An arXiv preprint proposes a statistical framework for stochastic optimization. The authors model the gradient as a normal distribution and argue that the stochastic gradient should be whitened using its covariance. This perspective is used to analyze and design methods such as Adam, Shampoo, and Muon.
The abstract is truncated, so the full derivation and experimental results are not yet available from the announcement. What is clear is the central claim: whitening by covariance is a principled way to treat noisy gradients, potentially unifying or improving existing optimizers.
Because only the abstract was provided, no comparison with other sources is possible. The paper's significance lies in offering a statistical justification for design choices in popular optimization algorithms.