Reproduce “Strictly Constrained Generative Modeling via Split Augmented Langevin Sampling” on OpenResearch

About this paper

Matthieu Blanke, Yongquan Qu, Sara Shamekh, Pierre Gentine

Researchers from NYU and Columbia University developed Constrained Alternated Split Augmented Langevin (CASAL), a framework enabling deep generative models to strictly satisfy mathematical constraints, including non-convex ones, by reformulating constrained sampling as a variational optimization with variable splitting. CASAL effectively explores complex distributions while enforcing hard constraints, demonstrating improved physical plausibility and exploration in scientific applications.

Key insights

• Standard constrained sampling suffers from a fundamental duality gap, meaning that direct Lagrangian approaches or penalty methods cannot strictly converge to the true conditional distribution; variable splitting addresses this.
• CASAL recovers the exact conditional distribution in the infinite penalty limit and provides non-asymptotic convergence guarantees for its mixing rate.
• The dual variable in CASAL plays a crucial role in bias correction, enabling unbiased exploration of complex, non-convex constrained distributions that projected methods often struggle with.

Summary by alphaXiv.