Benign overfitting is the phenomenon where a model that interpolates noisy training data still generalizes well. It is usually characterized by effective-rank conditions on the covariance spectrum. The authors of a new arXiv paper argue that these conditions are fragile: a perturbation of vanishing relative operator and Frobenius norm can turn a benign covariance into one whose minimum-norm interpolator has non-vanishing excess risk.

Instead, the paper proposes a natural geometry based on optimal transport of log-eigenvalues, measured in units of the sample size n. In log-spectral coordinates, the deterministic equivalent of the risk becomes a functional of two measures smoothed by a logistic kernel centered at the interpolation threshold. The exact first variation shows that the threshold moves by a sensitivity-weighted average of log-eigenvalue displacements, with each eigenvalue's influence decaying exponentially with its log-distance to the threshold.

The main results are global, dimension-free stability bounds. The variance functional is 8/(27n)-Lipschitz in the 1-Wasserstein distance between log-spectra, with a sharp constant, and an analogous bound controls the bias. Consequently, any rank-k update of the covariance, of arbitrary norm, changes the variance functional by at most 3k/n, and benign overfitting survives every perturbation whose localized transport cost is o(n). These results extend to ridge regression and infinite-dimensional trace-class covariances, and matching lower bounds show that the n-scaling is necessary and cannot be replaced by operator norms, Frobenius norms, or dimension-normalized spectral distances.

The paper also reports simulations, one-step feature learning in two-layer networks, and random features of real images that support the theory and delineate its limits. As the only source, this work offers a fresh perspective on when interpolation is safe, though it does not claim to overturn existing effective-rank intuition entirely—rather, it reframes the problem in a more robust geometric setting.