twice, we obtain the result. To see that the inequality
is sharp, consider doubly covered polygons inscribed into the equator of S2 .
For b), consider projections onto every F of all triangles (Ovw) as in the statement of the
exercise. The area of each projection is at most |vw|/2, since the distance from O to edge
(v, w) is at most 1. Now sum this over all edges e of P .
7.14. a) This was proved in [Fej1]. b) See [Lin].
7.15. a) See [BesE]. b) This was observed in [Bes2] (see also [She1]).
7.16. For a), b), c) and e) see the hint to Exercise 14.25, part a). In all cases, remove the
layers of parallelograms (parallelepipeds) one at a time and prove the claims by induction.
See [SCY], problems 112119, for the easy proofs and [Zie1, 7.3] for the references.
8.2. An elegant presentation of this is given in [Zie1, 9.2].
8.3. By the argument as in the section, all vertices on one side of the hyperplane H are
connected to a vertex lying on the nal face with respect to H.
8.4. See [BrM].
8.5. Part a) is given in [VasE], Problem 461. For part b), P must be a tetrahedron. Indeed,
if P has n faces, it has exactly (2n 4) vertices and (3n 6) edges. Since the number of
edges of each color must be at least 2n5, we have 2(2n5) 3n6, which implies n = 4.
For c), let v1 , . . . , vn be the set of vertices ordered according to some Morse function , so
that the simplex = (v1 , . . . , vd ) is the smallest facet and each vi , i d + 1, is connected
by at least d 1 edges to vertices vj with j < i. Color the vertices of with d 1 colors to
satisfy the requirements. Color the remaining vertices vi by coloring with dierent colors
the d 1 edges pointing to vj , j < i. This gives a coloring where every vi is connected to v1
by any of the d 1 colors, as desired.
8.6. The result is obvious for d = 2. For d 3, cut P with a hyperplane H around a
vertex v. By induction, the d 1-dimensional polytope Q = P H has an embedding
of Kd . Project away from v the vertices and edges of this embedding onto the vertices and

383

edges in . Connecting them by the edges with v gives the desired embedding of Kd+1 .
This result is attributed to Gr
unbaum, while our proof follows the outline in [Kuh, 2].
8.7. Observe that the total number of edges of each color must be the same, since it is equal
to one quarter the total number of sides of all faces. Now take P to be any polytope as
in the theorem with an odd number of edges. This elegant solution is given by S. Tokarev
in [Kvant], M1365 (1993, no. 2).
8.8. For a), use Eulers formula to conclude that the average degree of a vertex is strictly
smaller than 6. Parts b) and c) are due to Kotzig (1955, 1963) and use more delicate
applications of Eulers formula (see [Bor] for extensions and references). To obtain sharp
bounds in part d), take polytopes dual to the truncated dodecahedron (an Archimedean
solid with two decagons and one triangle adjacent to every vertex) and icosidodecahedron
(an Archimedean solid with two pentagons and two triangles adjacent to every vertex).
8.9. a) Denote by the line through two vertices of P at distance diam(P ). Every plane
H intersects at least three edges. Therefore, projections of edges of P onto cover every
point at least three times. Since the sum of these projections is at most L, this implies the
result.
b) By Mengers theorem, for every two vertices x, y there exists three non-intersecting
paths from x to y. Thus, for the diameter d of we have 3th(d 1) + 2 n. To see that
this is sharp, stack d 1 triangular prisms and two regular triangular pyramids attached on
the opposite ends. Now perturb the surface to make it strictly convex. See [GM1, JucM]
for complete proofs.
8.10. Part a) follows from Eulers formula. For b), consider paths on which are dened
by moving right and left, alternatively. Prove that every such path is closed and has even
length. Conclude from here the claim. We refer to [GM2] for details. Part c) is proved
in [Mot] (see also [Gr
u1]).
8.11. This ia variation on Exercise 24.2.
8.12. In R4 , take two equilateral triangles lying in two orthogonal 2-planes, with centers
at the origin. Their convex hull is the desired neighborly polytope. In Rd , d 5, take 6
points in a 4-dimensional subspace as before, and the remaining points in general position.
See [NagC] for the details.
8.13. The result is obvious in d = 2. In general, take any facet F P . If F is not simplex,
use inductive assumption. If all faces are simplices, choose any adjacent pair of facets.
See [Tve].
8.14. For b), in the same way as in a), consider two cases: when P is not simple and not
simplicial, and use induction on d. See [Dev] for the easy details.
8.15. For a) and b), assume every cross section of P is a k-gon. Intersect P by a plane near
a vertex. Then every vertex has degree k. Now intersect P by plane parallel to an edge e
and near e. The intersection if a (2k 2)-gon, a contradiction. This implies part c) as well.
For d), x a plane in general position and move it across the polytope. When crossing a
vertex, observe that the parity is unchanged.
8.16. For a), take P obtained by attaching triangular pyramids to the faces of an octahedron.
Check that P does not have a Hamiltonian cycle. For b), use an iterative construction.
8.17. For c), let : Qn,k Pn,k be dened by (x1 , . . . , xn ) = (y1 , . . . , yn ), where y1 = 1x1
and

xi1 xi if xi < xi1
yi =
for i 2 .
1 + xi1 xi if xi xi1

384

This elegant construction was discovered in [Sta1].

8.18. For a), prove combinatorially the recurrence relation Bn+1,k = (n k + 1)Bn,k1 +
(k + 1)Bn,k and use induction on n.
8.19. Part b) is a classical result of Birkho, (see [Barv, EKK]). For c), see [EKK]. For f )
and g), see [LovP].
8.20. Parts a) d) are due to Klee and Witzgall (1968). We refer to [EKK] for a nice
exposition. For e), see [Pak2].
8.21. Assume that no two particles enter vertices at the same time. Take a dual graph
to the graph of the polytope. For every vertex in corresponding to a face F , orient an
edge in along the edge where the particle is moving. Since the number of oriented edges
is equal to the number of vertices in , there is an oriented cycle. Consider how the cycle
changes as one particle passes through a vertex. Check that if there are no collisions, the
cycle gets smaller every time, a contradiction.
This result and its group theoretic applications are described in [FenR] (see also a friendly
exposition in [Ols]).
8.22. For part a), assume P is simplicial. Then it has n 6 vertices. By Eulers formula,
it has 3n 6 edges and 2n 4 faces. We obtain:
f0 + f1 + f2 + f3 = n + (3n 6) + (2n 4) + 1 = 6n 9 27.

If P is not simplicial, it has a non-triangular face F . Denote by F the centrally symmetric

face and by G the intersection of P with a hyperplane through the origin and parallel to F .
Since the vertices and edges of G correspond to the edges and faces of P , we obtain:
f0 (P ) f0 (F ) + f0 (F ) 4 + 4 = 8,

f1 (P ) f1 (F ) + f1 (F ) + f0 (G) 4 + 4 + 4 = 12,
f2 (P ) f2 (F ) + f2 (F ) + f1 (G) 1 + 1 + 4 = 6,
which implies the result. The above argument, the proof of part b), various extensions and
references are given in [SWZ]. The last part is an open problem due to Kalai.
9.1. a) Take e = (x1 , x2 ) to be the longest edge. Let 1 and 2 be the lines perpendicular
to e going through x1 and x2 , respectively. If no vertex projects onto e, at least one edge
intersects both 1 and 2 , contradicting the choice of e.
For b), note that the min max in () can never be achieved on a cut through two vertices.
For c), take, e.g., the great stellated dodecahedron. This problem appeared in [CroW].
9.2. For a), take a triangle of the largest area.
9.4. a) This is false as stated. Think of a generic curve C made of metal and place on the
oor. In the equilibrium C either has three tangent points, or two point which have contact
with C of order 2 and 1, or one point which has contact with C of order 3. For b), consider
a plane through any face of the convex hull of Q.
9.5. For b), consider a triangle = 2, and such that vertices of are midpoints of .
Use the fact that has maximal area to conclude that Q is inscribed into . Note also
that cm( ) = cm(). Now bound the desired ratio by that of inside and inside . For
details, see e.g., [Pro1, 9].
9.6. This is a discrete analogue of the result in [Tab3], which uses a version of the four
vertex theorem (Theorem 21.1). While the discrete result easily follows from the smooth
result by the limit argument, we do not know a simple direct proof.
9.7. These are discrete versions of several known results gathered in [Gug3].

385

9.8. This was proved in [DeoK].

9.10. For parts a) and b) see [CGG, Daw2]. See [Hep1] for part c). Part d) is a delicate
result due to Dawson [Daw2]. The main idea is to apply the Minkowski theorem 36.2 to
the simplices. Part e) is in two followup papers by Dawson and coauthors (1998, 2001).
9.11. Consider a degenerate (at) parallelepiped as in Figure 42.3. Perturb the dihedral
angles to make it non-degenerate (still nearly at). When placed on one of the non-square
faces, the center of mass lies outside the face, so the polytope rolls.

Figure 42.3. Parallelepiped which can stand only on two faces.

9.12. For b), start with a construction of a 17-gon using the idea of a heptagon in Figure 9.3,
optimizing it in such a way that point O can be taken as close to the center of mass
as possible. Now use this polygon as a cross-section of the slanted cylinder described in
Example 9.9. Make the sides slanted far enough so that the center of mass of the resulting
polyhedron projects onto O. We refer to [CGG] for details.
Parts c) and d) were recently obtained in [VD1, VD2]. In the plane, think of the projection
points as minima of the distance function from cm(Q). Now apply Lemma 9.6.
9.13. For a), this is called the orthic triangle (see [CoxG, 1.6]). For b), see [Phi].
9.15. This nice observation was communicated by Joe ORourke.
9.17. Denote by L = () the length of the longest triangle = [x1 x2 x3 ] inscribed into C.
For every point z C, denote by Rz the longest triangle inscribed into C with z as a vertex.
Clearly, (Rz ) = L when z = x1 and z = x2 . Take Rz with the smallest length (Rz ), where
z lies between x1 and x2 . Check that this is the desired billiard trajectory. See [Bir, 6.6],
[Tab6, 6] and [CroS].
9.18. This is proved in [Weg1].
9.19. A direct proof is given in [Kui]. By analogy with the 2-dimensional case, one can
reduce this problem to the number of simple closed geodesics on convex bodies in R4 , and
conclude that there exist at lest four double normals (see [Alb]). We do not know an
elementary proof of this result.
9.21. See [Tab2, 4] and [Tab6, 9], for these and other related results, background and
references (see also [Tab7]). Parts i) and j) are due to R. E. Schwartz (2007).
10.1. For a), consider a regular pyramid with a face angle on the bottom of side
triangles. The face angle at the top of side triangle is = 1 2 . Now, by the proof
above, we need only to check that 1 + 2 6= 2, i.e., 6= /3. For example, a standard
tetrahedron corresponding to = /4 gives the desired example (see Figure 10.5).
b) Consider a truncated regular pyramid as in Figure 42.5. Let the sums of face angles be
2
2
/ Q there are no
3 + for the top vertices, and 3 for the bottom three. When /
simple closed geodesics.

386

v1

v3

v4

v2

Figure 42.4. An example of a regular pyramid without simple closed

geodesics: = /4, = /2 (dierent views).

Figure 42.5. An example of a polytope P with only non-simple geodesics.

10.2. Consider a unit cube in each case and a geodesic starting with an interval [ab], where
1
, 0, 0), and k N.
a = (0, 12 , 0), b = ( 8k
10.3. These calculations were made in [FucF].
10.4. a) Recall that every geodesic divides the curvature into equal halves, and that pairs
of separated vertices are dierent for the three geodesics. We obtain three equation which
imply that the curvature of every vertex is . The result now follows from Exercise 25.12.
Parts b) and c) are given in [Pro2] (see also [StrL]).
d) For an equihedral tetrahedron , tile a plane with copies of its faces. Consider lines
which do not contain vertices and have irrational slope with the respect to the resulting grid.
Such lines when folded back onto give innite simple geodesics. In the other direction,
unfold the tetrahedron and extend the idea of the proof of Theorem 25.6 to show that the
curvature in every vertex is . For a complete proof of part d) and e), see [Post].
11.1. For c), see [Shar], Problem 221.
11.2. Start with two adjacent faces F1 and F2 which can be inscribed into a sphere S2 .
Check that all faces adjacent to both F1 , F2 are also inscribed into S2 (use the fact that P
is simple here). Continue with faces adjacent to two previously inscribed faces. Show that
this process halts when all faces are checked.
11.3. See [ShaS], Problems 5.21 and 5.35.
11.4. For a), assume there exists an inscribed sphere. For every face F , take a tangent
point aF of the sphere and triangulate F by taking cones from aF over the edges of F .
Now calculate the surface are of faces colored each color and compare the results. For b),
take a 1!!3!!3 box and color large squares black. For c), take a symmetric bipyramid Q
over 1!! rectangle, with > 0 suciently small, so that Q has no inscribed sphere. Then
there is a unique proper coloring with equal white and black areas by the symmetry. Now
perturb the vertices of Q to break the equality.
11.5. See [Gr
u4, 13.5] for the proof and some background.

387

11.6. Part a) is proved in [Bar1]. For b), start with a tetrahedron and construct a pyramid
on each face. The resulting polyhedron is dual to the truncated tetrahedron, has 8 vertices
and 12 faces. Show that it cannot project onto a regular octahedron. This solution is given
in [Bar2]. Note that this problem is a generalization of Exercise 14.16, corresponding to the
case where P is simplicial and C is a Hamiltonian cycle. Thus part a) of this exercise is a
generalization of c) of Exercise 14.16. Similarly, a counterexample described in the solution
of part b) of Exercise 14.16 also works for part b) in this exercise.
11.7. See [GMS] and [Schu1].
11.8. This is a special case of the main result in [Tho1].
12.2. This is due to Zamrescu and Gleason (see [Gle] and references therein).
12.3. Both parts are proved in [Schr2] (see also [Var2]).
12.5. Draw two parallel lines and take midpoints of intersections with the parabola. Show
that the line through the midpoints is parallel to the y axis. See [Vard], Problem 23.
12.6. Consider the right triangle with vertices: the center O, the vertex v and the
edge
midpoint z of a pentagon. If the pentagon is rational, then cos 310 = |vz|/|vO| = a for
some a Q, a contradiction.
12.8. See a construction in [Zie3].
13.4. This result is due to A. Galitzer (see [KM3, 3]).
14.4. See [Epp, ShaH].
14.5. For triangulations of the middle regions, encode them with the 01 sequences, which
corresponds
to k 1 up triangles and k 1 down triangles. The number of these sequences


is 2k2
,
and
two remaining Catalan numbers Ck2 correspond to triangulations of the top
k1
and bottom k-gons. See [HNU] for the details (see also [DRS, 3.4]).
14.6. This problem was on a 2007 Putnam Mathematical Competition. See a solution here:
http://tinyurl.com/2crvn9w
14.7. See [DRS, 3.3].
14.8. For a) see [HNU]. For b) and c), see [Gal+].
14.10. For part a), use at most n 2 moves to transform a given triangulation to a
triangulations with all diagonals meeting at the same point. Part b) is proved in [STT]
using hyperbolic geometry; a combinatorial proof is given in [Deho]. See also [DRS, San2]
for further references.
14.11. Part b) is false (see [BerE]).
14.12. For each of these functionals check that they behave accordingly along increasing
ips (essentially, prove the claim for all quadrilaterals). See [BerE] and [Mus2] for more on
this and further references, and [Lam] for g).
14.13. See [Tut, 10] (see also [PouS] for further references, applications and a bijective
proof).
14.14. For a) and b) use the idea in the example. For c), consider a grid ne enough and
add grid vertices. For parts c) e) and the references see [BZ4, Mae1, Sar, Zam]. For the
progress towards f ), see [ESU]. Part g) was proved in [Kri].
14.15. For a), triangulate the projection and lift it to X. For this and other parts, see [BDE].
14.16. For a), use a binary tree as in Section 8 and lift the triangles up depending how far
they are from the root. Alternatively, take the function as in Example 8.5. For b), take a
regular hexagon Q = [x1 . . . x6 ], triangulations T1 with diagonals (x1 , x5 ), (x2 , x4 ), (x2 , x5 ),

388

and T2 with diagonal (x1 , x3 ), (x1 , x4 ), (x4 , x6 ). See [Dek] for details and further references.
See also Exercise 11.6 for a special case of part b) and a generalization of part c).
14.17. For a), take six 1111 bricks and attach each of them in the middle to a unit cube.
Now slightly decrease the size of the bricks to make sure they do not intersect.
For b), one can use the construction in a). Here is an alternative approach. Solve a 2dimensional version rst and then add two large parallel slabs in R3 .
14.18. a) This is one of the oldest results in discrete geometry [Len]. The idea is to nd a
diagonal inside Q and then use induction. First, project Q onto the x axis and let v be the
leftmost vertex. Either the diagonal (u, w) between vertices neighboring v is inside Q, or
triangle (uvw) contains at least one vertex of Q. Then the diagonal (v, z) connecting the
leftmost such vertex z with v cannot intersect any edge of Q, and thus lies in the interior
of Q.
For parts b) and c), take the Sch
onhardts polyhedron [Sch
o] (see also an extension in [Ramb]
and [Bag]). Simply take a triangular prism and twist the top triangle (see Figure 42.6).
We refer to [BerE] and [DRS, 3.5] for the context and further references. Part d) is given
in [Bin].99
For part e), take e.g., six bricks as in part a) of the previous problem and connect them by
short tubes far away from the center.
14.19. See [RupS].

Figure 42.6. Schonhardts polyhedron and a non-regular triangulation.

14.20. For a), the standard example is given in Figure 42.6. For b), this is also false, e.g.,
for a cuboctahedron (see Figure 16.4). There is also an example with 6 vertices. For the
rest of the exercise see [San1].
14.21. See [Dil].
14.22. For a), denote by Rt the intersection of a simplex in the d-cube with a hyperplane
x1 = t. Show that vol(Rt ) = cti (1 t)di , for some c > 0, where i depends on the number
of vertices in with x1 = 1. Now use the fact that there is a unique linear combination of
polynomials ti (1 t)di equal to a constant. This argument is given in [Glaz]. For b) and
further references, see [DRS].
14.23. For a), here is a simple ad hoc argument: Draw the edges between centers of squares
inside each domino. Overlap two tilings. Find the innermost circle C. Since both tiling
coincide inside C, use 2-moves to remove this circle. Repeat this procedure until the tilings
coincide.
For b), the counterexamples to the rst two questions are given in Figure 42.7. For the nal
part, take an n n square region G in z = 0 level and in the checkerboard fashion add
99An

easier example can be found in http://tinyurl.com/radztn.

389

squares in z = 1 and z = 1 level, above or below every square except for the boundary
of G. This xes the (n2)2 vertical dominoes and allows only two possibilities the horizontal
dominoes in z = 0 level.

Figure 42.7. The 2-move connectivity fails for non-simply connected and
3-dimensional regions.
Parts c)g) are based on Chabouds exposition (1996) of Thurstons construction (1990).
For g), simply observe that height functions are exactly the integer functions on which
around every square are dierent by 1 on three edges and by 3 on one edge. Now check
that and preserve
Pthis.
For j), dene |h| = x h(x). Observe that under a 2-move |h| changes by 4 and that
|hmax | |hmin | = (n3 ) in this case. Part k) is due to C. and R. Kenyon (1993) and uses
height functions with values in the free product Zk Z . See [Pak3] for details, generalizations and references to all parts.
14.24. Part a) for rectangles is due to Pak (2000). Part b) is due to Sheeld (2002), who
used multidimensional height functions. Parts c)e) are due to Conway and Lagarias (1990).
See [Pak3] for the proof outline and the references.
14.25. Draw the dotted lines through the parallelograms with one side parallel to the same
side (see Figure 42.8). Number the lines from 1 to n. Dene a map from mosaics of a
2n-gon to certain reduced decompositions of a permutation = (n, n 1, . . . , 2, 1) Sn
into a product of adjacent transpositions si = (i, i + 1) by stretching them as in the gure
and projecting the intersection points on the horizontal line (transposition si corresponds
to the intersection of i-th and (i + 1)-th line from the top). These intersection points are
in bijection with pairs of elements
in {1, . . . , n} corresponding to lines, so the number of


n
possible orders is at most 2 !
5

2
4

1
2
3
4
5

5
4
3
2
1

Figure 42.8. Example of the map : D s4 s3 s2 s4 s1 s3 s2 s4 s3 s4 .

Observe that there is an ambiguity created by the order of transposition si and sj , |ij| 2.
Check that if reduced decompositions are taken modulo commutations as above, the map
is a bijection (Elnitsky, 1997). Also, the ips correspond to 3-relations si si+1 si = si+1 si si+1 .
These and commutativity relations give all Coxeter relations for Sn , which implies that ips
connect all mosaics. Observe that every reduced decomposition of must contain at least

390

one s1 , sn1 , and at least two si for all 2 i n 2, because all elements i must move
across i-th position. Use this to conclude that one can always apply at least n 2 dierent
3-relations.
For d) and e), one can nd a small non-regular mosaic by hand and then check that a
random mosaic contains lines with the same pattern. Here is a more general argument.
Consider n vectors v 1 , . . . , v n which project on the edges of P . Removing all lines but three
lines i, j, k gives one of the two mosaics of a hexagon. This submosaic gives an inequality
on
triples of vectors. Thus, for a xed
P , the space of these vectors is n-dimensional, so the
n
n
n
inequalities
correspond
to
=
eO(n log n) regions in Rn , since k hyperplanes in Rn
3
3
divide the space into kO(n) regions. In a dierent direction, use induction to show that the
2
number of mosaics is at least e(n ) .100 This implies that the number of regular mosaics is
asymptotically smaller than the number of all mosaics of 2n-gon.
15.1. Subdivide P into a large number of small cubes and approximate the boundary.
15.2. Extend the proof of Theorem 15.2.
15.3. It suces to prove the result for symmetric tetrahedra. For a tetrahedron =
(a1 , a2 , a3 , a4 ) R3 , let O be the center of the circumscribed sphere. Assume for now that
O . Denote by bi the orthogonal projections of O onto the face Fi opposite to ai , for
all 1 i 4. Subdivide into 12 tetrahedra with vertices at O, bi and two vertices of the
face Fi . Observe that each of the resulting tetrahedra is mirror symmetric, which implies
that is scissor congruent to its mirror image in this case. When O
/ , use the same
argument combined with Theorem 16.3. This proof is based on [Bri1].
15.4. Use Lemma 15.3 repeatedly or brute force to obtain scissor congruence of any parallelepiped with a brick 1 1 .

15.5. Since every polygon can be triangulated, it suces to prove the claim for the triangular
prisms. In the latter case, observe that two triangular prisms tile a parallelepiped and use
Theorem 16.4.
15.7. See [HadG] and [Bolt, 911, 19].

15.8. For both a) and b), start with the scissor congruence and x it by switching simplices
on the boundary with those in the interior. For c), take unit height prisms over a unit 3-cube
and a regular tetrahedron of the same volume. See [Schn1] for a related problem.
15.10. Place the centers of the cross tiles in the lattice points (a1 , . . . , ad ) Zd , where
a1 + 2a2 + . . . + dad = 0 mod 2d + 1. See [SteS, 3] for the complete proof and references.
15.11. For a) and b), see [Ste1] and [Schme], respectively. Part c) is given in [HocR].
15.12. For d), see [HocR]. Part g) follows from [Schme] or by a direct argument.
15.13. See [Ada1, Schmi] and references therein.
15.14. The idea of this problem is due to Yuri Rabinovich (unpublished). For a), note that
every translation of the unit square is a fundamental region of the natural action of Z2 .
Thus the orbit O(x) of every x Q has exactly one point in each Qi . Not all these points
are necessarily distinct, of course, but since the number of squares n is odd, at least one
point in O(x) is covered by an odd number of squares. In other words, for every point
x Q there exists a point y A in the orbit of x. Thus, area(A) area(Q).
100One

can obtain a much better bound using the exact formula due to Stanley (1984) for the
total number of reduced decompositions of (see [Sta3, Ex. 7.22]).

391

For c), take three subsets of a 1 7 rectangle, obtained by replacing 1 with unit squares
in the following sequences: 1101100, 0110110, 0011011. In this case area(A) = 2 while
area(Q) = 4. This construction is due to R. Connelly (personal communication).
For d), consider three symmetric translations of a triangle with unit sides, which have a common intersection: similar triangle with sides 1/2. This construction is due to A. Akopyan
(personal communication).
15.15. This follows by elementary counting [Pak6, 8.3]. See also [KonP] for generalizations
and bijective proofs.
16.1. Observe that Q() is the complement to eight copies of standard tetrahedra 1 . Use
complementarity and tiling lemmas to conclude that Q() is rectiable only when = 0.
Now assume that Q() cQ(), for some < . The equality of the volumes gives
c =

1 8(3 /6)
> 1.
1 8(3 /6)

Denote by D the regular octahedron comprised of eight tetrahedra 1 . From above, Q()
D H, where H = Q(0) is the unit cube. We have:
H cD Q() cD D cQ() cD D


c Q() D D cH D .

Since c > 1 as above, the equality of the volumes on both sides of the equation implies that
< c. Using s = /c, we obtain D R1 sD R2 , where s 6= 1 and R1 , R2 R.
Now Theorem 16.2 gives a contradiction.
16.2. For a), see [Ker]. For part b), see [Gol5]. See also similarly titled papers by M. Goldberg on convex tiles with other number of faces.
16.3. For a), take a circle of radius R and compute the average angle of polygons in two
dierent ways: one via the average in each octagon, and another from the fact that at each
vertex there are at least three edges meeting. Letting R , obtain a contradiction.
For b), tile a plane with T -tetrominoes. For more on tiling a plane with polygons of 7
sides see [Niv].
P
16.4. Consider the asymptotic behavior of the summations
dened in the proof of
Lemma 15.4. Calculate the asymptotics in two dierent ways (cf. Exercise 16.3). This
result is due to Debrunner (1980), and an elementary proof was given in [LM].
16.5. Take a rhombus R with irrational angles and consider two triangular prisms with
isosceles triangular faces orthogonal to R. Attach these prisms along the rhombus to form
a polytope P as in Figure 42.9. If the heights of the prisms are chosen generically, the only
way to tile the space is by attaching triangular faces to each other and forming layers as in
the gure. Since the layers are at irrational angle to each other, in the resulting tiling has
copies of P oriented in innitely many dierent way. This implies that all such tilings are
aperiodic. This example is due to J. Conway and outlined at http://tinyurl.com/3y4o4j

Figure 42.9. An example of a polytope which tiles the space aperiodically.

392

bn . Compare the second part with Exer16.6. For c), consider the ane Weyl group A
cise 7.16 a). For f ), the edges correspond to the vectors v = ei ej , for all i < j. These
vectors correspond to the edges in a complete graph Kn . Check that n 1 of these vectors
are independent if and only if they correspond to a spanning tree in Kn . Compute a determinant to show that the volume of every parallelepiped spanned by such vectors is 1. The
last two steps are exactly the same as in the matrix-tree theorem (see [Tut, 4]).
16.10. The rst such decomposition was given in [Syd2]. The decomposition in Figure 42.10
is given in [TVS]. Yet another decomposition (with only three pieces) is due to Schobi
(1985). See [SloV] for a description, references and generalizations.

Figure 42.10. A scissor congruence between tetrahedron 1 and the right

triangular prism (view from the top and from the bottom).
16.11. In fact, all these tetrahedra can tile the space, which is why they were discovered
by Sommerville (1923) and Goldberg [Gol3]. See also [Sene] for the history of these constructions and their underlying symmetries. For a), this tetrahedron tiles a unit cube. The
tiling lemma (Theorem 16.4) implies the result. For b), this tetrahedron can be tiled by two
tetrahedra in a). Similarly, four copies of a tetrahedron in c) or two copies of a tetrahedron
in d) tile 1 . Part e) is a special case of f ). For f ), observe that three such tetrahedra tile
a triangular prism and use the tiling lemma.
16.12. This result in part a) is due to Hadwiger (1951). We refer to [SloV] for a direct
proof, extensions and references. For b), see [Gol3].
16.13. See [Syd2] for details and references.
16.14 and 16.15. See [Fie] and [Cox2].
16.16. See [Deb2] (rediscovered by Brandts, Korotov and Krzek in 2007).
16.17. See [Cox1, 11] and [Cox2].
16.18. Sommervilles 1923 classication was recently nished in [Edm2].
16.19. See [Jes2].
16.20. See [Leb1] and [Jes1] for the necessary and sucient conditions on scissor congruence
of aggregates of Archimedean solids. For part c), see also [CRS].
16.22. For b) and d) see [Deb1]. For c), use the E8 lattice (see [Cox1]).
16.23. See [Bolt, 16].
16.25. An elementary proof is given in [Bolt, 20].
17.1. In higher dimension, extend Lemma 17.15 by taking height functions of star triangulations as follows. Set (a) = 1 for a center of D, (b) = for every center of a facet
triangulation, (c) = 2 for every center of a triangulation in a face of codimension 2, etc.
It is easy to see that when > 0 is small enough, the height function is as desired (use

393

induction on d). Check that when changes, the triangulation remains stable except for
2d moves inside a bipyramid in Rd .
17.4. For d), use induction on the dimension and number of facets. Suppose rst that P is
not simple. Cut P with a hyperplane near a vertex v with deg(v) > d. The polytope Q in
the intersection has a smaller dimension and by induction can be cut into simplices. Extend
these cuts to hyperplanes through v. Check that the resulting polytopes have fewer facets
and use inductive assumption. Suppose now that P is simple. By Exercise 8.13, we can
choose two adjacent facets F1 , F2 and a vertex v
/ F1 , F2 . Cut P with a hyperplane spanned
by v and F1 F2 . As in the proof of Proposition 17.12, observe that both P1 and P2 have
fewer facets than P . Use inductive assumption to nish the proof.
17.5. For a), observe that the graph dual to the triangulation is a tree, and the desired
triangles corresponds to its endpoints.
For b), use the following construction. Start with a region A1 Q adjacent to the boundary Q. If Q r Q1 is not simply connected, take a connected component B1 A1 with
the smallest number of regions. Let A2 B1 be a region adjacent to the boundary Q.
Again, if Q r A2 is not simply connected, take a connected component B1 Q r A2 with
the smallest number of regions. Repeat this procedure until one of the regions Q r Ai is
simply connected. Use induction on the number of regions in Bi . See [MucP] for a complete
proof.
For c), take three 3-dimensional tiles as in Figure 42.11. For d), a delicate construction of
one such triangulation is given in [Rud].

Figure 42.11. An impossible puzzle with three pieces.

17.6. See [IzmS] for an accessible proof of c). For d), suppose D1 , D2 P are two dissections
of P Rd . For every simplex D1 with points on the boundary, use elementary moves
to obtain a barycentric subdivision of . This gives a subdivision D1 P rening D2 .
Construct D2 similarly. By assumption, D1 and D2 are connected by stellar moves. Check
that each stellar move can be obtain as a composition of elementary moves.
18.1. It suces to show that the global change can be made with the same sequence
(, c, . . . , c), where c = min{ai , aj } > 0. For that, connect 1 to all other vertices in G and
make local exchanges to concentrate all values at 1 while keeping c elsewhere.
18.3 and 18.4. See [Sta2].
18.5. Part c) is proved in [Pak1] (see also [Pak6, 9.1]).
19.2. See [Kis, 86].
19.3. Parts a) and b) were communicated to me by Ezra Miller and G
unter Rote.
19.4. The rst unfolding is that of a polytope with two non-square rhombic faces. The
second and third are the Johnson solids called sphenocorona and disphenocingulum.
19.5. Part a) follows from the fact that the center of mass cm(P ) is a xed point of every
symmetry line. For part b), x a symmetry line . The remaining symmetry lines are split
into pairs symmetric with respect to , unless they are orthogonal to . In the latter case,

394

observe that if are symmetry lines, then the line is orthogonal to both and .
Therefore, the lines orthogonal to are split into pairs { , }, which implies the result.
This friendly solution is given by N. B. Vasiliev, V. A. Senderov, and A. B. Sossinsky
in [Kvant], M623 (1981, no. 3).
19.6. Consider six points on the edges of C at distance x from either (0, 0, 0) or (1, 1, 1).
Then use monotonicity (in fact, x = 3/4 works).101
19.7. See [Mill].
19.8. For a), take a regular bipyramid. For b), take the trapezohedron, a dual polytope to
the biprism. Finally, for c), consider a classication of nite subgroups of SO(3, R), the
automorphism group of P . We do not know a direct proof.
19.9. Consider a regular n-prism Pn (h) with varying height h. There are two types of cones
over the faces, and by continuity there exists h so that Pn (h) is fair (see [DiaK]). For n = 5,
see also the US Patent 6926275.102
19.12. a) Take a unit cube centered at the origin and attach to each face of the cube a
regular pyramid with dihedral angles /4 in the square faces. The resulting polytope P is
called rhombic dodecahedron, has 6 cube vertices of degree 3, and 8 new vertices of degree 4.
Note that P has 12 rhombic faces, is face-transitive and edge-transitive around the origin.
Thus, P is midscribed. On the other hand, P is clearly not inscribed.
b) Take the cuboctahedron (see Figure 16.4). This is a variation on Problem 15.10 in [PraS].
19.13. See a solution by V. A. Senderov in [Kvant], M1192 (1990, no. 4).
19.14. a) Such polytopes are called deltahedra. There are only eight of them (see [Cun]).
For b), note that there is only a nite number of combinations of faces around each vertex,
giving a lower bound i on the curvature of every vertex vi , for some constant =
(k) > 0. By the GaussBonnet theorem (Theorem 25.3), every such polytope has at most
4/ vertices. Thus, there is only a nite number of combinatorial types. By the Cauchy
theorem (Theorem 26.1), we conclude that there is only a nite number of such polytopes.
For c), these are prisms and antiprisms.
19.16. For a) and other polyhedra with symmetries see [Cox1, Crom, McS]. For b) and c),
see solution to Exercise 7.16 and the US Patent 3611620.103
19.17. See the Jessens original paper [Jes2].
19.18. For a), change the edge lengths one by one. For b), take 01 = 12 = . . . = d0 =
1 + and ij = 1 otherwise [MarW]. For c), consider combinatorics of dierent edge
lengths [Edm1].
20.1. See [Mus5] for references to optimal bounds.
20.2. The maximum is equal to six (take the longest diagonals in the icosahedron). To prove
this, x two lines, which we represent as two points x, y S2 on a sphere and their opposite
points x , y . Now consider all z S2 at equal distance to the closest of the pair, i.e., such
that min{|zx|, |zx |} = min{|zy|, |zy |}. Each of the four possible choices for the minima
corresponds to at most one additional lines, bringing the total to six. This observation is
due to I. F. Sharygin.
20.3. Denote by B1 the ball of radius 10 as in the theorem and suppose diam(P ) 21.
Then P lies inside the ball B2 of radius 11 with the same center as B1 . By Exercise 7.11,
101This

problem was given at the 28th Tournament of the Towns: http://www.turgor.ru/28/

online at http://www.google.com/patents?vid=USPAT6926275.
103Available at http://www.google.com/patents?vid=USPAT3611620.
102Available

395

the surface area of P must be greater than area(B1 ). On the other hand, each face F of P
lies in a circle H B2 , where H is a plane tangent to B1 . This gives an upper bound on
the area of F . Now use the condition that P has 19 faces to obtain a contradiction. Two
solutions of this problem (including this one) are given by A. G. Kushnirenko in [Kvant],
M35 (1971, no. 6).
20.4. Take spherical circles around each xi of radius /8 and use the area argument to show
that these circles cannot be disjoint. This solution is given by A. K. Tolpygo in [Kvant],
M656 (1981, no. 8).
20.6. For a), consider a line not parallel to the symmetry axis of parabolas. Check that
only a nite portion of the line is covered. For b), translate the cones C1 , . . . , Cn to have
the symmetry point at the origin. Observe that these cones are still disjoint except at the
boundary. Indeed, otherwise there exist two translations of a (small) cone which lie in two
cones Ci and Cj , a contradiction with Ci Cj = . This problem is given by A. Kuzminykh
in [Kvant], M748 (1982, no. 11).
20.7. Part b) was given at the Moscow Math. Olympiad in 1966. An easy solution is given
in http://tinyurl.com/n47h9l .
20.8. See [Daw3].
20.10. For a), take all cylinders to be parallel. For b), start with six parallel cylinders and
rotate three of them as a solid block. For c), consider the area of intersection or a cylinder
and a sphere of radius r. Optimize for r (see [BraW] for the full proof). This problem is
due to W. Kuperberg (1990).
20.11. For a), take two copies of B and put them together so some two balls touch. For b),
take the arrangements of balls in S3 at the vertices of the 600-cell. Make a stereographic
projection. For c), take the graph = (B) with vertices corresponding to balls and edges
if they touch. Orient the edges from balls of smaller to larger radius. Since the outdegree
of all vertices in is Kd , conclude the result. For more on this problem, see [KupS].

20.12. Part a) follows from part a) of Exercise 20.1 and the proof of part c) of the previous
exercise. For b), observe that n spheres in Rd are determined by their centers and radii,

and thus have nd + n degrees of freedom. For pairwise touching, they need to satisfy n2
equations. Now proceed as in Section 31. For c), make an inversion at a point where
two spheres are touching. They become two parallel hyperplanes, which implies that all
other spheres have the same radius. Parts a) and c) were communicated to me by Itai
Benjamini (2010).
21.1. b) An example is shown in Figure 42.12 (see [Mus1]). Here r1 < r2 < r3 < r4 .

v1
v2

v4
v3

Figure 42.12. A non-coherent quadrilateral.

21.2. This follows from Corollary 21.12. For a direct simple proof see [SCY], Problem 9.

396

21.4. The proof follows roughly along the same lines as the proof of Theorem 21.7. In this
case one needs to consider the set of points at equal distance to 3 or more lines and prove
that this is a binary tree.
21.6. See [Weg3].
21.10. For b), take an immersed closed curve C in Figure 9 in [Arn2]. Inscribe an equilateral
polygon Q and the same polygon Q with a shifted order of vertices.
21.11. See [OT1].
21.12. This is a discrete analogue of a result by Ghys (see [OT1] and references therein). To
see a connection to the four vertex theorem, consider a polygon Q R2 with vertices v1 =
(x1 , y1 ), . . . , vn = (xn , yn ). For every i, consider a unique projective linear transformation
i : RP1 RP1 , which maps xi1 , xi and xi+1 into yi1 , yi and yi+1 , respectively. Denote
by the graph (x, (x)) of function . Graph now plays the role of circumscribed circles
since by denition it contains vi1 , vi and vi+1 . Now check that i i is a local minimum
or maximum is the cyclic sequence if and only if the vertices vi2 and vi+2 of Q lie on the
same side of .
21.13. Part a) is a classical result of Szeg
o (1920) and S
uss (1924). We refer to [Ball] for
the related results and the references.
22.1. This proof is given in [A2, 9.1]. For a), see Figure 36.1. For b), take the signed area
of triangles spanned by the origin and the edges of P . For c), translate P so that the origin
in inside and dierentiate the area:
2 area(Pt ) = (h1 1 + . . . + hn n ) + (h1 1 + . . . + hn n ).

On the other hand, area(Pt ) = h1 1 + . . . + hn n by denition of the area, which implies

that area(Pt ) = h1 1 + . . . + hn n . Since area(Pt ) = 0 by assumption, and hi > 0 by
construction, we conclude that some i must be negative.
For d), suppose vertices v and w separate the increasing and decreasing edges in P . Place
the origin in the intersection of the lines supporting P at v and w. Now all hi i have the
same sign in the expression for area(Pt ) , a contradiction.
22.3. Two proofs are given in [OR].
22.4. In part a) use the arm lemma (Lemma 22.3) and in b) use the extended arm lemma
(Exercise 22.3). The proofs are given in [OR, ORS].
23.1. Consider separately all pairs of convex polygons in Y and then take a sum.
23.3. Part a) is a classical result on crossing numbers which can be found in [PacA, 14]
and [Mat1, 4.3]. Part b) was proved in [AloG].
23.5. For b), see [Hal4]. For parts c) and d), see [Oza1].
23.6. This is studied in [Oza].
23.7. An involved version of this result is proved in [BolG], where it was proved by using
the cut locus construction and deformations of polygons.
23.8. For a), an elegant smooth example is given in [Mor]. For b), see [Heil].
23.9. This is proved in [Hal3] via reduction to the smooth case.
23.10. This exercise follows closely [Carv].
23.11. The rst explicit construction of a stuck unknot (space hexagon which cannot
be rigidly atten) was constructed in [CanJ]. Another example is given in Figure 42.13
(see [Tou]).
23.12. This is proved in [CocO].

397

Figure 42.13. A stuck unknot.

23.13. This is a classical Erd
os problem. We refer to [DGOT] for an overview of the history,
several proofs and references.
24.1. a) Of all (n 1)! polygons on X, let Q have the shortest length. If two edges of Q
intersect, they can be replaced by two others to make a polygon of shorter length. This
implies that Q is simple. For b), project X onto a generic plane and take a polygon as in
part a). For c), project X and take the rst triple to form the rst triangle, the second
triple to form the second triangle, etc. For d), this is true in fact for all embeddings of K6
into R3 , not necessarily by straight lines (see [Ram]). The easiest way to prove this is by
deforming the conguration and checking that the number s of pairs of linked triangles is
always odd. Indeed, for every two edges there are two way these edges can connected to the
remaining two points. When two edges cross, both pairs of triangles change from linked to
unlinked or vice versa, so the parity of s is unchanged.
24.2. For two points x, y Q, denote by the shortest path between x and y which lies
inside Q. If (xy)
/ Q, then is not an interval, and thus must contain at least one vertex v
of Q. Now observe that locally around v we must have v > , a contradiction. For
another elementary proof of this classical result see [Cher, 1].

24.3. This is a classical corollary from a number of advanced results in knot theory (see,
e.g., [Ada2]). The following proof is given in [Fox, 7]. Denote by K#L the (connected)
sum of knots K and L (see Figure 42.14). Observe that K#L = L#K. Suppose K is nontrivial and K#L is an unknot. Consider a wild knot Q = K#L#K#L# . . . On one hand,
Q is isotopic to (K#L)#(K#L)# . . ., and thus an unknot. On the other hand, Q is isotopic
to K#(L#K)#(L#K)# . . . = K. Therefore, K is also an unknot, a contradiction.104
24.4. Follow the proof of Fenchels theorem, and use induction on the dimension for the
equality part.
24.5. For this result, generalizations and references see [Sant].
24.7. See [Sant, 7.5].

24.8. For a), this follows from the fact that the average length of a random projection has
length |Q|/ (see Exercise 24.6). See [Rama]. Parts b) and c) are presented in [Sul, 8].

24.9. For a), if A is a connected component of sh(Q) and x conv(A), then every plane H
through x intersects A at some point a. Since H goes through a, it contains at least four
points in Q, which implies that A is convex. The closure of A is a convex polytope by
piecewise linearity.
104The

reader may consider this method a cheating for we do not give a delicate convergence
argument implicitly used in the proof. There are several known discrete proofs of this exercise, but
none are elementary and suciently succinct to be included here.

398

For b), there are two cases. Either L has a knotted polygon, in which case the second
hull theorem implies the result. Otherwise, L has two polygons Q1 , Q2 whose convex hulls
intersect. The result now follows from conv(Q1 ) conv(Q2 ) sh(L).
24.10. For a), take the sum K#L of two non-trivial knots K and L as shown in Figure 42.14.
While sh(K), sh(L) 6= , the middle points x
/ sh(K#L).
For b), take an unknot Q and x as in the gure and observe that x sh(Q). For c),
deform Q appropriately.

K#L

x
Q

Figure 42.14. Sum of knots K#L with a disconnected second hull and
an unknot Q with a nonempty second hull.
24.11. This is proved in [CKKS, 6].
24.12. A stronger form of this result was proved in [Izm].
24.13. This inequality is due to A.D. Alexandrov (1947). See [Res2] for the proofs and
references.
24.14. Part a) goes back to Radon (1919) and is completely straightforward: the curvature
minimizes when C is convex, in which case, when the projection is simple, it becomes
an equality. Part b) is a corollary of Exercise 25.2 applied to the cone. For advanced
generalizations and references see [Res1].
For c), consider a point z sh(Q) and take Q to be the result of a translation of Q which
moves z to O. By denition of the second hull, all generic hyperplanes through O intersect Q
four or more times. Now use Croftons formula (Lemma 24.6) and the proof of Fenchels
theorem (Theorem 24.4) to conclude that (Q) (Q ) 4. We refer to [CKKS] for the
sketch of this proof and related results.
24.15. Let S be the set of midpoints of (x, y), where x, y Q. Following Subsection 5.3,
the surface S consists of triangles and parallelograms with the boundary at Q. Since Q is
knotted, the interior of S intersects with Q, giving the desired triple.
24.16. For a), given a point x Q consider all triangles spanned by x and edges of Q.
Denote by S the resulting surface. If S is embedded, then Q is unknotted, a contradiction.
For b) and c)
If two triangles T1 , T2 S intersect, the intersection is the desired line.
see [Den, Kup1].
24.17. Part a) is due to F
ary and has been generalized in various directions [BZ3, 28]. Three
elementary proofs are given in [Tab5]. Part b) is proved in [Cha], and part c) in [LR, NazP].
24.18. For b) and c), see [CSE]. For d), use a small circle C2 .
24.19. The net total curvature and the corresponding Croftons formula are proved in [Gul].
Part b) is proved in [AbrK] (see also [FSW] for a simple proof).
24.20. The idea is to adapt the proofs of the Fenchel and F
aryMilnor theorems in this
case [Gul].

399

24.21. Part a) was proved in [Bes1]. For b), take three symmetric geodesic arcs from the
North Pole of length and add a small geodesic equilateral triangle around the South
Pole. The resulting net has total length 3 + O(), where > 0. Part c) is a generalization
of a) given in [Crof].
25.3. Region A contains all vertices of P , has n = 2 |E| vertices, while A has r = |F|
connected components. Observe that every angle of A is a complement of the corresponding
face angle to 2, which gives
X
X
(A) = n (2)
(v) = (n |V |) 2 +
(2 (v))
vV

= 2 (2|E| |V |) +

vV

vV

(v) = 4 |E| 2 |V | + 4 .

On the other hand, by the previous exercise, we have (A) = (2|E| + 2|F| 4) . Together
these two equations imply Eulers formula.
25.5. For a), consider the dual cones C and D . Use the supporting hyperplanes to show
that C D . Conclude from here that (C) = (C ) (D ) = (D ).
25.6. For a), note that the centrally symmetric faces have at least four vertices. Now use
Eulers formula. For b), consider the Minkowski sum of n vectors in convex position (take
e.g., the side edges of a symmetric n-pyramid). The resulting zonotope has at least n faces,
all parallelograms (see Exercise 7.16).
25.8. For b), here is a solution from [Grub, 15.1]. Denote by ai the number of vertices of
degree i, and by fi the number of i-sided faces. For the numbers n, m and k of vertices,
edges and faces, respectively, we have n = a3 + a4 + a5 + . . ., k = f3 + f4 + f5 + . . ., and
2m = 3a3 + 4a4 + 5a5 + . . . = 3f3 + 4f4 + 5f5 + . . . By Eulers formula, we conclude:
8 = (4n 2m) + (4k 2m) = (4a3 + 4a4 + 4a5 + . . . 3a3 4a4 5a5 . . .)
+ (4f3 + 4f4 + 4f5 + . . . 3f3 4f4 5f5 . . .) a3 + f3 .

25.9. For a), observe that the curvature of all vertices are bounded: i /2. Now the
GaussBonnet theorem implies the result. Part b) is analogous. Note that both bounds
are tight: take the cube and the icosahedron, respectively.
25.10. Use Girards formula to write each solid angle in terms of dihedral angles. Summing
over all vertices, each dihedral angle will be counted twice. Then use Eulers formula.
25.11. This is a dual result to the GaussBonnet theorem (Theorem 25.3).
25.12. (ii) (i) Write out equations for the edge lengths.
(iii) (i) From the GaussBonnet theorem (Theorem 25.3), conclude that the curvature
of each vertex is equal to . Now write out the equations for the angle sums around each
vertex and inside each face.
(iv) (i) Use an argument based on Subsection 25.3.
(v) (iv) Write out equations for the solid angles in terms of dihedral angles.
25.13. This follows from Eulers formula. See Subsection 26.3 for the proof.
25.14. This is a restatement of Eulers formula. It is proved in [Gla] (see also [GGT, 2]).
25.15. Both parts easily follow from Exercise 25.14. For a), if there are no sinks, sources
and faces with oriented cycles, then the index of every vertex and face is 0, while by the
formula in the exercise, their sum is zero, a contradiction. Alternatively, one can prove
part a) in the same manner as Lemma 32.7. Since there are no sinks in the graph, every
oriented path can be extended until it self-intersects. If the resulting cycle is not along a

400

face, start moving inside the cycle until a smaller cycle (in the area) is obtained. Repeat
until the desired face cycle is found. This simple result seems to have been rediscovered a
number of times, but [Gla] is a denitive reference.
25.16. Take a triangular bipyramid where the top and bottom vertices have curvatures
and (2 ), respectively, for > 0 small enough.
25.18. a) Let C be a convex cone with n faces and acute dihedral angles i , 1 i n.
The sum of dihedral angles is equal to the sum of face angles of the dual cone C , i.e.,
n
X
i=1

i = n (C ) (n 2).

If i < /2, then n = 3 and all face angles of C are acute. The only polygons with acute
angles are triangles. We conclude that P is simple and all faces are acute triangles, i.e., P
is a tetrahedron.
b) By the same argument, P is simple and all faces are either non-obtuse triangles or
rectangles. Now apply Eulers formula.
c) This is false. Consider a prism of height h and two sides regular n-gons with side 1.
Connect adjacent midpoints as in Figure 42.15 and remove 2n resulting tetrahedra. Note
that the dihedral angle between triangles and n-gons decreases continuously with h
(0, ), from to /2. Take h to be such that = 3/2. It remains to check that the
dihedral angle between triangles and rhombi 3/2 as n . A more involved
example is given in [Pach] (see Exercise 40.12).

Figure 42.15. Truncated prism with dihedral angles 3/2.

25.20. See [Gr
u4, 13.3] for the elementary proof of b) and some background.
26.1. Parts b) and c) are given in [Karc].
26.2. This was proved in [Schl2].
27.1. See [DSS2].
27.2. Part a) is false. Take = (v1 v2 v3 v4 ) with |v1 v2 | = |v1 v3 | = |v4 v2 | = |v4 v3 |, an
obtuse dihedral angle at (v2 , v3 ) and equal other dihedral angles. This proof is given by
N. B. Vasiliev and V. A. Senderov in [Kvant], M1301 (1992, no.3). For b), observe that
equal dihedral angles imply equal face angles.
27.3. By denition, the local convexity condition divides all vertices into concave and convex
(via orientation of the cone at a vertices). Note that the edges in convex cones have dihedral
angles < , while the edges in concave cones have dihedral angles > . Thus, no edge can
connect concave and convex vertices. The connectivity of the edge graph implies that all
dihedral angles are < . Now use the local convexity criterion (Lemma 14.7).
28.1. To remove the second assumption, attach to two polygons the same big spherical
triangle so that the obtained polygons satisfy conditions of the theorem.

401

28.2. a) Let Pt = P + Bt , where Bt is a ball of radius t, so that P0 = P . Observe that

X
area(Pt ) = area(P ) + t
e ( e ) + 4 t2 .
e

On the other hand, since Pt BR+t , we have area(Pt ) area(BR+t ) = 4 (R + t)2 .

Therefore,
X
area(P ) + t
e ( e ) + 4 t2 4 (R + t)2 .
e

Letting t , we obtain the result.

b) Dene Pt = P + Bt , Qt = Q + Bt , and use area(Pt ) area(Qt ) as t .
28.4. See [CSE].
28.6. For a), move a vertex of a tetrahedron along one of the edges. Check the formula by a
direct calculation. Obtain the general formula from an (innitesimal) composition of such
transformations. For b), triangulate the polyhedron enclosed by the surface, apply a) and
use the additivity of the Schl
ai formula.
28.8 and 28.9. See [Kor2, Kor1].
30.1. For a), let us rst prove that tightness implies both conditions. Suppose the edge
condition fails for the edge e = (v, w) conv(P ), where v and w are vertices in P . Then
the plane along e and separating from other vertices of P divides P into at least three
parts. Similarly, if the vertex condition fails at v
/ conv(P ), we can separate v from other
vertices by a plane near v. Now observe that every such plane divides P into three or more
parts. In the opposite direction, suppose plane L intersects P and creates three or more
connected components. Choose a side with two or more such components. Recall that the
graph of edges of conv(P ) lying on one side of L is connected (see Exercise 8.3). If all
components contain vertices in conv(P ), then by the edge condition all components must
be connected to each other via the edges of conv(P ), a contradiction. Finally, if Q P is a
component without vertices in conv(P ), then the furthest from L vertex v Q violates the
vertex condition. This result is due to Bancho, while our proof follows [Kuh, 2].
b) The edge condition is clearly insucient as evident from a non-convex bipyramid in
Figure 30.1. Similarly, the vertex condition holds for a non-tight Jessens icosahedron (see
Exercise 19.17).
d) If an intersection Q = P L of P with a plane L is disconnected, then P has at least three
connected components, since n curves separate the surface homeomorphic to a sphere into
n + 1 components. Now, if every intersection is connected, the interior of A is a contractible
polygon, and the convexity criterion (Exercise 1.25) implies the result.
30.2. See [Spi] for one such construction.
30.3. Remember, it is very hard to make a exible polyhedron!
30.4. The rst part remains an open problem despite numerous attempts (and incorrect
constructions). The second part is a conjecture of Stachel [St1].
31.2. See [FedP, 7.5].
31.3. Parts a), b), c) are due to Robbins, who also conjectured parts d) and e, which were
proved in [FedP]. See [Pak4] for a survey and [Var1] for an independent approach.
31.4. See [Con1].
31.5. For a), let D be the square length of the diagonal and let R be the squared radius.
Clearly, D/4 + R = L is a constant given by the squared side edge length. Now make a

402

substitution of (L D/4) into the Sabitov polynomial for R. For b), use the fact that Zn
has exponentially many realizations. For c), further details and references see [FedP].
31.6. a) Use 12 = 23 = 13 = 2 and 01 = 02 = 03 = 1 + , for > 0 small enough. b)
Consider , R3 with edge lengths

01 = 02 = 03 = 1 ,
13 = 23 = 2 ,
12 = ,

01 = 02 = 03 = 1 ,
12 = 23 = 2 ,
13 = ,

where > 0 is small enough. Check that , with such edge lengths exist, and that there
is no tetrahedron with (ij + ij )/2 edge lengths. For c) and the references see [Riv].
34.8. Parts a) and b) are proved in [Mon1]. This technique was later extended in [Mon2]
to prove c) and d). Parts e) and f ) are proved in [Pok]. See [Ste2] for a survey.
32.2. This result is due to Pogorelov [Pog6].
32.3. From the contrary, suppose there exists a continuous deformation of C . Denote by G
the graph of P . As in the proof of the Cauchy theorem, consider a subgraph H G of
edges where the dihedral angles change. Check that every vertex of H has degree at least 4.
Use Exercise 25.7 to conclude that at least eight triangular faces of H (as a planar graph)
are triangular. In the other direction, show that even-sided faces of G cannot subdivide
these triangles. The second part is similar. See [DSS1] for details and references.
32.4. This can be proved using the combinatorial approach in Subsection 32.3. See [A2,
10.3] for the original proof.
33.2. Consider the following deformation {Pt } of the coordinates of Jessens orthogonal
icosahedron:






2, (1 + t), 0 ,
(1 + t), 0, 2 ,
0, 2, (1 + t) .
Observe that the lengths of the long edges are unchanged under the deformation, while the
lengths of the short edges are
p
p

(1 t)2 + (1 + t)2 = 2 + 2t2 = 2 + O(t2 ) as t 0.

This implies that the polyhedron P0 is not innitesimally rigid. Continuous rigidity follows
from the second order rigidity. First, check that the rigidity matrix has corank 1, i.e., the
above rst order deformation is the only possible (up to scaling). Then check that the
second order terms implied by O(t2 ) are strictly positive, for all t 6= 0.
33.5. Part d) is a proved in [Con3]. See [Con5, ConS, IKS] for more second order rigidity.
33.7 and 33.8. See [RotW, 6, 7].
33.10. See [Luo] for the proof, extensions and references.
31.7. For a), consider normals u i to the faces. Observe that u i u j = ij . Now take
w i = area(Fi ) u i , where Fi is the face of which does not contain vertex vi . Recall
that w 1 + w 2 + w 3 + w 4 = 0 (see Proposition 36.1). Thus there exists a closed space
quadrilateral Q with edges w i , and the sum of angles in Q is 2. Since the sum of angles
in Q is equal to 12 + 23 + 34 + 14 , we obtain the result.
For b), observe that 12 + 13 + 14 > , since the curvature of the dual cone (C1 ) =
( 12 ) + ( 13 ) + ( 14 ) < 2. Adding such inequalities for all four vertices we
obtain the lower bound. The upper bound follows by averaging of the inequalities in a).
For c), we have:
X
X
1
|r |2
cos ij =
hu i , u j i = 2 hu 1 + u 2 + u 3 + u 4 , u 1 + u 2 + u 3 + u 4 i = 2
,
2
2
ij

ij

403

where r = u 1 + u 2 + u 3 + u 4 . Since |r | 2, we immediately obtain both lower and upper

bounds. For d), observe that the equality holds only if r = 0. From the proof of the
previous exercise, we conclude that is equihedral.
Parts a)-d) are given by I. F. Sharygin in [Kvant], M353 (1976, no. 7). For part e), show that
one can monotonically increase the dihedral angles to obtain dihedral angles of a Euclidean
polytope in the limit.
33.11. This is a small special case of Schramms theorem in [Schra]. The proof uses an
advanced generalization of inversions and a similar inversion counting argument.
34.1. For part a), see [Hada, 255]. For part b), see [FedP, 8].
34.2. For b), write both sides of the Cayley-Menger determinants mod 8. Part c) was on
a Putnam Mathematical Competition in 1993.
34.3. For a), take two 4-tuples of points in the plane:

|a1 a2 | = |a1 a3 | = 2, |a2 a3 | = 2, |a2 a4 | = |a3 a4 | = 10, |a1 a4 | = 4.

|b1 b2 | = |b3 b4 | = 2, |b1 b3 | = 2, |b1 b4 | = |b2 b3 | = 10, |b2 b4 | = 4.

For b), x the rst coordinates of all points ai = (xi , yi ), 1 i n. For every six distances
which can potentially form a at 4-tuple, write out their Cayley-Menger determinant. Check
that the resulting polynomial equations on yi are nontrivial unless these distances form the
usual 4-point pattern. Conclude that for almost all values of yi the set {ai } has no equivalent
and non-congruent sets. See [BouK] for this and further results.
34.6. This was rst proved in [Var1]. A simple proof using the theory of places is given
in [Con6] (see also [Pak4]).
35.1. For c), use the fact that (C) is monotone on cones C. See [Pog3, 7.7] for the details
and further generalizations.
35.5. Take a set of points b1 , . . . , bk such that polygon B = [b1 . . . bk ] contains all ai in its
relative interior. Let Z = B. For every > 0, use use Theorem 35.7 to obtain P on
n + k vertices ai and bj , with boundary Z and curvatures 1 , . . . , n . Let P be a
shifted P with the sum of heights of interior vertices equal to 0. Check that P converges
to the desired convex cap as 0.
35.8. For a), we implicitly use combinatorics of P when lifting vi , since for > 0 small
enough the combinatorics is unchanged. For a greedy variational algorithm, start with the
upper surface of the convex hull conv(Z). Take the smallest i such that i < i . Lift it up
until the equality is reached. Repeat. For b), presumably the exponential upper bound on
the number of full triangulations must play a role (see Exercise 14.7).
35.9. Find an appropriate direction of rays and use Theorem 35.4. This result is obtained
in [Zil] (cf. [Thu, 7]).
36.2. Compare the corresponding faces in polytopes P , Q = 12 (P + P ) and P . Since
the normals to the corresponding faces are equal, these polygons are parallel (have parallel
corresponding edges). Here we allow zero edge lengths in case the parallel edge is not
present. Observe that the edge lengths in Q are the averages of the corresponding edge
lengths in P and P . Write signs (+), (0) and () on the edges of Q (including zero length
edges), if their edge lengths in P are greater, equal or smaller than that of edges in P ,
respectively.
Note that there are two types of faces in Q: those coming from faces in one polytope and
a vertex from another, and those coming from one edge in each. Apply Alexandrovs lemma
to show that around each face of Q, either all signs are zero or there are four sign changes.

404

Take the dual graph to the graph of Q and keep the same signs, so that there are at least
four sign changes around every vertex. Modify counting argument in Subsection 26.3, to
conclude that all signs must be zero.
36.3. In notation of Subsection 36.2, dene a map : P Z. Prove that Z W+ is a
convex polyhedron by a direct argument (cf. the proof of Lemma 35.3). Similarly, prove
directly that P is connected. Consider only polytopes P P of volume 1. In the topological
lemma, the injectivity follows from the Alexandrov theorem and the continuity is trivial.
To show that is proper, check that the limit polyhedron cannot become unbounded or
degenerate. A complete proof is given in [A2, 6.3].
36.5. Note that in the proof we use the equal facet areas only to avoid Fk Fk , and other
conditions can be used in its place. Use monotonicity of the perimeter, surface area, and
the mean curvature (see Exercises 7.11 and 28.2).
36.9. For a), consider normals u i to the faces and use the equation in Proposition 36.1
to conclude that their sum is zero. Check that the resulting 2-dimensional family of
{(u 1 , u 2 , u 3 , u 4 )} (up to rotations around O), is the same as the family of normals of
equihedral tetrahedra.
For b), recall that equihedral tetrahedra have equal opposite edges (see Exercise 25.12).
From a), this implies that all non-adjacent edges have equal length. This easily implies the
result.
36.10. Much of this exercise is due to Shephard (1963). See proofs and references in [McM1,
Mey]. Part e) is given in [Kall].
36.12. For a), translate P by a vector w . We have:
*
+
n
n
n
X
X
X

z1 + hw , u i i ai =
z1 ai + w ,
ai u i = (P, P ),
(P, P + w ) =
i=1

i=1

i=1

since the second term is zero. See [A2, 7.2] for the rest of this exercise.
36.13. Here (P, P ) is the usually called the mixed volume vol(P, P , . . . , P ), and the
exercise gives an equivalent way to state the BrunnMinkowski inequalities. This approach
leads to the further inequalities, as discussed in [BonF, BZ3, Schn2].
36.14. By the uniqueness part of Theorem 36.2, polytope P is a translate of (P ). This
implies the result.
36.15. A concise proof of part a is given in [Grub, 18.2].
36.16. For d) and e) see [A2, 7.1]. The truncated octahedron tiles the space and has 14
faces (see Exercise 16.6). The upper bound in f is due to Minkowski. We refer to [Grub,
32.2] for the proofs, references and the context.
36.18. See [BezB].
36.19. This example was recently constructed in [Pan].
37.1. For a), cut the surface of a convex cap by a hyperplane and combine two copies to
obtain a convex polytope with reection symmetry. Show that this can be done intrinsically.
For b), see [A2, 4.4].
37.4. See [ADO].
37.5. Part c) was proved by Harer and Zagier in (1986). See combinatorial proofs in [GouN,
Lass].
37.8. For a), it suces to check that the sum of angles around each vertex is < 2. Use the
law of cosines to obtain stability result for the angles (e.g., < 0.01 easily works). For b),

405

check what happens to the convexity of realizations when the inverse function theorem is
applied.
37.10. See [Vol3].
37.11. This is a restatement of an open ended problem communicated to me by Victor
Zalgaller.
38.2. For a), repeatedly use convex bendings in the proof of Leibins theorem (Theorem 38.2). Both parts a) and b) are outlined in the Appendix to [Pog3].
39.3. See [Ble].
39.4. The bendings were constructed in [Mi5].
39.6. See [Pau] and [MlaO].
39.8. Part a) is implicitly stated in the description part of the patent.
39.9. Parts b) and c) are due to Shtogrin in a series of papers [Sht1, Sht2, Sht3]
39.10. For a), we have |xy| = |xy|S1 = |(x)(y)|S2 |(x)(y)| = |(x)(y)|, for all
x, y C1 . Parts b) and c) are announced in [Tas1].
39.11. For a), use induction on the total number of vertices in C1 and C2 . Start at a
vertex v of C1 and consider a ray R starting at (v) such that the sum of angles R forms
with edges of C2 is equal to the angle at v. Extend the partial diagonal inside S1 as far as
possible, until one distance condition inequality becomes an equality. Use this construction
repeatedly until bigger polygons split into smaller polygons. Part a) and b) are outlined
in [Tas1], part c) in [Tas2].105
39.12. For a), once the surface S is folded onto a plane, continue folding it in halves until
it ts a disk of radius . This gives the desired realization.
39.13. See [BZ1] and [BZ4].
40.2. For a), see Figure 42.16. The outline of the rest in given in [Zal2].
v
w

Figure 42.16. Points in the shaded region have exactly two shortest paths
to w (they all go through vertex v). Non-convex polyhedral surface in R3
and shortest paths between points v and w.
40.3. For a), take a tetrahedron = [v1 v2 v3 v4 ] with edge lengths |v1 v3 | = |v2 v3 | = |v1 v4 | =
|v2 v4 | = 1, |v1 v2 | = , and |v3 v4 | = 2 , for > 0 suciently small.
For b), take a tetrahedron = [v1 v2 v3 v4 ] with edge lengths |v1 v2 | = |v1 v3 | = |v1 v4 | = 1,
|v2 v3 | = , and |v2 v4 | = |v3 v4 | = 2 , for > 0 suciently small.
40.4. For a), see [AHKN].
105The

inductive proof of b) in [Tas1] has a serious gap.

406

40.5. For a), see [PakP]; for b) see [Pinc].

40.6. The original proof of a) by D. M. Mount is unpublished and has a crucial gap. The
full proof of both parts is given in [MilP].
40.7. For b), see Figure 42.16 above. For a) and c) see [Aga+]. For a discussion on d)
see [MilP].
40.8. See [Aga+, ShaS] for the proofs and [SchrS] for further references.
40.9. The proof in [AroO] uses induction on the number of vertices by deforming the peels
(see Exercise 40.1) and using the Alexandrov existence theorem.
40.10. Use the higher dimensional generalization of the Mount lemma (Exercise 40.6) to
generalize Lemma 40.5. See the complete proof in [MilP].
40.11. Part a) is proved in [CanR] who proved it for equilateral triangles Qi . Reduction b)
is proved in [MilP, 9].
40.12. For a), consider an obtuse triangle and points x, y on side edges close to acute vertices
(see Figure 42.17), and take a tall prism with this triangle as an equator. For b) and a
partial result in the direction of d), see [BKZ]. Part c) was a problem of Pogorelov resolved
in [Pach] (cf. Exercise 25.18). Part d) is a conjecture of Har-Peled and Sharir (1996), still
open.

Figure 42.17. Shortest path between x and y.

40.13. The proof is outlined in [Vol2].
40.14. For a), use the idea in Example 40.11, i.e., glue the halves of the cube separated by
the plane through diagonals in the top and bottom faces. Same for the other diagonal. The
resulting 4-layered surface is already at. For the general surfaces, see the next exercise.
40.15. Here is a sketch of my solution of c). Use the realization of S given in the proof
of Lemma 37.5. Give an ad hoc construction of a layered (continuous) folding of three
faces of a tetrahedron onto the plane containing the fourth one, such that the boundary
is unchanged. Bend the folding so it ts inside triangle. Repeat until the whole surface is
folded. See [BerH] for another approach.
40.16. For a), observe that the interval [ab] is smaller than the covered part of the boundary,
which is a polygon between a and b. For c) see constructions in [Yas] and [Lang, 9.11].
Although stated in a continuous form, it is unclear if these constructions can be used to
obtain d), as the authors do not prove the isometry of intermediate stages. The most general
part e) was recently proved in [Tar2].
40.17. For a), place a small cube in the middle of the top face of a bigger cube. For b),
take a truncated cube (see Figure 16.4) and attach to each face a very tall triangular
pyramid [Tar1] (see also [Gr
u3] for further references).
40.20. See http://tinyurl.com/6jm5az .
42.1. Three simple proofs can be found in [Vard, pp. 1217].
42.2. a) Assume Q = [x1 . . . x5 ] R3 is regular and not at. Then the convex hull conv(Q)
is either a bipyramid or a pyramid over a 4-gon. Either way, we can always assume that

407

x1 and x4 lie on the same side of the plane spanned by x2 , x3 and x5 . By the assumptions,
all diagonals in Q have equal length. Thus tetrahedra (x4 x2 x3 x5 ) and (x1 x2 x3 x5 ) are
congruent. Since x1 and x4 lie on the same side of the isosceles triangle (x2 x3 x5 ), we
conclude that x1 and x4 are symmetric with respect to the plane bisecting (x2 , x3 ). Thus
points x1 , x2 , x3 and x4 lie in the same plane. The same argument shows that x5 , x2 , x3
and x4 lie on the same plane, and we obtain a contradiction. See [PraS], Problem 8.29.
b) For even n, consider the regular n-antiprism and a polygon obtained by the side edges.
For n odd, nd a non-planar regular 7-gon and use induction. To see why this is true for
odd n large enough, consider a long metal chain of small equilateral triangles attached as
in the regular antiprism, but make the chain open and exible along the edges. Clearly, we
can twist it and close it up if the chain is long enough.
42.4. This result is due to P. W. Jones (1990). We refer to [KenK] for an elegant presentation
and references.
42.5. Take the longest edge. See [Shar], Problem 170.
While there is a large literature on the subject, there is no standard name for these tetrahedra. We refer to [Arn3, pp. 2, 188191] for the references, and to [PraS, 6.4] (see
also [Shar], Problem 304) for simple proofs of these and other basic results on equihedral
tetrahedra. Let us mention that these tetrahedra are also called isosceles (see [Lee, The])
and equifacial (see [HajW]). See also Exercise 10.4 for more on this.
42.8. Consider a hyperplane H Rd dened by the equation
x1
xd
+ ... +
= d.
a1
ad
Clearly, a H. Since a is an interior point, there exists a vertex v of P on the same side
of H as the origin O. We conclude:
vi
v1
vd

+ ... +
d.
ai
a1
ad
This solution follows [SGK, 3.2].
42.10. See Problem 266 in [VasE].
42.11. The upper bound (2n 4) is sharp. See the solution to Problem 8.6 in [PraS].
42.12. This is false (see [PraS], Problem 13.12). Check that two vertices on the same edge
can lie in C. Extend the planes of C to divide the space into 8 cones. Conclude that at
least one of them will have two vertices of the cube.
42.13. Part a) is proved in [Klee]. Part b), was originally proposed in [She2] (see also [CFG],
section B10) and completely resolved in [CEG]. For c), start with a regular tetrahedron
and add a vertex at distance from the barycenter of every face. Then add a vertex at
distance 2 from the barycenter of every new face, etc. Repeat this O(n) times.
42.15. For a), the claim is false. Start with a unit circle C R3 dened by x2 + y 2 = 1
and z = 0. Take a cone over C from a = (1, 0, 1) and b = (1, 0, 1). The resulting
convex body B has extremal points at C (1, 0, 0) and a, b. To make only countably
many extremal points, modify this example by taking a convex hull of a, b and a converging
sequence to (1, 0, 0) of points in C. We are unaware of an elementary proof of b).
42.17. See [Zon2] for proofs, generalizations and references. For b), see also a hint on p. 74
in [Tao].
42.18. This problem was given at the Moscow Math. Olympiad in 2008. See http://tinyurl.com/ma34td
for a complete solution.

408

42.19. a) Take a projection of a tetrahedron onto a plane L containing face (abc). Assume
the remaining vertex d is projected into the interior of (abc). Choose two lines 1 = (ab)
and 2 1 . Rotate by an angle > 0 around 1 and then around 2 . The new projection
is a triangle (a b c ) that ts inside (abc), for small enough.
b) For the p
unit cube: the unit square projection ts inside the regular hexagonal projection
(with side 2/3).
c) For the regular octahedron with edge length 1: start with a unit square projection.
Rotate the octahedron around a diagonal of the square by an angle > 0. Rotate the
octahedron now around the second diagonal by an angle 2 . The resulting projection ts
inside the unit square, for small enough.
d) Although we have not checked, the regular icosahedron looks like a potential counterexample.
42.20. Part a) was proved by Kovalev (1984), while part c) was proved in [DML]. See [K
osT]
for a simple proof of a) and the references.
42.21. a) For Cd = {(x1 , . . . , xd ), 0 xi 1}, take a hyperplane x1 + . . . + xd = 2.
For b), c) and d), see [JocP].
42.22. Use the fact that K3,3 is not planar. See [Sany] for details and generalization.
42.23. See [Pol]. The example of cuboctahedron shows that the bound is tight.
42.24. This is called the Szilassi polyhedron. It is easy to see that this is possible only if it
has 14 vertices and 28 edges. See http://tinyurl.com/2qqzgc for the references.
42.25. This construction is called the Cs
asz
ar polyhedron [Cs
a].
42.26. Let = (v1 v2 v3 v4 ) and suppose (v1 v2 ) is the longest edge. We have:








2|v1 v2 | |v1 v3 | + |v2 v3 | + |v1 v4 | + |v2 v4 | = |v1 v3 | + |v1 v4 | + |v2 v3 | + |v2 v4 | .
Therefore, either |v1 v2 | |v1 v3 |+ |v1 v4 |, or |v2 v1 | |v2 v3 |+ |v2 v4 |, as desired. The problem
comes from the Tournament of Towns in 1994 (see http://tinyurl.com/lk6he3)
42.27. An elegant proof of this is given in [GriL, 2].

42.28. Let v i =
x
1 xi and w i = xi xi+1 , for all 1 i n. Clearly, v i = w 1 + . . . + w i1 . By
analogy with the volume formula in Subsection 34.1, we have 2area(X) = |u|, where
u =
=

n
X
i=1
n
X
i=1

v i v i+1 =

n
X
i=1

v i (v i + w i ) =

(w 1 + . . . + w i1 ) w i =

X
i<j

n
X
i=1

vi wi

wi wj .

Taking the norm of both sides immediately implies Sarrons formula.

42.31. See [Eri].
42.32. For this and other properties of the polar sine function see [LerW].
42.33. This is a result of Lagrange (1783). See [GooT] for a short proof.
42.36. See [Kah].
42.38. For b), this is false. Take a quadrilateral inscribed into a circle with O as a center.
Now permute triangles spanned by O and edges.
42.39. Part a) is usually attributed to Euler (1765), while parts b)-e) are due to L. FejesT
oth (1964). We refer to [KlaT] for references and a simple proof of b).
42.40. See [BolS, 22].

409

42.41. See the Math. Overow question http://tinyurl.com/yd998rw

42.42. For a), note that no cube can contain more than two vertices. For b), take the
smallest cube H1 on the bottom plane. Take then the smallest cube H2 on top of H1 , etc.
For c), see [Tut, 1] and [Yag1].
42.43. A complete solution is given by Vsevolod Lev in [Kvant], M607 (1980, no. 11).
42.44. For elementary proofs of b), see [Yag1, 4.5] and [Tuza]. For c), see [Sche]. For d)
and e), see [Kais]. Finally, for d), see [SST, 11].
42.45. For a), take two copies of four pairwise adjacent tetrahedra as shown in Figure 42.18.
Join them along the subdivided triangular faces after slightly twisting them. The bound in
part b) is not sharp and can be improved to 9 (see [Zak2]). Part c) is due to Tietze (1905),
while part d) (long conjectured to be false) was recently established in [EriK].

Figure 42.18. Four pairwise adjacent tetrahedra and the twisting direction.
42.46. See [Boll], Problem 8.
42.47. For a)c), one can always extract along a line: f (t) = t u for some unit vector u.
We prove part a). Dene Pi Pj if there is a point in Pi which lies directly below a point
in Pj . Check that denes a partial order on Pi . Now the largest element can be moved
upward.
d) Tile a plane L with unit squares. Place the regular tetrahedra with side 2 around L, one
per square, so that vertices of squares are midpoints of edges. There are two ways to t a
tetrahedra around a square then; use both ways alternatively, in a checkerboard fashion, so
that the tetrahedra t together. Shrink each tetrahedron around its center by , for some
small > 0. Check that for suciently small , none of these tetrahedra can be extracted
(by symmetry, it suces to check this for one tetrahedron).
e) To construct a nite family, take a 2m 2m square S region on L. Attach the opposite
sides of S and embed it into R3 . By taking m large enough, the resulting torus can be made
as at as desired, so that the images of the squares are at -distance from the unit squares.
Now arrange the tetrahedra as in part b). If is much smaller than , the tetrahedra will
be non-intersecting and neither one can be extracted.
f ) A similar construction, but in this case start with a tiling of L with regular hexagons.
Take a quotient S of L by two long vectors which preserve the grid. We omit the details.
Parts a)c) are given in [Daw1]. Construction in d) is given in [Kan] (see also [DEKP]).
The rst construction for e) is given in [Daw1] (see also [SnoS]). The constructions in e), f )
are new and based of innite constructions in [DEKP, Kan]. Part g) is open. For part h),
references and pictures see [SnoS].

410

References
[AbrK] J. Abrham and A. Kotzig, Transformations of Euler tours, Ann. Discrete Math. 8 (1980),
6569. [398]
[Ada1] C. C. Adams, Tilings of space by knotted tiles, Math. Intelligencer 17 (1995), no. 2, 4151.
[390]
[Ada2] C. C. Adams, The knot book, AMS, Providence, RI, 2004. [240, 397]
[Adl]
A. Adler, Die Theorie der geometrischen Konstruktionen, G. J. Goschensche Verlagshandulung, Leipzig, 1906; available at http://tinyurl.com/2kyaw3; (J). [116]
[Aga+] P. K. Agarwal, B. Aronov, J. ORourke and C. A. Schevon, Star unfolding of a
polytope with applications, SIAM J. Comput. 26 (1997), 16891713; available at
http://tinyurl.com/62dda4 [358, 406]
[AFF] F. Aguilo, M. A. Fiol and M. L. Fiol, Periodic tilings as a dissection method, Amer. Math.
Monthly 107 (2000), 341352; (). [147]
[AigZ] M. Aigner and G. Ziegler, Proofs from THE BOOK (Second edition), Springer, Berlin,
2003. [31, 215]
[AHKN] J. Akiyama, K. Hirata, M. Kobayashi and G. Nakamura, Convex developments of a regular
tetrahedron, Comput. Geom. 34 (2006), 210; [405]
[Alb]
S. I. Al ber, On periodicity problem in the calculus of large variations in the large, Amer.
Math. Transl. 14 (1960), 107172; (). [101, 385]
[ABe]
M. V. Alekhnovich and A. Ya. Belov, The complexity of algorithms of constructions by
compass and straightedge (in Russian), Fundam. Prikl. Mat. 7 (2001), 597614; available
at http://tinyurl.com/3dwerb [116]
` B. Vinberg and A. S. Solodovnikov, Geometry of spaces of constant
[AVS] D. V. Alekseevskij, E.
curvature, in Geometry II, 1138, Springer, Berlin, 1993; (). [272]
[Ale]
R. Alexander, Lipschitzian mappings and total mean curvature of polyhedral surfaces. I,
Trans. AMS 288 (1985), 661678; (). [271, 272]
[ADO] R. Alexander, H. Dyson and J. ORourke, The foldings of a square to convex polyhedra, in Lecture Notes in Comput. Sci. 2866, Springer, Berlin, 2003, 3850; available at
http://tinyurl.com/56c86v [404]
[ABi]
S. B. Alexander and R. L. Bishop, The FaryMilnor theorem in Hadamard manifolds, Proc.
AMS 126 (1998), 34273436; available at http://tinyurl.com/ytgl78 [240]
[A1]
A. D. Alexandrov, Intrinsic geometry of convex surfaces (translation of the 1948 Russian
original), CRC Press, Boca Raton, FL, 2005. [101, 262, 334, 358]
[A2]
A. D. Alexandrov, Convex polyhedra (translation of the 1950 Russian original), Springer,
Berlin, 2005. [215, 248, 256, 262, 271, 300, 321, 328, 334, 369, 396, 402, 404]
[AS]
A. D. Aleksandrov and E. P. Senkin, Supplement to the paper On the rigidity of convex
surfaces (in Russian), Vestnik Leningrad. Univ. 11 (1956), 104106. [277]
[AZa]
A. D. Aleksandrov and V. A. Zalgaller, Intrinsic Geometry of Surfaces, AMS, Providence,
RI, 1967; (); available at http://tinyurl.com/3xvtar [101]
[Ale1] V. A. Alexandrov, Remarks on Sabitovs conjecture on volume rigidity with innitesimal
bending of a surface, Sib. Math. J. 30 (1989), 678684; (). [314]
[Ale2] V. A. Alexandrov, A new example of a bendable polyhedron, Sib. Math. J. 36 (1995),
10491057; (), (). [285]
[Ale3] V. A. Alexandrov, An example of a exible polyhedron with nonconstant volume in the
spherical space, Beitr. Algebra Geom. 38 (1997), 1118; arXiv: math.MG/0111003. [285,
314]
[Ale4] V. A. Alexandrov, Flexible polyhedral surfaces, Quantum 9 (1998), no. 1, 46; (). [285]
[Ale5] V. A. Alexandrov, How one can crush a milk carton in such a way as to enlarge its
volume (in Russian), Soros Educational Journal 51 (2000), no. 2, 121127; available at
http://tinyurl.com/3acu49 [347]

411

[Ale6]
[AleC]
[AlmR]
[Alon]
[AloG]
[And+]

[Ani]
[Arn1]
[Arn2]
[Arn3]
[AroO]
[AroW]
[AtiM]
[Aur]
[AurK]

[BajB]
[Bag]
[BVK]
[Bak]
[Bali]
[Ball]
[Ban1]
[Ban2]
[BanG]
[Bar]
[BaFu]

V. A. Alexandrov, Inverse function theorems and their applications to the theory of polyhedra, to appear in Rev. Math. Math. Phys. [256, 285, 300, 314, 369]
V. Alexandrov and R. Connelly, Flexible suspensions with a hexagonal equator; arXiv:
0905.3683. [291]
F. J. Almgren and I. Rivin, The mean curvature integral is invariant under bending, in The
Epstein birthday schrift, Coventry, UK, 1998, 121; arXiv: math.DG/9810183. [272]
N. Alon, Splitting necklaces, Adv. in Math. 63 (1987), 247253. [38, 379]
N. Alon and E. Gyori, The number of small semispaces of a nite set of points in the plane,
J. Combin. Theory Ser. A 41 (1986), 154157; (). [396]
A. Andrzejak, B. Aronov, S. Har-Peled, R. Seidel and E. Welzl, Results on k-sets and jfacets via continuous motion, in Proc. 14th Symp. Comp. Geom. (1998), 192199; available
at http://tinyurl.com/25nqoo8 [228]
S. S. Anisov, Convex curves in RPn , Proc. Steklov Inst. Math. 221 (1998), 339; available
at http://tinyurl.com/2drfjw4 [378]
V. I. Arnold, Topological invariants of plane curves and caustics, AMS, Providence, RI,
1994. [206]
V. I. Arnold, Topological problems in the theory of wave propagation, Russian Math.
Surveys 51 (1996), 147; (), (). [94, 396]
V. I. Arnold, Arnolds problems, Springer, Berlin, 2004; (). [206, 358, 407]
B. Aronov and J. ORourke, Nonoverlap of the star unfolding, Discrete Comput. Geom. 8
(1992), 219250; (). [358, 406]
B. Aronov and E. Welzl, Identities involving crossings of j-edges, unpublished manuscript
(2004). [228]
M. F. Atiyah and I. G. Macdonald, Introduction to commutative algebra, Addison-Wesley,
Reading, MA, 1969; (). [314]
F. Aurenhammer, Voronoi diagrams-a survey of a fundamental geometric data structure,
ACM Comp. Surveys 23 (1991), 345405; available at http://tinyurl.com/24vkn6q [138]
F. Aurenhammer and R. Klein, Voronoi diagrams, in Handbook of computational geometry,
201290, North-Holland, Amsterdam, 2000; available at http://tinyurl.com/29y2k3b
[138, 206]
E. G. Bajmoczy and I. Bar
any, On a common generalization of Borsuks and Radons
theorem, Acta Math. Acad. Sci. Hungar. 34 (1979), 347350; (). [377]
F. Bagemihl, On indecomposable polyhedra, Amer. Math. Monthly 55 (1948), 411413; ().
[388]
I. Ya. Bakel man, A. L. Verner and B. E. Kantor, Introduction to dierential geometry in
the large (in Russian), Nauka, Moscow, 1973. [277]
M. J. C. Baker, A spherical Helly-type theorem, Pacic J. Math. 23 (1967), 13; available
at http://tinyurl.com/2wq9q5 [377]
M. L. Balinski, On the graph structure of convex polyhedra in n-space. Pac. J. Math. 11
(1961), 431434; available at http://tinyurl.com/33fbdel [82]
N. H. Ball, On ovals, Amer. Math. Monthly 37 (1930), 348353; (). [396]
T. F. Bancho, Non-rigidity theorems for tight polyhedra, Arch. Math. 21 (1970), 416423.
[285]
T. F. Bancho, Global geometry of polygons. I: The theorem of Fabricius-Bjerre, Proc.
AMS 45 (1974), 237241; (). [228]
T. F. Bancho and P. J. Giblin, Global theorems for symmetry sets of smooth curves and
polygons in the plane, Proc. Roy. Soc. Edinburgh 106 (1987), 221231. [206]
I. Barany, A generalization of Caratheodorys theorem, Discrete Math. 40 (1982), 141
152; (). [378]
I. Barany and K. Fukuda, A case when the union of polytopes is convex, Linear Algebra
Appl. 397 (2005), 381388; available at http://tinyurl.com/5ohvoo [378]

412

[BKP]

I. Barany, M. Katchalski and J. Pach, Hellys theorem with volumes, Amer. Math.
Monthly 91 (1984), 362365; (). [378]
[BKZ] I. Barany, K. Kuperberg and T. Zamrescu, Total curvature and spiralling shortest paths,
Discrete Comput. Geom. 30 (2003), 167176; arXiv: math.MG/0301338 [406]
[BarM] I. Barany and J. Matousek, Berges theorem, fractional Helly, and art galleries, Discrete
Math. 306 (2006), 23032313; available at http://tinyurl.com/254ydrs [377]
[BarV] I. Barany and A. M. Vershik, On the number of convex lattice polytopes, Geom. Funct.
Anal. 2 (1992), 381393; (). [328]
[BDE] G. Barequet, M. Dickerson and D. Eppstein, On triangulating three-dimensional polygons,
Comput. Geom. 10 (1998), no. 3, 155170; available at http://tinyurl.com/23caks5
[387]
[Bar1] D. W. Barnette, Projections of 3-polytopes, Israel J. Math. 8 (1970), 304308; (). [387]
[Bar2] D. W. Barnette, Preassigning the shape of projections of convex polytopes, J. Combin.
Theory Ser. A 42 (1986), 293295; (). [387]
[Bar3] D. W. Barnette, A short proof of the d-connectedness of d-polytopes, Discrete Math. 137
(1995), 351352. [25, 82]
[Barv] A. Barvinok, A course in convexity, AMS, Providence, RI, 2002. [5, 81, 176, 384]
[BecB] E. F. Beckenbach and R. Bellman, Inequalities, Springer, Berlin, 1961; (). [369]
[Ben]
D. Benko, A new approach to Hilberts third problem, Amer. Math. Monthly 114 (2007),
665676; (). [147]
[Ber1] M. Berger, Geometry I and II (translation of the 1977 French original), Springer, Berlin,
1987. [215, 248, 369]
[Ber2] M. Berger, Les paquets de cercles, in Dierential geometry and topology, 2364, World Sci.,
River Edge, NJ, 1993. [108]
[BerE] M. Bern and D. Eppstein, Mesh generation and optimal triangulation, in Computing in Euclidean geometry, 2390, World Sci., River Edge, NJ, 1992; available at
http://tinyurl.com/5gg8jw [387, 388]
[BerH] M. W. Bern and B. Hayes, Origami embedding of piecewise-linear two-manifolds, in Proc.
8th LATIN (2008), 617629; (). [406]
[BesE] A. S. Besicovitch and H. G. Eggleston, The total length of the edges of a polyhedron,
Quart. J. Math. Oxford Ser. (2) 8 (1957), 172190. [382]
[Bes1] A. S. Besicovitch, A net to hold a sphere, Math. Gaz. 41 (1957), 106107; (). [399]
[Bes2] A. S. Besicovitch, A cage to hold a unit-sphere, in Proc. Sympos. Pure Math., vol. VII,
AMS, Providence, RI, 1963, 1920. [382]
[BezB] A. Bezdek and K. Bezdek, When it is possible to translate a convex polyhedron in an other
one? Stud. Sci. Math. Hung. 21 (1986), 337342. [404]
[Bez]
K. Bezdek, Sphere packings revisited, Europ. J. Comb. 27 (2006), 864883; available at
http://tinyurl.com/2sxww6 [191]
[Bieb] L. Bieberbach, Theorie der geometrischen Konstruktionen, Birkhauser, Basel, 1952. [116]
[Biel]
A. Bielecki, Quelques remarques sur la note precedente, Ann. Univ. Mariae CurieSklodowska. Sect. A. 8 (1954), 101103. [381]
[Bin]
D. Binger, Polyhedra with unguarded interiors, Int. J. Comput. Geom. Appl. 12 (2002),
173179; (). [388]
[Bir]
G. D. Birkho, Dynamical Systems, AMS, Providence, RI, 1927; (). [385]
[Bjo+] A. Bjorner, M. Las Vergnas, B. Sturmfels, N. White and G. M. Ziegler, Oriented matroids,
Cambridge U. Press, Cambridge, 1993. [138]

[Bla1] W. Blaschke, Vorlesungen Uber

Dierentialgeometrie, Springer, Berlin, 1929; (). [206,
339]
[Bla2] W. Blaschke, Griechische und anschauliche Geometrie, R. Oldenbourg, M
unchen, 1953;
(). [248, 256]
[Bla3] W. Blaschke, Kreis und Kugel, de Gruyter, Berlin, 1956; (J). [70, 206]

413

[Ble]
[BobI]

[BojF]
[Boll]
[Bolt]
[BolG]
[BMS]
[BolS]
[BolY]
[BonF]
[Bor]
[BorF]
[BorB]
[BorM]
[BouK]
[BraW]
[Bre]
[Bri1]
[Bri2]
[Br]
[BrM]
[Buck]
[Bukh]
[BMN]
[BBI]
[BZ1]

D. D. Bleecker, Volume increasing isometric deformations of convex polyhedra, J. Di.

Geom. 43 (1996), 505526; available at http://tinyurl.com/37tcwh [347, 405]
A. I. Bobenko and I. Izmestiev, Alexandrovs theorem, weighted Delaunay triangulations,
and mixed volumes, Ann. Inst. Fourier 58 (2008), 447505; arXiv: math.DG/0609447.
[334]
V. Boju and L. Funar, The Math Problems Notebook, Birkhauser, Boston, MA, 2007. [378]
B. Bollobas, The Art of Mathematics: Coee Time in Memphis, Cambridge U. Press,
Cambridge, UK, 2007. [409]
V. G. Boltyanskii, Hilberts Third Problem, John Wiley, New York, 1978; (J). [5, 147, 157,
390, 392]
V. G. Boltjansky and I. Ts. Gohberg, Results and problems in combinatorial geometry,
Cambridge U. Press, Cambridge, UK, 1985; (J). [31, 378, 379, 396]
V. Boltyanski, H. Martini and P. S. Soltan, Excursions into combinatorial geometry,
Springer, Berlin, 1997. [31]
V. G. Boltjanskii and P. S. Soltan, Combinatorial geometry of various classes of convex

sets (in Russian), Stiinca,

Kishinev, 1978. [377, 408]
V. G. Boltyanski and I. M. Yaglom, Convex sets (in Russian), in Encyclopedia of Elementary
Mathematics, vol. V, Fizmatgiz, Moscow, 1966, 192269; (J). [71]
T. Bonnesen and W. Fenchel, Theory of convex bodies, BCS, Moscow, ID, 1987; ()[70,
328, 404]
O. Borodin, Joint extension of two theorems of Kotzig on 3-polytopes, Combinatorica 13
(1993), 121125; (). [383]
E. Boros and Z. F
uredi, The number of triangles covering the center of an n-set, Geom.
Dedicata 17 (1984), 6977; (). [25, 38, 379]
A. V. Borovik and A. S. Borovik, Mirrors and Reections: The Geometry of Finite and
Ane Reection Groups, Birkhauser, Boston, MA, 2004. [184]
P. Borwein and W. O. J. Moser, A survey of Sylvesters problem and its generalizations,
Aeq. Math. 40 (1990), 111135; available at http://tinyurl.com/63ttwy [25, 378]
M. Boutin and G. Kemper, On reconstructing n-point congurations from the distribution
of distances or areas, Adv. Appl. Math. 32 (2004), 709735; arXiv: math/0304192 [403]
P. Brass and C. Wenk, On the number of cylinders touching a ball, Geom. Dedicata 81
(2000), 281284; (). [395]
M. Breen, A Helly-type theorem for intersections of compact connected sets in the plane,
Geom. Dedicata 71 (1998), 111117; (). [377]
R. Bricard, Sur une question de geometrie relative aux poly`edres, Nouv. Ann. Math. 15
(1896), 331334; available at http://tinyurl.com/2s4rdp [147, 390]
R. Bricard, Memoire sur la theorie de locta`edre articule, J. Math. Pures Appl. 3 (1897),
113150; English translation is available at http://tinyurl.com/2pa63m [285]
A. Brndsted, An introduction to convex polytopes, Springer, New York, 1983; (). [81]
A. Brndsted and G. Maxwell, A new proof of the d-connectedness of d-polytopes, Canad.
Math. Bull. 32 (1989), 252254. [382]
R. C. Buck and E. F. Buck, Equipartition of convex sets, Math. Mag. 22 (1949), 195
198; (). [37, 379]
B. Bukh, A point in many triangles, Electron. J. Combin. 13 (2006), no. 1, Note 10, 3 pp.;
available at http://tinyurl.com/376l9h [38]
B. Bukh, J. Matousek and G. Nivasch, Stabbing simplices by points and ats; arXiv:
0804.4464 [38, 379]
D. Burago, Yu. Burago and S. Ivanov, A course in metric geometry, AMS, Providence, RI,
2001; (). [358]
Ju. D. Burago and V. A. Zalgaller, Polyhedral embedding of a net (in Russian), Vestnik
Leningrad. Univ. 15 (1960), no. 7, 6680. [405]

414

[BZ2]
[BZ3]
[BZ4]

[Bus]
[Caf]
[CanR]
[Can+]

[CanJ]

[CKKS]
[Car]
[Carv]
[Cha]
[ChaK]
[CEG]
[Chen]
[Cher]
[CocO]
[CSE]

[Cohn]
[CV]
[Con1]
[Con2]
[Con3]

[Con4]
[Con5]

Ju. D. Burago and V. A. Zalgaller, Sucient criteria of convexity J. Math. Sci. 10 (1978),
395435; (), (). [377]
Yu. D. Burago and V. A. Zalgaller, Geometric inequalities, Springer, Berlin, 1988; (). [71,
398, 404]
Yu. D. Burago and V. A. Zalgaller, Isometric piecewise-linear embeddings of twodimensional manifolds with a polyhedral metric into R3 , St. Petersburg Math. J. 7 (1996),
369385; (). [347, 387, 405]
H. Busemann, The geometry of geodesics, Academic Press, New York, 1955; (). [248]
L. Caarelli, The Monge-Amp`ere equation and optimal transportation, an elementary review, in Lecture Notes in Math. 1813, Springer, Berlin, 2003, 110; (). [176]
J. F. Canny and J. H. Reif, New lower bound techniques for robot motion planning problems, Proc. 28th FOCS (1987), 4960. [406]
J. Cantarella, J. H. G. Fu, R. Kusner, J. M. Sullivan and N. C. Wrinkle, Criticality for
the Gehring link problem, Geom. Topol. 10 (2006), 20552116; arXiv: math.DG/0402212
[240]
J. Cantarella and H. Johnston, Nontrivial embeddings of polygonal intervals and unknots in 3-space, J. Knot Theory Ramications 7 (1998), 10271039; available at
http://tinyurl.com/2f9s2pv [396]
J. Cantarella, G. Kuperberg, R. B. Kusner and J. M. Sullivan, The second hull of a knotted
curve, Amer. J. Math. 125 (2003), 13351348; arXiv: math/0204106 [240, 398]
P. Cartier, Decomposition des poly`edres: le point sur le troisi`eme probl`eme de Hilbert,
Asterisque 133134 (1986), 261288; (); available at http://tinyurl.com/3e4cnb [147]
F. J. C. de Carvalho, Characterizing maximally looped closed curves, Amer. Math.
Monthly 92 (1985), 202207; (). [396]
G. D. Chakerian, An inequality for closed space curves, Pacic J. Math. 12 (1962), 5357;
available at http://tinyurl.com/37v9yh [398]
G. D. Chakerian and M. S. Klamkin, Minimal covers for closed curves, Math. Mag. 46
(1973), 5561. [240]
B. Chazelle, H. Edelsbrunner and L. J. Guibas, The complexity of cutting complexes,
Discrete Comput. Geom. 4 (1989), 139181; (). [407]
W. Chen, Counterexamples to Knasters conjecture, Topology 37 (1998), 401405; (). [60]
S. S. Chern, Curves and surfaces in Euclidean space, in Studies in Global Geometry and
Analysis, MAA, 1967, 1656. [206, 240, 397]
R. Cocan and J. ORourke, Polygonal chains cannot lock in 4D, Comput. Geom. 20 (2001),
no. 3, 105129; arXiv: cs.CG/9908005 [396]
D. Cohen-Steiner and H. Edelsbrunner, Inequalities for the curvature of curves and surfaces,
Found. Comput. Math. 7 (2007), 391404; available at http://tinyurl.com/2f2tmqg [398,
401]
P. M. Cohn, Skew elds, Cambridge U. Press, Cambridge, 1995. [314]
S. Cohn-Vossen, K
urzeste Wege und Totalkrummung auf Flachen, Composito Math. 2
(1935), 63133; (), avaialable at http://tinyurl.com/3acm4p [248]
R. Connelly, The rigidity of suspensions, J. Dierential Geom. 13 (1978), 399408; see also
An attack on rigidity I, II (1974); (), available at http://tinyurl.com/3b58yg [401]
R. Connelly, Conjectures and open questions in rigidity, in Proc. ICM Helsinki, Acad. Sci.
Fennica, Helsinki, 1980, 407414. [285, 291, 314]
R. Connelly, The rigidity of certain cabled frameworks and the second-order rigidity of
arbitrarily triangulated convex surfaces, Adv. in Math. 37 (1980), 272299; available at
http://tinyurl.com/2fyyjph [402]
R. Connelly, Flexing surfaces, in The mathematical Gardner (Edited by D. A. Klarner),
7989, PWS Publishers, Boston, MA, 1981; (). [285, 291]
R. Connelly, Rigidity, in Handbook of Convex Geometry, vol. A, North-Holland, Amsterdam, 1993, 223271. [215, 253, 262, 293, 300, 302, 306, 402]

415

[Con6]
[CDR]

[CSW]
[ConS]

[CGG]
[CRS]
[ConS]
[CouR]
[Cox1]
[Cox2]
[CoxG]
[Crof]
[CFG]
[CroS]
[CroW]
[Crom]
[Csa]
[Cun]
[Daw1]
[Daw2]
[Daw3]
[DGK]
[Deb1]
[Deb2]
[DML]
[Dehn]
[Deho]

R. Connelly, Comments on generalized Heron polynomials and Robbins conjectures,

preprint; available at http://tinyurl.com/36zmzy [314, 403]
R. Connelly, E. D. Demaine and G. Rote, Straightening polygonal arcs and convexifying polygonal cycles, Discrete Comput. Geom. 30 (2003), 205239; available at
http://tinyurl.com/28uv3ou [228]
R. Connelly, I. Sabitov and A. Walz, The bellows conjecture, Beitr. Algebra Geom. 38
(1997), 110; available at http://tinyurl.com/3ylche [314]
R. Connelly and H. Servatius, Higher-order rigiditywhat is the proper denition?, Discrete Comput. Geom. 11 (1994), 193200; available at http://tinyurl.com/5tuf7q [306,
402]
J. H. Conway, M. Goldberg and R. K. Guy, Problem 66-12, SIAM Review 11 (1969), 7882.
[94, 385]
J. H. Conway, C. Radin and L. Sadun, On angles whose squared trigonometric functions
are rational, Discrete Comput. Geom. 22 (1999), 321332; arXiv: math-ph/9812019 [392]
J. H. Conway and N. J. A. Sloane, Sphere Packings, Lattices, and Groups (Third edition),
Springer, New York, 1999; (). [191]
R. Courant and H. Robbins, What is mathematics?, Oxford U. Press, New York, 1996; (J).
[37, 116, 123]
H. S. M. Coxeter, Regular polytopes (Third edition), Dover, New York, 1973. [184, 392,
394]
H. S. M. Coxeter, Orthogonal trees, Bull. Inst. Combin. Appl. 3 (1991), 8391; (). [392]
H. S. M. Coxeter and S. L. Greitzer, Geometry revisited, McGraw-Hill, New York, 1967; (J).
[385]
H. T. Croft, A net to hold a sphere, J. Lond. Math. Soc. 39 (1964), 14; (). [399]
H. T. Croft, K. J. Falconer and R. K. Guy, Unsolved problems in geometry, Springer, New
York, 1991. [94, 380, 407]
H. T. Croft and H. P. F. Swinnerton-Dyer, On the Steinhaus billiard table problem, Proc.
Cambridge Philos. Soc. 59 (1963), 3741. [385]
H. T. Croft and W. G. Wild, E2152, Amer. Math. Monthly 76 (1969), 1140; (). [384]
P. R. Cromwell, Polyhedra, Cambridge U. Press, Cambridge, UK, 1997. [184, 248, 394]
A. Csaszar, A polyhedron without diagonals, Acta Univ. Szeged. Math. 13 (1949), 140-142.
[408]
H. M. Cundy, Deltahedra, Math. Gaz. 36 (1952), 263266; (). [394]
R. Dawson, On removing a ball without disturbing the others, Math. Mag. 57 (1984),
2730; (). [409]
R. J. M. Dawson, Monostatic Simplexes, Amer. Math. Monthly 92 (1985), 541546; ().
[94, 385]
R. J. M. Dawson, Tilings of the sphere with isosceles triangles, Discrete Comput. Geom. 30
(2003), 467487; (). [395]
L. Danzer, B. Gr
unbaum and V. Klee, Hellys theorem and its relatives, in Convexity, AMS,
Providence, RI, 1963, 101180; (). [19, 377]
H. E. Debrunner, Zerlegungsrelationen zwischen regularen Polyedern des Ed , Arch.
Math. 30 (1978), 656-660. [392]
H. E. Debrunner, Dissecting orthoschemes into orthoschemes, Geom. Dedicata 33 (1990),
123152; (). [392]
H. E. Debrunner and P. Mani-Levitska, Can you cover your shadows? Discrete Comput.
Geom. 1 (1986), 4558; available at http://tinyurl.com/2v2rwy [408]

M. Dehn, Uber
die Starreit konvexer Polyeder, Math. Ann. 77 (1916), 466473; (); available at http://tinyurl.com/35kkkx [300]
P Dehornoy, On the rotation distance between binary trees, Adv. Math. 223/4 (2010),
13161355; arXiv: 0901.2557. [387]

416

[DRS]
[DGOT]
[Dev]
[Dek]
[Den]
[DeoK]
[DGPV]

[DiaK]
[Dil]
[Dol]
[DSS1]
[DSS2]
[Dup]
[Dvo]
[DEKP]
[Dys]
[Eck]
[EdeS]
[Edm1]
[Edm2]
[Egg1]
[Egg2]
[Emch]
[EKK]
[Epp]
[ESU]
[Erd]

J. A. De Loera, J. Rambau and F. Santos, Triangulations: Structures and Algorithms,

Springer, 2008. [138, 387, 388]
E. D. Demaine, B. Gassend, J. ORourke and G. T. Toussaint, Polygons ip nitely: aws
and a x, in Proc. CCCG 2006, 109112; available at http://tinyurl.com/2sdpt3 [397]
M. Develin, Disjoint faces of complementary dimension, Beitr. Algebra Geom. 45 (2004),
463464; available at http://tinyurl.com/2rdvl6 [383]
B. V. Dekster, Convex hulls of spatial polygons with a xed convex projection, Beitr.
Algebra Geom. 36 (1995), 123134; available at http://tinyurl.com/2vbedt [388]
E. Denne, Alternating quadrisecants of knots, arXiv: math.GT/0510561. [240, 398]
N. Deo and M. S. Klamkin, Existence of four concurrent normals to a smooth closed curve,
Amer. Math. Monthly 77 (1970), 10831084; (). [385]
D. DeTurck, H. Gluck, D. Pomerleano and D. S. Vick, The four vertex theorem and its
converse, Notices AMS 54 (2007), 192207; available at http://tinyurl.com/2kwjkn
[206]
P. Diaconis and J. B. Keller, Fair dice, Amer. Math. Monthly 96 (1989), 337339; (). [394]
M. B. Dillencourt, Realizability of Delaunay triangulations, Inform. Process. Lett. 33
(1990), no. 6, 283287; (). [388]
N. P. Dolbilin, Gems of Polytope Theory (in Russian), MCCME, Moscow, 2000; (J). [285]
N. P. Dolbilin, M. A. Shtan ko and M. I. Shtogrin, Nonbendability of polyhedral spheres
with even-sided faces, Russian Math. Surveys 51 (1996), 543544; (), (). [402]
N. P. Dolbilin, M. A. Shtan ko and M. I. Shtogrin, Rigidity of a quadrillage of a torus by
squares, Russian Math. Surveys 54 (1999), 839840; (), (). [400]
J. L. Dupont, Scissors congruences, group homology and characteristic classes, World Sci.,
River Edge, NJ, 2001; available at http://tinyurl.com/376ol6 [5, 147]
A. Dvoretzky, A converse of Hellys theorem on convex sets, Pacic J. Math. 5 (1955),
345350; available at http://tinyurl.com/2wjvsv [377]
A. V. Dyskin, Y. Estrin, A. J. Kanel-Belov and E. Pasternak, Topological interlocking of
platonic solids, Phil. Mag. Letters 83 (2003), 197203; (). [409]
F. J. Dyson, Continuous functions dened on spheres, Ann. of Math. 54 (1951), 534536;
(). [60]
J. Eckho, Helly, Radon, and Caratheodory type theorems, in Handbook of Convex Geometry, vol. A, 389448, North-Holland, Amsterdam, 1993. [19]
M. Edelstein and B. Schwarz, On the length of linked curves, Israel J. Math. 23 (1976),
9495; (). [240]
A. L. Edmonds, The geometry of an equifacetal simplex, Mathematika 52 (2005), 3145;
arXiv: math.MG/0408132 [394]
A. L. Edmonds, Sommervilles missing tetrahedra, Discrete Comput. Geom. 37 (2007),
287296; (). [392]
H. G. Eggleston, Convexity, Cambridge U. Press, Cambridge, UK, 1958. [71, 379]
H. G. Eggleston, Figures inscribed in convex sets, Amer. Math. Monthly 65 (1958), 7680;
(). [380]
A. Emch, Some properties of closed convex curves in a plane, Amer. J. Math. 35 (1913),
407412; (). [53, 379]
V. A. Emelichev, M. M. Kovalev and M. K. Kravtsov, Polytopes, Graphs, and Optimization,
Cambridge U. Press, New York, 1984; (). [82, 384]
D. Eppstein, The farthest point Delaunay triangulation minimizes angles, Comput. Geom. 1
(1992), no. 3, 143148; available at http://tinyurl.com/6fzxer [387]
or, Tiling space and slabs with acute tetrahedra,
D. Eppstein, J. M. Sullivan and A. Ung
Comput. Geom. 27 (2004), 237255; arXiv: cs.CG/0302027. [387]
P. Erdos, Personal reminiscences and remarks on the mathematical work of Tibor Gallai,
Combinatorica 2 (1982), 207212; available at http://tinyurl.com/6fr55a [25]

417

[ErdP]

P. Erdos and G. Purdy, Extremal problems in combinatorial geometry, in Handbook of

Combinatorics, Vol. 1, 809874, Elsevier, Amsterdam, 1995. [25, 377, 378]
[EriK] J. Erickson and S. Kim, Arbitrarily large neighborly families of congruent symmetric convex
3-polytopes, in Discrete Geometry: In Honor of W. Kuperbergs 60th Birthday, MarcelDekker, 2003, 267278; arXiv: math.CO/0106095. [409]
[Eri]
F. Eriksson, The law of sines for tetrahedra and n-simplices, Geom. Dedicata 7 (1978),
7180. [408]
[Euc]
Euclid, Elements, Books 1-13; available e.g., at http://tinyurl.com/62x9q5 ; (J). [256]
[Fab]
Fr. Fabricius-Bjerre, A proof of a relation between the numbers of singularities of a closed
polygon, J. Geom. 13 (1979), 126132. [228]
[Fed]
P. J. Federico, Descartes on polyhedra, Springer, Berlin, 1982. [248]
[FedP] M. Fedorchuk and I. Pak, Rigidity and polynomial invariants of convex polytopes, Duke
Math. J. 129 (2005), 371-404; available at http://tinyurl.com/3xvj96 [288, 291, 300,
401, 402, 403]
[Fej1]
L. Fejes Toth, E895, Amer. Math. Monthly 57 (1950), 419420; (). [382]
[Fej2]
L. Fejes Toth, Lagerungen in der Ebene, auf der Kugel und im Raum, Springer, Berlin,
1953; (J). [71, 184, 191]
[Fenn] R. Fenn, The table theorem, Bull. London Math. Soc. 2 (1970), 7376; (). [381]
[FenR] R. Fenn and C. Rourke, Klyachkos methods and the solution of equations over torsion-free
groups, LEns. Math. 42 (1996), 4974; available at http://tinyurl.com/374e58 [384]

[Fie]
M. Fiedler, Uber
qualitative Winkeleigenschaften der Simplexe Czechoslovak Math. J. 7
(1957), 463478; available at http://tinyurl.com/6pc3uj [392]
[FSW] H. Fleischner, G. Sabidussi and E. Wenger, Transforming Eulerian trails, Discrete
Math. 109 (1992), 103116; (). [398]
[Fle]
T. M. Flett, Amp`ere and the horizontal chord theorem, Bull. Inst. Math. Appl. 11 (1975),
no. 1-2, 34. [38]
[Flo]
E. E. Floyd, Real-valued mappings of spheres, Proc. AMS 6 (1955), 957959; (). [60]
[For]
R. Forman, A users guide to discrete Morse theory, Sem. Lothar. Combin. 48 (2002),
Art. B48c, 35 pp.; available at http://tinyurl.com/2o9zhu [82]
[Fox]
R. H. Fox, A quick trip through knot theory, in Topology of 3-manifolds and related topics,
120167, Prentice-Hall, Englewood Clis, NJ, 1962; (). [397]
[Fri]
E. J. Friedman, Finding a simple polytope from its graph in polynomial time, Discrete
Comput. Geom. 41 (2009), 249256; available at http://tinyurl.com/5om26t [77]
[FucF] D. Fuchs and E. Fuchs, Closed geodesics on regular polyhedra, Mosc. Math. J. 7 (2007),
265279. [386]
[FucT] D. Fuchs and S. Tabachnikov, Mathematical Omnibus. Thirty Lectures on Classic Mathematics, draft of a monograph; available at http://tinyurl.com/2u2c7v [285]
[Gal]
G. Galperin, Convex polyhedra without simple closed geodesics, Regular and Chaotic Dynamics 8 (2003), 4558; (); available at http://tinyurl.com/3xy8pv [101]
[Gal+] J. Galtier, F. Hurtado, M. Noy, S. Perennes and J. Urrutia, Simultaneous edge ipping
in triangulations, Internat. J. Comput. Geom. Appl. 13 (2003), 113133; available at
http://tinyurl.com/plazz8 [387]
[Gar]
R. J. Gardner, The BrunnMinkowski inequality, Bull. AMS 39 (2002), 355405; available
at http://tinyurl.com/2k8thm [71]
[GarM] R. J. Gardner and P. McMullen, On Hammers X-ray problem, J. London Math. Soc. 21
(1980), 171175; (). [381]
[GKZ] I. M. Gelfand, M. M. Kapranov and A. V. Zelevinsky, Discriminants, resultants, and multidimensional determinants, Birkhauser, Boston, MA, 1994. [82, 138]
[Ger]
M. L. Gerver, Width of polygons (in Russian), Matemat. Prosv. 2 (1999), 173191; (J).
[381]
[Gla]
L. Glass, A combinatorial analog of the Poincare index theorem, J. Combin. Theory Ser.
B 15 (1973), 264268. [399, 400]

418

[Glaz]

A. Glazyrin, Lower bounds for the simplexity of the d-cube, to appear in Discrete Math.
[388]
[Gle]
A. M. Gleason, Angle trisection, the heptagon, and the triskaidecagon, Amer. Math.
Monthly 95 (1988), 185194, 911; (). [387]
[Glu]
H. Gluck, Almost all simply connected closed surfaces are rigid, in Lecture Notes in
Math. 438, Springer, Berlin, 1975, 225239; (), (). [291]
[Gol1] M. Goldberg, Polyhedral linkages, Nat. Math. Mag. 16 (1942), 323332; (). [123]
[Gol2] M. Goldberg, Tubular linkages, J. Math. Phys. MIT 26 (1947), 1021. [285]
[Gol3] M. Goldberg, Three innite families of tetrahedral space-llers, J. Combin. Theory
Ser. A 16 (1974), 348354; (). [392]
[Gol4] M. Goldberg, Unstable polyhedral structures, Math. Mag. 51 (1978), 165170; available at
http://tinyurl.com/6funls [306]
[Gol5] M. Goldberg, On the space-lling heptahedra, Geom. Dedicata 7 (1978), 175184; (). [391]
[GooT] I. J. Good and T. Teichmann, E1923, Amer. Math. Monthly 75 (1968), 194195; (). [408]
[GPY] J. E. Goodman, J. Pach and C.-K. Yap, Mountain climbing, ladder moving, and the ringwidth of a polygon, Amer. Math. Monthly 96 (1989), 494510; (). [53, 381]
[Gor]
I. I. Gordon, Generalization of Kakutanis theorem on continuous functions given on a
sphere (in Russian), Uspehi Mat. Nauk 10 (1955), no. 1, 9799. [60]
[GGT] S. Gortler, C. Gotsman and D. Thurston, One-forms on meshes and applications to 3D
mesh parametrization, Comput. Aided Geom. Design 23 (2006), no. 2, 83112; available at
http://tinyurl.com/2rbe86 [399]
[GouN] I. P. Goulden and A. Nica, A direct bijection for the Harer-Zagier formula, J. Combin.
Theory Ser. A 111 (2005), 224238; (). [404]
[Gray] A. Gray, Tubes (Second edition), Birkhauser, Basel, 2004; (). [248]
[Gri]
H. B. Griths, The topology of square pegs in round holes, Proc. London Math. Soc. 62
(1991), 647672; (). [53, 380]
[GriH] P. Gritzmann and A. Hufnagel, On the algorithmic complexity of Minkowskis reconstruction theorem, J. London Math. Soc. 59 (1999), 10811100; (). [328]
[GriK] P. Gritzmann and V. Klee, On the complexity of some basic problems in computational
convexity. II. Volume and mixed volumes, in Polytopes: abstract, convex and computational,
Kluwer, Dordrecht, 1994, 373466; available at http://tinyurl.com/3xeyyh [369]
[GriL] P. Gritzmann and M. Lassak, Estimates for the minimal width of polytopes inscribed in convex bodies, Discrete Comput. Geom. 4 (1989), 627635; available at
http://tinyurl.com/65uycm [408]
[Gro]
H. Groemer, Stability of geometric inequalities, in Handbook of convex geometry, Vol. A,
125150, North-Holland, Amsterdam, 1993. [328]
[Grub] P. M. Gruber, Convex and discrete geometry, Springer, Berlin, 2007. [5, 70, 71, 108, 168,
328, 381, 399, 404]
[Gr
u1] B. Gr
unbaum, A simple proof of a theorem of Motzkin, Indag. Math. 26 (1964), 382384.
[383]
[Gr
u2] B. Gr
unbaum, Arrangements and spreads, AMS, Providence, RI, 1972. [37]
[Gr
u3] B. Gr
unbaum, No-net polyhedra, Geombinatorics 11 (2002), no. 4, 111114; available at
http://tinyurl.com/6f6849 [406]
[Gr
u4] B. Gr
unbaum, Convex Polytopes (Second edition), Springer, Berlin, 2003. [81, 108, 117,
248, 386, 400]
[Gr
u5] B. Gr
unbaum, Graphs of polyhedra; polyhedra as graphs, Discrete Math. 307 (2007),
445463; available at http://tinyurl.com/6s99oc [108]
[GMS] B. Gr
unbaum, P. Mani-Levitska and G. C. Shephard, Tiling three-dimensional space with
polyhedral tiles of a given isomorphism type, J. London Math. Soc. 29 (1984), 181191; ().
[387]
[GM1] B. Gr
unbaum and T. S. Motzkin, Longest simple paths in polyhedral graphs, J. London
Math. Soc. 37 (1962), 152160. [383]

419

[GM2]

[Gug1]
[Gug2]
[Gug3]
[Gug4]
[Gul]
[Hada]
[Had1]
[Had2]
[HDK]
[HadG]
[HLM]
[HajW]
[Hal1]
[Hal2]
[Hal3]
[Hal4]
[Hau]
[HMS]

[Heil]
[HenP]
[Hep1]
[Hep2]
[HerR]
[Hil]
[HilC]
[HocR]

B. Gr
unbaum and T. S. Motzkin, The number of hexagons and the simplicity of
geodesics on certain polyhedra, Canad. J. Math. 15 (1963), 744751; available at
http://tinyurl.com/6xu73m [383]
H. W. Guggenheimer, Dierential geometry, McGraw-Hill, New York, 1963. [206]
H. Guggenheimer, Finite sets on curves and surfaces, Israel J. Math. 3 (1965), 104112; ().
[53, 228, 381]
H. Guggenheimer, Research problems: Does there exist a four normals triangle? Amer.
Math. Monthly 77 (1970), 177179; (). [384]
H. Guggenheimer, On the converse Helly theorem, Amer. Math. Monthly 78 (1971), 278
279; (). [377]
R. Gulliver, Total curvature of graphs in space, Pure Appl. Math. Quarterly 3 (2007),
773783; avialable at http://tinyurl.com/2sn2ja [398]
J. Hadamard, Lecons de geometrie II. Geometrie dans lespace (reprint of the eighth edition), Sceaux, 1988; (J). [179, 215, 369, 403]
H. Hadwiger, Vorlesungen u
ber Inhalt, Ober
ache und Isoperimetrie, Springer, Berlin,
1957; (). [70]
H. Hadwiger, Homothetieinvariante und additive Polyederfunktionen, Arch. Math. 25
(1974), 203205; (). [157]
H. Hadwiger, H. Debrunner and V. Klee, Combinatorial Geometry in the Plane, Holt,
Rinehart and Winston, New York, 1964; (). [19, 31, 378]
H. Hadwiger and P. Glur, Zerlegungsgleichheit ebener Polygone, Elem. Math. 6 (1951),
97106; available at http://tinyurl.com/2mfwo4 [390]
H. Hadwiger, D. G. Larman and P. Mani, Hyperrhombs inscribed to convex bodies, J.
Combin. Theory Ser. B 24 (1978), 290293; (). [53, 380, 381]
M. Hajja and P. Walker, Equifacial tetrahedra, Internat. J. Math. Ed. Sci. Tech. 32 (2001),
501508; (). [407]
B. Halpern, The kernel of a starshaped subset of the plane, Proc. AMS 23 (1969), 692
696; (). [377]
B. Halpern, On the immersion of an n-dimensional manifold in n+1-dimensional Euclidean
space. Proc. AMS 30 (1971), 181184; (). [377]
B. Halpern, Double normals and tangent normals for polygons, Proc. AMS 51 (1975),
434437; (). [396]
B. Halpern, An inequality for double tangents, Proc. AMS 76 (1979), 133139; (). [396]
O. Haupt, Zur Verallgemeinerung des Zweischeitelsatzes bei ebenen Kurven, Arch. Math. 11
(1960), 294297; (). [206]
T. Hausel, E. Makai, and A. Sz
ucs, Inscribing cubes and covering by rhombic dodecahedra via equivariant topology, Mathematika 47 (2000), 371397; arXiv: math.MG/9906066.
[381]
E. Heil, Trefoil knots with tritangent planes, Beitr
age Algebra Geom. 45 (2004), 527530;
available at http://tinyurl.com/3adcs7 [396]
A. Henriques and I. Pak, Volume-preserving PL-maps between polyhedra, preprint; available at http://tinyurl.com/3x25su [176]
A. Heppes, A double-tipping tetrahedron, SIAM Review 9 (1967), 599600; (). [385]
A. Heppes, Covering a planar domain with sets of small diameter, Period. Math. Hungar. 53
(2006), 157168; (). [379]
E. Hertel and C. Richter, Squaring the circle by dissection, Beit. Algebra Geom. 44 (2003),
4755; available at http://tinyurl.com/34n269 [158]
D. Hilbert, Foundations of geometry, Open Court, LaSalle, Ill., 1971; (). [116]
D. Hilbert and S. CohnVossen, Geometry and the imagination, Chelsea, New York,
1952; (J). [25, 123]
R. Hochberg and M. Reid, Tiling with notched cubes, Discrete Math. 214 (2000), 255261;
available at http://tinyurl.com/6jmp6d [390]

420

H. Hopf, Uber
die Sehnen ebener Kontinuen und die Schleifen geschlossener Wege, Comment. Math. Helv. 9 (1936), 303319; (), available at http://tinyurl.com/343gta [38]
[Hop2] H. Hopf, Dierential Geometry in the Large, Lecture Notes in Math. 1000, Springer, Berlin,
1983. [147, 248]
[Hor]
L. Hormander, Notions of convexity, Birkhauser, Boston, MA, 1994. [70]
[Horn] R. A. Horn, On Fenchels theorem, Amer. Math. Monthly 78 (1971), 380381; (). [240]
[Hun]
N. Hungerbuhler, A short elementary proof of the MohrMascheroni theorem, Amer. Math.
Monthly 101 (1994), 784787; available at http://tinyurl.com/2uub8j [116]
[HNU] F. Hurtado, M. Noy and J. Urrutia, Flipping edges in triangulations, Discrete Comput.
Geom. 22 (1999), 333346; available at http://tinyurl.com/5m2ohq [387]
[IKS]
I. Ivanova-Karatopraklieva and I. Kh. Sabitov, Bending of surfaces. II, J. Math. Sci. 74
(1995), 9971043. [262, 306, 402]
[Izm]
I. Izmestiev, On the hull numbers of torus links, J. Knot Theory Ramications 15 (2006),
589600; arXiv: math.GT/0412139. [398]
[Izm]
I. Izmestiev, A variational proof of Alexandrovs convex cap theorem; arXiv:
math/0703169. [334]
[IzmS] I. Izmestiev and J.-M. Schlenker, On the innitesimal rigidity of polyhedra with vertices in
convex position, to appear in Pacic J. Math; arXiv: arXiv:0711.1981. [169, 393]
[Jac]
S. B. Jackson, Vertices for plane curves, Bull. AMS 50 (1944), 564478; available at
http://tinyurl.com/2psyqn [206, 380]
[Jer]
R. P. Jerrard, Inscribed squares in plane curves, Trans. AMS 98 (1961), 234241; (). [53]
[Jes1]
B. Jessen, On equivalence of aggregates of regular polyhedra (in Danish), Mat. Tidsskr. B.
(1946), 145148. [392]
[Jes2]
B. Jessen, Orthogonal icosahedra, Nordisk Mat. Tidskr. 15 (1967), 9096. [392, 394]
[JocP] W. Jockusch and N. Prabhu, Cutting a polytope, Geom. Dedicata 54 (1995), 307312.
[408]
[Joh]
N. W. Johnson, Convex polyhedra with regular faces, Canad. J. Math. 18 (1966), 169200.
[184]
[JucM] E. Jucovic and J. W. Moon, The maximum diameter of a convex polyhedron, Math. Mag. 38
(1965), 3132. [383]

[Kag1] V. F. Kagan, Uber

die Transformation der Polyeder, Math. Ann. 57 (1903), 421424;
available at http://tinyurl.com/396r64 [147]
[Kag2] V. F. Kagan, Foundations of geometry I, II (in Russian), GTI, Moscow, 1949, 1956. [116]
[Kah]
W. Kahan, What has the volume of a tetrahedron to do with computer programming
languages?, UC Berkeley preprint; available at http://tinyurl.com/3amrpg [408]
[KahK] J. Kahn and G. Kalai, A counterexample to Borsuks conjecture, Bull. AMS 29 (1993),
6062; arXiv: math.MG/9307229. [26, 31]
[Kaib] V. Kaibel, Reconstructing a simple polytope from its graph, in Combinatorial Optimization,
Springer, New York, 2002, 105118; arXiv: math/0202103. [77]
[Kais] H. Kaiser, Perfekte Dreieckzerlegungen, Elem. Math. 46 (1991), 106111; available at
http://tinyurl.com/3c887y [409]
[Kake] S. Kakeya, On the inscribed rectangles of a closed curvex curve, T
ohoku Math. J. 9 (1916),
163166. [380]
[Kaku] S. Kakutani, A proof that there exists a circumscribing cube around any bounded closed
convex set in R3 , Ann. of Math. 43 (1942), 739741; (). [60]
[Kal1] G. Kalai, A simple way to tell a simple polytope from its graph, J. Combin. Theory,
Ser. A 49 (1988), 381383; (). [82]
[Kal2] G. Kalai, Results and Problems around Borsuks Conjecture, a talk at PIMS, July 19, 2002;
available at http://tinyurl.com/2vlrul [31]
[Kall]
M. Kallay, Indecomposable polytopes, Israel J. Math. 41 (1982), 235243; (). [404]
[Kan]
A. Kanel, The story of one olympiad problem (in Russian), Matemat. Prosv. 5 (2001),
207208; (J). [409]
[Hop1]

421

[KM1]

M. Kapovich and J. J. Millson, On the moduli space of polygons in the Euclidean plane,
J. Di. Geom. 42 (1995), 430464; available at http://tinyurl.com/2d2uaj [228]
[KM2] M. Kapovich and J. J. Millson, Moduli spaces of linkages and arrangements, in Advances in geometry, Birkhauser, Boston, MA, 1999, 237270; available electronically at
http://tinyurl.com/2rb44h [123]
[KM3] M. Kapovich and J. J. Millson, On the moduli space of a spherical polygonal linkage,
Canad. Math. Bull. 42 (1999), 307320; available at http://tinyurl.com/2ka4bz [387]
[KM4] M. Kapovich and J. J. Millson, Universality theorems for conguration spaces of planar
linkages, Topology 41 (2002), 10511107; arXiv: math.AG/9803150. [123]
[Kara] R. N. Karasev, On a conjecture of A. Bezdek, Discrete Comput. Geom. 27 (2002), 419
439; (). [378]
[Karc] H. Karcher, Remarks on polyhedra with given dihedral angles, Comm. Pure Appl. Math. 21
(1968), 169174; (). [400]
[KarB] Y. Karshon and C. Bjorndahl, Revisiting Tietze-Nakajima - local and global convexity for
maps; arXiv: math/0701745. [138]
[KasS] B. S. Kashin and S. J. Szarek, The Knaster problem and the geometry of highdimensional cubes, C. R. Math. Acad. Sci. Paris 336 (2003), no. 11, 931936; available at
http://tinyurl.com/2by5ub [60]
[Kel]
T. Keleti, The mountain climbers problem, Proc. AMS 117 (1993), 8997; available at
http://tinyurl.com/25huj25 [53, 381]
[Kem] A. B. Kempe, How to draw a straight line; a lecture on linkages, Macmillan and Co.,
London, UK, 1877; available at http://tinyurl.com/2j8jkw [123]
[KenK] C. Kenyon and R. Kenyon, How to take short cuts, Discrete Comput. Geom. 8 (1992),
251264; available at http://tinyurl.com/2jckaz [407]
[Ker]
R. B. Kershner, On paving the plane, Amer. Math. Monthly 75 (1968), 839-844; (). [391]
[King] H. C. King, Planar linkages and algebraic sets, in Proc. 6th Gokova Geometry-Topology
Conf., Turkish J. Math. 23 (1999), 3356; arXiv: math.AG/9807023. [123]
[Kis]
A. P. Kiselev, Kiselevs Geometry. Book II. Stereometry (adapted from Russian by A.
Givental), Sumizdat, 2008; (J). [393]
[Kla]
D. A. Klain, The Minkowski problem for polytopes, Adv. Math. 185 (2004), 270288;
available at http://tinyurl.com/2mhrfh [328]
[KlaT] M. S. Klamkin and G. A. Tsintsifas, The Circumradius-Inradius Inequality for a Simplex
Math. Mag. 52 (1979), 2022; (). [408]
[Klee] V. Klee, Some characterizations of convex polyhedra. Acta Math. 102 (1959), 79107; ().
[407]
[KleR] V. Klee and J. R. Reay, A surprising but easily proved geometric decomposition theorem,
Math. Mag. 71 (1998), 311; (). [158]
[KleW] V. L. Klee and S. Wagon, Old and new unsolved problems in plane geometry and number
theory, MAA, Washington, DC, 1991. [53, 94, 381]
[Kli]
W. Klingenberg, Lectures on closed geodesics, Springer, Berlin, 1978; (). [101]
[Koe]
W. Koenen, Characterizing the circle, Amer. Math. Monthly 78 (1971), 993996; (). [380]
[KonP] M. Konvalinka and I. Pak, Geometry and complexity of OHaras algorithm, Adv. Applied
Math. 42 (2009), 157175; arXiv:0710.1459. [391]
[KosT] G. Kos and J. Torocsik, Convex disks can cover their shadow, Discrete Comput. Geom. 5
(1990), 529531; available at http://tinyurl.com/2jrtwz [408]
[Kos]
A. N. Kostovskij, Geometrical constructions using compasses only, Pergamon Press, Oxford,
1961; (J). [116]
[Kor1] I. G. Korepanov, Multidimensional analogues of geometric s t duality, Theor. Math.
Phys. 124 (2000), 9991005; (), arXiv: solv-int/9911008. [401]
[Kor2] I. G. Korepanov, Invariants of PL manifolds from metrized simplicial complexes, J. Nonlinear Math. Phys. 8 (2001), 196210; arXiv: math/0009225. [401]

422

[KorP]
[Kra]
[Kri]
[KroK]
[Kry]
[Kuh]
[Kui]
[Kup1]
[Kup2]
[KupS]
[Kvant]
[LM]

[LR]
[Lak]
[Lam]
[Lan]
[Lang]
[Lass]
[Leb1]
[Leb2]
[Leb3]
[Lee]
[Leib]
[Leic]
[LenW]
[Len]
[LerW]
[Lic]

M. Korn and I. Pak, Tilings of rectangle with T-tetrominoes, Theor. Comp. Sci. 319 (2004),
327; available at http://tinyurl.com/323wry [300]
H. Kramer, Hyperparallelograms inscribed to convex bodies, Mathematica (Cluj) 22 (1980),
no. 1, 6770. [53, 380]
M. Krzek, There is no face-to-face partition of R5 into acute simplices, Discrete Comput.
Geom. 36 (2006), 381390. [387]
E. H. Kronheimer and P. B. Kronheimer, The tripos problem, J. London Math. Soc. 24
(1981), 182192. [381]
D. A. Kryzhanovski, Isoperimeters (in Russian), Fizmatgiz, Moscow, 1959. [70]
W. K
uhnel, Tight polyhedral submanifolds and tight triangulations, Lecture Notes in
Math. vol1612, Springer, Berlin, 1995; (). [285, 383, 401]
N. H. Kuiper, Double normals of convex bodies, Israel J. Math. 2 (1964), 7180; (). [385]
G. Kuperberg, Quadrisecants of knots and links, J. Knot Theory Ramications 3 (1994),
4150; arXiv: math.GT/9712205. [398]
G. Kuperberg, A volume-preserving counterexample to the Seifert conjecture, Comment.
Math. Helv. 71 (1996), 7097; arXiv: math.DS/9504230. [176]
G. Kuperberg and O. Schramm, Average kissing numbers for non-congruent sphere packings, Math. Res. Lett. 1 (1994), 339344; arXiv: math/9405218. [395]
Kvant Magazine (in Russian), years 19702003; (J). [377, 378, 379, 381, 383, 394, 395, 400,
403, 409]
J. C. Lagarias and D. Moews, Polytopes that ll Rn and scissors congruence, Discrete
Comput. Geom. 13 (1995), 573583; and addendum 14 (1995), 359360; available at
http://tinyurl.com/2qpkkf and http://tinyurl.com/3cgrku [158, 391]
J. C. Lagarias and T. J. Richardson, Convexity and the average curvature of plane curves,
Geom. Dedicata 67 (1997), 130; available at http://tinyurl.com/2v9lxd [398]
I. Lakatos, Proofs and refutations, Cambridge U. Press, Cambridge, UK, 1976; (). [248]
T. Lambert, The delaunay triangulation maximizes the mean inradius, in Proc. 6th CCCG
(1994), 201206; available at http://tinyurl.com/5oce98 [387]
S. Lang, Introduction to algebraic geometry, Addison-Wesley, Reading, MA, 1972. [365]
R. J. Lang, Origami design secrets, A. K. Peters, Natick, MA, 2003. [358, 406]
B. Lass, Demonstration combinatoire de la formule de HarerZagier, C. R. Acad. Sci. Paris
Ser. I Math. 333 (2001), 155160; available at http://tinyurl.com/6phpk7 [404]
H. Lebesgue, Sur lequivalence des poly`edres..., Ann. Soc. Polon. Math. 17 (1939), 193226.
[392]
H. Lebesgue, Lecons sur les Constructions Geometriques, Gauthier-Villars, Paris, 1950.
[116]
H. Lebesgue, Octa`edres articules de Bricard, Enseignement Math. 13 (1967), 175185;
available at http://tinyurl.com/78ypk3 [285]
J. Leech, Some properties of the isosceles tetrahedron, Math. Gaz. 34 (1950), 269271; ().
[407]
A. S. Leibin, On the deformability of convex surfaces with a boundary (in Russian), Uspehi
Mat. Nauk 5 (1950), 149159. [339]
K. Leichtweiss, Konvexe Mengen, Springer, Berlin, 1980; (). [377]
W. J. Lenhart and S. H. Whitesides, Reconguring closed polygonal chains, Discrete Comput. Geom. 13 (1995), 123140; available at http://tinyurl.com/33tbzc [228]
N. J. Lennes, Theorems on the simple nite polygon and polyhedron, Amer. J. Math. 33
(1911), 3762; (). [388]
G. Lerman and J. T. Whitehouse, On d-dimensional d-semimetrics and simplex-type inequalities for high-dimensional sine functions; arXiv: 0805.1430. [408]
W. B. R. Lickorish, Simplicial moves on complexes and manifolds, in Geom. Topol. Monogr.,
vol. 2, 299320, Coventry, UK, 1999; arXiv: math.GT/9911256. [168]

423

[Lin]
[Liv]
[Lov]
[LovP]
[LudR]
[Luo]
[Lyu]
[Mae1]
[Mae2]
[Mak1]
[Mak2]
[Man]

[Mar]
[MarW]
[Mat1]
[Mat2]
[Mats]
[MazM]
[McM1]
[McM2]
[McS]
[Mel]
[Mey]
[Mey1]
[Mey2]
[Mey3]
[MeyA]
[Mi1]
[Mi2]
[Mi3]
[Mi4]

J. Linhart, Uber
die Kantenlangensumme von Dreieckspolyedern, Monatsh. Math. 83
(1977), 2536; available at http://tinyurl.com/2khfsh [382]
G. R. Livesay, On a theorem of F. J. Dyson, Ann. of Math. 59 (1954), 227229; (). [381]
L. Lovasz, On the number of halving lines, Ann. Univ. Sci. Budapest. E
otv
os Sect. Math. 14
(1971), 107108. [228]
L. Lovasz and M. D.Plummer, Matching Theory, AMS Chelsea, 2009; (). [384]
M. Ludwig and M. Reitzner, Elementary moves on triangulations, Discrete Comput.
Geom. 35 (2006), 527536; arXiv: math.MG/0506564. [168]
F. Luo, Rigidity of polyhedral surfaces, arXiv: math.GT/0612714. [402]
L. A. Lyusternik, Convex gures and polyhedra, Dover, New York, 1963; (). [38, 215, 328,
379]
H. Maehara, Acute triangulations of polygons, Europ. J. Comb. 23 (2002), 4555; (). [387]
H. Maehara, The problem of thirteen spheresa proof for undergraduates, Europ. J. Combin. 28 (2007), 17701778; (). [191]
V. V. Makeev, Inscribed simplices of a convex body, J. Math. Sci. 72 (1994), 31893190;
(), (). [381]
V. V. Makeev, On quadrangles inscribed in a closed curve and the vertices of a curve,
J. Math. Sci. 131 (2005), no. 1, 53955400; (), (). [380]
Yu. I. Manin, On solvability of ruler and compass construction problems (in Russian), in
Encyclopedia of Elementary Mathematics, vol. IV, 205228, Fizmatgiz, Moscow, 1963; (J).
[116]
G. E. Martin, Geometric constructions, Springer, New York, 1998. [116]
H. Martini and W. Wenzel, Simplices with congruent k-faces, J. Geom. 77 (2003), 136139;
(). [394]
J. Matousek, Lectures on Discrete Geometry, Springer, New York, 2002. [19, 25, 70, 396]
J. Matousek, Using the BorsukUlam theorem, Springer, Heidelberg, 2003. [31, 38]
H. Matsumura, Commutative ring theory, Cambridge U. Press, Cambridge, 1986. [314]
R. Mazzeo and G. Montcouquiol, Innitesimal rigidity of cone-manifolds and the Stoker
problem for hyperbolic and Euclidean polyhedra; arXiv: 0908.2981. [256]
P. McMullen, Representations of polytopes and polyhedral sets, Geom. Dedicata 2 (1973),
8399; (). [328, 404]
P. McMullen, Valuations and dissections, in Handbook of convex geometry, 933988, NorthHolland, Amsterdam, 1993. [168]
P. McMullen and E. Schulte, Abstract regular polytopes, Cambridge U. Press, Cambridge,
UK, 2002. [184, 394]
H. Melissen, Loosest circle coverings of an equilateral triangle, Math. Mag. 70 (1997),
118124; (). [379]
W. Meyer, Indecomposable polytopes, Trans. AMS 190 (1974), 7786; (). [404]
M. D. Meyerson, Equilateral triangles and continuous curves, Fund. Math. 110 (1980), 19;
available at http://tinyurl.com/2pvbwn [380]
M. D. Meyerson, Convexity and the table theorem, Pacic J. Math. 97 (1981), 167169;
available at http://tinyurl.com/2jsj3l [381]
M. D. Meyerson, Balancing acts, Topology Proc. 6 (1981), 5975. [53, 380]
M. D. Meyerson and M. D. Asic, 6385, Amer. Math. Monthly 90 (1983), 651. [380]
A. D. Milka, On a conjecture of Stoker (in Russian), Ukrainian Geom. Sb. 9 (1970), 8586.
[-]
A. D. Milka, The isoperimetry of polygons on the sphere (in Russian), Ukrainian Geom.
Sb. 10 (1971), 4950. [272]
A. D. Milka, What is geometry in the large? (in Russian), Znanie, Moscow, 1986. [256,
272, 328]
A. D. Milka, Unique determination of convex surfaces in a pseudo-Riemannian spherical
space. I, J. Soviet Math. 51 (1990), 22732276; (). [262]

424

[Mi5]
[Mi6]
[Mill]
[MilP]

[MilS]
[Mil1]
[Mil2]
[Min]

[MMP]
[MlaO]
[Mnev]
[Mon1]
[Mon2]
[Mor]
[Mot]
[MucP]
[M
ul]

[Mus1]
[Mus2]
[Mus3]
[Mus4]
[Mus5]
[Mus6]
[NagC]
[NazP]

A. D. Milka, Linear bendings of regular convex polyhedra (in Russian), Mat. Fiz. Anal.
Geom. 1 (1994), 116130. [347, 405]
A. D. Milka, Uniqueness theorems for spacelike convex polyhedra in a pseudo-Euclidean
space (in Russian), Mat. Fiz. Anal. Geom. 2 (1995), 232249. [262]
E. N. Miller, Icosahedra constructed from congruent triangles, Discrete Comput. Geom. 24
(2000), 437451; available at http://tinyurl.com/2lq5qu [394]
E. Miller and I. Pak, Metric combinatorics of convex polyhedra: Cut loci and nonoverlapping unfoldings, Discrete Comput. Geom. 39 (2008), 339388; arXiv: math.MG/0312253.
[352, 358, 406]
E. Miller and B. Sturmfels, Combinatorial commutative algebra, Springer, New York, 2005.
[5]
J. W. Milnor, On the total curvature of knots, Ann. of Math. 52 (1950), 248257; (). [240]
J. Milnor, The Schlai dierential equality, in Collected Papers, Vol. 1, Geometry, 281295,
Publish or Perish, Houston, TX, 1994. [272]
H. Minkowski, Allegemeine Lehrsatze u
ber die konvexen Polyeder, see Gesammelte Abhandlungen, vol. II (1897), 103121; (), available at http://tinyurl.com/33y28j [70,
328]
J. S. B. Mitchell, D. M. Mount and C. H. Papadimitriou, The discrete geodesic problem,
SIAM J. Comp. 16 (1987), 647668; (), available at http://tinyurl.com/2prkzg [358]
I. Mladenov and J. Oprea, The mylar balloon revisited, Amer. Math. Monthly 110 (2003),
761784; (). [405]
N. Mnev, The universality theorems..., in Lecture Notes in Math. 1346, Springer, Berlin,
1988, 527543; (). [117]
P. Monsky, On dividing a square into triangles, Amer. Math. Monthly 77 (1970), 161
164; (). [402]
P. Monsky, A conjecture of Stein on plane dissections, Math. Z. 205 (1990), 583592;
available at http://tinyurl.com/6463dx [402]
H. R. Morton, Trefoil knots without tritangent planes, Bull. London Math. Soc. 23 (1991),
7880; (). [396]
T. S. Motzkin, The evenness of the number of edges of a convex polyhedron, Proc. Nat.
Acad. Sci. USA 52 (1964), 4445; available at http://tinyurl.com/2q9p9t [383]
R. Muchnik and I. Pak, On tilings by ribbon tetrominoes, J. Combin. Theory Ser. A 88
(1999), 188193; available at http://tinyurl.com/3aubdc [393]
C. M
uller, Equidecomposability of polyhedra with reference to crystallographic groups,
Discrete Comput. Geom. 3 (1988), 383389; available at http://tinyurl.com/3dspo6
[147]
O. R. Musin, A four-vertex theorem for a polygon (in Russian), Kvant (1997), no. 2, 1113;
available at http://tinyurl.com/23trtc [206, 395]
O. R. Musin, Properties of the Delaunay Triangulation, in Proc. 13th ACM Sympos. Comput. Geom., 424426, ACM Press, 1997; arXiv: math.MG/0312050 [387]
O. R. Musin, Chebyshev systems and zeros of a function on a convex curve, Proc. Steklov
Inst. Math. 221 (1998), 236246; (). [-]
O. R. Musin, Curvature extrema and four-vertex theorems for polygons and polyhedra,
J. Math. Sci. 119 (2004), no. 2, 268277; (); arXiv: 0805.0629. [206]
O. R. Musin, The kissing number in four dimensions, Ann. Math. 168 (2008), 132; arXiv:
math.MG/0309430. [191, 394]
O. R. Musin, The kissing problem in three dimensions, Discrete Comput. Geom. 35 (2006),
375384; arXiv: math.MG/0410324. [191]
Bela de Sz. Nagy and H. S. M. Coxeter, 3909, Amer. Math. Monthly 48 (1941), 8182;
() (). [383]
A. I. Nazarov and F. V. Petrov, On Tabachnikovs conjecture, St. Petersburg Math. J. 19
(2008), 125135; () arXiv: math.MG/ 0507379. [398]

425

[Naz]

F. L. Nazarov, Arrangement of pyramids in a convex polyhedron (in Russian), Soobshch.

Akad. Nauk Gruzin. SSR 132 (1988), 2932. [378]
[Nie]
M. J. Nielsen, Figures Inscribed in Curves, website http://tinyurl.com/2mwt6y [53]
[NieW] M. J. Nielsen and S. E. Wright, Rectangles inscribed in symmetric continua, Geom. Dedicata 56 (1995), 285297; (). [381]
[Niv]
I. Niven, Convex polygons that cannot tile the plane, Amer. Math. Monthly 85 (1978),
785792; (). [391]
[Olo1] S. Olovianishniko, On the bending of innite convex surfaces (in Russian), Mat. Sbornik 18
(1946), 429440. [262]
[Olo2] S. Olovianishniko, Generalisation du theor`eme de Cauchy sur les poly`edres convexes (in
Russian), Mat. Sbornik 18 (1946), 441446. [262]
[Ols]
A. Yu. Olshanskii, Planar graphs (in Russian), Soros Educational Journal 2 (1996), no. 11,
117-122; available at http://tinyurl.com/29clsz [384]
[OR]
J. ORourke, On the development of the intersection of a plane with a polytope, Comput.
Geom. 24 (2003), 310; arXiv: cs.CG/0006035. [396]
[ORS] J. ORourke and C. Schevon, On the development of closed convex curves on 3-polytopes,
J. Geom. 35 (1989), 152157; (). [396]
[Oss]
R. Osserman, The four-or-more vertex theorem, Amer. Math. Monthly 92 (1985), 332
337; (). [206]
[OT1] V. Ovsienko and S. Tabachnikov, Projective geometry of polygons and discrete 4-vertex
and 6-vertex theorems, Enseign. Math. 47 (2001), 319; arXiv: math.DG/9909150. [396]
[OT2] V. Ovsienko and S. Tabachnikov, Projective dierential geometry old and new, Cambridge
U. Press, Cambridge, UK, 2005; available at http://tinyurl.com/3xhc8c [206]
[Oza1] T. Ozawa, On Halperns conjecture for closed plane curves, Proc. AMS 92 (1984), 554
560; (). [396]
[Oza]
T. Ozawa, On the numbers of n-tangencies of polygons in odd dimensional Euclidean spaces,
J. Geom. 32 (1988), 6985; (). [396]
[Pach] J. Pach, Folding and turning along geodesics in a convex surface, Geombinatorics 7 (1997),
6165; available at http://tinyurl.com/3556n2 [400, 406]
[PacA] J. Pach and P. K. Agarwal, Combinatorial Geometry, John Wiley, New York, 1995. [25,
108, 396]
[Pak1] I. Pak, Four questions on Birkho polytope, Ann. Comb. 4 (2000), 8390; available at
http://tinyurl.com/5lga57 [393]
[Pak2] I. Pak, On the number of faces of certain transportation polytopes, Europ. J. Combin. 21
(2000), 689694; available at http://tinyurl.com/23we4ct [384]
[Pak3] I. Pak, Tile invariants: new horizons, Theor. Comp. Sci. 303 (2003), 303331; available at
http://tinyurl.com/3bckk2 [138, 228, 300, 389]
[Pak4] I. Pak, The area of cyclic polygons: recent progress on Robbins conjectures, Adv. Applied
Math. 34 (2005), 690696; arXiv: math.MG/0408104. [291, 314, 401, 403]
[Pak5] I. M. Pak, A short proof of the rigidity of convex polytopes, Sib. Math. J. 47 (2006),
710713; (), arXiv: math.MG/0602579. [306]
[Pak6] I. Pak, Partition bijections, a survey, Ramanujan Journal 12 (2006), 575; available at
http://tinyurl.com/3bt9sk [391, 393]
[Pak7] I. Pak, Inating the cube without stretching, Amer. Math. Monthly 115 (2008), 443445;
arXiv: math.MG/0607754. [347]
[Pak8] I. Pak, Inating polyhedral surfaces, preprint; available at http://tinyurl.com/237lz7
[347]
[Pak9] I. Pak, The discrete square peg problem; arXiv: 0804.0657. [228]
[PakP] I. Pak and R. Pinchasi, How to cut out a convex polyhedron, preprint; available at
http://www.math.umn.edu/~pak [406]
[Pan]
G. Panina, A. D. Alexandrovs uniqueness theorem for convex polytopes and its renements
Beitr. Algebra Geom. 49 (2008), 5970; available at http://tinyurl.com/2laqc8 [404]

426

[Pau]

W. Paulsen, What is the shape of a mylar balloon?, Amer. Math. Monthly 101 (1994),
953958. [405]
[PfeZ] F. Pfender and G. M. Ziegler, Kissing numbers, sphere packings, and some unexpected
proofs, Notices AMS 51 (2004), 873883; available at http://tinyurl.com/2otq5e [191]
[Phi]
A. Philippakis, The orthic triangle and the o.k. quadrilateral, Amer. Math. Monthly 109
(2002), 704728; (). [385]
[Pinc] V. Pincu, On the Fewest Nets Problem for Convex Polyhedra, in Proc. 19th CCCG (2007),
2124; available at http://tinyurl.com/qolujx [406]
[Pink] U. Pinkall, On the four-vertex theorem, Aeq. Math. 34 (1987), 221230; available at
http://tinyurl.com/2ogbgz [206]
[Pog1] A. V. Pogorelov, Quasi-geodesic lines on a convex surface, in AMS Transl. 74 (1952), no. 74,
45 pp.; (). [101]
[Pog2] A. V. Pogorelov, A new proof of rigidity of convex polyhedra (in Russian), Uspehi Mat.
Nauk 11 (1956), 207208. [271]
[Pog3] A. V. Pogorelov, Extrinsic geometry of convex surfaces, AMS, Providence, RI, 1973; ().
[262, 321, 328, 339, 403, 405]
[Pog4] A. V. Pogorelov, The Minkowski multidimensional problem, John Wiley, New York,
1978; (). [328]
[Pog5] A. V. Pogorelov, Bendings of surfaces and stability of shells, AMS, Providence, RI,
1988; (). [347]
[Pog6] A. V. Pogorelov, On a problem of Stoker, Dokl. Akad. Nauk 385 (2002), no. 1, 2527; ().
[402]
[Pok]
V. G. Pokrovskii, Existence of polyhedra that do not decompose into equally large simplices,
Math. Notes 46 (1989), 785787; (), (). [402]
[Pol]
N. Polakova, On one property of convex polyhedra, Spisy Prrod. Fak. Univ. Brno 507
(1969), 337339. [408]
[Post] K. A. Post, Geodesic lines on a bounded closed convex polyhedron, Studia Sci. Math.
Hungar. 5 (1970), 411416. [386]
[PouS] D. Poulalhon and G. Schaeer, Optimal coding and sampling of triangulations, Algorithmica 46 (2006), 505527; available at http://tinyurl.com/22k57z [387]
` G. Poznjak, Non-rigid closed polyhedra (in Russian), Vestnik Moskov. Univ. Ser. I Mat.
[Poz]
E.
Meh. 3 (1960), 1419. [291]
[PraS] V. V. Prasolov and I. F. Sharygin, Problems in stereometry (in Russian), Nauka, Moscow,
1989; (J). [394, 407]
[Pro1] V. Yu. Protasov, Maxima and minima in geometry (in Russian), MCCME, Moscow,
2005; (J). [384]
[Pro2] V. Yu. Protasov, Closed geodesics on the surface of a simplex, Mat. Sbornik 198 (2007),
243260; (), (). [386]
[Rama] J. F. Ramaley, Buons noodle problem, Amer. Math. Monthly 76 (1969), 916918; ().
[397]
[Ramb] J. Rambau, On a generalization of Schonhardts polyhedron, in Combinatorial and computational geometry, 501516, Cambridge Univ. Press, Cambridge, 2005; available at
http://tinyurl.com/58np8m [388]
[Ram] J. L. Ramrez Alfonsn, Knots and links in spatial graphs: a survey, Discrete Math. 302
(2005), 225242; (). [397]
[Res1] Yu. G. Reshetnyak, Estimation on the visibility angle of a curve in a metric space of
nonpositive curvature, Sib. Math. J. 43 (2002), 562567; (), (). [398]
[Res2] Yu. G. Reshetnyak, The theory of curves in dierential geometry from the point of view
of the theory of functions of a real variable, Russian Math. Surveys 60 (2005), 11651181;
(), (). [398]
[Ric]
J. RichterGebert, Realization Spaces of Polytopes, in Lecture Notes in Math. 1643,
Springer, Berlin, 1997; available at http://tinyurl.com/2vs7ze [5, 108, 117]

427

[Riv]

I. Rivin, Some observations on the simplex, to appear in Bolyai conference proceedings;

arXiv: math.MG/0308239. [402]
[Ros]
J. T. Rosenbaum, Some consequences of the universal chord theorem, Amer. Math.
Monthly 78 (1971), 509513; (). [379]
[Roth] B. Roth, Rigid and exible frameworks, Amer. Math. Monthly 88 (1981), 621; (). [300]
[RotW] B. Roth and W. Whiteley, Tensegrity frameworks, Trans. AMS 265 (1981), 419446; ().
[402]
[RouS] C. Rourke and B. Sanderson, Introduction to Piecewise-Linear Topology, Springer, Berlin,
1972; (). [176]
[Rub]
V. S. Rublev, A characteristic property of the class of convex sets, Math. Notes 2 (1967),
677679; (), (). [377]
[Rud]
M. E. Rudin, An unshellable triangulation of a tetrahedron, Bull. Amer. Math. Soc. 64
(1958), 9091; available at http://tinyurl.com/2wd3vo [393]
[Rum] D. Rumsfeld, Rumsfelds Rules, Wall Street Journal, January 29, 2001; available at
http://tinyurl.com/5n46ux [5]
[RupS] J. Ruppert and R. Seidel, On the diculty of tetrahedralizing 3-dimensional
non-convex polyhedra, Discrete Comput. Geom. 7 (1992), 227253; available at
http://tinyurl.com/7bowj3 [388]
[Sab1] I. Kh. Sabitov, Local theory of bendings of surfaces, in Geometry III, 179250, Enc. Math.
Sci., vol. 48, Springer, Berlin, 1992; (). [262, 314, 339]
[Sab2] I. Kh. Sabitov, On the problem of invariance of the volume of exible polyhedron, Russian
Math. Surveys 50 (1995), 451452; (). [314]
[Sab3] I. Kh. Sabitov, The generalized HeronTartaglia formula and some of its consequences,
Sbornik Math. 189 (1998), 15331561; (). [314]
[Sab4] I. Kh. Sabitov, The volume as a metric invariant of polyhedra, Discrete Comput. Geom.
20 (1998), 405425; (). [314]
[Sab5] I. Kh. Sabitov, Volumes of polytopes (in Russian), MCCME, Moscow, 2002; (). [285]
[Sab6] I. Kh. Sabitov, Around the proof of the LegendreCauchy lemma on convex polygons, Sib.
Math. J. 45 (2004), 740762; (). [184, 215]
[SGK] V. A. Sadovnichii, A. A. Grigoryan and S. V. Konyagin, Student mathematical olympiad
problems (in Russian), Moscow U. Press, Moscow, 1987. [407]
[Sah]
C. H. Sah, Hilberts third problem: scissors congruence, Pitman, Boston, MA, 1979. [5, 147]
[Sant] L. A. Santalo, Integral geometry and geometric probability, Addison-Wesley, Reading, MA,
1976; (). [272, 381, 397]
[San1] F. Santos, On the renements of a polyhedral subdivision, Collect. Math. 52 (2001), 231
256; available at tt http://tinyurl.com/2t56lg [388]
[San2] F. Santos, Geometric bistellar ips: the setting, the context and a construction, in Proc.
ICM Madrid, Vol. III, EMS, Z
urich, 2006, 931962; arXiv: math.CO/0601746. [138, 166,
169, 387]
[Sany] R. Sanyal, Topological obstructions for vertex numbers of Minkowski sums; arXiv:
math.CO/0702717. [408]
[SWZ] R. Sanyal, A. Werner and G. M. Ziegler, On Kalais conjectures concerning centrally symmetric polytopes, arXiv:0708.3661. [384]
[Sar]
S. Saraf, Acute and non-obtuse triangulations of polyhedral surfaces, preprint (2007). [387]
[Sche] K. Scherer, The impossibility of a tessellation of the plane..., Elem. Math. 38 (1983), 14;
available at http://tinyurl.com/2uq87q [409]
[Schl1] J.-M. Schlenker, Convex polyhedra in Lorentzian space-forms, preprint, 1998; available at
http://tinyurl.com/33sz88 [262]
[Schl2] J.-M. Schlenker, Dihedral angles of convex polyhedra, Discrete Comput. Geom. 23 (2000),
409417; (). [256, 400]
[Schl3] J.-M. Schlenker, La conjecture des souets, dapr`es I. Sabitov, Asterisque 294 (2004),
7795; available at http://tinyurl.com/2mdo25 [314]

428

[Schl4]
[SchlS]
[Schme]
[Schmi]
[Schn1]
[Schn2]
[Scho]
[SchoZ]
[Scho]
[Schra]
[Schr1]
[Schr2]
[SchrS]

[Schu1]
[Schu1]
[Schw]
[Sed]
[Sene]
[Sen1]
[Sen2]
[ShaH]
[ShaS]
[Shar]
[She1]
[She2]
[SCY]

[Shn]

J.-M. Schlenker, Small deformations of polygons, Trans. AMS 359 (2007), 21552189;
arXiv: math.DG/0410058. [215, 271]
J.-M. Schlenker and R. Souam, Higher Schlai formulas and applications, Compositio
Math. 135 (2003), 124; available at http://tinyurl.com/2qnw3e [272]
J. H. Schmerl, Tiling space with notched cubes, Discrete Math. 133 (1994), 225235 [390]
P. Schmitt, Spacelling knots, Beitr. Algebra Geom. 38 (1997), 307313; arXiv:
math.MG/0107016. [390]
R. Schneider, Equidecomposable polyhedra, in Intuitive geometry, North-Holland, Amsterdam, 1987. [390]
R. Schneider, Convex bodies: the BrunnMinkowski theory, Cambridge U. Press, Cambridge, UK, 1993. [70, 328, 404]
I. J. Schoenberg, An isoperimetric inequality for closed curves convex in even-dimensional
Euclidean spaces, Acta Math. 91 (1954), 143164. [378]
I. J. Schoenberg and S. C. Zaremba, On Cauchys lemma concerning convex polygons,
Canad. J. Math. 19 (1967), 10621071; available at http://tinyurl.com/5zcag7 [215]

E. Schonhardt, Uber
die Zerlegung von Dreieckspolyedern in Tetraeder, Math. Ann. 98
(1928), 309312; available at http://tinyurl.com/65yked [388]
O. Schramm, How to cage an egg, Invent. Math. 107 (1992), 543560; available at
http://tinyurl.com/3bbr97 [108, 403]
P. Schreiber, Theorie der geometrischen Konstruktionen, Studienb
ucherei, VEB Deutscher
Verlag der Wissenschaften, Berlin, 1975. [116]
P. Schreiber, On the existence and constructibility of inscribed polygons, Beitr. Algebra
Geom. 34 (1993), 195199; available at http://tinyurl.com/2r7phu [387]
Y. Schreiber and M. Sharir, An optimal-time algorithm for shortest paths on a convex
polytope in three dimensions, Discrete Comput. Geom. 39 (2008), 500579; available at
http://tinyurl.com/62xlor [358, 406]
L. J. Schulman, An equipartition of planar sets, Discrete Comput. Geom. 9 (1993), 257266;
available at http://tinyurl.com/2rus7y [379]
E. Schulte, Tiling three-space by combinatorially equivalent convex polytopes, Proc. London Math. Soc. 49 (1984), 128140; available at http://tinyurl.com/pr5oul [387]
R. E. Schwartz, Billiards obtuse and irrational; available at http://tinyurl.com/652jmm
[94]
V. D. Sedykh, Discrete versions of the four-vertex theorem, in Topics in singularity theory,
AMS, Providence, RI, 1997, 197207. [206]
M. Senechal, Which tetrahedra ll space? Math. Mag. 54 (1981), 227243; (). [158, 392]
E. P. Senkin, Unique determination of convex polyhedra (in Russian), Uspehi Mat. Nauk 11
(1956), 211213. [277]
E. P. Senkin, Bending of convex surfaces, J. Soviet Math. 14 (1980), 12871305; (). [262,
339]
M. I. Shamos and D. Hoey, Closest-point problems, in Proc. 16th FOCS (1975), 151162,
IEEE, Long Beach, CA. [138, 387]
M. Sharir and A. Schorr, On shortest paths in polyhedral spaces, SIAM J. Comp. 15
(1986), 193215; (). [358, 386, 406]
I. F. Sharygin, Problems in geometry (in Russian), Nauka, Moscow, 1984; (J). [386, 407]
G. C. Shephard, A sphere in a crate, J. London Math. Soc. 40 (1965), 433434. [382]
G. C. Shephard, Twenty problems on convex polyhedra. II, Math. Gaz. 52 (1968), 359
367; (). [407]
D. O. Shklarsky, N. N. Chentzov and I. M. Yaglom, Selected problems and theorems of
elementary mathematics, vol. 3 (in Russian), GITTL, Moscow, 1954; (J). [248, 381, 382,
395]
L. G. Shnirelman, On certain geometrical properties of closed curves (in Russian), Uspehi
Matem. Nauk 10 (1944), 3444; available at http://tinyurl.com/28gsy3 [53, 228]

429

[Sho1]
[Sho2]
[Sht1]
[Sht2]
[Sht3]
[SST]
[STT]
[SloV]
[Smo]
[SnoS]
[Sos]
[Spi]
[Spr]
[St1]
[St2]
[Sta1]
[Sta2]
[Sta3]
[Ste1]
[Ste2]
[SteS]
[Sto]
[StrL]
[Stre]
[Stro]
[Stru]
[Sul]
[Sun]

L. A. Shor, On the exibility of convex polygons with a slit (in Russian), Ukrainian Mat.
16 (1964), 513520. [339]
Z.
L. A. Shor, The exibility of convex surfaces with boundary, Ukrainian Geom. Sb. 5-6
(1968), 177196; (), (). [339]
M. I. Shtogrin, Piecewise-smooth embedding of a cube, Russian Math. Surveys 59 (2004),
979981; (), (). [347, 405]
M. I. Shtogrin, Special isometric transformations of the surfaces of the Platonic solids,
Russian Math. Surveys 60 (2005), 799801; (), (). [405]
M. I. Shtogrin, Isometric embeddings of the surfaces of Platonic solids, Russian Math.
Surveys 62 (2007), 395397; (), (). [405]
J. D. Skinner, C. A. B. Smith and W. T. Tutte, On the dissection of rectangles into rightangled isosceles triangles, J. Combin. Theory Ser. B 80 (2000), 277319; (). [409]
D. D. Sleator, R. E. Tarjan and W. P. Thurston, Rotation distance, triangulations, and
hyperbolic geometry, J. Amer. Math. Soc. 1 (1988), 647681; (). [387]
N. J. A. Sloane and V. A. Vaishampayan, Generalizations of Schobis Tetrahedral Dissection, arXiv: 0710.3857. [392]
A. S. Smogorzhevskij, The ruler in geometrical constructions, Pergamon Press, Oxford,
1961; (J). [116]
J. Snoeyink and J. Stol, Objects that cannot be taken apart with two hands, Discrete
Comput. Geom. 12 (1994), 367384; available at http://tinyurl.com/32o5yy [409]
A. Sossinsky, Knots, Harvard U. Press, Cambridge, MA, 2002. [240]
A. Spiridonov, Rigidity of the n-dimensional frame, MIT preprint; available at
http://tinyurl.com/34ehud [401]
B. A. Springborn, Variational principles for circle patterns, Ph.D. thesis, TU Berlin, 2003;
arXiv: math.GT/0312363. [108]
H. Stachel, Flexible cross-polytopes in the Euclidean 4-space, J. Geom. Graph. 4 (2000),
159167; available at http://tinyurl.com/3au9np [285, 401]
H. Stachel, Flexible octahedra in the hyperbolic space, in Non-Euclidean Geometries,
Kluwer, Dordrecht, 2004, 209225; available at http://tinyurl.com/2mjc85 [285]
R. P. Stanley, Eulerian partitions of a unit hypercube, in Higher Combinatorics, Reidel,
Boston, 1977, p. 49; available at http://tinyurl.com/26panjb [384]
R. P. Stanley, Two poset polytopes, Discrete Comput. Geom. 1 (1986), 923; available at
http://tinyurl.com/2wzkfvq [393]
R. P. Stanley, Enumerative Combinatorics, vol. 1 and 2, Cambridge U. Press, Cambridge,
UK, 1997 and 1999; (). [82, 175, 390]
S. Stein, The notched cube tiles Rn . Discrete Math. 80 (1990), 335337, (). [390]
S. Stein, Cutting a polygon into triangles of equal areas, Math. Intelligencer 26 (2004),
no. 1, 1721. [402]
S. Stein and S. Szabo, Algebra and tiling, MAA, Washington, DC, 1994. [390]
J. J. Stoker, Geometrical problems concerning polyhedra in the large, Comm. Pure Appl.
Math. 21 (1968), 119168. [215, 256, 300]
J. Strantzen and Y. Lu, Regular simple geodesic loops on a tetrahedron, Geom. Dedicata 42
(1992), 139153; (). [386]
I. Streinu, Pseudo-triangulations, rigidity and motion planning, Discrete Comput. Geom. 34
(2005), 587635; available at http://tinyurl.com/3dr5yr [228]
W. Stromquist, Inscribed squares and square-like quadrilaterals in closed curves, Mathematika 36 (1989), 187197. [53]
D. J. Struik, Dierential geometry in the large, Bull. AMS 37 (1931), 4962; available at
http://tinyurl.com/2p3zb4 [379]
J. M. Sullivan, Curves of nite total curvature, in Discrete Dierential Geometry,
Birkhauser, Boston, MA, 2008, 137161; arXiv: math.GT/0606007. [397]
Sun Tzu, The Art of War, 6th century BC; available at http://tinyurl.com/397w6 [95]

430

[Svo]
[Syd1]

A. Svoboda, Computing Mechanisms and Linkages, McGraw-Hill, New York, 1948. [124]
J.-P. Sydler, Sur la decomposition des poly`edres, Comment. Math. Helv. 16 (1944), 266
273; available at http://tinyurl.com/2tg6c2 [157]
[Syd2] J.-P. Sydler, Sur les tetra`edres equivalent `a un cube, Elem. Math. 11 (1956), 7881; available at http://tinyurl.com/3dj8yr [392]
[Tab1] S. Tabachnikov, Around four vertices, Russian Math. Surveys 45 (1990), 229230; (), ().
[94]
[Tab2] S. Tabachnikov, Billiards, 1995; available at http://tinyurl.com/2jddgl [94, 385]
[Tab3] S. Tabachnikov, The four-vertex theorem revisitedtwo variations on the old theme, Amer.
Math. Monthly 102 (1995), 912916; available at http://tinyurl.com/26wd7o [384]
[Tab4] S. Tabachnikov, A four vertex theorem for polygons, Amer. Math. Monthly 107 (2000),
830833; available at http://tinyurl.com/33aan3 [206]
[Tab5] S. Tabachnikov, The tale of a geometric inequality, in MASS selecta, AMS, Providence, RI,
2003, 257262; available at http://tinyurl.com/2noaoh [398]
[Tab6] S. Tabachnikov, Geometry and billiards, AMS, Providence, RI, 2005; available at
http://tinyurl.com/349boy [94, 385]
[Tab7] S. Tabachnikov, A proof of Culters theorem on the existence of periodic orbits in polygonal
outer billiards, Geom. Dedicata 129 (2007), 8387; arXiv: 0706.1003. [385]
[Tao]
T. Tao, Solving mathematical problems: a personal perspective, Oxford U. Press, New
York, 2006. [407]
[Tar1] A. S. Tarasov, Polyhedra that do not admit natural unfoldings, Russian Math. Surveys 54
(1999), 656657; (), (). [406]
[Tar2] A. S. Tarasov, Solution of Arnolds folded ruble problem (in Russian), Chebysh. Sb. 5
(2004), no. 1, 174187. [358, 406]
[Tar3] A. S. Tarasov, Existence of a polyhedron which does not have a non-overlapping pseudoedge unfolding, preprint (2008). [358]
[Tas1] S. S. Tasmuratov, Deformation of a polygon into a polyhedron with a given boundary, Sib.
Math. J. 15 (1974), 947953; (). [347, 405]
[Tas2] S. S. Tasmuratov, Bending of a nitely connected polygon into a polyhedron with a given
boundary (in Russian), Geometry 6 (1977), 102107. [347, 405]
[The]
V. Thebault, On the isosceles tetrahedron, Amer. Math. Monthly 62 (1955), 356358; ().
[407]
[Tho1] C. Thomassen, Straight line representations of innite planar graphs, J. London Math.
Soc. 16 (1977), 411423; (). [387]
[Tho2] C. Thomassen, Tuttes spring theorem, J. Graph Theory 45 (2004), 275280; (). [108]
[ThoU] G. Thorbergsson and M. Umehara, Sextactic points on a simple closed curve, Nagoya Math.
J. 167 (2002), 5594; arXiv: math.DG/0008137. [206]
[Thu]
W. P. Thurston, Shapes of polyhedra and triangulations of the sphere, Geom. Topol.
Monogr. 1 (1998), 511549; arXiv: math.GT/9801088. [403]
[TVS] C. Tian, V. A. Vaishampayan and N. J. A. Sloane, Constant Weight Codes: A Geometric
Approach Based on Dissections, arXiv: 0706.1217. [392]
[Tot]
V. Totik, A tale of two integrals, Amer. Math. Monthly 106 (1999), 227240 [38, 379]
[Tou]
G. Toussaint, A new class of stuck unknots in Pol6 , Beitr. Algebra Geom. 42 (2001), 301
306; available at http://tinyurl.com/34t66s [396]
[Tru]
V. I. Trushkina, A coloring theorem and the rigidity of a convex polyhedron (in Russian),
Ukrainian Geom. Sb. 24 (1981), 116122. [306]
[Tut]
W. T. Tutte, Graph theory as I have known it, Oxford U. Press, New York, 1998. [387, 392,
409]
[Tuza] Zs. Tuza, Dissections into equilateral triangles, Elem. Math. 46 (1991), 153158; available
at http://tinyurl.com/3djukz [409]
[Tve]
H. Tverberg, How to cut a convex polytope into simplices, Geom. Dedicata 3 (1974), 239
240; (). [383]

431

[Ume]

M. Umehara, A unied approach to the four vertex theorems. I, in Dierential and symplectic topology of knots and curves, AMS, Providence, RI, 1999, 185228. [206]
[Vard] I. Vardi, Mekh-Mat entrance examination problems, IHES preprint M00-06; available at
http://tinyurl.com/34dx6a [387, 406]
[Var1] V. V. Varfolomeev, Inscribed polygons and Heron polynomials, Sb. Math. 194 (2003),
311331; (), (). [401, 403]
[Var2] V. V. Varfolomeev, Galois groups of the Heron-Sabitov polynomials for pentagons inscribed
in a circle, Sb. Math. 195 (2004), 149162; (), (). [387]
[VD1] P. L. Varkonyi and G. Domokos, Static equilibria of rigid bodies: dice, pebbles, and the
Poincare-Hopf theorem, J. Nonlinear Sci. 16 (2006), 255281; (). [94, 385]
[VD2] P. L. Varkonyi and G. Domokos, Mono-monostatic bodies: the answer to Arnolds question,
Math. Intelligencer 28 (2006), no. 4, 3438; available at http://tinyurl.com/2nvnmv
[385]
[VasE] N. B. Vasiliev and A. A. Egorov, The USSR olympiad problems (in Russian), Nauka,
Moscow, 1988. [228, 382, 407]
[Vol1] Ju. A. Volkov, On deformations of a convex polyhedral angle (in Russian), Uspehi Mat.
Nauk 11 (1956), 209210. [271]
[Vol2] Ju. A. Volkov, Existence of convex polyhedra with prescribed development. I (in Russian),
Vestnik Leningrad. Univ. 15 (1960), 7586; English translation is given in [A2]. [334, 406]
[Vol3] Ju. A. Volkov, An estimate of the deformation of a convex surface as a function of the
change in its intrinsic metric (in Russian), Ukrain. Geometr. Sb. 56 (1968), 4469. [335,
405]
[VP1] Ju. A. Volkov and E. G. Podgornova, Existence of a convex polyhedron with a given evolute
(in Russian), in Taskent. Gos. Ped. Inst. Ucen. Zap. 85 (1971), 354, 83. [334]
[VP2] Ju. A. Volkov and E. G. Podgornova, The cut locus of a polyhedral surface of positive
curvature (in Russian), Ukrainian Geom. Sb. 11 (1971), 1525. [358]
[Weg1] B. Wegner, Existence of four concurrent normals to a smooth closed hypersurface of E n .
Amer. Math. Monthly 80 (1973), 782785; (). [385]
[Weg2] B. Wegner, A cyclographic approach to the vertices of plane curves, J. Geom. 50 (1994),
186201; (). [206]
[Weg3] B. Wegner, On the evolutes of piecewise linear curves in the plane, Rad Hrvatske Akad.
Znan. Umjet. 467 (1994), 116. [396]
[Weg4] B. Wegner, Boses vertex theorem for simply closed polygons, Math. Pannon. 6 (1995),
121132. [206]
[Whi1] W. Whiteley, Innitesimally rigid polyhedra I. Statics of frameworks, Trans. AMS 285
(1984), 431465; (). [300]
[Whi2] W. Whiteley, Rigidity and scene analysis, in Handbook of discrete and computational geometry, CRC Press, Boca Raton, FL, 1997, 893916; (). [300, 302]
[Whit] S. Whitesides, Chain reconguration: the ins and outs, ups and downs of moving polygons
and polygonal linkages, in Lecture Notes in Comput. Sci. 2223, Springer, Berlin, 2001,
113; available at http://tinyurl.com/27pr3c [228]
[Wu]
Y.-Q. Wu, Inscribing smooth knots with regular polygons, Bull. LMS 36 (2004), 176180;
arXiv: math.GT/0610867. [380]
[Yag1] I. M. Yaglom, How to cut a square? (in Russian), Nauka, Moscow, 1968; (J). [409]
[Yag2] I. M. Yaglom, The problem of thirteen spheres (in Russian), Visha Shkola, Kiev, 1975; (J).
[191]
[YagB] I. M. Yaglom and V. G. Boltyanski, Convex gures, Holt, Rinehart and Winston, New
York, 1961; (), (J). [377, 379]
[YamY] H. Yamabe and Z. Yujobo, On the continuous function dened on a sphere, Osaka Math.
J. 2 (1950), 1922. [60, 381]
[Yas]
I. Yaschenko, Make your dollar bigger now!, Math. Intelligencer 20 (1998), no. 2, 3640.
[358, 406]

432

[Zak1]
[Zak2]
[Zal1]
[Zal2]
[Zal3]
[Zam]
[Zar]
[Zie1]
[Zie2]
[Zie3]
[Zil]
[Zon1]
[Zon2]
[Zyl]

J. Zaks, The chair theorem, in Proc. Second Louisiana Conference on Combinatorics, Graph
Theory and Computing, 557562, LSU, Baton Rouge, LA, 1971. [381]
J. Zaks, No nine neighborly tetrahedra exist, Mem. AMS 91 (1991), 106 pp. [409]
V. A. Zalgaller, On deformations of a polygon on a sphere (in Russian), Uspehi Mat.
Nauk 11 (1956), 177178. [277]
V. A. Zalgaller, Isometric imbedding of polyhedra (in Russian), Dokl. Akad. Nauk
SSSR 123 (1958), 599601. [358, 405]
V. A. Zalgaller, Convex Polyhedra with Regular Faces, Consultants Bureau, New York,
1969; (). [184]
T. Zamrescu, Acute triangulations: a short survey, in Proc. Sixth Conf. of R.S.M.S.,
Bucharest, Romania, 2003, 1018. [387]
K. Zarankiewicz, Un theor`eme sur luniformisation des fonctions continues ..., Bull. Acad.
Polon. Sci. 2 (1954), 117120. [381]
G. M. Ziegler, Lectures on Polytopes, Springer, New York, 1995. [81, 82, 103, 108, 117, 382]
G. M. Ziegler, Convex polytopes: extremal constructions and f -vector shapes, in Geometric
Combinatorics, AMS, Providence, RI, 2007, 617691; arXiv: math.MG/0411400. [104, 108]
G. M. Ziegler, Non-rational congurations, polytopes, and surfaces, Math. Intelligencer 30
(2008), no. 3, 3642; arXiv: 0710.4453. [387]
A. A. Zil berberg, On the existence of closed convex polyhedra with arbitrary prescribed
curvatures of vertices (in Russian), Uspehi Mat. Nauk 17 (1962), 119126. [403]
C. Zong, Sphere packings, Springer, New York, 1999. [191]
C. Zong, What is known about unit cubes, Bull. AMS 42 (2005), 181211; available at
http://tinyurl.com/2mx3lk [407]
V. B. Zylev, G-composedness and G-complementability, Soviet Math. Dokl. 9 (1968), 403
404; (). [157]

Notation for the references:

Whenever possible, we add free web links to publications or their arXiv numbers:
http://front.math.ucdavis.edu/
In all cases, try http://scholar.google.com search on the title.
()
()
(J)

means that the item in English is available electronically by subscription, either

from JSTOR or from the publishers web site
means that the item is available in Russian either as an original work or as a
translation, but may not be available electronically
means that the item in Russian is freely available from http://math.ru/

The web links are clickable in this le, and are either to web pages with unrestricted access to
papers, or to papers themselves (mostly in .pdf, but some in .ps or .ps.gz format). The links are
redirected via the http://TinyURL.com web site. If you wish to see the site addresses (and le
extensions) before opening them, go to the company web site and use the preview feature.

Index
2-move, see also ip
in domino tilings, 135
in planar triangulations, 125
3d -conjecture, 81

icosidodecahedron, 157, 383

rhombic dodecahedron, 394
truncated cube, 153
truncated cuboctahedron, 154
truncated dodecahedron, 383
truncated octahedron, 154
truncated tetrahedron, 154
Archimedes, 140
arm lemma, 208212, 215
extension to non-convex polygons, 214
weak version, 273, 276
Arnolds rouble problem, 358, see also
napkin folding problem
Arnold, V. I., 94, 358
aspect ratio (of a rectangle), 51, 58
associahedron, 75, 82, 126
Aumann, Georg, 377
average kissing number, 191
average width, 69

acyclic orientation, 72, 76

additive function, 161
ane Coxeter group, 156
Akopyan, Arseny, 391
Alexander, James Waddell, 168
Alexander, Ralph, 271
Alexander, Stephanie B., 240
Alexandrov unfolding, 349, 355
Alexandrovs existence theorem, 329
Alexandrovs lemma
general version, 213
local version, 214
Alexandrovs theorem
on convexly bendable surfaces, 338
on curvature of polytopes which lie on
rays, 317
on curvatures on convex caps, 318
on existence of the polyhedral surface, 329
on parallel polytopes, 253
on rigidity of triangulated convex surfaces,
300, 305
on uniqueness of polyhedral surfaces, 260
on uniqueness of polytopes with given face
normals, 326
the total curvature bound, 238
Alexandrov, A. D., 215, 262, 300, 321, 328,
330, 334, 339, 398
Alexandrov, V. A., 285
AlexandrovFenchel inequality, 71, 334
algorithm
continuous Dijkstra, 358
for geometric numbers, 110
Alon, Noga, 38
alternating permutation, 175
Amp`ere, AndreMarie, 38
angle
defect, 248, see also curvature
exterior, 230
solid, 243
visibility, of a path, 238
angular velocity equation, 264, 271, 295
antiprism, 255, 394
aperiodic tiling, 154, 158
Archimedean solid, 184
cuboctahedron, 153, 394

Baranys theorem, 20, 25

Barany, Imre, 25
Baker, M. J. C., 377
Balinskis theorem, 77, 82
BanachTarski paradox, 158
Bancho, Thomas F., 285, 401
Barbiers theorem, 379
bellows conjecture, 287, 307
to surfaces of higher genus, 313
bending construction
Bleeckers, 340
BuragoZalgallers, 340
Milkas, 342
bending of surfaces, 336
Benjamini, Itai, 395
Berge, Claude, 377
Bezdek, Karoly, 379
Bilinski, Stanko, 206
billiard trajectory
dense, 93
dual, 93
in a polygon, 90
periodic, 89
quasi, in a polygon, 90
Birkho polytope, 81, 175
Birkhos theorem
on billiard trajectories between points, 90
on periodic billiard trajectories, 89
on two geodesics, 97
Birkho, George David, 97
433

434

bisector, 36
Bishop, Richard L., 240
Blaschke selection theorem, 68, 71
Blaschke, Wilhelm, 206, 339, 379
BlindMani theorem, 76
blooming conjecture, 352
Boltyansky, Vladimir G., 147
BolyaiGerwien theorem, 139
Bonnesens inequality, 69
Bonnesen, Tommy, 381
BorosF
uredi theorem, 25, 33, 38
Borsuks conjecture, 26, 30, 31
Borsuks theorem, 27
Borsuk, Karol, 31, 240
Bose, Raj Chandra, 206
Brahmaguptas formula, 312
Breen, Marilyn, 377
Bricards condition, 142, 144
Bricards octahedron, 281
Bricard, Raoul, 147, 285
Brunns theorem, 64
BrunnMinkowski inequality, 64, 70, 328, 362
Buons needle, 237
Bugaenko, V. O., 381
Burago, Yu. D., 347
BuragoZalgallers theorem, 335, 347
Caratheodorys theorem, 20
colorful, 21, 25
Steinitz extension of, 23
Carpenters rule theorem, 223
Catalan number, 333
Catalan numbers, 75, 82
Catalan solid, 184
rhombic dodecahedron, 183
rhombic triacontahedron, 183
Cauchy problem (analytic), 321
Cauchys theorem, 249, 266, 273, 280
Alexandrovs extension to parallel
polytopes, 253
application to regular polytopes, 180
for non-strictly convex surfaces, 260
for spherical polyhedra, 257
for unbounded polyhedra in R3 , 259
in higher dimensions, 257
rigidity in Rd , 258
rigidity of convex polytopes in R3 , 249, 253
Stokers inverse, 255
Cauchy, Augustin Louis, 215, 256
CauchyAlexandrov theorem for polyhedral
surfaces, 260
caustic, 198, see also evolute

Cayleys formula, 81, 154

Cayley, Arthur, 206
CayleyMenger bideterminant, 374
CayleyMenger determinant, 310, 364
center of mass, 8, 47, 83, 85
Cevas theorem, 113
Chaboud, Thomas, 389
Chadwick, Lee S., 346
chain polytope, 175
chair theorem, 52
Chebyshev polynomials, 123, 289
Chebyshev, Pafnuty L., 123
circle
disjoint, 195
empty, 127, 195
equipartitioning the polygon, 50
full, 195
inscribed in a convex polygon, 197
intermediate, 195
neighboring, 195
osculating, 193, 203
circle packing theorem, 104
claw partition, 296
3-cycle reversal, 297
Cohn-Vossens theorem, 245
Cohn-Vossen, Stefan, 248, 339
Colin de Verdi`ere, Yves, 108
complementarity lemma, 149, 157
complete ag of a polytope, 178
concurrent lines, 32
cone
base, 244
continous deformation, 264
convex, 24
degenerate, 244
dual, 243, 266
geodesic, 350
simple, 24
conguration
double chain, 132
Fano, 112
Perles, 112
plane, 111
realization of, 111
Connellys conjecture, 291
Connellys second order rigidity theorem, 305
Connelly, Robert, 281, 290, 314, 352, 391
constant width set, 30, 237
convex cap, 244
convex hull, 8
convex polytope, 8
convexity criterion, 19, 284

435

Conway, John Horton, 158, 389, 391

Coxeter simplex, 156
Coxeter, H. S. M., 138, 184
Croftons formula, 231, 239
cross, 145
cross-polytope, 258
exible, 283
cross-ratio, 113, 205
Csaszar polyhedron, 408
cuboctahedron, 153, 394
curvature
absolute, 270
average total, 239
for spherical polyhedral surfaces, 276
gaussian, 248
mean, 168, 263, 270
net total, of graphs in R3 , 239
of a smooth curve, 193
total, 230, 232, 238
curve
smooth, 8
cusp in an evolute, 198
cut locus, 196, 349, 355
cut of a plane curve, 87
D-congruence, 145
Dahlberg, Bjorn E. J., 206
Dali, Salvador, 358
Dawson, Robert J. MacG., 385
de Guas formula, 247
Debrunner, Hans E., 158, 391
decomposition, 8
Dehn invariant, 147, 161
Dehns lemma, 293, 300
Dehns theorem, 140
general version, 162
Dehn, Max Wilhelm, 81, 147, 300
DehnSommerville equations, 72, 81
Delaunay triangulation, 126, 133
connection with Delaunay polygons, 201
in higher dimensions, 130, 132
inverse (furthest site), 132
optimal properties, 133
paraboloid construction, 127
Delone, Boris N., 138
deltahedron, 394
Denne, Elizabeth, 240
Descartes, Rene, 138, 248
Dey, Tamal K., 228
diameter, 26
Dirichlet, Lejeune, 138
discrete geodesic problem, 349, 356

disphenocingulum, 393
dissection, 159
dodecahedron, 167
dodecahedron construction, 181, 182
Dolbilin, Nikolai P., 138
domino tiling, 135
double normal
theorem, 87
to a smooth plane curve, 87
to smooth surfaces, 93
double quasi-normal to a polygon, 227
D
urers conjecture, 348, 358
D
urer, Albrecht, 358
Dvoretzky, Aryeh, 377
Dysons theorem, 55
Dyson, Freeman, 60
E8 lattice, 392
Eberhards theorem, 248
edge
extremal, 197
inection, 216, 219
edge summation lemma, 263, 271
Eggleston, Harold Gordon, 31
elementary move, 159
Emch, Arnold, 53
empty circle condition, 127, 128
empty sphere criterion, 130, 132
equipartition
of three polytopes, by volume, 36
of two polygons, by area, 36
with a cobweb, 37
with three lines, 32
with two lines, 32
Erdos ip problem, 227
Erdos, Paul, 25, 380, 397
Euclid, 180, 184, 256
Euler characteristic, 168
Eulers formula, 241, 248
Eulers inequality, 374
Euler, Leonard, 25, 408
Eulerian circuit, 239
Eulerian numbers, 80, 82
evolute
discrete, 198
dual, 198, 205
relative, 199
expansion of sets, 63
f -vector, 72
Fabricius-Bjerres formula, 219, 220
Fabricius-Bjerre, Fr., 228

436

fair division of a necklace

s colors, k pirates, 37, 38
two colors, two pirates, 35
Fary theorem, see also FaryMilnor theorem
on monotonicity of total curvature, 239
on planar graphs, 104
Fary, Istvan, 108, 240, 398
FaryMilnor theorem, 230
Fejes-Toth, 408
Fenchels theorem, 230
Fenchel, Werner, 240
Fermats last theorem, 25
at folding, 353
exible polyhedron, 265, 285
Bricards octahedron, 281
Connellys polyhedron, 281
Connellys suspensions, 290
homeomorphic to a torus, 282
spherical construction, 282
Walzs cross-polytope in R4 , 283, 285
exor, 281
ip, 125
bounds on the number of, 133
decreasing, 132
elementary move, 159
in convex polygon, 133
in the Erdos ip problem, 227
increasing, 129, 130
simultaneous, 133
stellar, 168
Floyd, Edwin E., 60
foldout, 348
Foppl polytope, 154
four cusps theorem, 199, 206
four edge theorem, 198, 206
four inection points theorem, 203
four support vertex theorem, 201, 206
four vertex theorem
converse, 204
discrete version, 194, 195, 202, 206
extremal circles, 194, 206
for Delaunay polygons, 201
for smooth convex curves, 193, 206
four edge theorem, 206
generalization, 203, 204, 206
inection points, 203
number of osculating circles, 194, 206
spherical, 203
support vertex theorem in R3 , 201
Frechet distance, 239, 270
full circle condition, 132
fundamental region, 140

Galitzer, Amy Jan, 387

Gausss ruler and the compass theorem, 109
GaussBonnet theorem, 242, 248, 254
for spherical polyhedral surfaces, 276
for unbounded convex polyhedra, 244
in higher dimensions, 243
Gehring, Frederick W., 240
geodesic, 96
simple closed, 97
geodesic distance, 8, 261
geodesics
isotopic, 100
simple closed, 244
Girards formula, 360
Gleason, Andrew M., 387
Glucks theorem, 287
Gluck, Herman, 206, 291
Goldberg, Charles, 38
Goldberg, Michael, 285, 391
golden ratio, 112, 179
GrahamNewman problem, 23
graph, 8
acyclic orientation of, 72
d-connected, 77
medial, 255
of a polytope, 72, see also polytope graph
regular, 72
Gr
unbaum, Branko, 31, 383
Gr
unbaumShephard theorem, 183
h-vector, 72
Haar measure, 60
Hadamard, Jacques, 138
Hadwiger invariant, 147
Hadwigers theorem, 28
HadwigerGlur theorem, 145
Hammers X-ray problem, 69
Har-Peled, Sariel, 406
Harer-Zagier formula, 333
Hauptvermutung, 176
height function, 130, 138
in a domino tiling, 136
Hellys theorem, 11, 19
d-dimensional, 12
colorful, 24
converse, 19
fractional, 15
generalized, 17
innite, 16
spherical, 17
topological, 18
weak converse, 18

437

Herons formula, 311

Hilberts third problem, 140
Hilbert, David, 140, 147, 339
Holmsen, Andreas, 378
honeycomb, 184
Hookers polyhedron, 284
Hopf, Heinz, 38
Horn, Roger A., 240
hypercube, 74
icosahedron, 167, 179, 181, 394
icosidodecahedron, 157, 383
increasing ip condition theorem, 129
innite descent method, 20
inection edge, 216
inscribed chord theorem, 34
integrality criterion, 308
isometry
global, 252, 257
intrinsic, of surfaces in R3 , 260
isoperimetric inequality
for spherical polygons in S2 , 269
in the plane, 61, 269
in Rd , 66
spherical (in S2 ), 69
isosceles trapezoid, 375
Jackson, Stanley B., 206
Jerrard, Richard, 53
Jessens orthogonal icosahedron, 157
construction, 183
innitesimal and continuous rigidity, 305
is vertex-transitive, 183
Jessens theorem, 157
Jessen, Brge, 147, 394
Johnson solid, 184
deltahedron, 394
disphenocingulum, 393
sphenocorona, 393
trapezohedron, 394
Johnson, Norman W., 184
Jones, Peter W., 407
k-line, for a set of points in R2 , 221
Kagan, Veniamin F., 147
Kahans formula, 374
Kahn, Je, 26
Kakutanis theorem, 56
in higher dimensions, 59
Kakutani, Shizuo, 60
Kalai, Gil, 19, 26, 31, 76, 82
Kapovich, Michael, 123
Kempe, Alfred, 123

kinematic density, 237

kinematics, 123
Kirchbergers theorem, 17
kissing cylinders, 190
kissing number, 185, 191
average, 191
Klyachko lemma, 81
Knasters problem, 58, 59
Knaster, Bronislaw, 60
Kneser, Adolf, 206
Kneser, Hellmuth, 206
knot
quadrisecants of knots, 238
tiling, 146
Koebe, Paul, 108
KoebeAndreevThurston theorem, 104, see
also circle packing theorem
Kotzig transformation, 239
Kotzig, Anton, 383
Kovalev, M. D., 408
Krasnoselskiis theorem, 18
Kuperberg, Greg, 176
Kuperberg, Wlodzimierz, 395
Kushnirenko, A. G., 395
Kuzminykh, A., 395
ladder problem, 52
Lagarias, Jerey C., 389
Lagrange, Joseph Louis, 408
Lang, Robert J., 358
law of cosines
for spherical polygons, 360
in R3 , 373
law of sines
for spherical polygons, 372
in R3 , 373
Lawrence construction, 115
Lebesgue, Henri, 123, 379
Leech, John, 191
Legendre, Adrien-Marie, 184, 215, 248, 256
LegendreCauchy lemma, 207209
spherical version, 207
Leibins theorem, 337
Lev, Vsevolod, 409
Levy, Paul, 38
Liebmann, Heinrich, 240, 339
limit argument, 46
Lindelofs theorem, 327
line development theorem, 215
linkage, 118
Chebyshevs, 123
mechanical, 119

438

pantograph, 120
Peaucellier inversor, 121
polyhedral, 123
realization, 118
spherical, 122
translator, 120
Watts, 123
linked tubes theorem, 232
Lipkin, Lippman, 123
Lipov, L., 378
local convexity criterion, 128
Lovasz, Laszlo, 228, 378
LyusternikShnirelman theorem, 97
MacMahon, Percy Alexander, 82
manifold
Riemannian, 248
unicoherent, 59
map
characteristic, 289
isometry, 252
Monge, 170
proper, 368
unfolding, 349
mapping lemma, 368
Marguliss napkin problem, 358, see also
napkin folding problem
mean curvature invariance theorem, 265
medial axis of a polygon, 206
Miller, Ezra, 358, 393
Millson, John J., 123
Milnors lemma, 231
Milnor, John W., 240
Minkowski formula, 270
Minkowski inequality, 361
Minkowski sum, 63
Minkowski theorem, 323, 328
Minkowski, Hermann, 328, 404
Mobius strip, 42
Mobius theorem, 203, 206
Mobius, August Ferdinand, 206
Molnar, Jozsef, 377
Monge equivalence, 170
of bipyramids in Rd , 172
of convex polytopes in Rd , 170
of general polygons in the plane, 174
Monge map, 170
Monge, Gaspard, 176
monotonicity
of cone curvature, 246
of the area, 69
of the average total curvature, 239

of the mean curvature, 269

of the total curvature of plane curves, 239
of the total curvature of space curves, 232
Morelli, Robert, 169
Morse function, 15, 7274, 77
Morse theory, 82
mosaic, 137, see also zonotopal tiling
Moscow Math. Olympiad, 382, 395, 407
Moser, J
urgen, 176
Mount, David M., 358, 406
mountain climbing lemma, 43
extension to general functions, 52
generalization to many climbers, 52
Mukhopadhyaya, Syamadas, 206
Musin, Oleg R., 191, 206
napkin folding problem, 354, 357
Narayana numbers, 80
Neville, Eric Harold, 379
Newman, M. H. A., 168
NewtonGregory problem, 191
normal
to a smooth curve, 84
to facets, 322
notched cube, 146
NP-hardness
of the discrete geodesic problem, 356
of the edge reconstruction problem, 334
of the polygon triangulability, 134
of the ray curvature conditions, 321
numbers
Catalan, 75, 82
Eulerian, 80, 82
kissing, 185
Narayana, 80
ORourke, Joseph, 385
octahedron, 167
Bricards exible, 281
equihedral, 51
Olovianishniko, S. P., 262
origami, 358
Ortel, Marvin, 240
orthic triangle, 385
orthoscheme, 156
osculating circle, 193
osculating conic, 204
Pach, Janos, 378
Pachner move, 169
Pak, Igor, 389
Pals theorem, 30
Pappus theorem, 116

439

paraboloid, 127
parallelepiped, 70, 144
Peaucellier, Charles-Nicolas, 123
penta-partition, 36
permutohedron, 74
f -vector, 74
tiling, 154
-congruence, 146
-congruence, 153
PL-manifold, 176
place (a ring valuation), 308, 365
plane
inection, of a polytope in R3 , 218
supporting vertex of a polytope, 242
tritangent, to a curve in R3 , 226
plane tessellation, 184
Platonic solid, 157, 184
Pogorelov uniqueness theorem, 261
Pogorelovs lemma, 266, 267, 271
Pogorelovs theorem
on curvatures on convex caps, 319
on octant shape polyhedra with given face
normals and areas, 325
on quasi-geodesics, 100
Pogorelov, A. V., 262, 271, 321, 328, 339,
402, 406
Pohlkes theorem, 371
Poincare conjecture, 97, see also
LyusternikShnirelman theorem
Poincare theorem
on closed geodesics, 97
on the hairy ball, 247
Poincare, Henri, 81, 97, 184
point
at, 193, 276
inection, 203
point and line arrangement, 111
polar sine function, 373
polygon, 8, see also space polygon
coherent, 194
Delaunay, 201
equiangular, 198
non-convex, 195
parallel, 199
regular, in R3 , 370
rigid deformation, 223
similar, 51
simple, 18, 54, 244
spherical, 360
star-shaped, 17
surface, 244
surrounds another, 199

tensegrity, 305
polygonal region, 8
polyhedron, 8, see also special polyhedron
convex cap, 244
exible, 280, see also exible polyhedron
orthant shape, 324
shaky, 306
tight, 279, 284, 285
unbounded convex, 244
polytope, 8, see also special polytope
adjacent, 375
circumscribed around a sphere, 183
decomposable, 326
edge-transitive, 178
face-transitive, 178, 182
fair (as in a fair die), 182
fortunate (as in Bricards condition), 142
indecomposable, 326
inscribed into a sphere, 183
irrational (with irrational coordinates), 115
midcribed around a sphere, 70, 183, 306
monostatic, 86
neighborly, 79
non-convex, 167
parallel, 253
passes through itself, 371
planted, 287
rectiable, 144, 148
regular, 177, see also regular polytope
self-similar, 149
shadow, 371
similar, 148
simple, 72
simplicial, 72
vertex-transitive, 178
polytope graph, 72
connectivity, 77
diameter, 79
edge coloring, 79
have embedded cliques, 79
vertex degrees, 79
poset polytope, 174
` G., 291
Poznjak, E.
Prekopa-Leindler inequality, 71
prism, 144, 394
Proizvolov, V., 377, 379
pseudoline arrangement, 138
Putnam Mathematical Competition, 387, 403
quadrisecants of knots, 238
quasi-billiard trajectory, 90
quasi-geodesic, 99

440

quasi-normal, 85, 227

Rabinovich, Yuri, 390
Rademacher, Hans, 108
Radons theorem, 12, 21
Radon, Johann, 398
ray, 244
coordinates, 315
realization
of a surface, 279
regular polytope, 177, see also Archimedean
solid and Platonic solid
in higher dimensions, 183, 184
antiprism, 394
cross-polytope, 258
prism, 394
regular triangulation, 130, 134, 135
connectivity theorem, 131
reptile, 146
Reuleaux triangle, 30, 50
rhombic dodecahedron, 183, 394
rhombic triacontahedron, 183
rigid motion, 139
rigidity
continous, 249, 293
framework, 305
innitesimal, 302
second order, 305
static, 292
ring width problem, 52
Robbinss problem, 289, 313, 314
Robbins, David P., 401
Rote, G
unter, 215
Rote, G
unter, 393
Russian Mathematical Olympiad, 378
S
uss, Wilhelm, 396
Sabitov polynomial relation, 288
Sabitov, Idjad Kh., 256, 314
Sarrons formula, 372
Schlai formula, 265, 271, 272
generalization, to all deformations, 271
spherical version, 271
Schlai simplex, 155
Schlai symbol, 184
Schlenker, Jean-Marc, 271
Schonhardts polyhedron, 388
Schramms theorem, 104, 306
Schramm, Oded, 108
Schreier, Otto, 377
Sch
utte, Kurt, 191
Schwartz, Richard E., 90, 385

scissor congruence, 139, 158

Bricards condition, 142
complementarity lemma, 149, 157
continuum equivalence classes, 153
for painted polytopes, 145
for unbounded polyhedra, 145
Jessens theorem, 157
mirror symmetry, 144
Sydlers main theorem, 162
Sydlers rectiability criteria, 149
Sydlers theorem on dihedral angles, 157
Sydlers theorem on regular tetrahedra, 148
tiling lemma, 149
under exing, 291
second hull theorem, 235
Sedykh, V. D., 206
Segre, Beniamino, 206, 240
Senderov, V. A., 394, 400
Senkin, E. P., 277
Sharir, Micha, 406
Sharygin, Igor F., 394, 403
shelling, 82
Shen, Alexander Kh., 381
Shephard, Georey C., 248, 358, 404
Shnirelman, Lev G., 53, 228
shortest path, 96, 356
Shtogrin, M. I., 347, 405
simple polygon, 8
simplex
Coxeter, 156
equifacetal, 184
monostatic, 86
path-simplex, 156
rainbow, 21
Schlai, 155
six vertex theorem, 204, 206
Sommerville, D. M. Y., 81, 392
source unfolding, 349
space polygon
triangulable, 134
weakly convex, 200
special polytopes and polyhedra
associahedron, 75, 82
Birkho polytope, 81, 175
chain polytope, 175
Csaszar polyhedron, 408
equiareal octahedron, 51
Fibonacci polytope, 175
golden tetrahedron, 155
Hookers polyhedron, 284
Jessens icosahedron, 157, 183, 305
permutohedron, 74, 154

441

plane partitions polytope, 175

poset polytope, 174
Schonhardts polyhedron, 388
Szilassi polyhedron, 408
transportation polytope, 81
sphenocorona, 393
spherical polyhedral surface, 276
splitting necklaces theorem, 35
square peg problem, 39, 49, 224
squaring the circle, 158
Stachel, Hellmuth, 285, 401
star unfolding, 349, see also Alexandrov
unfolding
Stashe polytope, see also associahedron
Steiner symmetrization, 68
Steiner, Jakob, 70
Steinitz theorem, 102, 104, 108
Steinitz, Ernst, 108, 215
Stokers conjecture, 256, 300
Stokers theorem, 255
Stoker, James J., 256, 300
subdivision, 8
Sullivan, Dennis P., 314
Sun Tzu, 95
supporting hyperplane, 8
surface
bendable, 336
Chadwick, 346
curvilinear, 345
distance, 260
star-shaped, 58
submetric (to another), 343
suspension, 290
Sydlers theorem, 162
on dihedral angles in a tetrahedron, 157
on regular tetrahedra, 148
the rectiability criteria, 149
the scissor congruence criterion, 162
Sydler, Jean-Pierre, 147, 157
Sylvesters problem, 98, 370
SylvesterGallai theorem, 20, 25
dual version, 23
Szego, Gabor, 396
Szilassi polyhedron, 408
Tabachnikov, Serge L., 206
table theorem, 52
Tasmuratov, S. S., 347
tennis ball theorem, 203
tensegrity polygon, 305
tetrahedron
equihedral, 51, 58, 100, 247, 326

golden, 155
Hills, 151, 155
Kahans volume formula, 374
monostatic, 86
random, 98
rectiable, 155
standard, 150, 151
Thom, Rene, 206
Thurston, William P., 108, 123, 389
Tietze, Heinrich, 138, 409
tiling
aperiodic, 154, 158
lemma, 149
periodic, 149
space with knotted tiles, 146
substitution, 146
the space with crosses, 145
the space with notched cubes, 146
with dominoes, 135, see also domino tiling
with ribbon trominoes, 136
zonotopal, 137, see also zonotopal tiling
Tolpygo, Alexei K., 395
Toponogovs theorem, 358
transportation polytope, 81
trapezohedron, 394
tree
binary, 80
minimum spanning, 133
orthogonal, 155
piecewise linear spanning, 348
triangulation, 8, 159
acute, 134
associated with height function, 130
combinatorial, 134
Delaunay, 126, see also Delaunay
triangulation
equiareal, 313
ip diameter, 133
full, 125, 130, 163
non-obtuse, 134
number of full triangulations, 132
of a convex polygon, 75, 125
of a non-convex polygon, 135
plane, as a graph, 286
regular, 130, see also regular triangulation
star, 163, 165
tripod
to curves in R2 , 91
to surfaces in R3 , 57
truncated cube, 153
truncated cuboctahedron, 154
truncated dodecahedron, 383

442

truncated octahedron, 154

truncated tetrahedron, 154, 387
Tuttes formula, 133
Tuttes spring theorem, 104
Tverbergs theorem
generalization of Radons theorem, 22
on how to cut a polytope, 167
two integrals theorem, 37
two vertex theorem, 206
unfolding
Alexandrov, 349, 355
non-overlapping, 349
source, in the plane, 349
universality theorem
of congurations, 113
of linkages, 119
Urysohns inequality, 69
Vainshtein, F. V., 378
valuation, 160, 168
Chevalley extension lemma, 314
symmetric valuation extension lemma, 160
van der Waerden, B. L., 191
Vasiliev, Nikolai B., 377, 378, 394, 400
vertex
convex, 299
index of, 72
of a smooth curve, 193
saddle, 299
sextactic, 204
spherical, 203
support, 201
Volkovs stability theorem, 334
Volkov, Yu. A., 271, 334
Voronoi diagram, 126
geodesic, 355
inverse (furthest site), 132, 197
of a convex polygon, 196
Voronoy, Georgy, 138
Wagner, Uli, 228
Walz, Anke, 285
Watt, James, 123
West, Douglas, 38
Weyl, Hermann, 339
Wienholtzs projection theorem, 237
Wlodarczyk, Jaroslaw, 169
Zalgallers theorem, 353, 355
Zalgaller, V. A., 184, 277, 347, 405
Zamrescu, Tudor I., 387
zonotopal tiling, 137, 138

zonotope, 70, 145, 154, 183, 327

Zylevs theorem, 153, 157