Let $X_1,...,X_N$ be random functions on an Hilbert space $\mathbb{H}$, following a Functional Autoregressive process of order 1: $$ X_{t+1} = \Psi X_t + \varepsilon_t $$ Where $\Psi$ is a linear bounded operator on $\mathbb{H}$ and $\{\varepsilon_t\} \subset \mathbb{H}$.

In the context of prediction for the above model, we aim to estimate $\Psi$.
Kargin and Onatski (2008) introduce the method of Predictive Factors.

In a nutshell, they propose to to estimate a low-rank approximation of $\Psi$.
In order to do so, a foundamental step consists in estimating eigenvalues and eigenfunctions of another operator $\hat{\Phi}_\alpha$: $$\hat{\Phi}_\alpha := \hat{C}^{-1/2}_\alpha \hat{C}_1^T \hat{C}_1 \hat{C}^{-1/2}_\alpha$$

Where $C_1$ is the lag-1 autocovariance, $C_\alpha = C + \alpha I$, $C$ is the autocovariance of the process and $I$ is the identity operator.
I have derived the following form for the estimators $\hat{C}^{-1/2}_\alpha, \hat{C}_1, \hat{C}_1^T$:

$$\hat{C}^{-1/2}_\alpha (x)= \sum_{j=1}^{+\infty} (\hat{\lambda}_j + \alpha)^{-1/2}\langle x, \hat{\xi}_j \rangle \hat{\xi}_j \quad x \in \mathbb{H}$$ $$\hat{C}_1 (x)= \sum_{j=1}^{N-1} \langle X_j, x \rangle X_{j+1} \quad x \in \mathbb{H}$$ $$\hat{C}_1^T (x)= \sum_{j=1}^{N-1} \langle X_{j+1}, x \rangle X_{j} \quad x \in \mathbb{H}$$

Where $\hat{\lambda}_j$ and $\hat{\xi}_j$ are the estimated eigenvalues and eigenfunctions of $C$.

I don't know how to proceed further and how to obtain the eigenvalues and eigenfunctions of the operator $\hat{\Phi}_\alpha$.

It is a rather technical question, but any help is greatly appreciated! Thank you!



Your Answer

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

Browse other questions tagged or ask your own question.