Note for Alan ’ s Class

Zhangzhang Si,Haifeng Gong,Song-Chun Zhu,Y. Wu

2009-01-01

Abstract:Wavelets have proven to be immensely useful for signal analysis and representation [7]. Various dictionaries of wavelets have been designed for different types of signals or function spaces [3, 13]. Two key factors underlying the successes of wavelets are the sparsity of the representation and the efficiency of the analysis. Specifically, a signal can typically be represented by a linear superposition of a small number of wavelet elements selected from an appropriate dictionary. The selection can be accomplished by efficient algorithms such as matching pursuit [10] and basis pursuit [4]. From a linear regression perspective, a signal can be considered a response variable, and the wavelet elements in the dictionary can be considered the predictor variables or regressors. The number of elements in a dictionary can often be much greater than the dimensionality of the signal, so this is the so-called “p > n” problem. The selection of the wavelet elements is the variable selection problem in linear regression. The matching pursuit algorithm [10] is the forward selection method, and the basis pursuit [4] is the lasso method [14].