Massively Parallel Exact Inference for Hawkes Processes

Researchers Ahmer Raza and Hudson Smith have published a breakthrough paper titled 'Massively Parallel Exact Inference for Hawkes Processes' that fundamentally changes how these important statistical models are computed. Hawkes processes are widely used for modeling self-exciting event sequences—think financial transactions, social media cascades, or earthquake aftershocks—but traditional maximum likelihood estimation scales quadratically (O(N²)) with the number of events, making large-scale analysis impractical. The researchers discovered that the Hawkes process intensity can be expressed as a product of sparse transition matrices that admit a linear-time associative multiply operation.

This mathematical insight enables computation via a parallel prefix scan algorithm, which is perfectly suited for modern GPU architectures. Their method reduces computational complexity to approximately O(N/P) with P parallel processors while maintaining exact likelihood computation—no approximations or simplifying assumptions required. This preserves the model's interpretability while achieving orders-of-magnitude speedups. The approach also includes a batching scheme that maintains constant memory usage, avoiding GPU memory constraints that typically limit large-scale inference.

The researchers demonstrate their method scales to thousands of nodes and tens of millions of events, substantially beyond what was previously possible. They've provided an open-source PyTorch library implementing these optimizations, making this breakthrough immediately accessible to practitioners in finance, epidemiology, social network analysis, and other fields where Hawkes processes are essential. This represents a significant advancement in both statistical methodology and computational efficiency for point process modeling.