While the method is designed for mesh-derived matrices, we also tested it on sparse matrices in the wild. Results are more variable, but large matrices often benefit, with up to 18.2× end-to-end…