Sharp oracle inequalities for low-complexity priors (Q2304249)

Let \(\overline{y}=(y_1, y_2,\dots, y_n)\) be \(n\) observations with a common marginal distribution, and let \(\mathbf{A}\) be a deterministic \(n\times p\) matrix. The authors consider problems related with the estimation of the parameter vector \(\overline{\theta}\in \mathbb{R}^p\) based on the data \(\overline{y}\) and matrix \(\mathbf{A}\). Let \(F:\mathbb{R}^n\times \mathbb{R}^n\rightarrow\mathbb{R}\) be a loss function supposed to be smooth and convex that assigns to each \(\overline{\theta}\in\mathbb{R}^p\) the cost \(F(\mathbf{A}\overline{\theta},\overline{y})\). Let \[ \overline{\theta}_0=\mathop{\mathrm{Argmin}}\limits_{\overline{\theta}\in\mathbb{R}^p}\mathbb{E}\left[F(\mathbf{A}\overline{\theta},\overline{y})\right] \] be a minimizer of the population risk. The authors provide general oracle inequalities in prediction for two estimators of \(\overline{\theta}_0\), the penalyzed estimator and the exponential weighted aggregation. Particular attention is paid to the case \(p>n\).