<i>L_p</i>–Estimators as Estimates of a Parameter of Location for a Sharp–pointed Symmetric Density (Q4702170)

The problem of location parameter \(\theta_0\) estimation is considered in the irregular case when the density \(f\) of the observations \(x_j\) is symmetric, continuous at \(\theta_0\), positive, non-increasing in \((\theta_0,+\infty)\) and has a cusp at \(\theta_0\), i.e. \[ \lim_{a\to 0+} a^q(f(\theta_0)-f(\theta_0+a))=\gamma_0>0. \] The author considers an \(L_p\) estimator \(\hat\theta_n=\hbox{arg inf}_{\theta\in R}\sum_{j=1}^n| x_j-\theta| ^p\) for some \(p\), where \(x_j\) are i.i.d. with density \(f(x)\). It is proved that for \(0<p<1/2\) \(\hat\theta_n\) converges to \(\theta_0\) with the rate \(n^{1/(2q+1)}\), and if \(q=1/2\) and \(p=1/2\) the rate is \((n\log n)^{1/2}\). For \(f(x)=C \exp(-| x-\theta| ^q)\) these rates are the rates of MLE convergence (best possible). The proof is based on general results of the author on M-estimates in \(R^d\) proved in this paper. The asymptotic distribution of normalized \(\hat\theta_n\) is described in terms of arguments of the minimum of a process \(c_1| \theta| ^s+c_2 B_{(2p+1)/2}(\theta)\), where \(c_1\), \(c_2\), \(s\) are some constants, and \(B_r\) is a fractional Brownian motion of order \(r\).