Answers are grounded in this paper and cite their sources.
LEMNA explains a security classifier by fitting a local mixture of simple regressions under a fused (total-variation) constraint. I adapted it to EMBER features, with 2,381 features and 1,000 samples in each local neighbourhood, and asked whether its regression step can be solved reliably. The fusion graph has 605 edges, so the matrix inside the solver has nullity of at least 776. On two saved inputs, the stacked matrix has a measured numerical rank of 1,020 and 681 out of 2,381 columns. On 16 saved subproblems that had failed my convergence rule, jittered Cholesky, LSMR and LSQR passed the residual tests in 4/16, 4/16 and 2/16 cases. No returned solution stayed inside the TV budget under a zero-tolerance check. For the 16 saved Cholesky outputs, I show that this check is stricter than the residual tests: the excess is bounded by √605 times the split gap ||Dβ − z||₂, and in all 16 cases it was inside that bound (at most 7.5 × 10⁻⁴). On two inputs, a larger budget did not help, but 1,000 or 2,000 iterations gave a residual certificate, and an independent OSQP solve reached objectives 1.3% and 0.10% below the 200-iteration result. Synthetic tests show that rank deficiency alone need not block convergence. I do not establish a cause, and I report no attack, calibration or official-test experiment.
No takes yet. Share an insight, caveat, or question.
Kottapalli Deepak Varma (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: