I'm a statistician not versed in functional analysis and operator theory. I wish that I might not find a wrong place for my question. All my questions are trivial in the scalar time series case, but the difficulty arises due to the infinite-dimensional functional(operator) setting. Any tiny or partial comment and insight would be enormously appreciated.

I consider a functional time series $(f_t)_{t=1}^{T}$ taking values in an infinite-dimensional separable Hilbert space $H$, specified by $f_t=Af_{t-1}+\varepsilon_t$, where $A$ is a Hilbert-Schmidt operator whose operator norm $\Vert A\Vert_{ \infty}$ is less than 1 so that $(f_t)_{t=1}^{T}$ is stationary.

Then, I think of the variance operator $\Sigma_f = \mathbb{E}(f_t\otimes f_t)$ of the time series, where $\otimes$ denotes the tensor product in $H$. Then, it is well-known that $\Sigma_f$ is a self-adjoint and trace-class operator, and it admits the spectral representation: $\Sigma_f = \sum_{i=1}^{\infty}\lambda_i(v_i\otimes v_i)$, where $(\lambda_i,v_i)_{i=1}^{\infty}$ is the eigenvalue and (orthonormalized) eigenfunction pairs of $\Sigma_f$, and $\lambda_i \downarrow 0$ as $i \rightarrow \infty$.

Thus, $\Sigma_f^{-1}$ is an unbounded operator, but it seems to be usually assumed in the literature that $\Vert \Sigma_f^{-\frac{1}{2}}A \Vert_{\infty}<\infty$.(If you should be interested, see this paper.)

Then, I am given consistent estimators $\hat{A}$ and $\hat\Sigma_f$ respectively for $A$ and $\Sigma_f$. Here, $\hat A$ is of finite rank for each $T$, and $\Vert \hat A -A\Vert_{\infty} \rightarrow 0$ in probability as $T \rightarrow \infty$. Also, $\hat \Sigma_f$ is self-adjoint and of finite rank for each $T$, and $\Vert\hat \Sigma_f - \Sigma_f \Vert_{\infty} \rightarrow 0$ in probability as $T \rightarrow \infty$.

Since both $\hat A$ and $\hat \Sigma_f$ are of finite rank for each $T$, it always holds true that $\Vert \hat\Sigma_f^{-\frac{1}{2}}\hat A \Vert_{\infty} < \infty$ for each T.

Then, my three questions are

Q1: Is it guaranteed that $\Vert \hat\Sigma_f^{-\frac{1}{2}}\hat A - \Sigma_f^{-\frac{1}{2}}A \Vert_{\infty} \rightarrow 0$ in probability as $T\rightarrow \infty$ ? If not, could we have at least that $\Vert \hat\Sigma_f^{-\frac{1}{2}}\hat A \Vert_{\infty} = O_p(1)$ as $T \rightarrow \infty$ ?

Q2: If another estimator $\tilde \Sigma_f$ for $\Sigma_f$ is such that $\Vert \Sigma_f - \tilde \Sigma_f + A\Sigma_f A^{*} - \hat A\tilde \Sigma_f \hat A^{*} \Vert_{\infty} \rightarrow 0$ in probability as $T \rightarrow \infty$, where $A^{*}$ and $\hat A^{*}$ are respectively the adjoints of $A$ and $\hat A$, then does it imply that $\Vert \Sigma_f - \tilde \Sigma_f \Vert_{\infty} \rightarrow 0$ in probability as $T \rightarrow \infty$ ?

Q3: Could you please let me know, if any, book, monograph, or paper dealing with such materials in a relatively easily accessible way ?

Any tiny or partial comment and insight would be enormously appreciated!



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.