Title:Tensor network complexity of multilinear maps

Abstract:We study tensor networks as a model of arithmetic computation for evaluating multilinear maps. This model turns out to be strong enough to capture the current best algorithms for a variety of problems, e.g., $O(n^{\omega + \epsilon})$ time matrix multiplication, $O(n \log n)$ time discrete Fourier transform, $O(n^{(\omega+\epsilon)t})$ time $3t$-clique counting and $O^*(2^n)$ time for computing the permanent of a matrix. For counting homomorphisms of a general pattern graph $P$ into a host graph on $n$ vertices we obtain an upper bound of $O(n^{(\omega+\epsilon)\operatorname{bw}(P)/2})$ where $\operatorname{bw}(P)$ is the branchwidth of $P$. This essentially matches the bound for counting cliques, and yields small improvements over previous algorithms for many choices of $P$.
While powerful, the model still has limitations, and we are able to show a number of unconditional lower bounds for various linear maps, including:
(a) an $\Omega(n^{\operatorname{bw}(P)})$ time lower bound for counting homomorphisms from $P$ to an $n$-vertex graph, matching the upper bound if $\omega = 2$. In particular for $P$ a $v$-clique this yields an $\Omega(n^{\lceil 2v/3 \rceil})$ time lower bound for counting $v$-cliques.
(b) an $\Omega(2^{0.918n})$ time lower bound for the permanent of an $n \times n$ matrix.
(c) an $\Omega(n^4)$ time lower bound for the trilinear map $D_{ijk} = \sum_{l} A_{ijl}B_{ikl} C_{jkl}$ generalizing matrix multiplication, taking three $n \times n \times n$ tensors $A, B, C$ and producing an $n \times n \times n$ tensor $D$, ruling out tensor networks as an approach to obtaining non-trivial algorithms for hyperclique counting and the Max-$3$-CSP problem.