Integral approximation by kernel smoothing (Q726736)

The main results of this paper are explicit rates of convergence in probability for an integral approximation problem, and corresponding central limit theorems. Let \(X_1,\ldots,X_n\) be a sequence of i\(.\)i\(.\)d\(.\) \(\mathbb{R}^d\)-valued random variables, each with density \(f\), and \(\varphi:\mathbb{R}^d\mapsto\mathbb{R}\). The authors show that, under certain regularity conditions, \[ n^{-1}\sum_{i=1}^n\frac{\varphi(X_i)}{\widehat{f}^{(i)}(X_i)}-\int\varphi(x)dx=o_\mathbb{P}(n^{-1/2})\,, \] where \(\widehat{f}^{(i)}\) is defined by \[ \widehat{f}^{(i)}(x)=\frac{1}{(n-1)h^d}\sum_{j\not=i}K\left(h^{-1}(x-X_j)\right)\, \] for each \(x\in\mathbb{R}^d\), some \(d\)-dimensional kernel \(K\) and bandwidth \(h\). The significance of this result is a faster convergence rate than would be expected by applying the central limit theorem with \(\widehat{f}\) replaced by \(f\). Explicit bounds on the rate of convergence accompany this result. A further, more refined, approximation for \(\int\varphi(x)dx\) is also defined, by modifying the previous approximation so that the leading error term in its expansion vanishes asymptotically. Corresponding central limit theorems are also given. The authors also explore a natural generalisation of the above result to a wider class of integrands, and show that for an analogous approximation of an integral \(\int T(x,f(x))dx\), it is only for \(T\) of the form \(T(x,y)=\varphi(x)\) for which faster convergence rates can be achieved. An application to estimating linear functionals of a regression function in a nonparametric regression model is also given, which can be used, for example, in the estimation of Fourier coefficients in various models. This includes a central limit theorem in which the asymptotic variance does not depend on the regression function. Finally, simulations are used to explore the quality of the proposed approximations, to suggest a suitable choice of the bandwidth, and to explore the problem of increasing the dimension \(d\). The simulations also suggest that at least some of the assumptions employed in proving the main results of the paper can be relaxed in practice.