Browse Prior Art Database

# Order Selection for AR Models by Predictive Least Squares

IP.com Disclosure Number: IPCOM000061857D
Original Publication Date: 1986-Sep-01
Included in the Prior Art Database: 2005-Mar-09
Document File: 2 page(s) / 27K

IBM

## Related People

Rissanen, J: AUTHOR [+2]

## Abstract

Disclosed is an algorithm for determining the order of autoregressive (AR) models based on the Predictive Least-Squares principle. The implementation of the order-determining algorithm uses predictive lattice filters, also called ladder forms, that have a modular structure amenable to VLSI implementation. AR models are used in speech modelling and synthesis, equalization of communication channels, spectral estimation, time series analysis and other areas. Given a sequence of t observations of y, we can predict the next observation t+1 of y using a so-called all-pole autoregressive (AR) model of order m as set out in equation (1), (Image Omitted) where the number m and the coefficients a are to be estimated from the t observations of y.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 68% of the total text.

Page 1 of 2

Order Selection for AR Models by Predictive Least Squares

Disclosed is an algorithm for determining the order of autoregressive (AR) models based on the Predictive Least-Squares principle. The implementation of the order-determining algorithm uses predictive lattice filters, also called ladder forms, that have a modular structure amenable to VLSI implementation. AR models are used in speech modelling and synthesis, equalization of communication channels, spectral estimation, time series analysis and other areas. Given a sequence of t observations of y, we can predict the next observation t+1 of y using a so-called all-pole autoregressive (AR) model of order m as set out in equation (1),

(Image Omitted)

where the number m and the coefficients a are to be estimated from the t observations of y. In particular, for a fixed order m the least squares estimates of the coefficients are given by the well-known formula set out in equation (2),

(Image Omitted)

where the matrices Y(t), Y'(t) are defined by the observed data. After having observed T data points, we can estimate an optimal order m by minimizing the accumulated prediction errors as set out in equation (3).

(Image Omitted)

To find this estimate requires the computation of the estimated coefficients according to equation (2), T times for each selected value of m until the minimum is found. But the sum of prediction errors (equation 3) can be calculated without specifically computing the coefficient estimates (equ...