Georgia Institute of Technology College of Sciences

The goal of this talk is to show recent advances regarding two important mathematical problems. The first one can be straightforwardly formulated in a graph theory language, but can be possibly applied in other fields. The second one was motivated by machine learning applications, but leads to graph theory techniques. The celebrated open conjecture of Erdos and Hajnal from 1989 states that families of graphs not having some given graph H as an induced subgraph contain polynomial-size cliques/stable sets (in the undirected setting) or transitive subsets (in the directed setting). Recent techniques developed over last few years provided the proof of the conjecture for new infinite classes of graphs (in particular the first infinite class of prime graphs). Furthermore, they gave tight asymptotics for the Erdos-Hajnal coefficients for many classes of prime tournaments as well as the proof of the conjecture for all but one tournament on at most six vertices and the proof of the weaker version of the conjecture for trees on at most six vertices. In this part of the talk I will summarize these recent achievements. Structured non-linear graph-based hashing is motivated by applications in neural networks, where matrices of linear projections are constrained to have a specific structured form. This drastically reduces the size of the model and speeds up computations. I will show how the properties of the underlying graph encoding correlations between entries of these matrices (such as its chromatic number) imply the quality of the entire non-linear hashing mechanism. Furthermore, I will explain how general structured matrices that very recently attracted researchers’ attention naturally lead to the underlying graph theory description.

A set F of graphs has the Erdos-Posa property if there exists a function f such that every graph either contains k disjoint subgraphs each isomorphic to a member in F or contains at most f(k) vertices intersecting all such subgraphs. In this talk I will address the Erdos-Posa property with respect to three closely related graph containment relations: minor, topological minor, and immersion. We denote the set of graphs containing H as a minor, topological minor and immersion by M(H),T(H) and I(H), respectively. Robertson and Seymour in 1980's proved that M(H) has the Erdos-Posa property if and only if H is planar. And they left the question for characterizing H in which T(H) has the Erdos-Posa property in the same paper. This characterization is expected to be complicated as T(H) has no Erdos-Posa property even for some tree H. In this talk, I will present joint work with Postle and Wollan for providing such a characterization. For immersions, it is more reasonable to consider an edge-variant of the Erdos-Posa property: packing edge-disjoint subgraphs and covering them by edges. I(H) has no this edge-variant of the Erdos-Posa property even for some tree H. However, I will prove that I(H) has the edge-variant of the Erdos-Posa property for every graph H if the host graphs are restricted to be 4-edge-connected. The 4-edge-connectivity cannot be replaced by the 3-edge-connectivity.

For integers k>=1 and n>=2k+1, the bipartite Kneser graph H(n,k) is defined as the graph that has as vertices all k-element and all (n-k)-element subsets of {1,2,...,n}, with an edge between any two vertices (=sets) where one is a subset of the other. It has long been conjectured that all bipartite Kneser graphs have a Hamilton cycle. The special case of this conjecture concerning the Hamiltonicity of the graph H(2k+1,k) became known as the 'middle levels conjecture' or 'revolving door conjecture', and has attracted particular attention over the last 30 years. One of the motivations for tackling these problems is an even more general conjecture due to Lovasz, which asserts that in fact every connected vertex-transitive graph (as e.g. H(n,k)) has a Hamilton cycle (apart from five exceptional graphs). Last week I presented a (rather technical) proof of the middle levels conjecture. In this talk I present a simple and short proof that all bipartite Kneser graphs H(n,k) have a Hamilton cycle (assuming that H(2k+1,k) has one). No prior knowledge will be assumed for this talk (having attended the first talk is not a prerequisite). This is joint work with Pascal Su (ETH Zurich).

We express weight enumerator of each binary linear code as a product. An analogous result was obtain by R. Feynman in the beginning of 60's for the speacial case of the cycle space of the planar graphs.