A note on the Alon-Saks-Seymour problem


Abstract

Let \(f(k)\) be the maximum possible chromatic number of a graph whose edge set can be partitioned into at most \(k\) complete bipartite graphs. Alon, Saks, and Seymour conjectured that \(f(k)=k+1\) for all \(k\). While the conjecture was verified for \(k \leq 9\) by Gao et al., it was disproved by Huang and Sudakov, and further Balodis et al. proved that \(f(k) \geq 2^{\widetilde{\Omega}((\log k)^2)}\).

In this note, we give a simple proof of the recursive upper bound \(f(k+1) \leq f(k)+f(\lfloor k/4 \rfloor)\). Consequently, \(f(k) \leq 2^{(\log_2 (4k))^2/4}\) for \(k \geq 1\). This improves the previous best known upper bound of Mubayi and Vishwanathan in the exponent by a factor which is asymptotically two. Note that these bounds are sharp up to a lower order factor in the exponent by the result of Balodis et al.

1 Introduction↩︎

For a positive integer \(k\), let \(f(k)\) be the maximum possible chromatic number of a graph whose edge set can be partitioned into at most \(k\) complete bipartite graphs. It is also convenient to set \(f(0)=1\). As a generalization of the Graham-Pollak theorem, Alon, Saks, and Seymour conjectured that \(f(k)=k+1\) for all \(k\). Gao, McKay, Naserasr, and Stevens [1] verified the Alon-Saks-Seymour conjecture for \(k \leq 9\).

If graphs \(G,G_1,G_2\) satisfy \(E(G) = E(G_1) \cup E(G_2)\), then \(\chi(G) \leq \chi(G_1)\chi(G_2)\). It easily follows that \(f(k+1) \leq 2f(k)\) and inductively \(f(k) \leq 2^k\). This was improved by Mubayi and Vishwanathan [2], who proved that \(f(k)\) is asymptotically at most \(2^{((\log k)^2+\log k)/2}\) (all logarithms in this paper are base \(2\)).

The Alon-Saks-Seymour conjecture was disproved by Huang and Sudakov [3]. They showed that \(f(k) \geq ck^{6/5}\) for an appropriate constant \(c\). The first super-polynomial lower bound on \(f(k)\) was proved by Göös [4]. This was subsequently improved by Balodis, Ben-David, Göös, Jain, and Kothari [5] to \(f(k) \geq 2^{\widetilde{\Omega}((\log k)^2)}\), which is sharp up to the lower order factor in the exponent. The Alon-Saks-Seymour problem has found interesting connections to communication complexity and learning theory (see [3], [5]).

Here we prove the following recursive upper bound. It improves the bound of Mubayi and Vishwanathan in the exponent by a factor \(2\).

Theorem 1. We have \(f(k+1) \leq f(k)+f(\lfloor k/4 \rfloor)\). Consequently, \(f(k) \leq 2^{(\log_2 (4k))^2/4}\) for \(k \geq 1\).

The proof of the latter bound in Theorem 1 follows from the former recursive bound by strong induction on \(k\), with the base cases \(1 \leq k \leq 3\) being easy to check. For \(k \geq 4\), by applying the recursive inequality for \(k-1,\ldots,\lfloor k/4 \rfloor\) and using monotonicity of \(f\), we get \(f(k) \leq (\lceil 3k/4 \rceil +1)f(\lfloor k/4 \rfloor) \leq kf(\lfloor k/4 \rfloor)\). Then, for \(k \geq 4\) we get by the induction hypothesis that \[f(k) \leq kf(\lfloor k/4 \rfloor) \leq k2^{(\log k)^2/4} = 2^{\log k + (\log k)^2/4} \leq 2^{(\log(4k))^2 /4}.\] For brevity, we omit the proof that the recursive bound implies \(f(k) \leq 2^{\frac{1}{4}(\log k)^2 - \frac{1}{2}(\log k)\log \log k +O(\log k)}\).

We are left to prove the first part of Theorem 1. As a warmup, we first give a very simple proof of a weaker recursive bound which can recover the Mubayi-Vishwanathan bound.

Proposition 1. We have \(f(k+1) \leq f(k)+f(\lfloor k/2 \rfloor)\).

Proof. Let \(G=(V,E)\) be a graph whose edge set can be partitioned into \(k+1\) complete bipartite graphs. Let \(A,B\) denote the parts of one of these complete bipartite graphs. Every other complete bipartite graph in the edge-partition cannot have an edge internal to \(A\) and an edge internal to \(B\) since otherwise it would also contain an edge in \(A \times B\). Hence, \(G[A]\) or \(G[B]\) (say \(G[A]\)) can be edge-partitioned into at most \(\lfloor k/2 \rfloor\) complete bipartite graphs. Thus, the chromatic number of the induced subgraph on \(A\) satisfies \(\chi(G[A]) \leq f(\lfloor k/2 \rfloor)\). Also, \(G[V \setminus A]\) can be edge-partitioned into at most \(k\) complete bipartite graphs, so \(\chi(G[V \setminus A]) \leq f(k)\). Finally, \(\chi(G) \leq \chi(G[A])+\chi(G[V \setminus A]) \leq f(k)+f(\lfloor k/2 \rfloor)\). ◻

Proof of Theorem 1. Let \(G=(V,E)\) be a graph whose edge set can be partitioned into \(k+1\) complete bipartite graphs. Let \(H_i\) denote the \(i^{\textrm{th}}\) complete bipartite graph of this edge partition and \(A_i,B_i\) denote its two parts. If \(H_i\) has an edge with vertices in a part of \(H_j\), then \(H_j\) cannot have an edge with vertices in a part of \(H_i\). Indeed, if \((a_i,b_i) \in A_i \times B_i\) is an edge of \(H_i\) in a part of \(H_j\) (say \(A_j\)), and \((a_j,b_j) \in A_j \times B_j\) is an edge of \(H_j\) in a part of \(H_i\) (say \(A_i\)), then \((b_j,b_i) \in A_i \times B_i\) and \((b_i,b_j) \in A_j \times B_j\), so \((b_i,b_j)\) is an edge of both \(H_i\) and \(H_j\), contradicting that these complete bipartite graphs are edge-disjoint.

Construct the auxiliary directed graph \(D\) with vertex set \([k+1]\) with an edge \(i \rightarrow j\) if the vertices of an edge of \(H_i\) are in a part of \(H_j\). By the above discussion, \(D\) is an oriented graph, and so it has a vertex \(j\) of indegree at most \(k/2\). As in the proof of Proposition 1, for each \(i\), \(H_i\) can have an edge in at most one of the two parts of \(H_j\). Hence, \(G[A_j]\) or \(G[B_j]\) (say \(G[A_j]\)) can be edge-partitioned into at most \(\lfloor k/4 \rfloor\) complete bipartite graphs. It follows that \(\chi(G) \leq \chi(G[V \setminus A_j])+\chi(G[A_j]) \leq f(k)+f(\lfloor k/4 \rfloor)\). ◻

References↩︎

[1]
Z. Gao, B. D. McKay, R. Naserasr, and B. Stevens, Bipartite edge partitions and the former Alon-Saks-Seymour conjecture, Australas. J. Combin.66(2016), 211–228.
[2]
D. Mubayi and S. Vishwanathan, Bipartite coverings and the chromatic number, Electron. J. Combin.16(2009), Note 34, 5 pp.
[3]
H. Huang and B. Sudakov, A counterexample to the Alon-Saks-Seymour conjecture and related problems, Combinatorica32(2012), 205–219.
[4]
M. Göös, Lower bounds for clique vs. independent set, in FOCS 2015, pp. 1066–1076.
[5]
K. Balodis, S. Ben-David, M. Göös, S. Jain, and R. Kothari, Unambiguous DNFs and Alon-Saks-Seymour, SIAM J. Comput., Special Section FOCS 2021 (2023), FOCS21-157–FOCS21-173.

  1. Department of Mathematics, Stanford University, Stanford, CA 94305. Email: jacobfox@stanford.edu. Research supported by NSF awards DMS-2452737 and DMS-2154129.↩︎