PAC-Bayesian Reconstruction Guarantees for Time Series Variational Autoencoders

While Variational Autoencoders are widely deployed in complex time-series domains like finance and healthcare, their theoretical generalization guarantees have remained largely tethered to i.i.d. assumptions. This work develops a PAC-Bayesian framework for latent variable models on sequential data by extending reconstruction-based bounds to Markovian latent structures. Under standard regularity conditions, the derived theoretical bounds remain independent of trajectory length, offering rigorous theoretical grounding for sequential generative forecasting.