You are currently browsing the tag archive for the ‘point processes’ tag.

Define a partition of ${1}$ to be a finite or infinite multiset ${\Sigma}$ of real numbers in the interval ${I \in (0,1]}$ (that is, an unordered set of real numbers in ${I}$, possibly with multiplicity) whose total sum is ${1}$: ${\sum_{t \in \Sigma}t = 1}$. For instance, ${\{1/2,1/4,1/8,1/16,\ldots\}}$ is a partition of ${1}$. Such partitions arise naturally when trying to decompose a large object into smaller ones, for instance:

1. (Prime factorisation) Given a natural number ${n}$, one can decompose it into prime factors ${n = p_1 \ldots p_k}$ (counting multiplicity), and then the multiset

$\displaystyle \Sigma_{PF}(n) := \{ \frac{\log p_1}{\log n}, \ldots,\frac{\log p_k}{\log n} \}$

is a partition of ${1}$.

2. (Cycle decomposition) Given a permutation ${\sigma \in S_n}$ on ${n}$ labels ${\{1,\ldots,n\}}$, one can decompose ${\sigma}$ into cycles ${C_1,\ldots,C_k}$, and then the multiset

$\displaystyle \Sigma_{CD}(\sigma) := \{ \frac{|C_1|}{n}, \ldots, \frac{|C_k|}{n} \}$

is a partition of ${1}$.

3. (Normalisation) Given a multiset ${\Gamma}$ of positive real numbers whose sum ${S := \sum_{x\in \Gamma}x}$ is finite and non-zero, the multiset

$\displaystyle \Sigma_N( \Gamma) := \frac{1}{S} \cdot \Gamma = \{ \frac{x}{S}: x \in \Gamma \}$

is a partition of ${1}$.

In the spirit of the universality phenomenon, one can ask what is the natural distribution for what a “typical” partition should look like; thus one seeks a natural probability distribution on the space of all partitions, analogous to (say) the gaussian distributions on the real line, or GUE distributions on point processes on the line, and so forth. It turns out that there is one natural such distribution which is related to all three examples above, known as the Poisson-Dirichlet distribution. To describe this distribution, we first have to deal with the problem that it is not immediately obvious how to cleanly parameterise the space of partitions, given that the cardinality of the partition can be finite or infinite, that multiplicity is allowed, and that we would like to identify two partitions that are permutations of each other

One way to proceed is to random partition ${\Sigma}$ as a type of point process on the interval ${I}$, with the constraint that ${\sum_{x \in \Sigma} x = 1}$, in which case one can study statistics such as the counting functions

$\displaystyle N_{[a,b]} := |\Sigma \cap [a,b]| = \sum_{x \in\Sigma} 1_{[a,b]}(x)$

(where the cardinality here counts multiplicity). This can certainly be done, although in the case of the Poisson-Dirichlet process, the formulae for the joint distribution of such counting functions is moderately complicated. Another way to proceed is to order the elements of ${\Sigma}$ in decreasing order

$\displaystyle t_1 \geq t_2 \geq t_3 \geq \ldots \geq 0,$

with the convention that one pads the sequence ${t_n}$ by an infinite number of zeroes if ${\Sigma}$ is finite; this identifies the space of partitions with an infinite dimensional simplex

$\displaystyle \{ (t_1,t_2,\ldots) \in [0,1]^{\bf N}: t_1 \geq t_2 \geq \ldots; \sum_{n=1}^\infty t_n = 1 \}.$

However, it turns out that the process of ordering the elements is not “smooth” (basically because functions such as ${(x,y) \mapsto \max(x,y)}$ and ${(x,y) \mapsto \min(x,y)}$ are not smooth) and the formulae for the joint distribution in the case of the Poisson-Dirichlet process is again complicated.

It turns out that there is a better (or at least “smoother”) way to enumerate the elements ${u_1,(1-u_1)u_2,(1-u_1)(1-u_2)u_3,\ldots}$ of a partition ${\Sigma}$ than the ordered method, although it is random rather than deterministic. This procedure (which I learned from this paper of Donnelly and Grimmett) works as follows.

1. Given a partition ${\Sigma}$, let ${u_1}$ be an element of ${\Sigma}$ chosen at random, with each element ${t\in \Sigma}$ having a probability ${t}$ of being chosen as ${u_1}$ (so if ${t \in \Sigma}$ occurs with multiplicity ${m}$, the net probability that ${t}$ is chosen as ${u_1}$ is actually ${mt}$). Note that this is well-defined since the elements of ${\Sigma}$ sum to ${1}$.
2. Now suppose ${u_1}$ is chosen. If ${\Sigma \backslash \{u_1\}}$ is empty, we set ${u_2,u_3,\ldots}$ all equal to zero and stop. Otherwise, let ${u_2}$ be an element of ${\frac{1}{1-u_1} \cdot (\Sigma \backslash \{u_1\})}$ chosen at random, with each element ${t \in \frac{1}{1-u_1} \cdot (\Sigma \backslash \{u_1\})}$ having a probability ${t}$ of being chosen as ${u_2}$. (For instance, if ${u_1}$ occurred with some multiplicity ${m>1}$ in ${\Sigma}$, then ${u_2}$ can equal ${\frac{u_1}{1-u_1}}$ with probability ${(m-1)u_1/(1-u_1)}$.)
3. Now suppose ${u_1,u_2}$ are both chosen. If ${\Sigma \backslash \{u_1,u_2\}}$ is empty, we set ${u_3, u_4, \ldots}$ all equal to zero and stop. Otherwise, let ${u_3}$ be an element of ${\frac{1}{1-u_1-u_2} \cdot (\Sigma\backslash \{u_1,u_2\})}$, with ech element ${t \in \frac{1}{1-u_1-u_2} \cdot (\Sigma\backslash \{u_1,u_2\})}$ having a probability ${t}$ of being chosen as ${u_3}$.
4. We continue this process indefinitely to create elements ${u_1,u_2,u_3,\ldots \in [0,1]}$.

We denote the random sequence ${Enum(\Sigma) := (u_1,u_2,\ldots) \in [0,1]^{\bf N}}$ formed from a partition ${\Sigma}$ in the above manner as the random normalised enumeration of ${\Sigma}$; this is a random variable in the infinite unit cube ${[0,1]^{\bf N}}$, and can be defined recursively by the formula

$\displaystyle Enum(\Sigma) = (u_1, Enum(\frac{1}{1-u_1} \cdot (\Sigma\backslash \{u_1\})))$

with ${u_1}$ drawn randomly from ${\Sigma}$, with each element ${t \in \Sigma}$ chosen with probability ${t}$, except when ${\Sigma =\{1\}}$ in which case we instead have

$\displaystyle Enum(\{1\}) = (1, 0,0,\ldots).$

Note that one can recover ${\Sigma}$ from any of its random normalised enumerations ${Enum(\Sigma) := (u_1,u_2,\ldots)}$ by the formula

$\displaystyle \Sigma = \{ u_1, (1-u_1) u_2,(1-u_1)(1-u_2)u_3,\ldots\} \ \ \ \ \ (1)$

with the convention that one discards any zero elements on the right-hand side. Thus ${Enum}$ can be viewed as a (stochastic) parameterisation of the space of partitions by the unit cube ${[0,1]^{\bf N}}$, which is a simpler domain to work with than the infinite-dimensional simplex mentioned earlier.

Note that this random enumeration procedure can also be adapted to the three models described earlier:

1. Given a natural number ${n}$, one can randomly enumerate its prime factors ${n =p'_1 p'_2 \ldots p'_k}$ by letting each prime factor ${p}$ of ${n}$ be equal to ${p'_1}$ with probability ${\frac{\log p}{\log n}}$, then once ${p'_1}$ is chosen, let each remaining prime factor ${p}$ of ${n/p'_1}$ be equal to ${p'_2}$ with probability ${\frac{\log p}{\log n/p'_1}}$, and so forth.
2. Given a permutation ${\sigma\in S_n}$, one can randomly enumerate its cycles ${C'_1,\ldots,C'_k}$ by letting each cycle ${C}$ in ${\sigma}$ be equal to ${C'_1}$ with probability ${\frac{|C|}{n}}$, and once ${C'_1}$ is chosen, let each remaining cycle ${C}$ be equalto ${C'_2}$ with probability ${\frac{|C|}{n-|C'_1|}}$, and so forth. Alternatively, one traverse the elements of ${\{1,\ldots,n\}}$ in random order, then let ${C'_1}$ be the first cycle one encounters when performing this traversal, let ${C'_2}$ be the next cycle (not equal to ${C'_1}$ one encounters when performing this traversal, and so forth.
3. Given a multiset ${\Gamma}$ of positive real numbers whose sum ${S := \sum_{x\in\Gamma} x}$ is finite, we can randomly enumerate ${x'_1,x'_2,\ldots}$ the elements of this sequence by letting each ${x \in \Gamma}$ have a ${\frac{x}{S}}$ probability of being set equal to ${x'_1}$, and then once ${x'_1}$ is chosen, let each remaining ${x \in \Gamma\backslash \{x'_1\}}$ have a ${\frac{x_i}{S-x'_1}}$ probability of being set equal to ${x'_2}$, and so forth.

We then have the following result:

Proposition 1 (Existence of the Poisson-Dirichlet process) There exists a random partition ${\Sigma}$ whose random enumeration ${Enum(\Sigma) = (u_1,u_2,\ldots)}$ has the uniform distribution on ${[0,1]^{\bf N}}$, thus ${u_1,u_2,\ldots}$ are independently and identically distributed copies of the uniform distribution on ${[0,1]}$.

A random partition ${\Sigma}$ with this property will be called the Poisson-Dirichlet process. This process, first introduced by Kingman, can be described explicitly using (1) as

$\displaystyle \Sigma = \{ u_1, (1-u_1) u_2,(1-u_1)(1-u_2)u_3,\ldots\},$

where ${u_1,u_2,\ldots}$ are iid copies of the uniform distribution of ${[0,1]}$, although it is not immediately obvious from this definition that ${Enum(\Sigma)}$ is indeed uniformly distributed on ${[0,1]^{\bf N}}$. We prove this proposition below the fold.

An equivalent definition of a Poisson-Dirichlet process is a random partition ${\Sigma}$ with the property that

$\displaystyle (u_1, \frac{1}{1-u_1} \cdot (\Sigma \backslash \{u_1\})) \equiv (U, \Sigma) \ \ \ \ \ (2)$

where ${u_1}$ is a random element of ${\Sigma}$ with each ${t \in\Sigma}$ having a probability ${t}$ of being equal to ${u_1}$, ${U}$ is a uniform variable on ${[0,1]}$ that is independent of ${\Sigma}$, and ${\equiv}$ denotes equality of distribution. This can be viewed as a sort of stochastic self-similarity property of ${\Sigma}$: if one randomly removes one element from ${\Sigma}$ and rescales, one gets a new copy of ${\Sigma}$.

It turns out that each of the three ways to generate partitions listed above can lead to the Poisson-Dirichlet process, either directly or in a suitable limit. We begin with the third way, namely by normalising a Poisson process to have sum ${1}$:

Proposition 2 (Poisson-Dirichlet processes via Poisson processes) Let ${a>0}$, and let ${\Gamma_a}$ be a Poisson process on ${(0,+\infty)}$ with intensity function ${t \mapsto \frac{1}{t} e^{-at}}$. Then the sum ${S :=\sum_{x \in \Gamma_a} x}$ is almost surely finite, and the normalisation ${\Sigma_N(\Gamma_a) = \frac{1}{S} \cdot \Gamma_a}$ is a Poisson-Dirichlet process.

Again, we prove this proposition below the fold. Now we turn to the second way (a topic, incidentally, that was briefly touched upon in this previous blog post):

Proposition 3 (Large cycles of a typical permutation) For each natural number ${n}$, let ${\sigma}$ be a permutation drawn uniformly at random from ${S_n}$. Then the random partition ${\Sigma_{CD}(\sigma)}$ converges in the limit ${n \rightarrow\infty}$ to a Poisson-Dirichlet process ${\Sigma_{PF}}$ in the following sense: given any fixed sequence of intervals ${[a_1,b_1],\ldots,[a_k,b_k] \subset I}$ (independent of ${n}$), the joint discrete random variable ${(N_{[a_1,b_1]}(\Sigma_{CD}(\sigma)),\ldots,N_{[a_k,b_k]}(\Sigma_{CD}(\sigma)))}$ converges in distribution to ${(N_{[a_1,b_1]}(\Sigma),\ldots,N_{[a_k,b_k]}(\Sigma))}$.

Finally, we turn to the first way:

Proposition 4 (Large prime factors of a typical number) Let ${x > 0}$, and let ${N_x}$ be a random natural number chosen according to one of the following three rules:

1. (Uniform distribution) ${N_x}$ is drawn uniformly at random from the natural numbers in ${[1,x]}$.
2. (Shifted uniform distribution) ${N_x}$ is drawn uniformly at random from the natural numbers in ${[x,2x]}$.
3. (Zeta distribution) Each natural number ${n}$ has a probability ${\frac{1}{\zeta(s)}\frac{1}{n^s}}$ of being equal to ${N_x}$, where ${s := 1 +\frac{1}{\log x}}$and ${\zeta(s):=\sum_{n=1}^\infty \frac{1}{n^s}}$.

Then ${\Sigma_{PF}(N_x)}$ converges as ${x \rightarrow \infty}$ to a Poisson-Dirichlet process ${\Sigma}$ in the same fashion as in Proposition 3.

The process ${\Sigma_{PF}(N_x)}$ was first studied by Billingsley (and also later by Knuth-Trabb Pardo and by Vershik, but the formulae were initially rather complicated; the proposition above is due to of Donnelly and Grimmett, although the third case of the proposition is substantially easier and appears in the earlier work of Lloyd. We prove the proposition below the fold.

The previous two propositions suggests an interesting analogy between large random integers and large random permutations; see this ICM article of Vershik and this non-technical article of Granville (which, incidentally, was once adapted into a play) for further discussion.

As a sample application, consider the problem of estimating the number ${\pi(x,x^{1/u})}$ of integers up to ${x}$ which are not divisible by any prime larger than ${x^{1/u}}$ (i.e. they are ${x^{1/u}}$-smooth), where ${u>0}$ is a fixed real number. This is essentially (modulo some inessential technicalities concerning the distinction between the intervals ${[x,2x]}$ and ${[1,x]}$) the probability that ${\Sigma}$ avoids ${[1/u,1]}$, which by the above theorem converges to the probability ${\rho(u)}$ that ${\Sigma_{PF}}$ avoids ${[1/u,1]}$. Below the fold we will show that this function is given by the Dickman function, defined by setting ${\rho(u)=1}$ for ${u < 1}$ and ${u\rho'(u) = \rho(u-1)}$ for ${u \geq 1}$, thus recovering the classical result of Dickman that ${\pi(x,x^{1/u}) = (\rho(u)+o(1))x}$.

I thank Andrew Granville and Anatoly Vershik for showing me the nice link between prime factors and the Poisson-Dirichlet process. The material here is standard, and (like many of the other notes on this blog) was primarily written for my own benefit, but it may be of interest to some readers. In preparing this article I found this exposition by Kingman to be helpful.

Note: this article will emphasise the computations rather than rigour, and in particular will rely on informal use of infinitesimals to avoid dealing with stochastic calculus or other technicalities. We adopt the convention that we will neglect higher order terms in infinitesimal calculations, e.g. if ${dt}$ is infinitesimal then we will abbreviate ${dt + o(dt)}$ simply as ${dt}$.

Given a set ${S}$, a (simple) point process is a random subset ${A}$ of ${S}$. (A non-simple point process would allow multiplicity; more formally, ${A}$ is no longer a subset of ${S}$, but is a Radon measure on ${S}$, where we give ${S}$ the structure of a locally compact Polish space, but I do not wish to dwell on these sorts of technical issues here.) Typically, ${A}$ will be finite or countable, even when ${S}$ is uncountable. Basic examples of point processes include

• (Bernoulli point process) ${S}$ is an at most countable set, ${0 \leq p \leq 1}$ is a parameter, and ${A}$ a random set such that the events ${x \in A}$ for each ${x \in S}$ are jointly independent and occur with a probability of ${p}$ each. This process is automatically simple.
• (Discrete Poisson point process) ${S}$ is an at most countable space, ${\lambda}$ is a measure on ${S}$ (i.e. an assignment of a non-negative number ${\lambda(\{x\})}$ to each ${x \in S}$), and ${A}$ is a multiset where the multiplicity of ${x}$ in ${A}$ is a Poisson random variable with intensity ${\lambda(\{x\})}$, and the multiplicities of ${x \in A}$ as ${x}$ varies in ${S}$ are jointly independent. This process is usually not simple.
• (Continuous Poisson point process) ${S}$ is a locally compact Polish space with a Radon measure ${\lambda}$, and for each ${\Omega \subset S}$ of finite measure, the number of points ${|A \cap \Omega|}$ that ${A}$ contains inside ${\Omega}$ is a Poisson random variable with intensity ${\lambda(\Omega)}$. Furthermore, if ${\Omega_1,\ldots,\Omega_n}$ are disjoint sets, then the random variables ${|A \cap \Omega_1|, \ldots, |A \cap \Omega_n|}$ are jointly independent. (The fact that Poisson processes exist at all requires a non-trivial amount of measure theory, and will not be discussed here.) This process is almost surely simple iff all points in ${S}$ have measure zero.
• (Spectral point processes) The spectrum of a random matrix is a point process in ${{\mathbb C}}$ (or in ${{\mathbb R}}$, if the random matrix is Hermitian). If the spectrum is almost surely simple, then the point process is almost surely simple. In a similar spirit, the zeroes of a random polynomial are also a point process.

A remarkable fact is that many natural (simple) point processes are determinantal processes. Very roughly speaking, this means that there exists a positive semi-definite kernel ${K: S \times S \rightarrow {\mathbb R}}$ such that, for any ${x_1,\ldots,x_n \in S}$, the probability that ${x_1,\ldots,x_n}$ all lie in the random set ${A}$ is proportional to the determinant ${\det( (K(x_i,x_j))_{1 \leq i,j \leq n} )}$. Examples of processes known to be determinantal include non-intersecting random walks, spectra of random matrix ensembles such as GUE, and zeroes of polynomials with gaussian coefficients.

I would be interested in finding a good explanation (even at the heuristic level) as to why determinantal processes are so prevalent in practice. I do have a very weak explanation, namely that determinantal processes obey a large number of rather pretty algebraic identities, and so it is plausible that any other process which has a very algebraic structure (in particular, any process involving gaussians, characteristic polynomials, etc.) would be connected in some way with determinantal processes. I’m not particularly satisfied with this explanation, but I thought I would at least describe some of these identities below to support this case. (This is partly for my own benefit, as I am trying to learn about these processes, particularly in connection with the spectral distribution of random matrices.) The material here is partly based on this survey of Hough, Krishnapur, Peres, and Virág.