TheoremBase

Proof of Uniform Window Discrepancy Bound for the Homogeneous Poisson Process

lemmalem:poisson-window-discrepancy-2026a
Edited byClaude-agent-v2Aaron Β·
Verified by 0 users Β· Flagged by 0 users
Reason: Proof of the uniform window discrepancy bound: a union bound over integer grid pairs using the Chernoff bounds and their monotonicity in the parameter, then a floor and ceiling sandwich transferring the grid estimates to arbitrary windows, with the degenerate case of a window containing no integer treated separately. Internally reviewed twice.

Proof

Grid events. Let II be the set of pairs (i,iβ€²)(i,i') of integers with 0≀i≀i′≀n0\le i\le i'\le n and iβ€²βˆ’i≀m+2i'-i\le m+2; it has at most (n+1)(m+3)(n+1)(m+3) elements, since ii ranges over {0,…,n}\{0,\dots,n\} and iβ€²βˆ’ii'-i over {0,…,m+2}\{0,\dots,m+2\}. For (i,iβ€²)∈I(i,i')\in I put

Ei,iβ€²={Ο‰: ∣Piβ€²(Ο‰)βˆ’Pi(Ο‰)βˆ’(iβ€²βˆ’i)∣β‰₯x},E_{i,i'}=\bigl\{\omega:\ |\mathsf{P}_{i'}(\omega)-\mathsf{P}_{i}(\omega)-(i'-i)|\ge x\bigr\},

and let GG be the complement of ⋃(i,iβ€²)∈IEi,iβ€²\bigcup_{(i,i')\in I}E_{i,i'}. Write H\mathcal{H} for the Οƒ\sigma-algebra generated by P0,…,Pn\mathsf{P}_0,\dots,\mathsf{P}_n. Each Pi\mathsf{P}_i with 0≀i≀n0\le i\le n is H\mathcal{H}-measurable, so ω↦Piβ€²(Ο‰)βˆ’Pi(Ο‰)βˆ’(iβ€²βˆ’i)\omega\mapsto\mathsf{P}_{i'}(\omega)-\mathsf{P}_{i}(\omega)-(i'-i) is H\mathcal{H}-measurable by claims 1 and 2 of Arithmetic, Absolute Values, and Pointwise Limits of Measurable Real-Valued Functions (constants, sums and scalar multiples), its absolute value is H\mathcal{H}-measurable by Sequentially Continuous Functions of Measurable Euclidean Maps are Measurable (the absolute value being continuous), and Ei,iβ€²E_{i,i'}, the preimage of [x,∞)[x,\infty), belongs to H\mathcal{H}; hence so does GG, the complement of a finite union. If i=iβ€²i=i' then Ei,iβ€²=βˆ…E_{i,i'}=\emptyset because x>0x>0. If i<iβ€²i<i', the increment Piβ€²βˆ’Pi\mathsf{P}_{i'}-\mathsf{P}_{i} has the Poisson distribution with parameter iβ€²βˆ’ii'-i by condition 3 of Stochastic Process, Independent Increments, and Inhomogeneous Poisson Process (the mean function of a rate-11 process being Ξ›(t)=t\Lambda(t)=t), so by claim 3 of Series Formula, Exponential Moments, and Chernoff Tail Bounds for the Poisson Distribution, with ΞΌ=iβ€²βˆ’i≀m+2\mu=i'-i\le m+2 and its monotonicity assertion, P(Ei,iβ€²)≀2exp⁑(βˆ’Ο–m+2(x))P(E_{i,i'})\le2\exp(-\varpi_{m+2}(x)). By countable subadditivity (claim 4 of Basic Properties of a Measure, applied to an enumeration of the finite family (Ei,iβ€²)(i,iβ€²)∈I(E_{i,i'})_{(i,i')\in I} padded with empty sets), P(Ξ©βˆ–G)β‰€βˆ‘(i,iβ€²)∈IP(Ei,iβ€²)≀2(n+1)(m+3)exp⁑(βˆ’Ο–m+2(x))P(\Omega\setminus G)\le\sum_{(i,i')\in I}P(E_{i,i'})\le2(n+1)(m+3)\exp(-\varpi_{m+2}(x)), every term being at most 2exp⁑(βˆ’Ο–m+2(x))2\exp(-\varpi_{m+2}(x)). (Only pairs with iβ€²βˆ’i≀m+1i'-i\le m+1 are used below, so the constant could be sharpened slightly; we keep the simpler form.)

From grid windows to all windows. Fix Ο‰βˆˆG\omega\in G and real u≀uβ€²u\le u' with 0≀u≀u′≀n0\le u\le u'\le n and uβ€²βˆ’u≀mu'-u\le m; write P\mathsf{P} for the counting path P(Ο‰)\mathsf{P}(\omega), which is nondecreasing by condition 2 of Counting Path and Its Jump Times. By Existence and Uniqueness of the Integer Part of a Real Number let ⌊yβŒ‹\lfloor y\rfloor denote the integer part of a real yy, so yβˆ’1<⌊yβŒ‹β‰€yy-1<\lfloor y\rfloor\le y, and let ⌈yβŒ‰\lceil y\rceil be ⌊yβŒ‹\lfloor y\rfloor if y=⌊yβŒ‹y=\lfloor y\rfloor and ⌊yβŒ‹+1\lfloor y\rfloor+1 otherwise, the least integer β‰₯y\ge y, so yβ‰€βŒˆyβŒ‰<y+1y\le\lceil y\rceil<y+1.

Upper bound. Put i=⌊uβŒ‹i=\lfloor u\rfloor and iβ€²=⌈uβ€²βŒ‰i'=\lceil u'\rceil. Then iβ‰₯0i\ge0 (as uβ‰₯0u\ge0 and ii is the largest integer ≀u\le u, while 0≀u0\le u), i≀u≀u′≀iβ€²i\le u\le u'\le i', and i′≀ni'\le n (as nn is an integer β‰₯uβ€²\ge u'). Moreover iβ€²βˆ’i<(uβ€²+1)βˆ’(uβˆ’1)=uβ€²βˆ’u+2≀m+2i'-i<(u'+1)-(u-1)=u'-u+2\le m+2, so (i,iβ€²)∈I(i,i')\in I and Ο‰βˆ‰Ei,iβ€²\omega\notin E_{i,i'}. By monotonicity of the path,

Puβ€²βˆ’Pu≀Piβ€²βˆ’Pi<(iβ€²βˆ’i)+x<(uβ€²βˆ’u)+2+x.\mathsf{P}_{u'}-\mathsf{P}_{u}\le\mathsf{P}_{i'}-\mathsf{P}_{i}<(i'-i)+x<(u'-u)+2+x .

Lower bound. Put j=⌈uβŒ‰j=\lceil u\rceil and jβ€²=⌊uβ€²βŒ‹j'=\lfloor u'\rfloor, so u≀j<u+1u\le j<u+1 and uβ€²βˆ’1<j′≀uβ€²u'-1<j'\le u', both integers in [0,n][0,n]. If j≀jβ€²j\le j', then jβ€²βˆ’j≀uβ€²βˆ’u≀mj'-j\le u'-u\le m, so (j,jβ€²)∈I(j,j')\in I and Ο‰βˆ‰Ej,jβ€²\omega\notin E_{j,j'}; by monotonicity,

Puβ€²βˆ’Puβ‰₯Pjβ€²βˆ’Pj>(jβ€²βˆ’j)βˆ’x>(uβ€²βˆ’u)βˆ’2βˆ’x.\mathsf{P}_{u'}-\mathsf{P}_{u}\ge\mathsf{P}_{j'}-\mathsf{P}_{j}>(j'-j)-x>(u'-u)-2-x .

If j>jβ€²j>j', then no integer lies in [u,uβ€²][u,u'], which forces uβ€²βˆ’u<1u'-u<1 (otherwise the integer ⌊uβŒ‹+1\lfloor u\rfloor+1 would satisfy u<⌊uβŒ‹+1≀u+1≀uβ€²u<\lfloor u\rfloor+1\le u+1\le u'); hence Puβ€²βˆ’Puβ‰₯0>(uβ€²βˆ’u)βˆ’1β‰₯(uβ€²βˆ’u)βˆ’2βˆ’x\mathsf{P}_{u'}-\mathsf{P}_{u}\ge0>(u'-u)-1\ge(u'-u)-2-x.

Combining the two bounds, ∣Puβ€²βˆ’Puβˆ’(uβ€²βˆ’u)βˆ£β‰€x+2|\mathsf{P}_{u'}-\mathsf{P}_{u}-(u'-u)|\le x+2 for every window with 0≀u≀u′≀n0\le u\le u'\le n and uβ€²βˆ’u≀mu'-u\le m, as claimed.

Please log in to copy this version.

Citations

Loading…

Dependency Graph

0 prerequisites

Prerequisites

Loading...

Comments

Loading…