Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
We introduce a broadly applicable approximate MCMC sampler for switching stochastic differential equations (SSDEs) by combining uniformization of the latent CTMC with time-conditioned factorized neural likelihood estimation (FNLE) to approximate transition densities. This approach overcomes limitations of existing SSDE inference methods, such as noise-free observations or state-independent diffusion, by not requiring analytically tractable transition densities. Experiments demonstrate successful recovery of regime paths and parameters for complex SSDE models and application to real-world data.
Bayesian inference for complex switching stochastic differential equations is now broadly tractable, even with non-analytic transition densities.
Switching stochastic differential equations (SSDEs) describe continuous-time dynamics whose parameters switch according to a latent regime process that follows a continuous-time Markov chain (CTMC). By allowing dynamics to change between regimes, SSDEs represent heterogeneous system behavior and have been applied across diverse fields. However, Bayesian inference for SSDEs remains difficult, and existing SSDE inference methods have limited applicability, with restrictions such as noise-free observations, univariate states, linear drift, or state-independent diffusion. In this study, we propose an approximate Markov chain Monte Carlo sampler for SSDEs using uniformization and factorized neural likelihood estimation (FNLE), a simulation-based inference method. Uniformization provides an exact representation of the CTMC but requires SDE transition densities over arbitrary time intervals. We approximate these densities by training a time-conditioned FNLE model. The resulting sampler is broadly applicable to SSDEs without requiring analytically tractable transition densities. In synthetic-data experiments, our method recovered regime paths and parameters for three SSDE models for which previous methods have limited applicability. We also applied our method to a real dataset and detected a regime transition.