Count on CFI graphs for #P-hardness

Given graphs $H$ and $G$, possibly with vertex-colors, a homomorphism is a function $f:V(H)\to V(G)$ that preserves colors and edges. Many interesting counting problems (e.g., subgraph and induced subgraph counts) are finite linear combinations $p(\cdot)=\sum_{H}\alpha_{H}\hom(H,\cdot)$ of homomorphism counts, and such linear combinations are known to be hard to evaluate iff they contain a large-treewidth graph $S$. The hardness can be shown in two steps: First, the problems $\hom(S,\cdot)$ for colorful (i.e., bijectively colored) large-treewidth graphs $S$ are shown to be hard. In a second step, these problems are reduced to finite linear combinations of homomorphism counts that contain the uncolored version $S^{\circ}$ of $S$. This step can be performed via inclusion-exclusion in $2^{|E(S)|}\mathrm{poly}(n,s)$ time, where $n$ is the size of the input graph and $s$ is the maximum number of vertices among all graphs in the linear combination. We show that the second step can be performed even in time $4^{\Delta(S)}\mathrm{poly}(n,s)$, where $\Delta(S)$ is the maximum degree of $S$. Our reduction is based on graph products with Cai-F\"urer-Immerman graphs, a novel technique that is likely of independent interest. For colorful graphs $S$ of constant maximum degree, this technique yields a polynomial-time reduction from $\hom(S,\cdot)$ to linear combinations of homomorphism counts involving $S^{\circ}$. Under certain conditions, it actually suffices that a supergraph $T$ of $S^{\circ}$ is contained in the target linear combination. The new reduction yields $\mathsf{\#P}$-hardness results for several counting problems that could previously be studied only under parameterized complexity assumptions. This includes the problems of counting, on input a graph from a restricted graph class and a general graph $G$, the homomorphisms or (induced) subgraph copies from $H$ in $G$.