An adaptive procedure for Fourier estimators: illustration to deconvolution and decompounding

TL;DR

This paper proposes an adaptive method for selecting cutoff parameters in Fourier density estimators, achieving near-optimal rates for inverse problems like deconvolution and decompounding, with applications to estimating jump densities of Poisson processes.

Contribution

It introduces a new adaptive procedure for Fourier estimators that is applicable to various inverse problems, including novel bounds for decompounding with different sampling rates.

Findings

Achieves rate optimality up to a logarithmic factor.

Provides a new upper bound for the L2-risk in decompounding.

Demonstrates the procedure's effectiveness across different sampling regimes.

Abstract

We introduce a new procedure to select the optimal cutoff parameter for Fourier density estimators that leads to adaptive rate optimal estimators, up to a logarithmic factor. This adaptive procedure applies for different inverse problems. We illustrate it on two classical examples: deconvolution and decompounding, i.e. non-parametric estimation of the jump density of a compound Poisson process from the observation of n increments of length $\Delta$ > 0. For this latter example, we first build an estimator for which we provide an upper bound for its L 2-risk that is valid simultaneously for sampling rates $\Delta$ that can vanish, $\Delta$ := $\Delta$ n $\rightarrow$ 0, can be fixed, $\Delta$ n $\rightarrow$ $\Delta$ 0 > 0 or can get large, $\Delta$ n $\rightarrow$ $\infty$ slowly. This last result is new and presents interest on its own. Then, we show that the adaptive procedure we present leads to an adaptive and rate optimal (up to a logarithmic factor) estimator of the jump density.