Rejection Sampling

Rejection Sampling is one of the simplest sampling algorithm. Every introductory text about Monte Carlo method use this algorithm as a first example. The de facto example of this is to approximate pi.

As with MCMC method like Metropolis-Hastings and Gibbs Sampling, Rejection Sampling is used to draw samples from a complicated target distribution where direct sampling is hard. This can be done by using a proposal distribution Q(x) that is easy to sample from. This Q(x) has to have an important property, namely, Q(x) has to envelope the target distribution P(x). That means, given a scaling factor k, it has to be kQ(x) > P(x), for all x. In other word, our target distribution has to be entirely under our scaled proposal distribution.

For clearer picture, let’s dive into the code.

import numpy as np
import scipy.stats as st
import seaborn as sns
import matplotlib.pyplot as plt


sns.set()


def p(x):
    return st.norm.pdf(x, loc=30, scale=10) + st.norm.pdf(x, loc=80, scale=20)


def q(x):
    return st.norm.pdf(x, loc=50, scale=30)


x = np.arange(-50, 151)
k = max(p(x) / q(x))


def rejection_sampling(iter=1000):
    samples = []

    for i in range(iter):
        z = np.random.normal(50, 30)
        u = np.random.uniform(0, k*q(z))

        if u <= p(z):
            samples.append(z)

    return np.array(samples)


if __name__ == '__main__':
    plt.plot(x, p(x))
    plt.plot(x, k*q(x))
    plt.show()

    s = rejection_sampling(iter=100000)
    sns.distplot(s)

The target distribution that I want to sample here is a mixture of Gaussian: N(30, 10) + N(80, 20). Take a note that this distribution is not normalized. Here’s what it looks like.

Now, I’ll pick N(50, 30) as my proposal distribution Q(x).

If we do Rejection Sampling on this Q(x) naively, it will surely failed, as those high probability areas of P(x) aren’t covered by Q(x). Intuitively, the acceptance rate will be lower if Q(x) isn’t enveloping P(x), which resulting in having a lot less samples in those high probability areas than it should.

To remedy that, we need to find scaling factor k, so that kQ(x) will envelop P(x) entirely. To find it, we need to get the maximum ratio of P(x) and Q(x). So it will just be k = max(P(x) / Q(x)) for all x.

The main sampling algorithm is to produce a lot of samples from proposal distribution Q(x), z ~ Q(x), then uniformly pick the height of the distribution, so that it will be u ~ Unif(0, kQ(z)). Now that sample of (z, u) will be uniform under the kQ(x) curve. Then, we should just need to evaluate the height of our target distribution P(x) at point z. And then very intuitively, we accept (z, u) samples that are under the P(x) curve.

Given a long time to run and generating samples, then it will converge to P(x). Let’s see the result.

Recall that our target distribution that we used in Rejection Sampling is unnormalized. However, given the samples, we could reconstruct the proper distribution. So normalization doesn’t really matter. The requirement of the target distribution hence relaxed, we just need a function proportional to our true target distribution.

Rejection Sampling sure is simple. But it has some drawback. We have to have pretty good heuristic on choosing the proposal distribution Q(x). It has to envelope the target distribution. Given a complicated target distribution, it’s hard to know the shape, and it’s hard to choose the proposal distribution.

Also when computing the scaling factor, we have to be careful to make it “just enough” enveloping the target distribution, as the rejection rate will be proportional to the ratio of P(x) and Q(x). Having a really big Q(x) sure envelop the P(x), but it also means that there are a lot of wasted space that the algorithm will reject, which means a wasted computational time.