Preliminaries. Since A is compact it is closed by Compact Subset of Rn is Closed and bounded by Compact Subset of Rn is Bounded, so there are x∈Rm and a real R′≥0 with dE(x,a)≤R′ for every a∈A, where dE is the Euclidean distance. Putting R=∣x∣+R′ and using the triangle inequality for the Euclidean norm (Elementary Properties of the Euclidean Norm on Rn) gives ∣a∣≤∣a−x∣+∣x∣=dE(x,a)+∣x∣≤R for every a∈A. Hence The Set of Controls with Values in a Closed Bounded Set is Nonempty, Bounded, Convex and Closed applies to A and R: UA is nonempty, bounded in norm by RT1/2, convex and closed. By claim 1 of Boundedness and Uniform Continuity of Population Cost Data on a Compact Control Set there is a real C≥0 with ∣L(Σ,α)∣≤C for all Σ∈Δl and α∈A. Next, every element of UA has a representative all of whose values lie in A; such representatives are called normalized below. Indeed, let ξ∈UA. By the definition of the set of A-valued controls there are a representative u of ξ and a set N∈B with λ(N)=0 such that u(t)∈A for every t∈[0,T]∖N. Since A is nonempty, fix a0∈A and define u′(t)=u(t) for t∈[0,T]∖N and u′(t)=a0 for t∈N. Then u′(t)∈A for every t∈[0,T]. Each component of u′ is measurable: for a Borel subset B of R, the preimage of B under the i-th component of u′ is ((ui)−1(B)∖N)∪NB, where NB=N if the i-th component of a0 lies in B and NB=∅ otherwise, and both parts lie in B. Since every value of u′ lies in A we have ∣u′(t)∣2≤R2 for every t, and λ([0,T])=T<∞, so u′ is square-integrable by claim 1 of Linearity and Monotonicity of the Lebesgue Integral applied to the nonnegative functions ∣u′∣2 and the constant R2. Finally u′ agrees with u off the null set N, so u′ represents the same element ξ. Finally, λ([0,T])=T by the interval toolkit, and modifying an integrand on a set of λ-measure zero changes neither its measurability nor its integral, again by that lemma.
Claim 1. Let S be an admissible state path and let ξ∈UA. By the Preliminaries, ξ has a normalized representative, that is, a representative u with u(t)∈A for every t∈[0,T]; this is the existence assertion of claim 1. Let u be such a representative. Put
E={(Σ,α)∈Rl+m:Σ∈Δl,α∈A},
a nonempty subset of Rl+m, and let g:E→R be given by g(Σ,α)=L(Σ,α). The map g is sequentially continuous on E: if (Σn,αn)∈E and (Σ,α)∈E with the Euclidean distancesd((Σn,αn),(Σ,α)) converging to 0, then, since the squared distance is the sum of the squared distances of the two blocks of coordinates, d(Σn,Σ) and d(αn,α) converge to 0 as well, so claim 1 of Population Cost Data gives L(Σn,αn)→L(Σ,α).
If u′ is another normalized representative of ξ, then u=u′ off a set of measure zero, so the two integrands agree off a set of measure zero and the two integrals coincide. Hence ΦS(ξ) is well defined.
with C as in the preliminaries. This bound does not depend on S or on ξ, so for sequences (Sn) of admissible state paths and (ξn) in UA one has ∣ΦSn(ξn)∣≤CT<CT+1 for every n, with CT+1>0; hence the real sequence (ΦSn(ξn))n is a bounded sequence, and the limit inferiors appearing in claims 3 and 4 are defined.
Claim 2, convexity. Let ξ,ξ′∈CM and let s be real with 0≤s≤1; let u,u′ be normalized representatives. For every t∈[0,T] the values u(t) and u′(t) lie in A, which is convex, so su(t)+(1−s)u′(t)∈A. The map sending a pair (a,b)∈R2m to sa+(1−s)b is sequentially continuous, and t↦(u(t),u′(t)) is componentwise measurable, so su+(1−s)u′ is measurable by Sequentially Continuous Functions of Measurable Euclidean Maps are Measurable; all its values lie in A, hence ∣su(t)+(1−s)u′(t)∣2≤R2 for every t, and it is square-integrable by claim 1 of Linearity and Monotonicity of the Lebesgue Integral as above. It represents sξ+(1−s)ξ′, which lies in UA because UA is convex by the Preliminaries. Hence su+(1−s)u′ is a normalized representative of sξ+(1−s)ξ′. By the convexity of L in the control, for every t
By claim 2 of Completeness of the Lebesgue Space of Square-Integrable Vector-Valued Functions there are natural numbers n1<n2<… and a set N∈B with λ(N)=0 such that (unj(t))j converges to u(t) in Rm for every t∈[0,T]∖N. Define u~j(t)=unj(t) for t∈/N and u~j(t)=u(t) for t∈N; each u~j is a normalized representative of [unj] (it differs from unj only on N, and all its values lie in A), and now u~j(t)→u(t) for everyt∈[0,T].
By the sequential continuity established in claim 1, L(St,u~j(t))→L(St,u(t)) for every t∈[0,T]; all these functions are measurable by claim 1 and satisfy L(St,u~j(t))≤C, and the constant function C is integrable on ([0,T],B,λ) because λ([0,T])=T<∞. By Dominated Convergence Theorem,
Each term satisfies ΦS([unj])≤M, so ΦS(ζ)≤M by Order Properties of Limits of Real Sequences, comparing with the constant sequence M. Hence ζ∈CM. So the closure of CM is contained in CM; as it also contains CM and is closed, CM is closed.
This holds for every real ε>0, so ΦS(ξ)≤ℓ: otherwise ε=(ΦS(ξ)−ℓ)/2 is positive and yields ΦS(ξ)≤ℓ+21(ΦS(ξ)−ℓ)<ΦS(ξ), which is false.
Claim 4. Put ℓ′=liminfnΦSn(ξn), which is defined by claim 1. Let ε>0 be real. By claim 2 of Boundedness and Uniform Continuity of Population Cost Data on a Compact Control Set there is a real δ>0 such that ∣L(Σ,α)−L(Σ′,α)∣≤ε whenever Σ,Σ′∈Δl satisfy ∣Σ−Σ′∣≤δ and α∈A. By hypothesis there is N0∈N with ∣Stn−St∣≤δ for every t∈[0,T] and every n≥N0. Fix n≥N0 and a normalized representative un of ξn. Then L(St,un(t))≤L(Stn,un(t))+ε for every t, so by claim 2 of Linearity and Monotonicity of the Lebesgue Integral, which supplies both the monotonicity and the linearity of the integral for signed integrable integrands, together with λ([0,T])=T,
Consider the sequence ζi=ξnN0+i−1 (i∈N), which lies in UA and satisfies ΦS(ζi)<ℓ′+ε(1+T) for every i. For every v∈H the real sequence (⟨ζi,v⟩L2)i is a subsequence of (⟨ξn,v⟩L2)n, the index map i↦nN0+i−1 being strictly increasing, hence has the same limit ⟨ξ,v⟩L2 by the fact that a subsequence of a convergent sequence has the same limit; so ζi⇀ξ. The level form established in claim 3, applied to (ζi)i with the bound ℓ′+ε(1+T), gives ξ∈UA and ΦS(ξ)≤ℓ′+ε(1+T).
This holds for every real ε>0. If ΦS(ξ)>ℓ′, then ε=(ΦS(ξ)−ℓ′)(2(1+T))−1 is positive and yields ΦS(ξ)≤ℓ′+21(ΦS(ξ)−ℓ′)<ΦS(ξ), which is false. Hence ΦS(ξ)≤ℓ′=liminfnΦSn(ξn).