← AI Techniques Radar
Inference efficiency

Preconditioned Implicit Midpoint Langevin (P-IMLA) for Non-Smooth Log-Concave Bayesian Imaging

Use a preconditioned implicit-midpoint discretization of Langevin dynamics, where an implicit midpoint update (via a Cayley-transform identity) is combined with matrix preconditioning to avoid the stationary-variance inflation attributed to preconditioned Euler-Maruyama. Optionally, replace the non-smooth regularization with a Moreau-envelope-based proximal form to handle non-smooth objectives.

Discovered 2026-11-10 · Sources checked 2026-11-10

This summary was generated with AI assistance.

Why it matters

For high-dimensional log-concave targets with non-smooth regularization, naïve preconditioning can introduce surplus stationary variance. The implicit midpoint step is reported to remove this surplus and improve bias/accuracy characteristics compared with explicit preconditioned Euler-style updates.

What changed

The method changes the discretization scheme from (preconditioned) Euler-Maruyama to an implicit midpoint update under preconditioning, and extends the sampler to a proximal non-smooth variant using an M-norm Moreau envelope.

Where it helps

  • Bayesian imaging with TV-like non-smooth regularization
  • Sampling from high-dimensional log-concave posteriors
  • Reducing stationary-variance inflation when using matrix preconditioning
  • Proximal handling of non-smooth objectives in Langevin sampling

Evidence & reproduction

Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.

Limitations & open questions

Reported guarantees rely on conditions stated in the source, including explicit smoothness assumptions for certain drift/bias results; the behavior under general non-smooth settings beyond the described assumptions is not specified in the text.

Sources