This is really intended as a draft chapter for our book. Given the diverse natures of the intended intended audiences, it is probably a bit light on explanation of the Haskell (use of monad transformers) for those with a background in numerical methods. It is hoped that the explanation of the mathematics is adequate for those with a background in Haskell but not necessarily in numerical methods. As always, any feedback is gratefully accepted.
Introduction
Imagine an insect, a grasshopper, trapped on the face of a clock which wants to visit each hour an equal number of times. However, there is a snag: it can only see the value of the hour it is on and the value of the hours immediately anticlockwise and immediately clockwise. For example, if it is standing on 5 then it can see the 5, the 4, and the 6 but no others.
It can adopt the following strategy: toss a fair coin and move anticlockwise for a head and move clockwise for a tail. Intuition tells us that over a large set of moves the grasshopper will visit each hour (approximately) the same number of times.
Can we confirm our intuition somehow? Suppose that the strategy has worked and the grasshopper is now to be found with equal probability on any hour. Then at the last jump, the grasshopper must either have been at the hour before the one it is now on or it must have been at the hour after the one it is now on. Let us denote the probability that the grasshopper is on hour by and the (conditional) probability that the grasshopper jumps to state given it was in state by . Then we have
Substituting in where is a normalising constant (12 in this case) we obtain
This tells us that the required distribution is a fixed point of the grasshopper’s strategy. But does the strategy actually converge to the fixed point? Let us perform an experiment.
First we import some modules from hmatrix.
> {# LANGUAGE FlexibleContexts #}
> module Chapter1 where
> import Data.Packed.Matrix
> import Numeric.LinearAlgebra.Algorithms
> import Numeric.Container
> import Data.Random
> import Control.Monad.State
> import qualified Control.Monad.Writer as W
> import qualified Control.Monad.Loops as ML
> import Data.Random.Source.PureMT
Let us use a clock with 5 hours to make the matrices sufficiently small to fit on one page.
Here is the strategy encoded as a matrix. For example the first row says jump to position 1 with probablity 0.5 or jump to position 5 with probability 0.5.
> eqProbsMat :: Matrix Double
> eqProbsMat = (5 >< 5)
> [ 0.0, 0.5, 0.0, 0.0, 0.5
> , 0.5, 0.0, 0.5, 0.0, 0.0
> , 0.0, 0.5, 0.0, 0.5, 0.0
> , 0.0, 0.0, 0.5, 0.0, 0.5
> , 0.5, 0.0, 0.0, 0.5, 0.0
> ]
We suppose the grasshopper starts at 1 o’clock.
> startOnOne :: Matrix Double
> startOnOne = ((1 >< 5) [1.0, 0.0, 0.0, 0.0, 0.0])
If we allow the grasshopper to hop 1000 times then we see that it is equally likely to be found on any hour hand with a 20% probability.
ghci> eqProbsMat
(5><5)
[ 0.0, 0.5, 0.0, 0.0, 0.5
, 0.5, 0.0, 0.5, 0.0, 0.0
, 0.0, 0.5, 0.0, 0.5, 0.0
, 0.0, 0.0, 0.5, 0.0, 0.5
, 0.5, 0.0, 0.0, 0.5, 0.0 ]
ghci> take 1 $ drop 1000 $ iterate (<> eqProbsMat) startOnOne
[(1><5)
[ 0.20000000000000007, 0.2, 0.20000000000000004, 0.20000000000000004, 0.2 ]]
In this particular case, the strategy does indeed converge.
Now suppose the grasshopper wants to visit each hour in proportion the value of the number on the hour. Lacking pen and paper (and indeed opposable thumbs), it decides to adopt the following strategy: toss a fair coin as in the previous strategy but only move if the number is larger than the one it is standing on; if, on the other hand, the number is smaller then choose a number at random from between 0 and 1 and move if this value is smaller than the ratio of the proposed hour and the hour on which it is standing otherwise stay put. For example, if the grasshopper is standing on 5 and gets a tail then it will move to 6 but if it gets a head then four fifths of the time it will move to 4 but one fifth of the time it will stay where it is.
Suppose that the strategy has worked (it is not clear that is has) and the grasshopper is now to be found at 12 o’clock 12 times as often as at 1 o’clock, at 11 o’clock 11 times as often as at 1 o’clock, etc. Then at the last jump, the grasshopper must either have been at the hour before the one it is now on, the hour after the one it is now on or the same hour it is now on. Let us denote the probability that the grasshopper is on hour by .
Substituting in at 4 say
The reader can check that this relationship holds for all other hours. This tells us that the required distribution is a fixed point of the grasshopper’s strategy. But does this strategy actually converge to the fixed point?
Again, let us use a clock with 5 hours to make the matrices sufficiently small to fit on one page.
Here is the strategy encoded as a matrix. For example the first row says jump to position 1 with probablity 0.5 or jump to position 5 with probability 0.5.
> incProbsMat :: Matrix Double
> incProbsMat = scale 0.5 $
> (5 >< 5)
> [ 0.0, 1.0, 0.0, 0.0, 1.0
> , 1.0/2.0, 1.0/2.0, 1.0, 0.0, 0.0
> , 0.0, 2.0/3.0, 1.0/3.0, 1.0, 0.0
> , 0.0, 0.0, 3.0/4.0, 1.0/4.0, 1.0
> , 1.0/5.0, 0.0, 0.0, 4.0/5.0, 1.0/5.0 + 4.0/5.0
> ]
We suppose the grasshopper starts at 1 o’clock.
If we allow the grasshopper to hop 1000 times then we see that it is equally likely to be found on any hour hand with a probability of times the probability of being found on 1.
ghci> incProbsMat
(5><5)
[ 0.0, 0.5, 0.0, 0.0, 0.5
, 0.25, 0.25, 0.5, 0.0, 0.0
, 0.0, 0.3333333333333333, 0.16666666666666666, 0.5, 0.0
, 0.0, 0.0, 0.375, 0.125, 0.5
, 0.1, 0.0, 0.0, 0.4, 0.5 ]
ghci> take 1 $ drop 1000 $ iterate (<> incProbsMat) startOnOne
[(1><5)
[ 6.666666666666665e2, 0.1333333333333333, 0.19999999999999996, 0.2666666666666666, 0.33333333333333326 ]]
In this particular case, the strategy does indeed converge.
Surprisingly, this strategy produces the desired result and is known as the Metropolis Algorithm. What the grasshopper has done is to construct a (discrete) Markov Process which has a limiting distribution (the stationary distribution) with the desired feature: sampling from this process will result in each hour being sampled in proportion to its value.
Markov Chain Theory
Let us examine what is happening in a bit more detail.
The grasshopper has started with a very simple Markov Chain: one which jumps clockwise or anticlockwise with equal probability and then modified it. But what is a Markov Chain?
A time homogeneous Markov chain is a countable sequence of random variables
such that
We sometimes say that a Markov Chain is discrete time stochastic process with the above property.
So the very simple Markov Chain can be described by
The grasshopper knows that so it can calculate without knowing . This is important because now, without knowing , the grasshopper can evaluate
where takes the maximum of its arguments. Simplifying the above by substituing in the grasshopper’s probabilities and noting that is somewhat obscure way of saying jump clockwise or anticlockwise we obtain
The Ergodic Theorem
In most studies of Markov chains, one is interested in whether a chain has a stationary distribution. What we wish to do is take a distribution and create a chain with this distribution as its stationary distribution. We will still need to show that our chain does indeed have the correct stationary distribution and we state the relevant theorem somewhat informally and with no proof.
Theorem
An irreducible, aperiodic and positive recurrent Markov chain has a unique stationary distribution.
Roughly speaking

Irreducible means it is possible to get from any state to any other state.

Aperiodic means that returning to a state having started at that state occurs at irregular times.

Positive recurrent means that the first time to hit a state is finite (for every state and more pedantically except on sets of null measure).
Note that the last condition is required when the state space is infinite – see Skrikant‘s lecture notes for an example and also for a more formal definition of the theorem and its proof.
Algorithm
Let be a probability distribution on the state space with for all and let be an ergodic Markov chain on with transition probabilities (the latter condition is slightly stronger than it need be but we will not need fully general conditions).
Create a new (ergodic) Markov chain with transition probabilities
where takes the maximum of its arguments.
Calculate the value of interest on the state space e.g. the total magnetization for each step produced by this new chain.
Repeat a sufficiently large number of times and take the average. This gives the estimate of the value of interest.
Convergence
Let us first note that the Markov chain produced by this algorithm almost trivially satisfies the detailed balance condition, for example,
Secondly since we have specified that is ergodic then clearly is also ergodic (all the transition probabilities are ).
So we know the algorithm will converge to the unique distribution we specified to provide estimates of values of interest.
Gibbs Sampling
Random Scan
For simplicity let us consider a model with two parameters and that we sample from either parameter with equal probability. In this sampler, We update the parameters in a single step.
The transition density kernel is then given by
where is the Dirac delta function.
Detailed balance
This sampling scheme satisifies the detailed balance condition. We have
In other words
Hand waving slightly, we can see that this scheme satisfies the premises of the ergodic theorem and so we can conclude that there is a unique stationary distribution and must be that distribution.
Systematic Scan
Most references on Gibbs sampling do not describe the random scan but instead something called a systematic scan.
Again for simplicity let us consider a model with two parameters. In this sampler, we update the parameters in two steps.
We observe that this is not timehomegeneous; at each step the transition matrix flips between the two transition matrices given by the individual steps. Thus although, as we show below, each individual transtion satisifies the detailed balance condition, we cannot apply the ergodic theorem as it only applies to timehomogeneous processes.
The transition density kernel is then given by
where .
Thus
Detailed balance
Suppose that we have two states and and that . Then . Trivially we have
Now suppose that
So again we have
Similarly we can show
But note that
whereas
and these are not necessarily equal.
So the detailed balance equation is not satisfied, another sign that we cannot appeal to the ergodic theorem.
An Example: The Bivariate Normal
Let us demonstrate the Gibbs sampler with a distribution which we actually know: the bivariate normal.
The conditional distributions are easily calculated to be
Let’s take a correlation of 0.8, a data point of (0.0, 0.0) and start the chain at (2.5, 2.5).
> rho :: Double
> rho = 0.8
>
> y :: (Double, Double)
> y = (0.0, 0.0)
>
> y1, y2 :: Double
> y1 = fst y
> y2 = snd y
>
> initTheta :: (Double, Double)
> initTheta = (2.5, 2.5)
We precalculate the variance needed for the sampler.
> var :: Double
> var = 1.0  rho^2
In Haskell and in the randomfu package, sampling from probability distributions is implemented as a monad. We sample from the relevant normal distributions and keep the trajectory using a writer monad.
> gibbsSampler :: Double > RVarT (W.Writer [(Double,Double)]) Double
> gibbsSampler oldTheta2 = do
> newTheta1 < rvarT (Normal (y1 + rho * (oldTheta2  y2)) var)
> lift $ W.tell [(newTheta1, oldTheta2)]
> newTheta2 < rvarT (Normal (y2 + rho * (newTheta1  y1)) var)
> lift $ W.tell [(newTheta1, newTheta2)]
> return $ newTheta2
It is common to allow the chain to “burn in” so as to “forget” its starting position. We arbitrarily burn in for 10,000 steps.
> burnIn :: Int
> burnIn = 10000
We sample repeatedly from the sampler using the monadic form of iterate. Running the monadic stack is slightly noisy but nonetheless straightforward. We use mersennerandompure64 (albeit indirectly via randomsource) as our source of entropy.
> runMCMC :: Int > [(Double, Double)]
> runMCMC n =
> take n $
> drop burnIn $
> snd $
> W.runWriter (evalStateT (sample (ML.iterateM_ gibbsSampler (snd initTheta))) (pureMT 2))
We can look at the trajectory of our sampler for various run lengths.
For bigger sample sizes, plotting the distribution sampled reassures us that we are indeed sampling from a bivariate normal distribution as the theory predicted.
Applications to Bayesian Statistics
Some of what is here and here excluding JAGS and STAN (after all this is a book about Haskell).
Applications to Physics
Applications to Physics
Most of what is here.
Pingback: Useful for referring—02202015  Honglang Wang's Blog