Optimal Non-Asymptotic Edgeworth Expansions for Multivariate Neural Network Outputs

Batch of new arXiv stat.ML papers (May 26, 2026) covering theoretical and methodological advances: non-asymptotic Edgeworth expansions quantifying finite-width deviations of Gaussian-initialized neural network outputs and posterior approximation error; a formal argument that causal inference is necessary for out-of-distribution generalization and a unifying framework connecting do-calculus, potential outcomes, DML, and IRM; a discriminative neural algorithm for detecting metastable basins via marginal trajectory distribution comparison; MEDAL, a distillation framework that converts manifold/DR