Strong consistency and rates for deconvolution of multivariate densities of stationary processes (Q689165)

The author studies the consistency of the estimators of a multivariate density when the observations are perturbed with a noise, i.e. \(Y_ i = X_ i + \varepsilon_ i\). Here \(\varepsilon_ i\) is the noise sequence, whose density is supposed to be a known function \(h\), the \(\varepsilon_ i\) are i.i.d. and independent of \(X=\{X_ i\}\). The process \(X\) is stationary and verifies an \(\alpha\)-mixing condition. He builds an estimator of \(f(x,p)\), the density of the vector \((X_ 1, X_ 2, \dots, X_ p)\), using a kernel \(K\) and the inverse Fourier transform of the noise density. Under certain decay conditions for the \(\alpha\)- mixing coefficient and smoothness assumptions for the function \(f\), he proves that: \[ \widehat f_ n(x;p) - f(x;p) = O \bigl( (\ln n/n)^ \gamma \bigr). \] The proof is based on an approximation lemma of mixing variables by independent ones, and on precise bounds for sums of independent random variables.