Gabriel Victorino Cardoso

Non-Asymptotic Error Bounds for SMC with Biased Proposals: Application to Conditional Diffusion Sampling

Stanislas Strasman, Gabriel Victorino Cardoso, Sylvain Le Corff, Vincent Lemaire, Antonio Ocello

Sequential Monte Carlo (SMC) methods are a natural tool for post-hoc conditioning of pretrained generative models, but in many applications the mutation kernels used by the particle system are biased approximations of an ideal Feynman--Kac flow. This paper develops a non-asymptot…Read more