On the Whittle estimator for linear random noise spectral density parameter in continuous-time nonlinear regression models (Q1984649): Difference between revisions

This paper deals with an estimation problem in a non-linear regression model in continuous time series. Noise is assumed to be a linear Lévy process. A brief description of the paper follows. Let's consider the following regression model. Let \((\Omega,F,P)\) be a complete probability space, \(L=(L(t))_{t\in\mathbb{R}}\) is a two-sided Lévy process, and \(a:\mathbb{R}\to[0,+\infty)\) is a function such that for each \(t\in\mathbb{R}\) is defined the random variable \[ \varepsilon(t)=\int_{\mathbb{R}}a(t-s)\cdot L(ds), \] For each $t\geq0$, $X(t)$ is a random variable defined over $\Omega$ such that \[ X(t)=g(t,\beta_0)+\varepsilon(t), \] where $g:(-\gamma,+\infty)\times U_{\gamma}\to\mathbb{R}$ is a continuous function, $\gamma$ is a positive real number, $U_{\gamma}$ is an open subset of $\mathbb{R}^{q}$, $q\geq 1$, defined by \[ U_{\gamma}=\bigcup_{||e||\leq 1}(U+\gamma e), \] where $U\subset\mathbb{R}^{q}$ is an open convex set, and $\beta_0\in U$ is an unknown parameter to be estimated. Let's see what is meant by two-sided Levy's process. A Lévy process, $L^1=(L^1(t))_{t\geq 0}$, is a stochastic process that satisfies the following properties: $L^1(t)$ is a random variable for all $t\geq 0$; $L^1(0)=0$ a.e.; $L^1$ has independent increments from the past, i.e. $L^1(s)-L^1(t)$ is independent from $L(u)$ if $0\leq u\leq t<s$; $L^1$ has stationary increments, i.e. $L^1(s)-L^1(t)$ and $L^1(s+u)-L^1(t+u)$ if $0\leq t<s$ and $u>0$ have the same distribution; $L^1$ is continuous in probability, i.e. $P(|L^1(s)-L^1(t)|>\varepsilon)\to_{s\to t}0$ if $\varepsilon>0$ and $0\leq s,t$; sample-paths are right-continuous with left limits (cádlág), i.e. for each $\omega\in\Omega$ \[ \lim_{s\downarrow t}L^1(s)(\Omega)=L^1(t)(\Omega),\text{ and there is }\lim_{s\uparrow t}L^1(s)(\Omega). \] There is a borelian Radon measure on $(\mathbb{R},B_1)$, $\Pi$, such that $\Pi({0})=0$, and \[ \int_{\mathbb{R}}\min(1,u^2)\cdot\Pi(du)<\infty, \] ($B_1$ is the Borel $\sigma$-field of $\mathbb{R}$). Let $a\in\mathbb{R}$, $b\geq 0$, $\tau:\mathbb{R}\to\mathbb{R}$ be defined for each $u\in\mathbb{R}$ by \[ \tau(u)=u\text{ if }|u|\leq 1 \quad (u/(|u|))\text{ if }|u|>1 \] and $\kappa:\mathbb{R}\to\mathbb{C}$ defined for each $z\in\mathbb{R}$ by \[ \kappa(z)=iaz-(1/2)bz^2+\int_{\mathbb{R}}(\exp(izu)-1-iz\tau(u))\cdot\Pi(du). \] The triple $(a,b,\Pi)$ is called a characteristic triplet of the Lévy process $L^1=(L^1(t))_{t\geq 0}$ if \[ \ln(E(\exp(izL^1(t))))=t\kappa(z). \] Be $L^2=(L^2(t))_{t\geq 0}$, the stochastic process defined by $L^2(t)=-L^1(t)$, $t\geq 0$. It can be seen that $L^2$ is a Lévy process with a characteristic triplet given by $(-a,b,\Pi)$ where $\Pi$ is given by $\Pi(B)=\Pi(-B)$, $B\in B_1$. In this case, the two-sided Levy's process $L=(L(t))_{t\in\mathbb{R}}$ is defined as \[ L(t)=L^1(t)\text{ if }t\geq 0\quad L^1(-t)\text{ if }t<0. \] Let $\{X(t) / 0\leq t\leq T\}$ be observations of the $X=(X(t))_{t\geq 0}$ process. Let $U$ be the closure of $U$ and $S_{T}:U\to[0,+\infty)$ be defined for each $\beta\in U$ by \[ S_{T}(\beta)=\int_0^{T}(X(t)-g(t,\beta))^2\cdot dt. \] The least-square estimator (LSE) of $\beta_0$ is any random vector $\beta_{T}=(\beta_{1,T},\dots,\beta_{q,T})\in U$ such that \[ S_{T}(\beta_{T})=\min({S_{T}(\beta) / \beta\in U}). \] In this paper under certain conditions of regularity it is said that there are the cumulants of all orders and also their respective spectral densities. This is for each integer $r\geq 2$, $M_{r}:\mathbb{R}^{r}\to\mathbb{R}$ and $\kappa_{r}:\mathbb{R}^{r}\to\mathbb{C}$ represent the moment and cumulant functions of order $r$ of $\varepsilon=(\varepsilon(t))_{t\in\mathbb{R}}$, i.e. \[ M_{r}(t_1,\dots,t_{r})=E(\varepsilon(t_1)\dots\varepsilon(t_{r})), \] and \[ \kappa_{r}(t_1,\dots,t_{r})=((1/i))^{r}((\partial^{r})/(\partial z_1\dots\partial z_{r}))K((z_1,\dots,z_{r}),(t_1,\dots,t_{r}))|_{z_1=\dots=z_{r}=0} \] where for each $((z_1,\dots,z_{r}),(t_1,\dots,t_{r}))\in\mathbb{R}^{r}\times\mathbb{R}^{r}$, is defined by \[ K((z_1,\dots,z_{r}),(t_1,\dots,t_{r}))=\ln(E(\exp(i\sum_{j=1}^{r}z_{j}\cdot\varepsilon(t_{j})))); \] and the cumulant spectral density is the function $f_{r}:\mathbb{R}^{r-1}\to\mathbb{C}$ such that \[ \kappa_{r}(t_1,\dots,t_{r}) = \kappa_{r}(t_1-t_{r},\dots,t_{r-1}-t_{r},0)=\int_{\mathbb{R}^{r-1}}\exp(i\sum_{j=1}^{r-1}\lambda_{j}(t_{j}-t_{r}))\cdot f_{r}(\lambda_1,\dots,\lambda_{r-1})\cdot\mu(d\lambda_1)\dots\mu(d\lambda_{r-1}), \] where $\mu$ is the Lebesgue measure on $\mathbb{R}$. Let $\theta^{(1)}\subset\mathbb{R}^{m_1}$ and $\theta^{(2)}\subset\mathbb{R}^{m_2}$ be open, bounded and convex sets, $\theta=\theta^{(1)}\times\theta^{(2)}$, and $\tau>0$ a real number, \[ \theta_{\tau}=\bigcup_{||e||\leq1}(\theta+\tau e), \] given the functions $d_2:\theta_{\tau}^{(2)}\to(0,\infty)$ and $a:\mathbb{R}\times\theta_{\tau}^{(1)}\to\mathbb C$, $f_2:\mathbb{R}\times\theta_{\tau}\to(0,\infty)$ is defined for each $(\lambda,\theta)=(\lambda,(\theta^{(1)},\theta^{(2)}))$ by \[ f_2(\lambda,\theta)=(1/(2\pi))d_2(\theta^{(2)})|a(\lambda,\theta^{(1)})|^2. \] From an earlier paper by other authors (cited in the References) with the addition of other conditions, there are $m_1\in\mathbb N$, $m_2\in\mathbb N$, $\theta, \tau>0$, functions $d_2:\theta_{\tau}^{(2)}\to(0,\infty)$, $a:\mathbb{R}\times\theta_{\tau}^{(1)}\to\mathbb C$, and $f_2$, as above, and $\theta_0=(\theta_0^{(1)},\theta_0^{(2)})\in\theta$ ($\theta$ is the closure of $\theta_0$) such that the spectral density of the order 2 cumulant is $f_2(\lambda,\theta_0)$. $\theta_0^{(1)}$ is a parameter of the function above mentioned, and $\theta_0^{(2)}$ represents a parameter of the Lévy process $L$. Let us now look at the definition of the minimum contrast estimator of the parameter $\theta_0$. Be $\beta_{T}$ a LSE of the parameter $\beta_0$. For each $\lambda\in\mathbb{R}$ be \[ I_{T}(\lambda,\beta_{T})=(1/(2\pi T))|\int_0^{T}(X(t)-g(t,\beta_{T}))\exp(-it\lambda)\cdot dt|^2, \] and is for each $\theta\in\theta_0$ \[ U_{T}(\theta,\beta_{T})=\int_{\mathbb{R}}(\ln f_2(\lambda,\theta)+((I_{T}(\lambda,\beta_{T}))/(f_2(\lambda,\theta))))w(\lambda)\cdot d\lambda, \] where $w:\mathbb{R}\to(0,\infty)$ is an even bounded measurable function such that the last integral is defined. The $U_{T}$ function is called ``Whittle contrast field''. The minimum contrast estimator (MCE) of $\theta_0$ is any random vector $\theta_{T}\in\theta$ such that \[ U_{T}(\theta_{T},\beta_{T})=\min({U_{T}(\theta,\beta_{T}) / \theta\in\theta_0}). \] In this paper the asymptotic consistency and normality of the MCE estimator is proved through various conditions and assumptions. Furthermore, an appendix presents the asymptotic properties already proven in previous works of the LSE estimator. This paper presents the generalization of a theorem used in the results referred to the LSE estimator. The two major results of this paper are being proved through several interesting lemmas in themselves. All proofs are rigorous, detailed and complemented by a long and updated bibliographical list. Application of the theoretical conditions and results obtained in a particular example of physics is shown in this paper. However, in each of the practical situations in which the results of this paper are to be applied, it will be necessary to consider whether the conditions to be met, as mentioned in this paper, can be imposed.