Nguyen and Vu prove particle mean-field variational inference stays stable without strong convexity

Vinh Nguyen and Truong Vu have released a new paper on arXiv (2608.11486) that tackles a long-standing limitation in particle-based variational inference. Mean-field variational inference (MFVI) approximates complex posterior distributions by evolving a system of interacting particles. Most existing convergence proofs rely on strong convexity—a restrictive condition that guarantees the optimization landscape is a 'bowl' with no tricky saddle points. In practice, many real-world target distributions violate this assumption, which has left a gap between theory and application.

The authors prove a non-asymptotic Wasserstein stability bound that works even when the target potential is merely smooth. They quantify the departure from convexity using a curvature defect, d_alpha(x,y), and show that if this defect is uniformly bounded by beta, then the particle iterates remain within O(sqrt(beta/alpha)) of any MFVI minimizer. Crucially, their error bound explicitly separates different error sources: initialization error, product-empirical approximation error, finite-batch drift error, and time discretization error. This decomposition lets practitioners see exactly how each algorithmic choice—particle number, batch size, step size—contributes to the overall approximation quality.

The proof technique is notable: the authors construct a 'stationary comparison array' whose population law is an MFVI minimizer but whose particle-level law is a random product empirical measure. This allows them to control the projected-drift discrepancy explicitly, leading to sharper bounds than previous work. They also provide coordinatewise defect estimates and structural conditions for dimension-independent sensitivity, meaning their bounds can scale to high-dimensional problems without the usual curse of dimensionality.

To demonstrate the applicability of their theory, the authors build an arbitrary-dimensional smooth nonconvex benchmark with a closed-form MFVI minimizer—a useful test case for future algorithms. They also explain why polynomially growing drifts require a modification of the standard explicit time-stepping scheme, warning practitioners about a subtle but important implementation pitfall. The 39-page paper, complete with 8 figures, is a significant step toward making variational inference guarantees realistic for modern machine learning workloads, where nonconvexity is the norm rather than the exception.