Convergence rate of kernel regression estimation for time series data when both response and covariate are functional (Q2189762): Difference between revisions

This paper addresses an issue, Functional Data Analysis (FDA), that is increasingly being researched and applied in the statistical community. Let's see a brief introduction to the subject of this paper. Let \(F\) be a set \(d:F\times F\to[0,+\infty)\) a semi-metric over \(F\), that is 1) \(d(x,x)=0\) for all \(x\in F\), 2) \(d(x,z)\leq d(x,y)+d(y,z)\) for all \((x,y,z)\in F\times F\times F\). It should be noted that there is no agreement on the definition of semi-metric in the specialized literature. The above definition was taken from a book that the authors of this paper mention with reference to semi-metric. Let \((\Omega,F,P)\) be a probability space, \(F\) is a semi-metric space as defined above, \(H\) is a separable Hilbert space, \(X:\Omega\to F\), \(Y:\Omega\to H\), and \(\varepsilon:\Omega\to H\) are random fields, \(m:F\to H\) is a functional operator such that \[ Y=m(X)+\varepsilon. \] The functional \(m\) is partly unknown and it is desired to estimate it, based on observations of \(X\) (explanatory functional variable) and \(Y\) (response functional variable). It is assumed that the observation process is \(\alpha\)-mixing. Now, let's look at its definition. For every \(i\in\mathbb{N}\) are \(X_{i}:\Omega\to F\), and \(Y_{i}: Ω\to H\) random fields. For each \(1\leq j\leq k\) be \(F_{j}^{k}\) the \(\sigma\)-field generated by \[ G_{j}^{k}=\{X_{i}^{-1}(B)\cap Y_{i}^{-1}(C)/j\leq i\leq k,\quad B\in\sigma(F),\quad C\in\sigma(H)\}, \] where \(\sigma(F)\) and \(\sigma(H)\) are the Borel \(\sigma\)-field of \(F\) and \(H\) respectively. For every \(k\) and \(n\) positive integers let \(F_{k+n}^{\infty}\) be the \(\sigma\)-field generated by \[ G_{k+n}^{\infty}=\{X_{i}v(B)\cap Y_{i}^{-1}(C)/k+n\leq i,\quad B\in\sigma(F),\quad C\in\sigma(H)\}. \] Let \(\alpha:\mathbb{N}\to[0,+\infty)\) be a function defined by \[ \alpha(n)=\sup_{k}\sup_{A\in F_1^{k}}\sup_{B\in F_{k+n}^{\infty}}|P(A\cap B)-P(A)P(B)|. \] The \((X_{i},Y_{i})_{i\in\mathbb{N}}\) stochastic process is said to be \(\alpha\)-mixing of order \(b>0\) if there is \(C>0\) such that \[ \alpha(n)\leq C\cdot n^{-b}. \] In the practical applications of disciplines such as economics, finance and others there are numerous examples of such stochastic processes. Now let's be \(n\geq 1\) and suppose that \(\{(X_{i},Y_{i})/1\leq i\leq n\}\) is a sample of an \(\alpha\)-mixing \((X_{i},Y_{i})_{i\in\mathbb{N}}\) process of order \(b>0\) such that \((X_{i},Y_{i})\) has the same distribution as \((X,Y)\). More precisely: \[ Y_{i}=m(X_{i})+\varepsilon_{i},\quad 1\leq i\leq n \] where \(E(\varepsilon_{i}|X_{i})=0\) a.e. The \(m\) estimator proposed in this paper is \[ m_{n}((X_{i},Y_{i})_{1\leq i\leq n})=((\sum_{i=1}^nY_{i}K(h_{n}^{-1}\cdot d(X_{i},X)))/(\sum_{i=1}^nK(h_{n}^{-1}\cdot d(X_{i},X)))), \] where \((h_{n})\) is a sequence of positive real numbers such that \(h_{n}\to_{n\to\infty}0\), and \(K\) is a function (kernel function) with certain properties. Under several conditions the authors probe two asymptotic properties of the estimator defined above. These properties are pointwise almost complete convergence and uniform almost complete convergence, which have already been defined by other authors in recent papers and which in this paper are somewhat modified. They can be seen as extensions of Borel Cantelli's lemma. A simulation study is used to test the performance of the proposed estimator for finite samples. The authors reveal the website from which the R-language routines that implement the study can be obtained free of charge. The list of references is extensive and well-updated. Proof of results (lemmas and theorems) are rigorously presented and can be studied with adequate background in Functional Analysis and Statistics.