Wavelet-based random densities (Q5943409)

Let \(\bigl\{ \psi_{jk} (x) = 2^{j/2}\psi(2^j x -k), (j, k) \in Z\times Z \bigr\}\) be a wavelet basis for the \(L_2\)-space. A wavelet based random density \(f(x)\) generated by \(\psi\) is defined by \(f(x) = \left(\sum_{j,k \in Z^2} d_{jk} \psi_{jk}(x)\right)^2\), where the coefficients \(d_{jk}\) are random up to the normalization constraint \(\sum_{j,k} d_{jk} =1.\) The authors present an algorithm for generation of random variables \(d_{jk}.\) They found the expectation of \(f(x)\), \[ Ef(x) = \lambda {\mathbf 1}(0\leq x \leq 1) +(1 - \lambda)^2 \sum_{j\geq 0}\biggl({\lambda\over 2}\biggr)^j \sum_k \psi_{jk}^2(x). \] They present some properties of random densities via wavelets. For example, let \(C^\alpha ([0, 1])\), \(0< \alpha <1\), be the Hölder space on \([0, 1]\), that is \[ C^\alpha ([0, 1]) = \left\{ f\in L^\infty([0, 1])\colon \sup_{x,h} {|f(x+h) -f(x)|\over |h |^\alpha} <\infty \right\}. \] Let the random density be generated by a sequence of i.i.d. random variables \(\{ u_{ik}\}\) such that \(Eu_{jk} =\lambda.\) Then \(f \in C^\alpha ([0, 1])\), \(\alpha = {1\over 2} \log_2 {1\over \lambda}\), with probability \(1.\) The authors propose an algorithm for generation of random variables with random density. Two application of random densities to minimax and \(\Gamma\)-minimax rules are considered.