Representative wide NIRS PCA qualification

Four real canonical public matrices were read from Datasets 0.3.14. Their X/M/Y/raw-manifest hashes were checked before and after; no files were modified. The executable benchmark is benchmarks/diagnostics/benchmark_real_pca.py, taking an explicit source/license/identity manifest. It performs three fitted-PCA runs, frozen query projection checks and an independent full NumPy SVD comparison of the exact diagnostic reference. Input reading and the NumPy oracle are excluded from native timing. Timings below are one Linux host, one numerical thread.

Matrix	Shape	Declared license	Median fast fit	Exact diagnostic reference
Figshare UVA illicit drugs, article 21252300	390 × 2151	CC-BY-4.0	0.017 s	0.098 s
EcoSIS NGEE Arctic Seward reflectance	511 × 2151	CC-BY-4.0	0.020 s	0.221 s
EcoSIS Cucurbita two stresses reflectance	541 × 2151	CC-BY-4.0	0.020 s	0.263 s
EcoSIS freeze-dried leaf reflectance	492 × 2151	ODbL-1.0	0.019 s	0.200 s

All have 2151 declared ascending bands from 350 to 2500 nm and finite values. The exact 390×2151 witness is Figshare, not EcoSIS; the historical issue’s stated shape is not evidence of that source identity. In the freeze-dried matrix, 492 rows have 82 declared IDs with multiplicity six. The nature of repetition is not established: these are not labelled technical replicates and do not imply train/test independence. These matrices qualify numerical behavior/performance, not predictive generalization.

Against independent SVD, the exact diagnostic reference’s largest absolute projector discrepancy was 1.8e-15 and its largest retained eigenvalue discrepancy 7.9e-10 (the NGEE eigenvalues are about 1.3e5–4.8e5). Changing query contents never changed calibration scores, means, loadings or eigenvalues. No full moments were requested or represented by a truncated sum.

The existing fast FlexiblePCA mode uses the native deterministic randomized SVD for these shapes. Its approximation is distinct: on NGEE, its projector differed by 0.00573 and the eigenvalue absolute discrepancy reached 116, about 0.09% of the smallest retained eigenvalue. This observation motivated exposing the existing exact Gram solver in the new diagnostic profile. The fast estimator and its serialized-state behavior were not silently changed.

RSS was sampled every 2 ms during the fast fit/projections; sampled incremental peaks ranged from roughly 4 KiB (allocator reuse) to 8.4 MiB, and are not rigorous allocation bounds. The JSON receipt retains baseline/peak RSS, repetitions, source hashes, library hash, exact timings and oracle errors. The native budget counts all live buffers and the eigensolver iteration ceiling; this benchmark supports these representative shapes without removing that bound. Windows and browser execution have separate qualification receipts; Linux timings are not transferred to those environments.