New XMSE-Aware Estimator Balances ML and EB for Robust Performance

A new paper from researchers Minghao Chen and Jiale Zheng introduces an XMSE-aware adaptive empirical Bayes (EB) estimator that dynamically balances maximum likelihood (ML) and EB shrinkage. Recent XMSE (excess mean squared error) analysis revealed that kernel-based EB can actually perform worse than ML when the kernel is poorly aligned with the true parameters. The authors turn this diagnostic into a design principle: they propose a mixed estimator with a fixed-weight XMSE that is a scalar quadratic function, leading to a closed-form oracle mixing weight. This weight ensures the estimator is no worse than both ML and the base EB at the XMSE scale. A plug-in implementation using finite-sample XMSE approximations is proven consistent, achieving a second-order oracle regret rate for an interior oracle weight.

The paper further establishes transfer of the regret bound to the fixed-weight risk curve evaluated at the selected weight, along with a thresholded boundary rule and extensions to compact kernel families and finite/growing kernel dictionaries with high-probability oracle bounds. Finite impulse response simulations using SURE-tuned, hard-selection, and trace-corrected baselines—plus public benchmarks Silverbox and Cascaded Tanks—show the estimator retains most regularization benefits when helpful and retreats toward ML under kernel misspecification. With 16 pages, 1 figure, and 14 tables, this work provides practical guidelines for building robust EB estimators in machine learning and systems control.