You Jiacheng Describes Optimizer Without Eig Inverse or Solve

You Jiacheng posted on X that combining the multiplicative update from PSGD-Kron with the KL objective from KL-Shampoo yields an optimizer without eigendecomposition, inverse, or solve steps. He specified a multiplicative update on P using the objective KL between N(0, Σ) and N(0, P^{-2}), where Σ equals E[gg.T]. Jiacheng credited the work to GPT Pro and attached an image of the derivation.