Sparse recovery in convex hulls via entropy penalization

Abstract

Let $(X,Y)$ be a random couple in $S\times T$ with unknown distribution $P$ and $(X_1,Y_1),...,(X_n,Y_n)$ be i.i.d. copies of $(X,Y).$ Denote $P_n$ the empirical distribution of $(X_1,Y_1),...,(X_n,Y_n).$ Let $h_1,...,h_N:S\mapsto [-1,1]$ be a dictionary that consists of $N$ functions. For $\lambda \in {\mathbb{R}}^N,$ denote $f_{\lambda}:=\sum_{j=1}^N\lambda_jh_j.$ Let $\ell:T\times {\mathbb{R}}\mapsto {\mathbb{R}}$ be a given loss function and suppose it is convex with respect to the second variable. Let $(\ell \bullet f)(x,y):=\ell(y;f(x)).$ Finally, let $\Lambda \subset {\mathbb{R}}^N$ be the simplex of all probability distributions on $\{1,...,N\}.$ Consider the following penalized empirical risk minimization problem \begin{eqnarray*}\hat{\lambda}^{\varepsilon}:={\mathop {argmin}_{\lambda\in \Lambda}}\Biggl[P_n(\ell \bullet f_{\lambda})+\varepsilon \sum_{j=1}^N\lambda_j\log \lambda_j\Biggr]\end{eqnarray*} along with its distribution dependent version \begin{eqnarray*}\lambda^{\varepsilon}:={\mathop {argmin}_{\lambda\in \Lambda}}\Biggl[P(\ell \bullet f_{\lambda})+\varepsilon \sum_{j=1}^N\lambda_j\log \lambda_j\Biggr],\end{eqnarray*} where $\varepsilon\geq 0$ is a regularization parameter. It is proved that the ``approximate sparsity'' of $\lambda^{\varepsilon}$ implies the ``approximate sparsity'' of $\hat{\lambda}^{\varepsilon}$ and the impact of ``sparsity'' on bounding the excess risk of the empirical solution is explored. Similar results are also discussed in the case of entropy penalized density estimation.