$ARMA(p,q)$, $\phi(B) x_t = \phi_0 + \theta(B)\epsilon_t$. Define $\pi(B) = \frac{\phi(B)}{\theta(B)}$ and the AR representation is

\[\pi(B) x_t = \frac{\phi_0}{\theta(B)} + \epsilon_t\]

and the MA model becomes

\[x_t = \frac{\phi_0}{\phi(B)} + \psi(B) \epsilon_t\]

Time domain representation:

\[x_t = \mu + \sum_{j=0}^\infty \psi_j \epsilon_{t-j}\]

$\epsilon_t$ iid mean-zero and $\text{var}(\epsilon_t) = \sigma^2$

Karhunen-Loeve theorem:

For any mean-zero, weakly stationary $x_t$, $t\in[a,b]$ , $x_t$ can be representated by

\[x_t = \sum_{k=1}^\infty z_k e_k(t)\]

$\mathbb{E}z_k = 0 $ and $\mathbb{E}z_iz_j = \lambda_j \delta_{ij}$ ($z_i$ are uncorrelated). Here, $e_k(t)_{k=1}^\infty$ continuous function on $[a,b]$, orthogonal in $L^2([a,b])$ , i.e.

\[\int_a^b e_k(t)e_j(t) dt = \delta_{kj}\]

The orthogonality of $e_k$ shows that

\[\int_a^b x_t e_j(t) dt = \int_a^b \sum_{k=1}^\infty z_k e_k(t) e_j(t) dt = \sum_{k=1}^\infty z_k\delta_{kj} = z_j\]

Now we consider the representation

\[x_t = \phi_0 + \int_0^\pi \alpha(\omega) \cos(\omega t) d\omega + \int_0^\pi \beta(\omega) \sin(\omega t) d\omega\]

This is the spectral frequency domain. With the property that

\[\alpha (\omega)\vert_{\omega_1}^{\omega_2}\]

and

\[\alpha (\omega)\vert_{\omega_3}^{\omega_4}\]

are uncorrelated when $\omega_1\leq \omega_2 \leq \omega_3\leq \omega_4 $.

Define autocovariance generating function (ACGF)

\[g(z) = g_x(z) = \sum_{j=\infty}^\infty \gamma_j z^j\]

where $\gamma_j = \text{Cov}(x_t,x_{t-j})$ for $(x_t)$ weakly stationary. The spectral,

\[\begin{aligned} s(\omega) &= \frac{1}{2\pi} g(e^{-i\omega})= \frac{1}{2\pi} \sum_{j\in\mathbb{Z}} \gamma_j e^{-i\omega j} \\ &= \frac{1}{2\pi} \sum_j \gamma_j[\cos(\omega j) - i\sin(\omega j)] \\ &= \frac{1}{2\pi} \left[\gamma_0 + 2\sum_{j=1}^\infty \gamma_j \cos(\omega j)\right] \end{aligned}\]

where we used the fact that $\gamma_j$ is an even function of $j$ while $\sin(\omega j)$ is odd.

Properties: If $ \sum_{j\in\mathbb{Z}} \vert\gamma_j \vert < \infty $

  1. the $s(\omega)$ exists
  2. $s(\omega)$ is continuous and differentiable in $\omega$.
  3. $s(-\omega) = s(\omega)$ symmetric around zero
  4. $s(\omega + 2\pi k) = s(\omega)$ for $\forall k \in \mathbb{Z}$.

Thus, we could only need to consider the $s(\omega)$ for $\omega \in [0, \pi]$.

Back to $MA(\infty)$ representation $(\mu = 0)$

\[x_t = \sum_{j=0}^\infty \psi_j \epsilon_{t-j}\]

Assume $\sum_{j=0}^\infty \vert \psi_j \vert <\infty$ and $\text{Var}(\epsilon_t) = \sigma^2$ ,

\[g(z) = \sigma^2 \psi(z) \psi(z^{-1})\]

Then

\[s(\omega) = \frac{1}{2\pi} g(e^{-i\omega}) = \frac{\sigma^2}{2\pi} \psi(e^{-i\omega}) \psi(e^{i\omega})\]

If $\psi_0 = 1$ and $\psi_1=\psi_2 = \dots = 0 $ , we have

\[\psi(z) = 1 - \psi_1(z) - \psi_2(z) - \dots = 1 \quad \Rightarrow s(\omega) = \frac{\sigma^2}{2\pi}\]

Thus this is called the white noise.

If $\psi_0 = 1$ and $\psi_i = 0$ for $i\geq 2$. Then

\[\begin{aligned} s(\omega) &= \frac{\sigma^2}{2\pi} (1-\psi_1 e^{-i\omega}) ( 1 - \psi_1 e^{i\omega}) &= \frac{\sigma^2}{2\pi} (1+2\cos(\omega) \psi_1 + \psi_1^2) \end{aligned}\]

Updated:

Leave a comment