On the Spectral Properties of Matrices Associated with Trend Filters

TL;DR

This paper analyzes the spectral properties of matrices used in trend filtering for time series, providing new theoretical insights and practical methods for filter design based on eigenvalues and eigenvectors.

Contribution

It offers a novel spectral analysis of trend filter matrices, including eigenstructure derivation and boundary condition effects, with new filter design strategies.

Findings

Eigenvectors approximate latent components of time series.

Eigenvalue distribution depends on boundary conditions and filter choice.

New estimators based on cut-off eigenvalues reduce variability.

Abstract

This paper is concerned with the spectral properties of matrices associated with linear filters for the estimation of the underlying trend of a time series. The interest lies in the fact that the eigenvectors can be interpreted as the latent components of any time series that the filter smooths through the corresponding eigenvalues. A difficulty arises because matrices associated with trend filters are finite approximations of Toeplitz operators and therefore very little is known about their eigenstructure, which also depends on the boundary conditions or, equivalently, on the filters for trend estimation at the end of the sample. Assuming reflecting boundary conditions, we derive a time series decomposition in terms of periodic latent components and corresponding smoothing eigenvalues. This decomposition depends on the local polynomial regression estimator chosen for the interior. Otherwise, the eigenvalue distribution is derived with an approximation measured by the size of the perturbation that different boundary conditions apport to the eigenvalues of matrices belonging to algebras with known spectral properties, such as the Circulant or the Cosine. The analytical form of the eigenvectors is then derived with an approximation that involves the extremes only. A further topic investigated in the paper concerns a strategy for a filter design in the time domain. Based on cut-off eigenvalues, new estimators are derived, that are less variable and almost equally biased as the original estimator, based on all the eigenvalues. Empirical examples illustrate the effectiveness of the method.