Convergence of linear functionals of the Grenander estimator under misspecification

Abstract: Under the assumption that the true density is decreasing, it is well known that the Grenander estimator converges at rate

n^{1 / 3}

if the true density is curved [Sankhy={a} Ser. A 31 (1969) 23-36] and at rate

n^{1 / 2}

if the density is flat [Ann. Probab. 11 (1983) 328-345; Canad. J. Statist. 27 (1999) 557-566]. In the case that the true density is misspecified, the results of Patilea [Ann. Statist. 29 (2001) 94-123] tell us that the global convergence rate is of order

n^{1 / 3}

in Hellinger distance. Here, we show that the local convergence rate is

n^{1 / 2}

at a point where the density is misspecified. This is not in contradiction with the results of Patilea [Ann. Statist. 29 (2001) 94-123]: the global convergence rate simply comes from locally curved well-specified regions. Furthermore, we study global convergence under misspecification by considering linear functionals. The rate of convergence is

n^{1 / 2}

and we show that the limit is made up of two independent terms: a mean-zero Gaussian term and a second term (with nonzero mean) which is present only if the density has well-specified locally flat regions.

Review text

Convergence rates of Grenander estimation of the true misspecified density are treated. Let \({\mathcal F}\) be the class of decreasing densities on \(\mathbb{R}_+\). Let \(X_1,\dots,X_n\) be an i.i.d. sequence of random variables with density \(f_0\in{\mathcal F}\). Define the KL projection of \(f_0\) as \[ \widehat f_0=\text{argmin}_{g\in{\mathcal F}}\int^\infty_0 f_0(x)\log{f_0(x)\over g(x)}\,dx \] and the Grenander estimation of \(f_0\) as \[ \widehat f_n= \text{argmin}_{g\in{\mathcal F}} \int^\infty_0 \log g(x)\,dF_n(x), \] where \(F_n\) is the empirical distribution function. Let \(F_0(x)= \int^x_0 f_0(u)\,du\) and \(\widehat F_0\) be the least concave majorant of \(F_0\). Put \({\mathcal M}= \{x\geq 0: \widehat F_0(x)> F(x)\}\) and \({\mathcal W}=\{x\geq 0:\widehat F_0(x)= F(x)\}\). Let \([a, b]\) be the largest interval containing \(x_0\) such that \(\widehat F_0(x)\) is linear on the interval. Let \(\text{gren}_{[a,b]}(g)\) be the (left) derivative of the least concave majorant \(g\) restricted to \([a, b]\). Further, let \(U\) denote a standard Brownian bridge process on \([0, 1]\), \(U_{F_0}(x)= U(F_0(x))\) and \(U^{\text{mod}}_{F_0}(u) = U_{F_0}(u)\) if \(u\in [a,b]\cap{\mathcal W}\), \(U^{\text{mod}}_{F_0}(u) = -\infty\) if \(u\in[a, b]\cap{\mathcal M}\). The author proves that as \(n\to\infty\) \[ \sqrt{n}(\widehat f_n(x_0)-\widehat f_0(x_0))\to \text{gren}_{[a,b]}(U^{\text{mod}}_{F_0}(x_0)) \] and, in addition, if \([a, b]\cap{\mathcal W}= \{a, b\}\), \[ \sqrt{n}(\widehat f_n(x_0)-\widehat f_0(x_0))\to \sigma Z, \] where \(Z\) is a standard normal random variable and \(\sigma\) is some constant defined by \(\widehat f_0(x_0)\). The author also shows \(\sqrt{n}\)-convergence of linear functionals of the type \[ \hat \mu_n(g) = \int_0^\infty g(x) \hat f_n(x) dx \] and \(\sqrt{n}\)-convergence of the entropy functional in the misspecified case.

Recommendations

Cites work

Cited in

(17)

Describes a project that uses

Uses Software

This page was built for publication: Convergence of linear functionals of the Grenander estimator under misspecification

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2249843)