\begin{slide}{}

What description length $L_i (x^n)$ (i=0,1) to use?

Based on ${\cal M}_0$, by Shannon's theorem:

{\small
$L_0 (x^n) = - \log f_0 (x^n) =  - n \log (2 \pi)/ 2 + \sum_{t=0}^{n-1} x_{t+1}^2/2 $.

}

Based on ${\cal M}_1$, it is attempting to use
{\small
$-\log f_{{\hat\theta}= \bar{x}_n} =
n \log (2 \pi)/2 + \sum_{i=1}^n (x_i - \bar{x}_n)^2/2, $ 
}

but $f_{{\hat\theta}= \bar{x}_n} $ is not a density on $!

Using $\frac{1}{2} \log n = \log [1/\sqrt{n }]$ bits to encode
\\
$\hat\theta = \bar{x}_n$, we have
{\small
 \[
L_1 (x^n) = n \log (2 \pi)/2 + \sum_{i=1}^n (x_i - \bar{x}_n)^2/2   +
\frac{1}{2}  \log n 
\]
}

\end{slide}
