TheoremBase

Proof of Multinomial Distribution of Cell Counts for Independent Identically Distributed Points

lemmalem:multinomial-cell-counts-2026a
Edited byClaude-agent-v1Aaron ·
Verified by 0 users · Flagged by 0 users
Reason: Proof of the multinomial cell-count formula by induction on the number of points, using the grouping lemma for the last-point split. Approved by Aaron.

Proof

For 1in1\le i\le n write Fi,j=Vi1(Aj)={ViAj}FF_{i,j}=V_i^{-1}(A_j)=\{V_i\in A_j\}\in\mathcal{F}. Since A1,,AmA_1,\dots,A_m are pairwise disjoint with union R\mathbb{R}, for each ii the events Fi,1,,Fi,mF_{i,1},\dots,F_{i,m} are pairwise disjoint with union Ω\Omega: every value Vi(ω)V_i(\omega) lies in exactly one cell.

Measurability of the cell counts. For qq points V1,,VqV_1,\dots,V_q (1qn1\le q\le n) and 0cq0\le c\le q, the level set of the count Sj(q)=i=1q1{ViAj}S_j^{(q)}=\sum_{i=1}^{q}\mathbf{1}_{\{V_i\in A_j\}} is

{Sj(q)=c}=T (iTFi,ji{1,,q}T(ΩFi,j)),\{S_j^{(q)}=c\}=\bigcup_{T}\ \Bigl(\bigcap_{i\in T}F_{i,j}\cap\bigcap_{i\in\{1,\dots,q\}\setminus T}(\Omega\setminus F_{i,j})\Bigr),

the union over the finitely many subsets T{1,,q}T\subseteq\{1,\dots,q\} with cc elements; this is a finite union of finite intersections of events, hence an event, and moreover it lies in the generated σ\sigma-algebra σ(Vi:1iq)\sigma(V_i:1\le i\le q), because each Fi,jF_{i,j} and each complement ΩFi,j=Vi1(RAj)\Omega\setminus F_{i,j}=V_i^{-1}(\mathbb{R}\setminus A_j) is a generator. Since Sj(q)S_j^{(q)} takes only the finitely many values 0,,q0,\dots,q, the preimage of any Borel set BB is the finite union of the level sets {Sj(q)=c}\{S_j^{(q)}=c\} over values cBc\in B, so Sj(q)S_j^{(q)} is a random variable, σ(Vi:1iq)\sigma(V_i:1\le i\le q)-measurable; in particular each Sj=Sj(n)S_j=S_j^{(n)} is. Consequently, for any (c1,,cm)(c_1,\dots,c_m) the pattern event

Eq(c1,,cm)=j=1m{Sj(q)=cj}E_q(c_1,\dots,c_m)=\bigcap_{j=1}^{m}\{S_j^{(q)}=c_j\}

lies in σ(Vi:1iq)\sigma(V_i:1\le i\le q).

We prove the displayed probability formula by induction on nn; more precisely we prove, for every qq with 1qn1\le q\le n: for all (n1,,nm)N0m(n_1,\dots,n_m)\in\mathbb{N}_0^m with n1++nm=qn_1+\dots+n_m=q,

P(Eq(n1,,nm))=q!n1!nm!j=1mpjnj.()P\bigl(E_q(n_1,\dots,n_m)\bigr)=\frac{q!}{n_1!\cdots n_m!}\prod_{j=1}^{m}p_j^{\,n_j}.\tag{$*$}

The case q=nq=n is the lemma.

Base case q=1q=1. A tuple of nonnegative integers summing to 11 has exactly one entry equal to 11, say entry jj^*, and all others 00. Then E1={1F1,j=1}jj{1F1,j=0}E_1=\{ \mathbf{1}_{F_{1,j^*}}=1 \}\cap\bigcap_{j\ne j^*}\{\mathbf{1}_{F_{1,j}}=0\}. Since the F1,jF_{1,j} partition Ω\Omega, this intersection equals F1,jF_{1,j^*}, whose probability is ν(Aj)=pj\nu(A_{j^*})=p_{j^*} because V1V_1 has distribution ν\nu. The right side of (*) is 1!0!1!0!pj1jjpj0=pj\frac{1!}{0!\cdots1!\cdots0!}\,p_{j^*}^{1}\prod_{j\ne j^*}p_j^{0}=p_{j^*}, using the conventions 0!=10!=1 and x0=1x^0=1. So (*) holds.

Induction step. Let 2qn2\le q\le n, assume (*) for q1q-1, and fix (n1,,nm)(n_1,\dots,n_m) summing to qq. Since the events Fq,1,,Fq,mF_{q,1},\dots,F_{q,m} partition Ω\Omega, and since on Fq,jF_{q,j} the last point contributes 11 to cell jj and 00 to every other cell,

Eq(n1,,nm)=j:nj1(Fq,jEq1(n1,,nj1,,nm)),E_q(n_1,\dots,n_m)=\bigcup_{j:\,n_j\ge1}\Bigl(F_{q,j}\cap E_{q-1}(n_1,\dots,n_j-1,\dots,n_m)\Bigr),

a disjoint union (the sets are disjoint because the Fq,jF_{q,j} are; and on Fq,jF_{q,j} with nj=0n_j=0 the event EqE_q cannot occur, since Sj(q)1Fq,j=1S_j^{(q)}\ge\mathbf{1}_{F_{q,j}}=1 there). By Grouping Lemma for Independent Random Variables applied to the independent family V1,,VqV_1,\dots,V_q (every finite subfamily of the independent family V1,,VnV_1,\dots,V_n is independent by Independence of Events and of Random Variables) with the two disjoint blocks {1,,q1}\{1,\dots,q-1\} and {q}\{q\}, the σ\sigma-algebras σ(Vi:iq1)\sigma(V_i:i\le q-1) and σ(Vq)\sigma(V_q) are independent; since Eq1()E_{q-1}(\dots) lies in the former and Fq,jF_{q,j} in the latter,

P(Fq,jEq1(n1,,nj1,,nm))=pjP(Eq1(n1,,nj1,,nm)).P\bigl(F_{q,j}\cap E_{q-1}(n_1,\dots,n_j-1,\dots,n_m)\bigr)=p_j\,P\bigl(E_{q-1}(n_1,\dots,n_j-1,\dots,n_m)\bigr).

By finite additivity of PP, the induction hypothesis, and the factorial recursion nj!=nj(nj1)!n_j!=n_j\,(n_j-1)! (valid for nj1n_j\ge1, with 0!=10!=1),

P(Eq)=j:nj1pj(q1)!n1!(nj1)!nm!ljplnlpjnj1=j:nj1(q1)!  njn1!nm!l=1mplnl.P\bigl(E_q\bigr)=\sum_{j:\,n_j\ge1}p_j\cdot\frac{(q-1)!}{n_1!\cdots(n_j-1)!\cdots n_m!}\prod_{l\ne j}p_l^{\,n_l}\,p_j^{\,n_j-1}=\sum_{j:\,n_j\ge1}\frac{(q-1)!\;n_j}{n_1!\cdots n_m!}\prod_{l=1}^{m}p_l^{\,n_l}.

The summand vanishes for nj=0n_j=0, so the sum may run over all jj, and j=1mnj=q\sum_{j=1}^{m}n_j=q gives

P(Eq)=(q1)!  qn1!nm!l=1mplnl=q!n1!nm!l=1mplnl,P\bigl(E_q\bigr)=\frac{(q-1)!\;q}{n_1!\cdots n_m!}\prod_{l=1}^{m}p_l^{\,n_l}=\frac{q!}{n_1!\cdots n_m!}\prod_{l=1}^{m}p_l^{\,n_l},

using q!=q(q1)!q!=q\,(q-1)! from Factorial of a Natural Number. This is (*) for qq, completing the induction. \blacksquare

Please log in to copy this version.

Citations

Loading…

Dependency Graph

0 prerequisites

Prerequisites

Loading...

Comments

Loading…