Past exam of the mathematics course of the University of Cambridge / 2004 / iii / Paper 29 (Poisson Processes)

For $\theta >0$, the one-parameter Poisson-Dirichlet distribution is a probability distribution on decreasing random mass sequences. A precise construction uses a Poisson random measure $\Xi$ on $(0,\infty )$ with intensityArrange its points as $J_1>J_2>\cdots$ and put $T={\sum }_i{J}_i$. The intensity is diffuse, so there are almost surely no ties. There are infinitely many points because the intensity has infinite mass near zero, but their sum is finite almost surely: the Campbell first-moment formula gives $\mathbb{E}T=\int u{\rho }_{\theta }(du)=\theta <\infty$. Also $T>0$ almost surely, since the probability of having no points is zero. DefineIts law is $PD(\theta )$, also written $PD(0,\theta )$. In particular, $p_1$ is the largest component; the ordering is part of the definition.

The Laplace functional of a Poisson random measure determines the total:For the integral identity, differentiate the integral with respect to $s$, obtaining $1/(1+s)$, and use its value zero at $s=0$. Thus $T$ has the Gamma distribution with shape $\theta$ and rate one, and density $t^{\theta -1}{e}^{-t}/\Gamma (\theta )$.

We use the Mecke formula for a Poisson random measure in its point-removal formHere the process inside the expectation on the right has the original law. To justify this identity, first take finite intensity. Conditional on a Poisson count $n$, its points are independent samples from normalized intensity. Expanding the expectation as a sum over $n$ and singling out one of the $n$ points replaces $n/n!$ by $1/(n-1)!$; the remaining points have the original Poisson law. For sigma-finite intensity, restrict the selected point to a finite-intensity set, condition on the independent outside process, and then exhaust the space. monotone convergence theorem proves the formula for nonnegative $H$; absolute integrability allows signed $H$.

Apply it with $H(u,\eta )=\phi (u/(u+\int v\eta (dv)))$. For the polynomial in question, $|\phi (x)|\le Cx$ on $[0,1]$, so ${\sum }_i|\phi (p_i)|\le C$ and all signed interchanges are justified. Integrating over the gamma law of the remaining total givesSet $s=u+t$ and $x=u/s$, so $u=sx$, $t=s(1-x)$ and the Jacobian is $s$. The integrand factors as $\phi (x)x^{-1}(1-x{)}^{\theta -1}{s}^{\theta -1}{e}^{-s}$. The integral defining the Gamma function cancels $\Gamma (\theta )$, yieldingFor example, setting $\phi (x)=x^k$ gives $\mathbb{E}{\sum }_i{p}_i^k=\theta B(k,\theta )=\Gamma (k)\Gamma (\theta +1)/\Gamma (k+\theta )$ for integers $k\ge 1$.

The same nonnegative calculation identifies the mean measure of Poisson-Dirichlet components for arbitrary nonnegative test functions. Write $N_a=\#\{i:p_i>a\}$. If $1/2\le a<1$, at most one component exceeds $a$, so $N_a=1_{\{p_1>a\}}$. Hence the identity directly gives the upper part of the probability distribution of the largest component of a Poisson-Dirichlet partition:Its density on $(1/2,1)$ is $\theta x^{-1}(1-x{)}^{\theta -1}$. For $\theta =1$, this tail is $-\log a$, and $\mathbb{P}(p_1>1/2)=\log 2$. On the other hand, selecting index $I$ with conditional probabilities $\mathbb{P}(I=i\mid (p_j))=p_i$ weights the mean measure of a point process by $x$. Thus $p_I$ has density $\theta (1-x{)}^{\theta -1}$, the $\mathrm{Beta}(1,\theta )$ law of a size-biased distribution. That beta law describes the selected component.

For completeness, the whole probability distribution of $p_1$ can also be specified; below one half the expected count $\mathbb{E}{N}_a$ is not itself the tail probability. Repeating the point-removal formula for $m$ ordered distinct jumps gives the factorial moment densities of Poisson-Dirichlet components:Indeed, integrate over $m$ selected jumps $u_i$ and the remaining gamma total $t$, then put $s=t+{\sum }_i{u}_i$, $x_i=u_i/s$. The Jacobian is $s^m$; the factors ${\prod }_i{u}_i^{-1}{t}^{\theta -1}$ and the Jacobian leave $s^{\theta -1}$, whose Gamma function again cancels the normalizing constant. Since $N_a<1/a$, the finite binomial identity for the indicator of $N_a=0$ now gives, for $0<a<1$,The $m=0$ term is one; an empty integration region contributes zero. This is inclusion-exclusion, since $1_{\{N_a=0\}}={\sum }_m(-1{)}^m(N_a{)}_m/m!$. Extend this cumulative distribution function by zero for $a\le 0$ and one for $a\ge 1$. For $a\ge 1/2$ only the terms $m=0,1$ can contribute, recovering the simpler formula above.

For $r>0$, let ${\kappa }_r$ be the probability law of a seed displacement from its tree. For the literal circular support, it is normalized arclength on $\{u:|u|=r\}$. The conditional mean measure of a point process isWe prove the conditional Poisson assertion directly by joint probability generating functions.

Fix disjoint bounded measurable sets $A_1,\dots ,A_k$. For a tree at $z$, put $q_j(z)={\kappa }_r(A_j-z)$ and $q_0(z)=1-{\sum }_j{q}_j(z)$. Conditional on its seed count, the allocation among these sets and their complement is multinomial distribution. Averaging over the Poisson seed count givesThus this tree contributes independent Poisson counts of means $\mu q_j(z)$ to the disjoint sets. Different trees contribute independently. Only trees within distance $r$ of the bounded union of the $A_j$ can contribute, and there are almost surely finitely many of them. Multiplying their probability generating functions givesThis is precisely the joint probability generating function of independent Poisson counts with the displayed means. Also ${\Lambda }_{\Pi }$ is locally finite, because each bounded set can receive seeds from only finitely many trees. This proves the conditional Poisson point process assertion and the Poisson offspring clusters directed by their parent process construction.

In arclength notation the directing measure is ${\Lambda }_{\Pi }=\mu {\sum }_{z\in \Pi }{\sigma }_{z,r}/(2\pi r)$, where ${\sigma }_{z,r}$ is arclength on the circle centered at $z$. If “circle” is used to mean the filled disc instead, take ${\kappa }_r(du)=1_{\{|u|\le r\}}du/(\pi r^2)$; thenThe generating-function proof and the following unconditional conclusion hold for either reading. The circular version has a singular conditional directing measure, giving the random directing measure of a Cox process formulation.

Unconditionally this is a Cox process, specifically a Poisson-offspring Neyman-Scott process. Its mean measure of a point process is nevertheless spatially uniform. With $q_A(z)={\kappa }_r(A-z)$, the Campbell first-moment formula and translation invariance givesince $\int q_A(z)dz=\int {\kappa }_r(du)\int 1_A(z+u)dz=|A|$. Uniform mean alone does not establish a Poisson law.

For a bounded set $A$ of positive area, the conditional count has mean and variance ${\Lambda }_{\Pi }(A)$. The law of total variance givesA Poisson integral of a deterministic function $q$ has variance $\lambda \int q^{2}dz$: for a simple function on disjoint sets this follows from independent Poisson counts, and square-integrable approximation gives the general formula. Here $0\le q_A\le 1$ and it has bounded support, soThe strict inequality holds when $\lambda ,\mu >0$, because $\int q_A=|A|>0$ implies $\int q_A^2>0$. A Poisson random variable has equal mean and variance. Hence the unconditional process is not a Poisson process in the nontrivial model. If $\lambda =0$ or $\mu =0$, there are no seeds and the process is the degenerate empty Poisson process. The overdispersion of nondegenerate mixed Poisson counts quantifies the additional clustering induced by the random parents.

Let $\sigma$ denote surface measure on a sphere on $S^2$ and $\nu$ volume measure on the unit ball. The Poisson mapping theorem can be proved directly here: counts of image points in disjoint sets are counts of the original process in their disjoint preimages. They are therefore independent Poisson variables, and their means are the volumes of those preimages.

For the radius map, a measurable set $A\subset [0,1]$ has meanThis follows from spherical coordinates or by subtracting concentric-ball volumes. The disjoint-preimage argument proves that the radius image is a Poisson process with this mean measure of a point process. There is no mass at zero or at one.

The original process has no point at the origin almost surely, since its mean count there is zero. Thus division by $r$ in the direction map is well-defined almost surely. For a measurable surface set $C\subset S^2$, spherical coordinates giveAgain disjoint surface sets have disjoint cone preimages, giving the required independent Poisson counts. Both pushforward measures are diffuse, so there are almost surely no image collisions; they describe the indicated point sets as well as counting measures.

The two image processes are not independent. They have exactly the same total number of points,Indeed,This alone disproves process independence.

More generally, split the original points into those whose radius lies in $A$ and direction lies in $C$, those satisfying only the first condition, and those satisfying only the second. Counts in these three disjoint regions are independent. The count shared by both images is Poisson with mean $\sigma (C){\int }_A{r}^{2}dr$, soThe fact that Poisson polar projections share their total count explains the dependence despite the product polar intensity $r^{2}dr\sigma (d\omega )$. Conditional on $N$, the points are independent uniform samples from the ball. Their radius density is $3r^2$ and their direction law is $\sigma /(4\pi )$, independently for each point. Thus the two image configurations are independent conditional on $N$; unconditionally the shared nonconstant count couples them.