σ-Automata and Chebyshev-Polynomials

Klaus Sutner
Carnegie Mellon University
Pittsburgh, PA 15213
[email protected]
http://www.cs.cmu.edu/~sutner

Abstract
A σ-automaton is an additive, binary cellular automaton on a graph. For product graphs
such as a grids and cylinders, reversibility and periodicity properties of the corresponding σ-
automaton can be expressed in terms of a binary version of Chebyshev polynomials. We will
give a detailed analysis of the divisibility properties of these polynomials and apply our results
to the study of σ-automata.

1
1 Introduction

A σ-automaton is a simple, non-uniform, binary cellular automaton on a directed graph. These

automata were first studied by Lindenmayer in [8], and later in [1, 10, 13, 14, 2]. Briefly, a σ-
automaton consists of a directed graph G = hV, Ei together with a global rule given by
X
σ(X)(v) := X(u) mod 2.
u∈N (v)

Here X : V → {0, 1} is a pattern or configuration of the automaton and N (v) denotes the open
neighborhood { u ∈ V  (u, v) ∈ E } of vertex v. If the underlying graph is not obvious from context
we will write σG or σ(G) for emphasis. Note that rule σ is well-defined for any locally finite graph.
For example, the σ-automaton on the undirected biinfinite path P∞ is none other than the standard
linear CA with rule number 90, see [16]. As usual, we interpret an undirected edge as a pair of
opposite directed edges. A closely related class of automata is obtained by modifying rule σ to
rule σ + : the summation is now over the closed neighborhood N + (v) = N (v) ∪ {v}. Alternatively,
one can think of attaching self-loops to all nodes in the graph. On P∞ , rule σ + corresponds to the
linear CA with rule number 150. As observed by Fredkin, both rules have the property that any
finite configuration (i.e., a configuration with finite support) will reproduce itself after a sufficient
number of iterations.
There is a peculiar difference between σ-automata and σ + -automata with respect to their reversibil-
ity properties on simple product graphs such as n by m grids, cylinders and tori. For σ-automata,
there is a simple characterization of reversibility in terms of the parameters n and m. For example,
the n by m grid under rule σ is reversible iff n + 1 and m + 1 are coprime. No such character-
izations are known for σ + -automata and it appears difficult to determine reversibility by means
other than brute force calculation. The purpose of this paper is to trace the source of these diffi-
culties to determining the position of irreducible polynomials over the 2-element field in a certain
canonical enumeration (πn ) of polynomials associated with σ/σ + -automata on product graphs. For
σ-automata, it suffices to deal with π-polynomials directly, but for σ + -automata one has to con-
sider the images of the irreducible factors of these polynomials under the involution x 7→ x + 1. It
appears that this involution alters the position of an irreducible polynomial in a rather complicated
fashion, and is responsible for the erratic behavior of σ + -automata.
The dynamics of a σ-automaton on a graph G can be summarized conveniently by the state tran-
sition diagram T (G), also referred to as the phase space of the automaton, a directed graph whose
vertex set is CG , the collection of all patterns over G, and whose edges are of the form (X, σ(X)).
For our discussion, it is helpful to interpret the pattern space CG as a vector space over the two-
element field F2 . From the definitions, it follows that both rules σ and σ + = σ + id are linear
endomorphisms of these pattern spaces. Over undirected graphs the rules are also self-adjoint, a
property that was used in [14, 13] to analyze the phase space of these automata. Due to the lin-
earity of rule σ, the structure is fairly simple: every vertex has out-degree 1 and in-degree either 0
or corank of σG , since the predecessors of a vertex form an affine subspace of the pattern space CG .
The diagram thus consists of cycles and trees which are anchored on these cycles. In particular, for
finite graphs G, the σ-automaton over G is reversible iff the diagram is a disjoint union of cycles.
The same comments apply to σ + -automata. A detailed analysis of the state transition diagrams of
rules σ and σ + on undirected cycles, as well as higher dimensional analogues, can be found in [10].

We write d(G) for the corank of σG : C → C , and likewise d+ (G) for rule σ + . If the graph G is

2
given explicitly, say, as an adjacency matrix, then it is trivial to determine the corank of the linear
maps σ and σ + over C in time polynomial in the size of G. However, one is usually interested
in parameterized families of graphs, such as grids, cylinders, tori and so forth. For these graphs,
there is a natural succinct representation in terms of their defining parameters. For example, an
n by m grid can be specified in log n + log m bits. Therefore, one would like to determine the
crucial properties of a σ-automaton or σ + -automaton in time polynomial in the size of the succinct
representation. For example, the σ-automaton on an m by n grid Pm,n has kernel dimension

d(Pm,n ) = gcd(m + 1, n + 1) − 1,

and therefore the automaton is reversible iff m + 1 and n + 1 are coprime; a property that is
easily tested in time polynomial in the size of n and m. Similar characterizations exist for cylinders
Cm ×Pn and tori Cm ×Cn , see [13]. As we already mentioned, no comparable simple characterization
are known for the corresponding σ + -automata.

In this paper we will focus on determining the reversibility of σ + -automata on product graphs of
the form G = H × P where P is a path or a cycle. Let us write Pn and Cn for the undirected path
and cycle of length n, respectively, and Pm,n = Pm × Pn for m by n grids. For product graphs
G = H × Pn it is demonstrated in [13] that

d(G) = cork πn+1 (σ(H)),

where πi is a binary version of the Chebyshev polynomials of the second kind. More precisely,
π0 = 0 and πi is a polynomial defined by

πi (x) := Ui−1 (x/2) mod 2.

Alternatively, the generating function G of the sequence (πi ) is given by

z
G(z) =
z2 + xz + 1
For our purposes, the most convenient representation of these polynomials is in terms of a homo-
geneous second order recurrence over F2 [x]:

πn+2 = x · πn+1 + πn

with initial conditions π0 = 0, π1 = 1. Note that second order recurrences similar to the last one
can also be used to construct reversible cellular automata, a trick going back to Fredkin, see for
example [15]. The reversibility of the recurrence in the sense that πn = x πn+1 + πn+2 will be
important later.
In their recent paper [2], Barua and Ramakrishnan show that m by n grids under rule σ + are
reversible iff the two polynomials πm+1 (x) and πn+1 (1 + x) are coprime. We will extend their
results to show that
d+ (Pm,n ) = deg gcd πm+1 (x), πn+1 (1 + x) .

Moreover, it can be seen that πm+1 is the minimal polynomial of the linear map σ(Pm ). Thus, the
reversibility of σ + -automata on grids depends on divisibility properties of polynomials, rather than
integers as for σ-automata. In fact, reversibility of σ-automata can also be expressed in terms of π-
polynomials: gcd(m+1, n+1)−1 is none other than the degree of gcd(πm+1 , πn+1 ) = πgcd(m+1,n+1) .
Extensions of these results to higher-dimensional automata can be found in the forthcoming [3].

3
In order to shed some light on reversibility conditions, we will study the divisibility properties
Q of π-
polynomials in some detail. As it turns out, for odd n, πn can be written as a product πn = d|n ρd
where ρd is the square of a product of certain irreducible polynomials, referred to as the critical
factors of πn . A similar representation can be found for even n. In our setting, it is crucial that
every irreducible polynomial τ occurs as a factor of some π-polynomial. Hence, we can define the
depth of τ to be the least n such that τ divides πn . We will show that the depth d of τ divides
2k ± 1, where k is the degree of τ . In fact, the degree of τ turns out to be the suborder of 2 in the
multiplicative group Z∗d . Hence, the depth of τ is bounded by 2k + 1 where k is the degree of τ . As
we will see, this bound is tight.
Note that as a consequence of the characterization of the degree in terms of the suborder of 2, all
irreducible polynomials of the same depth must have the same degree; but not conversely. This
turns out to be the major obstacle in generalizing the description of the kernel dimension in grids
under rule σ to rule σ + . The involution x 7→ 1 + x changes the depth of the irreducible factors
of πn+1 in a rather complicated fashion, see the table 2 in the appendix for the depth values of
irreducible polynomials of degree 8. Consequently, there appears to be no easy way to compute the
GCD of two polynomials πm+1 (x) and πn+1 (1 + x).
The reader may also wish to compare these polynomials to another variation of binary Chebyshev
polynomials discussed in [4] on page 118. The inductive definition there takes the form ρn = ρn−1 +
xρn−2 . Many of the divisibility properties of the π-polynomials are shared by the ρ-polynomials.

Binary Chebyshev polynomials are also useful in the study of automata on infinite grids of the form
H∞ := H × P∞ . We can think of the σ-automaton H∞ as an additive one-dimensional cellular
automaton of width 3 over the alphabet {0, 1}V (H) , where V (H) denotes the vertex set of H. One
can easily show that d(H∞ ) = 2 · d(H). As it turns out, all patterns in the kernel of σ(H∞ ) are
periodic. We denote the maximal period of any kernel pattern by per(H) and likewise by per+ (H)
for σ + -automata. Computing the period of a σ-automaton with H = Pm is easy: it is always equal
to 2m + 2. For the analogous σ + -automata, on the other hand, periods and weak periods are much
harder to describe and we can only give a somewhat indirect description in terms of the depth
function mentioned above.

More precisely, define the weak period wper(H) of H to be the least p such that πp (σ(H)) = 0;
wper+ (H) is defined analogously. It is easy to see that per(H) = wper(H) or per(H) = 2 wper(H).
In particular for σ-automata with H = Pm we have per(Pm ) = 2(m + 1) = 2 wper(Pm ). Barua and
Ramakrishnan verify a conjecture from [13] which states that

wper+ (P2k −1 ) = 3 · 2k−1

for all k. We will show that per+ (P2k −1 ) = 3 · 2k . In fact, the last assertion is a simple corollary to
theorem 3.1 below, which shows how to express the periods of paths Pm under rule σ + in terms of
the depth function.

This paper is organized as follows. In section 2 we introduce π-polynomials in F2 [x] and determine
their basic divisibility properties. In the next section, we consider related shift register sequences
over Fq . In particular we will show that every irreducible polynomial occurs as a factor of one of the
π-polynomials. We also obtain simple representations of these shift register sequences. The next
section deals with linear operators of the form πn (σ(Pm )) and shows how to compute their coranks.
We then apply these results to the problem of determining the reversibility of the σ-automaton on
Pm,n in section 5. In section 6 we briefly indicate how our results can be generalized to σ-automata

4
on cylinders. The last section concludes with a few open problems. The appendix contains tables
of a the π-polynomials (or rather: their essential irreducible factors, see below) up to π51 , their
counterparts under the involution x 7→ 1 + x on F2 [x], and their depths.
Background information from linear algebra and the theory of finite fields can be found in, say, [6],
[11] or [7]. The second and third reference and Berlekamp’s classic text on coding theory [4] both
contain a careful discussion of the relationship between shift registers and finite fields.

2 Binary Chebyshev Polynomials

Consider the sequence of π-polynomials πn , n ≥ 0, over F2 [x] given by π0 = 0 and

πi (x) := Ui−1 (x/2) mod 2,

where Un denotes the n-th Chebyshev polynomial of the second kind. A little calculation shows
that
Xn + i
πn (x) = xi mod 2.
2i + 1
i

Thus, the coefficients of the π-polynomials are closely related to to the well-known pattern of
binomial coefficients modulo 2. Figure 1 below shows a plot of the first 100 π-polynomials.
By Lucas’ theorem [9], xy mod 2 =

Q xi

yi mod 2, where xi is the i-th digit in the binary expansion
of x, and likewise for y. Thus, the coefficients of πn can be computed in constant time, at least for
machine sized integers n. This representation can be used to establish proposition 2.1 below, but
k
the necessary calculations are rather tedious. For example, one can see that π2k = x2 −1 .
For our purposes, the most useful representation of π-polynomials is in terms of a second order
homogeneous recurrence over F2 [x]:

π0 = 0,
π1 = 1,
πn = x · πn−1 + πn−2 . (1)

As mentioned earlier, the recurrence is reversible in the sense that πn−2 = x πn−1 + πn . We will also
have occasion to study the same recurrence over other algebraic structures such as finite fields F2k
and endomorphism rings End(C) where C is a finite vector space over F2 . Note that the reversibility
of equation (1) is preserved over these structures, and therefore the corresponding sequences must
all be periodic. For the time being, though, we only consider polynomials over F2 .
We hasten to point out that the numbering of the polynomials differs from the one used in [13] and
[2]; the polynomials there begin with π0 = 1 (i.e., in the old numbering πi (x) = Ui (x/2) mod 2).
As it turns out, the current numbering makes it easier to state some of the divisibility properties.

Our interest in these polynomials in connection with σ-automata comes from the fact that for
product graphs G = H × Pn , every pattern X in the kernel of σG is already completely determined
by its first row X1 = row1 (X). Here we assume that patterns are represented by m by n matrices
over F2 , where m is the cardinality of H. The partial patterns are given by Xi = πi (X1 ), and the
complete pattern (X1 , X2 , . . . , Xn ) lies in the kernel of σG iff πn+1 (σH )(X1 ) = 0. In particular, σG
is reversible iff the map πn+1 (σH ) is injective.

5
Figure 1: A plot of the coefficients of the first 100 π-polynomials. A box indicates a 1, and a blank
a 0. The low-order terms are on the left.

The basic divisibility properties of binary Chebyshev polynomials are shared by a whole class of
recurrences over F [x]. In fact, the following proposition holds in any Euclidean domain; see also
[2] for a direct proof for π-polynomials.

Proposition 2.1 Let a, b ∈ F [x] be coprime, where F is an arbitrary finite field. Define a sequence
(τn ) in F [x] by τ0 = 0, τ1 = 1, and τn = a · τn−1 + b · τn−2 , for all n ≥ 2. Then

τp = τq+1 τp−q + b · τq τp−q−1 ,

gcd(τn , τm ) = τgcd(n,m) ,
k −1 k
τ2k n = a2 τn2 ,
2
τ2n+1 = τn+1 + b · τn2 .

where p ≥ q + 1.

The proofs are straightforward by induction and will be omitted. Note that for the binary Cheby-
shev polynomials we have a = x and b = 1, so that the last two equations can also be written
as
k −1 k
π2k n = x2 πn2 ,
2
π2n+1 = πn+1 + πn2
= (πn+1 + πn )2 .

Also note that as a consequence of the second equation, m | n ⇐⇒ πm | πn .

6
2.1 Factoring π-Polynomials

From the last proposition, it is trivial to determine the GCD of π-polynomials. As we will
show in section 5, in connection with σ + -automata it is necessary to compute GCDs of the form
gcd(πi (x), πj (x + 1)). To this end, we will now determine how π-polynomials factor into irreducible
components. We will see later that indeed all irreducible polynomials over F2 occur as factors of
some π-polynomial. For the time being, note that except for τ = x = π2 , all irreducible factors
τ must first occur at odd levels by proposition 2.1. Furthermore, since π2n+1 is a square by the
same corollary, all irreducible factors in π2n+1 must occur at least squared. Define the depth of an
irreducible polynomial τ ∈ F2 [x] to be


dep(τ ) := min n  τ divides πn ,

whenever such an n exists, and let Y

ρn := τ2
dep(τ )=n

be the squared product of all irreducible factors that occur at level n for the first time. For the sake
of completeness, let ρ1 = π1 = 1. We will refer to ρn as the critical term

2 of πn . Note that ρn
may
2
be a product of squares of irreducible factors; e.g., ρ17 = 1 + x + x4 1 + x + x2 + x3 + x4 .
The remainder of this section is devoted to a proof of the following characterization of the π-
polynomials in terms of their irreducible factors. Let φ denote Euler’s totient function.

Theorem 2.1 For all positive n = 2k p, where p is odd,

k
Y k k
Y k
πn (x) = x2 −1 ρ2d (x) = x2 −1 ρd (x2 ).
d|p d|p

Furthermore, deg ρd = φ(d) unless d = 1.

Möbius inversion allows one to compute the critical factors in terms of the π-polynomials: ρn =
Q µ(d)
d|n πn/d , again for all odd numbers n, where µ denotes the Möbius function.

Proof. (Of theorem 2.1.)

The first step is to show that whenever a critical term ρd divides a π-polynomial πn , then all of πd
also divides ρn .

Claim 1: Let τ be an irreducible factor of depth d. Then τ | πn implies that πd | πn .

If τ | πn , then, by the last proposition, τ | gcd(πd , πn ) = πgcd(d,n) . But then, by the definition of
depth, d ≤ gcd(n, d), and it follows that indeed d = gcd(d, n), whence πd | πn .

As a consequence, ρd | πn iff d | n. It follows that every π-polynomial πn is a product of powers of

terms ρd where d > 1 ranges over the divisors of n, and it remains to determine the exponents. For
even n we can use proposition 2.1 to reduce the problem of finding a decomposition for πn to the
problem of finding a decomposition for πp where n = 2k p, p odd. Thus, we
only have to show that
multρd (πp ) = 1 for p odd, d > 1. Here multa (b) := max i  ai divides b denotes the multiplicity
of a in b. To this end, write ∆n := πn+1 + πn so that π2n+1 = ∆2 . We have to show that the
polynomials ∆n contain no square factors.

Claim 2: The polynomials ∆n are square-free, for all n.

7
It suffices to prove that polynomial ∆n and its formal derivative ∆0n are coprime, for all n. Here is
the argument for n even, the odd case is entirely similar. For n = 2m,

∆0n = πn+1
0
+ πn0 = πn0 = πm
2
.

Now suppose τ is an irreducible factor dividing both ∆n and ∆0n . Then τ divides πm and therefore
πn . But then τ must also divide πn+1 , and, by the last proposition, dep τ divides both n and n + 1.
Thus, dep τ = 1, τ = 1, and we are done.

To complete the proof of the theorem it now remains to verify that the degree
Q of ρn is φ(n), for all
odd n > 1, the case n = 2 being trivial. We have already seen that πn = d|n ρd for odd n. By
induction we get
X X
deg ρn = n − 1 − ρd = n − 1 − φ(d) = φ(n),
d|n,1<d<n d|n,1<d<n
P
since d|n φ(d) = n. 2

As an example of theorem 2.1, we consider π300 . We have the factorization

π300 = x3 (1 + x)8 (1 + x + x2 )8 (1 + x3 + x4 )8
(1 + x + x2 + x3 + x5 + x6 + x10 )8 (1 + x3 + x4 + x5 + x9 + x12 + x13 + x15 + x20 )8

where the irreducible terms are associated with divisors 2, 3, 5, 15, 25, and 75, respectively. All
critical factors are squares of just one irreducible term in this case.

3 The Depth of an Irreducible Polynomial

We now show that all irreducible polynomials over F2 occur as factors of some π-polynomial, so
that the depth function from the last section is indeed well-defined for all irreducible polynomials.
To this end, consider τ ∈ F2 [x] irreducible of degree k. Set q := 2k and let α be a root of τ in the
splitting field Fq . Substituting, we obtain a sequence si := πi (α) in Fq given by the second order
homogeneous recurrence

sn+2 = α sn+1 + sn (2)

with initial conditions s0 = 0, s1 = 1. Thus, (si ) is a feedback shift register sequence or linear
recurrent sequence, see [4] and [7] for a wealth of background information. We will frequently use
results from these sources without further mention. Since the coefficient of sn in the recurrence is
1, the sequence must be periodic. By our choice of initial conditions, our sequence is the impulse
response sequence associated with recurrence (2), and therefore maximizes the period of all such
sequences. Needless to say, the period of sequence (2) is none other than the depth d of polynomial
τ : 0 = sd = π(α) ∈ Fq ' F2 [x]/(τ ). Thus we have the following lemma.

Lemma 3.1 Every irreducible polynomial in F2 [x] occurs as a factor of some π-polynomial.

8
Note that the basic building block of the sequence is a palindrome: si = sd−i for all 0 ≤ i ≤ d.
To obtain more information about the depth of an irreducible polynomial over F2 , consider
  the
companion matrix of recurrence (1). Over Fq , the companion matrix takes the form 01 α1 , and
the depth of τ must be a divisor of the order of A in GL(2; Fq ). The latter is well-known to be
q(q − 1)(q 2 − 1). In fact, one can obtain a slightly stronger result: the order of A is a divisor of
2 lcm(q − 1, q + 1), see [12].
Since we can rule out the factor 2, the depth of an irreducible polynomial must actually be a divisor
of q − 1 or q + 1. To determine which irreducible polynomials give rise to depths dividing q − 1 and
which have depth dividing q + 1, consider the characteristic polynomial of recurrence (2):

f (z) := z 2 + αz + 1 ∈ Fq [z]. (3)

We have chosen z as the indeterminate to avoid confusion with the irreducible polynomial τ (x).
In the splitting field of f we can obtain the following simple representation of the shift register
sequence (sn ). Since f is an affine polynomial, its roots form an affine subspace of the splitting
field. In this case, it is easy to see that the roots must have the form β1 and β2 = α + β1 = 1/β1 .
If f splits over Fq , then the shift register sequence (si ) can be represented thus, see [7]:

sn = (β1n + β2n )/α. (4)

It follows immediately that the period of (si ), and thus the depth of τ , the order of β1 , and in
particular a divisor of q − 1, the order of the multiplicative subgroup of Fq .

Suppose, on the other hand, that f is irreducible over Fq . Then f splits over Fq2 , and the shift
register sequence can be represented in the form sn = θβ1n + (θβ1n )q where β1 is a root of f and θ
an element of Fq2 . Note that the sum is the trace of θβ1n from Fq2 to Fq and the conjugate β1q is
none other than the second root β2 = 1/β1 . A little calculation shows that θ = 1/α ∈ Fq , so that
the representation from equation (4) still holds. Hence, the depth of τ is the order of β1 in the
multiplicative subgroup of Fq2 , and, since β1q = 1/β1 , a divisor of q + 1.

Given τ , it is easy to check whether f splits over Fq .

Lemma 3.2 The characteristic polynomial of recurrence (2) is irreducible over Fq iff the linear
term in τ is 0.

Proof. Let f be the characteristic polynomial as in equation (3). By the characteristic 2 version
of Stickelberger’s theorem, see [4], f has two irreducible factors iff the trace of 1/α is 0. Lastly, the
absolute trace over Fq of 1/α is the coefficient of xk−1 in the minimal polynomial of 1/α, which is
none other than the coefficient of x in the reciprocal polynomial τ . 2

We note in passing that one can find a reasonably explicit description of the roots of the charac-
teristic polynomial f in the case where f splits over Fq . Consider any element θ ∈ Fq such that
Tr θ = 1 and let γ := 1/α2 . Define
m−1
X iX i−1
0 2 j
β := θ γ2
i=1 j=0

9
Then αβ 0 is a root of f , as one can verify via a simple calculation. In particular for m odd we can
choose θ = 1, so that
m−1
α(α−4 + α−16 + . . . + α−2 )
is a root of f .

It remains to pin down the relationship between the degree of an irreducible polynomial and its
depth a little more carefully. Recall ∗
 i that the suborder

of 2 in the multiplicative group Zn , n odd,
is defined as sordn (2) = min i  2 ≡ ±1 (mod n) . Clearly, sordn (2) is either the same as the
standard order of 2 in Z∗n , or half that value. We will show that degree of an irreducible polynomial
is the suborder of 2 in Z∗d , where d is the depth of the polynomial.

Theorem 3.1 Let τ ∈ F2 [x] be an irreducible polynomial of degree k and d its depth. Letting
q = 2k , d divides q − 1 iff the linear term in τ vanishes, and q + 1 otherwise. In either case, k is the
suborder of 2 in the multiplicative group Z∗d .

Proof. The first claim of the theorem follows immediately from the last lemma and the preceding
comments. Letting l := sordd (2), we conclude that l ≤ k.
Setting r := 2l , a simple induction shows that
i
X
πr−1 = xr−2
1≤i≤l
πr+1 = πr−1 + xr .

We will need the following technical claim.

Claim: Let K ⊆ L ⊆ M be a tower of finite fields where K = F2 , L = F2l and M = Fq . Let b be
an element of M such that TrL
K b = c ∈ K. Then b ∈ L.
To verify the claim, consider the Frobenius homomorphism h(z) = z 2 of L/K. Thus,
X
c = TrL
K b = hi (b).
0≤i<l

Applying h we obtain X
h(c) = c = hi (b).
1≤i≤l

2l
Adding, we find b = b . But then b must lie in L, the fixed field of hl , the Frobenius homomorphism
of M over L.

We can now establish the second part of the theorem as follows. Suppose for the sake of a contra-
diction that l = sordd (2) < k. As before, we write α rather than x to denote a root of τ in the
splitting field Fq .
First assume that d divides 2l − 1. It is easy to see that l must then divide k. Hence, F2l is a
subfield of Fq . Then
l i l i
X X
0 = s2l −1 = α2 −2 = α2 (1/α)2 .
1≤i≤l 1≤i≤l

Thus, the trace of 1/α from F2l to F2 is 0. From the claim, it follows that 1/α lies in the proper
subfield F2l , therefore α lies in the same subfield, and we have the desired contradiction.

10
 l
It remains to deal with the situation d | 2l + 1 for some l < k. In this case, 0 = s2l +1 = s2l −1 + α2 .
l i
Hence, 0 = α2 (1 + 1≤i≤l (1/α)2 ) and again we obtain a contradiction via our claim.
P
2

As a consequence of the last theorem, all irreducible factors of a critical factor ρn must have the
same degree. Furthermore, all linear terms must have the same coefficient. Since ρn has degree
φ(n), the number of irreducible factors in ρn is φ(n)/(2 sordn (2)). The relation deg τ = sorddep τ (2)
leaves an exponential range of possibilities for the depth of an irreducible polynomial. The only
bounds immediately available are 2k + 1 ≤ dep τ ≤ 2k + 1 where k is the degree of τ . The upper
bound is tight. It appears that for each d such that sordd (2) = k there exists an irreducible
polynomial of degree k whose depth is d. For example, table 3 in the appendix shows all 30
irreducible polynomials of degree 8 and their depths. The possible choices for d are 51, 85, 255 and
257 in this case.

4 The Minimal Polynomial of σ(Pm )

We now return to the discussion of σ-automata over product graphs G = H × Pn . Let C be the
pattern space of H and let f be any endomorphism of C as an F2 vector space. Then the map
F2 [x] → End(C), x 7→ f , is a homomorphism into a finite ring. Hence, the sequence (πi (f )) is
periodic. We are here interested in the case f = σ(Pm ) or f = σ + (Pm ). To lighten notation, we
write Sm for σ(Pm ) and Sm+ = S + id for σ + (P ). The following basic properties of π (S ) are
m m n m
obvious from our previous discussion.

Proposition 4.1 For all n ≥ m ≥ 1 we have the following basic symmetry properties:
cork πn+1 (Sm ) = cork πm+1 (Sn ),
cork πn+1 (Sm ) = cork π(n+1)−(m+1) (Sm )
= cork πgcd(n+1,m+1) (Sm ).

The crucial connection between binary Chebyshev polynomials and rule σ on paths is the fact
that the minimal polynomial of σ(Pm ) is πm+1 . Using a geometric argument, it is shown in [13]
that πm+1 is an annihilator of Sm . As pointed out in [2], this fact follows immediately from the
Caley-Hamilton theorem since πm+1 is none other than the characteristic polynomial of Sm . We
will now show that πm+1 is indeed the minimal polynomial of Sm .
For the proof of minimality, it is convenient to use the following notation system for vectors in
v ∈ F2m . Let {1, 2, . . . , m} be the standard basis and write correspondingly for any non-zero vector
v: v = v[1] + v[2] + . . . + v[k], where 1 ≤ v[1] < v[2] < . . . < v[k] ≤ m. We will refer to k as the
length of v.

Lemma 4.1 πm+1 is the minimal polynomial of Sm .

Proof. From the preceding discussion, it suffices to show that no polynomial of degree d < m
annihilates Sm . To see this, note that Smi (1) = v[1] + v[2] + . . . + v[r] + (i + 1) for all 0 ≤ i < m: S
m
is the sum of a right and a left shift with fixed boundary conditions. But then, for any polynomial
τ of degree d < m, we have τ (Sm )(1) = v[1] + v[2] + . . . + v[r] + d 6= 0, since the contribution of
the leading term cannot be canceled by lower order terms. 2

11
The critical terms of πm+1 are coprime by definition, hence we have the following fact.

Proposition 4.2 Let m + 1 = 2k p where p is odd. Then

k −1 k
X
m = cork πm+1 (Sm ) = cork x2 (Sm ) + cork ρ2d (Sm ).
d|p

Lemma 4.2 Let τ be any polynomial of degree d where 1 ≤ d ≤ m. Then cork τ (Sm ) ≤ deg τ .

Proof. Suppose that v1 , . . . , vd is a collection of linearly independent elements of the kernel of

τ (Sm ). As in lemma 4.1, it follows that 1 ≤ vi [1] ≤ d for all i = 1, . . . , d . From linear algebra, we
obtain a new basis v10 , . . . , vd0 such that vi0 [1] = i, and d < vi0 [2] ≤ m or the length of vi0 is only 1.
Now suppose u is a kernel element not contained in the linear hull of v10 , . . . , vd0 . Then there is a
linear combination w of v10 , . . . , vd0 such that (u + w)[1] > d. But then, by the standard argument,
τ (Sm )(u + v) 6= 0, contradiction. 2

Lemma 4.3 For any factor τ of πm+1 , we have cork τ (Sm ) = deg τ .

Proof. First note that it suffices to establish the claim for factors τ = τ0e where τ0 is irreducible,
since corank is additive on orthogonal factors. So let m + 1 = 2k p where p is odd. From proposition
4.2 and lemma 4.2 we have
k k+1
X X
m = cork x2 −1 (Sm ) + cork τ 2 (Sm )
d|p dep τ =d
X
k k
≤ 2 −1+2 φ(d) = m.
d|p

k k+1
Hence, equality holds everywhere and we have cork x2 −1 (Sm ) = 2k − 1 and cork τ 2 (Sm ) =
2k deg τ . For the sake of simplicity we will only consider the irreducible terms τ 6= x, the argument
for x is entirely similar.
So suppose τ is any irreducible factor of πm+1 and write f := τ (Sm ) to lighten notation. It follows
from linear algebra that cork f j ≤ j cork f for all j. But then cork f j = j cork f = j deg τ for all
j ≤ 2k+1 , as required. 2

Since Pm is undirected, the map f = τ (Sm ) in the last proof is self-adjoint. It follows that
cork f 2 = 2 cork f iff ker f ⊆ rg f = (ker f )⊥ , and likewise for higher powers of f . This will be used
again in theorem 5.2.
We can now determine the corank of any map τ (Sm ) in terms of the GCD of τ and the minimal
polynomial of Sm .

Theorem 4.1 Let τ be an arbitrary polynomial in F2 [x]. Then

cork τ (Sm ) = deg gcd(τ, πm+1 ). (5)

Proof. We will first show that factors of τ that are orthogonal to πm+1 do not contribute to the
corank of τ (Sm ).

12
Claim 1: cork τ (Sm ) = 0 iff τ and πm+1 are coprime.
First suppose that τ and πm+1 have a nontrivial common factor τ0 . Then, by the last lemma,
cork τ (Sm ) ≥ τ0 (Sm ) = deg τ0 > 0. On the other hand, suppose that τ and πm+1 are coprime.
Since τ πm+1 is an annihilator of Sm we have m = cork τ πm+1 (Sm ) = cork τ (Sm ) + cork πm+1 (Sm ).
But the last term is m, so that cork τ (Sm ) = 0, as required.

It follows from claim 1 that the corank of τ (Sm ) is the same as the corank of δ e (Sm ) where the
Q
product is over all irreducible divisors δ of πm+1 , and e = multδ (τ ). By orthogonality, cork τ (Sm ) =
cork δ e (Sm ).
P

Claim 2: For all irreducible divisors δ of πm+1 : cork δ e (Sm ) = min(e, multδ (τ )) deg δ.
As in the last lemma, write f := δ(Sm ). We have already shown that cork δ e (Sm ) = e deg δ as long
as e ≤ e0 := multδ (τ ). For e ≥ e0 we have m = cork δ e (Sm ) + cork θ(Sm ) where θ := πm+1 /δ e0 .
But then cork δ e (Sm ) = e0 deg δ and we are done.
P
By the second claim, cork τ (Sm ) = δ min(multδ (τ ), multδ (πm+1 )) deg δ = deg gcd(τ, πm+1 ). This
completes the proof of the theorem. 2

Since πm+1 is the minimal polynomial of Sm , we can decompose the pattern space F2m of Pm into a
direct sum of subspaces Ei , associated with the invariant factors of πm+1 . For example, if m = 2p is
even, the invariant factors take the form (τ1 . . . τr )(Sm ) and (τ1 . . . τr )2 (Sm ). The invariant subspace
of the first map has dimension p and consists of all symmetric patterns X such that X(p + 1) = 0.
As a F2 [x]-submodule, it is generated by the pattern 1 + m. For m + 1 = 2k p we have symmetric
patterns associated with π(m+1)/2 (Sm ), double symmetric ones associated with π(m+1)/4 (Sm ), and
so forth. Invariant subspaces will be used in the proof of theorem 5.2 below.

5 π + -Polynomials

The last theorem gives a complete characterization of the coranks of a σ-automata on a grid.
Indeed, the description of the corank of Sn,m in [13, 2] can be rephrased as follows:

d+ (Pn,m ) = deg πn+1 (Sm ) = deg gcd(πn+1 , πm+1 ) = deg πgcd(n+1,m+1) = gcd(n + 1, m + 1) − 1.

In order to extend this result to σ + -automata, first note the following basic symmetry properties.

Proposition 5.1
+ +
cork Sn,m = cork πn+1 (Sm ) = cork πm+1 (Sn+ )
+ +
= cork πn+1 (Sm ) = cork πm+1 (Sn ).

Here πn+ (x) := πn (1 + x). It is advantageous to consider πn+ as the result of applying the endomor-
phism F2 [x] → F2 [x] induced by x 7→ 1 + x to πn . In keeping with previous notation, we will write
τ + for the image of polynomial τ under this map. Now + is an involution, so the π + -polynomials
inherit the divisibility properties of the π-polynomials. For example, πm + | π + iff m | n. Moreover,
n
+
all irreducible polynomials occur as factors of some π -polynomial. We can think of the involution

13
+ as acting on the endomorphism ring End(C): σ + = σ + id. Hence, it follows from lemma 4.1
+ +:
that πm+1 is the minimal polynomial of Sm

τ + (Sm
+
) = τ + (Sm + id) = τ + (1 + x)(Sm ) = τ ++ (Sm ) = τ (Sm ).

+ as follows.
Theorem 4.1 allows one to determine the corank of Sn,m

Theorem 5.1 For all positive n and m:

+ +
d+ (Pn,m ) = deg gcd(πn+1 , πm+1 ) = deg gcd(πn+1 , πm+1 ).

Proof. Since the second claim follows from the first by symmetry, it suffices to prove the first
+ = cork π + (S ) = deg gcd(π + , π
equation. But cork Sn,m n+1 m n+1 m+1 ), by the last theorem, and we are
done. 2

For a few special values of n and m one can determine d+ (Pm,n ) completely from theorem 5.1, see
also theorem 4.6 of [2].

Corollary 5.1 For m + 1 = 2k and n + 1 = 2l p, p odd, we have


 0 if 3 6 | n + 1,
d+ (Pn,m ) = 2l+1 if l < k − 1,
 k
2 − 1 otherwise.

Corollary 5.2 For m + 1 = 2k · 3 and n + 1 = 2l p, p odd, we have


 0 if 2, 3 6 | n + 1,
l



 2 −1 if l ≤ k + 1, 3 6 | n + 1
 k+1

2 if l > k + 1, 3 6 | n + 1
d+ (Pn,m ) = l

 3·2 −1 if l + 1 < k, 3 | n + 1


 2l − 1 + 2k − 1 if k − 1 ≤ l ≤ k + 1, 3 | n + 1

3 · 2k − 1 l ≥ k + 2, 3 | n + 1

if

In general, though, there seems to be no simple way to determine GCDs of the π + -polynomials and
π-polynomials as a function of the parameters m and n.

5.1 Reversible Grids

Consider the somewhat easier problem of determining all reversible σ + -automata on an m by n

grid where m is fixed. As we have seen, reversibility of the σ + -automaton on Pm,n = Pm × Pn is
+
equivalent to πm+1 being coprime to πn+1 . The latter condition can be expressed as a divisibility
condition on n + 1.

Lemma 5.1 Fix m ≥ 1. Then there are positive integers t1 , . . . , tr such that the σ + -automaton
on Pm,n , n ≥ m, is reversible iff ti does not divide n + 1, for all i = 1, . . . , r .

Proof. As usual, write m + 1 = 2k p where p is odd and let q = 2k . Then

Y 2q
πm+1 = xq−1 τi
1≤i≤r

14
for certain irreducible polynomials τ1 , . . . , τr . Set τ0 := x and let ti := dep τi+ for i = 0, . . . , r .
Then the σ + -automaton on Pm,n is reversible iff ti does not divide n + 1, for all i = 0, . . . , r .
+
To see this, first assume that σ + (Pm,n ) is reversible. Then, by the last theorem, πm+1 must divide
+
πn+1 . Therefore τi divides πn+1 and, by claim 1 in the proof of theorem 2.1, πti must divide πn+1 .
By proposition 2.1, ti divides n + 1, as required.
On the other hand, suppose that none of the ti divides n + 1, for i = 0, . . . , r . But then none of
+
the τi+ divides πn+1 either and we must have gcd(πn+1 , πm+1 ) = 1. Hence σ + (Pm,n ) is reversible
and we are done. 2

5.2 Periodicity Properties of Grids

We now turn to linear cellular automata of the form G = H × P∞ where the local rule is given
by rule σ + and the alphabet is F2V (H). Using the same extension argument as in [13], it is not
hard to see that d+ (G) = 2d+ (H). It follows that the global map of G is injective iff the σ + (H)
is injective; furthermore, the global map is always open. By an argument similar to the one in
lemma 3.1, we can see that all kernel patterns are periodic and the periods are uniformly bounded.
Denote the maximal period of the σ + -automaton G by per+ (H). In order to determine per+ (H),
suppose X is a kernel pattern σ + (G) of period p such that X0 = 0. Then the partial configuration
(X1 , . . . , Xp−1 ) is clearly in the kernel of σ + (H × Pp−1 ).
This leads the following definition. Suppose H is a graph on m points. A σ + -automaton on H × Pn
+
is totally irreversible if d+ (H ×Pn ) = m. Hence, in a totally irreversible automaton πn+1 (σ(H)) = 0.
+
Note that whenever πp (σ(H)) = 0, it follows from the symmetry of equation (1) that

πp−i (σ(H)) = πp+i (σ(H)), for all 0 ≤ i ≤ p.

Thus, any kernel pattern on the infinite cellular automaton G must consist of identical blocks of
length p − 1, possibly reversed, and separated by 0-patterns. The pattern is symmetric with respect
to each of these 0-patterns. This suggests to define the weak period of H under σ + , in symbols
per+ (H), to be the least p > 0 such that H × Pp is totally irreversible.
It follows that πp+ (σ(H)) = 0 iff p is a multiple of per+ (H) iff the minimal polynomial of σ(H)
divides πp+ . Specifically, if H is a path, it follows that
 +
wper+ (Pm ) = min n  πm+1


divides πn+1 .

The basic properties of the period and weak period are expressed in the next proposition.

Proposition 5.2 For all graphs H we have:

• per+ (H) = wper+ (H) or per+ (H) = 2 wper+ (H).

• If per+ (H) is odd, then per+ (H) = wper+ (H).
• If per+ (H) = 2p is even, then πp (σ + (H))(X) lies in the kernel of σ + (H), for all patterns X.

Proof.
It is clear that wper+ (H) | per+ (H), so suppose p := wper+ (H) < per+ (H). Then by the symmetry
of equation (1) we have πp−i (σ(H)) = πp+i (σ(H)) for all 0 ≤ i ≤ p. In particular, π2p (σ(H)) = 0

15
and π2p−1 (σ(H)) = id. But then π2p+1 (σ(H)) = id and we are done with the first claim. The
second claim follows immediately from the first.
+
For the third claim note that 0 = π2p + (π (S + )).
(Sm ) = ((1 + x)πp+ )(Sm ) = Sm p m 2

Returning to paths, we can describe the period of Pm under rule σ + as follows. First, define the
following analogue to the depth function

dep+ (τ ) := min i  τ divides πi+ ,



where τ is irreducible. Of course, dep+ (τ ) = dep(τ + ) since F is an involution. Second, to simplify

the statement of the next theorem, let us adopt Knuth’s convention to write [ϕ] for the boolean
value, interpreted as 0 or 1, of any predicate ϕ, see [5].

Theorem 5.2 Let m + 1 = 2k p where p is odd and set

D := lcm dep+ (τ )  τ | πm+1 , τ 6= x, 1 + x, τ irreducible .



Here we assume D = 1 if no such factor τ exists. Then


 2k · 3 if p = 1,
+
wper (Pm ) = (6)
 2k+2[3|m+1] lcm(D, 3[2|m+1] ) otherwise.

Moreover, per+ (Pm ) = wper+ (Pm ) iff 3 | m + 1, and per+ (Pm ) = 2 wper+ (Pm ) otherwise.

Proof. To determine the weak period of Sm + we have to find the least n > 0 such that π annihilates
n
+ +
Sm . Equivalently, πn must annihilate Sm . Since πm+1 is the minimal polynomial of Sm , it follows
+
that πm+1 must divide πn+ , or, equivalently, πm+1 must divide πn .
As we have seen,   Y k+1
k −1 k+1
πm+1 = x2 · (1 + x)2 · τ2
where the product is over all irreducible τ 6= x, 1 + x that divide πp and the parenthesized middle
term only occurs if 3 | p. Hence,
2k+1
k
 k+1  Y
+
πm+1 = (1 + x)2 −1 · x2 · τ+ .

A straightforward, if tedious, calculation now shows that wper(σ(Pm )) has the form as claimed in
the theorem.
It remains to show that wper+ (Pm ) = per+ (Pm ) iff 3 | m + 1. First, suppose that 3 | m + 1.
Then the weak period of m must be even, say, 2p. A little calculation shows that cork πp (Sm +) =
+ + +
deg gcd(πp , πm+1 ) = πm+1 /x = m − 1. Hence, the range of πp (Sm ) is one-dimensional. Now the
+
P
kernel of Sm is Z := 1 + 2 + 4 + 5 + . . . + (m − 1) + m and it is easy to see that i Z × {2i + 1}
is a pattern in the kernel of σ(Pm × P2p−1 ). But then the range of πp (Sm + ) is the kernel of S + .
m
It now follows immediately from the symmetry of equation (1) that πp−1 (Sm +) = π +
p+i (Sm ) for all
+) = π
i = 0, . . . , p . In particular, id = π1 (Sm +
2p−1 (Sm ), whence the period of m is 2p.

The second case, 3 6 | m + 1, requires a slightly more complicated argument. Note, though, that
as long as the period of m is even, say, 2p, we can proceed as in the previous case: the range of
+ ) must be the kernel of S + , which is trivial since 3 does not divide m + 1. Thus, the weak
πp (Sm m
period of m must be p.

16
Hence, it suffices to show that Pm cannot have odd period. So suppose for the sake of a contradiction
that per+ (Pm ) = 2p + 1 is odd. It follows that the weak period of m must agree with the full
period. From the first part of the theorem, neither 2 nor 3 can divide m + 1 and we must have
πp = (τ1 . . . τr )2 , where all the τi are irreducible polynomials and different from x and 1 + x. We
can enumerate these irreducible factors in such a way that πm+1 = (τ1+ . . . τs+ )2 for some s ≤ r.
We will write τ := τ1 . . . τs and f := τ (Sm + ). Thus, f 2 = τ 2 (S + ) = τ + 2 (S ) = π
m m m+1 (Sm ) = 0. We
will now show that the invariant subspaces corresponding to f and f 2 are the symmetric patterns
in F2m , and the asymmetric patterns, respectively. More information on orthogonal subspaces in
connection with σ-automata can be found in [13, 14].

Claim 1: F2m = E1 ⊕ E2 where E1 is the kernel of f = τ (Sm

+ ). Then E consists precisely of all
1
symmetric patterns in F2m and has dimension m/2.
+ . Furthermore,
It is clear from the definition that E1 is closed under reversal and also closed under Sm
E1 is self-orthogonal. To see this, note that f 2 = 0, whence ker f ⊆ rg f = (ker f )⊥ , where the last
equality follows from the self-adjointness of f . It follows that E contains only symmetric patterns.
For suppose X = X[1] + X[2] + . . . + X[k] is a pattern in E. k must be even, by orthogonality.
Since every basis vector except for 1 and m maps to an odd-cardinality pattern under Sm + it follows
+
that X[1] + X[k] = m + 1. Otherwise, the orbit of X under Sm would contain an odd-cardinality
pattern. Repeating the argument for X + X[1] + X[k] shows that X is symmetric. The degree of
τ is m/2, and it follows by theorem 4.1 that the dimension of E is also m/2. Our claim follows.

Let us write Z for the kernel of σ(Pm,2p ). We will think of these patterns as m by 2p matrices over
F2 . Let V := { row1 (K)  K ∈ Z } ⊆ F22p be the subspace consisting of the top rows of patterns in


Z.
Claim 2: V = V1 ⊕ V2 where V1 = ker τ (S2p ) contains precisely all the symmetric patterns in V .
Moreover, col1 (K) ∈ E1 iff row1 (K) ∈ V1 .
For the proof of the second claim, it is best to write the action of σ(Pm,2p ) on any two-dimensional
kernel pattern K as a matrix equation
+
Sm · K + K · S2p = 0.

It follows immediately that

+
τ (Sm ) · K + K · τ (S2p ) = 0.
Since the subspaces E1 ⊆ F2m and V1 ⊆ F22p are both closed under Sm and S2p , respectively, we
have, for any two-dimensional pattern K in Z: τ (Sm + ) annihilates the first column of K iff τ (S + )
m
+
annihilates all columns of K iff τ (S2p ) annihilates all rows of K iff τ + (S2p ) annihilates all rows of
+
K iff τ + (S2p ) annihilates the first row of K. In other words, col1 (K) lies in E1 iff row1 (K) lies in
V1 .
We can now repeat the argument of claim 1 for the map g = τ (Sm+ ) : V → V . To see that every

symmetric pattern in V must lie in V1 note that ker g = (ker g)⊥ : every symmetric pattern is
orthogonal to the kernel, and therefore an element of it.

From claims 1 and 2, we can write Z as a direct sum Z1 ⊕Z2 , where both spaces have dimension m/2.
Moreover, Z1 consists of all patterns that have symmetric columns, or, equivalently, symmetric rows.
In particular, there are 2m/2 patterns in Z whose rows fail to be symmetric. It follows that the
period of m must be larger than wper(m) and we have per(m) = 2 wper(m), as desired. 2

17
The analogue of the following corollary for weak periods was conjectured in [13] and first proved
in [2].

Corollary 5.3 For all k ≥ 0: per+ (2k − 1) = 3 · 2k .

6 Cylinders

In this section we will briefly show how to adapt our results for σ-automata on grids to σ-automata
on cylinders of the form Cm × Pn . First, we have to determine the minimal polynomial for the
maps σ(Cm ).
p
Lemma 6.1 The minimal polynomial of σ(Cm ) is x πm/2 if m is even, and x πm for m odd.

Proof. Consider the case where m = 2m0 is even. Since σ(Cm ) commutes with the shift, we have
to determine the non-trivial polynomial π of lowest degree such that π(σ(Cm ))(m) = 0, where m
denotes a one-point pattern, as in section 4. Note that X := σ(Cm )(m) = 1 + (m − 1) because
of the cyclic boundary conditions. Furthermore, since m is even, every pattern X 0 = σ(Cm )t (X),
t ≥ 1, has the properties X 0 (m0 ) = X 0 (m) = 0 and is symmetric with respect to m0 . Hence, we can
simulate the evolution of pattern X on the σ-automaton Cm on the σ-automaton Pm−1 . Since the
patterns on the second automaton are all symmetric, it follows from 5.2 that the the least degree
polynomial π such that π(σ(Cm ))(X) = 0 is none other than πm/2 .
The argument for m odd is entirely similar and will be omitted. 2

It is now straightforward to establish the analogue of lemma 5.1 and characterize the reversible and
totally irreversible cylinders. For example, for odd m, all those cylinders Cm × Pn are reversible
for which n is also odd but gcd(m, n + 1) = 1.
Note, though, there is a slight complication in computing the coranks of rule σ on a cylinder. The
degree of the minimal polynomial of σ(Cm ) is (m + 1)/2 or m/2, depending on whether m is odd
or even. Thus, the arguments of lemmata 4.1 and 4.2 have to be adjusted. In particular, for even
m, the corank of τ (σ(Cm )) is 2 deg τ for all divisors τ of the minimal polynomial x πm/2 . For odd
q
m, the corank of τ (σ(Cm )) is 2 deg τ for all divisors τ of πm/2 , but 1 + 2 deg τ for (xτ )(σ(Cm )).
The following table shows the degree of gcd(π, πn+1 ), where π is the minimal polynomial of σ(Cm ),
as well as the coranks of σ(Cm × Pn ). We write k = mult2 (m) and l = mult2 (n + 1).

degree corank
1/2(gcd(m, n + 1) − 1) gcd(m, n + 1) − 1 0 = k = l,
1/2(gcd(m, n + 1) − 1) + 1 gcd(m, n + 1) 0 = k < l,
1/2 gcd(m, n + 1) gcd(m, n + 1) 0 < k ≤ l,
gcd(m, n + 1) − 1 2(gcd(m, n + 1) − 1) l < k.

The weak period as well as the full period of rule σ on Cm is m for m even, and 2m for m odd.
To determine the weak and full period of rule σ + on cylinders, we have to compute the depth of
τ + , where τ is an irreducible factor of the minimal polynomial of σ(Cm ). Note that x is always
such a factor, hence the weak period of rule σ + on a cylinder must always be divisible by three,
the depth of x+ = 1 + x.

18
Using the same notation as in theorem 5.2, we have for m = 2k p, p odd,


 2[3|m] lcm(3, D) if k = 0,
wper+ (Cm ) = (7)
 2k−1+2[3|m] lcm(3, D) otherwise.

Also, for cylinders the periods are determined thus: per+ (Cm ) = 2 wper+ (Cm ) iff m ≡ 2, 4 (mod 6),
and per+ (Cm ) = wper+ (Cm ) otherwise.

7 Open Problems

We conclude by stating a few open problems about the dynamics of binary σ + -automata that can
be expressed in terms of Chebyshev polynomials.

Reversible Squares
As far as squares are concerned, it follows from our results that in order for the m by m grid to
be reversible under rule σ + we must have 6 6 | m + 1 and for all odd e > 3 such that e | m + 1 and
τ | ρe irreducible: dep(τ + ) 6 | m + 1. Is there a simple algorithm to test the second property? Are
there any totally irreversible squares other than 4 × 4? Equivalently, is there any m > 4 such that
+
πm+1 = πm+1 ? We suspect that the answer to the last question is no.

Computing Depth and Period

Lastly, the most interesting question in connection with binary Chebyshev polynomials is whether
there is an algorithm to compute the depth of an irreducible polynomial τ over F2 , other than the
obvious brute-force approach. In particular, is there an algorithm which is polynomial in the degree
of τ ? Note that the depth of many irreducible polynomials realizes the upper bound 2deg τ ± 1, so
that one cannot enumerate in polynomial time all π-polynomials that are potential multiples of τ .
Is it the case that for each d such that sordd (2) = k there exists an irreducible polynomial of degree
k depth is d? What is the distribution of these polynomials for the possible choices of d? See table
3 in the appendix for a complete listing of the depths of all irreducible polynomials τ of degree 8
as well as their counterparts τ + .
An obviously related problem is the computation of periods per+ (m). Note, though, that there
might be an alternative approach to computing periods that does not use the depth function.

19
References

[1] B. Andrasfai. Cellular automata in trees. In Finite and Infinite Sets, volume 37, pages 35–45.
Colloquia Mathematica Societatis Janos Bolyai, Eger, Hungary, 1981.

[2] R. Barua and S. Ramakrishnan. σ-game, σ + -game, and two-dimensional cellular automata.
Theoretical Computer Science, 154(2):349–366, 1996.

[3] P. Sarkar and R. Barua. Multidimensional σ-automata, π-polynomials and generalized S-

matrices. To appear.

[4] E. R. Berlekamp. Algebraic Coding Theory. McGraw-Hill, 1968.

[5] R. L. Graham, D. E. Knuth, and O. Patashnik. Concrete Mathematics: A Foundation for

Computer Science. Addison-Wesley, 1988.

[6] T. W. Hungerford. Algebra. Springer-Verlag, 1974.

[7] R. Lidl and H. Niederreiter. Finite Fields, volume 20 of Encyclopedia of Mathematics and its
Applications. Cambridge University Press, 1984.

[8] A. Lindenmayer. Mathematical models for cellular interactions in development. Journal of

Theoretical Biology, 18:280–299, 1968.

[9] M. E. Lucas. Sur les congruences des nombres Euleriennes et des coefficients différentiels des
functions trigonométrique, suivant un module premier. Bull. Soc. Math. France, 6:49–54, 1878.

[10] O. Martin, A. M. Odlyzko, and S. Wolfram. Algebraic properties of cellular automata. Com-
mun. Math. Phys., 93:219–258, 1984.

[11] R. J. McEliece. Finite Fields for Computer Scientists and Engineers. Kluwer, 1987.

[12] I. Niven. Fermat’s theorem for matrices. Duke Math. J., 15:823–826, 1948.

[13] K. Sutner. On σ-automata. Complex Systems, 2(1):1–28, 1988.

[14] K. Sutner. Linear cellular automata and the Garden-of-Eden. Mathematical Intelligencer,
11(2):49–53, 1989.

[15] T. Toffoli and Margolus. Injective cellular automata. Physica D, 45(1–3):386–395, 1990.

[16] S. Wolfram. Theory and Applications of Cellular Automata. World Scientific, 1986.

20
8 Appendix

The first table below contains all the critical factors ρe for e ≤ 51 and the second shows the images
of the irreducible factors τ of ρe under the involution x+ = 1 + x. The second column in this table
gives the depth of τ + . A dash indicates a value larger than 2000.

From the table, we can calculate the periods per+ (Pm ) and per+ (Cm ) for some values of m. For
example, per+ (P20 ) = 22 lcm(9, 65) = 2340. By contrast, per+ (P40 ) = 2 · lcm(1025, 1023) =
2097150.

Lastly, in table three, the depth of all irreducible polynomials of degree 8 are shown. Note that all
d such that sordd (2) = 8 occur (d = 51, 85, 255, 257). Those irreducibles whose roots α give rise
to a reducible characteristic polynomial f (z) = z 2 + αz + z of the linear recurrence sequence (si )
have depths dividing 255, and the ones for which f splits only over Fq2 all have depth 257, since
the latter is a Fermat prime.

21
 e ρe
2 x
3 1+x
5 1 + x + x2
7 1 + x2 + x3
9 1 + x + x3
11 1 + x + x2 + x4 + x5
13 1 + x + x4 + x5 + x6
15 1 + x3 + x4
17 1 + x + x4
1 + x + x2 + x3 + x4
19 1 + x + x4 + x5 + x6 + x8 + x9
21 1 + x5 + x6
23 1 + x2 + x3 + x4 + x8 + x10 + x11
25 1 + x + x2 + x3 + x5 + x6 + x10
27 1 + x + x5 + x7 + x9
29 1 + x + x8 + x9 + x12 + x13 + x14
31 1 + x2 + x5
1 + x3 + x5
1 + x2 + x3 + x4 + x5
33 1 + x + x2 + x3 + x5
1 + x + x3 + x4 + x5
35 1 + x5 + x6 + x7 + x9 + x11 + x12
37 1 + x + x2 + x8 + x9 + x10 + x12 + x13 + x16 + x17 + x18
39 1 + x3 + x4 + x5 + x6 + x7 + x8 + x11 + x12
41 1 + x + x3 + x4 + x7 + x8 + x10
1 + x + x4 + x9 + x10
43 1 + x + x7
1 + x + x2 + x3 + x4 + x5 + x7
1 + x + x2 + x4 + x5 + x6 + x7
45 1 + x3 + x4 + x5 + x7 + x9 + x12
47 1 + x4 + x6 + x7 + x8 + x16 + x20 + x22 + x23
49 1 + x2 + x3 + x7 + x9 + x10 + x11 + x14 + x17 + x19 + x21
51 1 + x2 + x3 + x4 + x8
1 + x2 + x3 + x7 + x8

Table 1: The critical factors ρe for e ≤ 51.

22
 e depth ρ+
e
2 3 1+x
3 2 x
5 5 1 + x + x2
7 9 1 + x + x3
9 7 1 + x2 + x3
11 31 1 + x2 + x5
13 63 1 + x2 + x4 + x5 + x6
15 17 1 + x + x2 + x3 + x4
17 17 1 + x + x4
15 1 + x3 + x4
19 513 1 + x + x2 + x4 + x5 + x6 + x9
21 65 1 + x + x2 + x5 + x6
23 – 1 + x4 + x8 + x9 + x11
25 1025 1 + x + x3 + x5 + x6 + x8 + x10
27 511 1 + x2 + x3 + x6 + x7 + x8 + x9
29 – 1 + x + x2 + x4 + x5 + x6 + x8 + x10 + x12 + x13 + x14
31 11 1 + x + x2 + x4 + x5
31 1 + x2 + x3 + x4 + x5
31 1 + x3 + x5
33 33 1 + x + x3 + x4 + x5
33 1 + x + x2 + x3 + x5
35 455 1 + x2 + x7 + x8 + x10 + x11 + x12
37 – 1 + x2 + x5 + x8 + x10 + x13 + x16 + x17 + x18
39 585 1 + x3 + x4 + x7 + x8 + x9 + x10 + x11 + x12
41 1025 1 + x + x2 + x5 + x6 + x7 + x10
1023 1 + x2 + x4 + x9 + x10
43 127 1 + x2 + x3 + x4 + x5 + x6 + x7
127 1 + x2 + x4 + x6 + x7
129 1 + x + x2 + x3 + x7
45 1365 1 + x6 + x7 + x9 + x12
47 – 1 + x8 + x17 + x19 + x20 + x21 + x23
49 – 1 + x + x2 + x4 + x7 + x10 + x11 + x12 + x14 + x16 + x17 + x18 + x19 + x20 + x21
51 257 1 + x + x3 + x4 + x8
85 1 + x2 + x4 + x5 + x6 + x7 + x8

Table 2: The irreducible polynomials τ + where τ divides a critical factor ρe for e ≤ 51. The second
column shows the depth of τ + (a dash indicates a value larger than 2000).

23
 number image depth irreducible polynomial
1 17 257 1 + x + x3 + x4 + x8
2 12 51 1 + x2 + x3 + x4 + x8
3 15 257 1 + x + x3 + x5 + x8
4 29 255 1 + x2 + x3 + x5 + x8
5 13 255 1 + x3 + x4 + x5 + x8
6 10 257 1 + x + x2 + x3 + x4 + x5 + x8
7 16 255 1 + x2 + x3 + x6 + x8
8 23 257 1 + x + x2 + x3 + x4 + x6 + x8
9 9 257 1 + x + x5 + x6 + x8
10 6 255 1 + x2 + x5 + x6 + x8
11 28 255 1 + x3 + x5 + x6 + x8
12 2 85 1 + x4 + x5 + x6 + x8
13 5 257 1 + x + x2 + x4 + x5 + x6 + x8
14 24 257 1 + x + x3 + x4 + x5 + x6 + x8
15 3 257 1 + x + x2 + x7 + x8
16 7 257 1 + x + x3 + x7 + x8
17 1 51 1 + x2 + x3 + x7 + x8
18 19 257 1 + x + x2 + x3 + x4 + x7 + x8
19 18 257 1 + x + x5 + x7 + x8
20 21 85 1 + x3 + x5 + x7 + x8
21 20 255 1 + x4 + x5 + x7 + x8
22 22 255 1 + x2 + x3 + x4 + x5 + x7 + x8
23 8 257 1 + x + x6 + x7 + x8
24 14 257 1 + x + x2 + x3 + x6 + x7 + x8
25 30 257 1 + x + x2 + x4 + x6 + x7 + x8
26 27 85 1 + x2 + x3 + x4 + x6 + x7 + x8
27 26 257 1 + x + x2 + x5 + x6 + x7 + x8
28 11 257 1 + x + x4 + x5 + x6 + x7 + x8
29 4 85 1 + x2 + x4 + x5 + x6 + x7 + x8
30 25 255 1 + x3 + x4 + x5 + x6 + x7 + x8

Table 3: The depths of all irreducible binary polynomials τ of degree 8. The image column gives
the index of τ + with respect to the numbering in the first column. All polynomials of the form
1 + x + . . . + x8 have depth 257.

24