Counting large patterns in degenerate graphs

TL;DR

This paper improves algorithms for counting large patterns in $d$-degenerate graphs by introducing the concept of $(c,d)$-locatable graphs, achieving more efficient runtimes with polynomial dependence on pattern size.

Contribution

The paper introduces $(c,d)$-locatable graphs, extending previous work, and provides improved algorithms with polynomial dependence on pattern size for counting subgraphs in $d$-degenerate graphs.

Findings

Improved subgraph counting algorithms for $(c,d)$-locatable graphs.

Characterization of $(1,d)$-locatable graphs with linear runtime dependence.

Lower bounds showing hardness for barely non-$(c,d)$-locatable graphs.

Abstract

The problem of subgraph counting asks for the number of occurrences of a pattern graph $H$ as a subgraph of a host graph $G$ and is known to be computationally challenging: it is $\#W[1]$-hard even when $H$ is restricted to simple structures such as cliques or paths. Curticapean and Marx (FOCS'14) show that if the graph $H$ has vertex cover number $\tau$, subgraph counting has time complexity $O(|H|^{2^{O(\tau)}} |G|^{\tau + O(1)})$. This raises the question of whether this upper bound can be improved for input graphs $G$ from a restricted family of graphs. Earlier work by Eppstein~(IPL'94) shows that this is indeed possible, by proving that when $G$ is a $d$-degenerate graph and $H$ is a biclique of arbitrary size, subgraph counting has time complexity $O(d 3^{d/3} |G|)$. We show that if the input is restricted to $d$-degenerate graphs, the upper bound of Curticapean and Marx can be improved for a family of graphs $H$ that includes all bicliques and satisfies a property we call $(c,d)$-locatable. Importantly, our algorithm's running time only has a polynomial dependence on the size of~$H$. A key feature of $(c,d)$-locatable graphs $H$ is that they admit a vertex cover of size at most $cd$. We further characterize $(1,d)$-locatable graphs, for which our algorithms achieve a linear running time dependence on $|G|$, and we establish a lower bound showing that counting graphs which are barely not $(1,d)$-locatable is already $\#\text{W}[1]$-hard. We note that the restriction to $d$-degenerate graphs has been a fruitful line of research leading to two very general results (FOCS'21, SODA'25) and this creates the impression that we largely understand the complexity of counting substructures in degenerate graphs. However, all aforementioned results have an exponential dependency on the size of the pattern graph $H$.