Fast Covariance Estimation for High-dimensional Functional Data

TL;DR

This paper introduces two scalable algorithms, FACE and SVDS, for fast covariance estimation in high-dimensional functional data, enabling practical analysis of large covariance matrices in medical imaging and sensor data.

Contribution

The paper presents two novel algorithms, FACE and SVDS, that significantly improve the speed and memory efficiency of covariance smoothing for very large matrices.

Findings

FACE is an order of magnitude faster than existing methods.

Both algorithms handle matrices up to 100,000 dimensions.

FACE enhances functional regression speed in R.

Abstract

For smoothing covariance functions, we propose two fast algorithms that scale linearly with the number of observations per function. Most available methods and software cannot smooth covariance matrices of dimension $J \times J$ with $J>500$; the recently introduced sandwich smoother is an exception, but it is not adapted to smooth covariance matrices of large dimensions such as $J \ge 10,000$. Covariance matrices of order $J=10,000$, and even $J=100,000$, are becoming increasingly common, e.g., in 2- and 3-dimensional medical imaging and high-density wearable sensor data. We introduce two new algorithms that can handle very large covariance matrices: 1) FACE: a fast implementation of the sandwich smoother and 2) SVDS: a two-step procedure that first applies singular value decomposition to the data matrix and then smoothes the eigenvectors. Compared to existing techniques, these new algorithms are at least an order of magnitude faster in high dimensions and drastically reduce memory requirements. The new algorithms provide instantaneous (few seconds) smoothing for matrices of dimension $J=10,000$ and very fast ($<$ 10 minutes) smoothing for $J=100,000$. Although SVDS is simpler than FACE, we provide ready to use, scalable R software for FACE. When incorporated into R package {\it refund}, FACE improves the speed of penalized functional regression by an order of magnitude, even for data of normal size ($J <500$). We recommend that FACE be used in practice for the analysis of noisy and high-dimensional functional data.