Predictive density estimators with integrated $L_1$ loss

TL;DR

This paper investigates the efficiency of predictive density estimators under integrated $L_1$ loss, demonstrating that scale expansion improves performance universally across various settings, with numerical and theoretical support.

Contribution

It establishes the universal dominance of scale-expanded densities over plug-in estimators for a broad class of distributions and loss functions, extending previous results.

Findings

Scale expansion improves predictive density estimation for $q$ decreasing and absolutely continuous.

Universal dominance of scale-expanded densities over plug-in estimators for $d \,\geq\, 2$.

Unimodality of $q$ is necessary; in some cases, plug-in estimators are optimal.

Abstract

This paper addresses the problem of an efficient predictive density estimation for the density $q(\|y-\theta\|^2)$ of $Y$ based on $X \sim p(\|x-\theta\|^2)$ for $y, x, \theta \in \mathbb{R}^d$. The chosen criteria are integrated $L_1$ loss given by $L(\theta, \hat{q}) \, =\, \int_{\mathbb{R}^d} \big|\hat{q}(y)- q(\|y-\theta\|^2) \big| \, dy$, and the associated frequentist risk, for $\theta \in \Theta$. For absolutely continuous and strictly decreasing $q$, we establish the inevitability of scale expansion improvements $\hat{q}_c(y;X)\,=\, \frac{1}{c^d} q\big(\|y-X\|^2/c^2 \big) $ over the plug-in density $\hat{q}_1$, for a subset of values $c \in (1,c_0)$. The finding is universal with respect to $p,q$, and $d \geq 2$, and extended to loss functions $\gamma \big(L(\theta, \hat{q} ) \big)$ with strictly increasing $\gamma$. The finding is also extended to include scale expansion improvements of more general plug-in densities $q(\|y-\hat{\theta}(X)\|^2 \big)$, when the parameter space $\Theta$ is a compact subset of $\mathbb{R}^d$. Numerical analyses illustrative of the dominance findings are presented and commented upon. As a complement, we demonstrate that the unimodal assumption on $q$ is necessary with a detailed analysis of cases where the distribution of $Y|\theta$ is uniformly distributed on a ball centered about $\theta$. In such cases, we provide a univariate ($d=1$) example where the best equivariant estimator is a plug-in estimator, and we obtain cases (for $d=1,3$) where the plug-in density $\hat{q}_1$ is optimal among all $\hat{q}_c$.