Researchers propose λ-PSD for scalable statistical testing
New method λ-PSD boosts statistical test power 10x while cutting compute time linearly
Deep Dive
λ-PSD redefines Stein discrepancy construction as an explicit signal-to-noise ratio (SNR) optimization problem, avoiding the exponential SNR collapse observed in standard polynomial Stein discrepancies. By applying a covariance-aware reweighting in a low-dimensional subspace, λ-PSD achieves stable SNR and substantially improves test power while preserving linear-time complexity in sample size.
Key Points
- λ-PSD optimizes SNR² explicitly, avoiding exponential decay seen in traditional polynomial Stein discrepancies
- Maintains linear O(n) time complexity while achieving 10x higher test power in empirical evaluations
- Validated under Gaussian settings with a 15-page paper including 5 figures and theoretical guarantees
Why It Matters
Enables faster, more reliable statistical testing for machine learning and scientific applications without sacrificing computational efficiency