level: research
this paper examines the fluctuation behavior of stochastic gradient euler-maruyama estimators used in generalized non-reversible langevin dynamics. the focus is on the leading-order variance when the step size is very small. the analysis assumes an unbiased stochastic gradient oracle and structural conditions that allow a central limit theorem to hold. the empirical average, computed over a time horizon proportional to the inverse squared step size, is shown to satisfy a central limit theorem as the step size goes to zero.
the limiting variance is expressed using the poisson equation of the full-gradient diffusion process. the authors rewrite this variance in an operator form that connects it to the continuous-time asymptotic variance. under standard operator-theoretic assumptions, they derive a condition where adding an anti-symmetric perturbation to the drift strictly reduces the leading-order fluctuation constant compared to the reversible case. this means the non-reversible dynamics can yield more stable estimates.
the results apply to bounded smooth predictive observables, which are common in bayesian inference and optimization tasks. the work provides theoretical backing for using non-reversible langevin algorithms in stochastic gradient settings. it clarifies how the choice of dynamics affects the accuracy of long-run averages. the findings may guide the design of more efficient sampling methods in machine learning.
why it matters: it gives a theoretical reason to use non-reversible dynamics for lower variance in stochastic gradient mcmc, which can improve sampling efficiency in ai models.