# Understanding MCMC: what would the alternative be?

Learning Bayesian stats for the first time; as an angle towards understanding MCMC I wondered: is it doing something that fundamentally can't be done another way, or is it just doing something far more efficiently than the alternatives?

By way of illustration, suppose we're trying to compute the probability of our parameters given the data $P(x,y,z|D)$ given a model that computes the opposite, $P(D|x,y,z)$. To calculate this directly with Bayes' theorem we need the denominator $P(D)$ as pointed out here. But could we compute that by integration, say as follows:

p_d = 0.
for x in range(xmin,xmax,dx):
for y in range(ymin,ymax,dy):
for z in range(zmin,zmax,dz):
p_d_given_x_y_z = cdf(model(x,y,z),d)
p_d += p_d_given_x_y_z * dx * dy * dz


Would that work (albeit very inefficiently with higher numbers of variables) or is there something else that would cause this approach to fail?

• Integration would work in many cases, but it would take too long (i.e., it is inefficient). MCMC is a way to estimate the posterior efficiently. – Mark White Nov 6 '17 at 15:32
• Not relevant for the question, but I think your are missing the prior over x, y, z in your integral (it appears in the numerator of Bayes' formula) – alberto Nov 6 '17 at 15:44

You are describing a grid approximation to the posterior, and that is a valid approach, allthough not the most popular. There are quite a few cases in which the posterior distribution can be computed analytically. Monte Carlo Markov Chains, or other approximate methods, are methods to obtain samples of the posterior distribution, that sometimes work when the analytical solution cannot be found.

The analytical solutions that can be found are typically cases of "conjugate" families, and you can find more about that by googling, see for example https://en.wikipedia.org/wiki/Conjugate_prior.

As a first example, if your prior on p is uniform on [0, 1], where p is a success parameter in a simple binomial experiment, the posterior is equal to a Beta distribution. Integration, or summation, can be done explicitly in this case.

If you have finitely many parameter choices, or you use a grid approximation as in your example, a simple summation may be all you need. The number of computations can explode quickly however, if you have a couple of variables and want to use a dense grid.

There are several algorithms for sampling from the posterior. Hamiltonian Monte Carlo, specifically the NUTS sampler, is now popular and used in stan and PyMC3, Metropolis Hastings is the classic. Variational Inference is a relative newcomer, not a sampling method actually but a different way of obtaining an approximation. At the moment, none of the methods, including analytical solutions, are the best, they all work well in specific cases.

• Good answer, but your last paragraph seems to imply that variational inference is a sampling method, which it isn't. You might consider correcting that. – Ruben van Bergen Nov 7 '17 at 8:57

Calculating the denominator does not help in understanding the nature of the posterior distribution (or of any distribution). As discussed in a recent question, to know that the density of a d-dimensional vector $\theta$ is $$π(θ|x)∝\exp\{−||θ−x||^2−||θ+x||^4−||θ−2x||^6\},\qquad x,θ∈ℝ^d,$$ does not tell me where are the regions of interest for this posterior distribution.

Monte Carlo methods are techniques that make use of random numbers. The goal is to find samples $x$ that are distributed according $P(x)$ and it is assumed that $P(x)$ is complex. This means that we cannot evaluate it directly. If this is not the case, you can just compute it analytically. As in your example this would be $P(D)$.

What you propose is essentially a grid search through the space of $x$ and $y$. This can be very exhaustive if $x$ and $y$ are high dimensional and infeasible if they are continuous. Another problem is that you have to compute the cdf in each step.

MCMC methods try to solve this by proposing candidate samples $c_i$ and then accepting or rejecting them depending on some measure. This can in theory be faster then going through all possible combinations. so basically you find samples that are drawn from the prior $P(D)$. A theoretical problem here is that this is only the case in the limit number of samples drawn, i.e. after $\infty$ samples. So you don't know when to stop the Markov Chain.