Counting Small Induced Subgraphs: Hardness via Fourier Analysis

For a fixed graph property $\Phi$ and integer $k \geq 1$, the problem $\#\mathrm{IndSub}(\Phi,k)$ asks to count the induced $k$-vertex subgraphs satisfying $\Phi$ in an input graph $G$. If $\Phi$ is trivial on $k$-vertex graphs (i.e., if $\Phi$ contains either all or no $k$-vertex graphs), this problem is trivial. Otherwise we prove, among other results: - If $\Phi$ is edge-monotone (i.e., closed under deleting edges), then $\#\mathrm{IndSub}(\Phi,k)$ cannot be solved in time $n^{o(k)}$ assuming ETH. This strengthens a result by D\"oring, Marx and Wellnitz [STOC 2024] that only ruled out an exponent of $o(\sqrt{\log k}/ \log \log k)$. Our results also hold when counting modulo fixed primes. - If there is some fixed $\varepsilon > 0$ such that at most $(2-\varepsilon)^{\binom{k}{2}}$ graphs on $k$ vertices satisfy $\Phi$, then $\#\mathrm{IndSub}(\Phi,k)$ cannot be solved in time $n^{o(k/\sqrt{\log k})}$ assuming ETH. Our results hold even when each of the graphs in $\Phi$ may come with an arbitrary individual weight. This generalizes previous results for hereditary properties by Focke and Roth [SIAM J.\ Comput.\ 2024] up to a $\sqrt{\log k}$ factor in the exponent of the lower bound. - If $\Phi$ only depends on the number of edges, then $\#\mathrm{IndSub}(\Phi,k)$ cannot be solved in time $n^{o(k)}$ assuming ETH. This improves on a lower bound by Roth, Schmitt and Wellnitz [FOCS 2020] that only ruled out an exponent of $o(k / \sqrt{\log k})$. In all cases, we also obtain $\mathsf{\#W[1]}$-hardness if $k$ is part of the input and the problem is parameterized by $k$. We also obtain lower bounds on the Weisfeiler-Leman dimension. Our results follow from relatively straightforward Fourier analysis, and our paper subsumes most of the known $\mathsf{\#W[1]}$-hardness results known in the area, often with tighter lower bounds under ETH.