Density deconvolution under general assumptions on the distribution of measurement errors (Q2039776)

Let us formulate, according to the Introduction to this article, the problem of density deconvolution. Assume that \(X_j\), $j=1,\dots,n$, are i.i.d. random variables with the density \(f\) and \(Y_j\), $j=1,\dots,n$, observed random samples generated from \(Y_j=X_j +\epsilon _j\), for $j=1,\dots,n$. The measurement errors \(\epsilon _ j\), $j=1,\dots,n$, are i.i.d. random variables, independent of \(X_j\), with a known distribution $G$. If one denotes by \(f_Y\) the probability density of $Y=X+\epsilon\), defined as the convolution \(f_Y(y)= [f{\ast}dG](y)\), the aim is to estimate $f$ from the observations \(Y_j\), $j=1,\dots,n$. The accuracy of the estimator \(\tilde {f}\) of $f$ will be measured by the minimax risk. The estimation methods introduced in this paper are based on the Laplace transform. The construction of the estimators is based on the linear functional strategy, used mostly in solving ill-posed inverse problems and is presented in the second section of the paper. The third section is devoted to the introduction of some assumptions on the general distribution of the measurement errors. As examples of distributions satisfying these assumptions the authors analyse the uniform distribution, convolution of uniform distributions, discrete distributions and convolution of uniform and smooth densities. Problems related to Infinite series representation and kernel representation are discussed in the fourth section and finally the introduction of the estimator of $f$ and a result on the upper bound on risk are presented in the fifth section of the article. The sixth section contains results on specific distributions of measurement errors. The result in the seventh section should prove that the considered moment conditions are necessary. Concluding remarks can be found in section eight. All proofs of the reported results are contained in the Appendix.