Reason: Proof of the weighted compensated observation-sums lemma via the multiplier identities and interval estimates for the compensated counters, the compensated-martingales lemma, and the counter moment partition technique; includes the N-uniform fourth-moment bound.
Proof
Fix a solution and adopt the notation of the statement. Throughout, Ω0 is the regular event; it has probability 1, so expectations are unchanged when integrands are modified off Ω0, and we use this silently. Let Ξt be the sum of all counters as in part (c) of the multiplier lemma, so that c~t≤Ξt pointwise (c~t omits the nonnegative transition counters) and E[Ξtp]<∞ for every natural p≥1 by that part. For a real δ>0 set
and likewise ωG. Each entry of F is continuous on the compact interval [0,T], hence uniformly continuous by the Heine-Cantor theorem, so ωF(δ)→0 as δ→0; also ωF≤2Fˉ and ωG≤2Gˉ everywhere.
Step 0: part (a). Work at a fixed ω∈Ω0 for the pathwise claims. By condition 3 of the solution definition, each observation counter t↦N~ti,υ and the observation total t↦c~t agree on [0,T] with restrictions of counting paths; in particular all are nondecreasing, right-continuous, integer-valued, and vanish at 0 (so J~0F=J0F=0). As recorded in condition 5, at each jump time τj of the observation total exactly one observation counter jumps, its jump is exactly 1, and its channel is υj. Fix υ and consider t↦Πtυ=∑iN~ti,υ. Every point of increase of Πυ is a point of increase of the observation total (the remaining summands of c~ being nondecreasing), hence one of the τj; and at τj the jump of Πυ equals 1 if υj=υ and 0 otherwise. This is the first claim. Consequently, for 0≤r≤t≤T,
For the compensator identity: by condition 2 of the solution definition, at every ω, A~ti,υ is the Lebesgue integral over [0,t] of s↦1Ω0β~(σsi,υ,Σs). On Ω0, pointwise in s, using the occupation indicators ηsi,σ of the derived notation of the solution definition (each agent occupying exactly one state) and the formula of the aggregate observation drift,
and summing the integrals over i (linearity of the Lebesgue integral) gives ∑iA~ti,υ=N∫[0,t]b~υ(Σs)ds on Ω0. By part (a) of the martingale decomposition theorem, 0≤b~υ(Σs)≤B~; and 1Ω0b~υ(Σs)=N−1∑i=1N1Ω0β~(σsi,υ,Σs) pointwise in (s,ω) (both sides equal N−1 times the displayed identity above when multiplied by 1Ω0, since the identity holds on Ω0, and both sides vanish off Ω0); condition 2 of the solution definition already asserts product-measurability, for the product σ-algebra of the trace Borel σ-algebra on [0,T] and F, of each map (s,ω)↦1Ω0β~(σsi,υ,Σs) (this is exactly what makes A~ti,υ a well-defined Lebesgue integral for every t and a random variable, per that condition); summing over i and dividing by N, (s,ω)↦1Ω0b~υ(Σs) is product-measurable, being a finite sum of product-measurable maps. Hence the compensator increment obeys, at every ω and for every γ,
and combining (0.1) and (0.2) gives the stated increment bound for J~F,γ and the bound ∣JtF,γ∣≤FˉN−1/2c~t (case r=0). Right-continuity on Ω0: the finitely many jump times being isolated, for t∈[0,T) and t′↓t the sum in Jt′F,γ eventually acquires no new terms (c~t′=c~t for t′ below the next jump time, by right-continuity and integrality of the counting path), while the compensator is Lipschitz in t by (0.2).
Measurability and moments. Each c~s is a finite sum of counters, hence Fssys-measurable. For j≥1 define τj on all of Ω as the j-th jump time of the counting path agreeing with c~ (equal to +∞ off {c~T≥j}); then {τj≤s}={c~s≥j}∈Fssys for every s∈[0,T]. By part (iv) of the existence theorem, the observation-event count, the event times, and the channels up to any time t are measurable for the observation filtration, which is contained in Ftsys; hence on {τj≤t} the variables τj and υj are Ftsys-measurable. Writing
JtF,γ=N−1/2j≥1∑1{τj≤t}υ=1∑l~1{υj=υ}Fτjγυ(at most c~T nonzero terms),
fix n≥0: on {c~T=n}∈Ftsys the sum has at most n nonzero terms, indexed by j≤n, with {τj≤t}∩{c~T=n}∈Ftsys and {υj=υ}∩{τj≤t}∩{c~T=n}∈Ftsys (both τj and υj being Ftsys-measurable on {τj≤t}, as just recorded), so JtF,γ1{c~T=n} is, on {c~T=n}, a finite sum of products of the Ftsys-measurable indicators 1{τj≤t}, 1{υj=υ} (j≤n) with Fτjγυ (measurable by composition of the continuous entry with the measurable min(τj,T)), hence Ftsys-measurable; since the events {c~T=n} (n≥0) partition Ω, JtF,γ=∑n≥01{c~T=n}JtF,γ is Ftsys-measurable, this sum having, at each ω, exactly one nonzero summand.
For the compensator: the preceding paragraph shows (s,ω)↦1Ω0b~υ(Σs) is product-measurable and bounded by B~; hence (s,ω)↦∑υFsγυ1Ω0b~υ(Σs) is product-measurable, being a finite sum of products of this map with the continuous (hence Borel-measurable) maps s↦Fsγυ, and bounded by Fˉl~B~; by Tonelli's theorem applied on [0,T]×[0,T]×Ω to its positive and negative parts, the map (t,ω)↦N1/2∫[0,t]∑υFsγυ1Ω0b~υ(Σs)ds=N1/2∫[0,T]1{s≤t}∑υFsγυ1Ω0b~υ(Σs)ds (the integrand (s,t,ω)↦1{s≤t}∑υFsγυ1Ω0b~υ(Σs) being product-measurable, as a product of the Borel-measurable 1{s≤t} with the map just shown) is product-measurable in (t,ω), and in particular, for each fixed t, Ftsys-measurable in ω (the states at times ≤t being adapted by part (iv) of the existence theorem, so the restriction of the integrand to s≤t only involves Ftsys-measurable data). Hence each J~tF,γ is a random variable, Ftsys-measurable, and (t,ω)↦1Ω0(ω)J~tF,γ(ω) is product-measurable, being the difference of the product-measurable compensator just constructed and the map (t,ω)↦1Ω0(ω)JtF,γ(ω), itself product-measurable: for each j, the set {(t,ω):τj(ω)≤t} is the preimage of the closed half-plane {(t,x)∈[0,T]×R:x≤t} (closed, hence Borel, in the product) under the map (t,ω)↦(t,τj(ω)), product-measurable because (t,ω)↦t and (t,ω)↦τj(ω) are each measurable (composition with the coordinate projections); so (t,ω)↦1{τj≤t} is product-measurable, hence so is (t,ω)↦1{τj≤t}1{υj=υ} (the second factor not depending on t), and summing the finitely many nonzero terms on each {c~T=n}×[0,T]×Ω as above shows (t,ω)↦JtF,γ(ω), and a fortiori 1Ω0JtF,γ, is product-measurable. Finally, at every ω,
and for each natural p, (x+y)p≤2p(xp+yp) for x,y≥0, so E[∣J~tF,γ∣p]<∞ by the moments of ΞT. This proves (a).
Step 1: partitions and the increment decomposition. Fix 0≤r<t≤T, a natural n≥1, and set δ=(t−r)/n, tk=r+kδ (0≤k≤n), Ik=(tk,tk+1]. Write ΔkX=Xtk+1−Xtk for any process X, Δkc~=c~tk+1−c~tk, and define
On Ω0: by the jump correspondence of Step 0, ΔkJF,γ=N−1/2∑j:τj∈IkFτjγυj, while N−1/2∑υFtkγυ∑iΔkN~i,υ=N−1/2∑j:τj∈IkFtkγυj, and each ∣τj−tk∣≤δ for τj∈Ik; by the compensator identity of Step 0, N−1/2∑υFtkγυ∑iΔkA~i,υ=N1/2∑υFtkγυ∫Ikb~υ(Σs)ds. Hence, with λk=N−1/2Δkc~+N1/2l~B~δ,
on Ω0, using ∑k(Δkc~)q≤(∑kΔkc~)q−1maxkΔkc~≤c~Tq for the first part and nδq=(t−r)δq−1 for the second. All variables λk, c~T, and ΞT have moments of every order (Step 0).
Step 2: part (b). Let Z be Frsys-measurable and square-integrable. Integrability of Z(J~tF,γ−J~rF,γ) holds by the Cauchy-Schwarz inequality for the mean-square norm and (a). Telescoping and using the decomposition of Step 1,
For each k, Z is Ftksys-measurable (the filtration is increasing) and square-integrable, so E[ZΔkM(i,υ)]=0 for every observation clock label by part (a) of the multiplier lemma; by linearity E[ZSkF,γ]=0. By (1.1) and the Cauchy-Schwarz inequality,
which is finite and tends to 0 as n→∞ since ωF(δ)→0. The left-hand side does not depend on n, so it vanishes. Taking Z=1D with D∈Frsys yields the averaged martingale property of the definition of a square-integrable martingale; adaptedness and square-integrability hold by (a), and J~0F=0 on Ω0, an event of probability 1. This proves (b).
Step 3: part (c). Let Z be Frsys-measurable with Z2 square-integrable; then Z is square-integrable (E[Z2]≤1+E[Z4]). Set Xk=J~tkF,γ−J~rF,γ and Yk=J~tkG,δ−J~rG,δ, so X0=Y0=0 and Xn, Yn are the full increments. All products handled below are integrable: for instance E[∣ZXY∣]≤∥ZX∥2∥Y∥2 and E[(ZX)2]=E[Z2X2]≤∥Z2∥2∥X2∥2, finite by hypothesis and (a), with the Cauchy-Schwarz inequality. Telescoping,
Multiply by Z and take expectations termwise. The multipliers ZXk and ZYk are Ftksys-measurable (by (a)) and square-integrable (as just computed), so E[ZXkΔkY]=0 and E[ZYkΔkX]=0 by part (b) applied on [tk,tk+1]. For the quadratic term, write ΔkX=SkF,γ+ρkF,γ and ΔkY=SkG,δ+ρkG,δ. First,
by part (b) of the multiplier lemma, whose hypotheses hold: Z, ZMtk(i,υ), and ZMtk(i′,υ′) are square-integrable, since E[(ZMtka)2]≤∥Z2∥2∥(Mtka)2∥2<∞ by part (c) of that lemma. By the compensator identity of Step 0 (on Ω0),
Summing over k and using (1.2) with q=2 together with the Cauchy-Schwarz inequality (E[∣Z∣c~T2]≤∥Z∥2∥c~T2∥2<∞), the total error from both sources is at most
since ωF(δ)+ωG(δ)→0 and the bracket stays bounded. The main terms add up, by additivity of the Lebesgue integral over adjacent intervals and linearity of the expectation, to E[Z∫[r,t]∑υFsγυGsδυb~υ(Σs)ds]. As the left-hand side E[ZXnYn] does not depend on n, the identity of (c) follows. The recorded special case is the case Z=1, r=0, rewritten with the diagonal matrix D(Σs) (the (γ,δ) entry of FsD(Σs)Gs⊤ being exactly ∑υFsγυGsδυb~υ(Σs) by the formulas for matrix products and the transpose), and the second-moment bound follows from 0≤b~υ≤B~ and monotonicity of the integral.
Step 4: part (d). Let Z be as in (c). Recall from part (b) of the martingale decomposition that Mtγ=Σtγ−Σ0γ−∫[0,t]bγ(Σs,αs)ds, so that ∣Mtγ−Msγ∣≤N1ΔΞ(s,t]+2(l−1)B(t−s) at every ω∈Ω0 and all s≤t, where ΔΞ(s,t]=Ξt−Ξs: indeed each state-transition event changes each Σγ by at most 1/N (condition 6 of the solution definition), and ∣bγ∣≤2(l−1)B by part (a) of the decomposition theorem. In particular Mγ is bounded by 2+2(l−1)BT on Ω0. By part (c) of the counter moment lemma, almost surely NMtγ=∑σ:σ=γ(Mtσγ−Mtγσ) for all t, where Mtσγ=∑iMti,σγ is the aggregate compensated counter over the transition clock labels; hence, for any square-integrable Fssys-measurable W and t≥s,
by part (a) of the multiplier lemma. Now telescope with Xk=Mtkγ−Mrγ and Yk=J~tkF,δ−J~rF,δ as in Step 3. The terms E[ZXkΔkY] vanish by part (b) with the square-integrable multiplier ZXk (Xk bounded, and Ftksys-measurable by adaptedness of the martingale Mγ); the terms E[ZYkΔkX] vanish by (4.1) with W=ZYk (square-integrable as in Step 3, Ftksys-measurable by (a)). For the quadratic terms, write ΔkY=SkF,δ+ρkF,δ and expand ΔkX by the almost sure representation above:
since every pair consists of a transition clock label and an observation clock label, which are distinct, so each expectation vanishes by part (b) of the multiplier lemma (hypotheses verified as in Step 3). Finally, by the pathwise bounds and (1.1),
and summing over k, using Δkc~≤ΔkΞ, ∑k(ΔkΞ)2≤ΞT2, ∑kΔkΞ≤ΞT, nδ2=(t−r)δ, and the Cauchy-Schwarz inequality, the total is at most ωF(δ) times a finite constant (depending on N, T, B, B~, l~, l, E[Z2], and the moments of ΞT, but not on n), which tends to 0. Since E[ZXnYn] does not depend on n, part (d) follows.
Step 5: part (e). Fix γ and abbreviate Xs=J~sF,γ. By (c), m2=sup{E[Xs2]:s∈[0,T]}≤Fˉ2l~B~T. Fix t∈(0,T] (the case t=0 is trivial) and partition [0,t] as in Step 1 with r=0. Telescoping fourth powers and expanding by the binomial theorem,
all terms integrable by (a). We bound the four groups; write O for the set of the Nl~ observation clock labels and fa=Ftkγυ for a=(i,υ)∈O, and let C⋆ be the constant of the multiplier lemma.
Cubic multiplier.E[Xtk3ΔkX]=0 by part (b) with the square-integrable Ftksys-measurable multiplier Xtk3.
Quadratic multiplier. By part (c) with Z=Xtk2 (its square Xtk4 is square-integrable by (a)) on [tk,tk+1], and 0≤b~υ≤B~:
Linear multiplier. Write ΔkX=Sk+ρk with Sk=SkF,γ, ρk=ρkF,γ. Then (ΔkX)3=Sk3+(3Sk2ρk+3Skρk2+ρk3), and by (1.1) with ωF≤2Fˉ, ∣3Sk2ρk+3Skρk2+ρk3∣≤7(3Fˉ)2ωF(δ)λk3. Using (1.2) with q=3 and the Cauchy-Schwarz inequality, the total contribution of these remainder terms over all k is at most ωF(δ) times a finite constant independent of n, hence tends to 0. Expand Sk3=N−3/2∑a,b,c∈OfafbfcΔkMaΔkMbΔkMc. For the triples not all equal, part (f) of the multiplier lemma with the integrable multiplier Xtk bounds each term by C⋆E[∣Xtk∣]δ2≤C⋆m21/2δ2; there are fewer than (Nl~)3 of them, so with the prefactor N−3/2Fˉ3 their total over all k is at most N3/2l~3Fˉ3C⋆m21/2tδ→0. For the Nl~ diagonal triples, split E[Xtk(ΔkMa)3]=E[Xtk((ΔkMa)3−ΔkAa)]+E[XtkΔkAa] (writing Aa=A~i,υ for a=(i,υ)): by part (e) of the multiplier lemma with the square-integrable multiplier Xtk, the first expectation is at most C⋆(1+m2)δ3/2 in absolute value, so these parts total at most N−1/2l~Fˉ3C⋆(1+m2)tδ1/2→0; for the second parts, ∑a∈OΔkAa≤Nl~B~δ pathwise on Ω0 (compensator identity and b~υ≤B~), so
using N≥1 and E[∣Xtk∣]≤m21/2 (Cauchy-Schwarz against the constant 1).
Constant multiplier.(ΔkX)4=Sk4+((ΔkX)4−Sk4), and ∣(ΔkX)4−Sk4∣≤15(3Fˉ)3ωF(δ)λk4 by (1.1); by (1.2) with q=4 the total of these remainders tends to 0 as before. Expand Sk4=N−2∑a,b,c,d∈OfafbfcfdΔkMaΔkMbΔkMcΔkMd. Quadruples not all equal: part (f) with Z=1 bounds each by C⋆δ2; fewer than (Nl~)4 of them, prefactor N−2Fˉ4, total over k at most N2l~4Fˉ4C⋆tδ→0. Diagonal quadruples: by part (d) of the multiplier lemma with Z=1 (bounded by ζ=1) and k=4, E[(ΔkMa)4]≤E[ΔkAa]+C⋆δ2, so their total per interval is at most N−2Fˉ4(Nl~B~δ+Nl~C⋆δ2)≤Fˉ4l~B~δ+Fˉ4l~C⋆δ2 (N≥1), and over all k at most Fˉ4l~B~t+Fˉ4l~C⋆tδ.
Summation. Summing the surviving bounds over k (each carries a factor δ and there are n=t/δ intervals) and letting n→∞, the vanishing groups disappear and
E[Xt4]≤6Fˉ2l~B~tm2+4Fˉ3l~B~tm21/2+Fˉ4l~B~t.
With y=l~B~T≥l~B~t and m2≤Fˉ2y, m21/2≤Fˉy1/2, the right-hand side is at most Fˉ4(6y2+4y3/2+y). If y≥1 this is at most 11Fˉ4y2; if y<1 it is at most 11Fˉ4y; in either case at most 11Fˉ4(1+y)2≤11(1+Fˉ)4(1+l~B~T)2. The constants involved depend only on Fˉ, l~, B~, and T, proving (e). ■