A homomorphism from a graph $G$ to a graph $H$ is an edge-preserving mapping from $V(G)$ to $V(H)$. In the graph homomorphism problem, denoted by $Hom(H)$, the graph $H$ is fixed and we need to determine if there exists a homomorphism from an instance graph $G$ to $H$. We study the complexity of the problem parameterized by the cutwidth of $G$. We aim, for each $H$, for algorithms for $Hom(H)$ running in time $c_H^k n^{\mathcal{O}(1)}$ and matching lower bounds that exclude $c_H^{k \cdot o(1)}n^{\mathcal{O}(1)}$ or $c_H^{k(1-\Omega(1))}n^{\mathcal{O}(1)}$ time algorithms under the (Strong) Exponential Time Hypothesis. In the paper we introduce a new parameter that we call $\mathrm{mimsup}(H)$. Our main contribution is strong evidence of a close connection between $c_H$ and $\mathrm{mimsup}(H)$: * an information-theoretic argument that the number of states needed in a natural dynamic programming algorithm is at most $\mathrm{mimsup}(H)^k$, * lower bounds that show that for almost all graphs $H$ indeed we have $c_H \geq \mathrm{mimsup}(H)$, assuming the (Strong) Exponential-Time Hypothesis, and * an algorithm with running time $\exp ( {\mathcal{O}( \mathrm{mimsup}(H) \cdot k \log k)}) n^{\mathcal{O}(1)}$. The parameter $\mathrm{mimsup}(H)$ can be thought of as the $p$-th root of the maximum induced matching number in the graph obtained by multiplying $p$ copies of $H$ via certain graph product, where $p$ tends to infinity. It can also be defined as an asymptotic rank parameter of the adjacency matrix of $H$. Our results tightly link the parameterized complexity of a problem to such an asymptotic rank parameter for the first time.
