Computational and Statistical Aspects of High-Dimensional Structured Estimation

Abstract

Modern statistical learning often faces high-dimensional data, for which the number of features that should be considered is very large. In consideration of various constraints encountered in data collection, such as cost and time, however, the available samples for applications in certain domains are of small size compared with the feature sets. In this scenario, statistical estimation becomes much more challenging than in the large-sample regime. Since the information revealed by small samples is inadequate for finding the optimal model parameters, the estimator may end up with incorrect models that appear to fit the observed data but fail to generalize to unseen ones. Owning to the prior knowledge about the underlying parameters, additional structures can be imposed to effectively reduce the parameter space, in which it is easier to identify the true one with limited data. This simple idea has inspired the study of high-dimensional statistics since its inception. Over the last two decades, sparsity has been one of the most popular structures to exploit when we estimate a high-dimensional parameter, which assumes that the number of nonzero elements in parameter vector/matrix is much smaller than its ambient dimension. For simple scenarios such as linear models, L1-norm based convex estimators like Lasso and Dantzig selector, have been widely used to find the true parameter with reasonable amount of computation and provably small error. Recent years have also seen a variety of structures proposed beyond sparsity, e.g., group sparsity and low-rankness of matrix, which are demonstrated to be useful in many applications. On the other hand, the aforementioned estimators can be extended to leverage new types of structures by finding appropriate convex surrogates like the L1 norm for sparsity. Despite their success on individual structures, current developments towards a unified understanding of various structures are still incomplete in both computational and statistical aspects. Moreover, due to the nature of the model or the parameter structure, the associated estimator can be inherently non-convex, which may need additional care when we consider such unification of different structures. In this thesis, we aim to make progress towards a unified framework for the estimation with general structures, by studying the high-dimensional structured linear model and other semi-parametric and non-convex extensions. In particular, we introduce the generalized Dantzig selector (GDS), which extends the original Dantzig selector for sparse linear models. For the computational aspect, we develop an efficient optimization algorithm to compute the GDS. On statistical side, we establish the recovery guarantees of GDS using certain geometric measures. Then we demonstrate that those geometric measures can be bounded by utilizing simple information of the structures. These results on GDS have been extended to the matrix setting as well. Apart from the linear model, we also investigate one of its semi-parametric extension -- the single-index model (SIM). To estimate the true parameter, we incorporate its structure into two types of simple estimators, whose estimation error can be established using similar geometric measures. Besides we also design a new semi-parametric model called sparse linear isotonic model (SLIM), for which we provide an efficient estimation algorithm along with its statistical guarantees. Lastly, we consider the non-convex estimation for structured multi-response linear models. We propose an alternating estimation procedure to estimate the parameters. In spite of dealing with non-convexity, we show that the statistical guarantees for general structures can be also summarized by the geometric measures.