Rate of convergence of the empirical Radon transform (Q1209609): Difference between revisions

Let \(X\) be a \(d\)-variate random variable. Consider the family \(\{\theta\cdot X\}\) of the \(X\) projections onto directions determined by unit vectors \(\theta\in R^ d\). Assume that \(X\) has the density \(f(x)\). Denote by \(f_ \theta(x)\) the density of \(\theta\cdot X\) which is called the Radon transform of \(f(x)\). The author deals with asymptotic properties of the family of the kernel estimators \[ \widehat f_ \theta(u)= (nh)^{-1} \sum^ n_{j=1} K((u-\theta\cdot X_ j)/h),\quad -\infty< u<\infty, \] where \(K(.)\) is a kernel, \(h\) is a bandwidth and \(X_ j\) are i.i.d., \(X_ j\overset {d} = X\). In particular, the asymptotics of the functional \(D=\sup_ \theta|\widehat f_ \theta(0)- f_ \theta(0)|\) and the lower and upper sequences for \(D_ n=\sup_{\theta,u}|\widehat f_ \theta(u)- f_ \theta(u)|\) are established. In addition, the following estimator of \(f(x)\) based on the indirect observations \(\theta\cdot X_ j\), \[ \widehat f(x)= -(8\pi^ 2 n h^ 3)^{-1} \sum^ n_{j=1} \int_{|\theta|= 1} K''(\theta\cdot(x-X_ j)/h)d\theta \] is suggested. Under some restrictions the rate of convergence of \(E\|\widehat f- f\|^ 2\) to zero is estimated.