The block generalized-power method is a common heuristic for sparse spectral optimization, but the standard block formulation mixes two nonequivalent objectives and leaves support selection details unspecified. The author rebuilds the model from the primal, deriving two exact identities: A-SUM, a sum of soft-thresholded Euclidean norms, and A-SQ, the squared-threshold objective obtained by a quadratic lift. A two-component example shows their maximizers can disagree, so the choice matters.

On the theory side, the paper separates the historical full-polar iteration from a safeguarded Riemannian ascent for A-SQ and proves whole-sequence convergence using sufficient-ascent and relative-error lemmas plus the Kurdyka-Lojasiewicz property. A strict threshold margin forces finite identification of support and sign pattern; after identification, the map reduces to ordinary power iteration on the active covariance for rank-one hard thresholding, and a radius-based safe-screening rule follows.

Experiments on planted-spike data match predicted identification time and local rates. On GPT-2 residual streams, ViT CLS streams, and Pitprops, the thresholded block-sparse components match dense PCA on reconstruction and improve monosemanticity and class specialization. On three single-cell RNA-seq datasets, A-SQ beats dense PCA on cluster-recovery AMI by +0.023 to +0.044 at matched sparsity.