Rare-event sampling for stochastic dynamics in network systems using cluster updates

What the math gives to ML

The paper introduces conditional-path Monte Carlo (CPMC), a trajectory-space sampler that replaces local rejected path changes with non-local spacetime-cluster updates while enforcing a rare macroscopic boundary condition at every sampled trajectory. Its transferable asset is a Markov chain over event trajectories whose states all satisfy a target event, avoiding the exponentially small acceptance rates and genealogical collapse of forward simulation or splitting. A direct neural-network use is rare-event data generation for stochastic RNNs, neural state-space models, or neural simulators: sample trajectories conditioned on a rare terminal failure, constraint violation, or high-loss event, then train on these paths. The key falsifiable prediction is that conditional cluster updates preserve nonzero effective sample size as the target-event probability becomes exponentially small, whereas rejection-based trajectory sampling loses usable samples exponentially.