Autoregressive Moving Average Graph Filter Design

Abstract

To accurately match a finite-impulse response (FIR) graph filter to a desired response, high filter orders are generally required leading to a high implementation cost. Autoregressive moving average (ARMA) graph filters can alleviate this problem but their design is more challenging. In this paper, we focus on ARMA graph filter design for a known graph. The fundamental aim of our ARMA design is to create a good match to the desired response but with less coefficients than a FIR filter. Our design methods are inspired by Prony’s method but using proper modifications to fit the design to the graph context. Compared with FIR graph filters, our ARMA graph filters show better results for the same number of coefficients.