Spiking neural networks, which communicate through discrete spikes, have an input space that can be divided into 'causal pieces' — distinct regions that determine how the network responds. The number of these pieces is a natural measure of a network's complexity, and prior work has shown that it also provides a lower bound on approximation error for some spiking networks: a network with more causal pieces can, in principle, approximate a wider class of functions.

The new arXiv preprint complements that result by establishing upper bounds on the number of causal pieces. This gives researchers a more complete picture of the trade-off between network size and expressivity, and helps characterise the fundamental limits of what single-spike neural networks can compute. The exact form of the bound is not detailed in the abstract, but the result narrows the theoretical gap between known lower and upper limits.

As a preprint, the work has not yet undergone peer review, but it adds to a growing body of theory on the computational power of spiking neural networks.