Dice Question Streamline Icon: https://streamlinehq.com

Detection of subcritical signals in CCA beyond spectral methods

Develop and analyze procedures for detecting and estimating subcritical signals in canonical correlation analysis—when the population squared canonical correlation ρ^2 lies below the spectral threshold 1/sqrt((τ_M−1)(τ_K−1))—particularly for non-Gaussian noise, for example via entrywise nonlinear transformations, and establish whether such methods can provably improve upon the spectral detectability threshold.

Information Square Streamline Icon: https://streamlinehq.com

Background

In the one-spike CCA model, spectral methods cannot detect signals when the population strength is below the BBP-type threshold depending on τ_K and τ_M. Under Gaussian noise, recent results indicate the likelihood ratio does not diverge, suggesting impossibility of reliable detection by spectral statistics.

The survey suggests that for non-Gaussian noise, improved detection may be possible using non-spectral transformations akin to methods in PCA, but a complete theory for CCA is not yet available.

References

The details have not yet been figured out in the CCA setting, however, in the related PCA setup improvement can be achieved by applying a certain function (which depends on the distribution on the noise and is the identical function for the Gaussian noise --- leading to no improvements in the latter case) to each matrix element before using spectral methods, see, e.g., \citet{perry2018optimality} and references therein.

Canonical Correlation Analysis: review (2411.15625 - Bykhovskaya et al., 23 Nov 2024) in Section “Subcritical signals and regularizations” (Chapter 4)