All-Pairs Shortest Paths with Few Weights per Node

We study the central All-Pairs Shortest Paths (APSP) problem under the restriction that there are at most $d$ distinct weights on the outgoing edges from every node. For $d=n$ this is the classical (unrestricted) APSP problem that is hypothesized to require cubic time $n^{3-o(1)}$, and at the other extreme, for $d=1$, it is equivalent to the Node-Weighted APSP problem. We present new algorithms that achieve the following results: 1. Node-Weighted APSP can be solved in time $\tilde{O}(n^{(3+\omega)/2}) = \tilde{O}(n^{2.686})$, improving on the 15-year-old subcubic bounds $\tilde{O}(n^{(9+\omega)/4}) = \tilde{O}(n^{2.843})$ [Chan; STOC '07] and $\tilde{O}(n^{2.830})$ [Yuster; SODA '09]. This positively resolves the question of whether Node-Weighted APSP is an ``intermediate'' problem in the sense of having complexity $n^{2.5+o(1)}$ if $\omega=2$, in which case it also matches an $n^{2.5-o(1)}$ conditional lower bound. 2. For up to $d \leq n^{3-\omega-\epsilon}$ distinct weights per node (where $\epsilon > 0$), the problem can be solved in subcubic time $O(n^{3-f(\epsilon)})$ (where $f(\epsilon) > 0$). In particular, assuming that $\omega = 2$, we can tolerate any sublinear number of distinct weights per node $d \leq n^{1-\epsilon}$, whereas previous work [Yuster; SODA '09] could only handle $d \leq n^{1/2-\epsilon}$ in subcubic time. This promotes our understanding of the APSP hypothesis showing that the hardest instances must exhaust a linear number of weights per node. Our result also applies to the All-Pairs Exact Triangle problem, thus generalizing a result of Chan and Lewenstein on "Clustered 3SUM" from arrays to matrices. Notably, our technique constitutes a rare application of additive combinatorics in graph algorithms.