On Graph Isomorphism Problem

TL;DR

This paper introduces a deterministic algorithm for the Graph Isomorphism problem that leverages geometric and combinatorial properties of graph automorphisms, achieving a runtime of n^{O(log n)}.

Contribution

The paper presents a novel deterministic algorithm for GI that combines geometric subspace analysis with combinatorial partitioning techniques.

Findings

Algorithm solves GI in n^{O(log n)} time.

Uses geometric features of automorphism groups.

Employs partition refinement based on subspace decompositions.

Abstract

Let $G$ and $H$ be two simple graphs. A bijection $\phi:V(G)\rightarrow V(H)$ is called an isomorphism between $G$ and $H$ if $(\phi v_i)(\phi v_j)\in E(H)$ $\Leftrightarrow$ $v_i v_j\in E(G)$, $\forall v_i,v_j \in V(G)$. In the case that $G = H$, we say $\phi$ an automorphism of $G$ and denote the group consisting of all automorphisms of $G$ by $\mathrm{Aut}~G$. As well-known, the problem of determining whether or not two given graphs are isomorphic is called Graph Isomorphism Problem (GI). One of key steps in resolving GI is to work out the partition $\Pi^*_G$ of $V(G)$ composed of orbits of $\mathrm{Aut}~G$. By means of geometric features of $\Pi^*_G$ and combinatorial constructions such as the multipartite graph $[\Pi^*_{t_1},\cdots,\Pi^*_{t_s}]$, we can reduce the problem of determining $\Pi_G^*$ to that of working out a series of partitions of $V(G)$ each of which consists of orbits of a stabilizer that fixes a sequence of vertices of $G$, and thus the determination of the partition $\Pi^*_v$ is a critical transition. On the other hand, we have for a given subspace $U \subseteq \mathbb{R}^n$ a permutation group $\mathrm{Aut}~U := \{ \sigma \in S_n : \sigma ~ U = U \}$. As a matter of fact, $\mathrm{Aut}~G = \cap_{\lambda \in \mathrm{spec} \mathbf{A}(G) } \mathrm{Aut}~V_{\lambda}$, and moreover we can obtain a good approximation $\Pi[ \oplus V_{\lambda} ; v ]$ to $\Pi_v^*$ by analyzing a decomposition of $V_{\lambda}$ resulted from the division of $V_{\lambda}$ by subspaces $\{ \mathrm{proj}[ V_{\lambda} ]( \pmb{e}_v )^{\perp} : v \in V(G) \}$. In fact, there is a close relation among subspaces spanned by cells of $\Pi[ \oplus V_{\lambda} ; v ]$ of $G$, which enables us to determine $\Pi_v^*$ more efficiently. In virtue of that, we devise a deterministic algorithm solving GI in time $n^{ O( \log n ) }$.