AR1模型首个元素y(1)的确定方法问询

Great question! Let's walk through the standard, practical approaches to compute or handle $y(1)$ in an AR(1) process defined by:

$y(t) = a \cdot y(t-1) + e(t)$

when the initial value $y(0)$ isn't available. These methods are widely used in statistical software and time series practice:

1. Simplified Approximation: Assume $y(0)$ equals the process mean #

For a stationary AR(1) process (where $|a| < 1$), the mean $\mu$ is $\frac{E[e(t)]}{1-a}$. If the error term $e(t)$ has a mean of 0 (the most common case), $\mu = 0$.

This means you can set $y(0) = 0$, so:
y(1) = e(1)

If your process has a non-zero mean, substitute $\mu$ for $y(0)$ instead. This is a quick, practical approach, and when your sample size is large, the impact of this initial assumption on downstream analysis is negligible.

2. Treat $y(1)$ as a Fixed Initial Condition (Conditional Likelihood) #

Instead of trying to "compute" $y(1)$, many software tools treat the first observation $y(1)$ as a given starting point. When fitting the AR(1) model, they only use observations from $t=2$ to $t=T$ to estimate the coefficient $a$.

For example, in R's arima() function, the default method="CSS" (Conditional Sum of Squares) uses this logic—it fixes $y(1)$ as the initial value and minimizes the sum of squared errors for the rest of the series.

3. Use the Stationary Initial Distribution (Unconditional Likelihood) #

For a stationary AR(1) process, $y(0)$ follows the process's stationary distribution. If $e(t)$ is a normal white noise with variance $\sigma_e^2$, this distribution is:

$y(0) \sim N\left(\mu, \frac{\sigma_e^2}{1-a2}\right)$

When fitting the model with maximum likelihood (e.g., method="ML" in R's arima()), the software estimates the parameters $a$, $\sigma_e^2$, and accounts for the uncertainty in $y(0)$ using this distribution. This is more statistically rigorous, especially for small sample sizes, as it doesn't treat $y(1)$ as a fixed value.

Quick Code Example (R) #

# Generate AR(1) data with a=0.6, e~N(0,1)
set.seed(123)
e <- rnorm(100)
y <- numeric(100)
y[1] <- e[1]  # Using y(0)=0 approximation
for(t in 2:100) {
  y[t] <- 0.6*y[t-1] + e[t]
}

# Fit with conditional least squares (fixed y(1))
fit_css <- arima(y, order=c(1,0,0), method="CSS")

# Fit with maximum likelihood (accounts for initial distribution)
fit_ml <- arima(y, order=c(1,0,0), method="ML")

To recap:

• If you just need a single value for $y(1)$, use y(1) = e(1) (assuming zero mean).
• If fitting a model, choose conditional likelihood for simplicity or unconditional likelihood for better small-sample accuracy.

内容的提问来源于stack exchange，提问作者Fabio Capezzuoli