Confidence Intervals for the Savitzky-Golay Filter with an Application to the Keeling Data for Atmospheric CO2

TL;DR

This paper develops a statistical method to estimate noise variance and confidence intervals for the Savitzky-Golay filter, demonstrated through atmospheric CO2 data analysis, improving the filter's interpretability.

Contribution

It introduces a novel approach to derive reliable noise variance estimates and confidence intervals for the Savitzky-Golay filter, enhancing its statistical foundation.

Findings

Reliable noise variance estimation method proposed

Confidence intervals for filter output established

Application to Keeling CO2 data demonstrated effectiveness

Abstract

The Savitzky-Golay FIR digital filter is based on a least-squares polynomial fit to a sample of equally spaced data. The polynomial fit gives the filter the ability to preserve moments of features in the data like peak width. However the S-G filter is not generally regarded as having a sound statistical basis. This puts the filter in the category of smoothing filters where the degree of smoothing depends on the somewhat arbitrary choice of the filter parameters. This arbitrariness makes the variance of the residuals between the filter input and output an unreliable estimate of the variance of the noise in the filter input. And without a reliable estimate of the input noise variance there is no basis for determining statistically meaningful confidence intervals on the filter output. This paper proposes a method of using the S-G filter to determine a reliable estimate of the variance of the noise in the data. This estimate is then used as the basis for selecting appropriate filter parameters and determining statistically meaningful confidence intervals on the filter output. To illustrate the proposed method an analysis of the Keeling measurements of atmospheric CO2 concentration is presented.