Preconditioner Structures

Kronecker and E-KFAC structures more consistently improved convergence and final accuracy than diagonal structures in CIFAR experiments. The paper tested diagonal, K-FAC, and E-KFAC block structures for the inverse preconditioner. E-KFAC is treated as an architectural bias for learned inverse actions rather than as requiring an exact curvature eigenbasis. Bias and normalization-scale parameters are handled diagonally in the tested structures. Diagonal LLQR is cheap but often lacks enough expressivity for standard CIFAR classification convergence.