H. Paul Keeler

Creating a reversible Markov chain using acceptance(-rejection)

The study of Markov chains is generally the study of their long term behaviour, which, under certain conditions, is captured by them having a unique stationary distribution. Stationarity is an important property. It is, in a sense, a local property of a Markov chain.

For a Markov chain, a more global property is something called reversibility. Markov chains with this property must possess a stationary distribution, which, we see below, is an immediate consequence of reversibility. Reversible Markov chains (or processes) with discrete time¹ are the cornerstone of Markov chain Monte Carlo (MCMC) methods.

In this post we look at how a reversible Markov chain is constructed from a non-reversible (but irreducible) Markov chain by introducing an acceptance-rejection step. This post complements another post I wrote on the Metropolis-Hastings algorithm.

Reversibility

A Markov process on state space $\mathbb{X}$ with kernel $K$ is (time) reversible with respect to the distribution $\mu$ if the following holds

$$ \mu(x)K(x,y) = \mu (y) K(y,x)\quad x,y\in\mathbb{X}\,.$$

This reversibility condition is also called the detailed balance equation. If this condition is met, then the Markov process will have a stationary distribution $\mu$. By summing over $x$, we can verify this because we obtain

$$ \sum_{x\in\mathbb{X}}\mu(x)K(x,y) =\mu(y)\sum_{x\in\mathbb{X}} K(y,x)=\mu(y)\,.$$

This is just the balance equation, often written as $\mu=K\mu$, which says that the transition kernel $K$ has a stationary distribution $\mu $.

First Markov chain

We consider a Markov chain with kernel $J$ defined on a finite state space $\mathbb{X}$. If the Markov chain is at state $x\in \mathbb{X}$, it visits another state $y\in \mathbb{X}$ with the probability $J(x,y)$. This is a simple time-homogeneous finite Markov chain.

Irreducibility

For our Markov chain, we assume that every state $x$ in $\mathbb{X}$ where $\pi(x)>0$ is reachable with positive probability in a single step. This implies the easy-to-achieve condition $J(x,y)>0$ where $\pi(x)>0$ for all points $x,y \in \mathbb{X}$. This requirement is a stronger form of irreducibility.

Creating a new Markov chain with acceptance

We create a new Markov chain by introducing an acceptance step. For the Markov chain with kernel $J$, we assume that each time step, after choosing the jump direction but before jumping, a biased coin is flipped . The success probability $\alpha(x,y)$ depends on the current position $x\in \mathbb{X}$ and the (potential) next position $y\in \mathbb{X}$.

Transition kernel

For our new Markov chain, we can quickly reason the transition kernel $M$. We first look at the off-diagonal elements of the kernel (matrix) $M$. To go from state $x$ and to another state $y\neq x$, the probability is simply

$$ M(x,y) = \alpha(x,y) J(x,y), \quad x\neq y\,.$$

The transition matrix $M$ needs to be stochastic, so the rows sum to one, so $\sum_{y\in\mathbb{X}}M(x,y)=1$. That gives us the diagonal elements of $M$, although their exact form is not needed to show reversibility.

$\alpha(x,y)$ needs a symmetric function $s(x,y)$

For reversibility, we just need to swap rows and columns. Clearly we only need to look at the off-diagonal entries, which implies the requirement

$$ \pi(x)M(x,y) = \pi (y) M(y,x)\quad x\neq y\,.$$

Both sides are symmetric in $x$ and $y$, meaning they are equal to some non-negative symmetric $s(x,y)=s(y,x)$. Looking at the right-hand side, we get

$$\begin{aligned}\pi(y) M(y,x)&=\pi(y)J(y,x) \alpha(y,x)\\ &= s(y,x)\,.\end{aligned}$$

This implies that the function $\alpha$ is a non-negative function such that $\alpha\leq 1$, to ensure it’s a probability, with the form

$$ \alpha(x,y)=\frac{s(x,y)}{\pi(x)J(x,y)}\,.$$

The only task remaining now is to choose a reasonable symmetric function $s$ such that $\alpha\leq 1$, ensuring $\alpha$ is a probability. Of course, our choice for the symmetric function $s$ should also be a function of the stationary distribution $\pi$ and the underlying kernel $J$.

Examples

I’ll give two principal examples of the symmetric function $s(x,y)$. Working in reverse chronological order, I’ll give the simpler of the two examples first.

Barker

A somewhat natural example is

$$s(x,y) = \frac{\pi(x)J(x,y)\pi(y)J(y,x)}{\pi(x)J(x,y)+\pi(y)J(y,x)}\,.$$

This is clearly a symmetric function, which only has the terms $\pi$ and $J$. The acceptance probability becomes

$$\alpha(x,y) = \frac{\pi(y)J(y,x)}{\pi(x)J(x,y)+\pi(y)J(y,x)}\,.$$

A.A. Barker proposed this function in a 1965 paper as part of his PhD work in mathematical physics at the University of Adelaide. Barker had been inspired by a previous 1953 paper, which brings us to the next example.

Metropolis(-Rosenbluth-Rosenbluth-Teller-Teller)-Hastings

The now most important example is

$$s(x,y) = \min[\pi(x)J(x,y),\pi(y)J(y,x)]\,.$$

We can see that this is a symmetric function. The acceptance probability becomes

$$\alpha(x,y) = \min[1,\frac{\pi(y)J(y,x)}{\pi(x)J(x,y)}]\,.$$

This example is very famous in the world of Markov chain Monte Carlo methods. It is the main part of the so-called Metropolis-Hastings algorithm, which comes from a 1953 paper by Nicholas Metropolis, Arianna W. Rosenbluth, Marshall Rosenbluth, Augusta H. Teller, and Edward Teller (two husband-wife pairs), who looked at a special case, and a 1970 paper by W.K. Hastings, who generalized the method.

Markov goes to Monte Carlo

Numerical and scientific fields such as statistics, computational physics, and optimization methods heavily rely upon Markov chain Monte Carlo methods. These simulation techniques use the power of Markov chains to sample general probability distributions, which in turn give Monte Carlo methods for estimating integrals and optimal solutions.

This is the third part of a multi-part series of posts. The first part covers Markov chains. The second part covers the basics of Monte Carlo methods. This post combines the ideas from the first two parts. Overall, the posts sketch the mechanics of Markov chain Monte Carlo (MCMC) methods.

In a later post I detail the principal MCMC method known as the Metropolis(-Rosenbluth-Rosenbluth-Teller-Teller)-Hastings method, where you can see these mechanics come together. I’ve also given posts presenting code (in Julia, Python, MATLAB, R, and C) for this method here and here.

Here we will only examine Markov chains with countable state spaces. In this setting, we only need standard probability and matrix knowledge. But the results extend to general state spaces, such as Euclidean space.

By the way, I doubt that Andrey A. Markov ever went to Monte Carlo in the small country of Monaco.¹

Monte Carlo conditions

Sum

In the first post we covered Markov chains with countable state spaces, because the mathematics is more transparent. Consequently, we can use such Markov chains to estimate the sum of a function $f$ over a countable (possibly infinite) set $\mathbb{S}$. We can write the sum as

$$\begin{align}S(f)&=\sum_{x\in \mathbb{S}} f(x) \\&= \sum_{x\in\mathbb{S}} \frac{f(x)q(x)}{q(x)}\\& = \mathbb{E}_{q}[ \frac{f(Y)}{q(Y)}]\\& = \mathbb{E}_{q}[ g(Y)]\,,\end{align}$$

where $g=f/q$ and $Y$ is a suitable random variable that has a probability mass function $q$ with support $\mathbb{S}$.

Sampling the random variable (or point) $Y$, gives the samples $y_1,\dots,y_n$, which yields an unbiased Monte Carlo estimate of the sum, namely

$$S_n (f)=\frac{1}{n}\sum_{i=1}^n \frac{f(y_i)}{q(y_i)}.$$

The closer the function $q$ is to $|f|$, the better the estimate.

Integral

We can write an integral over some region $\mathbb{S}$ as

$$\begin{align}\int_{\mathbb{S}} f(x) dx &= \int_{\mathbb{S}} \frac{f(x)}{p(x)}p(x)dx\\ & =\mathbb{E}_p[\frac{f(Y)}{p(Y)}]\,,\end{align}$$

where $p$ is the probability density of a random variable (or point) $Y$ with support $A$. Then by sampling $Y$, the resulting samples $y_1, \dots, y_n\in [0,1]$ give the unbiased Monte Carlo estimate

$$I_n (f)=\frac{1}{n}\sum_{i=1}^n \frac{f(y_i)}{p(y_i)}\,.$$

Ergodic sequence

To estimate the sum or integral, we’ll need a sequence of ergodic random variables $\{Y_i\}_{i\in\mathbb{N}}$, which we can achieve with an ergodic Markov chain with a state space $\mathbb{S}$.

Markov chain conditions

In the context of Markov chains and, more generally, stochastic processes, we can think of ergodicity as convergence and a slightly more general version of the law of large numbers. A Markov chain with a countable state space needs some conditions to ensure ergodicity.

Regularity conditions for countable a Markov chain

A stationary distribution $\pi$

Aperiodicity

Irreducibility

Postive recurrence

In general, most of the needed structure for the Markov chain comes from irreducibility and a stationary distribution existing. Sometimes the above conditions are redundant. For example, an aperiodic, irreducible Markov chain with a finite state space is always positive recurrent. We look at ways on how to satisfy these conditions in the Monte Carlo setting when the underlying state space is countable.

Stationary distribution $\pi$

A stationary distribution $\pi$ satisfies equation

$$ \pi=P\pi.$$

We know exactly what our stationary distribution $\pi$ needs to be. It is the one used in the Monte Carlo method, namely the distribution of $Y$. The challenge is constructing a Markov chain with such a stationary distribution. This will of course depend on the Markov (transition) kernel $P$ of the Markov chain.

We’ll look at the other conditions on the Markov kernel $P$.

Aperiodicity

We recall that the period $d_x$ of a state $x\in \mathbb{X}$ is the greatest common divisor of all $n$ values such that $P(x,x)^n>0$. We need every state of the countable Markov chain to be aperiodic, meaning $d_x=1$ for all $x\in\mathbb{X}$.

One sufficient way to achieve this requirement is to first have the countable Markov chain to be irreducible, which we’ll cover in the next section. Then if there exists at least one state $x\in \mathbb{x}$ such that $P(x,x)>0$, then the Markov chain is aperiodic.²

Clearly the last condition is not difficult to satisfy, putting the focus now on the irreducibility requirement.

Irreducibility

The irreducibility property says that it is possible for the Marko chain to get from any point to another point in a finite number of steps. More precisely, for an irreducible Markov chain with countable state space, there must exist for all $x,y\in\mathbb{X}$ a natural number $s$ (possibly depending $x$ and $y$) such that $P(x,y)^s>0$.

We can achieve this requirement by introducing a stronger requirement, which is easier to verify. A countable Markov chain with a transition $P$ will be irreducible if for all $x,y\in\mathbb{X}$ the condition $P(x,y)>0$ holds. We have simply required that the above number $s=1$ for all $x,y\in\mathbb{X}$.

Furthermore, if $P(x,x)>0$ for all $x\in\mathbb{X}$, then our irreducible countable Markov chain is also aperiodic.

Positive recurrence

For a point or state $x\in\mathbb{X}$, we recall its first return time being defined as

$$ T_x^+=\min\{ t\geq 1: X_t=x\} \,.$$

A state $x$ is called positive recurrent if the expected value of its first return time is finite, meaning $\mathbb{E}_x(T_x^+)<\infty$. For a countable Markov chain, if all the states in the state space are positive recurrent, then we say the Markov chain is positive recurrent.

A countable irreducible Markov chain is positive recurrent if (and only if) it has a stationary distribution $\pi$. Furthermore, that stationary distribution will be unique and nonzero, meaning $\pi(x)>0$ for all $x\in\mathbb{X}$.³

For positive recurrence, we need to prove that the Markov chain has a stationary distribution.

Ergodicity

A countable Markov chain that is aperiodic and irreducible is ergodic. In other words, an ergodic Markov process $X$ with stationary distribution $\pi$ is a random sequence $X_0, X_1,\dots$ such that the following holds almost surely

$$ \frac{1}{t}\sum_{i=0}^{t-1} g(X_i) \rightarrow \sum_{x\in\mathbb{X}}g(x)\pi(x)$$

as $t\rightarrow\infty$ for all bounded, nonnegative functions $g$. Conveniently, wllle can use such an ergodic Markov chain to make statistical estimates.

Sampling with a Markov chain

For the Monte Carlo sample $y_1,\dots,y_n$, we simply use the Markov chain samples $x_0,x_1,\dots,x_{n-1}$, giving the Monte Carlo estimate. In the discrete case, this is simply

$$S_n (f)=\frac{1}{n}\sum_{i=1}^n g(x_{i-1}).$$

which, due to ergodicity, converges to $S(f)=\mathbb{E}_{Y}[g(Y)]$.

Summary

In summary, a sufficient way to have a countable ergodic Markov chain is for the transition kernel to be such that $P(x,y)>0$ for all $x\in\mathbb{X}$ and there exists a (stationary) distribution $\pi=P\pi$. For the first requirement, one imagines using a non-negative function such as $e^{-(x-y)^2}/C$, where $C>0$ is a suitable normalization constant.

The requirements seem possible, but to bring them all together is still a great challenge for the uninitiated. Fortunately, some great minds almost seven decades ago proposed the first Markov chain that meets all the requirements for a specific stationary distribution, which was used for a Monte Carlo estimate.

The first such Markov chain Monte Carlo method appeared in a paper written by five scientists, including two husband-wife pairs. It became known as the Metropolis or Metropolis-Hastings algorithm, and it will be the subject of a future post.

The second MC in MCMC methods

In calculus differentiating a function is easy. Integrating one is hard. Much research has gone into finding ways for calculating integrals. But often functions cannot be integrated, at least analytically.

No problem. You do it numerically. You take the function’s curve, split it up into nice rectangles or, even better, trapezoids, and calculate their areas, and add them up. Not accurate enough? Split the curve up even finer until you’re happy. Done.

That works fine in one dimension and even two dimensions. But then, as the number of dimensions increases, the amount of splitting and adding grows exponentially, soon hitting unwieldy numbers. This is a mountain too steep for even the fastest computers. The problem is what applied mathematics guru Richard E. Bellman famously called the curse of dimensionality.

But thankfully randomness offers a path around this mountain. The approach is called Monte Carlo integration for estimating sums and integrals

This is the second part of a series of posts on Mark chain Monte Carlo methods. The first part covers Markov chains. This posts covers the basics of Monte Carlo methods. The third part will combine the ideas from the first two parts. Overall, the three posts will sketch the mechanics of Markov chain Monte Carlo (MCMC) methods.

We’ll focus on estimating sums, but the same logic applies to integrals.

Monte Carlo estimates

In the discrete (or countable) case, consider a function $f$ that maps from a countable state space $\mathbb{S}$. This state space maybe, for example, the lattice $\mathbb{Z}^d$ or some subset of it. We write the sum as

$$S(f)=\sum_{x\in \mathbb{S}} f(x) $$

Such sums arise in many fields, particularly statistical physics, which is where the (Markov chain) Monte Carlo methods were born.¹

The idea behind a Monte Carlo estimate of a sum is remarkably simple. You randomly sample the function $f$, find the average, and use it as an estimate of the sum.

Estimating with uniform variables

For a finite set $\mathbb{S}=\{x_1,\dots,x_m\}$, we write the sum of $f$ over the set $\mathbb{S}$ as

$$\begin{align}S(f)&=\sum_{x\in \mathbb{S}} f(x) \frac{m}{m} \\&= \sum_{x\in\mathbb{S}} \frac{f(x)p(x)}{p(x)}\\& = \mathbb{E}_{p}[ \frac{f(U)}{p(U)}]\,,\end{align}$$

where $U$ is a discrete uniform random variables, which has the probability mass function $p$ with support $\mathbb{S}$, meaning $p(x)=1/m$ if $x\in \mathbb{S}$ and $p(x)=0$ otherwise. But using a uniform random variable is a naive way, as there are better random variables we can use to sample $f$.

Estimating with general variables

More generally, we can write the sum of a countable (possibly infinite) set $\mathbb{S}$ as

$$\begin{align}S(f)&= \sum_{x\in\mathbb{S}} \frac{f(x)q(x)}{q(x)}\\& = \mathbb{E}_{q}[ \frac{f(Y)}{q(Y)}]\\& = \mathbb{E}_{q}[ g(Y)]\,,\end{align}$$

where $g=p/q$ and $Y$ is a suitable random variable that has a probability mass function $q$ with support $\mathbb{S}$.

Sampling the random variable (or point) $Y$, gives the samples $y_1,\dots,y_n$, which yields an unbiased Monte Carlo estimate of the sum

$$S_n (f)=\frac{1}{n}\sum_{i=1}^n \frac{f(y_i)}{q(y_i)}.$$

If we can (quickly) sample $Y$, then we can (quickly) estimate the integral. The closer the (probability mass) function $q$ is to $|f|$, the better the estimate.

We can also use the above reasoning with integrals, replacing the sum and probability mass function with an integral and a probability density.

Error rate

If the samples $y_1, \dots, y_n $ are random and independent, the (strong) law of large numbers guarantees that the unbiased Monte Carlo estimate converges (almost surely) to the correct answer.

Further assuming $S(f^2)<\infty$, the error of the integral estimate $|S(f)-S_n(f)|$ is proportional to

$$\frac{\mathbb{V}_q(f)}{n ^{1/2}},
$$
where
$$\mathbb{V}(f)_q=E_q[f^2(Y)]-(E_q[f(Y)])^2,$$
is the variance of the function $f$. (This is not the variation of a function, which is something different.) The error $|S(f)-S_n(f)|$ reduces at the rate $O(n^{-1/2})$. The Monte Carlo approach allows for continual updating until the estimate is sufficiently converged.

The above logic applies to integrals of functions on general spaces.

The dimension $d$ of the underlying space, as $\mathbb{S}^d$, does not appear in the error. The independence of the dimension is precisely why we tackle high-dimensional sums and integrals with Monte Carlo methods. The general rule of thumb is to use Monte Carlo integration for functions of four or more dimensions.

Losing the independence

Again, a sequence of independent random variables allow our Monte Carlo estimate to converge. We can loosen this requirement and allow some dependence between variables. To obtain a Monte Carlo estimate, we just need to use a sequence of random variables that are ergodic. Then we recover the (strong) law of large numbers that gives our Monte Carlo estimate.

Ergodicity

The word ergodicity is one of those technical terms you’ll see and hear in different fields of mathematics and physics, often when randomness is involved. The concept is also central to deterministic dynamical systems.

There are different versions and strengths of ergodicity, depending on the system and applications. But loosely speaking, an ergodic system is one in which temporal averages and spatial (or ensemble) averages converge to the same value.

That means when you want to find an average, you have two options in an ergodic system. The temporal option: You can measure something over a long time. The spatial (or ensemble) option: You can measure many instances of the same something. The two different averages will eventually converge to the same value.

Ergodic sequences

In the context of Markov processes and, more generally, stochastic processes, we can think of ergodicity as convergence and a slightly more general version of the law of large numbers.

Consider a stochastic process $X$, which is a sequence of random variables $X_0,X_1,\dots$ indexed by some parameter $t$. We assume the probability distributions $\pi_0,\pi_1,\dots$ of the random variables converge to some final stationary distribution $\pi$. For a countable state space, the (probability) distribution $\pi$ corresponds to a probability mass function.

We then say a stochastic process $X$ state space is ergodic if the following holds almost surely

$$ \frac{1}{t}\sum_{i=0}^t g(X_i)\rightarrow \sum_{x\in \mathbb{Z}}g(x)\pi(x)=\mathbb{E}_{\pi} [g(X)]\,,$$

as $t\rightarrow\infty$ for all bounded functions, nonnegative functions $g$. The subscript of the expectation simply denotes that the expectation is taken with respect to the (stationary) distribution $\pi$.

The equation tells the story. On the left-hand side you have a temporal average. On the right-hand side you have a spatial (or ensemble) average. As time proceeds, the first average will converge to the second average. Ergodicity.

The first MC in MCMC methods

Markov chains form a fundamentally important class of stochastic processes. It would be hard to over stress their importance in probability, statistics, and, more broadly, science and technology. They are indispensable in random simulations, particularly those based on the Markov chain Monte Carlo methods. In this post, we’ll have a look some Markov chain basics needed for such simulation methods.

This is the first part of a series of posts on Mark chain Monte Carlo methods. This post covers the basics of Markov chains, which is the more involved part. The second part will cover Monte Carlo methods. The third part will combine the ideas from the first two parts. Overall, the three posts will sketch the mechanics of Markov chain Monte Carlo (MCMC) methods.

Markov chains vs Markov processes

All Markov chains are Markov processes. Some people use the term Markov chain to refer to discrete-time Markov processes with general state spaces. Other people prefer the term Markov chain for continuous-time Markov processes with countable state spaces.¹

Nevertheless, the first MC in the MCMC suggests the Markov chain Monte Carlo crowd prefers the former sense of Markov chain, given the use of discrete-time Markov processes in their simulations.

Markov the frog

Some writers introduce Markov chains with a mental image of a frog jumping around lily pads scattered over a pond. (Presumably the frog never misses a lily pad.) We assume the frog randomly chooses the next lily pad through some random mechanism. Perhaps the distances between lily pads or their sizes influence the chances that frog will jump between them.

We further assume that the frog is a bit particular, preferring to jump in certain directions more than others. More precisely, the probability of our frog jumping from a lily pad labelled $x$ to another labelled $y$ is $P(x,y)$. But jumping in the opposite direction happens with probability $P(y,x)$, which in general is not equal to $P(x,y)$.

I typically use the term points, but the Markov literature usually says that the Markov chain visits states.

State space

We can interpret a Markov chain, a type of stochastic process, as a collection or sequence of random variables. ²The values of the random variables are points in some mathematical space $\mathbb{X}$. This space can be quite abstract, but in practice it’s usually the lattice $\mathbb{Z}^n$, Euclidean space $\mathbb{R}^n$, or a subset of one of these two spaces. For our frog example, all the lily pads in the pond form the state space.

We’ll only consider countable Markov chains where the number of points in the state space $\mathbb{X}$ is countable. Although the results and theory generally hold for more general state spaces, the accompanying work requires more technical mathematics. For finite and countable state spaces, we can use standard probability and matrix knowledge. But when we use uncountable state spaces such as $\mathbb{R}^n$, we enter the world of measure theory and functional analysis.

I will often write a point $x$ in a (state) space $\mathbb{X}$. But you can say an element $x$ of a set $\mathbb{X}$. Many authors refers to the points or elements as states of the Markov chain. In the frog example, each lily pad is a different state.

Markov property

A discrete-time countable Markov chain is a random process that jumps between points of some countable mathematical space $\mathbb{X}$ such that, when at point $x \in \mathbb{X}$, the next position is chosen according to a probability distribution $P(x,·)$ depending only on $x$.

More specifically, a sequence of random variables $(X_0, X_1, . . .)$ is a discrete-time Markov chain $X$ with a countable state space
$\mathbb{X}$ and kernel $P$ if for all $x,y \in \mathbb{X}$ and all $t \geq 1$ satisfying $\mathbb{P}[X_{t−1}=x_{t-1},\dots,X_0=x_0]>0$, we have

$$ \begin{align}\mathbb{P}[X_{t+1} =y|&X_{t}=x,X_{t−1}=x_{t-1},\dots,X_0=x_0]\\&=\mathbb{P}[X_{t+1} =y|X_t =x]\\&=P(x,y)\,.\end{align}$$

This equation is often called the Markov property.

The Markov property says that the conditional probability of jumping from point $x$ to $y$ remains the same, regardless of which points or states $x_0,x_1,\dots,x_{t-1}$ were previously visited. This is precisely why the kernel $P$ contains all the information needed to describe the future random evolution of the Markov chain.

We have assumed the probabilities given by $P$ are fixed, meaning we have described a homogeneous Markov chain.

Markov kernel

The kernel $P$ is called the Markov (transition) kernel or probability kernel. Assuming a countable state space $\mathbb{X}$, we can reference any probability value of the kernel $P$ with two variables $x,y\in\mathbb{X}$. If we assume a finite state space $\mathbb{X}$, then the kernel $P$ becomes a regular matrix taught in linear algebra. An infinite but countable state space gives an infinite matrix $P$. The rows of the kernel matrix $P$ must add up to one, because each row is a probability measure.

A more general space, such as Euclidean space $\mathbb{R}^n$, results in a more general kernel with respect to a suitable measure. In this setting, $P(x,·)$ is no longer a probability mass function, but a general probability measure.

Initial distribution

At time $t=0$ we describe the random initial configuration of a Markov process with a probability distribution $\mu_0$. For a finite or countable Markov chain, this initial distribution $\mu_0$ corresponds to a probability mass function encoded as a row vector.

Jumping from $x$ to $y$

The probability distribution $\mu_0$ gives the probability of starting in state (or at point) $x\in\mathbb{X}$. After one time step, we can write down the probability distribution $\mu_1$ that gives us the different probabilities of the Markov chain being at different states. At $n=1$, basic matrix algebra and probability rules give us the matrix equation

$$\mu_1=\mu_0 P$$

By induction, after $t$ time steps we have the expression

$$\mu_n=\mu_0 P^n\,.$$

where the superscript $n$ denotes matrix power. We can write the $n$-time step kernel as $P_{(n)}$, which for a finite Markov chain is given by the matrix equation $P_{(n)}=P^n$.

Seeing how $P_{(n)}$ behaves as $n$ approaches infinity forms part of work that studies the convergence and ergodicity properties of Markov chains. I’ll make these concepts clearer below. But first I’ll give some conditions that are typically needed.

Regularity conditions

A Markov chain with a countable state space needs some conditions to ensure convergence and ergodicity.

Regularity conditions

A stationary distribution $\pi$

Aperiodicity

Irreducibility

Postive recurrence

The nature of the state space and the kernel will dictate these conditions. These conditions are also not necessarily logically distinct. For example, on a finite state space, you’ll get positive recurrence for free, because an aperiodic, irreducible Markov chain with a finite state space is always positive recurrent.

We now briefly detail these conditions and in another post I’ll give examples how the conditions can be met.

Stationary distribution $\pi$

It’s possible to encounter a probability distribution $\pi$ where applying the kernel $P$ returns the same distribution $\pi$, meaning

$$ \pi=\pi P\,.$$

This (fixed-point) equation is called the balance equation.

The distribution $\pi$ is called the stationary, invariant or steady-state distribution. A Markov chain does not need to have a stationary distribution. And if a Markov chain does have one, it may not be unique. Its existence and uniqueness will depend on the Markov kernel $P$.

Showing that a unique stationary distribution exists and it is possible to reach it with probability one is the stuff of Markov convergence results. Markov chain Monte Carlo methods hinge upon these results .

Aperiodicity

It is possible for a Markov chain to get trapped in a loop, periodically visiting the same states. The period $d_x$ of a state $x\in \mathbb{x}$ is the greatest common divisor of all $n$ values such that $P(x,x)^n>0$. If the period of a point is $d_x=1$, then we say it’s aperiodic. If every state of a Markov chain is aperiodic, we says it’s an aperiodic Markov chain.

Aperiodicity means there are no loops to trap the Markov chain. This property is typically needed for convergence results.

Irreducibility

A Markov chain with a countable state space $\mathbb{X}$ is irreducible if the Markov chain can go from any point $x\in\mathbb{X}$ to another other point $x\in\mathbb{X}$ with a positive probability in a finite number of time steps. In other words, there exists a natural number $s$ such that $P(x,y)^s>0$ for all $x,y\in\mathbb{X}$.

Irreducibility ensures that a Markov chain will visit all the states in its state space. This property is also needed for convergence results.

Recurrence

When studying Markov processes, a quantity of interest is how much time it takes to return to a state or point. For a point $x\in\mathbb{X}$, we define its first return time as

$$ T_x^+=\min\{ t\geq 1: X_t=x\} \,.$$

As the name suggests, this random variable is the number of time steps for the Markov process return to state $x$, taking whichever path, conditioned on it starting at $x$.

We call a state $x$ recurrent if the probability of its first return time being finite is one, meaning $\mathbb{P}_x(T_x^+<\infty)=1$. Otherwise the state $x$ is said to be transient.

Positive recurrence

We can classify different types of recurrence based on the expected value of the first return times. A state $x$ is called positive recurrent if the expected value of its first return time is finite, meaning $\mathbb{E}_x(T_x^+)<\infty$. Otherwise state $x$ is null recurrent.

For a countable Markov chain, if all the states in the state space are (positive) recurrent, so $\mathbb{E}_x(T_x^+)<\infty$ for all $x\in\mathbb{X}$, then we say the Markov chain is (positive) recurrent.

Again, the concept of positive recurrence is needed for convergence results.

Ergodicity

We say a countable Markov chain is ergodic if it is irreducible, aperiodic and positive recurrent.³ Ergodicity allows one to find averages by employing a more general form of the law of large numbers, which Monte Carlo methods rely upon. We stress that definitions of ergodicity vary somewhat, but in general it means convergence and laws of large numbers exists.

The acceptance(-rejection) method for simulating random variables

In a previous post, I covered a simple but much used method for simulating random variables or, rather, generating random variates. To simulate a random variable, the method requires, in an easy fashion, calculating the inverse of its cumulative distribution function. But you cannot always do that.

In lieu of this, the great John von Neumann wrote in a 1951 paper that you can sample a sequence of values from another probability distribution, accepting only the values that meet a certain condition based on this other distribution and the desired distribution, while rejecting all the others. The accepted values will follow the desired probability distribution. This method of simulation or sampling is called the rejection method, the acceptance method, and it has even the double-barrelled name the acceptance-rejection (AR) method.

Details

Let $X$ be a continuous random variable with a (probability) density $p(x)$, which is the derivative of its cumulative probability distribution $P(X\leq x)$. The density $p(x)$ corresponds to the desired or target distribution from which we want to sample. For whatever reason, we cannot directly simulate the random variable $X$. (Maybe we cannot use the inverse method because $P(X\leq x)$ is too complicated.)

The idea that von Newman had was to assume that we can easily simulate another random variable, say, $Y$ with the (probability) density $q(x)$. The density $q(x)$ corresponds to a proposal distribution that we can sample (by using, for example, the inverse method).

Now we further assume that there exists some finite constant $M>0$ such that we can bound $p(x)$ by $Mq(x)$, meaning

$$ p(x) \leq M q(x), \text{ for all } x . $$

Provided this, we can then sample the random variable $Y$ and accept a value of it (for a value of $X$) with probability

$$\alpha = \frac{p(Y)}{Mq(Y)}.$$

If the sampled value of $Y$ is not accepted (which happens with probability $1-\alpha$), then we must repeat this random experiment until a sampled value of $Y$ is accepted.

Algorithm

We give the pseudo-code for the acceptance-rejection method suggested by von Neumann.

Random variable $X$ with density $p(x)$

Sample a random variable $Y$ with density $q(x)$, giving a sample value $y$.

Calculate the acceptance probability $\alpha = \frac{p(y)}{Mq(y)}$.

Sample a uniform random variable $U\sim U(0,1)$, giving a sample value $u$.

Return the value $y$ (for the value of $X$) if $u\leq \alpha$, otherwise go to Step 1 and repeat.

As covered in a previous post, Steps 3 and 4 are equivalent to accepting the value $y$ with probability $\alpha$.

Point process application

In the context of point processes, this method is akin to thinning point processes independently. This gives a method for positioning points non-uniformly by first placing the points uniformly. The method then thins points based on the desired intensity function. As I covered in a previous post, this is one way to simulate an inhomogeneous (or nonhomogeneous) Poisson point process.

Efficiency

Basic probability theory tells us that the number of experiment runs (Steps 1 to 3) until acceptance is a geometric variable with parameter $\alpha$. On average the acceptance(-rejection) method will take $1/\alpha$ number of simulations to sample one value of the random $X$ of the target distribution. The key then is to make the proposal density $q(x)$ as small as possible (and adjust $M$ accordingly), while still keeping the inequality $p(x) \leq M q(x)$.

Higher dimensions

The difficulty of the acceptance(-rejection) method is finding a good proposal distribution such that the product $Mq(x)$ is not much larger than the target density $p(x)$. In one-dimension, this can be often done, but in higher dimensions this becomes increasingly difficult. Consequently, this method is typically not used in higher dimensions.

Another approach with an acceptance step is the Metropolis-Hastings method, which is the quintessential Markov chain Monte Carlo (MCMC) method. This method and its cousins have become exceedingly popular, as they give ways to simulate collections of dependent random variables that have complicated (joint) distributions.

The Box-Muller method for simulating normal variables

In the previous post, I covered a simple but much used method for simulating random variables or, rather, generating random variates. To simulate a random variable, the method requires writing down, in a tractable manner, the inverse of its cumulative distribution function.

But in the case of the normal (or Gaussian) distribution, there is no closed-form expression for its cumulative distribution function nor its inverse. This means you cannot, in an elegant and fast way at least, generate with the inverse method a single normal random variable using a single uniform random variable.

Interestingly, however, you can generate two (independent) normal variables with two (independent) uniform variables using the Box-Muller method, originally proposed by George Box and Mervin E. Muller. This approach uses the inverse method, but in practice it’s not used much (see below). I detail this method because I find it neat and it highlights the connection between the normal distribution and rotational symmetry, which has been the subject of some recent 3Blue1Brown videos on YouTube.

(This method was also used to simulate the Thomas point process, which I covered in a previous post.)

Incidentally, this connection is also mentioned in a previous post on simulating a Poisson point process on the surface of a sphere. In that method post, Method 2 uses an observation by the Muller that normal random variables can be used to position points uniformly on spheres.

I imagine this method was first observed by transforming two normal variables, instead of guessing various distribution pairs that would work. Then I’ll sketch the proof in the opposite direction, though it works in both directions.

Proof outline

The joint probability density of two independent variables is simply the product of the two individual probabilities densities. Then the joint density of two standard normal variables is

$$\begin{align}f_{X,Y}(x,y)&=\left[\frac{1}{\sqrt{2\pi}}e^{-x^2/2}\right]\left[\frac{1}{\sqrt{2\pi}}e^{-y^2/2}\right]\\&=\frac{1}{{2\pi}}e^{-(x^2+y^2)/2}\,.\end{align}$$

Now it requires a change of coordinates in two dimensions (from Cartesian to polar) using a Jacobian determinant, which in this case is $|J(\theta,r)=r|$.¹ giving a new joint probability density

$$f_{\Theta,R}(\theta,r)=\left[\frac{1}{\sqrt{2\pi}}\right]\left[ r\,e^{-r^2/2}\right]\,.$$

Now we just identify the two probability densities. The first probability density corresponds to a uniform variable on $[0, 2\pi]$, whereas the second is that of a Rayleigh variable with parameter $\sigma=1$. Of course the proof works in the opposite direction because the transformation (between Cartesian and polar coordinates) is a one-to-one function.

Algorithm

Here’s the Box-Muller method for simulating two (independent) standard normal variables with two (independent) uniform random variables.

Two (independent) standard normal random variable $Z_1$ and $Z_2$

Generate two (independent) uniform random variables $U_1\sim U(0,1)$ and $U_2\sim U(0,1)$.

Return $Z_1=\sqrt{-2\ln U_1}\cos(2\pi U_2)$ and $Z_2=\sqrt{-2\ln U_1}\sin(2\pi U_2)$.

The method effectively samples a uniform angular variable $\Theta=2\pi U_2$ on the interval $[0,2\pi]$ and a radial variable $R=\sqrt{-2\ln U_1}$ with a Rayleigh distribution.

The algorithm produces two independent standard normal variables. Of course, as many of us learn in high school, if $Z$ is a standard normal variable, then the random variable $X=\sigma Z +\mu$ is a normal variable with mean $\mu$ and standard deviation $\sigma>0$ .

The fall of the Box-Muller method

Sadly this method was typically not used, as historically computer processors were slow at doing calculations involving the necessary mathematical functions. To avoid these functions researchers developed and employed other methods such as the ziggurat algorithm.

Also, although processors can now do such calculations much faster, many languages, not just scientific ones, come with functions for generating normal variables. Consequently, there had not been much need in implementing this method.

Update: The return of the Box-Muller method

The above conventional wisdom has changed in recent years as processors can now (on a hardware level) readily evaluate such functions. (I had been waiting to see if certain libraries would be re-written by using the Box-Muller method, but why bother if the old ones work so well?) When I used the term “processors”, I had central processor units (CPUs) in mind, but in recent years graphically processor units (GPUs) have become widely popular.

In a comment on this post, it pointed out that the Box-Muller method is the preferred choice for GPUs, as evidenced by its implementation in Nvidia’s CUDA library. The reason is GPUs do not handle well loops and branches in algorithms, so you should use methods that avoid these algorithmic steps. And the Box-Muller method is one that does just that.

The NVDIA website says:

Because GPUs are so sensitive to looping and branching, it turns out that the best choice for the Gaussian transform is actually the venerable Box-Muller transform

The inverse method for simulating random variables

We will cover a simple but much used method for simulating random variables or, rather, random variates. Although the material here is found in introductory probability courses, it frequently forms the foundation of more advance stochastic simulation techniques, such as Markov chain Monte Carlo methods.

Details

The basics of probability theory tell us that any random variable can, in theory, be written as a function of a uniform random variable $U$ distributed on the interval $(0,1)$, which is usually written as $U\sim U(0,1)$. All one needs is the inverse of the cumulative distribution function of the desired random variable.

More specifically, let $X$ be a random variable with a cumulative distribution function $F(x)=\mathbb{P}(X\leq x)$. The function $F$ is nondecreasing in $x$, so its inverse can be defined as $F^{-1}(y)=\inf\{x:F(x)\geq y\}$, which is known as the generalized inverse of $F(x)$.

Some authors assume the minimum is attained so the infimum is replaced with the minimum, giving $F^{-1}(y)=\min\{x:F(x)\geq y\}$.

In short, the following result is all that we need.

Transform of a uniform variable $U\sim U(0,1)$

For a uniform random variable $U\sim U(0,1)$, the random variable $F^{-1}(U)$ has the cumulative distribution function $\mathbb{P}(F^{-1}(U)\leq x)=P(U\leq F(x))=F(x)$.

Algorithm

The above observation gives a method, which I like to call the direct method, for exactly simulating a random variable $X$ with the (cumulative) distribution (function) $F$.

Random variable $X$ with distribution $F$

Sample a uniform random variable $U\sim U(0,1)$, giving a value $u$.

Return the value $x=F^{-1}(u)$ as the sampled value of $U$.

But this approach only works if we can write down (in a relatively straightforward way) the inverse $F^{-1}$, which is usually not the case. This means you cannot generate, for example, simulate a normal random variable with a single uniform random variable by using just the inverse method, as we cannot write down the inverse of its cumulative distribution function.

(Interestingly, with two (independent) uniform random variables, we can use the transform method to simulate two (independent) normal (or Gaussian) random variables. This approach is called the Box-Muller transform, which I’ll cover in another post.)

Nevertheless, we can apply the inverse method to some useful distributions.

Examples

Warning: The following examples are only for illustration purposes. Except for the Bernoulli example, you would never use them in standard scientific languages such as MATLAB, Python (with NumPy), R or Julia, because those languages already have much better functions for simulating these and many other random variables (or variates). If you are writing a function in a language that lacks such functions, I would consult one of the references mentioned below. Although the inverse method is usually intuitive and elegant, it is often not the fastest method.

Bernoulli distribution

The simplest random variable is that with the Bernoulli distribution. With probability $p$, a Bernoulli random variable $X$ takes the value one. Otherwise, $X$ takes the value zero (with probability $1-p$). This gives the (cumulative) distribution (function):

$$ F_B(x)=\begin{cases}
0 & \text{if } x < 0 \\
1 – p & \text{if } 0 \leq x < 1 \\
1 & \text{if } x \geq 1
\end{cases}$$

This gives a very simple way to simulate (or sample) a Bernoulli variable $X$ with parameter $p$.

Bernoulli random variable $X$ with parameter $p$

Sample a uniform random variable $U\sim U(0,1)$, giving a value $u$.

If $u\leq p$, return $x=1$; otherwise return $x=0$.

Application: Acceptance simulation methods

In random simulation code, whenever you do something (or not) with some probability $p$ (or probability $1-p$), then the code will perform the above step. Consequently, you see this in the (pseudo-)code of many stochastic simulations with random binary choices, particularly schemes that have an acceptance step such the Metropolis-Hastings method and other Markov chain Monte Carlo (MCMC) methods.

In MCMC schemes, a random (binary) choice is proposed and it is accepted with a certain probability, say, $\alpha $. This is the equivalent of accepting the proposed choice if some uniform random variable $U$ meets the condition $U\leq \alpha$.

This explains why the pseudo-code of the same algorithm can vary. Some pseudo-code will say accept with probability $\alpha$, while other pseudo-code will say do if $U\leq \alpha$. It’s two equivalent formulations.

Exponential distribution

The cumulative distribution function of an exponential variable with mean $1/\lambda$ is $F_E(x)= 1-e^{-\lambda x}$, which has the inverse $F^{-1}_E(y)=-(1/\mu)\ln[1-y]$. We can use the fact that on the interval $(0,1)$, a uniform variable $U\sim U(0,1)$ and $1-U$ have the same distribution. Consequently, the random variables $\ln [1-U]$ and $\ln U$ are equal in distribution.

This gives a method for simulating exponential random variables.

Exponential random variable $X$ with mean $1/\lambda$

Sample a uniform random variable $U\sim U(0,1)$, giving a value $u$.

Return $x=-(1/\lambda)\ln u$.

Application: Poisson simulation method

Of course you can use this method to simulate exponential random variables, but it has another application. In a previous post on simulating Poisson variables, I mentioned that exponential random variables can be used to simulate a Poisson random variable in a direct (or exact) manner. That method is based on the distances between the points of a homogeneous Poisson point process (on the real line) being exponential random variables.

But this method is only suitable for low values of $\lambda$, less than say fifteen.

Rayleigh distribution

The Rayleigh distribution is $\mathbb{P}(X\leq x)= (x/\sigma^2)e^{-x^2/(2\sigma^2)}$, where $\sigma>0$ is its scale parameter. The square root of an exponential variable with mean $1/\lambda$ has a Rayleigh distribution with scale parameter $\sigma=1/\sqrt{2\lambda}$.

Consequently, the generation method is similar to the previous example.

Rayleigh random variable $Y$ with scale parameter $\sigma>0$

Sample a uniform random variable $U\sim U(0,1)$, giving a value $u$.

Return $y=\sigma\sqrt{-2\ln u}$.

Other methods

The inverse method is intuitive and often succinct. But most functions for simulating random variables (or, more correctly, generating random variates) do not use these methods, as they are not fast under certain parameter regimes, such as large means. Consequently, other method are used such as approximations (with, say, normal random variables), such as the ones I outlined in this post on simulating Poisson random variables.

More complicated random systems, such as collections of dependent variables, can be simulated using Markov chain Monte Carlo methods, which is the direction we’ll take in a couple posts after this one.

New link – math3ma.com

In my internet wanderings, I stumbled upon this blog:

https://almostsuremath.com/

The writer, Tae-Danae Bradley, wrote a PhD applying category theory (a field of mathematics that strives for abstraction) to problems in machine learning.

The Newman-Ziff algorithm for simulating percolation models

Imagine you’re trying to estimate some statistics about a certain percolation system. For example, estimating the probability of a given bond being open, which means connected, when a giant component forms. The system or model is too complicated for analytic results, which is true for all percolation models with the exception of a handful.

So you code up the percolation model in your favourite programming language, run a large number of stochastic simulations, collect the statistics, and then look at the results. Percolation systems are, by definition, large, but with fast computers, the simulation method works. You get your statistics.

Easy.

But that’s just for one parameter. What if you want to see what happens with a range of parameter values? Well, you do the same thing, but now just run the large number of simulations for a range of different parameter values, right?

No.

You could do that, and it would work. But it would be slow. How to make it faster?

Newman and Ziff proposed and implemented a fast percolation algorithm based on the simple ideas of keeping old simulations and using conditional probabilities.

Algorithm

The algorithm was presented in this paper:

2000 – Newman, Ziff – Efficient Monte Carlo Algorithm and High-Precision Results for Percolation

There exists a preprint with a slightly different title.

2001 – Newman, Ziff – A fast Monte Carlo algorithm for site or bond percolation.

Overview

The simulation method consists of three components.

Components of the Newman-Ziff algorithm

Re-use parts of previous simulation outcomes by shuffling.

Use fast union-find algorithms.

Find a smooth curve of results by using conditional probabilities.

None of them are entirely revolutionary, but put together they form a fast (and popular) simulation method for studying (monotonic) percolation models.

1. Randomly sampling sequences

The main insight was to sample in an incremental fashion, adding an open bond to the percolation model, such as a lattice, during each simulation run, without throwing away the previous simulation. More specifically, let’s say they sample a bond model with $k$ open bonds during one simulation run. Then in the next simulation, use the same configuration, and simply add one bond.

This is a nice trick, but there’s no free lunch here. On one hand, they don’t waste results. But on the other hand, they do throw away independence (and ergodicity) by doing this. But perhaps the numerical results don’t care.

To do this step, Newman and Ziff use a random shuffling algorithm called the Fisher-Yates shuffle. But it’s not clear if Newmand and Ziff thought of this shuffling algorithm independently or not. They never call it by its name nor do they cite any work on this well-known shuffling method. Both MATLAB and Python (SciPy) have functions for doing random shuffles.

2. Using union-find

Imagine you have a collection of things. Some of these things are connected to other things in your collection. You want to understand how these clusters of connected things behave.

To find the clusters, looking for ultimately the biggest one, Newman and Ziff use a union-find method. There are different types of these alogrithms, often hinging upon the use of recursion. They were developed mostly in the 1980s, particularly in work by Robert Tarjan.

These methods are very efficient at finding clusters. The algorithm speed or, rather, complexity is given in relation to the inverse of an Ackermann function, which is a famously fast growing function. These methods rely upon the monotonicity of growing unions.

(If you have a non-monotonic percolating system, then unfortunately you can’t use these algorithms, which is the case for percolation based on signal-to-interference-plus-noise ratio (SINR). )

3. Filling in the parameter gaps

For any random sample of a percolation model, the number of open bonds or sites is a natural number. But the parameter of the percolation model, such as the probability of a site or bond being open, is a real number.

Newman and Ziff use a simple conditioning argument to fill the gaps.

For any statistic $S$ of of a given percolation model, the Newman-Ziff algorithm finds the expected statistic conditioned on $n$ number of open bonds (or equivalent objects in other models). In other words, the algorithm finds the expected conditional statistics $E[S_n|N=n]$. Then we just need the probability of $N$ being $n$ as a function of the parameter we’re trying to vary, such as the bond probability $p$. With this probability $P(N=n)$, we immediately arrive at the expression for the statistic $S$ with introductory probability

$$ E(S_n)=E_N [E(S_n|N=n)].$$

For a nice discrete model with $m$ total, we get the expression

$$ E(S_n)=\sum _{n=1}^{m}[P(N=n)S_n].$$

But we know the probability (distribution) of $N$, as it’s simply the binomial variable. Assuming there are $m$ total bonds (or equivalent objects), then we arrive at

$$P(N=n)={m\choose n} p^n(1-p)^{m-n} .$$

Then the final statistic or result is simply

$$E(S_n)=\sum _{n=1}^{m} {m\choose n} p^n(1-p)^{m-n} S_n.$$

Code

I wrote the algorithm in MATLAB and Python. The code can be found here. I’ve included a script for plotting the results (for small square lattices).

MATLAB

It turns out that MATLAB doesn’t like recursion too much. Well, at least that’s my explanation for the slow results when I use the union-find method with recursion. So I used the union-find method without recursion.

Python

My Python code is mostly for illustration purposes. Check out this Python library. That website warns of the traps that multithreading in NumPy can pose.

C

In the other original paper, Newman and Ziff included C code of their algorithm. You can find it on Newman’s website, but I have also a copy of it (with added comments from me) located here.

Future work

I plan to implement this algorithm using the NVIDIA CUDA library. This is exactly the type of algorithm that we can implement and run using graphical processing units (GPUs), which is the entire point of the CUDA library.

At first glance, the spatial dependence of the problem suggests using texture memory to speed up the memory accessing on the GPUs. But I will explain why that’s probably not a good idea for this specific algorithm

New link: gregorygundersen.com

I came across this blog:

http://gregorygundersen.com/blog

The writer now focuses mostly on financial models and techniques, but earlier posts cover topics in probability and statistics.

Reversibility

First Markov chain

Irreducibility

Creating a new Markov chain with acceptance

Transition kernel

\(\alpha(x,y)\) needs a symmetric function \(s(x,y)\)

Examples

Barker

Metropolis(-Rosenbluth-Rosenbluth-Teller-Teller)-Hastings

Further reading

Articles

Historical

Introductory

History

Books

Websites

Monte Carlo conditions

Sum

Integral

Ergodic sequence

Markov chain conditions

Stationary distribution \(\pi\)

Aperiodicity

Irreducibility

Positive recurrence

Ergodicity

Sampling with a Markov chain

Summary

Further reading

Books

Markov chains

Monte Carlo methods

Monte Carlo estimates

Estimating with uniform variables

Estimating with general variables

Error rate

Losing the independence

Ergodicity

Ergodic sequences

Further reading

Markov chains vs Markov processes

Markov the frog

State space

Markov property

Markov kernel

Initial distribution

Jumping from \(x\) to \(y\)

Regularity conditions

Stationary distribution \(\pi\)

Aperiodicity

Irreducibility

Recurrence

Positive recurrence

Ergodicity

Further reading

Details

Algorithm

Point process application

Efficiency

Higher dimensions

Further reading

Proof outline

Algorithm

The fall of the Box-Muller method

Update: The return of the Box-Muller method

Further reading

Websites

Papers

Books

Details

Algorithm

Examples

Bernoulli distribution

Application: Acceptance simulation methods

Exponential distribution

Application: Poisson simulation method

Rayleigh distribution

Other methods

Further reading

Algorithm