arXiv:2607.01805 2026 Training 2 ideas extracted · analyzed Aug 30, 2026
The paper introduces a three-parameter divergence obtained by replacing the logarithm in the quantum alpha-z Renyi functional with the r-logarithm, while retaining matrix-power structure and data-processing guarantees in specified parameter ranges. The transferable asset is a tunable power-law deformation of the usual log-sum-exp/Renyi objective: it changes gradient weighting between high- and low-probability events and provides an explicit mixing inequality. A practical first use is as a drop-in classification or consistency loss on categorical distributions, where the quantum expression reduces to a computable classical weighted power sum. The data-processing inequality also motivates regularizing predictions across stochastic augmentations or bottlenecks.
Use the divergence's data-processing principle as a consistency objective between predictions before and after a stochastic augmentation or feature bottleneck. Penalize disagreement under transformations while retaining the asymmetric power-law weighting of the r-deformed divergence.
Replace cross-entropy or ordinary Renyi loss between a target distribution and a model distribution with the paper's r-deformed alpha-z divergence. The deformation parameter r provides a controllable power-law alternative to the logarithm, allowing experiments that emphasize hard, low-probability target events differently from standard log losses.