Appendix A Poisson Process and Poisson Random Measure Reference [130] is the main source for the material on Poisson process and Poisson random measure. A.1 Definitions Definition A.1. Let (Ω,F ,P) be a probability space, and S be a locally compact, separable metric space with Borel σ-algebra B.ThesetS denotes the collec- tion of all countable subsets of S.APoisson process with state space S, defined on (Ω,F ,P),isamap from (Ω,F ,P) to S satisfying: (a) for each B in B, N(B)=#{ ∩ B} is a Poisson random variable with parameter μ(B)=E[N(B)]; (b) for disjoint sets B1,...,Bn in B, N(B1),...,N(Bn) are independent. If B1,B2,... are disjoint, then, by definition, we have ∞ ∪∞ N( i=1Bi)=∑ N(Bi), i=1 and ∞ μ ∪∞ μ ( i=1Bi)=∑ (Bi). i=1 Hence, N is a random measure, and μ is a measure, both on (S,B). The random measure N can also be written in the following form: S. Feng, The Poisson–Dirichlet Distribution and Related Topics, 199 Probability and its Applications, DOI 10.1007/978-3-642-11194-5, © Springer-Verlag Berlin Heidelberg 2010 200 A Poisson Process and Poisson Random Measure N = ∑ δς . ς∈ Definition A.2. The measure μ is called the mean measure of the Poisson process , and N is called a Poisson random measure associated with the Poisson process . The measure μ is also called the mean measure of N. Remark: Let S =[0,∞) and μ be the Lebesgue measure on S. Then the Poisson random measure associated with the Poisson process with state space S and mean measure μ is just the one-dimensional time-homogeneous Poisson process that is defined as a pure-birth Markov chain with birth rate one. The random set is com- posed of all jump times of the process. Definition A.3. Assume that S = Rd for some d ≥ 1. Then the mean measure is also called the intensity measure. If there exists a positive constant c such that for any measurable set B, μ(B)=c|B|, |B| = Lebesgue measure of B, then the Poisson process is said to be homogeneous with intensity c. A.2 Properties Theorem A.1. Let μ be the mean measure of a Poisson process with state space S. Then μ is diffuse; i.e., for every x in S, μ({x})=0. Proof. For any fixed x in S,seta = μ({x}). Then, by definition, 2 a − P{N({x})=2} = e a = 0, 2 which leads to the result. 2 The next theorem describes the close relation between a Poisson process and the multinomial distribution. Theorem A.2. Let be a Poisson process with state space S and mean measure μ. Assume that the total mass μ(S) is finite. Then, for any n ≥ 1,1 ≤ m ≤ n, and any set partition B1,...,Bm of S, the conditional distribution of the random vector (N(B1),...,N(Bm)) given N(S)=n is a multinomial distribution with parameters n and μ(B ) μ(B ) 1 ,..., m . μ(S) μ(S) A.2 Properties 201 Proof. For any partitions n1,...,nm of n, P{N(B1)=n1,...,N(Bm)=nm | N(S)=n} P{N(B )=n ,...,N(B )=n } = 1 1 m m P{N(S)=n} −μ μ n (Bi) m (Bi)i e ∏i=1 = ni! μ(S)ne−μ(S) n! m ni n μ(Bi) = ∏ μ . n1,...,nm i=1 (S) 2 Theorem A.3. (Restriction and union) (1) Let be a Poisson process with state space S. Then, for every B in B, ∩ B is a Poisson process with state space S and mean measure μB(·)=μ(·∩B). Equivalently, ∩B can also be viewed as a Poisson process with state space B with mean measure given by the restriction of μ on B. (2) Let 1 and 2 be two independent Poisson processes with state space S, and respective mean measures μ1 and μ2. Then 1 ∪ 2 is a Poisson process with state space S and mean measure μ = μ1 + μ2. Proof. Direct verification of the definition of Poisson process. 2 The remaining theorems in this section are stated without proof. The details can be found in [130]. Theorem A.4. (Mapping) Let be a Poisson process with state space S and σ- finite mean measure μ(·). Consider a measurable map h from S to another locally compact, separable metric space S. If the measure μ(·)=μ( f −1(·)) is diffuse, then f ()={ f (ς) : ς ∈ } is a Poisson process with state space S and mean measure μ. Theorem A.5. (Marking) Let be a Poisson process with state space S and mean measure μ. The mark of each point ς in , denoted by mς , is a random variable, tak- ing values in a locally compact, separable metric space S, with distribution q(z,·). Assume that: (1) for every measurable set B in S,q(·,B) is a measurable function on (S,B); (2) given , the random variables {mς : ς ∈ } are independent; 202 A Poisson Process and Poisson Random Measure then ˜ = {(ς,mς ) : ς ∈ } is a Poisson process on with state space S×S and mean measure μ˜ (dx,dm)=μ(dx)q(x,dm). The Poisson process ˜ is aptly called a marked Poisson process. Theorem A.6. (Campbell) Let be a Poisson process on space (S,B) with mean measure μ. Then for any non-negative measurable function f , − E exp − ∑ f (ς) = exp (e f (x) − 1)μ(dx) . ς∈ S If f is a real-valued measurable function on (S,B) satisfying min(| f (x)|,1)μ(dx) < ∞, S then for any complex number λ such that the integral (eλ f (x) − 1)μ(dx) S converges, we have λ E exp λ ∑ f (ς) = exp (e f (x) − 1)μ(dx) . ς∈ S Moreover, if | f (x)|μ(dx) < ∞, (A.1) S then E ∑ f (ς) = f (x)μ(dx), ς∈ S Var ∑ f (ς) = f 2(x)μ(dx). ς∈ S In general, for any n ≥ 1, and any real-valued measurable functions f1,..., fn sat- isfying (A.1), we have n E ∑ f1(ς1)··· fn(ςn) = ∏E ∑ fi(ςi) . (A.2) ς ∈ distinct ς1,...,ςn∈ i=1 i Appendix B Basics of Large Deviations In probability theory, the law of large numbers describes the limiting average or mean behavior of a random population. The fluctuations around the average are characterized by a fluctuation theorem such as the central limit theorem. The theory of large deviations is concerned with the rare event of deviations from the average. Here we give a brief account of the basic definitions and results of large deviations. Everything will be stated in a form that will be sufficient for our needs. All proofs will be omitted. Classical references on large deviations include [30], [50], [168], and [175]. More recent developments can be found in [46], [28], and [69]. The formulations here follow mainly Dembo and Zeitouni [28]. Theorem B.6 is from [157]. Let E be a complete, separable metric space with metric ρ. Generic elements of E are denoted by x,y, etc. Definition B.1. A function I on E is called a rate function if it takes values in [0,+∞] and is lower semicontinuous. For each c in [0,+∞),theset {x ∈ E : I(x) ≤ c} is called a level set. The effective domain of I is defined as {x ∈ E : I(x) < ∞}. If all level sets are compact, the rate function is said to be good. Rate functions will be denoted by other symbols as the need arises. Let {Xε : ε > 0} be a family of E-valued random variables with distributions {Pε : ε > 0}, defined on the Borel σ-algebra B of E. Definition B.2. The family {Xε : ε > 0} or the family {Pε : ε > 0} is said to satisfy a large deviation principle (LDP) on E as ε converges to zero, with speed ε and a good rate function I if S. Feng, The Poisson–Dirichlet Distribution and Related Topics, 203 Probability and its Applications, DOI 10.1007/978-3-642-11194-5, © Springer-Verlag Berlin Heidelberg 2010 204 B Basics of Large Deviations for any closed set F, limsup ε logPε {F}≤−inf I(x), (B.1) ∈ ε→0 x F for any open set G, liminf ε logPε {G}≥−inf I(x). (B.2) ε→0 x∈G Estimates (B.1) and (B.2) are called the upper bound and lower bound, respectively. Let a(ε) be a function of ε satisfying a(ε) > 0, lim a(ε)=0. ε→0 If the multiplication factor ε in front of the logarithm is replaced by a(ε), then the LDP has speed a(ε). It is clear that the upper and lower bounds are equivalent to the following state- ment: for all B ∈ B, − inf I(x) ≤ liminfε logPε {B} x∈B◦ ε→0 limsupε logPε {B}≤−inf I(x), ε→0 x∈B where B◦ and B denote the interior and closure of B respectively. An event B ∈ B satisfying inf◦ I(x)=inf I(x) x∈B x∈B is called a I-continuity set. Thus for a I-continuity set B, we have that lim ε logPε {B} = − inf I(x). ε→0 x∈B If the values for ε are only {1/n : n ≥ 1}, we will write Pn instead of P1/n. If the upper bound (B.1) holds only for compact sets, then we say the family {Pε : ε > 0} satisfies the weak LDP.