Minimax Estimation of Kernel Stein Discrepancy: Trace versus Hilbert-Schmidt Scales

What the math gives to ML

The paper distinguishes estimating the scalar KSD from estimating the full Stein mean embedding and shows that the optimal scalar estimator is controlled by the Hilbert–Schmidt norm of a Stein covariance operator. The positive-part square-root of an unbiased U-statistic avoids the diagonal contribution that makes the usual V-statistic fluctuate at the larger trace scale, with the gap growing with effective rank and potentially exponentially in dimension. The most direct neural-network transfer is a lower-variance KSD evaluation or training loss for samplers, score models, and variational generators whose target score is available.