Last modified: January 27, 2025

This article is written in: 🇺🇸

Yule-Walker Equations

The Yule-Walker equations are a set of linear equations that relate the autocorrelations of an autoregressive (AR) process to its parameters. These equations are crucial for estimating the parameters of AR models and for understanding the autocorrelation structure of the process.

Definition

The Yule-Walker equations provide a way to estimate the parameters of an AR(p) process by relating them to the autocorrelations of the process. The key idea is to derive a system of equations that link the parameters ${\varphi }_1,{\varphi }_2,\dots ,{\varphi }_p$ with the autocorrelation function (ACF) at different lags $k$.

For an AR(p) process, the Yule-Walker equations are:

$$\rho \left(k\right)={\varphi }_1\rho (k-1)+{\varphi }_2\rho (k-2)+\cdots +{\varphi }_p\rho (k-p)\text{for}k=1,2,\dots ,p$$

Where:

• $\rho \left(k\right)$ is the autocorrelation at lag $k$.
• $\rho \left(0\right)=1$, since the autocorrelation at lag 0 is always 1.

These equations can be used to estimate the parameters ${\varphi }_1,{\varphi }_2,\dots ,{\varphi }_p$ of the AR process from the sample autocorrelations.

Deriving the Yule-Walker Equations

Assumptions

1. The stationarity assumption for an AR process means that the mean and variance of the process remain constant over time, and the autocovariance function depends only on the lag, not on the time.
2. In the AR process, the white noise error term $Z_t$ has zero mean and constant variance ${\sigma }_Z^2$, meaning $Z_t$ is uncorrelated with past values of the process.

Derivation Steps

1. Multiply the AR(p) model by $X_{t-k}$ for a given lag $k$:

$$X_t={\varphi }_1{X}_{t-1}+{\varphi }_2{X}_{t-2}+\cdots +{\varphi }_p{X}_{t-p}+Z_t$$

Multiply both sides by $X_{t-k}$ for $k=1,2,\dots ,p$.

1. Take the expectation of both sides: Use the definition of the autocovariance function $\gamma \left(k\right)=E\left[X_t{X}_{t-k}\right]$, and replace the expectations accordingly. Since the white noise $Z_t$ is uncorrelated with $X_{t-k}$, the expected value of $Z_t{X}_{t-k}$ is zero for all $k\ge 1$.

2. Use covariance definitions: Express the expectations as autocovariances $\gamma \left(k\right)$ and normalize by $\gamma \left(0\right)$, which is the variance of the process, to convert them into autocorrelations $\rho \left(k\right)$.

3. Formulate the Yule-Walker equations: You will end up with a system of linear equations that relate the autocorrelations at different lags $\rho \left(1\right),\rho \left(2\right),\dots ,\rho \left(p\right)$ to the AR model parameters ${\varphi }_1,{\varphi }_2,\dots ,{\varphi }_p$.

Example: Yule-Walker Equations for an AR(2) Process

Consider the AR(2) process:

$$X_t=3X_{t-1}+2X_{t-2}+Z_t$$

Where $Z_t$ is white noise with variance ${\sigma }_Z^2$.

Deriving the Yule-Walker Equations

I. Multiply by $X_{t-1}$ and take the expectation:

$$E\left[X_t{X}_{t-1}\right]=3E\left[X_{t-1}^2\right]+2E\left[X_{t-2}{X}_{t-1}\right]$$

Using $\gamma \left(1\right)=E\left[X_t{X}_{t-1}\right]$ and $\gamma \left(0\right)=E\left[X_{t-1}^2\right]$, we get:

$$\gamma \left(1\right)=3\gamma \left(0\right)+2\gamma \left(1\right)$$

II. Multiply by $X_{t-2}$ and take the expectation:

$$E\left[X_t{X}_{t-2}\right]=3E\left[X_{t-1}{X}_{t-2}\right]+2E\left[X_{t-2}^2\right]$$

Using $\gamma \left(2\right)=E\left[X_t{X}_{t-2}\right]$, we get:

$$\gamma \left(2\right)=3\gamma \left(1\right)+2\gamma \left(0\right)$$

III. Express the autocovariances as autocorrelations: Normalize by the variance $\gamma \left(0\right)$, to convert autocovariances into autocorrelations $\rho \left(k\right)$:

$$\rho \left(1\right)=\frac{3}{1+2}=\frac{3}{3}=1,\rho \left(2\right)=\frac{3\cdot 1+2}{\gamma \left(0\right)}$$

Solving the Difference Equations for AR(2)

To solve the Yule-Walker equations for the AR(2) process, we assume a solution of the form $\rho \left(k\right)={\lambda }^k$. Substituting into the equation:

$${\lambda }^2=3\lambda +2$$

The characteristic equation is:

$${\lambda }^2-3\lambda -2=0$$

Solving the quadratic equation:

$$\lambda =\frac{3\pm \sqrt{9+8}}{2}=\frac{3\pm \sqrt{17}}{2}$$

Thus, the roots are:

$${\lambda }_1=\frac{3+\sqrt{17}}{2},{\lambda }_2=\frac{3-\sqrt{17}}{2}$$

The general solution for the autocorrelation function $\rho \left(k\right)$ is:

$$\rho \left(k\right)=c_1{\lambda }_1^k+c_2{\lambda }_2^k$$

Finding Coefficients c_1 and c_2

Use $\rho \left(0\right)=1$

At $k=0$, we know $\rho \left(0\right)=1$. This gives the equation:

$$c_1+c_2=1$$

Use $\rho \left(1\right)$

At $k=1$, we use the relationship $\rho \left(1\right)=3\rho \left(0\right)+2\rho \left(1\right)$ to find:

$$\rho \left(1\right)=3\cdot 1+2\cdot \rho \left(1\right)$$

Solving this, we get:

$$\rho \left(1\right)=\frac{3}{2}$$

Substituting into the general solution:

$$c_1{\lambda }_1+c_2{\lambda }_2=\frac{3}{2}$$

Solving the system of equations:

$$\begin{array}{cc}{c}_1+c_2& =1\\ c_1{\lambda }_1+c_2{\lambda }_2& =\frac{3}{2}\end{array}$$

This gives the values for $c_1$ and $c_2$, which can then be substituted back into the general solution for $\rho \left(k\right)$.

Matrix Form of Yule-Walker Equations for AR(p)

For an AR(p) process, the Yule-Walker equations can be written in matrix form. Define:

Vector of autocorrelations:

$$r=\left[\begin{array}{c}\rho \left(1\right)\rho \left(2\right)⋮\rho \left(p\right)\end{array}\right]$$

Vector of AR coefficients:

$$\varphi =\left[\begin{array}{c}{\varphi }_1{\varphi }_2⋮{\varphi }_p\end{array}\right]$$

The autocorrelation matrix $R$ is a Toeplitz matrix:

$$R=\left[\begin{array}{ccccc}1& \rho \left(1\right)& \rho \left(2\right)& \dots & \rho (p-1)\\ \rho \left(1\right)& 1& \rho \left(1\right)& \dots & \rho (p-2)\\ \rho \left(2\right)& \rho \left(1\right)& 1& \dots & \rho (p-3)\\ ⋮& ⋮& ⋮& \ddots & ⋮\\ \rho (p-1)& \rho (p-2)& \rho (p-3)& \dots & 1\end{array}\right]$$

The Yule-Walker equations are then written as:

$$R\varphi =r$$

To solve for the AR coefficients $\varphi$:

$$\varphi =R^{-1}r$$