Sharp hypocoercive convergence estimates for underdamped Langevin dynamics with specular reflection

What the math gives to ML

The paper gives a constructive hypocoercive mechanism for constrained underdamped Langevin dynamics: momentum transport can converge at a rate scaling like the square root of the position-space Poincare constant, rather than the linear rate of overdamped reflected Langevin. The transferable asset is the combination of Hamiltonian transport, friction, Gaussian noise, and exact specular boundary reflection, which preserves a Gibbs distribution while enforcing hard convex constraints. This suggests a constrained stochastic optimizer or Bayesian neural-network sampler whose parameters remain inside a box, simplex, or other convex set without projection bias. The first useful test is whether specular momentum improves mixing or optimization progress over projected overdamped Langevin at equal gradient evaluations.