TheoremBase

Maximising couplings exist by weak-star compactness of bounded laws, gluing along the last marginal gives a Lipschitz estimate for the test function and a first-order expansion yielding the plan subjet, and the localisation hypothesis forces the test function and the displacements of maximising couplings to vanish.

Proof

Each result cited is universally quantified over the data in its own statement.

Notation. For γ∈Σ3d2\gamma\in\Sigma^{2}_{3d} put Ψ(γ)=p(γ)−h(s(γ))−θ s(γ)2\Psi(\gamma)=p(\gamma)-h(s(\gamma))-\theta\,s(\gamma)^{2}, where s(γ)s(\gamma) and p(γ)p(\gamma) are the displacement and momentum pairing of Couplings of a Square-Integrable Plan with a Law: Displacement and Momentum Pairing §displacement-pairing; by that clause a realisation of γ\gamma exists and s(γ)s(\gamma), p(γ)p(\gamma) may be computed from any realisation. Put aπ=∣π∣moma_{\pi}=|\pi|_{\mathrm{mom}} and K=aπ+L+4θrdK=a_{\pi}+L+4\theta r\sqrt{d}, a real number with K>0K>0. By The Coupling Test Function of a Square-Integrable Plan §test, Ψ(γ)≤Φ(λ)\Psi(\gamma)\le\Phi(\lambda) whenever λ∈Σd2\lambda\in\Sigma^{2}_{d} and γ∈C(λ)\gamma\in\mathcal{C}(\lambda); and since Φ(λ)\Phi(\lambda) is a least upper bound in the sense of The Real Numbers: Standing Notation and Background §bounds, for every real ε>0\varepsilon>0 some γ∈C(λ)\gamma\in\mathcal{C}(\lambda) has Ψ(γ)>Φ(λ)−ε\Psi(\gamma)>\Phi(\lambda)-\varepsilon. For L2L^{2} dd-tuples of one tracial W*-probability space (H,M,Ω)(H,M,\Omega), the pairing is the inner product and ∥⋅∥2\lVert\cdot\rVert_{2} the norm of the complex Hilbert space HdH^{d}, sums and real multiples are those of HdH^{d}, and the pairing is real, by Sums, Real Multiples and the Pairing of Square-Integrable Tuples in a Tracial W*-Probability Space §operations and Sums, Real Multiples and the Pairing of Square-Integrable Tuples in a Tracial W*-Probability Space §pairing. Hence, by the results in force by Plan Jets and Hamiltonians on Square-Integrable Noncommutative Laws: Standing Notation §background, for all L2L^{2} dd-tuples U1,U2,U3U_{1},U_{2},U_{3} of one space and real tt we have ∣⟨U1,U2⟩2∣≤∥U1∥2∥U2∥2|\langle U_{1},U_{2}\rangle_{2}|\le\lVert U_{1}\rVert_{2}\lVert U_{2}\rVert_{2}, ∥U1+U2∥2≤∥U1∥2+∥U2∥2\lVert U_{1}+U_{2}\rVert_{2}\le\lVert U_{1}\rVert_{2}+\lVert U_{2}\rVert_{2}, ∥tU1∥2=∣t∣ ∥U1∥2\lVert tU_{1}\rVert_{2}=|t|\,\lVert U_{1}\rVert_{2}, ⟨U1,U2+tU3⟩2=⟨U1,U2⟩2+t⟨U1,U3⟩2\langle U_{1},U_{2}+tU_{3}\rangle_{2}=\langle U_{1},U_{2}\rangle_{2}+t\langle U_{1},U_{3}\rangle_{2}, ∥U1+U2∥22=∥U1∥22+2⟨U1,U2⟩2+∥U2∥22\lVert U_{1}+U_{2}\rVert_{2}^{2}=\lVert U_{1}\rVert_{2}^{2}+2\langle U_{1},U_{2}\rangle_{2}+\lVert U_{2}\rVert_{2}^{2}, and U1=0U_{1}=0 whenever ∥U1∥2=0\lVert U_{1}\rVert_{2}=0. These rules are used below without further mention.

Step A (the gauge). First, h′(s)≥0h'(s)\ge0 for every real s≥0s\ge0. Indeed, fix a real η>0\eta>0, let rηr_{\eta} be as in the hypothesis at the point ss, and put t=s+rη/2t=s+r_{\eta}/2. Since hh is nondecreasing in the sense of Monotone Real Function §nondecreasing, 0≤h(t)−h(s)≤h′(s)(t−s)+η(t−s)0\le h(t)-h(s)\le h'(s)(t-s)+\eta(t-s), and dividing by t−s>0t-s>0 gives h′(s)≥−ηh'(s)\ge-\eta. As η>0\eta>0 was arbitrary, h′(s)≥0h'(s)\ge0.

Second, ∣h(t)−h(s)∣≤L∣t−s∣|h(t)-h(s)|\le L|t-s| for all real s,t≥0s,t\ge0. The hypothesis gives, for every real c≥0c\ge0 and η>0\eta>0, a real rη(c)>0r_{\eta}(c)>0 such that

∣h(t)−h(c)∣≤(∣h′(c)∣+η)∣t−c∣≤(L+η)∣t−c∣for every real t≥0 with ∣t−c∣<rη(c).|h(t)-h(c)|\le(|h'(c)|+\eta)|t-c|\le(L+\eta)|t-c|\qquad\text{for every real }t\ge0\text{ with }|t-c|<r_{\eta}(c).

Fix reals 0≤a<b0\le a<b and η>0\eta>0, and let AA be the set of real tt with a≤t≤ba\le t\le b and h(t)−h(a)≤(L+η)(t−a)h(t)-h(a)\le(L+\eta)(t-a). Then a∈Aa\in A and bb is an upper bound of AA, so AA has a least upper bound cc with a≤c≤ba\le c\le b, by The Real Numbers: Standing Notation and Background §bounds. We show c∈Ac\in A. If c=ac=a this is clear. Otherwise, by the least-upper-bound property there is t∈At\in A with c−rη(c)<t≤cc-r_{\eta}(c)<t\le c; the displayed estimate at cc gives h(c)−h(t)≤(L+η)(c−t)h(c)-h(t)\le(L+\eta)(c-t), and adding h(t)−h(a)≤(L+η)(t−a)h(t)-h(a)\le(L+\eta)(t-a) gives c∈Ac\in A. Next, c=bc=b: otherwise t1=min⁡(b,c+rη(c)/2)t_{1}=\min(b,c+r_{\eta}(c)/2) satisfies c<t1≤bc<t_{1}\le b, the displayed estimate at cc gives h(t1)−h(c)≤(L+η)(t1−c)h(t_{1})-h(c)\le(L+\eta)(t_{1}-c), and adding h(c)−h(a)≤(L+η)(c−a)h(c)-h(a)\le(L+\eta)(c-a) gives t1∈At_{1}\in A, contradicting t1>ct_{1}>c. Hence b∈Ab\in A, that is, 0≤h(b)−h(a)≤(L+η)(b−a)0\le h(b)-h(a)\le(L+\eta)(b-a), the first inequality because hh is nondecreasing. As η>0\eta>0 was arbitrary, 0≤h(b)−h(a)≤L(b−a)0\le h(b)-h(a)\le L(b-a), which gives the claim.

Step B (bounds on couplings). Let ν∈Σd,r\nu\in\Sigma_{d,r}, let γ∈C(κd(ν))\gamma\in\mathcal{C}(\kappa_{d}(\nu)) and let (X,P,X′)(X,P,X') be a realisation of γ\gamma. We show

law(X,P)=κ2d(π),law(X)=κd(μ),∥P∥2=aπ,∥X∥2≤rd,∥X′∥2≤rd,\mathrm{law}(X,P)=\kappa_{2d}(\pi),\qquad\mathrm{law}(X)=\kappa_{d}(\mu),\qquad\lVert P\rVert_{2}=a_{\pi},\qquad\lVert X\rVert_{2}\le r\sqrt{d},\qquad\lVert X'\rVert_{2}\le r\sqrt{d},

and consequently s(γ)≤2rds(\gamma)\le2r\sqrt{d} and p(γ)≤aπ s(γ)p(\gamma)\le a_{\pi}\,s(\gamma). Since B(X,P,X′)=(X,P)B(X,P,X')=(X,P) and C(X,P,X′)=X′C(X,P,X')=X', Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §push-forward and Couplings of a Square-Integrable Plan with a Law: Displacement and Momentum Pairing §couplings give law(X,P)=B#γ=κ2d(π)\mathrm{law}(X,P)=B_{\#}\gamma=\kappa_{2d}(\pi) and law(X′)=C#γ=κd(ν)\mathrm{law}(X')=C_{\#}\gamma=\kappa_{d}(\nu). By Every Square-Integrable Noncommutative Law is the Law of a Square-Integrable Tuple; Realisation of Couplings and of Almost Optimal Pairs §coupling, law(X)=pr#1κ2d(π)\mathrm{law}(X)=\mathrm{pr}^{1}_{\#}\kappa_{2d}(\pi) and law(P)=pr#2κ2d(π)\mathrm{law}(P)=\mathrm{pr}^{2}_{\#}\kappa_{2d}(\pi). By Calculus of Square-Integrable Noncommutative Laws: Agreement on Bounded Laws, Lipschitz Estimates, Functoriality of Push-Forwards, Moment Formulas, Positivity, the Cost and the Diagonal Coupling §bounded and Affine Substitutions of Noncommutative Laws: Self-Adjointness, Composition, Moment Formulas and Positivity, and the Coordinate Data §coordinate, these equal κd(π∘ι1)\kappa_{d}(\pi\circ\iota^{1}) and κd(π∘ι2)\kappa_{d}(\pi\circ\iota^{2}) (here π∈Σ2d\pi\in\Sigma_{2d} by Marginal Isometries, Bounded Plans and Displacement Pairings for Noncommutative Laws §plans, so π∘ι1,π∘ι2∈Σd\pi\circ\iota^{1},\pi\circ\iota^{2}\in\Sigma_{d} by the preamble of Marginal Isometries, Bounded Plans and Displacement Pairings for Noncommutative Laws), and π∘ι1=μ\pi\circ\iota^{1}=\mu because π\pi is a bounded plan at μ\mu in the sense of Marginal Isometries, Bounded Plans and Displacement Pairings for Noncommutative Laws §plans. By Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §moments and Calculus of Square-Integrable Noncommutative Laws: Agreement on Bounded Laws, Lipschitz Estimates, Functoriality of Push-Forwards, Moment Formulas, Positivity, the Cost and the Diagonal Coupling §bounded, ∥X∥22=M(μ)\lVert X\rVert_{2}^{2}=M(\mu), ∥X′∥22=M(ν)\lVert X'\rVert_{2}^{2}=M(\nu) and ∥P∥22=M(π∘ι2)\lVert P\rVert_{2}^{2}=M(\pi\circ\iota^{2}). Since μ,ν∈Σd,r\mu,\nu\in\Sigma_{d,r}, each of the real numbers μ(xjxj)\mu(x_{j}x_{j}) and ν(xjxj)\nu(x_{j}x_{j}) has modulus at most r2r^{2} by Noncommutative Laws of Finitely Many Self-Adjoint Variables with a Norm Bound §norm-bound (words of length 22), so M(μ)≤dr2M(\mu)\le dr^{2} and M(ν)≤dr2M(\nu)\le dr^{2}. Since ι2\iota^{2} substitutes xd+jx_{d+j} for xjx_{j} by Couplings of Two Noncommutative Laws and Their Quadratic Cost §marginals and is multiplicative by Substitution is the Unique Unital Homomorphism with Prescribed Values on the Variables: Monomials, Products, Adjoints and Composition §homomorphism, M(π∘ι2)=∑j=1dπ(xd+jxd+j)=aπ2M(\pi\circ\iota^{2})=\sum_{j=1}^{d}\pi(x_{d+j}x_{d+j})=a_{\pi}^{2} by Marginal Isometries, Bounded Plans and Displacement Pairings for Noncommutative Laws §plans. Finally s(γ)=∥X′−X∥2≤∥X′∥2+∥X∥2≤2rds(\gamma)=\lVert X'-X\rVert_{2}\le\lVert X'\rVert_{2}+\lVert X\rVert_{2}\le2r\sqrt{d} and p(γ)=⟨P,X′−X⟩2≤∥P∥2 s(γ)=aπ s(γ)p(\gamma)=\langle P,X'-X\rangle_{2}\le\lVert P\rVert_{2}\,s(\gamma)=a_{\pi}\,s(\gamma).

Step C (law invariance). Let k,n∈Nk,n\in\mathbb{N}, let ZZ and Z1Z_{1} be L2L^{2} kk-tuples of tracial W*-probability spaces (possibly different) with law(Z)=law(Z1)\mathrm{law}(Z)=\mathrm{law}(Z_{1}), and let T=(E,0)T=(E,0) be an affine datum from kk to nn variables. By Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §push-forward, law(TZ)=T#law(Z)=T#law(Z1)=law(TZ1)\mathrm{law}(TZ)=T_{\#}\mathrm{law}(Z)=T_{\#}\mathrm{law}(Z_{1})=\mathrm{law}(TZ_{1}). By Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §moments, ∥TZ∥22\lVert TZ\rVert_{2}^{2} and ∥TZ1∥22\lVert TZ_{1}\rVert_{2}^{2} both equal the second moment of this law, so ∥TZ∥2=∥TZ1∥2\lVert TZ\rVert_{2}=\lVert TZ_{1}\rVert_{2}; and if n=2dn=2d and TZ=(A,A′)TZ=(A,A'), TZ1=(A1,A1′)TZ_{1}=(A_{1},A_{1}') with L2L^{2} dd-tuples A,A′,A1,A1′A,A',A_{1},A_{1}', then by the same clause and Sums, Real Multiples and the Pairing of Square-Integrable Tuples in a Tracial W*-Probability Space §pairing both ⟨A,A′⟩2\langle A,A'\rangle_{2} and ⟨A1,A1′⟩2\langle A_{1},A_{1}'\rangle_{2} equal ∑i=1dmi,d+i(law(TZ))\sum_{i=1}^{d}\mathrm{m}_{i,d+i}(\mathrm{law}(TZ)), so they are equal. By Square-Integrable Tuples in a Tracial W*-Probability Space: Their Norm, Affine Images, Pairs, Embedded Images and Laws §operations, every tuple formed below from the coordinates of a tuple ZZ by selecting or reordering blocks and taking real linear combinations of blocks is TZTZ for the affine datum T=(E,0)T=(E,0) whose coefficients are read off from the formula; pairs and triples are tuples by the same clause. We use this without writing out the matrices EE.

Step D (gluing). Let k,m,n∈Nk,m,n\in\mathbb{N}, let Z,WZ,W be L2L^{2} tuples of lengths k,mk,m of one tracial W*-probability space and Z1,VZ_{1},V L2L^{2} tuples of lengths k,nk,n of another, with law(Z)=law(Z1)\mathrm{law}(Z)=\mathrm{law}(Z_{1}). With Fm,FnF^{m},F^{n} as in Gluing Two Square-Integrable Noncommutative Laws along a Common Marginal, Fm(Z,W)=ZF^{m}(Z,W)=Z and Fn(Z1,V)=Z1F^{n}(Z_{1},V)=Z_{1}, so F#mlaw(Z,W)=law(Z)=law(Z1)=F#nlaw(Z1,V)F^{m}_{\#}\mathrm{law}(Z,W)=\mathrm{law}(Z)=\mathrm{law}(Z_{1})=F^{n}_{\#}\mathrm{law}(Z_{1},V) by Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §push-forward. By Gluing Two Square-Integrable Noncommutative Laws along a Common Marginal §glue there are a tracial W*-probability space and L2L^{2} tuples Z^,W^,V^\hat{Z},\hat{W},\hat{V} of it, of lengths k,m,nk,m,n, with law(Z^,W^)=law(Z,W)\mathrm{law}(\hat{Z},\hat{W})=\mathrm{law}(Z,W) and law(Z^,V^)=law(Z1,V)\mathrm{law}(\hat{Z},\hat{V})=\mathrm{law}(Z_{1},V).

Step E (transfer). Let ν,ν′∈Σd,r\nu,\nu'\in\Sigma_{d,r}, put w=W2(ν,ν′)w=W_{2}(\nu,\nu'), and let γ∈C(κd(ν))\gamma\in\mathcal{C}(\kappa_{d}(\nu)). We show that there is γ′′∈C(κd(ν′))\gamma''\in\mathcal{C}(\kappa_{d}(\nu')) with ∣s(γ′′)−s(γ)∣≤w|s(\gamma'')-s(\gamma)|\le w and Ψ(γ′′)≥Ψ(γ)−Kw\Psi(\gamma'')\ge\Psi(\gamma)-Kw. Let (X,P,X′)(X,P,X') be a realisation of γ\gamma. By The Noncommutative Wasserstein Distance: Existence of Optimal Couplings, Symmetry, Separation, a Moment Bound, Weak-Star Lower Semicontinuity, and Displacement Interpolation §attained (with R=rR=r) there is an optimal coupling ϱ∈Π(ν,ν′)\varrho\in\Pi(\nu,\nu'), so ϱ∈Σ2d\varrho\in\Sigma_{2d} by Marginal Isometries, Bounded Plans and Displacement Pairings for Noncommutative Laws §bounded-couplings and I(ϱ)=w2I(\varrho)=w^{2} by The Noncommutative Quadratic Wasserstein Distance and Optimal Couplings §optimal; by Calculus of Square-Integrable Noncommutative Laws: Agreement on Bounded Laws, Lipschitz Estimates, Functoriality of Push-Forwards, Moment Formulas, Positivity, the Cost and the Diagonal Coupling §bounded, κ2d(ϱ)∈Π2(κd(ν),κd(ν′))\kappa_{2d}(\varrho)\in\Pi^{2}(\kappa_{d}(\nu),\kappa_{d}(\nu')) and I(κ2d(ϱ))=w2\mathcal{I}(\kappa_{2d}(\varrho))=w^{2}. By Every Square-Integrable Noncommutative Law is the Law of a Square-Integrable Tuple; Realisation of Couplings and of Almost Optimal Pairs §coupling and Square-Integrable Noncommutative Laws: the Wasserstein Completion of the Laws, Affine Push-Forwards, Moments, Couplings and Cost §couplings there are L2L^{2} dd-tuples Y,Y′′Y,Y'' of a tracial W*-probability space with law(Y)=κd(ν)\mathrm{law}(Y)=\kappa_{d}(\nu), law(Y′′)=κd(ν′)\mathrm{law}(Y'')=\kappa_{d}(\nu') and ∥Y−Y′′∥2=w\lVert Y-Y''\rVert_{2}=w. Since law(X′)=κd(ν)\mathrm{law}(X')=\kappa_{d}(\nu) by Step B, Step D with Z=X′Z=X', W=(X,P)W=(X,P), Z1=YZ_{1}=Y and V=Y′′V=Y'' gives L2L^{2} dd-tuples X^′,X^,P^,Y^′′\hat{X}',\hat{X},\hat{P},\hat{Y}'' of one tracial W*-probability space with law(X^′,(X^,P^))=law(X′,(X,P))\mathrm{law}(\hat{X}',(\hat{X},\hat{P}))=\mathrm{law}(X',(X,P)) and law(X^′,Y^′′)=law(Y,Y′′)\mathrm{law}(\hat{X}',\hat{Y}'')=\mathrm{law}(Y,Y''). By Step C, law(X^,P^,X^′)=law(X,P,X′)=γ\mathrm{law}(\hat{X},\hat{P},\hat{X}')=\mathrm{law}(X,P,X')=\gamma, law(Y^′′)=law(Y′′)=κd(ν′)\mathrm{law}(\hat{Y}'')=\mathrm{law}(Y'')=\kappa_{d}(\nu') and ∥Y^′′−X^′∥2=∥Y′′−Y∥2=w\lVert\hat{Y}''-\hat{X}'\rVert_{2}=\lVert Y''-Y\rVert_{2}=w. Put γ′′=law(X^,P^,Y^′′)\gamma''=\mathrm{law}(\hat{X},\hat{P},\hat{Y}''). By Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §push-forward, B#γ′′=law(X^,P^)=B#γ=κ2d(π)B_{\#}\gamma''=\mathrm{law}(\hat{X},\hat{P})=B_{\#}\gamma=\kappa_{2d}(\pi) and C#γ′′=law(Y^′′)=κd(ν′)C_{\#}\gamma''=\mathrm{law}(\hat{Y}'')=\kappa_{d}(\nu'), so γ′′∈C(κd(ν′))\gamma''\in\mathcal{C}(\kappa_{d}(\nu')) by Couplings of a Square-Integrable Plan with a Law: Displacement and Momentum Pairing §couplings.

Put a^=X^′−X^\hat{a}=\hat{X}'-\hat{X} and v=Y^′′−X^′v=\hat{Y}''-\hat{X}', so that Y^′′−X^=a^+v\hat{Y}''-\hat{X}=\hat{a}+v and ∥v∥2=w\lVert v\rVert_{2}=w. Computing from the realisations (X^,P^,X^′)(\hat{X},\hat{P},\hat{X}') of γ\gamma and (X^,P^,Y^′′)(\hat{X},\hat{P},\hat{Y}'') of γ′′\gamma'', we get s(γ)=∥a^∥2s(\gamma)=\lVert\hat{a}\rVert_{2}, s(γ′′)=∥a^+v∥2s(\gamma'')=\lVert\hat{a}+v\rVert_{2}, hence ∣s(γ′′)−s(γ)∣≤w|s(\gamma'')-s(\gamma)|\le w, and p(γ′′)−p(γ)=⟨P^,v⟩2≥−∥P^∥2 w=−aπwp(\gamma'')-p(\gamma)=\langle\hat{P},v\rangle_{2}\ge-\lVert\hat{P}\rVert_{2}\,w=-a_{\pi}w by Step B. By Step A, h(s(γ′′))−h(s(γ))≤Lwh(s(\gamma''))-h(s(\gamma))\le Lw. By Step B, s(γ)+s(γ′′)≤4rds(\gamma)+s(\gamma'')\le4r\sqrt{d}, so θs(γ′′)2−θs(γ)2=θ(s(γ′′)−s(γ))(s(γ′′)+s(γ))≤4θrd w\theta s(\gamma'')^{2}-\theta s(\gamma)^{2}=\theta(s(\gamma'')-s(\gamma))(s(\gamma'')+s(\gamma))\le4\theta r\sqrt{d}\,w. Adding, Ψ(γ′′)≥Ψ(γ)−(aπ+L+4θrd)w=Ψ(γ)−Kw\Psi(\gamma'')\ge\Psi(\gamma)-(a_{\pi}+L+4\theta r\sqrt{d})w=\Psi(\gamma)-Kw.

Proof of clause 1. Fix ν∈Σd,r\nu\in\Sigma_{d,r}. Since π∈Σ2d\pi\in\Sigma_{2d} by Marginal Isometries, Bounded Plans and Displacement Pairings for Noncommutative Laws §plans, there is a real Rπ>0R_{\pi}>0 with π∈Σ2d,Rπ\pi\in\Sigma_{2d,R_{\pi}} by Noncommutative Laws of Finitely Many Self-Adjoint Variables with a Norm Bound §law; put R′=max⁡(Rπ,r)R'=\max(R_{\pi},r).

(a) Every γ∈C(κd(ν))\gamma\in\mathcal{C}(\kappa_{d}(\nu)) equals κ3d(λ)\kappa_{3d}(\lambda) for some λ∈Σ3d,R′\lambda\in\Sigma_{3d,R'}. Indeed, let (X,P,X′)(X,P,X') be a realisation of γ\gamma in (H,M,Ω)(H,M,\Omega). By Step B, law(X,P)=κ2d(π)\mathrm{law}(X,P)=\kappa_{2d}(\pi) and law(X′)=κd(ν)\mathrm{law}(X')=\kappa_{d}(\nu), so by Square-Integrable Tuples with a Bounded Law are Vacuum Tuples of Bounded Self-Adjoint Operators §operator and Square-Integrable Tuples with a Bounded Law are Vacuum Tuples of Bounded Self-Adjoint Operators §bound (for 2d2d and dd variables) there are a self-adjoint 2d2d-tuple uu and a self-adjoint dd-tuple u′u' in MM with uΩ=(X,P)u\Omega=(X,P), u′Ω=X′u'\Omega=X', and all operator norms of their entries at most R′R'. The concatenation (u,u′)(u,u') is a self-adjoint 3d3d-tuple in MM in the sense of Self-Adjoint Tuples in a Tracial W*-Probability Space and Their Laws §tuple, and its vacuum tuple is (X,P,X′)(X,P,X') by Square-Integrable Tuples in a Tracial W*-Probability Space: Their Norm, Affine Images, Pairs, Embedded Images and Laws §operations. By Laws of Self-Adjoint Tuples in a Tracial W*-Probability Space: Moments, Affine Images, Couplings, Embeddings and L^2 Approximation §law, λ=λ(u,u′)∈Σ3d,R′\lambda=\lambda_{(u,u')}\in\Sigma_{3d,R'}, and by Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §bounded, γ=law(X,P,X′)=κ3d(λ)\gamma=\mathrm{law}(X,P,X')=\kappa_{3d}(\lambda).

(b) Let σB\sigma_{B}, σC\sigma_{C} and σU\sigma_{U} be the affine substitutions of the affine data BB, CC and UU of Couplings of a Square-Integrable Plan with a Law: Displacement and Momentum Pairing. For λ∈Σ3d\lambda\in\Sigma_{3d} we have λ∘σB∈Σ2d\lambda\circ\sigma_{B}\in\Sigma_{2d} and λ∘σC∈Σd\lambda\circ\sigma_{C}\in\Sigma_{d} by Affine Substitutions of Noncommutative Laws: Self-Adjointness, Composition, Moment Formulas and Positivity, and the Coordinate Data §self-adjoint, and κ3d(λ)∈C(κd(ν))\kappa_{3d}(\lambda)\in\mathcal{C}(\kappa_{d}(\nu)) if and only if λ∘σB=π\lambda\circ\sigma_{B}=\pi and λ∘σC=ν\lambda\circ\sigma_{C}=\nu. Indeed B#κ3d(λ)=κ2d(λ∘σB)B_{\#}\kappa_{3d}(\lambda)=\kappa_{2d}(\lambda\circ\sigma_{B}) and C#κ3d(λ)=κd(λ∘σC)C_{\#}\kappa_{3d}(\lambda)=\kappa_{d}(\lambda\circ\sigma_{C}) by Calculus of Square-Integrable Noncommutative Laws: Agreement on Bounded Laws, Lipschitz Estimates, Functoriality of Push-Forwards, Moment Formulas, Positivity, the Cost and the Diagonal Coupling §bounded, and the canonical maps κ2d,κd\kappa_{2d},\kappa_{d} of the completions of Square-Integrable Noncommutative Laws: the Wasserstein Completion of the Laws, Affine Push-Forwards, Moments, Couplings and Cost §laws are injective by The Metric Completion is a Complete Metric Space with a Dense Isometric Copy of the Space, and Maps Preserving Cauchy Sequences Extend to It §isometry; now apply Couplings of a Square-Integrable Plan with a Law: Displacement and Momentum Pairing §couplings.

(c) For λ∈Σ3d\lambda\in\Sigma_{3d} and i,j∈[2d]i,j\in[2d] put cij(λ)=λ(σU(xixj))c_{ij}(\lambda)=\lambda(\sigma_{U}(x_{i}x_{j})). Since λ∘σU\lambda\circ\sigma_{U} is a tracial state by Affine Substitutions of Noncommutative Laws: Self-Adjointness, Composition, Moment Formulas and Positivity, and the Coordinate Data §self-adjoint, cij(λ)=mij(λ∘σU)c_{ij}(\lambda)=\mathrm{m}_{ij}(\lambda\circ\sigma_{U}) is real by Affine Substitutions of Noncommutative Laws: Self-Adjointness, Composition, Moment Formulas and Positivity, and the Coordinate Data §moments, and mij(U#κ3d(λ))=mij(κ2d(λ∘σU))=cij(λ)\mathrm{m}_{ij}(U_{\#}\kappa_{3d}(\lambda))=\mathrm{m}_{ij}(\kappa_{2d}(\lambda\circ\sigma_{U}))=c_{ij}(\lambda) by Calculus of Square-Integrable Noncommutative Laws: Agreement on Bounded Laws, Lipschitz Estimates, Functoriality of Push-Forwards, Moment Formulas, Positivity, the Cost and the Diagonal Coupling §bounded. Hence, by the formulas of Couplings of a Square-Integrable Plan with a Law: Displacement and Momentum Pairing §displacement-pairing,

s(κ3d(λ))2=∑i=1dcd+i,d+i(λ),p(κ3d(λ))=∑i=1dci,d+i(λ).s(\kappa_{3d}(\lambda))^{2}=\sum_{i=1}^{d}c_{d+i,d+i}(\lambda),\qquad p(\kappa_{3d}(\lambda))=\sum_{i=1}^{d}c_{i,d+i}(\lambda).

Now choose, for every n∈Nn\in\mathbb{N}, some γn∈C(κd(ν))\gamma_{n}\in\mathcal{C}(\kappa_{d}(\nu)) with Ψ(γn)>Φ(κd(ν))−2−n\Psi(\gamma_{n})>\Phi(\kappa_{d}(\nu))-2^{-n} (Notation), and by (a) some λn∈Σ3d,R′\lambda_{n}\in\Sigma_{3d,R'} with γn=κ3d(λn)\gamma_{n}=\kappa_{3d}(\lambda_{n}). By Sequential Weak-Star Compactness of the Noncommutative Laws with a Given Norm Bound §compact there are a subsequence (λnk)k∈N(\lambda_{n_{k}})_{k\in\mathbb{N}} in the sense of Subsequence of a Sequence in a Set and λ∈Σ3d,R′\lambda\in\Sigma_{3d,R'} with λnk→λ\lambda_{n_{k}}\to\lambda weak-star; by Weak-Star Convergence of Noncommutative Laws §weak-star, for every q∈P3dq\in\mathcal{P}_{3d} the real and imaginary parts of λnk(q)\lambda_{n_{k}}(q) converge to those of λ(q)\lambda(q). For q′∈P2dq'\in\mathcal{P}_{2d}, (b) gives λnk(σB(q′))=π(q′)\lambda_{n_{k}}(\sigma_{B}(q'))=\pi(q') for every kk, so λ(σB(q′))=π(q′)\lambda(\sigma_{B}(q'))=\pi(q'); thus λ∘σB=π\lambda\circ\sigma_{B}=\pi, and likewise λ∘σC=ν\lambda\circ\sigma_{C}=\nu. By (b), γ∗=κ3d(λ)∈C(κd(ν))\gamma^{*}=\kappa_{3d}(\lambda)\in\mathcal{C}(\kappa_{d}(\nu)). The real numbers cij(λnk)c_{ij}(\lambda_{n_{k}}) converge to cij(λ)c_{ij}(\lambda), so by (c) s(γnk)2→s(γ∗)2s(\gamma_{n_{k}})^{2}\to s(\gamma^{*})^{2} and p(γnk)→p(γ∗)p(\gamma_{n_{k}})\to p(\gamma^{*}). Since ∣x−y∣2≤∣x−y∣(x+y)=∣x2−y2∣|x-y|^{2}\le|x-y|(x+y)=|x^{2}-y^{2}| for reals x,y≥0x,y\ge0, also s(γnk)→s(γ∗)s(\gamma_{n_{k}})\to s(\gamma^{*}), and by Step A ∣h(s(γnk))−h(s(γ∗))∣≤L∣s(γnk)−s(γ∗)∣→0|h(s(\gamma_{n_{k}}))-h(s(\gamma^{*}))|\le L|s(\gamma_{n_{k}})-s(\gamma^{*})|\to0. Hence Ψ(γnk)→Ψ(γ∗)\Psi(\gamma_{n_{k}})\to\Psi(\gamma^{*}) by the limit rules of The Real Numbers: Standing Notation and Background §sequences. Since nk≥kn_{k}\ge k, we have Φ(κd(ν))−2−k<Ψ(γnk)≤Φ(κd(ν))\Phi(\kappa_{d}(\nu))-2^{-k}<\Psi(\gamma_{n_{k}})\le\Phi(\kappa_{d}(\nu)) for every kk, so Ψ(γ∗)=Φ(κd(ν))\Psi(\gamma^{*})=\Phi(\kappa_{d}(\nu)) and γ∗\gamma^{*} is a maximising coupling for ν\nu.

Proof of clause 2. We show ∣Φ(κd(ν))−Φ(κd(ν′))∣≤K W2(ν,ν′)|\Phi(\kappa_{d}(\nu))-\Phi(\kappa_{d}(\nu'))|\le K\,W_{2}(\nu,\nu') for all ν,ν′∈Σd,r\nu,\nu'\in\Sigma_{d,r}. Let γ\gamma be a maximising coupling for ν\nu, which exists by clause 1. Step E gives γ′′∈C(κd(ν′))\gamma''\in\mathcal{C}(\kappa_{d}(\nu')) with Φ(κd(ν′))≥Ψ(γ′′)≥Ψ(γ)−K W2(ν,ν′)=Φ(κd(ν))−K W2(ν,ν′)\Phi(\kappa_{d}(\nu'))\ge\Psi(\gamma'')\ge\Psi(\gamma)-K\,W_{2}(\nu,\nu')=\Phi(\kappa_{d}(\nu))-K\,W_{2}(\nu,\nu'). Exchanging ν\nu and ν′\nu' and using W2(ν′,ν)=W2(ν,ν′)W_{2}(\nu',\nu)=W_{2}(\nu,\nu') from The Noncommutative Wasserstein Distance: Existence of Optimal Couplings, Symmetry, Separation, a Moment Bound, Weak-Star Lower Semicontinuity, and Displacement Interpolation §symmetry gives the claim. Consequently, for ν∈Σd,r\nu\in\Sigma_{d,r} and real ε>0\varepsilon>0, the choice δ=ε/K\delta=\varepsilon/K gives ∣Φ(κd(ν))−Φ(κd(ν′))∣<ε|\Phi(\kappa_{d}(\nu))-\Phi(\kappa_{d}(\nu'))|<\varepsilon whenever ν′∈Σd,r\nu'\in\Sigma_{d,r} and W2(ν,ν′)<δW_{2}(\nu,\nu')<\delta, which is continuity in the sense of Continuous Map Between Metric Spaces for the metric space of The Noncommutative Laws with a Norm Bound Form a Complete Bounded Metric Space with Interpolation Points and the metric of Plan Jets and Hamiltonians on Square-Integrable Noncommutative Laws: Standing Notation §metrics.

Proof of clause 3. Put c=h′(s)/s+2θc=h'(s)/s+2\theta if s>0s>0 and c=0c=0 if s=0s=0, so that S=P−caS=P-ca in both cases (for s=0s=0, S=PS=P), and SS is an L2L^{2} dd-tuple by Sums, Real Multiples and the Pairing of Square-Integrable Tuples in a Tracial W*-Probability Space §operations. Since (X,P,X′)(X,P,X') realises γ\gamma, s=s(γ)s=s(\gamma) and p(γ)=⟨P,a⟩2p(\gamma)=\langle P,a\rangle_{2}. By Step B, law(X′)=κd(ν)\mathrm{law}(X')=\kappa_{d}(\nu), and by Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §push-forward, pr#1law(X′,S)=law(X′)\mathrm{pr}^{1}_{\#}\mathrm{law}(X',S)=\mathrm{law}(X'), so law(X′,S)\mathrm{law}(X',S) is a plan at κd(ν)\kappa_{d}(\nu) in the sense of Plans at a Square-Integrable Noncommutative Law §plan. By Step A, h′(s)≥0h'(s)\ge0, so c≥0c\ge0; if s>0s>0, ∥S−P∥2=c s=h′(s)+2θs\lVert S-P\rVert_{2}=c\,s=h'(s)+2\theta s, and if s=0s=0, ∥S−P∥2=0=h′(0)+2θ⋅0\lVert S-P\rVert_{2}=0=h'(0)+2\theta\cdot0 because h′(0)=0h'(0)=0. This proves the norm bound.

Fix a real η>0\eta>0. The radius is chosen as follows, in this order. If s>0s>0, let ρ>0\rho>0 be the radius rη/2r_{\eta/2} of the hypothesis on hh at the point ss with η/2\eta/2 in place of η\eta, and put r∗=min⁡(ρ, η/(2(L/(2s)+θ)))r_{*}=\min\bigl(\rho,\ \eta/(2(L/(2s)+\theta))\bigr). If s=0s=0, let ρ>0\rho>0 be the radius rη/2r_{\eta/2} of the hypothesis on hh at the point 00, and put r∗=min⁡(ρ, η/(2θ))r_{*}=\min(\rho,\ \eta/(2\theta)). We verify the inequality of Plan Superdifferentials, Plan Subdifferentials and Plan Jets with Slack on Square-Integrable Noncommutative Laws §sub with slack 00 and radius r∗r_{*}. Let (H1,M1,Ω1)(H_{1},M_{1},\Omega_{1}) be a tracial W*-probability space and Y,S~,Y′′Y,\tilde{S},Y'' L2L^{2} dd-tuples of it with law(Y,S~)=law(X′,S)\mathrm{law}(Y,\tilde{S})=\mathrm{law}(X',S) and ∥Y′′−Y∥2<r∗\lVert Y''-Y\rVert_{2}<r_{*}.

Step D with Z=(X′,S)Z=(X',S), W=(X,P)W=(X,P), Z1=(Y,S~)Z_{1}=(Y,\tilde{S}) and V=Y′′V=Y'' gives L2L^{2} dd-tuples X^′,S^,X^,P^,Y^′′\hat{X}',\hat{S},\hat{X},\hat{P},\hat{Y}'' of one tracial W*-probability space with law((X^′,S^),(X^,P^))=law((X′,S),(X,P))\mathrm{law}((\hat{X}',\hat{S}),(\hat{X},\hat{P}))=\mathrm{law}((X',S),(X,P)) and law((X^′,S^),Y^′′)=law((Y,S~),Y′′)\mathrm{law}((\hat{X}',\hat{S}),\hat{Y}'')=\mathrm{law}((Y,\tilde{S}),Y''). By Step C: law(X^,P^,X^′)=law(X,P,X′)=γ\mathrm{law}(\hat{X},\hat{P},\hat{X}')=\mathrm{law}(X,P,X')=\gamma; ∥S^−P^+c(X^′−X^)∥2=∥S−P+ca∥2=0\lVert\hat{S}-\hat{P}+c(\hat{X}'-\hat{X})\rVert_{2}=\lVert S-P+ca\rVert_{2}=0, so S^=P^−ca^\hat{S}=\hat{P}-c\hat{a} with a^=X^′−X^\hat{a}=\hat{X}'-\hat{X}; and, with v=Y^′′−X^′v=\hat{Y}''-\hat{X}', law(Y^′′)=law(Y′′)\mathrm{law}(\hat{Y}'')=\mathrm{law}(Y''), ∥v∥2=∥Y′′−Y∥2\lVert v\rVert_{2}=\lVert Y''-Y\rVert_{2} and ⟨S^,v⟩2=⟨S~,Y′′−Y⟩2\langle\hat{S},v\rangle_{2}=\langle\tilde{S},Y''-Y\rangle_{2}. Computing from the realisation (X^,P^,X^′)(\hat{X},\hat{P},\hat{X}') of γ\gamma, s=∥a^∥2s=\lVert\hat{a}\rVert_{2} and p(γ)=⟨P^,a^⟩2p(\gamma)=\langle\hat{P},\hat{a}\rangle_{2}. Put γ′′=law(X^,P^,Y^′′)\gamma''=\mathrm{law}(\hat{X},\hat{P},\hat{Y}''). As in Step E, B#γ′′=law(X^,P^)=B#γ=κ2d(π)B_{\#}\gamma''=\mathrm{law}(\hat{X},\hat{P})=B_{\#}\gamma=\kappa_{2d}(\pi) and C#γ′′=law(Y^′′)C_{\#}\gamma''=\mathrm{law}(\hat{Y}'') by Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §push-forward, so γ′′∈C(law(Y^′′))\gamma''\in\mathcal{C}(\mathrm{law}(\hat{Y}'')) by Couplings of a Square-Integrable Plan with a Law: Displacement and Momentum Pairing §couplings, and s(γ′′)=ts(\gamma'')=t with t=∥a^+v∥2t=\lVert\hat{a}+v\rVert_{2} and p(γ′′)=⟨P^,a^+v⟩2p(\gamma'')=\langle\hat{P},\hat{a}+v\rangle_{2}. We have ∣t−s∣≤∥v∥2<r∗≤ρ|t-s|\le\lVert v\rVert_{2}<r_{*}\le\rho and t2−s2=2⟨a^,v⟩2+∥v∥22t^{2}-s^{2}=2\langle\hat{a},v\rangle_{2}+\lVert v\rVert_{2}^{2}.

Case s>0s>0. Since 0≤(t−s)20\le(t-s)^{2} gives 2s(t−s)≤t2−s22s(t-s)\le t^{2}-s^{2}, we get t−s≤⟨a^,v⟩2/s+∥v∥22/(2s)t-s\le\langle\hat{a},v\rangle_{2}/s+\lVert v\rVert_{2}^{2}/(2s). By the choice of ρ\rho, h(t)−h(s)≤h′(s)(t−s)+η2∣t−s∣h(t)-h(s)\le h'(s)(t-s)+\tfrac{\eta}{2}|t-s|, and, since 0≤h′(s)≤L0\le h'(s)\le L and ∣t−s∣≤∥v∥2|t-s|\le\lVert v\rVert_{2},

h(t)−h(s)≤h′(s)s⟨a^,v⟩2+L2s∥v∥22+η2∥v∥2.h(t)-h(s)\le\frac{h'(s)}{s}\langle\hat{a},v\rangle_{2}+\frac{L}{2s}\lVert v\rVert_{2}^{2}+\frac{\eta}{2}\lVert v\rVert_{2}.

Therefore

Ψ(γ′′)−Ψ(γ)=⟨P^,v⟩2−(h(t)−h(s))−θ(2⟨a^,v⟩2+∥v∥22)≥⟨P^−ca^,v⟩2−(η2+(L2s+θ)∥v∥2)∥v∥2,\Psi(\gamma'')-\Psi(\gamma)=\langle\hat{P},v\rangle_{2}-\bigl(h(t)-h(s)\bigr)-\theta\bigl(2\langle\hat{a},v\rangle_{2}+\lVert v\rVert_{2}^{2}\bigr)\ge\langle\hat{P}-c\hat{a},v\rangle_{2}-\Bigl(\frac{\eta}{2}+\Bigl(\frac{L}{2s}+\theta\Bigr)\lVert v\rVert_{2}\Bigr)\lVert v\rVert_{2},

and (L/(2s)+θ)∥v∥2<η/2(L/(2s)+\theta)\lVert v\rVert_{2}<\eta/2 by the choice of r∗r_{*}. Hence Ψ(γ′′)≥Ψ(γ)+⟨S^,v⟩2−η∥v∥2\Psi(\gamma'')\ge\Psi(\gamma)+\langle\hat{S},v\rangle_{2}-\eta\lVert v\rVert_{2}.

Case s=0s=0. Then a^=0\hat{a}=0, S^=P^\hat{S}=\hat{P}, t=∥v∥2t=\lVert v\rVert_{2}, p(γ)=0p(\gamma)=0 and Ψ(γ)=−h(0)=0\Psi(\gamma)=-h(0)=0. By the choice of ρ\rho and h(0)=h′(0)=0h(0)=h'(0)=0, h(t)≤η2th(t)\le\tfrac{\eta}{2}t; and θt2≤η2t\theta t^{2}\le\tfrac{\eta}{2}t since t<η/(2θ)t<\eta/(2\theta). Hence Ψ(γ′′)=⟨P^,v⟩2−h(t)−θt2≥Ψ(γ)+⟨S^,v⟩2−η∥v∥2\Psi(\gamma'')=\langle\hat{P},v\rangle_{2}-h(t)-\theta t^{2}\ge\Psi(\gamma)+\langle\hat{S},v\rangle_{2}-\eta\lVert v\rVert_{2}.

In both cases, since γ′′∈C(law(Y^′′))\gamma''\in\mathcal{C}(\mathrm{law}(\hat{Y}'')), law(Y^′′)=law(Y′′)\mathrm{law}(\hat{Y}'')=\mathrm{law}(Y'') and Ψ(γ)=Φ(κd(ν))\Psi(\gamma)=\Phi(\kappa_{d}(\nu)), the Notation and the identities from Step C give

Φ(law(Y′′))≥Ψ(γ′′)≥Φ(κd(ν))+⟨S~,Y′′−Y⟩2−(0+η)∥Y′′−Y∥2.\Phi(\mathrm{law}(Y''))\ge\Psi(\gamma'')\ge\Phi(\kappa_{d}(\nu))+\langle\tilde{S},Y''-Y\rangle_{2}-(0+\eta)\lVert Y''-Y\rVert_{2}.

The left side is ΦM1(Y′′)\Phi_{M_{1}}(Y'') by Plan Jets and Hamiltonians on Square-Integrable Noncommutative Laws: Standing Notation §lifts and Lifts of Functions on Square-Integrable Noncommutative Laws to Square-Integrable Tuples §lift. So law(X′,S)\mathrm{law}(X',S) is a plan subdifferential of Φ\Phi at κd(ν)\kappa_{d}(\nu) with slack 00 in the sense of Plan Superdifferentials, Plan Subdifferentials and Plan Jets with Slack on Square-Integrable Noncommutative Laws §sub, that is, law(X′,S)∈J−Φ(κd(ν))\mathrm{law}(X',S)\in J^{-}\Phi(\kappa_{d}(\nu)) by Plan Superdifferentials, Plan Subdifferentials and Plan Jets with Slack on Square-Integrable Noncommutative Laws §subjet.

Proof of clause 4. First, Ψ(γ)≤−θ2s(γ)2\Psi(\gamma)\le-\tfrac{\theta}{2}s(\gamma)^{2} for every γ∈C(κd(μ))\gamma\in\mathcal{C}(\kappa_{d}(\mu)). Write s=s(γ)s=s(\gamma). If s<r0s<r_{0}, the hypothesis gives p(γ)≤12h(s)p(\gamma)\le\tfrac12h(s), so Ψ(γ)≤−12h(s)−θs2≤−θ2s2\Psi(\gamma)\le-\tfrac12h(s)-\theta s^{2}\le-\tfrac{\theta}{2}s^{2} because h≥0h\ge0. If s≥r0s\ge r_{0}, Step B (with ν=μ∈Σd,r\nu=\mu\in\Sigma_{d,r}) gives p(γ)≤aπs≤θr02s≤θ2s2p(\gamma)\le a_{\pi}s\le\tfrac{\theta r_{0}}{2}s\le\tfrac{\theta}{2}s^{2}, so Ψ(γ)≤θ2s2−h(s)−θs2≤−θ2s2\Psi(\gamma)\le\tfrac{\theta}{2}s^{2}-h(s)-\theta s^{2}\le-\tfrac{\theta}{2}s^{2}. Hence Φ(κd(μ))≤0\Phi(\kappa_{d}(\mu))\le0.

Second, by Every Square-Integrable Noncommutative Law is the Law of a Square-Integrable Tuple; Realisation of Couplings and of Almost Optimal Pairs §coupling there are L2L^{2} dd-tuples X,PX,P of a tracial W*-probability space with law(X,P)=κ2d(π)\mathrm{law}(X,P)=\kappa_{2d}(\pi), and then law(X)=pr#1κ2d(π)=κd(μ)\mathrm{law}(X)=\mathrm{pr}^{1}_{\#}\kappa_{2d}(\pi)=\kappa_{d}(\mu) as computed in Step B. Put γ0=law(X,P,X)\gamma_{0}=\mathrm{law}(X,P,X). By Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §push-forward, B#γ0=law(X,P)=κ2d(π)B_{\#}\gamma_{0}=\mathrm{law}(X,P)=\kappa_{2d}(\pi) and C#γ0=law(X)=κd(μ)C_{\#}\gamma_{0}=\mathrm{law}(X)=\kappa_{d}(\mu), so γ0∈C(κd(μ))\gamma_{0}\in\mathcal{C}(\kappa_{d}(\mu)) by Couplings of a Square-Integrable Plan with a Law: Displacement and Momentum Pairing §couplings. From this realisation, s(γ0)=∥X−X∥2=0s(\gamma_{0})=\lVert X-X\rVert_{2}=0 and p(γ0)=⟨P,X−X⟩2=0p(\gamma_{0})=\langle P,X-X\rangle_{2}=0, so Ψ(γ0)=−h(0)=0\Psi(\gamma_{0})=-h(0)=0 and Φ(κd(μ))≥0\Phi(\kappa_{d}(\mu))\ge0. Therefore Φ(κd(μ))=0\Phi(\kappa_{d}(\mu))=0.

Proof of clause 5. Put wn=W2(νn,μ)w_{n}=W_{2}(\nu_{n},\mu) and sn=s(γn)s_{n}=s(\gamma_{n}). The hypotheses of clause 4 hold, so Φ(κd(μ))=0\Phi(\kappa_{d}(\mu))=0, and the estimate proved for clause 2 gives ∣Φ(κd(νn))∣≤Kwn|\Phi(\kappa_{d}(\nu_{n}))|\le Kw_{n} for every nn; hence (Φ(κd(νn)))n(\Phi(\kappa_{d}(\nu_{n})))_{n} converges to 00. For every nn, Step E (with ν=νn\nu=\nu_{n}, ν′=μ\nu'=\mu, γ=γn\gamma=\gamma_{n}) gives γn′′∈C(κd(μ))\gamma''_{n}\in\mathcal{C}(\kappa_{d}(\mu)) with ∣s(γn′′)−sn∣≤wn|s(\gamma''_{n})-s_{n}|\le w_{n} and Ψ(γn′′)≥Ψ(γn)−Kwn=Φ(κd(νn))−Kwn≥−2Kwn\Psi(\gamma''_{n})\ge\Psi(\gamma_{n})-Kw_{n}=\Phi(\kappa_{d}(\nu_{n}))-Kw_{n}\ge-2Kw_{n}. By the first part of the proof of clause 4, Ψ(γn′′)≤−θ2s(γn′′)2\Psi(\gamma''_{n})\le-\tfrac{\theta}{2}s(\gamma''_{n})^{2}, so s(γn′′)2≤4Kwn/θs(\gamma''_{n})^{2}\le4Kw_{n}/\theta. Given a real ε>0\varepsilon>0, choose N∈NN\in\mathbb{N} such that wn<min⁡(ε/2, θε2/(16K))w_{n}<\min\bigl(\varepsilon/2,\ \theta\varepsilon^{2}/(16K)\bigr) for all n≥Nn\ge N, by the convergence of (wn)n(w_{n})_{n} to 00 in the sense of The Real Numbers: Standing Notation and Background §sequences; then for n≥Nn\ge N, s(γn′′)2<ε2/4s(\gamma''_{n})^{2}<\varepsilon^{2}/4, so s(γn′′)<ε/2s(\gamma''_{n})<\varepsilon/2, and 0≤sn≤s(γn′′)+wn<ε0\le s_{n}\le s(\gamma''_{n})+w_{n}<\varepsilon. Hence (sn)n(s_{n})_{n} converges to 00.

For the last assertion, let (Xn,Pn,Xn′)(X_{n},P_{n},X'_{n}) and SnS_{n} be as stated. Since γn\gamma_{n} is a maximising coupling for νn\nu_{n}, clause 3 gives ∥Sn−Pn∥2≤h′(sn)+2θsn\lVert S_{n}-P_{n}\rVert_{2}\le h'(s_{n})+2\theta s_{n}, and h′(sn)≥0h'(s_{n})\ge0 by Step A. Given a real ε>0\varepsilon>0, let rh>0r_{h}>0 be the radius of the hypothesis on h′h' for ε/2\varepsilon/2 in place of η\eta, and then choose N∈NN\in\mathbb{N} with sn<min⁡(rh, ε/(4θ))s_{n}<\min\bigl(r_{h},\ \varepsilon/(4\theta)\bigr) for n≥Nn\ge N. Then 0≤∥Sn−Pn∥2<ε/2+ε/2=ε0\le\lVert S_{n}-P_{n}\rVert_{2}<\varepsilon/2+\varepsilon/2=\varepsilon for n≥Nn\ge N, so (∥Sn−Pn∥2)n(\lVert S_{n}-P_{n}\rVert_{2})_{n} converges to 00.

Citations

Loading…

Dependencies

Uses0

Loading…

Comments

Log in to comment.

Loading…