BDA3 Chapter 2 Exercise 14

Here’s my solution to exercise 14, chapter 2, of Gelman’s Bayesian Data Analysis (BDA), 3rd edition. There are solutions to some of the exercises on the book’s webpage.

\(\DeclareMathOperator{\dbinomial}{binomial} \DeclareMathOperator{\dbern}{Bernoulli} \DeclareMathOperator{\dpois}{Poisson} \DeclareMathOperator{\dnorm}{normal} \DeclareMathOperator{\dcauchy}{Cauchy} \DeclareMathOperator{\dgamma}{gamma} \DeclareMathOperator{\invlogit}{invlogit} \DeclareMathOperator{\logit}{logit} \DeclareMathOperator{\dbeta}{beta}\)

Suppose we have a normal prior \(\theta \sim \dnorm (\mu_0, \frac{1}{\tau_0})\) and a normal sampling distribution \(y \mid \theta \sim \dnorm(\theta, \sigma)\), where the variance is known. We will show by induction that the posterior is \(\theta \mid y_1, \dotsc, y_{n} \sim \dnorm(\mu_{n}, \frac{1}{\tau_{n}})\) where

\[ \frac{1}{\tau_{n}^2} = \frac{1}{\tau_0^2} + \frac{n}{\sigma^2} \quad \text{and} \quad \mu_n = \frac{\frac{1}{\tau_0^2}\mu_0 + \frac{n}{\sigma^2}\bar y_n}{\frac{1}{\tau_0^2} + \frac{n}{\sigma^2}} \]

for \(n = 1, \dotsc, \infty\).

Base case

The case \(n = 1\) can be reexpressed as

\[ \begin{align} \frac{1}{\tau_1^2} = \frac{\sigma^2 + \tau_0^2}{\sigma^2\tau_0^2} \quad \text{and} \quad \mu_1 = \frac{\sigma_0^2\mu_0 + \tau_0^2y}{\sigma^2 + \tau_0^2} . \end{align} \]

Now we can combine the fractions in the exponent, expand the brackets, collect the terms as \(\theta\)-coefficients, rewrite the coefficients in terms of \(\mu_1\) and \(\tau_1\), then complete the square in terms of \(\theta\):

\[\begin{align} \frac{(y - \theta)^2}{\sigma^2} + \frac{(\theta - \mu_0)^2}{\tau_0^2} &= \frac{(y^2 + \theta^2 - 2\theta y) \tau_0^2 + (\theta^2 + \mu_0^2 - 2 \theta \mu_0) \sigma^2}{\sigma^2 \tau_0^2} \\ &= \frac{\theta^2 (\tau_0^2 + \sigma^2) -2 \theta (\sigma^2\mu_0 + \tau_0^2y) + (y^2\tau_0^2 + \mu_0^2\sigma^2)}{\sigma^2\tau_0^2} \\ &= \frac{\theta^2 (\tau_0^2 + \sigma^2) -2 \theta \mu_1 (\sigma^2 + \tau_0^2) + \mu_1 (\sigma^2 + \tau_0^2) }{\sigma^2\tau_0^2} \\ &= \theta^2\frac{1}{\tau_1^2} -2\theta \mu_1 \frac{1}{\tau_1^2} + \mu_1 \frac{1}{\tau_1^2} \\ &= \frac{\theta^2 -2 \mu_1 \theta + \mu_1^2}{\tau_1^2} \\ &= \frac{(\theta - \mu_1)^2}{\tau_1^2} . \end{align}\]

Induction step

The induction hypothesis is that the variance and mean are given by

\[ \frac{1}{\tau_n^2} = \frac{1}{\tau_0^2} + \frac{n}{\sigma^2} \quad \text{and} \quad \mu_n = \frac{\frac{1}{\tau_0^2}\mu_0 + \frac{n}{\sigma^2}\bar y}{\frac{1}{\tau_0^2} + \frac{n}{\sigma^2}} \]

for some \(n \ge 1\).

Starting with the variance, the base step can be reexpressed as

\[ \frac{1}{\tau_{n+1}^2} = \frac{1}{\tau_n^2} + \frac{1}{\sigma^2} . \]

Now apply the induction hypothesis to get

\[ \frac{1}{\tau_{n+1}^2} = \frac{1}{\tau_0^2} + \frac{n}{\sigma^2} + \frac{1}{\sigma^2} = \frac{1}{\tau_0^2} + \frac{n+1}{\sigma^2} . \]

For the mean we apply the same strategy. The base step can be reexpressed as

\[ \mu_{n+1} = \frac{ \frac{1}{\tau_n^2}\mu_n + \frac{1}{\sigma^2}y_{n+1} }{ \frac{1}{\tau_n^2} + \frac{1}{\sigma^2} } . \]

Applying the induction hypothesis then gives

\[\begin{align} \mu_{n+1} &= \frac{ \frac{1}{\tau_n^2} \left( \frac{\frac{1}{\tau_0^2}\mu_0 + \frac{n}{\sigma^2}\bar y_n}{\frac{1}{\tau_0^2} + \frac{n}{\sigma^2}} \right) + \frac{1}{\sigma^2}y_{n+1} }{ \frac{1}{\tau_0^2} + \frac{n+1}{\sigma^2}} \\ &= \frac{ \frac{1}{\tau_0^2}\mu_0 + \frac{n}{\sigma^2}\bar y_n + \frac{1}{\sigma^2}y_{n+1} }{ \frac{1}{\tau_0^2} + \frac{n+1}{\sigma^2} } \\ &= \frac{ \frac{1}{\tau_0^2}\mu_0 + \frac{n+1}{\sigma^2}\bar y_{n+1} }{ \frac{1}{\tau_0^2} + \frac{n+1}{\sigma^2} } , \end{align}\]

since

\[ \begin{align} (n + 1) \bar y_{n + 1} &:= \sum_1^{n+1} y_i \\ &= y_{n+1} + \sum_1^n y_i \\ &= y_{n+1} + n\bar y_n . \end{align} \]