Comonadic Markov Chain Monte Carlo

Some time ago I came across a way to in-principle perform inference on certain probabilistic programs using comonadic structures and operations.

I decided to dig it up and try to use it to extend the simple probabilistic programming language I talked about a few days ago with a stateful, experimental inference backend. In this post we’ll

  • Represent probabilistic programs as recursive types parameterized by a terminating instruction set.
  • Represent execution traces of probabilistic programs via a simple transformation of our program representation.
  • Implement the Metropolis-Hastings algorithm over this space of execution traces and thus do some inference.

Let’s get started!

Representing Programs That Terminate

I like thinking of embedded languages in terms of instruction sets. That is: I want to be able to construct my embedded language by first defining a collection of abstract instructions and then using some appropriate recursive structure to represent programs over that set.

In the case of probabilistic programs, our instructions are probability distributions. Last time we used the following simple instruction set to define our embedded language:

data ModelF r =
    BernoulliF Double (Bool -> r)
  | BetaF Double Double (Double -> r)
  deriving Functor

We then created an embedded language by just wrapping it up in the higher-kinded Free type to denote programs of type Model.

data Free f a =
    Pure a
  | Free (f (Free f a))

type Model = Free ModelF

Recall that Free represents programs that can terminate, either by some instruction in the underlying instruction set, or via the Pure constructor of the Free type itself. The language defined by Free ModelF is expressive enough to easily construct a ‘forward-sampling’ interpreter, as well as a simple rejection sampler for performing inference.

Notice that we don’t have a terminating instruction in ModelF itself - if we’re using it, then we need to rely on the Pure constructor of Free to terminate programs. Otherwise they’d just have to recurse forever. This can be a bit limiting if we want to transform a program of type Free ModelF to something else that doesn’t have a notion of termination baked-in (Fix, for example).

Let’s tweak the ModelF type to get the following:

data ModelF a r =
    BernoulliF Double (Bool -> r)
  | BetaF Double Double (Double -> r)
  | NormalF Double Double (Double -> r)
  | DiracF a
  deriving Functor

Aside from adding another foundational distribution - NormalF - we’ve also added a new constructor, DiracF, which carries a parameter with type a. We need to incorporate this carrier type in the overall type of ModelF as well, so ModelF itself also gets a new type parameter to carry around.

The DiracF instruction is a terminating instruction; it has no recursive point and just terminates with a value of type a when reached. It’s structurally equivalent to the Pure a branch of Free that we were relying on to terminate our programs previously - the only thing we’ve done is add it to our instruction set proper.

Why DiracF? A Dirac distribution places the entirety of its probability mass on a single point, and this is the exact probabilistic interpretation of the applicative pure or monadic return that one encounters with an appropriate probability type. Intuitively, if I sample a value \(x\) from a uniform distribution, then that is indistinguishable from sampling \(x\) from said uniform distribution and then sampling from a Dirac distribution with parameter \(x\).

Make sense? If not, it might be helpful to note that there is no difference between any of the following (to which uniform and dirac are analogous):

> action :: m a
> action >>= return :: m a
> action >>= return >>= return >>= return :: m a

Wrapping ModelF a up in Free, we get the following general type for our programs:

type Program a = Free (ModelF a)

And we can construct a bunch of embedded language terms in the standard way:

beta :: Double -> Double -> Program a Double
beta a b = liftF (BetaF a b id)

bernoulli :: Double -> Program a Bool
bernoulli p = liftF (BernoulliF p id)

normal :: Double -> Double -> Program a Double
normal m s = liftF (NormalF m s id)

dirac :: a -> Program a b
dirac x = liftF (DiracF x)

Program is a general type, capturing both terminating and nonterminating programs via its type parameters. What do I mean by this? Note that in Program a b, the a type parameter can only be concretely instantiated via use of the terminating dirac term. On the other hand, the b type parameter is unaffected by the dirac term; it can only be instantiated by the other nonterminating terms: beta, bernoulli, normal, or compound expressions of these.

We can thus distinguish between terminating and nonterminating programs at the type level, like so:

type Terminating a = Program a Void

type Model b = forall a. Program a b

Void is the uninhabited type, brought into scope via Data.Void or simply defined via data Void = Void Void. Any program that ends via a dirac instruction must be Terminating, and any program that doesn’t end with a dirac instruction can not be Terminating. We’ll just continue to call a nonterminating program a Model, as before.

Good. So if it’s not clear: from a user’s perspective, nothing has changed. We still write probabilistic programs using simple monadic language terms. Here’s a Gaussian mixture model where the mixing parameter follows a beta distribution, for example:

mixture :: Double -> Double -> Model Double
mixture a b = do
  prob   <- beta a b
  accept <- bernoulli prob
  if   accept
  then normal (negate 2) 0.5
  else normal 2 0.5

Meanwhile the syntax tree generated looks something like the following. It’s more or less a traditional probabilistic graphical model description of our program:

It’s important to note that in this embedded framework, the only pieces of the syntax tree that we can observe are those related directly to our primitive instructions. For our purposes this is excellent - we can focus on programs entirely at the level of their probabilistic components, and ignore the deterministic parts that would otherwise be distractions.

To collect samples from mixture, we can first interpret it into a sampling function, and then simulate from it. The toSampler function from last time doesn’t change much:

toSampler :: Program a a -> Prob IO a
toSampler = iterM $ \case
  BernoulliF p f -> Prob.bernoulli p >>= f
  BetaF a b f    -> Prob.beta a b >>= f
  NormalF m s f  -> Prob.normal m s >>= f
  DiracF x       -> return x

Sampling from mixture 2 3 a thousand times yields the following

> simulate (toSampler (mixture 2 3))

Note that the rightmost component gets more traffic due to the hyperparameter combination of 2 and 3 that we provided to mixture.

Also, a note - since we have general recursion in Haskell, so-called ‘terminating’ programs here can actually.. uh, fail to terminate. They must only terminate as far as we can express the sentiment at the embedded language level. Consider the following, for example:

foo :: Terminating a
foo = (loop 1) >>= dirac where
  loop a = do
    p <- beta a 1
    loop p

foo here doesn’t actually terminate. But at least this kind of weird case can be picked up in the types:

> :t simulate (toSampler foo)
simulate (toSampler foo) :: IO Void

If you try to sample from a distribution over Void or forall a. a then I can’t be held responsible for what you get up to. But there are other cases, sadly, where we’re also out of luck:

trollGeometric :: Double -> Model Int
trollGeometric p = loop where
  loop = do
    accept <- return False
    if   accept
    then return 1
    else fmap succ loop

A geometric distribution that actually used its argument \(p\), for \(0 < p \leq 1\), could be guaranteed to terminate with probability 1. This one doesn’t, so trollGeometric undefined >>= dirac won’t.

At the end of the day we’re stuck with what our host language offers us. So, take the termination guarantees for our embedded language with a grain of salt.

Stateful Inference

In the previous post we used a simple rejection sampler to sample from a conditional distribution. ‘Vanilla’ Monte Carlo algorithms like rejection and importance sampling are stateless. This makes them nice in some ways - they tend to be simple to implement and are embarrassingly parallel, for example. But the curse of dimensionality prevents them from scaling well to larger problems. I won’t go into detail on that here - for a deep dive on the topic, you probably won’t find anything better than this phenomenal couple of talks on MCMC that Iain Murray gave at a MLSS session in Cambridge in 2009. I think they’re unparalleled to this day.

The point is that in higher dimensions we tend to get a lot out of state. Essentially, if one finds an interesting region of high-dimensional parameter space, then it’s better to remember where that is, rather than forgetting it exists as soon as one stumbles onto it. The manifold hypothesis conjectures that interesting regions of space tend to be near other interesting regions of space, so exploring neighbourhoods of interesting places tends to pay off. Stateful Monte Carlo methods - namely, the family of Markov chain Monte Carlo algorithms - handle exactly this, by using a Markov chain to wander over parameter space. I’ve written on MCMC in the past - you can check out some of those articles if you’re interested.

In the stateless rejection sampler we just performed conditional inference via the following algorithm:

  • Sample from a parameter model.
  • Sample from a data model, using the sample from the parameter model as input.
  • If the sample from the data model matches the provided observations, return the sample from the parameter model.

By repeating this many times we get a sample of arbitrary size from the appropriate conditional, inverse, or posterior distribution (whatever you want to call it).

In a stateful inference routine - here, the good old Metropolis-Hastings algorithm - we’re instead going to do the following repeatedly:

  • Sample from a parameter model, recording the way the program executed in order to return the sample that it did.
  • Compute the cost, in some sense, of generating the provided observations, using the sample from the parameter model as input.
  • Propose a new sample from the parameter model by perturbing the way the program executed and recording the new sample the program outputs.
  • Compute the cost of generating the provided observations using this new sample from the parameter model as input.
  • Compare the costs of generating the provided observations under the respective samples from the parameter models.
  • With probability depending on the ratio of the costs, flip a coin. If you see a head, then move to the new, proposed execution trace of the program. Otherwise, stay at the old execution trace.

This procedure generates a Markov chain over the space of possible execution traces of the program - essentially, plausible ways that the program could have executed in order to generate the supplied observations.

Implementations of Church use variations of this method to do inference, the most famous of which is a low-overhead transformational compilation procedure described in a great and influential 2011 paper by David Wingate et al.

Representing Running Programs

To perform inference on probabilistic programs according to the aforementioned Metropolis-Hastings algorithm, we need to represent executing programs somehow, in a form that enables us to examine and modify their internal state.

How can we do that? We’ll pluck another useful recursive structure from our repertoire and consider the humble Cofree:

data Cofree f a = a :< f (Cofree f a)

Recall that Cofree allows one to annotate programs with arbitrary information at each internal node. This is a great feature; if we can annotate each internal node with important information about its state - its current value, the current state of its generator, the ‘cost’ associated with it - then we can walk through the program and examine it as required. So, it can capture a ‘running’ program in exactly the way we need.

Let’s describe running programs as values having the following Execution type:

type Execution a = Cofree (ModelF a) Node

The Node type is what we’ll use to describe the internal state of each node on the program. I’ll define it like so:

data Node = Node {
    nodeCost    :: Double
  , nodeValue   :: Dynamic
  , nodeSeed    :: MWC.Seed
  , nodeHistory :: [Dynamic]
  } deriving Show

I’ll elaborate on this type below, but you can see that it captures a bunch of information about the state of each node.

One can mechanically transform any Free-encoded program into a Cofree-encoded program, so long as the original Free-encoded program can terminate of its own accord, i.e. on the level of its own instructions. Hence the need for our Terminating type and all that.

In our case, setting everything up just right takes a bit of code, mainly around handling pseudo-random number generators in a pure fashion. So I won’t talk about every little detail of it right here. The general idea is to write a function that takes instructions to the appropriate state captured by a Node value, like so:

initialize :: Typeable a => MWC.Seed -> ModelF a b -> Node
initialize seed = \case
  BernoulliF p _ -> runST $ do
    (nodeValue, nodeSeed) <- samplePurely (Prob.bernoulli p) seed
    let nodeCost    = logDensityBernoulli p (unsafeFromDyn nodeValue)
        nodeHistory = mempty
    return Node {..}

  BetaF a b _ -> runST $ do
    (nodeValue, nodeSeed) <- samplePurely (Prob.beta a b) seed
    let nodeCost    = logDensityBeta a b (unsafeFromDyn nodeValue)
        nodeHistory = mempty
    return Node {..}


You can see that for each node, I sample from it, calculate its cost, and then initialize its ‘history’ as an empty list.

Here it’s worth going into a brief aside.

There are two mildly annoying things we have to deal with in this situation. First, individual nodes in the program typically sample values at different types, and second, we can’t easily use effects when annotating. This means that we have to pack heterogeneously-typed things into a homogeneously-typed container, and also use pure random number generation facilities to sample them.

A quick-and-dirty answer for the first case is to just use dynamic typing when storing the values. It works and is easy, but of course is subject to the standard caveats. I use a function called unsafeFromDyn to convert dynamically-typed values back to a typed form, so you can gauge the safety of all this for yourself.

For the second case, I just use the ST monad, along with manual state snapshotting, to execute and iterate a random number generator. Pretty simple.

Also: in terms of efficiency, keeping a node’s history on-site at each execution falls into the ‘completely insane’ category, but let’s not worry much about efficiency right now. Prototypes gonna prototype and all that.


Given this initialize function, we can transform a terminating program into a running program by simple recursion. Again, we can only transform programs with type Terminating a because we need to rule out the case of ever visiting the Pure constructor of Free. We handle that by the absurd function provided by Data.Void:

execute :: Typeable a => Terminating a -> Execution a
execute = annotate defaultSeed where
  defaultSeed         = (42, 108512)
  annotate seeds term = case term of
    Pure r -> absurd r
    Free instruction ->
      let (nextSeeds, generator) = xorshift seeds
          seed  = MWC.toSeed (V.singleton generator)
          node  = initialize seed instruction
      in  node :< fmap (annotate nextSeeds) instruction

And there you have it - execute takes a terminating program as input and returns a running program - an execution trace - as output. The syntax tree we had previously gets turned into something like this:

Perturbing Running Programs

Given an execution trace, we’re able to step through it sequentially and investigate the program’s internal state. But to do inference we also need to modify it as well. What’s the answer here?

Just as Free has a monadic structure that allows us to write embedded programs using built-in monadic combinators and do-notation, Cofree has a comonadic structure that is amenable to use with the various comonadic combinators found in Control.Comonad. The most important for our purposes is the comonadic ‘extend’ operation that’s dual to monad’s ‘bind’:

extend :: Comonad w => (w a -> b) -> w a -> w b
extend f = fmap f . duplicate

To perturb a running program, we can thus write a function that perturbs any given annotated node, and then extend it over the entire execution trace.

The perturbNode function can be similar to the initialize function from earlier; it describes how to perturb every node based on the instruction found there:

perturbNode :: Execution a -> Node
perturbNode (node@Node {..} :< cons) = case cons of
  BernoulliF p _ -> runST $ do
    (nvalue, nseed) <- samplePurely (Prob.bernoulli p) nodeSeed
    let nscore   = logDensityBernoulli p (unsafeFromDyn nvalue)
    return $! Node nscore nvalue nseed nodeHistory

  BetaF a b _ -> runST $ do
    (nvalue, nseed) <- samplePurely (Prob.beta a b) nodeSeed
    let nscore   = logDensityBeta a b (unsafeFromDyn nvalue)
    return $! Node nscore nvalue nseed nodeHistory


Note that this is a very crude way to perturb nodes - we’re just sampling from whatever distribution we find at each one. A more refined procedure would sample from each node on a more local basis, sampling from its respective domain in a neighbourhood of its current location. For example, to perturb a BetaF node we might sample from a tiny Gaussian bubble around its current location, repeating the process if we happen to ‘fall off’ the support. I’ll leave matters like that for another post.

Perturbing an entire trace is then as easy as I claimed it to be:

perturb :: Execution a -> Execution a
perturb = extend perturbNode

For some comonadic intuition: when we ‘extend’ a function over an execution, the trace itself gets ‘duplicated’ in a comonadic context. Each node in the program becomes annotated with a view of the rest of the execution trace from that point forward. It can be difficult to visualize at first, but I reckon the following image is pretty faithful:

Each annotation then has perturbNode applied to it, which reduces the trace back to the standard annotated version we saw before.

Iterating the Markov Chain

So: to move around in parameter space, we’ll propose state changes by perturbing the current state, and then accept or reject proposals according to local economic conditions.

If you already have no idea what I’m talking about, then the phrase ‘local economic conditions’ probably didn’t help you much. But it’s a useful analogy to have in one’s head. Each state in parameter space has a cost associated with it - the cost of generating the observations that we’re conditioning on while doing inference. If certain parameter values yield a data model that is unlikely to generate the provided observations, then those observations will be expensive to generate when measured in terms of log-likelihood. Parameter values that yield data models more likely to generate the supplied observations will be comparatively cheaper.

If a proposed execution trace is significantly cheaper than the trace we’re currently at, then we usually want to move to it. We allow some randomness in our decision to keep everything nice and measure-preserving.

We can thus construct the conditional distribution over execution traces using the following invert function, using the same nomenclature as the rejection sampler we used previously. To focus on the main points, I’ll elide some of its body:

  :: (Eq a, Typeable a, Typeable b)
  => Int -> [a] -> Model b -> (b -> a -> Double)
  -> Model (Execution b)
invert epochs obs prior ll = loop epochs (execute terminated) where
  terminated = prior >>= dirac
  loop n current
    | n == 0    = return current
    | otherwise = do
        let proposal = perturb current

            -- calculate costs and movement probability here

        accept <- bernoulli prob
        let next = if accept then proposal else stepGenerators current
        loop (pred n) (snapshot next)

There are a few things to comment on here.

First, notice how the return type of invert is Model (Execution b)? Using the semantics of our embedded language, it’s literally a standard model over execution traces. The above function returns a first-class value that is completely uninterpreted and abstract. Cool.

We’re also dealing with things a little differently from the rejection sampler that we built previously. Here, the data model is expressed by a cost function; that is, a function that takes a parameter value and observation as input, and returns the cost of generating the observation (conditional on the supplied parameter value) as output. This is the approach used in the excellent Practical Probabilistic Programming with Monads paper by Adam Scibior et al and also mentioned by Dan Roy in his recent talk at the Simons Institute. Ideally we’d just reify the cost function here from the description of a model directly (to keep the interface similar to the one used in the rejection sampler implementation), but I haven’t yet found a way to do this in a type-safe fashion.

Regardless of whether or not we accept a proposed move, we need to snapshot the current value of each node and add it to that node’s history. This can be done using another comonadic extend:

snapshotValue :: Cofree f Node -> Node
snapshotValue (Node {..} :< cons) = Node { nodeHistory = history, .. } where
  history = nodeValue : nodeHistory

snapshot :: Functor f => Cofree f Node -> Cofree f Node
snapshot = extend snapshotValue

The other point of note is minor, but an extremely easy detail to overlook. Since we’re handling random value generation at each node purely, using on-site PRNGs, we need to iterate the generators forward a step in the event that we don’t accept a proposal. Otherwise we’d propose a new execution based on the same generator states that we’d used previously! For now I’ll just iterate the generators by forcing a sample of a uniform variate at each node, and then throwing away the result. To do this we can use the now-standard comonadic pattern:

stepGenerator :: Cofree f Node -> Node
stepGenerator (Node {..} :< cons) = runST $ do
  (nval, nseed) <- samplePurely (Prob.beta 1 1) nodeSeed
  return Node {nodeSeed = nseed, ..}

stepGenerators :: Functor f => Cofree f Node -> Cofree f Node
stepGenerators = extend stepGenerator

Inspecting Execution Traces

Alright so let’s see how this all works. Let’s write a model, condition it on some observations, and do inference.

We’ll choose our simple Gaussian mixture model from earlier, where the mixing probability follows a beta distribution, and cluster assignment itself follows a Bernoulli distribution. We thus choose the ‘leftmost’ component of the mixture with the appropriate mixture probability.

We can break the mixture model up as follows:

prior :: Double -> Double -> Model Bool
prior a b = do
  p <- beta a b
  bernoulli p

likelihood :: Bool -> Model Double
likelihood left
  | left      = normal (negate 2) 0.5
  | otherwise = normal 2 0.5

Let’s take a look at some samples from the marginal distribution. This time I’ll flip things and assign hyperparameters of 3 and 2 for the prior:

> simulate (toSampler (prior 3 2 >>= likelihood))

It looks like we’re slightly more likely to sample from the left mixture component than the right one. Again, this makes sense - the mean of a beta(3, 2) distribution is 0.6.

Now, what about inference? I’ll define the conditional model as follows:

posterior :: Model (Execution Bool)
posterior = invert 1000 obs prior ll where
  obs = [ -1.7, -1.8, -2.01, -2.4
        , 1.9, 1.8

  ll left
    | left      = logDensityNormal (negate 2) 0.5
    | otherwise = logDensityNormal 2 0.5

Here we have four observations that presumably arise from the leftmost component, and only two that match up with the rightmost. Note also that I’ve replaced the likelihood model with its appropriate cost function due to reasons I mentioned in the last section. (It would be easy to reify this model as its cost function, but doing it for general models is trickier)

Anyway, let’s sample from the conditional distribution:

> simulate (toSampler posterior)

Sampling returns a running program, of course, and we can step through it to examine its structure. We can use the supplied values recorded at each node to ‘automatically’ step through execution, or we can supply our own values to investigate arbitrary branches.

The conditional distribution we’ve found over the mixing probability is as follows:

Looks like we’re in the right ballpark.

We can examine the traces of other elements of the program as well. Here’s the recorded distribution over component assignments, for example - note that the rightmost bar here corresponds to the leftmost component in the mixture:

You can see that whenever we wandered into the rightmost component, we’d swiftly wind up jumping back out of it:


This is a fun take on probabilistic programming. In particular I find a few aspects of the whole setup to be pretty attractive:

We use a primitive, limited instruction set to parameterize both programs - via Free - and running programs - via Cofree. These off-the-shelf recursive types are used to wrap things up and provide most of our required control flow automatically. It’s easy to transparently add structure to embedded programs built in this way; for example, we can statically encode independence by replacing our ModelF a type with something like:

data InstructionF a = Coproduct (ModelF a) (Ap (ModelF a))

This can be hidden from the user so that we’re left with the same simple monadic syntax we presently enjoy, but we also get to take independence into account when performing inference, or any other structural interpretation for that matter.

When it comes to inference, the program representation is completely separate from whatever inference backend we choose to augment it with. We can deal with traces as first-class values that can be directly stored, inspected, manipulated, and so on. And everything is done in a typed and purely-functional framework. I’ve used dynamic typing functionality from Data.Dynamic to store values in execution traces here, but we could similarly just define a concrete Value type with the appropriate constructors for integers, doubles, bools, etc., and use that to store everything.

At the same time, this is a pretty early concept - doing inference efficiently in this setting is another matter, and there are a couple of computational and statistical issues here that need to be ironed out to make further progress.

The current way I’ve organized Markov chain generation and iteration is just woefully inefficient. Storing the history of each node on-site is needlessly costly and I’m sure results in a ton of unnecessary allocation. On a semantic level, it also ‘complects’ state and identity: why, after all, should a single execution trace know anything about traces that preceded it? Clearly this should be accumulated in another data structure. There is a lot of other low-hanging fruit around strictness and PRNG management as well.

From a more statistical angle, the present implementation does a poor job when it comes to perturbing execution traces. Some changes - such as improving the proposal mechanism for a given instruction - are easy to implement, and representing distributions as instructions indeed makes it easy to tailor local proposal distributions in a context-independent way. But another problem is that, by using a ‘blunt’ comonadic extend, we perturb an execution by perturbing every node in it. In general it’s better to make small perturbations rather than large ones to ensure a reasonable acceptance ratio, but to do that we’d need to perturb single nodes (or at least subsets of nodes) at a time.

There may be some inroads here via comonad transformers like StoreT or lenses that would allow us to zoom in on a particular node and perturb it, rather than perturbing everything at once. But my comonad-fu is not yet quite at the required level to evaluate this, so I’ll come back to that idea some other time.

I’m interested in playing with this concept some more in the future, though I’m not yet sure how much I expect it to be a tenable way to do inference at scale. If you’re interested in playing with it, I’ve dumped the code from this post into this gist.

Thanks to Niffe Hermansson and Fredrik Olsen for reviewing a draft of this post and providing helpful comments.

A Simple Embedded Probabilistic Programming Language

What does a dead-simple probabilistic programming language look like? The simplest thing I can imagine involves three components:

  • A representation for probabilistic models.
  • A way to simulate from those models (‘forward’ sampling).
  • A way to sample from a conditional model (‘backward’ sampling).

Rob Zinkov wrote an article on this type of thing around a year ago, and Dan Roy recently gave a talk on the topic as well. In the spirit of unabashed unoriginality, I’ll give a sort of composite example of the two. Most of the material here comes directly from Dan’s talk; definitely check it out if you’re curious about this whole probabilistic programming mumbojumbo.

Let’s whip together a highly-structured, typed, embedded probabilistic programming language - the core of which will encompass a tiny amount of code.

Some preliminaries - note that you’ll need my simple little mwc-probability library handy for when it comes time to do sampling:

{-# LANGUAGE DeriveFunctor #-}
{-# LANGUAGE LambdaCase #-}

import Control.Monad
import Control.Monad.Free
import qualified System.Random.MWC.Probability as MWC

Representing Probabilistic Models

Step one is to represent the fundamental constructs found in probabilistic programs. These are abstract probability distributions; I like to call them models:

data ModelF r =
    BernoulliF Double (Bool -> r)
  | BetaF Double Double (Double -> r)
  deriving Functor

type Model = Free ModelF

Each foundational probability distribution we want to consider is represented as a constructor of the ModelF type. You can think of them as probabilistic instructions, in a sense. A Model itself is a program parameterized by this probabilistic instruction set.

In a more sophisticated implementation you’d probably want to add more primitives, but you can get pretty far with the beta and Bernoulli distributions alone. Here are some embedded language terms, only two of which correspond one-to-one with to the constructors themselves:

bernoulli :: Double -> Model Bool
bernoulli p = liftF (BernoulliF p id)

beta :: Double -> Double -> Model Double
beta a b = liftF (BetaF a b id)

uniform :: Model Double
uniform = beta 1 1

binomial :: Int -> Double -> Model Int
binomial n p = fmap count coins where
  count = length . filter id
  coins = replicateM n (bernoulli p)

betaBinomial :: Int -> Double -> Double -> Model Int
betaBinomial n a b = do
  p <- beta a b
  binomial n p

You can build a lot of other useful distributions by just starting from the beta and Bernoulli as well. And technically I guess the more foundational distributions to use here would be the Dirichlet and categorical, of which the beta and Bernoulli are special cases. But I digress. The point is that other distributions are easy to construct from a set of reliable primitives; you can check out the old lambda-naught paper by Park et al for more examples.

See how binomial and betaBinomial are defined? In the case of binomial we’re using the property that models have a functorial structure by just mapping a counting function over the result of a bunch of Bernoulli random variables. For betaBinomial we’re directly making use of our monadic structure, first describing a weight parameter via a beta distribution and then using it as an input to a binomial distribution.

Note in particular that we’ve expressed betaBinomial by binding a parameter model to a data model. This is a foundational pattern in Bayesian statistics; in the more usual lingo, the parameter model corresponds to the prior distribution, and the data model is the likelihood.

Forward-Mode Sampling

So we have our representation. Next up, we want to simulate from these models. Thus far they’re purely abstract, and don’t encode any information about probability or sampling or what have you. We have to ascribe that ourselves.

mwc-probability defines a monadic sampling-based probability distribution type called Prob, and we can use a basic recursion scheme on free monads to adapt our own model type to that:

toSampler :: Model a -> MWC.Prob IO a
toSampler = iterM $ \case
  BernoulliF p f -> MWC.bernoulli p >>= f
  BetaF a b f    -> MWC.beta a b >>= f

We can glue that around the relevant mwc-probability functionality to simulate from models directly:

simulate :: Model a -> IO a
simulate model = MWC.withSystemRandom . MWC.asGenIO $
  MWC.sample (toSampler model)

And this can be used with standard monadic combinators like replicateM to collect larger samples:

> replicateM 10 $ simulate (betaBinomial 10 1 4)

Reverse-Mode Sampling

Now. Here we want to condition our model on some observations and then recover the conditional distribution over its internal parameters.

This part - inference - is what makes probabilistic programming hard, and doing it really well remains an unsolved problem. One of the neat theoretical results in this space due to Ackerman, Freer, and Roy is that in the general case the problem is actually unsolvable, in that one can encode as a probabilistic program a conditional distribution that computes the halting problem. Similarly, in general it’s impossible to do this sort of thing efficiently even for computable conditional distributions. Consider the case of a program that returns the hash of a random n-long binary string, and then try to infer the distribution over strings given some hashes, for example. This is never going to be a tractable problem.

For now let’s use a simple rejection sampler to encode a conditional distribution. We’ll require some observations, a proposal distribution, and the model that we want to invert:

invert :: (Monad m, Eq b) => m a -> (a -> m b) -> [b] -> m a
invert proposal model observed = loop where
  loop = do
    parameters <- proposal
    generated  <- replicateM (length observed) (model parameters)
    if   generated == observed
    then return parameters
    else loop

Let’s use it to compute the posterior or inverse model of an (apparently) biased coin, given a few observations. We’ll just use a uniform distribution as our proposal:

posterior :: Model Double
posterior = invert [True, True, False, True] uniform bernoulli

Let’s grab some samples from the posterior distribution:

> replicateM 1000 (simulate posterior)

The central tendency of the posterior floats about 0.75, which is what we’d expect, given our observations. This has been inferred from only four points; let’s try adding a few more. But before we do that, note that the present way the rejection sampling algorithm works is:

  • Propose a parameter value according to the supplied proposal distribution.
  • Generate a sample from the model, of equal size to the supplied observations.
  • Compare the collected sample to the supplied observations. If they’re equal, then return the proposed parameter value. Otherwise start over.

Rejection sampling isn’t exactly efficient in nontrivial settings anyway, but it’s supremely inefficient for our present case. The random variables we’re interested in are exchangeable, so what we’re concerned about is the total number of True or False values observed - not any specific order they appear in.

We can add an ‘assistance’ function to the rejection sampler to help us out in this case:

  :: (Monad m, Eq c) => ([a] -> c) -> m b -> (b -> m a) -> [a] -> m b
invertWithAssistance assister proposal model observed = loop where
  loop = do
    parameters <- proposal
    generated  <- replicateM (length observed) (model parameters)
    if   assister generated == assister observed
    then return parameters
    else loop

The assister summarizes both our observations and collected sample to ensure they’re efficiently comparable. In our situation, we can use a simple counting function to tally up the number of True values we observe:

count :: [Bool] -> Int
count = length . filter id

Now let’s create another posterior by conditioning on a few more observations:

posterior0 :: Model Double
posterior0 = invertWithAssitance count uniform bernoulli obs where
  obs =
    [True, True, True, False, True, True, False, True, True, True, True, False]

and collect another thousand samples from it. This would likely take an annoying amount of time without the use of our count function for assistance above:

> replicateM 1000 (simulate posterior0)

Note that with more information to condition on, we get a more informative posterior.


This is a really basic formulation - too basic to be useful in any meaningful way - but it illustrates some of the most important concepts in probabilistic programming. Representation, simulation, and inference.

I think it’s also particularly nice to do this in Haskell, rather than something like Python (which Dan used in his talk) - it provides us with a lot of extensible structure in a familiar framework for language hacking. It sort of demands you’re a fan of all these higher-kinded types and structured recursions and all that, but if you’re reading this blog then you’re probably in that camp anyway.

I’ll probably write a few more little articles like this over time. There are a ton of improvements that we can make to this basic setup - encoding independence, sampling via MCMC, etc. - and it might be fun to grow everything out piece by piece.

I’ve dropped the code from this post into this gist.

Randomness in Haskell

Randomness is a constant nuisance point for Haskell beginners who may be coming from a language like Python or R. While in Python you can just get away with something like:

In [2]: numpy.random.rand(3)
Out[2]: array([ 0.61426175,  0.05309224,  0.38861597])

or in R:

> runif(3)
[1] 0.49473012 0.68436352 0.04135914

In Haskell, the situation is more complicated. It’s not too much worse when you get the hang of things, but it’s certainly one of those things that throws beginners for a loop - and for good reason.

In this article I want to provide a simple guide, with examples, for getting started and becoming comfortable with randomness in Haskell. Hopefully it helps!

I’m writing this from a hotel during my girlfriend’s birthday, so it’s being slapped together very rapidly with a kind of get-it-done attitude. If anything is unclear or you have any questions, feel free to shoot me a ping and I’ll try to improve it when I get a chance.

Randomness on Computers in General

Check out the R code I posted previously. If you just open R and type runif(3) on your machine, then odds are you’ll get a different triple of numbers than what I got above.

These numbers are being generated based on R’s global random number generator (RNG), which, absent any fiddling by the user, is initialized as needed based on the system time and ID of the R process. So: if you open up the R interpreter and call runif(3), then behind the scenes R will initialize the RNG based on the time and process ID, and then use a particular algorithm to generate random numbers based on that initialized value (called the ‘seed’).

These numbers aren’t truly random - they’re pseudo-random, which means they’re generated by a deterministic algorithm such that the resulting values appear random over time. The default algorithm used by R, for example, is the famous Mersenne Twister, which you can verify as follows:

> RNGkind()
[1] "Mersenne-Twister" "Inversion"

You can also set the seed yourself in R, using the set.seed function. Then if you type something like runif(3), R will use this initialized RNG rather than coming up with its own seed based on the time and process ID. Setting the seed allows you to reproduce operations involving pseudo-random numbers; just re-set the seed and perform the same operations again:

> set.seed(42)
> runif(3)
[1] 0.9148060 0.9370754 0.2861395
> set.seed(42)
> runif(3)
[1] 0.9148060 0.9370754 0.2861395

(It’s good practice to always initialize the RNG using some known seed before running an experiment, simulation, and so on.)

So the big thing to notice here, in any case, is that R uses a global RNG. It maintains the state of this RNG implicitly and behind the scenes. When you type runif(3), R consults this implicit RNG, gives you your pseudo-random numbers based on its value, and updates the global RNG without you needing to worry about any of this plumbing yourself. The same is generally true for randomness in most programming languages - Python, C, Ruby, and so on.

Explicit RNG Management

But let’s come back to Haskell. Haskell, unlike R or Python, is purely-functional. State, or effects in general, are never implicit in the same way that R updates its global RNG. We need to either explicitly pass around a RNG ourselves, or at least allow some explicit monad to do it for us.

Passing around a RNG manually is annoying, so in practice this means everyone uses a monad to handle RNG state. This means that one needs to be comfortable working with monadic code in order to practically use random numbers in Haskell, which presents a big hurdle for beginners who may have been able to ignore monads thus far on their Haskell journey.

Let’s see what I mean by all of this by going through a few examples. Make sure you have stack installed, and then grab a few libraries that we’ll make use of in the remainder of this post:

$ stack install random mwc-random primitive

The Really Annoying Method - Manual RNG Management

Let me demonstrate the simplest conceptual method for dealing with random numbers: manually grabbing and passing around a RNG without involving any monads whatsoever.

First, open up GHCi:

$ stack ghci

And let’s also get some quick preliminaries out of the way:

Prelude> :set prompt "> "
> import System.Random
> import Control.Monad
> let runif_pure = randomR (0 :: Double, 1)
> let runif n = replicateM n (randomRIO (0 :: Double, 1))
> let set_seed = setStdGen . mkStdGen

We’ll first use the basic System.Random module for illustration. To initialize a RNG, we can make one by providing the mkStdGen function with an integer seed:

> let rng = mkStdGen 42
> rng
43 1

We can use this thing to generate random numbers. A simple function to do that is randomR, which will generate pseudo-random values for some ordered type in a given range. We’ll use the runif_pure alias for it that we defined previously, just to make things look similar to the previous R example and also emphasize that this one is a pure function:

> runif_pure rng
(1.0663729393723398e-2,2060101257 2103410263)

You can see that we got back a pair of values, the first element of which is our random number 1.0663729393723398e-2. Cool. Let’s try to generate another:

> runif_pure rng
(1.0663729393723398e-2,2060101257 2103410263)

Hmm. We generated the same number again. This is because the value of rng hasn’t changed - it’s still the same value we made via mkStdGen 42. Since we’re using the same random number generator to generate a pseudo-random value, we get the same pseudo-random value.

If we want to make new random numbers, then we need to use a different generator. And the second element of the pair returned from our call to runif_pure is exactly that - an updated RNG that we can use to generate additional random numbers.

Let’s try that all again, using the generator we get back from the first function call as an input to the second:

> let (x, rng1) = runif_pure rng
> x
> let (y, rng2) = runif_pure rng1
> y


I mean.. sort of. It works and all, and it does constitute a general-purpose solution. But manually binding updated RNG states to names and swapping those in for new values is still pretty annoying.

You could also generate an infinite list of random numbers using the randomRs function and just take from it as needed, but you still probably need to manage that list to make sure you don’t re-use any numbers. You kind of trade off managing the RNG for managing an infinite list of random numbers, which isn’t much better.

The Less-Annoying Method - Get A Monad To Do It

The good news is that we can offload the job of managing the RNG state to a monad. I won’t actually explain how that works in detail here - I think most people facing this problem are initially more concerned with getting something working, rather than deeply grokking monads off the bat - so I’ll just claim that we can get a monad to handle the RNG state for us, and that will hopefully (mostly) suffice for now.

Still rolling with the System.Random module for the time being, we’ll use the runif alias for the randomRIO function that we defined previously to generate some new random numbers:

> runif 3
> runif 3

Simpler! Notice we haven’t had to do anything with a generator manually - we just ask for random numbers and then get them, just like in R. And if we want to set the value of the RNG being used here, we can use the setStdGen function with an RNG that we’ve already created. Here let’s just use the set_seed alias we defined earlier, to mimic R’s set.seed function:

> set_seed 42
> runif 3
> set_seed 42
> runif 3

So things are similar to how they work in R here - we have a global RNG of sorts, and we can set its state as desired using the set_seed function. But since this is Haskell, the effects of creating and updating the generator state must still be explicit. And they are explicit - it’s just that they’re explicit in the type of runif:

> :t runif
runif :: Int -> IO [Double]

Note that runif returns a value that’s wrapped up in IO. This is how we indicate explicitly - at the type level - that something is being done with the generator in the background. IO is a monad, and it happens to be the thing that’s dealing with the generator for us here.

What this means for you, the practitioner, is that you can’t just mix values of some type a with values of type IO a willy-nilly. You may be writing a function f with type [Double] -> Double, where the input list of doubles is intended to be randomly-generated. But if you just go ahead and generate a list xs of random numbers, they’ll have type IO [Double], and you’ll stare in confusion at some type error from GHC when you try to apply f to xs.

Here’s what I mean. Take the example of just generating some random numbers and then summing them up. First, in R:

> xs = runif(3)
> sum(xs)
[1] 1.20353

And now in Haskell, using the same mechanism we tried earlier:

> let xs = runif 3
> :t xs
xs :: IO [Double]
> sum xs
    No instance for (Num [Double]) arising from a use of ‘sum’
    In the expression: sum xs
    In an equation for ‘it’: it = sum xs

This means that to deal with the numbers we generate, we have to treat them a little differently than we would in R, or compared to the situation where we were managing the RNG explicitly in Haskell. Concretely: if we use a monad to manage the RNG for us, then the numbers we generate will be ‘tagged’ by the monad. So we need to do something or other to make those tagged numbers work with ‘untagged’ numbers, or functions designed to work with ‘untagged’ numbers.

This is where things get confusing for beginners. Here’s how we could add up some random numbers in GHCi:

> xs <- runif 3
> sum xs

We’ve used the <- symbol to bind the result of runif 3 to the name xs, rather than let xs = .... But this is sort of particular to running code in GHCi; if you try to do this in a generic Haskell function, you’ll possibly wind up with some more weird type errors. To do this in regular ol’ Haskell code, you need to both use <--style binding and also acknowledge the ‘tagged’ nature of randomly-generated values.

The crux is that, when you’re using a monad to generate random numbers in Haskell, you need to separate generating them from using them. Rather than try to explain what I mean here precisely, let’s rely on example, and implement a simple Metropolis sampler for illustration.

A Metropolis Sampler

The Metropolis algorithm will help you approximate expectations over certain probability spaces. Here’s how it works. Picture yourself strolling around some bumpy landscape; you want to walk around it in such a fashion that you visit regions of it with probability proportional to their altitude. To do that, you can repeatedly:

  1. Pick a random point near your current location.
  2. Compare your present altitude to the altitude of that point you picked. Calculate a probability based on their ratio.
  3. Flip a coin where the chance of observing a head is equal to that probability. If you get a head, move to the location you picked. Otherwise, stay put.

Let’s implement it in Haskell, using a monadic random number generator to do so. This time we’re going to use mwc-random - a more industrial-strength randomness library that you can confidently use in production code.

mwc-random uses Marsaglia’s multiply-with-carry algorithm to generate pseudo-random numbers. It requires you to explicitly create and pass a RNG to functions that need to generate random numbers, but it uses a monad to update the RNG state itself. This winds up being pretty nice; let’s dive in to see.

Create a module called Metropolis.hs and get some imports out of the way:

module Metropolis where

import Control.Monad
import Control.Monad.Primitive
import System.Random.MWC as MWC
import System.Random.MWC.Distributions as MWC

Step One

The first thing we want to do is implement is point (1) from above:

Pick a random point near your current location.

We’ll just use a standard normal distribution of the appropriate dimension to do this - we just want to take a location, perturb it, and return the perturbed location.

propose :: [Double] -> Gen RealWorld -> IO [Double]
propose location rng = traverse (perturb rng) location where
  perturb gen x = MWC.normal x 1 gen

So at finer detail: we’re walking over the coordinates of the current location and generating a normally-distributed value centered at each coordinate. The MWC.normal function will do this for a given mean and standard deviation, and we can use the traverse function to walk over each coordinate.

Note that we pass a mwc-random RNG - the value with type Gen RealWorld - to the propose function. We need to supply this generator anywhere we want to generate random numbers, but we don’t need to manually worry about tracking and updating its state. The IO monad will do that for us. The resulting randomly-generated values will be tagged with IO, so we’ll need to deal with that appropriately.

Step Two

Now let’s implement point (2):

Compare your present altitude to the altitude of that point you picked. Calculate a probability based on their ratio.

So, we need a function that will compare the altitude of our current point to the altitude of a proposed point and compute a probability from that. The following will do: it takes a function that will compute a (log-scale) altitude for us, as well as the current and proposed locations, and returns a probability.

moveProbability :: ([Double] -> Double) -> [Double] -> [Double] -> Double
moveProbability altitude current proposed =
    whenNaN 0 (exp (min 0 (altitude proposed - altitude current)))
    whenNaN val x
      | isNaN x   = val
      | otherwise = x

Step Three

Finally, the third step of the algorithm:

Flip a coin where the chance of observing a head is equal to that probability. If you get a head, move to the location you picked. Otherwise stay put.

So let’s get to it:

decide :: [Double] -> [Double] -> Double -> Gen RealWorld -> IO [Double]
decide current proposed prob rng = do
  accept <- MWC.bernoulli prob rng
  return $
    if   accept
    then proposed
    else current

Here we need to flip a coin, so we require a source of randomness again. The decide function thus takes another generator of type Gen RealWorld that we then supply to the MWC.bernoulli function, and the result - the final location - is once again wrapped in IO.

This function clearly demonstrates the typical way that you’ll deal with random numbers in Haskell code. decide is a monadic function, so it proceeds using do-notation. When you need to generate a random value - here we generate a random True or False value according to a Bernoulli distribution - you bind the result to a name using the <- symbol. Then afterwards, in the scope of the function, you can use the bound value as if it were pure. But the entire function must still return a ‘wrapped-up’ value that makes the effect of passing the generator explicit at the type level; right here, that means that the value will be wrapped up in IO.

Putting Everything Together

The final Metropolis transition is a combination of steps one through three. We can put them together like so:

metropolis :: ([Double] -> Double) -> [Double] -> Gen RealWorld -> IO [Double]
metropolis altitude current rng = do
  proposed <- propose current rng
  let prob = moveProbability altitude current proposed
  decide current proposed prob rng

Again, metropolis is monadic, so we start off with a do to make monadic programming easy on us. Whenever we need a random value, we bind the result of a random number-returning function using the <- notation.

The propose function returns a random location, so we bind its result to the name proposed using the <- symbol. The moveProbability function, on the other hand, is pure - so we bind that using a let prob = ... expression. The decide function returns a random value, so we can just plop it right on the end here. The entire result of the metropolis function is random, so it is wrapped up in IO.

The result of metropolis is just a single transition of the Metropolis algorithm, which involves doing this kind of thing over and over. If we do that, we observe a bunch of points that trace out a particular realization of a Markov chain, which we can generate as follows:

  :: Int -> ([Double] -> Double) -> [Double] -> Gen RealWorld -> IO [[Double]]
chain epochs altitude origin rng = loop epochs [origin] where
  loop n history@(current:_)
    | n <= 0    = return history
    | otherwise = do
        next <- metropolis altitude current rng
        loop (n - 1) (next:history)

An Example

Now that we have our chain function, we can use it to trace out a collection of points visited on a realization of a Markov chain. Remember that we’re supposed to be wandering over some particular abstract landscape; here, let’s stroll over the one defined by the following function:

landscape :: [Double] -> Double
landscape [x0, x1] =
  -0.5 * (x0 ^ 2 * x1 ^ 2 + x0 ^ 2 + x1 ^ 2 - 8 * x0 - 8 * x1)

What we’ll now do is pick an origin to start from, wander over the landscape for some number of steps, and then print the resulting realization of the Markov chain to stdout. We’ll do all that through the following main function:

main :: IO ()
main = do
  rng <- MWC.createSystemRandom
  let origin = [-0.2, 0.3]
  trace <- chain 1000 landscape origin rng
  mapM_ print trace

Running that will dump a trace to stdout. If you clean it up and plot it, you’ll see that the visited points have traced out a rough approximation of the landscape:


Hopefully this gives a broad idea of how to go about using random numbers in Haskell. I’ve talked about:

  • Why randomness in Haskell isn’t as simple as randomness in (say) Python or R.
  • How to handle randomness in Haskell, either by manual generator management or by offloading that job to a monad.
  • How to get thing done when a monad manages the generator for you - separating random number generation from random number processing.
  • Doing all the above with an industrial-strength RNG, using a simple Metropolis algorithm as an example.

Hopefully the example gives you an idea of how to work with random numbers in practice.

I’ll be the first to admit that randomness in Haskell requires more work than randomness in a language like R, which to this day remains my go-to interactive data analysis language of choice. Using randomness effectively in Haskell requires a decent understanding of how to work with monadic code, even if one doesn’t quite understand monads entirely yet.

What I can say is that when one has developed some intuition for monads - acquiring a ‘feel’ for how to work with monadic functions and values - the difficulty and awkwardness drop off a bit, and working with randomness feels no different than working with any other effect.

Happy generating! I’ve dumped the code for the Metropolis example into a gist.

For a more production-quality Metropolis sampler, you can check out my mighty-metropolis library, which is a member of the declarative suite of MCMC algos.

On Measurability

.. this one is pretty dry, I’ll admit. David Williams said it best:

.. Measure theory, that most arid of subjects when done for its own sake, becomes amazingly more alive when used in probability, not only because it is then applied, but also because it is immensely enriched.

Unfortunately for you, dear reader, we won’t be talking about probability.

Moving on. What does it mean for something to be measurable in the mathematical sense? Take some arbitrary collection \(X\) and slap an appropriate algebraic structure \(\mathcal{X}\) on it - usually an algebra or \(\sigma\)-algebra, etc. Then we can refer to a few different objects as ‘measurable’, going roughly as follows.

The elements of the structure \(\mathcal{X}\) are called measurable sets. They’re called this because they can literally be assigned a notion of measure, whcih is a kind of generalized volume. If we’re just talking about some subset of \(X\) out of the context of its structure then we can be pedantic and call it measurable with respect to \(\mathcal{X}\), say. You could also call a set \(\mathcal{X}\)-measurable, to be similarly precise.

The product of the original collection and its associated structure \((X, \mathcal{X})\) is called a measurable space. It’s called that because it can be completed with a measuring function \(\mu\) - itself called a measure - that assigns notions of measure to measurable sets.

Now take some other measurable space \((Y, \mathcal{Y})\) and consider a function \(f\) from \(X\) to \(Y\). This is a measurable function if it satisfies the following technical requirement: that for any \(\mathcal{Y}\)-measurable set, its preimage under \(f\) is an element of \(\mathcal{X}\) (so: the preimage under \(f\) is \(\mathcal{X}\)-measurable).

The concept of measurability for functions probably feels the least intuitive of the three - like one of those dry taxonomical classifications that we just have to keep on the books. The ‘make sure your function is measurable and everything will be ok’ heuristic will get you pretty far. But there is some good intuition available, if you want to look for it.

Here’s an example: define a set \(X\) that consists of the elements \(A\), \(B\), and \(C\). To talk about measurable functions, we first need to define our measurable sets. The de-facto default structure used for this is a \(\sigma\)-algebra, and we can always generate one from some underlying class of sets. Let’s do that from the following plain old partition that splits the original collection into a couple of disjoint ‘slices’:

\[H = \{\{A, B\}, \{C\}\}\]

The \(\sigma\)-algebra \(\mathcal{X}\) generated from this partition will just be the partition itself, completed with the whole set \(X\) and the empty set. To be clear, it’s the following:

\[\mathcal{X} = \left\{\{A, B, C\}, \{A, B\}, \{C\}, \emptyset\right\}\]

The resulting measurable space is \((X, \mathcal{X})\). So we could assign a notion of generalized volume to any element of \(\mathcal{X}\), though I won’t actually worry about doing that here.

Now. Let’s think about some functions from \(X\) to the real numbers, which we’ll assume to be endowed with a suitable \(\sigma\)-algebra of their own (one typically assumes the standard topology on \(\mathbb{R}\) and the associated Borel \(\sigma\)-algebra).

How about this - a simple indicator function on the slice containing \(C\):

\[f(x) = \begin{cases} 0, \, x \in \{A, B\} \\ 1, \, x \in \{C\} \end{cases}\]

Is it measurable? That’s easy to check. The preimage of \(\{0\}\) is \(\{A, B\}\), the preimage of \(\{1\}\) is \(\{C\}\), and the preimage of \(\{0, 1\}\) is \(X\) itself. Those are all in \(\mathcal{X}\), and the preimage of the empty set is the empty set, so we’re good.

Ok. What about this one:

\[g(x) = \begin{cases} 0, \, x \in \{A\} \\ 1, \, x \in \{B\} \\ 2, \, x \in \{C\} \end{cases}\]

Check the preimage of \(\{1, 2\}\) and you’ll find it’s \(\{B, C\}\). But that’s not a member of \(\mathcal{X}\), so \(g\) is not measurable!

What happened here? Failing to satisfying technical requirements aside: what, intuitively, made \(f\) measurable where \(g\) wasn’t?

The answer is a problem of resolution. Look again at \(\mathcal{X}\):

\[\left\{\{A, B, C\}, \{A, B\}, \{C\}, \emptyset\right\}\]

The structure \(\mathcal{X}\) that we’ve endowed our collection \(X\) with is too coarse to permit distinguishing between elements of the slice \(\{A, B\}\). There is no measurable set \(A\), nor a measurable set \(B\) - just a measurable set \(\{A, B\}\). And as a result, if we define a function that says something about either \(A\) or \(B\) without saying the same thing about the other, that function won’t be measurable. The function \(f\) assigned the same value to both \(A\) and \(B\), so we didn’t have any problem there.

If we want to be able to distinguish between \(A\) and \(B\), we’ll need to equip \(X\) with some structure that has a finer resolution. You can check that if you make a measurable space out of \(X\) and its power set (the set of all subsets of \(X\)) then \(g\) will be measurable there, for example.

So if we’re using partitions to define our measurable sets, we get a neat little property: for any measurable function, elements in the same slice of the partition must have the same value when passed through the function. So if you have a function \(h : X \to H\) that takes an element to its respective slice in a partition, you know that \(h(x_{0}) = h(x_{1})\) for any \(x_{0}\), \(x_{1}\) in \(X\) implies that \(f(x_{0}) = f(x_{1})\) for any measurable function \(f\).


Whipping together a measurable space using a \(\sigma\)-algebra generated by a partition of sets occurs naturally when we talk about correlated equilibrium, a solution concept in non-cooperative game theory. It’s common to say a function - in that context a correlated strategy - must be measurable ‘with respect to the partition’, which sort of elides the fact that we still need to generate a \(\sigma\)-algebra from it anyway.

Some oldschool authors (Halmos, at least) developed their measure theory using \(\sigma\)-rings, but this doesn’t seem very popular nowadays. Since a ring doesn’t require including the entire set \(X\), you need to go through an awkward extra hoop when defining measurability on functions. But regardless, it’s interesting to think about what happens when one uses different structures to define measurable sets!

Making a Market

Suppose you’re in the derivatives business. You are interested in making a market on some events; say, whether or not your friend Jay will win tomorrow night’s poker game, or that the winning pot will be at least USD 100. Let’s examine some rules about how you should do business if you want this venture to succeed.

What do I mean by ‘make a market’? I mean that you will be willing to buy and sell units of a particular security that will be redeemable from the seller at some particular value after tomorrow’s poker game has ended (you will be making a simple prediction market, in other words). You can make bid offers to buy securities at some price, and ask offers to sell securities at some price.

To keep things simple let’s say you’re doing this gratis; society rewards you extrinsically for facilitating the market - your friends will give you free pizza at the game, maybe - so you won’t levy any transaction fees for making trades. Also scarcity isn’t a huge issue, so you’re willing to buy or sell any finite number of securities.

Consider the possible outcomes of the game (one and only one of which must occur):

  1. (A) Jay wins and the pot is at least USD 100.
  2. (B) Jay wins and the pot is less than USD 100.
  3. (C) Jay loses and the pot is at least USD 100.
  4. (D) Jay loses and the pot is less than USD 100.

The securities you are making a market on pay USD 1 if an event occurs, and USD 0 otherwise. So: if I buy 5 securities on outcome \(A\) from you, and outcome \(A\) occurs, I’ll be able to go to you and redeem my securities for a total of USD 5. Alternatively, if I sell you 5 securities on outcome \(A\), and outcome \(A\) occurs, you’ll be able to come to me and redeem your securities for a total of USD 5.

Consider what that implies: as a market maker, you face the prospect of making hefty payments to customers who redeem valuable securities. For example, imagine the situation where you charge USD 0.50 for a security on outcome \(A\), but outcome \(A\) is almost certain to occur in some sense (Jay is a beast when it comes to poker and a lot of high rollers are playing); if your customers exclusively load up on 100 cheap securities on outcome \(A\), and outcome \(A\) occurs, then you stand to owe them a total payment of USD 100 against the USD 50 that they have paid for the securities. You thus have a heavy incentive to price your securities as accurately as possible - where ‘accurate’ means to minimize your expected loss.

It may always be the case, however, that it is difficult to price your securities accurately. For example, if some customer has more information than you (say, she privately knows that Jay is unusually bad at poker) then she potentially stands to profit from holding said information in lieu of your ignorance on the matter (and that of your prices). Such is life for a market maker. But there are particular prices you could offer - independent of any participant’s private information - that are plainly stupid or ruinous for you (a set of prices like this is sometimes called a Dutch book). Consider selling securities on outcome \(A\) for the price of USD -1; then anyone who buys one of these securities not only stands to redeem USD 1 in the event outcome \(A\) occurs, but also gains USD 1 simply from the act of buying the security in the first place.

Setting a negative price like this is irrational on your part; customers will realize an arbitrage opportunity on securities for outcome \(A\) and will happily buy as many as they can get their hands on, to your ruin. In other words - and to nobody’s surprise - by setting a negative price, you can be made a sure loser in the market.

There are other prices you should avoid setting as well, if you want to avoid arbitrage opportunities like these. For starters:

  • For any outcome \(E\), you must set the price of a security on \(E\) to be at least USD 0.
  • For any certain outcome \(E\), you must set the price of a security on \(E\) to be exactly USD 1.

The first condition rules out negative prices, and the second ensures that your books balance when it comes time to settle payment for securities on a certain event.

What’s more, the price that you set on any given security doesn’t exist in isolation. Given the outcomes \(A\), \(B\), \(C\), and \(D\) listed previously, at least one must occur. So as per the second rule, the price of a synthetic derivative on the outcome “Jay wins or loses, and the pot is any value” must be set to USD 1. This places constraints on the prices that you can set for individual securities. It suffices that:

  • For any countable set of mutually exclusive outcomes \(E_{1}, E_{2}, \ldots\), you must set the price of the security on outcome “\(E_{1}\) or \(E_{2}\) or..” to exactly the sum of the prices of the individual outcomes.

This eliminates the possibility that your customers will make you a certain loser by buying elaborate combinations of securities on different outcomes.

There are other rules that your prices must obey as well, but they fall out as corollaries of these three. If you broke any of them you’d also be breaking one of these.

It turns out that you cannot be made a sure loser if, and only if, your prices obey these three rules. That is:

  • If your prices follow these rules, then you will offer customers no arbitrage opportunities.
  • Any market absent of arbitrage opportunities must have prices that conform to these rules.

These prices are called coherent, and absence of coherence implies the existence of arbitrage opportunities for your customers.

But Why Male Models

The trick, of course, is that these prices correspond to probabilities, and the rules for avoiding arbitrage correspond to the standard Kolmogorov axioms of probability theory.

The consequence is that if your description of uncertain phenomena does not involve probability theory, or does not behave exactly like probability theory, then it is an incoherent representation of information you have about those phenomena.

As a result, probability theory should be your tool of choice when it comes to describing uncertain phenomena. Granted you may not have to worry about market making in return for pizza, but you’d like to be assured that there are no structural problems with your description.


This is a summary of the development of probability presented in Jay Kadane’s brilliant Principles of Uncertainty. The original argument was developed by de Finetti and Savage in the mid-20th century.

Kadane’s book makes for an exceptional read, and it’s free to boot. I recommend checking it out if it has flown under your radar.

An interesting characteristic of this development of probability is that there is no way to guarantee the nonexistence of arbitrage opportunities for a countably infinite number of purchased securities. That is: if you’re a market maker, you could be made a sure loser in the market when it came time for you to settle a countably infinite number of redemption claims. The quirk here is that you could also be made a sure winner as well; whether you win or lose with certainty depends on the order in which the claims are settled! (Fortunately this doesn’t tend to be an issue in practice.)

Thanks to Fredrik Olsen for reviewing a draft of this post.