Optimal Quantization of Signals for System Identification

TL;DR

This paper investigates optimal quantization schemes for signals in system identification, deriving solutions for high and low resolution cases to minimize estimation errors under resource constraints.

Contribution

It introduces a method to derive optimal quantizers for system identification, including solutions for both high and low resolution scenarios, improving estimation accuracy.

Findings

Optimal quantizer in high-resolution case derived from Euler-Lagrange equations.

Low-resolution optimal quantizer found via recursive minimization, coarse near origin.

Quantization impacts data requirements for reducing estimation errors.

Abstract

In this paper, we examine the optimal quantization of signals for system identification. We deal with memoryless quantization for the output signals and derive the optimal quantization schemes. The objective functions are the errors of least squares parameter estimation subject to a constraint on the number of subsections of the quantized signals or the expectation of the optimal code length for either high or low resolution. In the high-resolution case, the optimal quantizer is found by solving Euler-Lagrange's equations and the solutions are simple functions of the probability densities of the regressor vector. In order to clarify the minute structure of the quantization, the optimal quantizer in the low resolution case is found by solving recursively a minimization of a one-dimensional rational function. The solution has the property that it is coarse near the origin of its input and becomes dense away from the origin in the usual situation. Finally the required quantity of data to decrease the total parameter estimation error, caused by quantization and noise, is discussed.