Density estimation by truncated wavelet expansion (Q1293840)

The paper under review deals with the estimation of a density function using a sample of i.i.d. random variables \(X_1,\dots,X_n\), having a common density \(f\). The method used is the approximation by a wavelet basis. The main difference to the paper of \textit{G. Kerkyacharian} and \textit{D. Picard} [ibid. 18, No. 4, 327-336 (1993; Zbl 0793.62019)] is that here the author uses a truncated expansion instead of an infinite one. Let have for \(f\) the expansion \[ f(x)=\sum_{k=-\infty}^{\infty}\alpha_{j,k}\varphi_{j,k}(x)+\sum_{l=j}^{\infty}\sum_{k=-\infty}^{\infty}\beta_{l,k}\psi_{l,k}(x), \] \[ \text{where} \varphi_{j,k}(x)=2^{j/2}\varphi(2^{j/2}x-k),\;k\in Z,\;\psi_{l,k}(x)=2^{l/2}\psi(2^{l/2}x-k),\;l\geq j,\;j\in Z. \] The functions \(\varphi\) and \(\psi\) are the scaling and wavelet, respectively. In the present paper the author defines the truncated version of the projection estimator, namely: \[ \hat{f} _{j(n),k(n)}(x)=\sum_{|k|\leq k(n)} \hat{\alpha}_{j(n),k}\varphi_{j(n),k}(x), \] where \(\hat{\alpha}_{j,k}\) is the empirical estimator of \(\alpha_{j,k}\), in the above cited paper \(k(n)=\infty\). The optimal choice for \(j(n)\) gives that \(2^{j(n)}\simeq (2sn)^{1/(1+2s)}\) if \(f\) belongs to the Besov space \(B_{s,2,q}\) with parameters \(s,p,q\), and for \(k(n)\) is \(O(n^{1/(1+2s)})\) if \(f\) has bounded support, and \(O(n^{\frac{2(r+s)+1}{2r(2s+1)}})\) otherwise, where \(r\) is such that \(E_f[|X|^r]<\infty\). The main result is the following M.I.S.E. evaluation: \[ M_{2}n^{-2s/(2s+1)}\leq E_f\|\hat{f} _{j(n),k(n)}-f\|^2_{L^2}\leq M_1n^{-2s/(2s+1)}. \]