Stochastic Thresholding: An approach to Estimator Optimization via Fisher Information Maximization

In stochastic thresholding, the threshold for quantization of a signal is randomized. An estimator based on quantized signal data can be optimized through stochastic thresholding. By controlling certain parameters of the probability distribution function of the threshold, we can achieve a gain in Fisher information, a measure of efficacy of any unbiased estimator. Thus any conceivable estimator based on quantized signal will perform better than an estimator operating on original signal provided the stochastic thresholding scheme is followed. Both bi-level and tri-level quantization cases are discussed. The optimization is illustrated by a Maximum Likelihood Estimator to estimate the amplitude of a sinusoid drowned in heavy noise. Stochastic thresholding can also be used to maximize output SNR. This is illustrated by applying 3-level quantization with stochastic thresholding on a digital image. Contrary to intuition, it is seen that loss of information through quantization can be minimized via randomizing the threshold of comparison.