J. fixed point theory appl.

2 (2007), 215–220
c 2007 Birkhäuser Verlag Basel/Switzerland
1661-7738/020215-6, published online 20.12.2007 Journal of Fixed Point Theory
DOI 10.1007/s11784-007-0042-5 and Applications

Euler’s fixed point theorem:

The axis of a rotation
Bob Palais and Richard Palais

Dedicated to the memory of Leonhard Euler, “The Master of us all”, on the occasion of the
300th anniversary of his birth

Abstract. We give an elementary proof of what is perhaps the earliest fixed

point theorem; namely Leonhard Euler’s theorem of 1775 on the existence of
an axis v for any three-dimensional rotation R. The proof is constructive and
shows that no multiplications are required to compute v.

Mathematics Subject Classification (2000). Primary 55M20, 01A50.

Keywords. Rotation, axis.

In what is perhaps the historically earliest fixed point theorem, Leonhard

Euler [1] stated in 1775 that in three dimensions, every rotation has an axis.
Euler’s original formulation of the result is that if a sphere is rigidly rotated about
its center then there is a diameter that remains fixed. A modern reformulation is:

Euler’s Theorem. If R is a 3 × 3 matrix satisfying RT R = RRT = I and

det R = +1, then there is a non-zero vector v satisfying Rv = v.

Counterexamples are easy to find in two or other even dimensions. Even in

three dimensions, it is not immediately obvious that the composition of rotations
about distinct axes is equivalent to a rotation about a single axis. This important
fact has a myriad of applications in pure and applied mathematics, and as a result
there are many known proofs. In this 300th anniversary year of Euler’s birth, we
give a constructive proof of the fixed vector v that appears to be new.

Proof. Our construction is this: if A := 12 (R − RT ) is the skew-symmetric part

of the matrix R, then the vector v := (a23 , a31 , a12 ) is left fixed by R. This
follows from RART = A, which is immediate from the definition of A. To see
that RAR−1 = RART = A is equivalent to Rv = v, we invoke the linear
isomorphism v → Jv between R3 and the skew-symmetric 3 × 3 matrices, given
216 B. Palais and R. Palais JFPTA

explicitly by ⎛ ⎞
0 −v3 v2
Jv := ⎝ v3 0 −v1 ⎠ .
−v2 v1 0
Jv is just the matrix implementing cross-product with v; i.e., Jv (w) := v × w for
all w ∈ R3 . An essential fact about the cross-product operation is its invariance
under any such rotation matrix R, that is, R(v × w) = Rv × Rw. In terms of
the map J, this says JRv (Rw) = RJv (w) or simply JRv R = RJv . Rearranging
this, we arrive at JRv = RJv R−1 . (In the language of representation theory, this
says that J is an equivalence or “intertwining operator” between the standard
representation of SO(3) on R3 and the adjoint representation on its Lie algebra,
so(3), the skew-adjoint 3 × 3 matrices. Or, in more elementary terms, the matrix
for cross-product with a rotated vector is obtained by conjugating (by the rotation)
the matrix for cross-product with the un-rotated vector.) It is now immediate that
a vector v is fixed by a rotation R if and only if the skew-adjoint matrix Jv is
fixed under conjugation by R.
But we are not through yet! What if A = 0? Such orthogonal matrices—the
symmetric ones—have measure zero, and correspond to angles of rotation 0 and π.
But still they are important, and we have a similar result for them. In this (non-
generic) case, R = RT and so R2 = I. Therefore, R(I + R) = R + R2 = I + R,
so the columns of I + R are fixed. Since R is proper (det R = +1), R = −I, so
I + R has a non-zero column that is fixed by R. 
 cos θ sin θ 0 
Examples. If R = Rθ = − sin θ cos θ 0 is rotation about the z-axis by an angle θ,
 0 sin θ 0  0 0 1
0

then A = − sin θ 0 0 so v = 0 . If θ is not 0 or π then we are in the
0 0 0 sin θ
generic case, and we find that the z-axis is a fixed axis of the  x1 rotation.
 If θ= 0,

 0
then R = I, so every vector is fixed, while if θ = π and x = x2 then v = 0
x
3 x3
(which is non-zero ifwe choose x3 = 0) and again we find that the z-axisis fixed.
2/3 2/3 −1/3 0 1/2 −1/2 1/2
If R = RG = −1/3 2/3 2/3 , then A = −1/2 0 1/2 , so 1/2 is
2/3 −1/3 2/3 1/2 −1/2 0 1/2
fixed.  −1/32/3 2/3
 2/3 2/3 2/3
If R = RN = 2/3 −1/3 2/3 , then A = 0, and R + I = 2/3 2/3 2/3 , so
 2/3 2/3 2/3 −1/3 2/3 2/3 2/3

2/3 is fixed.
2/3

Remarks. 1) Our proof of Euler’s theorem uses one non-trivial result, the “intrin-
sic” nature of the cross-product operation in a three-dimensional, oriented, real
inner-product space. But this is obvious from the usual characterization of u × v,
namely that its length is the area of the parallelogram spanned by u and v, that
it is orthogonal to u and v, and that (u, v, u × v) is an oriented triple.
2) In the generic case our key fixed-point fact, RART = A, requires only
the orthogonality of R. However, if R is improper then R(v × w) = −Rv × Rw,
Vol. 2 (2007) Euler’s fixed point theorem 217

so our argument shows that the vector v corresponding to A is not fixed by R,

but rather is “reversed”, i.e., mapped to −v (but note that the axis through v is
still fixed). In the non-generic case we are using another elementary fixed-point
result; if a linear operator T on a vector space V satisfies T2 = I, and we define
P := 12 (I + T), then P2 = P, so P is the projection on the image of P along
its kernel. But if w = Pv = 12 (v + Tv), then Tw = 12 (Tv + v) = w, so P
projects V onto the fixed-point set V+ (the +1 eigenspace) of T. And if Pv = 0,
then Tv = −v, so P projects V along the space V− of reversed vectors (the −1
eigenspace of T). Symmetrically, I − P = 12 (I − T) is the projection on V− along
V+ . In particular V = V+ ⊕ V− , and of course when T is symmetric this is an
orthogonal decomposition. So regardless of whether R is proper or improper, in
the non-generic orthogonal case in which R = RT , we have R2 = I, so the two
projections 12 (I + R) and 12 (I − R) give an orthogonal decomposition of R3 into
fixed and reversed subspaces. We may classify four cases up to O(3) conjugacy by
their complementary dimensions between 0 and 3. We also note that proper and
improper examples are in one-to-one correspondence under composition with −I.
3) Euler’s original demonstration never invokes the distinction between proper
and improper transformations, apparently leaving it implicit in the concept of a
rigid motion. Moreover, Euler’s construction of a fixed axis for a non-generic ro-
tation fails (in different ways) for both the proper and improper cases, so it does
require modification to be valid by modern standards. (Indeed, there have been
many other purported proofs of Euler’s theorem over the years that also fail for
similar reasons, and this is why we have paid such detailed attention to the non-
generic case above). Remarkably, Euler’s demonstration does contain constructions
foreshadowing both cases of the proof we have given here, but in his more geomet-
ric language! We address these considerations, provide a (previously unavailable)
English translation of the original proof, and survey other proofs in a forthcoming
paper [8].
4) We note that, using our construction, no multiplications are required to
find an axis in either case. Similarly, the cosine of the angle of rotation about this
axis can easily be found by taking the trace of R. The trace is independent of
basis, and in a proper orthonormal basis containing the axis, the matrix for R is
given by the matrix Rθ in the example above, whose trace is clearly 1+ 2 cos θ. For
comparison, other approaches have been suggested to compute the axis involving
from 1 to 3 cross-products or up to 18 multiplications, followed by more cross-
products, normalizations and projections for up to about 40 multiplications to
find the cosine of the angle (see for example [3, Chapter 12, Section 9: Finding the
axis and angle]).

Final remarks suggested by the referee’s comments

The fact that the axis of a generic rotation R is a multiple of the vector correspond-
ing to the anti-symmetric part of R has been observed and used in other contexts
in the past. There is also another construction that does not require any multi-
218 B. Palais and R. Palais JFPTA

plications to obtain the axis. As we will outline below, all of these constructions
assume the existence of an axis, and seem to require appeals to less elementary
concepts than our approach above to turn them into proofs of Euler’s theorem.
The axis-angle representation of a rotation R about a unit axis v by an angle
θ is based upon the decomposition of a vector x into a component along v, and a
component orthogonal to v: x = (x · v)v + (x − (x · v)v). Since v × ((x − (x · v)v))
is the image of x − (x · v)v under a quarter-turn counterclockwise rotation of v⊥ ,
and Rv = v, we can check that
Rx = (x · v)v + (cos θ)(x − (x · v)v) + (sin θ)(v × (x − (x · v)v)).
Using v × v = 0, this simplifies to
Rx = x + (sin θ)(v × x) + (1 − cos θ)((x · v)v − x).
If Jv is the matrix defined earlier by Jv (x) = v × x, the fact that v is a unit vector
implies J2v = vvT − I, or J2v x = (x · v)v − x, which leads to the Euler–Rodrigues
rotation formula:
R = I + (sin θ)Jv + (1 − cos θ)J2v .
In [1], Alperin derives this result from Lagrange’s well-known triple prod-
uct formula p × (q × r) = (p · r)q − (p · q)r, while Fillmore [5] and Kahan [6]
derive it for R(θ) = exp(θJv ) using the series representation for exp and the
recurrence Jvn+2k = (−1)k Jvn+2k , n > 0. Fillmore takes the series as the defi-
nition of exp(θJv ) while Kahan solves the initial value problem dR/dθ = Jv R,
R(0) = I. Kahan observes that this formula implies 12 (R − RT ) = (sin θ)Jv which
can be used (as we did) to recover the axis in the generic case, and points out this
can fail numerically when sin θ is small. Fillmore uses the formula to show that
if we define S := R + RT + (1 − trace(R))I (clearly a symmetric matrix) then
S = (3 − trace(R))vvT . This gives another method to obtain the axis without
multiplication, as the right-hand side exhibits this matrix as a rank one symmet-
ric matrix, any of whose non-zero columns are multiples of v. Kalman [7] makes
the related observation that for any x, R(Sx) = Sx. He shows that this is equiv-
alent to R3 − trace(R)R2 + trace(R)R − I = 0, and confirms this by applying
the Cayley–Hamilton theorem. Fillmore and Kalman both use the fact that the
eigenvalues of R are {1, cos θ + i sin θ, cos θ − i sin θ} to show trace(R) = 1 + 2 cos θ.
Note that all of these approaches start with the assumption that R is a rotation
about an axis v through an angle θ, and so cannot be used as a proof of Euler’s
theorem without considerable modification. (Moreover they all invoke results like
the Cayley–Hamilton theorem, the power series for exp, and facts about solutions
of the characteristic equation of a matrix that seem less elementary than the ap-
proach we have given above.) However, granting the truth of Euler’s theorem, they
do give alternative derivations for our construction of the axis of a rotation in the
generic case, and recognizing this, Kahan presents the classical proof based on
characteristic polynomials that 1 is an eigenvalue of R, and suggests finding the
corresponding eigenvector by elimination or computing an appropriate column of
Adj(R − I). Fillmore also examines the improper case, and the representation of
Vol. 2 (2007) Euler’s fixed point theorem 219

orthogonal transformations as the product of two (in the proper case) or three (in
the improper case) reflections.
There have also been several algorithms proposed to compute the unit quater-
nions ±Q corresponding to R, [2, 10–12]. Most are based upon a modified ver-
sion of the Euler–Rodrigues rotation formula, which has a nice expression in
terms of the components of the unit quaternion associated with R, Q = [q, q] =
[cos θ2 , sin θ2 v]. Using sin θ = 2 cos θ2 sin θ2 , we get sin θJv = 2qJq , and since 1 −
cos θ = 1 − cos(2 θ2 ) = 1 − (cos2 θ2 − sin2 θ2 ) = 2 sin2 θ2 we get (1 − cos θ)J2v = 2J2q .
This leads to the well-known quaternion-to-matrix formula
R = I + 2qJq + 2J2q .
Because the vector part of Q is a unit axis of R, scaled by sin θ2 (rather than sin θ as
in our construction), matrix-to-quaternion algorithms compute the anti-symmetric
part of R, then perform square roots and division to accomplish the desired scaling.
In another forthcoming paper [9], we present new and efficient algorithms for
implementing and interpolating rotations based on a novel realization of a rotation
and unit quaternions without reference to an axis.

Acknowledgments
We are indebted to the referee for pointing out to us references [1] and [5].

References
[1] R. Alperin, The matrix of a rotation. College Math. J. 20 (1989), 230.
[2] I. Y. Bar-Itzhack, New method for extracting the quaternion from a rotation matrix.
AIAA J. Guidance Control Dynamics 23 (2000), 1085–1087.
[3] B. Casselman, A Manual of Geometry and PostScript. Cambridge Univ. Press, 2005;
http://www.math.ubc.ca/∼cass/graphics/manual/pdf/ch12.pdf.
[4] L. Euler, Formulae generales pro translatione quacunque corporum rigidorum. Novi
Commentarii Academiae Scientiarum Petropolitanae 20 (1776), 189–207. Reprinted
in Opera Omnia, Series 2, Volume 9, 84–98, and in Theoria motus corporum solido-
rum seu rigidorum, ed. nova, 1790, 449–460.
[5] J. P. Fillmore, A note on rotation matrices. IEEE J. Computer Graphics Appl. 4
(1984), 30–33.
[6] W. Kahan, Computing cross-products and rotations. UC Berkeley Math H110 Lecture
Notes; http://www.cs.berkeley.edu/∼wkahan/MathH110/Cross.pdf.
[7] D. Kalman, The axis of a rotation: analysis, algebra, geometry. Math. Mag. 62 (1989),
248–252.
[8] B. Palais, R. Palais and S. Rodi, Euler’s theorem on the axis of a rotation: proofs
old and new. To appear.
[9] B. Palais and R. Palais, Reflections on rotations. To appear.
[10] S. W. Shepperd, Quaternion from rotation matrix. J. Guidance Control 1 (1978),
223–224.
220 B. Palais and R. Palais JFPTA

[11] K. Shoemake, Quaternion calculus and fast animation. In: Computer Animation:
3-D Motion Specification and Control, SIGGRAPH 1987 Tutorial (1987), 101–121.
[12] R. A. Spurrier, Comment on singularity-free extraction of a quaternion from a
direction-cosine matrix. J. Spacecraft Rockets 15 (1978), 255.

Bob Palais
Department of Mathematics
The University of Utah
Salt Lake City, UT 84112, USA
e-mail: [email protected]
Richard Palais
Department of Mathematics
103 MSTB University of California at Irvine
Irvine, CA 92697, USA
e-mail: [email protected]

To access this journal online:

www.birkhauser.ch/jfpta