Researchers develop GPU-accelerated algorithm that reduces Hawkes process computation from O(N²) to O(N/P) using parallel prefix scanning

Multivariate Hawkes processes are self-exciting point processes widely used in modeling, but maximum likelihood estimation traditionally scales as O(N²) with the number of events

New method expresses Hawkes process intensity as a product of sparse transition matrices, enabling linear-time associative multiplication and parallel computation

Algorithm achieves near-linear speedup with P parallel processors on modern GPUs while maintaining constant memory usage and computing exact likelihood values

Approach combines parallel prefix scan techniques with a natural batching scheme to avoid GPU memory constraints typical in large-scale inference problems