Fusion of Hard and Soft Information in Nonparametric Density Estimation

Abstract

Abstract.This article discusses univariate density estimation in situations when the sample (hard information) is supplemented by "soft" information about the random phenomenon. These situations arise broadly in operations research and management science where practical and computational reasons severely limit the sample size, but problem structure and past experiences could be brought in. In particular, density estimation is needed for generation of input densities to simulation and stochastic optimization models, in analysis of simulation output, and when instantiating probability models. We adopt a constrained maximum likelihood estimator that incorporates any, possibly random, soft information through an arbitrary collection of constraints. We illustrate the breadth of possibilities by discussing soft information about shape, support, continuity, smoothness, slope, location of modes, symmetry, density values, neighborhood of known density, moments, and distribution functions. The maximization takes place over spaces of extended real-valued semicontinuous functions and therefore allows us to consider essentially any conceivable density as well as convenient exponential transformations. The infinite dimensionality of the optimization problem is overcome by approximating splines tailored to these spaces. To facilitate the treatment of small samples, the construction of these splines is decoupled from the sample. We discuss existence and uniqueness of the estimator, examine consistency under increasing hard and soft information, and give rates of convergence. Numerical examples illustrate the value of soft information, the ability to generate a family of diverse densities, and the effect of misspecification of soft information.

Stability and Error Analysis for Optimization and Generalized Equations

“…It anyhow furnishes accurate estimates of the Attouch-Wets distance [32,33]. This global perspective based on set distances provides foundations for computationally attractive approximations of functions [35,33,34] and formulations of function identification problems [35], especially in nonparametric statistics [38,37].…”