Skip to content

Sparse MFPCA observation stress and performance

This page closes the final pre-release-candidate evidence gate for the public native sparse MFPCA estimator. It keeps observation-process sensitivity and runtime/memory characterization separate from native known-truth qualification.

Observation-process stress

The stress runner reuses the frozen pre-0.12 row-loss mechanisms and applies them to paired planar observations before fitting the public fit_sparse_mfpca() estimator.

The supported contract remains unchanged: x and y are jointly observed at the same retained native timestamps. Row deletion can depend on x, y, eccentricity, velocity, or phase, but it does not create coordinate-specific missingness.

The deterministic matrix uses rho_xy=0.6, three seeded replicates, and:

none
mcar
signal_dependent_x_loss
signal_dependent_y_loss
eccentricity_dependent_loss
velocity_dependent_loss
phase_dependent_loss

Each fit retains the same declared grid, bandwidths, component count, measurement-error variances, PSD action, and joint-PACE contract used by the native recovery qualification.

The evidence records actual loss, mean/covariance/Cxy ISE, joint-subspace cosines, score correlation, reconstruction ISE, score-system failures and conditioning, and PSD correction.

This is descriptive stress evidence:

threshold_gate_applied = false
parameter_selection_performed = false

Signal-dependent observation loss can change finite-sample recovery and is not silently reclassified as an estimator defect. The stress table does not select bandwidths, ridge, grid size, component count, or public defaults.

Dedicated sparse-MFPCA performance envelope

A separate runner measures only the public native sparse MFPCA/joint-PACE pipeline. Synthetic data generation occurs before the timed fit, and every timing repetition is executed in a fresh Python process with BLAS/OpenMP thread counts fixed to one.

Three planar workloads are declared:

very_sparse_planar
moderate_planar
irregular_rich_planar

For each workload the evidence retains:

  • number of curves;
  • median native samples per curve;
  • ordered off-diagonal covariance-pair count per block;
  • total xx/xy/yy latent pair products;
  • evaluation-grid size;
  • returned component count;
  • runtime median, IQR and range over at least three repetitions;
  • process peak RSS median/max when available;
  • score-failure rate;
  • relative joint-PSD correction;
  • Python/package/platform/CPU/commit/run provenance.

The envelope is non-comparative and threshold-free:

comparative_benchmark = false
speed_threshold_applied = false

It does not claim that eyetrajectoriespy is faster than another package, does not define a universal workstation runtime, and does not weaken scientific computation to satisfy a speed target.

Observed PR evidence

All 21 public-estimator fits in the three-replicate stress matrix completed successfully, with score-failure rate 0 for every mechanism.

The no-loss baseline had mean ISE about 0.055, joint covariance ISE 0.239, minimum joint-subspace cosine 0.982, median separated score correlation 0.951, and reconstruction ISE 0.0216. MCAR at about 20.4% loss remained close to that baseline: joint covariance ISE 0.272, minimum subspace cosine 0.980, median score correlation 0.941, and reconstruction ISE 0.0254.

The informative mechanisms were visibly more sensitive in different ways. Signal-dependent x/y loss increased mean ISE to about 0.328/0.378 and reconstruction ISE to about 0.070/0.054 while retaining minimum joint-subspace cosines around 0.975. Eccentricity-dependent loss produced the largest covariance/reconstruction change in this matrix: joint covariance ISE about 1.45, Cxy ISE about 0.255, and reconstruction ISE about 0.144, while the minimum joint-subspace cosine remained about 0.964. Velocity- and phase-dependent loss were closer to the baseline on the principal recovery metrics. These differences are retained as sensitivity evidence, not reclassified as estimator defects or used to select tuning parameters.

The dedicated performance run used GitHub-hosted Linux/CPython 3.12 on an AMD EPYC 9V74 runner with numerical-library thread counts fixed to one. Median public-fit runtimes over three fresh-process repetitions were approximately 0.51 s for 80 curves with median 8 native samples and a 21-point grid, 1.35 s for 60 curves with median 14 samples and a 31-point grid, and 2.98 s for 36 curves with median 23.5 samples and a 41-point grid. Peak process RSS medians were approximately 226, 227 and 229 MiB, respectively. All three workloads had score-failure rate 0. These values define a recorded environment-specific performance envelope only; no speed threshold or cross-package claim follows from them.

RC interpretation

After this tranche, the 0.12 methodological evidence sequence contains:

  1. analytical/operator checks;
  2. native known-truth recovery;
  3. grid/ridge sensitivity;
  4. internal two-stage marginal-basis sensitivity;
  5. external mGSFPCA 0.2.2 sensitivity;
  6. public-estimator observation-process stress;
  7. dedicated sparse-MFPCA runtime/peak-memory characterization;
  8. existing documentation, portability, release-readiness, examples and cross-platform package CI.

An RC decision should still be made explicitly after reviewing the complete evidence. This page does not itself change the package version or publish 0.12.