pp-eigentest
Posterior predictive eigenvalue testing for signal rank determination. Three-layer consensus architecture with FWER and FDR control; NumPy, C++, JAX, and sparse backends.
pp-eigentest is a posterior predictive eigenvalue testing framework for determining signal rank in high-dimensional datasets. It uses an INID bootstrap over Gram spectra as the null model and tests eigenvalue ratios via a three-layer consensus architecture:
- Dimensionality heuristics — parallel analysis variants, adjacent ratio/gap statistics, calibrated thresholds
- Adaptive thresholding — data-driven cutoffs that adjust to spectral structure
- Multiple testing correction — fixed-sequence FWER, Holm step-down, Benjamini–Hochberg FDR
Supports NumPy, C++, JAX (GPU-accelerated), and sparse backends. The rank-selection logic is model-agnostic: it consumes posterior mean/variance summaries from any generative engine — vbpca-py, probabilistic PCA, or an empirical-Bayes Gaussian model — rather than depending on one specific upstream model.
In development; pre-release. Source at yoavram-lab/pp-eigentest.