ICML2026

Second-Order Smooth Planning with Optimal-Transport Bellman Smoothing

Tuan Dam

Abstract

Planning with a generative model aims to estimate state values using minimal oracle calls. For entropy-regularized MDPs, SmoothCruiser exploits the smoothness of the LogSumExp\operatorname{LogSumExp} Bellman operator to achieve O~(ε4)\widetilde{\mathcal{O}}(\varepsilon^{-4}) sample complexity, but its first-order Taylor approximation limits the rate. We develop a curvature--complexity theory showing that if a Bellman aggregator has Taylor remainder of order β2\beta \ge 2, the optimal oracle complexity exponent is 2+2/(β1)2 + 2/(\beta-1)---recovering O~(ε4)\widetilde{\mathcal{O}}(\varepsilon^{-4}) for β=2\beta=2 and predicting O~(ε3)\widetilde{\mathcal{O}}(\varepsilon^{-3}) for β=3\beta=3. To achieve β=3\beta=3, we introduce an entropic optimal-transport regularizer over action distributions. The resulting OT-smoothed Bellman operator admits a closed-form expression, explicit gradient policy, and Lipschitz Hessian. We derive an unbiased estimator of the quadratic Taylor term via cross-product debiasing, enabling a second-order SmoothCruiser with O~(ε3)\widetilde{\mathcal{O}}(\varepsilon^{-3}) complexity. We further propose gap-dependent variants and provide a complexity analysis and show advantage of our method.