On signal and extraneous roots in Singular Spectrum Analysis

TL;DR

This paper analyzes the roots of characteristic polynomials in Singular Spectrum Analysis, revealing how signal and extraneous roots influence component separation, forecasting, and parameter estimation in time series analysis.

Contribution

It provides a detailed characterization of signal and extraneous roots in SSA, including conditions for exact separability and the asymptotic distribution of extraneous roots.

Findings

Characterization of exact separability cases

Asymptotic distribution of extraneous roots

Impact of roots on SSA forecasting accuracy

Abstract

In the present paper we study properties of roots of characteristic polynomials for the linear recurrent formulae (LRF) that govern time series. We also investigate how the values of these roots affect Singular Spectrum Analysis implications, in what concerns separation of components, SSA forecasting and related signal parameter estimation methods. The roots of the characteristic polynomial for an LRF comprise the signal roots, which determine the structure of the time series, and extraneous roots. We show how the separability of two time series can be characterized in terms of their signal roots. All possible cases of exact separability are enumerated. We also examine properties of extraneous roots of the LRF used in SSA forecasting algorithms, which is equivalent to the Min-Norm vector in subspace-based estimation methods. We apply recent theoretical results for orthogonal polynomials on the unit circle, which enable us to precisely describe the asymptotic distribution of extraneous roots relative to the position of the signal roots.