Strategies to summarize posterior

1. If we have familiar distributions, we can simply simulate samples in R. What if there isn’t a built in function for doing this?

2. Grid approximation.

   • Compute values of the posterior on a grid of points.
   • Approximate the continuous posterior by a discrete posterior.
   • Applicable for one- and two-parameter problems.

3. If there are more parameters, we consider deterministic tools.

   • MAP
   • Bayesian CLT
   • Numerical Integration

4. Monte Carlo methods are more general.

   • Independent Monte Carlo
   • Markov Chain Monte Carlo

Inverse CDF transformation

• Suppose that we are given $U \sim U(0,1)$ and want to simulate a continuous r.v. $X$ with cdf $F_X$. Put $Y = F_{X}^{-1}(U)$, we have $$F_{Y}(y)=\mathbb{P}(Y \leq y)=\mathbb{P}\left(F_{X}^{-1}(U) \leq y\right)=\mathbb{P}\left(U \leq F_{X}(y)\right)=F_{X}(y).$$

• For any continuous r.v. $X$, $F_{X}(X) \sim U(0,1)$. Why?

Example: exponential distribution

• If $X \sim \exp(\lambda)$, then $$F_{X}(x)=\left\{\begin{array}{ll}{0} & {\text { for } x<0} \\ {1-e^{-\lambda x}} & {\text { for } x \geq 0}\end{array}\right.$$

• How to work with inverse CDF transformation?

Example: exponential distribution

lambda <- 1
u <- runif(20)
x <- -1/lambda * log(1 - u)
curve(1 - exp(-x), 0, 5, ylab = "u", main = "Inverse CDF transformation for exp(1)")
for (i in 1:length(x)) {
    lines(c(x[i], x[i]), c(u[i], 0), lty = "dashed")
    lines(c(x[i], 0), c(u[i], u[i]), lty = "dashed")
}

Example: exponential distribution

# Direct sampling
lambda <- 1
u <- runif(1000)
x <- -1/lambda * log(1 - u)
hist(x)

What if we can not write the inverse?

Rejection sampling

• Although we cannot easily sample from $f(\theta|y)$, there exists another density $g(\theta)$ (called proposal density), like a Normal distribution or perhaps a $t$-distribution, from which it is easy for us to sample.

• Then we can sample from $g(\theta)$ directly and then "reject" the samples in a strategic way to make the resulting "non-rejected" samples look like they came from $f(\theta|y)$.

Rejection sampling

1. Sample $U \sim \operatorname{Unif}(0,1)$.

2. Sample $\theta$ at random from the probability density proportional to $g(\theta)$.

3. If $U \leq \frac{f(\theta|y)}{M g(\theta)}$, accept $\theta$ as a draw from $f$. If the drawn $\theta$ is rejected, return to step 1. (With probability $\frac{f(\theta | y)}{M g(\theta)}$), accept $\theta$).

The algorithm can be repeated until the desired number of samples from the target density $f$ has been accepted.

Illustration

What should $g(\theta)$ look like?

1. $\mathcal{X}_f \subset \mathcal{X}_g$, where $\mathcal{X}_f$ is the support of $f(\theta|y)$.

2. For any $\theta \in \mathcal{X}_f$, $\frac{f(\theta|y)}{g(\theta)} \leq M$, that is $$M=\sup _{\theta \in \mathcal{X}_{f}} \frac{f(\theta|y)}{g(\theta)}<\infty.$$

A simple choice is choose $g(\theta)$ that heavier tails.

Example

curve(dnorm(x), -6, 6, xlab = "x", ylab = "Density", n = 200)
curve(dt(x, 2), -6, 6, add = TRUE, col = 4, n = 200)
legend("topright", c("Normal density", "t density"), col = c(1, 4), bty = "n", lty = 1)

Probability of $\theta$ accepted

$$\begin{aligned} \mathbb{P}(\theta \text { accepted }) &=\mathbb{P}\left(U \leq \frac{f(\theta|y)}{M g(\theta)}\right) \\ &=\int \mathbb{P}\left(U \leq \frac{f(\theta|y)}{M g(\theta)} | \theta\right) g(\theta) d \theta \\ &=\int \frac{f(\theta|y)}{M g(\theta)} g(\theta) d \theta \\ &=\frac{1}{M} \end{aligned}$$

1. This has implications for how we choose the proposal density $g$.

2. We want to choose $g$ so that it matches $f$ as closely as possible.

Back to the example

Why rejection sampling work?

It can be shown that the distribution function of the accepted values is equal to the distribution function corresponding to the target density: $$P(\Theta \leq \theta | \Theta \text{ accepted}) = F(\theta).$$

What is the problem of rejection sampling?

Rejection sampling

• Suppose we are interested in $\mathbb{E}_{f}[h(\theta|y)]$, what do we do?

• Note that in most cases, the candidates generated from $g$ fall within the domain of $f$, so that they are in fact values that could plausibly come from $f$.

• They are simply over- or under-represented in the frequency with which they appear.

• For example, if $g$ has heavier tails than $f$, then there will be too many extreme values generated from $g$. Rejection sampling simply thins out those extreme values to obtain the right proportion.

• But what if we could take those rejected values and, instead of discarding them, simply downweight or upweight them in a specific way?

Lab 10

Use the rejection sampling algorithm to simulate from $N(0, 1)$. Try the following proposal densities.

1. Logistic Distribution.
2. Cauchy distribution.
3. Student $t_5$ distribution.
4. $N(1,2)$ distribution.

For each proposal record the percentage of iterates that are accepted. Which is the best proposal?

Importance sampling

• We can rewrite the target $$\mathbb{E}_{f}[h(\theta|y)]=\mathbb{E}_{g}\left[\frac{f(\theta|y)}{g(\theta)} h(\theta|y)\right].$$
• If $\theta_{1}, \dots, \theta_{n} \sim g$, we get $$\tilde{\mu}_{n}=\frac{1}{n} \sum_{i=1}^{n} \frac{f\left(\theta_{i}|y\right)}{g\left(\theta_{i}\right)} h\left(\theta_{i}|y\right)=\frac{1}{n} \sum_{i=1}^{n} w_{i} h\left(\theta_{i}|y\right) \approx \mathbb{E}_{f}[h(\theta|y)].$$
• $w_i$ are referred to as the importance weights.

Summary

• Direct sampling
• Rejection sampling
• Importance sampling

References and further readings

References

• Chapter 10 of the BDA book by Gelman et.al.

Further readings

• Chapter 6 of Advanced Statistical Computing by Roger Peng.