Stability and Risk Bounds of Iterative Hard Thresholding

The Iterative Hard Thresholding (IHT) algorithm is one of the most popular and promising greedy pursuit methods for high-dimensional statistical estimation under cardinality constraint. The existing analysis of IHT mostly focuses on parameter estimation and sparsity recovery consistency. From the perspective of statistical learning theory, another fundamental question is how well the IHT estimation would perform on unseen samples. The answer to this question is important for understanding the generalization ability of IHT yet has remaind elusive. In this paper, we investigate this problem and develop a novel generalization theory for IHT from the viewpoint of algorithmic stability. Our theory reveals that: 1) under natural conditions on the empirical risk function over n samples of dimension p, IHT with sparsity level k enjoys an (O) over tilde (n(-1/2)root k log(n) log(p)) rate of convergence in sparse excess risk; and 2) a fast rate of order (O) over tilde (n(-1) k(log(3) (n) + log(p))) can be derived for strongly convex risk function under certain strong-signal conditions. The results have been substantialized to sparse linear regression and logistic regression models along with numerical evidence provided to support our theory.