数学杂志  2017, Vol. 37 Issue (2): 340-346   PDF    
扩展功能
加入收藏夹
复制引文信息
加入引用管理器
Email Alert
RSS
本文作者相关文章
胡宏昌
弱相依半参数回归模型的加权小波估计的渐近正态性
胡宏昌     
湖北师范学院数学与统计学院, 湖北 黄石 435002
摘要:本文研究了半参数回归模型yi=X'iβ + gti)+ ei,i=1,2,…,n,其中{ei}为ψ -弱相依随机误差序列.利用小波估计的方法得到了参数、非参数的加权小波估计量.在相当一般的条件下,获得了这些小波估计量的渐近正态性,不仅推广了半参数回归模型的相应结果,而且在一定程度上统一了相依半参数回归模型的渐近正态性的理论.
关键词半参数回归模型    ψ-弱相依    小波估计    渐近正态性    
ASYMPTOTIC NORMALITY OF WEIGHTED WAVELET ESTIMATORS IN A SEMIPARAMETRIC REGRESSION MODEL WITH WEAKLY DEPENDENT ERRORS
HU Hong-chang     
School of Mathematics and Statistics, Hubei Normal University, Huangshi 435002, China
Abstract: Consider the following semiparametric regression model yi=X'iβ+g (ti)+ei, i=1, 2, …, n, where random errors{ei}are ψ-weakly dependent.Using the wavelet method, we obtain some estimators of the parametric component and the nonparametric component.Under general conditions, we investigate the asymptotic normality of these wavelet estimators.We not only generalize the corresponding conclusions of semiparametric regression models, but also unify asymptotic normal theorey of semiparametric regression models with dependent errors to some certain degree.
Key words: semiparametric regression model     ψ-weakly dependent     wavelet estimation     asymptotic normality    
1 引言

考虑如下半参数回归模型

$ y_i=X^{'}_i \beta+g(t_i)+e_i, i=1, 2, \cdots, n, $ (1.1)

其中$"'"$表示向量或矩阵的转置,随机设计点列$X_i=(x_{i1}, x_{i2}, \cdots, x_{id})^{'}$,非随机点列$t_i\in [0,1]$$\beta=(\beta_1, \beta_2, \cdots, \beta_d)^{'}\in \mathbb{R}^d$$d$维未知参数,$g(t)$是定义在$[0,1]$上的未知函数,$\{e_i\}$为零均值的$\psi-$弱相依(定义见下文)序列.

半参数回归模型自从Engle等人提出后(见文献[1]),取得了丰硕的研究成果.独立误差下的半参数回归模型的统计推断理论比较完善和系统(如文献[2-3]).对于各种相依误差下的半参数回归模型(1.1)及其特殊情形人们也进行了研究,并取得了丰富的成果,如:误差为鞅差的情形参见文献[4], NA误差的情形参见文献[5-6];文献[7]讨论了一般相依误差(包括混合误差)下非参数回归模型最小二乘估计的相合性;文献[8]考虑了一类$\psi-$弱相依误差的半参数回归模型,得到参数和非参数的广义最小二乘估计量(基于核方法)的渐近正态性.由于$\psi-$弱相依随机序列包含了Gaussian序列、各种混合序列、相伴序列、Bernoulli漂移、Markov链等(见文献[9-10]等)在内的相依序列,因此模型(1.1) 在一定程度上是各种误差情形下的半参数回归模型的较一般形式.我们试图寻找半参数回归模型的小波估计的统一推断理论,本文只研究加权小波估计量的渐近正态性.

2 $\psi-$弱相依序列和小波估计

为了下文的需要,先引入一些记号和给出两类函数.记向量$v=(v_1, v_2, \cdots, v_n)'\in \mathbb{R}^n$的范数为${\left\| v \right\|_p} = {(\sum\limits_{i = 1}^n | {v_i}{|^p})^{\frac{1}{p}}},1 \leqslant p < \infty $${\left\| v \right\|_\infty } = \mathop {\max }\limits_{1 \leqslant i \leqslant n} |{v_i}|$,以及矩阵$A=(a_{ij})_{n\times n}$的范数为$\|A\|_p=\max_{\|v\|_p>0}(\frac{\|Av\|_p}{\|v\|_p}), 1\leq p < \infty$${\left\| A \right\|_\infty } = \mathop {\max }\limits_{1 \leqslant i \leqslant n} \sum\limits_{j = 1}^n | {a_{ij}}|$.记$L^\infty(\mathbb{R}^n)$为空间$\mathbb{R}^n$上实值有界函数的集合,$L^\infty=\cup_{n=1}^\infty L^\infty(\mathbb{R}^n)$$\mathbb{R}^n$上的$l_1$范数为${\left\| {({u_1},{u_2}, \cdots ,{u_n})} \right\|_1} = \sum\limits_{n = 1}^n | {u_i}|$,函数$h:\mathbb{R}^n\rightarrow \mathbb{C}$的Lipschitz模为

$ {\text{Lip}} = \mathop {\sup }\limits_{u \ne v} \frac{{|h(u) - h(v)|}}{{{{\left\| {u - v} \right\|}_1}}}. $

又记$\mathbb{F}_n=\{h\in L^\infty(\mathbb{R}):\text{Lip}(h) < \infty, \|h\|_\infty=\sup_u|h(u)|\leq 1\}$$\mathbb{F}=\cup_{n=1}^\infty\mathbb{F}_n$上的两类函数为

$ \psi_1(h, \kappa, n, m)=\min(n, m)\text{Lip}(h)\text{Lip}(\kappa) $

$ \psi_2(h, \kappa, n, m)=4(n+m)\min\{\text{Lip}(h), \text{Lip}(\kappa)\}, $

其中$h, \kappa\in \mathbb{R}$.

定义2.1  [9]随机序列$\{\varepsilon_i, i\in\mathbb{N}\}$称为$(\theta, \mathbb{F}, \psi)$-弱相依(简称$\psi$-弱相依),如果存在当$r\rightarrow \infty$时递减到零的序列$\theta=\{\theta_r, r\in\mathbb{N}\}$和函数$\psi$,使得对任何$(i_1, i_2, \cdots, i_n)$$(j_1, j_2, \cdots, j_m)$

$ \left|\text{Cov}(h(\varepsilon_{i_1}, \varepsilon_{i_2}, \cdots, \varepsilon_{i_n}), \kappa(\varepsilon_{j_1}, \varepsilon_{j_2}, \cdots, \varepsilon_{j_m}))\right|\leq\psi(h, \kappa, n, m)\theta_r, $

其中$i_1\leq i_2\leq \cdots\leq i_n < i_n+r\leq j_1\leq j_2\leq\cdots\leq j_m$.

设有一个给定的刻度函数$\phi(x)\in \mathbb{S}_l$(阶为$l$的Schwartz空间),相伴$L^2(\mathbb{R})$的多尺度分析为$\{V_m\}$,其再生核为

$ {E_m}(t,s) = {2^m}{E_0}({2^m}t,{2^m}s) = {2^m}\sum\limits_{k \in \mathbb{Z}} \phi ({2^m}t - k)\phi ({2^m}s - k). $

$A_i$$[0,1]$上的分割且$t_i\in A_i, 1\leq i\leq n$.采用通常的估计方法,可以得到参数$\beta$和非参数$g(t)$的加权小波估计量分别为

$ \hat{\beta}_n=(\tilde{X}'\Psi^{-1}\tilde{X})^{-1}\tilde{X}'\Psi^{-1}\tilde{Y} $ (2.1)

$ {\hat g_n}(t) = {\hat g_0}(t,{\hat \beta _n}) = \sum\limits_{i = 1}^n {({y_i} - \tilde X{{\hat \beta }_n})} \int_{{A_i}} {{E_m}} (t,s)ds, $ (2.2)

其中$\Psi=Var(e)=(\psi_{ij})_{n\times n}$$n$阶正定矩阵,

$ \tilde{X}=(I-W)X, X=(x_{ij})_{n\times d}, W=(w_{ij})_{n\times n}, w_{ij}=\int_{A_j}E_m(t_i, s)ds, $
$ \tilde{Y}=(I-W)Y, Y=(y_1, y_2, \cdots, y_n)', e=(e_1, e_2, \cdots, e_n)', $
$ {\hat g_0}(t,\beta ) = \sum\limits_{i = 1}^n {({y_i} - \tilde X\beta )} \int_{{A_i}} {{E_m}} (t,s)ds. $
3 主要结果

本节将给出参数$\beta$和非参数$g(t)$的加权小波估计量的渐近正态性,为此先给出如下的一些基本假设条件:

(Ⅰ)存在光滑函数$f_j(t), t\in [0,1]$,使得

$ x_{ij}=f_j(t_i)+\eta_{ij}, 1\leq i\leq n, 1\leq j\leq d, $ (3.1)

其中$\{\eta_{ij}\}$独立同分布,且$E\eta_{ij}=0$$\{\eta_{ij}\}$$\{e_i\}$相互独立.

(Ⅱ)$g(\cdot), f_j(\cdot)\in \mathbb{H}^\alpha, \alpha>\frac{1}{2}$.

(Ⅲ)$g(\cdot), f_j(\cdot)$满足$\gamma$阶Lipschitz条件,$\gamma>0, 1\leq j\leq d$.

(Ⅳ)$\phi(\cdot)\in\mathbb{S}_l(l\geq\alpha)$$\phi$满足1阶Lipschitz条件且具有紧支撑,当$\xi\rightarrow 0$时,$|\hat{\phi}(\xi)-1|=O(\xi)$,其中$\hat{\phi}$$\phi$的Fourier变换.

(Ⅴ)$\mathop {\max }\limits_{1 \leqslant i \leqslant n} ({s_i} - {s_{i - 1}}) = O({n^{ - 1}})$,且$2^m=O(n^{\frac{1}{3}})$.

(Ⅵ)对于矩阵$\Psi=(Q'Q)^{-1}$假定:$\|\Psi\|_2=O(1), \|\Psi^{-1}\|_\infty=O(1)$$\mathop {\lim }\limits_{n \to \infty } ({n^{ - 1}}\eta '{\Psi ^{ - 1}}\eta ) = V$,其中$\eta=(\eta_{ij})_{n\times d}$$V$$d$阶正定矩阵;$\|W'\Psi^{-1}\eta_j\|=o_P(n^{-\frac{1}{2}})$.

注1 对于随机设计情形下的半参数回归模型的小波估计,假设条件(Ⅰ)-(Ⅴ)是相当弱的条件,如文献[4][6]等均使用这些条件.假设(Ⅵ)比文献[8]的假设4要弱,且去掉了文献[8]的假设2中的(b)和(c).

基于上面的基本假设,可以得到如下的主要结果.

定理3.1 如基本假设条件(Ⅰ)-(Ⅵ)成立,又设$\{e_i\}$$(\theta, \mathbb{F}, \psi_1)$$(\theta, \mathbb{F}, \psi_2)$-弱相依序列且$\theta_r=O(r^{-\rho})$$\rho\geq2$$\mathop {\sup }\limits_i E|{e_i}{|^{2 + \delta }} < \infty (0 < \delta < 1)$, 则对于$\alpha>\frac{3}{2}, \gamma\geq\frac{1}{3}$,有

$ n^{\frac{1}{2}}(\hat{\beta}_n-\beta)\longrightarrow_DN(0, V^{-1}), n\rightarrow \infty, $

其中$"D"$表示依分布收敛.

定理3.2 如满足定理3.1的所有条件,则

$ \frac{\hat{g}_n(t)-\bar{g}(t)}{(Var(\hat{g}_n(t)))^{\frac{1}{2}}}\longrightarrow_DN(0, 1), \forall t\in [0,1], n\rightarrow \infty, $

其中$\bar g(t) = \sum\limits_{i = 1}^n {\int_{{A_i}} {{E_m}} } (t,s)dsg({t_i})$$\text{Var}({\hat g_n}(t)) \to \sum\limits_{i = 1}^n {\sum\limits_{j = 1}^n {{\psi _{ij}}} } \int_{{A_j}} {{E_m}} (t,s)ds\int_{{A_i}} {{E_m}} (t,s)ds$.

注2 当误差$\{e_i\}$为Gaussian序列、各种混合序列、相伴序列、Bernoulli漂移、Markov链等时,上述结论显然成立.由此可知,本文的结论不仅推广了独立误差情形的半参数回归模型的相应结论,而且在一定程度上统一了相依半参数回归模型的渐近正态性的理论.

4 主要结果的证明

为了证明本文的主要结果,我们先给出若干必要的引理,其中引理4.1和4.2是小波估计常用的两个结论(见文献[11]等),引理4.3参见文献[6],引理4.4是仿照文献[8]的方法而得到.下文中的$C$表示常数,且在不同的地方可以取不同的值.

引理4.1 [11] 若条件(Ⅰ)-(Ⅳ)成立,则

$ \mathop {\sup }\limits_t |{f_j}(t) - \sum\limits_{i = 1}^n {(\int_{{A_i}} {{E_m}} (t,s)ds){f_j}({t_i})| = O({n^{ - \gamma }}) + O({\tau _m})} $

$ \mathop {\sup }\limits_t |g(t) - \sum\limits_{i = 1}^n {(\int_{{A_i}} {{E_m}} (t,s)ds)g({t_i})| = O({n^{ - \gamma }}) + O({\tau _m})} , $

其中

$ \tau_m=\left\{ \begin{aligned} 2^{-m(\alpha-\frac{1}{2})}, \frac{1}{2} < \alpha < \frac{3}{2}, \\ \frac{\sqrt{m}}{2^m}, \alpha=\frac{3}{2}, \\ 2^{-m}, \alpha>\frac{3}{2}. \end{aligned} \right. $

引理4.2  [11] 若条件(Ⅳ)成立,则

$ \mathop {\sup }\limits_{0 \le s \le 1} {\mkern 1mu} |{E_m}(t,s)| = O({2^m});\mathop {\sup }\limits_t {\mkern 1mu} \int_0^1 | {E_m}(t,s)|ds \le C;\int_0^1 {{E_m}} (t,s)ds \to 1,n \to \infty . $

引理4.3 [6]若条件(Ⅲ)(Ⅳ)成立,且$E\eta_{1j}^2 < \infty$,则

$ \mathop {\sup }\limits_t {\mkern 1mu} |\sum\limits_{i = 1}^n {{\eta _{ij}}} \int_{{A_i}} {{E_m}} (t,s)ds| = O({n^{ - \frac{1}{3}}}\log n),a.s.,n \to \infty . $

引理4.4 如基本假设条件(Ⅰ)-(Ⅵ)成立,则依概率有

$ \mathop {\lim }\limits_{n \to \infty } {\mkern 1mu} {n^{ - 1}}(\tilde X'{\Psi ^{ - 1}}\tilde X) = V,n \to \infty . $

 令$\tilde{\eta}_i=\eta_i-\sum_{i=1}^n\eta_j\int_{A_j}E_m(t_i, s)ds$$F_j=(f_j(t_1), f_j(t_2), \cdots, f_j(t_n))'$$\tilde{F}_j=(I-W)F_j$${\tilde f_j}({t_i}) = {f_j}({t_i}) - \sum\limits_{l = 1}^n {\int_{{A_l}} {{E_m}} } ({t_i},s)ds{f_j}({t_l})$, 则$n^{-1}(\tilde{X}'\Psi^{-1}\tilde{X})$的第$(i, j)$元素为

$ \begin{align} {{n}^{-1}}{{({\tilde{X}}'{{\Psi }^{-1}}\tilde{X})}_{ij}}={{n}^{-1}}({{{{\tilde{X}}'}}_{i}}{{\Psi }^{-1}}{{{\tilde{X}}}_{j}}) \\ ={{n}^{-1}}({{{{\tilde{F}}'}}_{i}}{Q}'Q{{{\tilde{F}}}_{j}})+{{n}^{-1}}({{{{\tilde{\eta }}'}}_{i}}{Q}'Q{{{\tilde{\eta }}}_{j}})+2{{n}^{-1}}({{{{\tilde{F}}'}}_{i}}{Q}'Q{{{\tilde{\eta }}}_{j}}) \\ ={{I}_{1}}+{{I}_{2}}+2{{I}_{3}}. \\ \end{align} $ (4.1)

由假设(Ⅵ)及引理4.1,得

$ \begin{align} {{I}_{1}}={{n}^{-1}}({{{{\tilde{F}}'}}_{i}}{Q}'Q{{{\tilde{F}}}_{j}})={{n}^{-1}}\|Q{{{\tilde{F}}}_{j}}\|_{2}^{2}\le {{n}^{-1}}\|Q\|_{2}^{2}\|{{{\tilde{F}}}_{j}}\|_{2}^{2} \\ \le \|{{\Psi }^{-1}}{{\|}_{\infty }}(O({{n}^{-2\gamma }})+O(\tau _{m}^{2}))\to 0. \\ \end{align} $ (4.2)

由引理4.3容易得到$\|W\eta_j\|_2=O_P(n^{\frac{1}{6}}\log n)$,于是由此及假设(Ⅵ),得

$ |n^{-1}\eta'_iW'\Psi^{-1}W\eta_j|=n^{-1}\|QW\eta_j\|^2_2\leq \|\Psi^{-1}\|_\infty n^{-1}\|W\eta_j\|^2_2\longrightarrow_P 0 $ (4.3)

$ |n^{-1}\eta'_i\Psi^{-1}W\eta_j|\leq \|\Psi^{-1}\|_\infty \|W\eta_j\|_2\|n^{-1}\eta'_i\|_2\leq Cn^{-\frac{2}{3}}\log^2n\|\eta'_i\|_2\longrightarrow_P 0. $ (4.4)

因而由(4.3)和(4.4)式可以得到

$ I_2=n^{-1}\tilde{\eta}'_i\Psi^{-1}\tilde{\eta}_j =n^{-1}\eta'_i(I-W)'\Psi^{-1}(I-W)\eta_j\nonumber \\ =n^{-1}\eta'_i\Psi^{-1}\eta_j+n^{-1}\eta'_iW'\Psi^{-1}W\eta_j-2n^{-1}\eta'_i\Psi^{-1}W\eta_j\longrightarrow_P V_{ij}. $ (4.5)

由假设(Ⅵ),引理4.1,得

$ I_3\leq \|\tilde{F}'_i\|_2\|\Psi^{-1}\|_2\cdot n^{-1}\|(I-W)\eta_j\|_2\longrightarrow_P 0. $ (4.6)

由(4.1), (4.2), (4.5)和(4.6)式即可证明该引理.

定理4.1的证明 由(2.1)式得

$ n^{\frac{1}{2}}(\hat{\beta}_n-\beta)=n^{\frac{1}{2}}(\tilde{X}'\Psi^{-1}\tilde{X})^{-1} \tilde{X}'\Psi^{-1}(\tilde{X}\beta+\tilde{e})-n^{\frac{1}{2}}\beta\nonumber \\ =n^{\frac{1}{2}}(\tilde{X}'\Psi^{-1}\tilde{X})^{-1} \tilde{X}'\Psi^{-1}(I-W)e. $ (4.7)

$B=(B_1, B_2, \cdots, B_d)'=(\tilde{X}'\Psi^{-1}\tilde{X})^{-1}$$C'_n=(c_{1n}, c_{2n}, \cdots, c_{nn})=B_i\tilde{X}'\Psi^{-1}(I-W)$,则$n^{\frac{1}{2}}(\hat{\beta}_n-\beta)$的第$i$个分量为$\sqrt{n}S_n=C'_ne=\sum_{i=1}^nc_{ni}e_i$.注意到

$ \sigma^2_n=Var(C'_ne)=C'_nVar(e)C_n\nonumber \\ =B_i\tilde{X}'\Psi^{-1}(I-W)\sigma^2\Psi (I-W)'\Psi^{-1}\tilde{X}B'_i\nonumber \\ =\sigma^2B_i\tilde{X}'\Psi^{-1}\tilde{X}B'_i+\sigma^2B_i\tilde{X}'\Psi^{-1}W\Psi W'\Psi^{-1}\tilde{X}B'_i-2\sigma^2B_i\tilde{X}'W'\Psi^{-1}\tilde{X}B'_i. $ (4.8)

考虑$\tilde{X}'\Psi^{-1}W\Psi W'\Psi^{-1}\tilde{X}$的第$(i, j)$元素,由引理4.1和假设(Ⅵ),得

$ \begin{gathered} |{{\tilde X'}_i}{\Psi ^{ - 1}}W\Psi W'{\Psi ^{ - 1}}{{\tilde X}_j}| \leqslant {\left\| {{{\tilde X'}_i}{\Psi ^{ - 1}}W} \right\|_2}{\left\| \Psi \right\|_2}{\left\| {W'{\Psi ^{ - 1}}{{\tilde X}_j}} \right\|_2} = {\left\| \Psi \right\|_2}\left\| {W'{\Psi ^{ - 1}}{{\tilde X}_j}} \right\|_2^2 \hfill \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \leqslant C{\left\| \Psi \right\|_2}(\left\| {W'{\Psi ^{ - 1}}{{\tilde F}_j}} \right\|_2^2 + \left\| {W'{\Psi ^{ - 1}}{{\tilde \eta }_j}} \right\|_2^2) \hfill \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \leqslant C{\left\| \Psi \right\|_2}(\left\| {W'{\Psi ^{ - 1}}{{\tilde F}_j}} \right\|_2^2 + \left\| {W'{\Psi ^{ - 1}}{\eta _j}} \right\|_2^2 + \left\| {W'{\Psi ^{ - 1}}W{\eta _j}} \right\|_2^2) \hfill \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; = {o_P}(n). \hfill \\ \end{gathered} $ (4.9)

考虑$\tilde{X}'W'\Psi^{-1}\tilde{X}$的第$(k, j)$元素的平方

$ \begin{gathered} |{{\tilde X'}_i}W'{\Psi ^{ - 1}}{{\tilde X}_j}{|^2} \leqslant \left\| {QW{{\tilde X}_i}} \right\|_2^2\left\| {Q{{\tilde X}_j}} \right\|_2^2 \leqslant {{\tilde X'}_j}{\Psi ^{ - 1}}{{\tilde X}_j}{\left\| \Psi \right\|_2}\left\| {W({{\tilde F}_j} + {{\tilde \eta }_j})} \right\|_2^2 \hfill \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \leqslant {n^{ - 1}}{{\tilde X'}_j}{\Psi ^{ - 1}}{{\tilde X}_j}{\left\| \Psi \right\|_2}({n^{1 - 2\gamma }} + n\tau _m^2 + n{({n^{ - \frac{1}{3}}}\log n)^2}) \hfill \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \leqslant C{({n^{\frac{1}{6}}}\log n)^2}. \hfill \\ \end{gathered} $ (4.10)

于是由(4.8)-(4.10)式可得

$ n^{-1}\sigma^2_n \longrightarrow_P B'_iVB_i=\Delta^2. $ (4.11)

利用文献[8]的方法容易证明

$ S_n=n^{-\frac{1}{2}}\sum^n_{i=1}c_{ni}e_i \longrightarrow_D N(0, \Delta^2). $ (4.12)

由(4.7), (4.11)和(4.12)式即可得到定理1.

定理4.2的证明 注意到

$ {\hat g_n}(t) - \sum\limits_{i = 1}^n g ({t_i})\int_{{A_i}} {{E_m}} (t,s)ds = (\beta - {\hat \beta _n})'\sum\limits_{i = 1}^n {{{\tilde X}_i}} \int_{{A_i}} {{E_m}} (t,s)ds + \sum\limits_{i = 1}^n {{e_i}} \int_{{A_i}} {{E_m}} (t,s)ds = {I_1} + {I_2}. $ (4.13)

由定理4.1和引理4.3可得

$ \sum\limits_{i = 1}^n {{{\eta '}_i}} (\beta - {\hat \beta _n})\int_{{A_i}} {{E_m}} (t,s)ds{n^{ - 1}} = {O_P}({n^{ - \frac{1}{2}}} \cdot {n^{ - \frac{1}{3}}}\log n) = {O_P}({n^{ - \frac{5}{6}}}\log n). $ (4.14)

从而有

$ \begin{gathered} {\text{Var}}({{\hat g}_n}(t)) = {\text{Var}}\left( {\sum\limits_{i = 1}^n {{{\tilde X'}_i}} (\beta - {{\hat \beta }_n})\int_{{A_i}} {{E_m}} (t,s)ds + \sum\limits_{i = 1}^n {{e_i}} \int_{{A_i}} {{E_m}} (t,s)ds} \right) \hfill \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\; = {\text{Var}}\left( {\sum\limits_{i = 1}^n {{{\tilde \eta '}_i}} (\beta - {{\hat \beta }_n})\int_{{A_i}} {{E_m}} (t,s)ds + \sum\limits_{i = 1}^n {{e_i}} \int_{{A_i}} {{E_m}} (t,s)ds} \right) \hfill \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \to {\text{Var}}\left( {\sum\limits_{i = 1}^n {{e_i}} \int_{{A_i}} {{E_m}} (t,s)ds} \right) \hfill \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; = \sum\limits_{i = 1}^n {\sum\limits_{j = 1}^n {{\psi _{ij}}} } \int_{{A_j}} {{E_m}} (t,s)ds\int_{{A_i}} {{E_m}} (t,s)ds. \hfill \\ \end{gathered} $ (4.15)

由定理4.1和引理4.1-4.2容易得到

$ I_1=o_P(1). $ (4.16)

从而由(4.13)-(4.16) 式知,要证明定理4.2成立,只需证明下式成立即可

$ {\left( {Var(\sum\limits_{i = 1}^n {{e_i}} \int_{{A_i}} {{E_m}} (t,s)ds)} \right)^{ - \frac{1}{2}}}\sum\limits_{i = 1}^n {{e_i}} \int_{{A_i}} {{E_m}} (t,s)ds{ \to _D}N(0,1). $ (4.17)

类似于(4.12)式的证明方法,我们很容易得到(4.17)式,在此略.

参考文献
[1] Engle R F, Granger W J, Rice J, Weiss A. Semiparametric estimates of the relation between weather and electricity sales[J]. J. American Stati. Assoc., 1986, 80: 310–319.
[2] You J, Chen G. Testing heteroscedasticity in partially linear regression models[J]. Stati. Prob. Lett., 2005, 73: 61–70. DOI:10.1016/j.spl.2005.03.002
[3] Härdle W, Liang H, Gao J. Partially linear models[M]. New York: Physica-Verlag Heidelberg, 2000.
[4] 蔡择林, 胡宏昌. 半参数回归模型小波估计的弱收敛速度[J]. 数学杂志, 2011, 31(2): 331–340.
[5] Baek J I, Liang H Y. Asymptotics of estimators in semi-parametric model under NA samples[J]. J. Stati. Plan. Infer., 2006, 136: 3362–3382. DOI:10.1016/j.jspi.2005.01.008
[6] Hu H C, Wu L. Convergence rate of wavelet estimators in semiparametric regression models under NA samples[J]. Chinese Annals Math., Series B, 2012, 33(4): 609–624. DOI:10.1007/s11401-012-0718-z
[7] Hu S H. Consistency for the least squares estimator in nonlinear regression model[J]. Stati. Prob. Lett., 2004, 67: 183–192. DOI:10.1016/j.spl.2003.11.020
[8] Hwang E, Shin D. Semiparametric estimation for partially linear models with ψ -weak dependent errors[J]. J. Korean Stati. Soc., 2011, 40: 411–424. DOI:10.1016/j.jkss.2011.01.002
[9] Doukhan P, Louhichi S. A new weak dependence condition and applications to moment inequalities[J]. Stoch. Proc. Their Appl., 1999, 84: 313–342. DOI:10.1016/S0304-4149(99)00055-1
[10] Dedecker J, Doukhan P, Lang G, Leon J R, Louhichi S, Prieur C. Weak dependence:with examples and applications[M]. New York: Springer-Verlag, 2007.
[11] Antoniads A, Gregoire G, Mckeague IW. Wavelet methods for curve estimation[J]. J. American Stati. Assoc., 1994, 89: 1340–1353. DOI:10.1080/01621459.1994.10476873