Mutually Orthogonal Latin Squares based on Cellular Automata

TL;DR

This paper explores how Cellular Automata over finite fields can generate mutually orthogonal Latin squares, providing conditions for orthogonality, enumeration methods, and constructions for maximal sets of MOLS.

Contribution

It establishes a link between Linear Bipermutive CA and orthogonal Latin squares, offering new enumeration and construction techniques for MOLS based on polynomial properties.

Findings

Orthogonality of Latin squares corresponds to coprimality of local rule polynomials.

Enumeration of MOLS pairs via counting coprime polynomials over finite fields.

Construction of maximal MOLS sets using LBCA with irreducible polynomials.

Abstract

We investigate sets of Mutually Orthogonal Latin Squares (MOLS) generated by Cellular Automata (CA) over finite fields. After introducing how a CA defined by a bipermutive local rule of diameter $d$ over an alphabet of $q$ elements generates a Latin square of order $q^{d-1}$, we study the conditions under which two CA generate a pair of orthogonal Latin squares. In particular, we prove that the Latin squares induced by two Linear Bipermutive CA (LBCA) over the finite field $\mathbb{F}_q$ are orthogonal if and only if the polynomials associated to their local rules are relatively prime. Next, we enumerate all such pairs of orthogonal Latin squares by counting the pairs of coprime monic polynomials with nonzero constant term and degree $n$ over $\mathbb{F}_q$. Finally, we present a construction of MOLS generated by LBCA with irreducible polynomials and prove the maximality of the resulting sets, as well as a lower bound which is asymptotically close to their actual number.

Equations

$$(L_1(i_1,j_1),L_2(i_1,j_1))\ne (L_1(i_2,j_2),L_2(i_2,j_2))$$

$$F(x_0,\cdots ,x_{n-1})=(f(x_0,\cdots ,x_{d-1}),f(x_1,\cdots ,x_d),\cdots ,f(x_{n-d},\cdots ,x_{n-1})).$$

$$f(x_0,x_1,\cdots ,x_{d-1})=x_0\oplus \phi (x_1,\cdots ,x_{d-1}),$$

$$f(x_0,x_1,\cdots ,x_{d-1})=x_0\oplus \phi (x_1,\cdots ,x_{d-2})\oplus x_{d-1},$$

$$f(x_0,\cdots ,x_{d-1})=a_0{x}_0+a_1{x}_1+\cdots +a_{d-1}{x}_{d-1},$$

$$M_F=\left(\begin{array}{ccccccccc}{a}_0& \cdots & a_{d-1}& 0& \cdots & \cdots & \cdots & \cdots & 0\\ 0& a_0& \cdots & a_{d-1}& 0& \cdots & \cdots & \cdots & 0\\ ⋮& ⋮& ⋮& \ddots & ⋮& ⋮& ⋮& \ddots & ⋮\\ 0& \cdots & \cdots & \cdots & \cdots & 0& a_0& \cdots & a_{d-1}\end{array}\right).$$

$$p_f(X)=a_0+a_{1}X+\cdots +a_{d-1}{X}^{d-1}\in {\mathbb{F}}_q[X].$$

$$x_{d-1}$$

$$x_0$$

$${\mathcal{S}}_F(i,j)=\varphi (F(\psi (i)||\psi (j))),$$

$$f(x_0,\cdots ,x_{d-1})$$

$$g(x_0,\cdots ,x_{d-1})$$

$$\mathcal{H}(x,y)=(F(z),\mathcal{G}(z))=(\tilde{x},\tilde{y})$$

$$\{\begin{array}{l}F(z)=M_F{z}^{\top }=\tilde{x}\\ \mathcal{G}(z)=M_G{z}^{\top }=\tilde{y}\end{array}.$$

$$M=\left(\begin{array}{ccccccccc}{a}_0& \cdots & a_{d-1}& 0& \cdots & \cdots & \cdots & \cdots & 0\\ 0& a_0& \cdots & a_{d-1}& 0& \cdots & \cdots & \cdots & 0\\ ⋮& ⋮& ⋮& \ddots & ⋮& ⋮& ⋮& \ddots & ⋮\\ 0& \cdots & \cdots & \cdots & \cdots & 0& a_0& \cdots & a_{d-1}\\ b_0& \cdots & b_{d-1}& 0& \cdots & \cdots & \cdots & \cdots & 0\\ 0& b_0& \cdots & b_{d-1}& 0& \cdots & \cdots & \cdots & 0\\ ⋮& ⋮& ⋮& \ddots & ⋮& ⋮& ⋮& \ddots & ⋮\\ 0& \cdots & \cdots & \cdots & \cdots & 0& b_0& \cdots & b_{d-1}\end{array}\right),$$

$$S_n=\{f(X)=a_0+a_{1}X+\cdots +a_{n-1}{X}^{n-1}+X^n:a_0\ne 0\}$$

$$f(X)$$

$$g(X)$$

$$A_n=\{(f,g)\in S_n^2:\gcd (f,g)=1\},$$

$$B_n=\{(f,g)\in S_n^2:\gcd (f,g)\ne 1\}.$$

$$a_n=q(q-1{)}^3\frac{q^{2n-2}-1}{q^2-1}+(q-1)(q-2).$$

$$S_1=\{f(X)=a_0+X:a_0\ne 0\}.$$

$$a_1=(q-1)(q-2),$$

$$s_n^2=a_n+b_n.$$

$$b_n=\sum _{e=0}^{n-1}{a}_e{s}_{n-e}.$$

$$a_n$$

$$a_{n-1}$$

$$q{a}_{n-1}=q{s}_{n-1}^2-\sum _{e=0}^{n-2}{a}_e(q{s}_{n-1-e}).$$

$$q{a}_{n-1}=q{s}_{n-1}^2-\sum _{e=0}^{n-2}{a}_e{s}_{n-e}.$$

$$a_n-q{a}_{n-1}$$

$$=s_n^2-q{s}_{n-1}^2-a_{n-1}{s}_1.$$

$$a_n-q{a}_{n-1}=(q-1{)}^2(q^{2n-2}-q^{2n-3})-a_{n-1}(q-1),$$

$$a_n$$

Full text

**Keywords ** Mutually orthogonal Latin squares, cellular automata, Sylvester matrices, polynomials

1 Introduction

A Latin square of order $N\in\mathbb{N}$ is a $N\times N$ matrix where each number from $1$ to $N$ appears exactly once in each row and column. Two Latin squares $L_{1}$ and $L_{2}$ of order $N$ are orthogonal if by superimposing them one obtains all ordered pairs $(i,j)$ of numbers from $1$ to $N$, and Mutually Orthogonal Latin Squares (MOLS) are sets of Latin squares that are pairwise orthogonal.

Despite their simple definition, the construction of MOLS is a notoriously difficult combinatorial problem and it is one of the most studied research topics in design theory. This interest is also due to the numerous applications that MOLS have in other fields such as cryptography (for example in the design of authentication codes [29] and multipermutations [30]), coding theory (see e.g. the Golomb-Posner code [9]) and statistics (particularly in the design of experiments [23]). Some of the best known constructions of MOLS include MacNeish’s theorem [16] and Wilson’s construction [31] (see [14, 4] for a more complete overview of construction methods).

The goal of this paper is to investigate a new construction of MOLS based on Cellular Automata (CA), a particular kind of discrete dynamical system described by a regular lattice of cells, where each cell synchronously updates its state by applying a local rule to itself and its neighboring cells. The motivation for studying this construction of MOLS spawned from the question of designing a threshold secret sharing scheme based on CA without adjacency constraints on the shares, as in the schemes proposed in [28, 19].

To this end, we first isolate a particular subclass of CA – namely, those defined by bipermutive local rules of diameter $d$ – and remark that the Cayley tables of their global rules are Latin squares of order $s^{d-1}$, where $s$ is the size of the CA alphabet. We then narrow our attention to the case where the local rules are linear over the finite field $\mathbb{F}_{q}$, characterizing the pairs of Linear Bipermutive CA (LBCA) that produce orthogonal Latin squares. In particular, we prove that the Latin squares generated by two LBCA are orthogonal if and only if the polynomials associated to their local rules are relatively prime over $\mathbb{F}_{q}$. This is done by observing that the orthogonality of the squares is equivalent to the invertibility of the Sylvester matrix obtained from the transition matrices of the LBCA. Subsequently, we determine the number of pairs of orthogonal Latin squares generated by LBCA with rules of a fixed diameter $d$. Due to the aforementioned characterization, this actually amounts to counting the number of pairs of coprime monic polynomials with nonzero constant term and degree $n=d-1$ over $\mathbb{F}_{q}$. Although the enumeration of coprime polynomial pairs over finite fields is a well-studied problem [27, 3], to the best of our knowledge the case where both polynomials have a nonzero constant term has not been addressed before. We thus solve this counting problem through a recurrence equation, remarking that for $q=2$ the resulting integer sequence is already known in the OEIS for other combinatorial and number-theoretic facts [1]. Finally, we present a construction of MOLS based on LBCA whose rules are defined by the product of two irreducible polynomials, and we prove that the size of the MOLS resulting from this construction is maximal, meaning that they cannot be extended by adding another Latin square generated by LBCA. Further, we count how many maximal MOLS can be produced by our construction, and we prove that the corresponding lower bound is asymptotically close to the actual number of maximal MOLS which can be generated by LBCA.

The present paper is an extended version of [17], a work that was informally presented at AUTOMATA 2016. In particular, the new original contributions of this paper concern the counting results of coprime polynomials and the construction of MOLS based on irreducible polynomials.

The rest of this paper is organized as follows. Section 2 covers the basic background definitions about Latin squares and cellular automata. Section 3 focuses on the characterization of orthogonal Latin squares generated by linear bipermutive CA. Section 4 addresses the enumeration of coprime polynomials with nonzero constant term, which are in one-to-one correspondence with orthogonal Latin squares generated by LBCA. Section 5 describes a construction of MOLS based on LBCA with irreducible polynomials, proves the maximality of the resulting MOLS sizes and provides a lower bound for their number. Finally, Section 6 summarizes the contributions of this paper, and discusses some interesting avenues for future research on this topic.

2 Preliminaries on Latin Squares and Cellular Automata

In this section, we gather all the basic definitions that will be used to describe our results, referring the reader to [14] and [13] for further information about Latin squares and cellular automata, respectively.

We start by giving the formal definition of a Latin square:

Definition 1**.**

Let $X$ be a finite set of cardinality $|X|=N\in\mathbb{N}$, and let $[N]=\{1,\cdots,N\}$. A Latin square of order $N$ is a $N\times N$ square matrix $L$ with entries from $X$ such that, for all $i,j,k\in[N]$ with $k\neq j$, it holds that $L(i,j)\neq L(i,k)$ and $L(j,i)\neq L(k,i)$.

In other words, Definition 1 states that each row and each column of a Latin square is a permutation of the support set $X$. The concept of Latin square is equivalent to that of quasigroup:

Definition 2**.**

A quasigroup of order $N\in\mathbb{N}$ is a pair $\langle X,\circ\rangle$ where $X$ is a finite set of $N$ elements and $\circ$ is a binary operation over $X$ such that for all $x,y\in X$ the two equations $x\circ z=y$ and $z\circ x=y$ admit a unique solution for all $z\in X$.

Indeed, an algebraic structure $\langle X,\circ\rangle$ is a quasigroup if and only if its Cayley table is a Latin square [14]. In what follows, we will assume that the support set is always $X=[N]=\{1,\cdots,N\}$.

We now introduce the orthogonality property of Latin squares:

Definition 3**.**

Two Latin squares $L_{1}$ and $L_{2}$ of order $N$ are called orthogonal Latin squares (OLS) if

[TABLE]

for all distinct pairs of coordinates $(i_{1},j_{1}),(i_{2},j_{2})\in[N]\times[N]$.

Equivalently, $L_{1}$ and $L_{2}$ are orthogonal if their superposition yields all the ordered pairs of the Cartesian product $[N]\times[N]$. A set of $k$ Latin squares which are pairwise orthogonal is denoted as a $k$-MOLS, where the acronym stands for Mutually Orthogonal Latin Squares.

Cellular Automata (CA) are a particular kind of discrete dynamical systems defined by shift-invariant local functions. In particular, a CA is composed of a lattice of cells whose states range over a finite alphabet $A$. Each cell updates in parallel its state by applying a local rule $f:A^{\nu}\rightarrow A$ to itself and $\nu-1$ surrounding cells. One of the most common studied settings is that of one-dimensional infinite CA, where the lattice is the full shift space $A^{\mathbb{Z}}$. The Curtis-Hedlund-Theorem [11] topologically characterizes such CA in terms of global maps $F:A^{\mathbb{Z}}\rightarrow A^{\mathbb{Z}}$ that are both shift-invariant and uniformly continuous with respect to the Cantor distance.

For our work, we are interested in one-dimensional finite CA. This case leads to the problem of updating the cells at the boundaries, since they do not have enough neighbors upon which the local rule can be applied. In this paper we focus on No Boundary CA (NBCA), which we define as follows:

Definition 4**.**

Let $A$ be a finite alphabet and $n,d\in\mathbb{N}$ with $n\geq d$. The No Boundary Cellular Automaton (NBCA) $F:A^{n}\rightarrow A^{n-d+1}$ of length $n$ and diameter $d$ determined by the local rule $f:A^{d}\rightarrow A$ is the vectorial function defined for all $x\in A^{n}$ as

[TABLE]

In other words, in a NBCA of length $n$, each output coordinate with index $0\leq i\leq n-d$ is determined by evaluating the local rule $f$ of diameter $d$ on the neighborhood formed by the $i$-th input coordinate $x_{i}$ and the $d-1$ coordinates to its right, i.e. $x_{i+1},\cdots,x_{i+d}$.

The NBCA model has been investigated in [22] for the design of S-boxes. There, the authors considered the case where the alphabet is $A=\mathbb{F}_{2}$, so that a NBCA corresponds to a particular kind of vectorial Boolean function defined by shift-invariant coordinate functions. When the CA alphabet is $\mathbb{F}_{2}$ the local rule $f:\mathbb{F}_{2}^{d}\rightarrow\mathbb{F}_{2}$ can be represented by its truth table, and its decimal representation is referred to as the Wolfram code of the rule. In this paper, we will mainly consider the setting where the alphabet is the finite field $\mathbb{F}_{q}$, with $q$ being any power of a prime number. In order to avoid burdening notation we will use CA and NBCA interchangeably, since NBCA is the only model considered in the remainder of this work.

The following example grounds Definition 4 for the case of binary CA (i.e. when $A=\mathbb{F}_{2}$):

Example 1**.**

Let $A=\mathbb{F}_{2}$, and consider a CA $F:\mathbb{F}_{2}^{6}\rightarrow\mathbb{F}_{2}^{4}$ of length $n=6$ and diameter $d=3$ with local rule $f:\mathbb{F}_{2}^{3}\rightarrow\mathbb{F}_{2}$ defined as $f(x_{0},x_{1},x_{2})=x_{0}\oplus x_{1}\oplus x_{2}$. Figure 1 depicts the application of the CA global function $F$ over the vector $x=(0,1,0,1,0,0)$ and reports the truth table of the local rule $f$. The Wolfram code of $f$ is $150$, since it corresponds to the decimal encoding of the output column $(0,1,1,0,1,0,0,1)$ of the table.

This paper focuses on the class of bipermutive CA, formally defined below:

Definition 5**.**

A CA $F:A^{n}\rightarrow A^{n-d+1}$ induced by a local rule $f:A^{d}\rightarrow A$ is called left permutive (respectively, right permutive) if, for all $z\in A^{d-1}$, the restriction $f_{R,z}:A\rightarrow A$ (respectively, $f_{L,z}:A\rightarrow A$) obtained by fixing the first (respectively, the last) $d-1$ coordinates of $f$ to the values specified in $z$ is a permutation on $A$. A CA which is both left and right permutive is said to be a bipermutive CA (BCA).

Remark that when $A=\mathbb{F}_{2}=\{0,1\}$ a local rule $f:\mathbb{F}_{2}^{d}\rightarrow\mathbb{F}_{2}$ is left permutive if and only if there exists a generating function $\varphi:\mathbb{F}_{2}^{d-1}\rightarrow\mathbb{F}_{2}$ such that

[TABLE]

and symmetrically for right permutive rules. Thus, bipermutive CA over $\mathbb{F}_{2}$ are those induced by local rules of the form

[TABLE]

where $\varphi$ is a Boolean function of $d-2$ variables. Considering Example 1, one can see that rule 150 is bipermutive, since it corresponds to the case where $\varphi$ is the identity function over the second variable of the neighborhood.

Most of the results stated in this paper concern CA that, beside being bipermutive, are also linear over the finite field $\mathbb{F}_{q}$. A CA $F:\mathbb{F}_{q}^{n}\rightarrow\mathbb{F}_{q}^{n-d+1}$ of diameter $d$ is called linear if its local rule $f:\mathbb{F}_{q}^{d}\rightarrow\mathbb{F}_{q}$ is a linear combination of the cells in the neighborhood, i.e. there exist $a_{0},\cdots,a_{d-1}\in\mathbb{F}_{q}$ such that

[TABLE]

for all $x\in\mathbb{F}_{q}^{d}$, where sum and product are the field operations of $\mathbb{F}_{q}$. For $q=2$, these respectively correspond to the logical operations XOR ($\oplus$) and AND ($\land$). A linear CA can be seen as a linear transformation over $\mathbb{F}_{q}$-vector spaces described by the following $n\times(n-d+1)$ transition matrix:

[TABLE]

In particular, the CA global rule is defined as the matrix-vector multiplication $F(x)=M_{F}\cdot x^{\top}$ for all $x\in\mathbb{F}_{q}^{n}$. As remarked in [20], the matrix $M_{F}$ in Equation (6) is the generator matrix of a cyclic code. Hence, one can naturally define the polynomial $p_{f}(X)\in\mathbb{F}_{q}[X]$ associated to a linear CA $F$ as the generator polynomial of degree $n\leq d-1$ of the corresponding cyclic code:

[TABLE]

It is easy to see that a linear CA is bipermutive if and only if both $a_{0}$ and $a_{d-1}$ are not null. Indeed, the inverse functions of the right and left restrictions $f_{R,z}$ and $f_{L,z}$ can be defined for all $z\in\mathbb{F}_{q}^{d-1}$ and $y\in\mathbb{F}_{q}$ as follows:

[TABLE]

Following the notation in [21], we denote by LBCA a CA $F$ which is defined by a rule which is both linear and bipermutive. In what follows, we will consider mainly the situation where $a_{d-1}=1$, which means that the polynomial $p_{f}(X)$ associated to the LBCA is monic of degree $n=d-1$.

3 Characterization Results

In this section, we first observe that any bipermutive CA can be used to generate a Latin square. We then prove a necessary and sufficient condition which characterizes orthogonal Latin squares generated by pairs of LBCA.

3.1 Latin Squares from Bipermutive CA

We begin by showing that any BCA of diameter $d$ and length $2(d-1)$ generates a Latin square of order $N=q^{d-1}$, where $q$ is the size of the CA alphabet. To this end, we first need some additional notation and definitions.

Given an alphabet $A$ of $q$ symbols, in what follows we assume that a total order $\leq$ is defined over $A^{d-1}$ and $\phi:A^{d-1}\rightarrow[N]$ is a monotone one-to-one mapping between $A^{d-1}$ and $[N]=\{1,\cdots,q^{d-1}\}$, where $[N]$ is endowed with the usual order of natural numbers. We denote by $\psi$ the inverse mapping of $\phi$.

We now formally define the notion of square associated to a CA:

Definition 6**.**

Let $A$ be an alphabet of $q$ symbols. The square associated to the CA $F:A^{2(d-1)}\rightarrow A^{d-1}$ defined by rule $f:A^{d}\rightarrow A$ is the square matrix $\mathcal{S}_{F}$ of size $q^{d-1}\times q^{d-1}$ with entries from $[q^{d-1}]$ defined for all $1\leq i,j\leq q^{d-1}$ as

[TABLE]

where $\psi(i)||\psi(j)\in A^{2(d-1)}$ denotes the concatenation of $\psi(i),\psi(j)\in A^{d-1}$.

Hence, the square $\mathcal{S}_{F}$ is defined by encoding the first half of the CA configuration as the row coordinate $i$, the second half as the column coordinate $j$ and the output $F(\psi(i)||\psi(j))$ as the entry at $(i,j)$.

As an example, for $A=\mathbb{F}_{2}$ and diameter $d=3$, Figure 2 depicts the square $\mathcal{S}_{F}$ associated to the CA $F:\mathbb{F}_{2}^{4}\rightarrow\mathbb{F}_{2}^{2}$ defined by rule $150$. The mapping $\phi$ is defined as $\phi(00)\mapsto 1$, $\phi(10)\mapsto 2$, $\phi(01)\mapsto 3$ and $\phi(11)\mapsto 4$. Notice that in this particular case $\mathcal{S}_{F}$ is a Latin square.

We remark that this representation has been adopted in several works in the CA literature, even though under a different guise. Indeed, one can consider the square associated to a CA as the Cayley table of an algebraic structure $\langle S,\circ\rangle$, where $S$ is a set of size $|A|^{d-1}$ isomorphic to $A^{d-1}$, and $\circ$ is a binary operation over $S$. The two operands $x,y\in S$ are represented by the vectors respectively composed of the leftmost and rightmost $d-1$ input cells of the CA, while the $d-1$ output cells represent the result $z=x\circ y$. To the best of our knowledge, the first researchers who employed this algebraic characterization of cellular automata were Pedersen [26] and Eloranta [6], respectively for investigating their periodicity and partial reversibility properties. Other works in this line of research include Moore and Drisko [25], who studied the algebraic properties of the square representation of CA, and Moore [24], who considered the computational complexity of predicting CA whose local rules define solvable and nilpotent groups.

As noticed above, the square associated to the CA defined by rule 150 is actually a Latin square. We will now show that this holds in general for all bipermutive CA. To this end, we first recall a Lemma proved in [19], which states that fixing $d-1$ adjacent cells in a bipermutive CA yields a permutation between the remaining variables and the output:

Lemma 1**.**

Let $F:A^{n}\rightarrow A^{n-d+1}$ be a BCA defined by local rule $f:A^{d}\rightarrow A$. Given $\tilde{x}\in A^{d-1}$ and $i$ with $0\leq i\leq n-d+1$, let $F|_{\tilde{x},i}:A^{n-d+1}\rightarrow A^{n-d+1}$ be the restriction of $F$ obtained by fixing to $\tilde{x}$ the block of $d-1$ consecutive coordinates starting in $i$ of the BCA input vector, i.e. $x_{i}=\tilde{x}_{0}$, $x_{i+1}=\tilde{x}_{1}$, $\cdots$, $x_{i+d-2}=\tilde{x}_{d-2}$. Then, $F|_{\tilde{x},i}$ is a permutation over $A^{n-d+1}$.

On account of Lemma 1, we can prove that the squares associated to bipermutive CA are indeed Latin squares:

Lemma 2**.**

Let $A$ be an alphabet of $q$ symbols, and $d\geq 2$. Then, the square $L_{F}$ of the BCA $F:A^{2(d-1)}\rightarrow A^{d-1}$ defined by local rule $f:A^{d}\rightarrow A$ is a Latin square of order $N=q^{d-1}$.

Proof.

Let $i\in[N]$ be a row of $L_{F}$, and let $\psi(i)=(x_{0},\cdots,x_{d-2})\in A^{d-1}$ be the vector associated to $i$ with respect to the total order $\leq$ on $A^{d-1}$. Consider now the set $C=\{c\in A^{2(d-1)}:(c_{0},\cdots,c_{d-2})=\psi(i)\}$, i.e. the set of configurations of length $2(d-1)$ whose first $d-1$ coordinates coincide with $\psi(i)$, and let $F_{\psi(i),0}:A^{d-1}\rightarrow A^{d-1}$ be the restriction of $F$ determined by $\psi(i)$. By Lemma 1, the function $F_{\psi(i),0}$ is a permutation over $A^{d-1}$. So, the $i$-th row of $L_{F}$ is a permutation of $[N]$. A symmetric argument holds for any column $j$ of $L_{F}$, with $1\leq j\leq N$, which fixes the rightmost $d-1$ variables of $F$ to $\psi(j)$. Hence, every column of $L_{F}$ is also a permutation of $[N]$, and thus $L_{F}$ is a Latin square of order $q^{d-1}$.∎∎

3.2 Orthogonal Latin Squares from Linear Bipermutive CA

In the next result, we prove a characterization of orthogonal Latin squares generated by LBCA in terms of their associated polynomials:

Theorem 1**.**

Let $F,G:\mathbb{F}_{q}^{2(d-1)}\rightarrow\mathbb{F}_{q}^{d-1}$ be two LBCA of length $2(d-1)$, respectively defined by the local rules $f,g:\mathbb{F}_{q}^{d}\rightarrow\mathbb{F}_{q}$ defined as:

[TABLE]

Then, the Latin squares $L_{F}$ and $L_{G}$ of order $q^{d-1}$ generated by $F$ and $G$ are orthogonal if and only if the polynomials $p_{f}(X),p_{g}(X)\in\mathbb{F}_{q}[X]$ associated to $f$ and $g$ are relatively prime.

Proof.

Denote by $z=x||y$ the concatenation of vectors $x$ and $y$. We show that the function $\mathcal{H}:\mathbb{F}_{q}^{2(d-1)}\times\mathbb{F}_{q}^{2(d-1)}\rightarrow\mathbb{F}_{q}^{2(d-1)}\times\mathbb{F}_{q}^{2(d-1)}$, defined for all $(x,y)\in\mathbb{F}_{q}^{2(d-1)}\times\mathbb{F}_{q}^{2(d-1)}$ as

[TABLE]

is bijective if and only if the polynomials $p_{f}(X)$ and $p_{g}(X)$ associated to $F$ and $G$ are coprime. Given the transition matrices $M_{F}$ and $M_{G}$ respectively associated to $F$ and $G$, one can rewrite Equation (13) as a system of two equations:

[TABLE]

Since both $M_{F}$ and $M_{G}$ have size $(d-1)\times 2(d-1)$, Equation (14) is a linear system of $2(d-1)$ equations and $2(d-1)$ unknowns, defined by the following $2(d-1)\times 2(d-1)$ square matrix:

[TABLE]

i.e., $M$ is obtained by superposing the transition matrices $M_{F}$ and $M_{G}$. Thus $\mathcal{H}(x,y)=Mz^{\top}$ and $\mathcal{H}$ is bijective if and only if the determinant of $M$ is not null. Remark that matrix $M$ in Equation (15) is a Sylvester matrix, and its determinant is the resultant of the two polynomials $p_{f}(X)$ and $p_{g}(X)$ associated to the LBCA $F$ and $G$, respectively. It is well known (see for instance [15]) that the resultant of two polynomials is nonzero if and only if they are relatively prime. Hence, $\mathcal{H}$ is bijective (or equivalently, the Latin squares $L_{F}$ and $L_{G}$ are orthogonal) if and only if the polynomials $p_{f}(X)$ and $p_{g}(X)$ are relatively prime.∎∎

The next result immediately follows from the above theorem:

Corollary 1**.**

A family $p_{1}(X),\cdots,p_{k}(X)\in\mathbb{F}_{q}[X]$ of $k\in\mathbb{N}$ pairwise coprime polynomials of degree $n=d-1$ is equivalent to a set of $k$ MOLS of order $q^{n}$ generated by LBCA.

For alphabet $A=\mathbb{F}_{2}$ and diameter $d=3$ there exist only two linear bipermutive rules, i.e. rule $150$ and rule $90$, the latter defined as $f_{90}(x_{0},x_{1},x_{2})=x_{0}\oplus x_{2}$. As shown in Figure 3, the Latin squares of order $N=4$ defined by the LBCA $F_{150}$ and $F_{90}$ respectively induced by $f_{150}$ and $f_{90}$ are orthogonal, since the associated polynomials $p_{150}(X)=1+X+X^{2}$ and $p_{90}(X)=1+X^{2}$ are coprime over $\mathbb{F}_{2}$.

4 Counting Coprime Polynomial Pairs

By Corollary 1, one can generate a set of $k$ MOLS of order $q^{d-1}$ through LBCA of diameter $d$ by finding $k$ pairwise relatively prime polynomials of degree $n=d-1$. The problem of counting the number of pairs of relatively prime polynomials over finite fields has been investigated in several papers (see e.g. [27, 2, 3, 12]). However, notice that determining the number of pairs of linear CA inducing orthogonal Latin squares entails counting only specific pairs of polynomials, namely those whose constant term is not null. This is due to the requirement that the CA local rules must be bipermutive. To the best of our knowledge, this particular counting problem has not been considered in the literature, for which reason we address it in this section.

Formally, for $n\geq 1$ let

[TABLE]

be the set of monic polynomials in $\mathbb{F}_{q}[X]$ of degree $n$ and with nonzero constant term. For all $n\geq 1$ we have that $s_{n}=|S_{n}|=(q-1)q^{n-1}$. Moreover, we define $S_{0}=\{1\}$ (the unique monic polynomial of degree zero), and hence $s_{0}=1$.

Recall that the greatest common divisor of two polynomials $f,g\in\mathbb{F}_{q}[X]$ is the unique monic polynomial of highest degree $h$ such that

[TABLE]

for some $i,j\in\mathbb{F}_{q}[X]$. We remark that if $f,g\in S_{n}$, then $i,j\in S_{e}$ for some $0\leq e\leq n$ and $h\in S_{n-e}$.

Additionally, let us define the following subsets of $S_{n}^{2}=S_{n}\times S_{n}$:

[TABLE]

In other words, $A_{n}$ and $B_{n}$ are respectively the sets of pairs of coprime and non-coprime monic polynomials of degree $n$ with nonzero constant term. Similarly as above, let $a_{n}=|A_{n}|$ and $b_{n}=|B_{n}|$. We are interested in determining $a_{n}$, since by Theorem 1 the cardinality of $A_{n}$ corresponds to the number of orthogonal Latin squares of order $q^{n}$ generated by LBCA pairs of diameter $d=n+1$. For $n=0$, we clearly have $a_{0}=1$ and $b_{0}=0$. The following result characterizes $a_{n}$ for all $n\geq 1$:

Theorem 2**.**

Let $n\geq 1$. Then, the number of pairs of coprime monic polynomials of degree $n$ with nonzero constant term is

[TABLE]

Proof.

Let us first settle the case $n=1$, for which we have

[TABLE]

It is clear that $\gcd(f,g)=1$ for any $f,g\in S_{1}$ if and only if $f\neq g$. Thus, it follows that

[TABLE]

which proves Equation (17) for the case $n=1$. For $n>1$, remark first that

[TABLE]

Moreover, any pair $(f,g)$ with $\deg(\gcd(f,g))=n-e$ ($0\leq e\leq n-1$) can be uniquely expressed as a pair $(h,(i,j))$, where $h\in S_{n-e}$ and $(i,j)\in S_{e}$. Hence,

[TABLE]

Combining Equations (18) and (19) we have

[TABLE]

Multiplying both sides of (21) by $q$ we obtain

[TABLE]

Remark that $qs_{n-1-e}=q(q-1)q^{n-2-e}=(q-1)q^{n-1-e}=s_{n-e}$. Hence, Equation (22) can be rewritten as

[TABLE]

By subtracting Equations (20) and (23) we thus have

[TABLE]

Since $s_{n}^{2}=(q-1)^{2}q^{2n-2}$, while $qs_{n-1}^{2}=(q-1)^{2}q^{2n-3}$ and $s_{1}=(q-1)$, Equation (24) becomes

[TABLE]

from which it follows that

[TABLE]

By iterating the above procedure to determine the term $a_{n-1}$ from the difference $a_{n-1}-qa_{n-2}$, the term $a_{n-2}$ from $a_{n-2}-qa_{n-3}$, and so on until $a_{2}$ from $a_{2}-qa_{1}$, one has

[TABLE]

from which we finally obtain the result.∎∎

Remark 1**.**

Notice that in Theorem 2 we count all ordered coprime polynomial pairs. To get the number of distinct pairs, one simply has to divide Equation (17) by $2$, thus obtaining $\tilde{a}_{n}=\frac{1}{2}a_{n}$. In particular, for $q=2$ the formula for $\tilde{a}_{n}$ becomes

[TABLE]

The first terms of this sequence for $n\geq 1$ are:

[TABLE]

which is a shifted version of OEIS sequence A002450 [1], defined by

[TABLE]

It is easily seen that $c_{n}=\tilde{a}_{n+1}$, i.e. $c_{n}$ corresponds to the number of distinct coprime pairs of polynomials of degree $n+1$ over $\mathbb{F}_{2}$ where both polynomials have nonzero constant term. We remark that sequence A002450 is known for several other combinatorial facts not related to polynomials or orthogonal Latin squares arising from LBCA, for which we refer the reader to [1].

5 A Construction of MOLS based on LBCA

In this section, we tackle the question of determining the maximum number of MOLS generated by linear bipermutive CA over $\mathbb{F}_{q}$ of a given order. Given $n\in\mathbb{N}$ we consider in particular the following two problems:

Problem 1**.**

What is the maximum number $N_{n}$ of LBCA over $\mathbb{F}_{q}$ of diameter $n+1$ whose Latin squares are mutually orthogonal? From Section 3, this actually amounts to compute the maximum number of monic pairwise coprime polynomials of degree $n$ and nonzero constant term over $\mathbb{F}_{q}$.

Problem 2**.**

What is the number $T_{n}$ of maximal sets of $N_{n}$ MOLS generated by LBCA?

In the remainder of this section we present a construction for sets of MOLS based on LBCA defined by pairwise coprime polynomials over $\mathbb{F}_{q}$. Moreover, we solve Problem 1 by proving that the size of MOLS resulting from this construction is maximal. We also determine the number $D_{n}$ of MOLS that can be generated through this construction, and show that it is asymptotically close to $T_{n}$.

Recall from Section 4 that $S_{n}$ denotes the set of all degree $n$ monic polynomials $f\in\mathbb{F}_{q}[X]$ with nonzero constant term $a_{0}$. Additionally, let

[TABLE]

In other words, $\mathcal{M}_{n}$ is the family of subsets of $S_{n}$ of pairwise coprime polynomials. In order to solve Problem 1, we have to determine the maximal cardinality of the subsets in $\mathcal{M}_{n}$, that is

[TABLE]

On the other hand, for Problem 2 we want to count how many sets in $\mathcal{M}_{n}$ have cardinality $N_{n}$:

[TABLE]

We begin by considering the set $\mathcal{I}_{n}$ of irreducible polynomials of degree $n$ over $\mathbb{F}_{q}$ with nonzero constant term, all of which are trivially pairwise coprime. Hence, $\mathcal{I}_{n}$ is included in all subsets having maximum cardinality $N_{n}$. Denoting by $I_{n}$ the cardinality of $\mathcal{I}_{n}$, one has that $I_{0}=1$ and $I_{1}=q-1$, while for $n\geq 2$ $I_{n}$ is given by Gauss’s formula [7]:

[TABLE]

where $d$ ranges over all positive divisors of $n$ (including $1$ and $n$), while $\mu$ denotes the Möbius function. Let $d=\varrho_{1}^{\alpha_{1}}\varrho_{2}^{\alpha_{2}}\cdots\varrho_{k}^{\alpha_{k}}$ be the prime factorization of $d\in\mathbb{N}$. Then, $d$ is called square-free (s.f.) if $\alpha_{i}=1$ for all $i\in\{1,\cdots,k\}$, i.e. if $d$ is not divisible by any prime power with exponent higher than $1$. The Möbius function of $d$ is defined as:

[TABLE]

We thus have that

[TABLE]

In order to refine this lower bound, we have to determine how many other (reducible) polynomials of degree $n$ one can add to $\mathcal{I}_{n}$ so that the resulting set only includes pairwise coprime polynomials. To this end, we first need a side result which shows that the sequence of the numbers of monic irreducible polynomials is non-decreasing in the degree $n$. As a preliminary remark, observe that any polynomial $f$ in $S_{n}$ is either irreducible and hence belongs to $\mathcal{I}_{n}$, or all its irreducible factors belong to $\mathcal{J}_{n}=\bigcup_{k=1}^{\lfloor\frac{n}{2}\rfloor}\mathcal{I}_{k}$.

Lemma 3**.**

For all $q\geq 2$ powers of a prime number and $n\geq 1$, $I_{n}$ is a non-decreasing function of $n$.

Proof.

We want to show that $I_{n}\geq I_{n-1}$ for all $n\geq 2$. Note that $I_{1}=q-1$ since we do not consider the polynomial $X$ (its constant term being null), while $I_{n}$ is given by Gauss’s formula for $n\geq 2$.

The claim is easily proved for $n\leq 4$, since

[TABLE]

We now assume $n\geq 5$. We first prove that $I_{n}\geq\frac{1}{n}(q^{n}-q^{n-2})$. It is easily checked for $n=5$, since $I_{5}=\frac{1}{5}(q^{5}-q)$; for $n\geq 6$, consider the sum

[TABLE]

Remark that, for all $d|n$ with $d\neq 1$, the term $q^{\frac{n}{d}}$ in the sum of Gauss’s formula occurs in the sum of Equation (37), i.e. for $i=\frac{n}{d}$. Since in Gauss’s formula one always adds or subtracts the term $q^{\frac{n}{d}}$ depending on the value of $\mu(d)$, by Equations (34) and (37) we have the following inequality:

[TABLE]

from which it follows that

[TABLE]

We now prove that $I_{n-1}\leq\frac{1}{n-1}q^{n-1}$. Similarly to the previous inequality, let us denote by $p$ the smallest prime divisor of $n-1$. Then, in the sum of Gauss’s formula for $I_{n-1}$ we subtract $q^{\frac{n-1}{p}}$, since $\mu(p)=-1$. Consider now the sum

[TABLE]

Again, each term $q^{\frac{n-1}{d}}$ in Gauss’s formula for $I_{n-1}$ occurs in (40) for $i=\frac{n-1}{d}$. Thus, the following inequality holds:

[TABLE]

Therefore,

[TABLE]

Combining the lower bound in (39) and the upper bound in (42), we obtain

[TABLE]

Thus, since $q\geq 2$ and $n>5$, it follows that

[TABLE]

∎∎

Consider now the following construction for a family of pairwise coprime polynomials parameterized on the degree $n$:

Construction-Irreducible$(n)$

Initialization:

Initialize set $\mathcal{P}_{n}$ to $\mathcal{I}_{n}$

Loop:

For all $1\leq k\leq\left\lfloor\frac{n}{2}\right\rfloor$ do:

1. 1.

Build set $\mathcal{P}^{\prime}_{k}$ by multiplying each polynomial in $\mathcal{I}_{k}$ with a distinct polynomial in $\mathcal{I}_{n-k}$ 2. 2.

Add set $\mathcal{P}^{\prime}_{k}$ to $\mathcal{P}_{n}$

Output:

return $\mathcal{P}_{n}$

Hence, set $\mathcal{P}_{n}$ is constructed by first adding all irreducible polynomials of degree $n$, then by adding the set of all irreducible polynomials of degree $1$ multiplied by as many irreducible polynomials of degree $n-1$, and so on. In particular, notice that step 1 in the loop of Construction-Irreducible is possible since by Lemma 3 one has that $I_{n-k}\geq I_{k}$ for all $k\leq\left\lfloor\frac{n}{2}\right\rfloor$. Further, all polynomials added to $\mathcal{P}_{n}$ through Construction-Irreducible are pairwise coprime, since they all have distinct irreducible factors.

Remark that the procedure Construction-Irreducible can be iterated only up to $k\leq\left\lfloor\frac{n}{2}\right\rfloor$, because by symmetry the irreducible polynomials of degree $n-k$ with $k>\left\lfloor\frac{n}{2}\right\rfloor$ correspond to those of degree $k\leq\left\lfloor\frac{n}{2}\right\rfloor$. Notice also that, when $n$ is even, the last step of the procedure consists of squaring all irreducible polynomials of degree $\frac{n}{2}$.

Hence, we have shown that the set $\mathcal{P}_{n}$ which is generated by procedure Construction-Irreducible is indeed a member of the family $\mathcal{M}_{n}$. The cardinality of such set is given by

[TABLE]

In fact, beside the initial step when one adds all irreducible polynomials of degree $n$ to $\mathcal{P}_{n}$, in each iteration $k$ of the loop the number of polynomials that one can obtain by multiplying two irreducible factors is bounded by the number of irreducible polynomials of degree $k$, which is $I_{k}$. We have thus obtained the following result, which gives a more precise lower bound on $N_{n}$:

[TABLE]

A natural question arising from Inequality (47) is whether the above construction is optimal, i.e. if the maximum number of pairwise coprime polynomials $N_{n}$ is actually equal to $C_{n}$. In the next theorem, we prove that the MOLS produced by Construction-Irreducible are indeed maximal, and we characterize the families of $\mathcal{T}_{n}$.

Theorem 3**.**

For any $n$ and $q$, the maximum number of MOLS generated by LBCA of diameter $d=n+1$, or equivalently the maximum number of pairwise coprime monic polynomials of degree $n$ and nonzero constant term is:

[TABLE]

Moreover, let $A\subseteq S_{n}$. Then, $A\in\mathcal{T}_{n}$ if and only if the following hold:

$A$* contains $\mathcal{I}_{n}$;* 2. 2.

if $n$ is even then $A$ contains $\{g^{2}:g\in\mathcal{I}_{n/2}\}$; 3. 3.

for every $g\in\mathcal{I}_{k}$ with $k<n/2$, there exists a unique $f\in A$ such that $g|f$. This $f$ is either of the form $f=g^{a}$ with $a=n/k$, or of the form $f=g^{b}h$ where $bk<n/2$ and $h\in\mathcal{I}_{n-bk}$, in which case $h$ does not divide any other $f^{\prime}\in A$.

Proof.

We first determine the value of $N_{n}$. By inequality (47) we have that $N_{n}\geq C_{n}$. Conversely, let $A\in\mathcal{M}_{n}$ be a maximum collection of mutually coprime polynomials in $S_{n}$, i.e. with cardinality $N_{n}$. Clearly, $A$ must contain all irreducible polynomials $\mathcal{I}_{n}$, so let $B=A\setminus\mathcal{I}_{n}$ be the set of reducible polynomials in $A$. For any $f\in B$, denote the irreducible polynomial of lowest degree in the factorization of $f$ as $T_{f}$ (if there are several, choose the first one in lexicographic order). Note that $T_{f}$ has degree at most $n/2$. Now if $f,g\in B$ satisfy $T_{f}=T_{g}$, then $f$ and $g$ are not coprime, therefore the map $f\mapsto T_{f}$ is an injection from $B$ to $\mathcal{J}_{n}$. Thus $|A|\leq C_{n}$ and $N_{n}=C_{n}$.

We now characterize the families of cardinality $N_{n}$. We begin by showing that any such family must satisfy the three properties of the theorem. Firstly, as seen above, such a family $A$ must contain $\mathcal{I}_{n}$, so let us focus on $B=A\setminus\mathcal{I}_{n}$. This time, the mapping $f\mapsto T_{f}$ is a bijection from $B$ to $\mathcal{J}_{n}$, hence let $g\mapsto F_{g}$ be its inverse. Secondly, if $g\in\mathcal{I}_{n/2}$, then $F_{g}=gh$ for some $h\in\mathcal{I}_{n/2}$. If $h\neq g$, then $F_{h}\neq F_{g}$ but $\gcd(F_{g},F_{h})=h$, which violates coprimality; thus $h=g$ and $F_{g}=g^{2}$. Thirdly, if $g\in\mathcal{I}_{k}$, then either $F_{g}=g^{a}$ for $a=n/k$ or $F_{g}=g^{b}H_{g}$ for $bk<n$ and $\gcd(H_{g},g)=1$. If $H_{g}$ is reducible, then its factor of lowest degree $h\in\mathcal{J}_{n}$ is a common divisor of $F_{h}$ and $F_{g}$, which again violates coprimality. Thus $H_{g}\in\mathcal{I}_{n-bk}$ for $bk<n/2$. Finally, if $H_{g}=H_{g^{\prime}}$ for another $g^{\prime}\in\mathcal{J}_{n}$, then again coprimality is violated. Conversely, it is easily checked that any family satisfying all three properties is a family of $N_{n}$ coprime polynomials in $S_{n}$. ∎∎

We now determine how many maximal sets of pairwise coprime polynomials of degree $n$ one can obtain through Construction-Irreducible, thus providing a lower bound on $T_{n}$. In particular, this corresponds to the case of families $A\in\mathcal{T}_{n}$ where the polynomial $f\in A$ in the third condition of Theorem 3 is of the form $f=gh$ (i.e. $b=1$) and $h\in\mathcal{I}_{n-bk}$. Moreover, we show that this lower bound is asymptotically close to the actual value of $T_{n}$. Before proving this result, we first need the following asymptotic estimate of $I_{n}$:

Lemma 4**.**

Let $I_{n}$ be defined as in Equation (34). Then, as $n$ tends to infinity,

[TABLE]

Proof.

Let us rewrite Equation (34) by extracting the terms $d=1$ and $d=p$ from the sum, where $p$ is the smallest prime divisor of $n$. Since $\mu(1)=1$ and $\mu(p)=-1$, we have

[TABLE]

The smallest divisor of $n$ which is strictly greater than $p$ can be at most $p+1$. Thus, each term in the sum of Equation (50) is limited in absolute value by $q^{\frac{n}{p+1}}$. In particular, as $d$ grows the value $q^{\frac{n}{d}}$ decreases, hence we can bound the sum in (50) with the geometric series $\sum_{i=0}^{\infty}q^{\frac{n}{p+1}-i}$:

[TABLE]

Since $q\geq 2$, we have that $\sum_{i=0}^{\infty}q^{-i}\leq 2$. Thus, we obtain

[TABLE]

Consider now the difference $q^{\frac{n}{p}}-2\cdot q^{\frac{n}{p+1}}$:

[TABLE]

Clearly, $q^{-\frac{n}{p(p+1)}}\rightarrow 0$ for $n\rightarrow\infty$. Hence, we have that $q^{\frac{n}{p}}-2\cdot q^{\frac{n}{p+1}}=\mathcal{O}\left(q^{\frac{n}{p}}\right)$, and by Inequality (52) it follows that

[TABLE]

Therefore, Equation (34) can be rewritten as

[TABLE]

where the rightmost equality follows from the fact that $p\geq 2$ for all $n\in\mathbb{N}$. ∎∎

We can now prove our lower bound on $T_{n}$. In what follows, we denote by $D_{n}$ the number of maximal sets produced by Construction-Irreducible.

Theorem 4**.**

For all $n$, it holds that

[TABLE]

Moreover, as $n$ tends to infinity, we have

[TABLE]

Proof.

Let us first prove the formula for $D_{n}$. For all $1\leq k\leq\lfloor\frac{n}{2}\rfloor$, the set $\mathcal{P}^{\prime}_{k}$ in step 1 of the loop of Construction-Irreducible is obtained by first taking an irreducible polynomial $f_{1}\in\mathcal{I}_{k}$ and multiplying it by an irreducible polynomial $g_{1}\in\mathcal{I}_{n-k}$. Hence, the choices for $g_{1}$ are $I_{n-k}$. Then, one takes another irreducible polynomial $f_{2}\in\mathcal{I}_{k}$ and multiplies it by an irreducible polynomial $g_{2}\in\mathcal{I}_{n-k}$, with $g_{2}\neq g_{1}$. Thus, the possible choices for $g_{2}$ are $I_{n-k}-1$. Since the choices for the polynomials in $\mathcal{I}_{n-k}$ are independent, and since we have to select $I_{k}$ of them, we have that the number of choices for constructing $\mathcal{P}_{k}^{\prime}$ is

[TABLE]

Further, since for $1\leq k_{1},k_{2}\leq\lfloor\frac{n}{2}\rfloor$ with $k_{1}\neq k_{2}$ the choices for constructing $\mathcal{P}_{k_{1}}^{\prime}$ and $\mathcal{P}_{k_{2}}^{\prime}$ are independent, we obtain that

[TABLE]

We now prove that $\log_{q}D_{n}=\Theta(q^{\frac{n}{2}})$, starting with some estimates for $\log_{q}I_{n-k}$ and $\log_{q}(I_{n-k}-I_{k})$. As a first remark, observe that Equation (49) in Lemma 4 shows that

[TABLE]

It is then easy to prove that for $n$ large enough and $k<n/2$, one has

[TABLE]

for $\delta<1$, e.g. $\delta=\frac{1}{q-1/2}$. Combining Inequality (59) with Equation (49) we obtain

[TABLE]

and hence

[TABLE]

Equation (61) thus yields the following estimate for $\log_{q}(I_{n-k}-I_{k})$:

[TABLE]

Consider now $\log_{q}E_{k}$. By Equation (56), it is easy to see that

[TABLE]

Since by (58) we have that $\log_{q}I_{n-k}\leq n-k$, while by (62) it holds that $\log_{q}E_{k}\geq I_{k}(n-k-\mathcal{O}(\log(n-k)))$, the inequalities in (63) can be rewritten as follows:

[TABLE]

Consequently, we obtain the following estimate for $\log_{q}E_{k}$:

[TABLE]

Denoting $\sigma=\sum_{i=0}^{\infty}q^{-i}$, we have

[TABLE]

Therefore,

[TABLE]

We finally prove that $\log_{q}T_{n}=\log_{q}D_{n}+\mathcal{O}(q^{\frac{n}{3}})$. By Theorem 3, in any maximal family of polynomials in $\mathcal{T}_{n}$, and any irreducible $g\in\mathcal{I}_{k}$ for $n/3<k<n/2$, the corresponding $f$ must be $f=gh$ for some $h\in\mathcal{I}_{n-k}$, thus there are $E_{k}$ choices for the polynomials in the family that have an irreducible factor of degree $k$. If $k\leq n/3$, then for any $g\in\mathcal{I}_{k}$ there are at most $1+\sum_{d=\left\lfloor\frac{n}{2}\right\rfloor+1}^{n-k}I_{d}$ choices for the corresponding polynomial $f$ in the family. Altogether, we obtain

[TABLE]

Define now $B_{k}$ as

[TABLE]

Again, this yields $\sum_{k=1}^{\left\lfloor\frac{n}{3}\right\rfloor}\log_{q}B_{k}=\mathcal{O}(q^{\frac{n}{3}})$, which in turn gives us

[TABLE]

∎∎

6 Conclusions and Perspectives

In this paper we undertook an investigation of mutually orthogonal Latin squares generated through linear bipermutive CA. First, we proved that any bipermutive CA of diameter $d$ and length $2(d-1)$ can be used to generate a Latin square of order $N=q^{d-1}$, with $q$ being the size of the CA alphabet. We then focused on orthogonal Latin squares generated by LBCA, showing a characterization result based on the Sylvester matrix induced by two linear local rules. In particular, we proved that two LBCA generate orthogonal Latin squares if and only if the polynomials associated to their local rules are relatively prime. In the second part of the paper, we determined the number of LBCA pairs over $\mathbb{F}_{q}$ generating orthogonal Latin squares, i.e. the number of coprime polynomial pairs $(f,g)$ of degree $n$ over $\mathbb{F}_{q}$ where both $f$ and $g$ have nonzero constant term. In particular, we remarked that the integer sequence generated by the closed-form formula of the recurrence equation for $q=2$ corresponds to A002450, a sequence which is already known in the OEIS for several other facts not related to polynomials or orthogonal Latin squares. In the last part of the paper, we presented a construction of MOLS generated by LBCA based on irreducible polynomials, and we proved that the resulting families are maximal. Finally, we also derived a lower bound on the number of maximal MOLS induced by the proposed construction, and showed that this bound is asymptotically close to the actual number of maximal MOLS produced by LBCA.

There are several opportunities for further improvements on the results presented in this paper.

A first direction for future research is to generalize the study to MOLS generated by nonlinear bipermutive CA. In this case, one obviously cannot rely on the characterization result of Theorem 1, since this crucially depends on the use of the Sylvester matrix defined from the transition matrices of linear CA. Preliminary work led by some of the authors of the present paper showed that a necessary condition for a pair of BCA (either linear or nonlinear) to generate orthogonal Latin squares is that their local rules must be pairwise balanced, meaning that each of the four pair of bits must occur equally often in the juxtaposition of their truth tables [18]. We believe it is still possible to use the theory of resultants to characterize orthogonal Latin squares generated by nonlinear BCA. As a matter of fact, the main difference between linear and nonlinear pairs is that in the former case the system of equations (14) concerns univariate polynomials. On the other hand, in the nonlinear case one can associate multivariate polynomials to the local rules, and then use the tools of elimination theory (to which the concept of resultant belongs) to study the invertibility of the resulting systems.

A second extension worth exploring, especially concerning the possible applications related to secret sharing, is to investigate the structure of the inverse of a Sylvester matrix. As described in [17], a family of $k$ MOLS generated by LBCA can be used to design a $(2,k)$-threshold secret where the dealing phase corresponds to evaluating the global rules of the $k$ LBCA to an initial configuration whose left half is the secret, while the right half is randomly chosen. The outputs of the LBCA will be the shares distributed to the $k$ players. In order to reconstruct the secret, any two out of $k$ players must invert the Sylvester matrix associated to their CA (which are assumed to be public) and then multiply it by the vector obtained by concatenating their shares. Hence, an interesting question is whether the reconstruction phase can be carried out again by CA computation, which means that the inverse of the Sylvester matrix related to two LBCA must be of Sylvester type as well. This question has been answered in negative during the Fifth International Students’ Olympiad in Cryptography – NSUCRYPTO [10] for LBCA over the finite field $\mathbb{F}_{2}$. In particular, it has been proved that the only Sylvester matrix over $\mathbb{F}_{2}$ satisfying this condition is the one defined by the polynomials $X^{n}$ and $1+X^{n}$, which does not correspond to a pair of LBCA since $X^{n}$ has null constant term. However, the existence question for Sylvester matrices whose inverses are of Sylvester type remains open for larger finite fields.

Finally, another interesting idea would be to extend our investigation to Mutually Orthogonal Latin Hypercubes generated by CA, i.e. the generalization of MOLS to higher dimensions. This would be equivalent to study the conditions under which CA can be used to construct orthogonal arrays with strength higher than $2$. A characterization result for such kind of orthogonal arrays would allow one to design a general $(t,n)$-threshold secret sharing scheme based on CA, or equivalently to design linear MDS codes through CA. A possible idea to achieve this result would be to first characterize which subclasses of bipermutive CA generate Latin hypercubes. From there, the next step would be to characterize sets of linear CA inducing Orthogonal Latin Hypercubes, which are equivalent to orthogonal arrays [14]. However, we note that there are no straightforward ways to generalize the concept of resultant to more than two polynomials [8]. As a matter of fact, some of the existing generalizations involve matrices which do not correspond to those related to hypercubes generated by CA. To the best of our knowledge, the only resultant matrix for several polynomials that most resemble the CA hypercube case has been defined in [5], which could thus represent a starting point for future work on the subject.

Acknowledgements

The authors wish to thank Arthur Benjamin, Curtis Bennett and Igor Shparlinski for insightful comments on how to count the number of pairs of coprime polynomials with nonzero constant term.