arXiv++ Combinatorics

Inapproximability results for $\mathsf{Max\,k\,CSP\!-\!q}$ have been traditionally established using balanced $t$-wise independent distributions, which are closely related to orthogonal arrays, a famous family of combinatorial designs. In this work, we investigate the role of these combinatorial structures in the context of the differential approximability of $\mathsf{k\,CSP\!-\!q}$, providing new structural insights and approximation bounds. We first establish a direct connection between the average differential ratio on $\mathsf{k\,CSP\!-\!q}$ instances and orthogonal arrays. This allows us to derive the new differential approximability bounds of $1/q^k$ for $(k +1)$-partite instances, $Ω(1/n^{\lfloor k/2\rfloor})$ for Boolean instances, $Ω(1/n)$ when $k =2$, and $Ω(1/n^{k -\lceil\log_{Θ(q)}k\rceil})$ when $k, q\geq 3$. We then introduce families of array pairs, called {\em alphabet reduction pairs of arrays}, that are still related to balanced $k$-wise independence. Using these pairs of arrays, we establish a reduction from $\mathsf{k\,CSP\!-\!q}$ to $\mathsf{k\,CSP\!-\!k}$ (where $q >k$), with an expansion factor of $1/(q -k/2)^k$ on the differential approximation guarantee. Combining this with a 1998 result by Yuri Nesterov, we conclude that $\mathsf{2\,CSP\!-\!q}$ is approximable within a differential factor of $0.429/(q -1)^2$. Finally, using similar Boolean array pairs, {\em called cover pairs of arrays}, we prove that every Hamming ball of radius $k$ provides a $Ω(1/n^k)$-approximation of the instance diameter. Thus, our work highlights the relevance of combinatorial designs for establishing structural differential approximation guarantees for CSPs.

In [1], we introduced a family of combinatorial designs, which we call "alphabet reduction pairs of arrays", ARPAs for short. These designs depend on three integer parameters $q, p \leq q, k\leq p$: $q$ is the size of the symbol set $\{0, 1 ,\ldots, q -1\}$ in which the coefficients of the arrays take their values; $p$ is the maximum number of distinct symbols allowed in a row of the second array of the pair; $k$ is the larger integer for which the two arrays of the pair coincide -- up to the order of their rows -- on any $k$-ary subset of their columns. The first array must contain at least one occurrence of the word $0\ 1\ \ldots\ q -1$. Intuitively, the idea is to cover "as many as possible" occurrences of this word of $q$ symbols with "as few as possible" words of at most $p$ different symbols. These designs are related to the approximability of "Constraint Satisfaction Problems with bounded constraint arity", known as $k\,$CSPs. In this context, we are particularly interested in ARPAs in which the frequency of the word $0\ 1\ \ldots\ q -1$ is maximal. We introduce a seemingly simpler family of combinatorial designs as "Cover pairs of arrays" (CPAs). The arrays of a CPA take Boolean coefficients, and must still coincide on any $k$-ary subset of their columns. The purpose is, as it were, to cover "as many as possible" occurrences of the word of $q$ ones using "as few as possible" $q$-length Boolean words of weight at most $p$. We show that, when it comes to maximizing the frequency of the words $0\ 1\ \ldots\ q -1$ in ARPAs and $1\ 1\ \ldots\ 1$ in CPAs, ARPAs and CPAs are equivalent. We prove the optimality of the ARPAs given in [1] for the case $p =k$. In addition, we provide optimal ARPAs for the cases $k =1$ and $k =2$. We emphasize the fact that both families of combinatorial designs are related to the approximability of $k\,$CSPs.