Each result cited is universally quantified over the data in its own statement. Write λ = l a w ( X ^ ) \lambda=\mathrm{law}(\hat{X}) λ = law ( X ^ ) , ν = l a w ( Y ^ ) \nu=\mathrm{law}(\hat{Y}) ν = law ( Y ^ ) and D = ∥ X ^ − Y ^ ∥ 2 D=\lVert\hat{X}-\hat{Y}\rVert_{2} D = ∥ X ^ − Y ^ ∥ 2 .
Step 0 (tools). (a) Tuple algebra. Fix a tracial W*-probability space ( K , N , Ψ ) (K,N,\Psi) ( K , N , Ψ ) . By Plan Jets and Hamiltonians on Square-Integrable Noncommutative Laws: Standing Notation §pairing and Sums, Real Multiples and the Pairing of Square-Integrable Tuples in a Tracial W*-Probability Space , its L 2 L^{2} L 2 d d d -tuples lie in the complex Hilbert space K d K^{d} K d , sums and real multiples are those of K d K^{d} K d , the pairing ⟨ ⋅ , ⋅ ⟩ 2 \langle\cdot,\cdot\rangle_{2} ⟨ ⋅ , ⋅ ⟩ 2 is the inner product of K d K^{d} K d and takes real values on L 2 L^{2} L 2 d d d -tuples, and ∥ ⋅ ∥ 2 \lVert\cdot\rVert_{2} ∥ ⋅ ∥ 2 is the norm of K d K^{d} K d ; the difference Z − W Z-W Z − W of Square-Integrable Tuples in a Tracial W*-Probability Space: Their Norm, Affine Images, Pairs, Embedded Images and Laws §tuples is Z + ( − 1 ) W Z+(-1)W Z + ( − 1 ) W . By conjugate symmetry and linearity in the second argument (Complex Inner Product Space , conditions 1, 2 and 3) and realness, ⟨ ⋅ , ⋅ ⟩ 2 \langle\cdot,\cdot\rangle_{2} ⟨ ⋅ , ⋅ ⟩ 2 is symmetric and real bilinear on L 2 L^{2} L 2 d d d -tuples, so for all L 2 L^{2} L 2 d d d -tuples Z , W Z,W Z , W ,
∥ Z + W ∥ 2 2 = ∥ Z ∥ 2 2 + 2 ⟨ Z , W ⟩ 2 + ∥ W ∥ 2 2 . (E) \lVert Z+W\rVert_{2}^{2}=\lVert Z\rVert_{2}^{2}+2\langle Z,W\rangle_{2}+\lVert W\rVert_{2}^{2}.\tag{E} ∥ Z + W ∥ 2 2 = ∥ Z ∥ 2 2 + 2 ⟨ Z , W ⟩ 2 + ∥ W ∥ 2 2 . ( E )
Moreover ∣ ⟨ Z , W ⟩ 2 ∣ ≤ ∥ Z ∥ 2 ∥ W ∥ 2 |\langle Z,W\rangle_{2}|\le\lVert Z\rVert_{2}\lVert W\rVert_{2} ∣ ⟨ Z , W ⟩ 2 ∣ ≤ ∥ Z ∥ 2 ∥ W ∥ 2 by Cauchy-Schwarz Inequality in a Complex Inner Product Space , and ∥ t Z ∥ 2 = ∣ t ∣ ∥ Z ∥ 2 \lVert tZ\rVert_{2}=|t|\,\lVert Z\rVert_{2} ∥ tZ ∥ 2 = ∣ t ∣ ∥ Z ∥ 2 and ∥ Z + W ∥ 2 ≤ ∥ Z ∥ 2 + ∥ W ∥ 2 \lVert Z+W\rVert_{2}\le\lVert Z\rVert_{2}+\lVert W\rVert_{2} ∥ Z + W ∥ 2 ≤ ∥ Z ∥ 2 + ∥ W ∥ 2 by claim 2 of The Induced Norm is a Norm, and Induces a Metric .
(b) Φ \Phi Φ on pairs. For L 2 L^{2} L 2 d d d -tuples X , Y X,Y X , Y of ( K , N , Ψ ) (K,N,\Psi) ( K , N , Ψ ) , Every Square-Integrable Noncommutative Law is the Law of a Square-Integrable Tuple; Realisation of Couplings and of Almost Optimal Pairs §coupling gives p r # 1 l a w ( X , Y ) = l a w ( X ) \mathrm{pr}^{1}_{\#}\mathrm{law}(X,Y)=\mathrm{law}(X) pr # 1 law ( X , Y ) = law ( X ) , p r # 2 l a w ( X , Y ) = l a w ( Y ) \mathrm{pr}^{2}_{\#}\mathrm{law}(X,Y)=\mathrm{law}(Y) pr # 2 law ( X , Y ) = law ( Y ) and I ( l a w ( X , Y ) ) = ∥ X − Y ∥ 2 2 \mathcal{I}(\mathrm{law}(X,Y))=\lVert X-Y\rVert_{2}^{2} I ( law ( X , Y )) = ∥ X − Y ∥ 2 2 , and Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §moments gives M ^ ( l a w ( Z ) ) = ∥ Z ∥ 2 2 \widehat{M}(\mathrm{law}(Z))=\lVert Z\rVert_{2}^{2} M ( law ( Z )) = ∥ Z ∥ 2 2 . Writing g ( Z ) = log ( 1 + ∥ Z ∥ 2 2 ) g(Z)=\log(1+\lVert Z\rVert_{2}^{2}) g ( Z ) = log ( 1 + ∥ Z ∥ 2 2 ) , so that G ( l a w ( Z ) ) = g ( Z ) G(\mathrm{law}(Z))=g(Z) G ( law ( Z )) = g ( Z ) , and using the lifts ,
Φ ( l a w ( X , Y ) ) = u N ( X ) − v N ( Y ) − 1 2 ε ∥ X − Y ∥ 2 2 − β g ( X ) − β g ( Y ) . (F) \Phi(\mathrm{law}(X,Y))=u_{N}(X)-v_{N}(Y)-\tfrac{1}{2\varepsilon}\lVert X-Y\rVert_{2}^{2}-\beta g(X)-\beta g(Y).\tag{F} Φ ( law ( X , Y )) = u N ( X ) − v N ( Y ) − 2 ε 1 ∥ X − Y ∥ 2 2 − β g ( X ) − β g ( Y ) . ( F )
If X ′ X' X ′ is a further L 2 L^{2} L 2 d d d -tuple of ( K , N , Ψ ) (K,N,\Psi) ( K , N , Ψ ) , the pairs ( X ′ , Y ) (X',Y) ( X ′ , Y ) and ( X , Y ) (X,Y) ( X , Y ) are L 2 L^{2} L 2 2 d 2d 2 d -tuples (Square-Integrable Tuples in a Tracial W*-Probability Space: Their Norm, Affine Images, Pairs, Embedded Images and Laws §operations ) whose difference has coordinates X j ′ − X j X'_{j}-X_{j} X j ′ − X j (j ∈ [ d ] j\in[d] j ∈ [ d ] ) followed by d d d zeros, so its L 2 L^{2} L 2 norm is ∥ X ′ − X ∥ 2 \lVert X'-X\rVert_{2} ∥ X ′ − X ∥ 2 by the definition of the norm in Square-Integrable Tuples in a Tracial W*-Probability Space: Their Norm, Affine Images, Pairs, Embedded Images and Laws §tuples , and Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §lipschitz gives W ^ 2 ( l a w ( X ′ , Y ) , l a w ( X , Y ) ) ≤ ∥ X ′ − X ∥ 2 \widehat{W}_{2}(\mathrm{law}(X',Y),\mathrm{law}(X,Y))\le\lVert X'-X\rVert_{2} W 2 ( law ( X ′ , Y ) , law ( X , Y )) ≤ ∥ X ′ − X ∥ 2 ; likewise W ^ 2 ( l a w ( X , Y ′ ) , l a w ( X , Y ) ) ≤ ∥ Y ′ − Y ∥ 2 \widehat{W}_{2}(\mathrm{law}(X,Y'),\mathrm{law}(X,Y))\le\lVert Y'-Y\rVert_{2} W 2 ( law ( X , Y ′ ) , law ( X , Y )) ≤ ∥ Y ′ − Y ∥ 2 .
(c) Logarithm. For reals s , s ′ ≥ 0 s,s'\ge0 s , s ′ ≥ 0 , put t = ( 1 + s ′ ) / ( 1 + s ) > 0 t=(1+s')/(1+s)>0 t = ( 1 + s ′ ) / ( 1 + s ) > 0 . Then log ( 1 + s ′ ) = log ( 1 + s ) + log t \log(1+s')=\log(1+s)+\log t log ( 1 + s ′ ) = log ( 1 + s ) + log t by the product rule in The Natural Logarithm , and log t ≤ t − 1 = ( s ′ − s ) / ( 1 + s ) \log t\le t-1=(s'-s)/(1+s) log t ≤ t − 1 = ( s ′ − s ) / ( 1 + s ) by The Function s log s s\log s s log s : Continuity, Young's Inequality and Lower Bounds, with the Elementary Bounds for the Exponential and the Logarithm §log . For L 2 L^{2} L 2 d d d -tuples Z , Z ′ Z,Z' Z , Z ′ of ( K , N , Ψ ) (K,N,\Psi) ( K , N , Ψ ) , apply this with s = ∥ Z ∥ 2 2 s=\lVert Z\rVert_{2}^{2} s = ∥ Z ∥ 2 2 , s ′ = ∥ Z ′ ∥ 2 2 s'=\lVert Z'\rVert_{2}^{2} s ′ = ∥ Z ′ ∥ 2 2 ; by (E) with W = Z ′ − Z W=Z'-Z W = Z ′ − Z , s ′ − s = 2 ⟨ Z , Z ′ − Z ⟩ 2 + ∥ Z ′ − Z ∥ 2 2 s'-s=2\langle Z,Z'-Z\rangle_{2}+\lVert Z'-Z\rVert_{2}^{2} s ′ − s = 2 ⟨ Z , Z ′ − Z ⟩ 2 + ∥ Z ′ − Z ∥ 2 2 , and 1 / ( 1 + s ) ≤ 1 1/(1+s)\le1 1/ ( 1 + s ) ≤ 1 , so
g ( Z ′ ) − g ( Z ) ≤ ⟨ 2 1 + ∥ Z ∥ 2 2 Z , Z ′ − Z ⟩ 2 + ∥ Z ′ − Z ∥ 2 2 . (L) g(Z')-g(Z)\le\Bigl\langle\frac{2}{1+\lVert Z\rVert_{2}^{2}}Z,\;Z'-Z\Bigr\rangle_{2}+\lVert Z'-Z\rVert_{2}^{2}.\tag{L} g ( Z ′ ) − g ( Z ) ≤ ⟨ 1 + ∥ Z ∥ 2 2 2 Z , Z ′ − Z ⟩ 2 + ∥ Z ′ − Z ∥ 2 2 . ( L )
(d) Norms are law data. If Z , Z ′ Z,Z' Z , Z ′ are L 2 L^{2} L 2 d d d -tuples (of possibly different spaces) with l a w ( Z ) = l a w ( Z ′ ) \mathrm{law}(Z)=\mathrm{law}(Z') law ( Z ) = law ( Z ′ ) , then ∥ Z ∥ 2 = ∥ Z ′ ∥ 2 \lVert Z\rVert_{2}=\lVert Z'\rVert_{2} ∥ Z ∥ 2 = ∥ Z ′ ∥ 2 by Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §moments and Monotonicity of Squaring on the Nonnegative Elements of an Ordered Field (claim 3). If l a w ( X , P ) = l a w ( X ′ , P ′ ) \mathrm{law}(X,P)=\mathrm{law}(X',P') law ( X , P ) = law ( X ′ , P ′ ) , then l a w ( X ) = l a w ( X ′ ) \mathrm{law}(X)=\mathrm{law}(X') law ( X ) = law ( X ′ ) and l a w ( P ) = l a w ( P ′ ) \mathrm{law}(P)=\mathrm{law}(P') law ( P ) = law ( P ′ ) by Every Square-Integrable Noncommutative Law is the Law of a Square-Integrable Tuple; Realisation of Couplings and of Almost Optimal Pairs §coupling .
(e) The Ekeland inequality. By hypothesis, Φ ( γ ) ≤ Φ ( γ ^ ) + δ W ^ 2 ( γ , γ ^ ) \Phi(\gamma)\le\Phi(\hat{\gamma})+\delta\,\widehat{W}_{2}(\gamma,\hat{\gamma}) Φ ( γ ) ≤ Φ ( γ ^ ) + δ W 2 ( γ , γ ^ ) for every γ ∈ Σ 2 d 2 \gamma\in\Sigma^{2}_{2d} γ ∈ Σ 2 d 2 .
Clause 1. Put s = ∥ X ^ ∥ 2 2 s=\lVert\hat{X}\rVert_{2}^{2} s = ∥ X ^ ∥ 2 2 and κ = 2 β ε / ( 1 + s ) \kappa=2\beta\varepsilon/(1+s) κ = 2 βε / ( 1 + s ) , so q = ( κ / ε ) X ^ q=(\kappa/\varepsilon)\hat{X} q = ( κ / ε ) X ^ . Let π u = l a w ( X ^ , p + q ) \pi_{u}=\mathrm{law}(\hat{X},p+q) π u = law ( X ^ , p + q ) ; by Every Square-Integrable Noncommutative Law is the Law of a Square-Integrable Tuple; Realisation of Couplings and of Almost Optimal Pairs §coupling , p r # 1 π u = λ \mathrm{pr}^{1}_{\#}\pi_{u}=\lambda pr # 1 π u = λ , so π u \pi_{u} π u is a plan at λ \lambda λ . Let T = ( B , 0 ) T=(B,0) T = ( B , 0 ) be the affine datum from 2 d 2d 2 d to 2 d 2d 2 d variables with, for i , j ∈ [ d ] i,j\in[d] i , j ∈ [ d ] , B i j = 1 B_{ij}=1 B ij = 1 if i = j i=j i = j and 0 0 0 otherwise, B i , d + j = 0 B_{i,d+j}=0 B i , d + j = 0 , B d + i , j = 1 + κ B_{d+i,j}=1+\kappa B d + i , j = 1 + κ if i = j i=j i = j and 0 0 0 otherwise, and B d + i , d + j = − ε B_{d+i,d+j}=-\varepsilon B d + i , d + j = − ε if i = j i=j i = j and 0 0 0 otherwise. By Square-Integrable Tuples in a Tracial W*-Probability Space: Their Norm, Affine Images, Pairs, Embedded Images and Laws §operations , for all L 2 L^{2} L 2 d d d -tuples X , P X,P X , P of any tracial W*-probability space, T ( X , P ) = ( X , ( 1 + κ ) X − ε P ) T(X,P)=(X,(1+\kappa)X-\varepsilon P) T ( X , P ) = ( X , ( 1 + κ ) X − εP ) (the terms with coefficient 0 0 0 vanish). Since ε p = X ^ − Y ^ \varepsilon p=\hat{X}-\hat{Y} εp = X ^ − Y ^ and ε q = κ X ^ \varepsilon q=\kappa\hat{X} εq = κ X ^ , we get ( 1 + κ ) X ^ − ε ( p + q ) = Y ^ (1+\kappa)\hat{X}-\varepsilon(p+q)=\hat{Y} ( 1 + κ ) X ^ − ε ( p + q ) = Y ^ , so T ( X ^ , p + q ) = ( X ^ , Y ^ ) T(\hat{X},p+q)=(\hat{X},\hat{Y}) T ( X ^ , p + q ) = ( X ^ , Y ^ ) .
Let η > 0 \eta>0 η > 0 and put r = η ( 1 / ( 2 ε ) + β ) − 1 > 0 r=\eta\,(1/(2\varepsilon)+\beta)^{-1}>0 r = η ( 1/ ( 2 ε ) + β ) − 1 > 0 . Let ( K , N , Ψ ) (K,N,\Psi) ( K , N , Ψ ) be a tracial W*-probability space and X , P , X ′ X,P,X' X , P , X ′ be L 2 L^{2} L 2 d d d -tuples of it with l a w ( X , P ) = π u \mathrm{law}(X,P)=\pi_{u} law ( X , P ) = π u and ∥ X ′ − X ∥ 2 < r \lVert X'-X\rVert_{2}<r ∥ X ′ − X ∥ 2 < r . Put Y = ( 1 + κ ) X − ε P Y=(1+\kappa)X-\varepsilon P Y = ( 1 + κ ) X − εP . By Calculus of Laws of Square-Integrable Tuples: Bounded Tuples, the Lipschitz Bound, Moments, Affine Push-Forwards, Couplings and Embeddings §push-forward (twice),
l a w ( X , Y ) = l a w ( T ( X , P ) ) = T # π u = l a w ( T ( X ^ , p + q ) ) = l a w ( X ^ , Y ^ ) = γ ^ . \mathrm{law}(X,Y)=\mathrm{law}(T(X,P))=T_{\#}\pi_{u}=\mathrm{law}(T(\hat{X},p+q))=\mathrm{law}(\hat{X},\hat{Y})=\hat{\gamma}. law ( X , Y ) = law ( T ( X , P )) = T # π u = law ( T ( X ^ , p + q )) = law ( X ^ , Y ^ ) = γ ^ .
By (d), l a w ( X ) = λ \mathrm{law}(X)=\lambda law ( X ) = λ and ∥ X ∥ 2 2 = s \lVert X\rVert_{2}^{2}=s ∥ X ∥ 2 2 = s ; and X − Y = ε P − κ X X-Y=\varepsilon P-\kappa X X − Y = εP − κ X , so
P = 1 ε ( X − Y ) + 2 β 1 + ∥ X ∥ 2 2 X . (P) P=\tfrac{1}{\varepsilon}(X-Y)+\tfrac{2\beta}{1+\lVert X\rVert_{2}^{2}}X.\tag{P} P = ε 1 ( X − Y ) + 1 + ∥ X ∥ 2 2 2 β X . ( P )
By (e) with γ = l a w ( X ′ , Y ) \gamma=\mathrm{law}(X',Y) γ = law ( X ′ , Y ) and (b), Φ ( l a w ( X ′ , Y ) ) ≤ Φ ( l a w ( X , Y ) ) + δ ∥ X ′ − X ∥ 2 \Phi(\mathrm{law}(X',Y))\le\Phi(\mathrm{law}(X,Y))+\delta\lVert X'-X\rVert_{2} Φ ( law ( X ′ , Y )) ≤ Φ ( law ( X , Y )) + δ ∥ X ′ − X ∥ 2 . Expanding both sides by (F) and cancelling v N ( Y ) v_{N}(Y) v N ( Y ) and β g ( Y ) \beta g(Y) β g ( Y ) ,
u N ( X ′ ) ≤ u N ( X ) + 1 2 ε ( ∥ X ′ − Y ∥ 2 2 − ∥ X − Y ∥ 2 2 ) + β ( g ( X ′ ) − g ( X ) ) + δ ∥ X ′ − X ∥ 2 . u_{N}(X')\le u_{N}(X)+\tfrac{1}{2\varepsilon}\bigl(\lVert X'-Y\rVert_{2}^{2}-\lVert X-Y\rVert_{2}^{2}\bigr)+\beta\bigl(g(X')-g(X)\bigr)+\delta\lVert X'-X\rVert_{2}. u N ( X ′ ) ≤ u N ( X ) + 2 ε 1 ( ∥ X ′ − Y ∥ 2 2 − ∥ X − Y ∥ 2 2 ) + β ( g ( X ′ ) − g ( X ) ) + δ ∥ X ′ − X ∥ 2 .
By (E) with Z = X − Y Z=X-Y Z = X − Y and W = X ′ − X W=X'-X W = X ′ − X , ∥ X ′ − Y ∥ 2 2 − ∥ X − Y ∥ 2 2 = 2 ⟨ X − Y , X ′ − X ⟩ 2 + ∥ X ′ − X ∥ 2 2 \lVert X'-Y\rVert_{2}^{2}-\lVert X-Y\rVert_{2}^{2}=2\langle X-Y,X'-X\rangle_{2}+\lVert X'-X\rVert_{2}^{2} ∥ X ′ − Y ∥ 2 2 − ∥ X − Y ∥ 2 2 = 2 ⟨ X − Y , X ′ − X ⟩ 2 + ∥ X ′ − X ∥ 2 2 ; by (L) with Z = X Z=X Z = X , Z ′ = X ′ Z'=X' Z ′ = X ′ , β ( g ( X ′ ) − g ( X ) ) ≤ ⟨ 2 β 1 + s X , X ′ − X ⟩ 2 + β ∥ X ′ − X ∥ 2 2 \beta(g(X')-g(X))\le\langle\frac{2\beta}{1+s}X,X'-X\rangle_{2}+\beta\lVert X'-X\rVert_{2}^{2} β ( g ( X ′ ) − g ( X )) ≤ ⟨ 1 + s 2 β X , X ′ − X ⟩ 2 + β ∥ X ′ − X ∥ 2 2 . Adding, using bilinearity, (P) and u N ( X ) = u ( λ ) u_{N}(X)=u(\lambda) u N ( X ) = u ( λ ) ,
u N ( X ′ ) ≤ u ( λ ) + ⟨ P , X ′ − X ⟩ 2 + ( 1 2 ε + β ) ∥ X ′ − X ∥ 2 2 + δ ∥ X ′ − X ∥ 2 . u_{N}(X')\le u(\lambda)+\langle P,X'-X\rangle_{2}+\Bigl(\tfrac{1}{2\varepsilon}+\beta\Bigr)\lVert X'-X\rVert_{2}^{2}+\delta\lVert X'-X\rVert_{2}. u N ( X ′ ) ≤ u ( λ ) + ⟨ P , X ′ − X ⟩ 2 + ( 2 ε 1 + β ) ∥ X ′ − X ∥ 2 2 + δ ∥ X ′ − X ∥ 2 .
Since ∥ X ′ − X ∥ 2 < r \lVert X'-X\rVert_{2}<r ∥ X ′ − X ∥ 2 < r , ( 1 2 ε + β ) ∥ X ′ − X ∥ 2 2 ≤ η ∥ X ′ − X ∥ 2 (\frac{1}{2\varepsilon}+\beta)\lVert X'-X\rVert_{2}^{2}\le\eta\lVert X'-X\rVert_{2} ( 2 ε 1 + β ) ∥ X ′ − X ∥ 2 2 ≤ η ∥ X ′ − X ∥ 2 , whence u N ( X ′ ) ≤ u ( λ ) + ⟨ P , X ′ − X ⟩ 2 + ( δ + η ) ∥ X ′ − X ∥ 2 u_{N}(X')\le u(\lambda)+\langle P,X'-X\rangle_{2}+(\delta+\eta)\lVert X'-X\rVert_{2} u N ( X ′ ) ≤ u ( λ ) + ⟨ P , X ′ − X ⟩ 2 + ( δ + η ) ∥ X ′ − X ∥ 2 . Thus π u \pi_{u} π u is a plan superdifferential of u u u at λ \lambda λ with slack δ \delta δ (Plan Superdifferentials, Plan Subdifferentials and Plan Jets with Slack on Square-Integrable Noncommutative Laws §super ), and π u ∈ J δ + u ( λ ) \pi_{u}\in J^{+}_{\delta}u(\lambda) π u ∈ J δ + u ( λ ) by Plan Superdifferentials, Plan Subdifferentials and Plan Jets with Slack on Square-Integrable Noncommutative Laws §superjet .
Clause 2. Put t = ∥ Y ^ ∥ 2 2 t=\lVert\hat{Y}\rVert_{2}^{2} t = ∥ Y ^ ∥ 2 2 and κ ′ = 2 β ε / ( 1 + t ) \kappa'=2\beta\varepsilon/(1+t) κ ′ = 2 βε / ( 1 + t ) , so q ′ = ( κ ′ / ε ) Y ^ q'=(\kappa'/\varepsilon)\hat{Y} q ′ = ( κ ′ / ε ) Y ^ , and let π v = l a w ( Y ^ , p − q ′ ) \pi_{v}=\mathrm{law}(\hat{Y},p-q') π v = law ( Y ^ , p − q ′ ) , a plan at ν \nu ν as before. Let T ′ T' T ′ be the affine datum from 2 d 2d 2 d to 2 d 2d 2 d variables with T ′ ( Y , P ) = ( ( 1 + κ ′ ) Y + ε P , Y ) T'(Y,P)=((1+\kappa')Y+\varepsilon P,\,Y) T ′ ( Y , P ) = (( 1 + κ ′ ) Y + εP , Y ) for all L 2 L^{2} L 2 d d d -tuples Y , P Y,P Y , P (entries 1 + κ ′ 1+\kappa' 1 + κ ′ and ε \varepsilon ε on the diagonals of the upper blocks, 1 1 1 on the diagonal of the lower left block, 0 0 0 elsewhere, translation part 0 0 0 ; Square-Integrable Tuples in a Tracial W*-Probability Space: Their Norm, Affine Images, Pairs, Embedded Images and Laws §operations ). Since ( 1 + κ ′ ) Y ^ + ε ( p − q ′ ) = Y ^ + ( X ^ − Y ^ ) = X ^ (1+\kappa')\hat{Y}+\varepsilon(p-q')=\hat{Y}+(\hat{X}-\hat{Y})=\hat{X} ( 1 + κ ′ ) Y ^ + ε ( p − q ′ ) = Y ^ + ( X ^ − Y ^ ) = X ^ , T ′ ( Y ^ , p − q ′ ) = ( X ^ , Y ^ ) T'(\hat{Y},p-q')=(\hat{X},\hat{Y}) T ′ ( Y ^ , p − q ′ ) = ( X ^ , Y ^ ) .
Let η > 0 \eta>0 η > 0 , r r r as above, and let Y , P , Y ′ Y,P,Y' Y , P , Y ′ be L 2 L^{2} L 2 d d d -tuples of a tracial W*-probability space ( K , N , Ψ ) (K,N,\Psi) ( K , N , Ψ ) with l a w ( Y , P ) = π v \mathrm{law}(Y,P)=\pi_{v} law ( Y , P ) = π v and ∥ Y ′ − Y ∥ 2 < r \lVert Y'-Y\rVert_{2}<r ∥ Y ′ − Y ∥ 2 < r . Put X = ( 1 + κ ′ ) Y + ε P X=(1+\kappa')Y+\varepsilon P X = ( 1 + κ ′ ) Y + εP . As in clause 1, l a w ( X , Y ) = T # ′ π v = γ ^ \mathrm{law}(X,Y)=T'_{\#}\pi_{v}=\hat{\gamma} law ( X , Y ) = T # ′ π v = γ ^ , l a w ( Y ) = ν \mathrm{law}(Y)=\nu law ( Y ) = ν , ∥ Y ∥ 2 2 = t \lVert Y\rVert_{2}^{2}=t ∥ Y ∥ 2 2 = t , and P = 1 ε ( X − Y ) − 2 β 1 + t Y P=\frac{1}{\varepsilon}(X-Y)-\frac{2\beta}{1+t}Y P = ε 1 ( X − Y ) − 1 + t 2 β Y . By (e) with γ = l a w ( X , Y ′ ) \gamma=\mathrm{law}(X,Y') γ = law ( X , Y ′ ) , (b) and (F), after cancelling u N ( X ) u_{N}(X) u N ( X ) and β g ( X ) \beta g(X) β g ( X ) ,
v N ( Y ′ ) ≥ v ( ν ) − 1 2 ε ( ∥ X − Y ′ ∥ 2 2 − ∥ X − Y ∥ 2 2 ) − β ( g ( Y ′ ) − g ( Y ) ) − δ ∥ Y ′ − Y ∥ 2 . v_{N}(Y')\ge v(\nu)-\tfrac{1}{2\varepsilon}\bigl(\lVert X-Y'\rVert_{2}^{2}-\lVert X-Y\rVert_{2}^{2}\bigr)-\beta\bigl(g(Y')-g(Y)\bigr)-\delta\lVert Y'-Y\rVert_{2}. v N ( Y ′ ) ≥ v ( ν ) − 2 ε 1 ( ∥ X − Y ′ ∥ 2 2 − ∥ X − Y ∥ 2 2 ) − β ( g ( Y ′ ) − g ( Y ) ) − δ ∥ Y ′ − Y ∥ 2 .
By (E) with Z = X − Y Z=X-Y Z = X − Y , W = − ( Y ′ − Y ) W=-(Y'-Y) W = − ( Y ′ − Y ) , ∥ X − Y ′ ∥ 2 2 − ∥ X − Y ∥ 2 2 = − 2 ⟨ X − Y , Y ′ − Y ⟩ 2 + ∥ Y ′ − Y ∥ 2 2 \lVert X-Y'\rVert_{2}^{2}-\lVert X-Y\rVert_{2}^{2}=-2\langle X-Y,Y'-Y\rangle_{2}+\lVert Y'-Y\rVert_{2}^{2} ∥ X − Y ′ ∥ 2 2 − ∥ X − Y ∥ 2 2 = − 2 ⟨ X − Y , Y ′ − Y ⟩ 2 + ∥ Y ′ − Y ∥ 2 2 , and by (L), g ( Y ′ ) − g ( Y ) ≤ ⟨ 2 1 + t Y , Y ′ − Y ⟩ 2 + ∥ Y ′ − Y ∥ 2 2 g(Y')-g(Y)\le\langle\frac{2}{1+t}Y,Y'-Y\rangle_{2}+\lVert Y'-Y\rVert_{2}^{2} g ( Y ′ ) − g ( Y ) ≤ ⟨ 1 + t 2 Y , Y ′ − Y ⟩ 2 + ∥ Y ′ − Y ∥ 2 2 . Hence
v N ( Y ′ ) ≥ v ( ν ) + ⟨ P , Y ′ − Y ⟩ 2 − ( 1 2 ε + β ) ∥ Y ′ − Y ∥ 2 2 − δ ∥ Y ′ − Y ∥ 2 ≥ v ( ν ) + ⟨ P , Y ′ − Y ⟩ 2 − ( δ + η ) ∥ Y ′ − Y ∥ 2 , v_{N}(Y')\ge v(\nu)+\langle P,Y'-Y\rangle_{2}-\Bigl(\tfrac{1}{2\varepsilon}+\beta\Bigr)\lVert Y'-Y\rVert_{2}^{2}-\delta\lVert Y'-Y\rVert_{2}\ge v(\nu)+\langle P,Y'-Y\rangle_{2}-(\delta+\eta)\lVert Y'-Y\rVert_{2}, v N ( Y ′ ) ≥ v ( ν ) + ⟨ P , Y ′ − Y ⟩ 2 − ( 2 ε 1 + β ) ∥ Y ′ − Y ∥ 2 2 − δ ∥ Y ′ − Y ∥ 2 ≥ v ( ν ) + ⟨ P , Y ′ − Y ⟩ 2 − ( δ + η ) ∥ Y ′ − Y ∥ 2 ,
so π v ∈ J δ − v ( ν ) \pi_{v}\in J^{-}_{\delta}v(\nu) π v ∈ J δ − v ( ν ) by Plan Superdifferentials, Plan Subdifferentials and Plan Jets with Slack on Square-Integrable Noncommutative Laws §sub and Plan Superdifferentials, Plan Subdifferentials and Plan Jets with Slack on Square-Integrable Noncommutative Laws §subjet .
Clause 3. Let a = ∥ X ^ ∥ 2 ≥ 0 a=\lVert\hat{X}\rVert_{2}\ge0 a = ∥ X ^ ∥ 2 ≥ 0 . From 0 ≤ ( a − 1 ) 2 0\le(a-1)^{2} 0 ≤ ( a − 1 ) 2 we get 2 a ≤ 1 + a 2 2a\le1+a^{2} 2 a ≤ 1 + a 2 , and trivially a 2 ≤ 1 + a 2 a^{2}\le1+a^{2} a 2 ≤ 1 + a 2 . Hence ∥ q ∥ 2 = 2 β a / ( 1 + a 2 ) ≤ β \lVert q\rVert_{2}=2\beta a/(1+a^{2})\le\beta ∥ q ∥ 2 = 2 β a / ( 1 + a 2 ) ≤ β and ( 1 + a ) ∥ q ∥ 2 = 2 β a / ( 1 + a 2 ) + 2 β a 2 / ( 1 + a 2 ) ≤ β + 2 β = 3 β (1+a)\lVert q\rVert_{2}=2\beta a/(1+a^{2})+2\beta a^{2}/(1+a^{2})\le\beta+2\beta=3\beta ( 1 + a ) ∥ q ∥ 2 = 2 β a / ( 1 + a 2 ) + 2 β a 2 / ( 1 + a 2 ) ≤ β + 2 β = 3 β . The same computation with ∥ Y ^ ∥ 2 \lVert\hat{Y}\rVert_{2} ∥ Y ^ ∥ 2 bounds q ′ q' q ′ .
Clause 4. Let B ≥ 0 B\ge0 B ≥ 0 be real with ∣ u ∣ ≤ B |u|\le B ∣ u ∣ ≤ B and ∣ v ∣ ≤ B |v|\le B ∣ v ∣ ≤ B on Σ d 2 \Sigma^{2}_{d} Σ d 2 (Bounded Real-Valued Function on a Set ). Since I ≥ 0 \mathcal{I}\ge0 I ≥ 0 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 §cost and G ≥ 0 G\ge0 G ≥ 0 (as log ( 1 + s ) ≥ 1 − ( 1 + s ) − 1 ≥ 0 \log(1+s)\ge1-(1+s)^{-1}\ge0 log ( 1 + s ) ≥ 1 − ( 1 + s ) − 1 ≥ 0 for s ≥ 0 s\ge0 s ≥ 0 by The Function s log s s\log s s log s : Continuity, Young's Inequality and Lower Bounds, with the Elementary Bounds for the Exponential and the Logarithm §log ), Φ ≤ 2 B \Phi\le2B Φ ≤ 2 B and Φ ′ ≤ 2 B \Phi'\le2B Φ ′ ≤ 2 B , so both suprema exist (Approximation Property of the Supremum and the Infimum in R \mathbb{R} R ). By the definitions of Φ \Phi Φ and Φ ′ \Phi' Φ ′ and Every Square-Integrable Noncommutative Law is the Law of a Square-Integrable Tuple; Realisation of Couplings and of Almost Optimal Pairs §coupling (I ( γ ^ ) = D 2 \mathcal{I}(\hat{\gamma})=D^{2} I ( γ ^ ) = D 2 ), Φ ′ ( γ ^ ) = Φ ( γ ^ ) + ( 1 2 ε − 1 4 ε ) D 2 = Φ ( γ ^ ) + 1 4 ε D 2 \Phi'(\hat{\gamma})=\Phi(\hat{\gamma})+(\frac{1}{2\varepsilon}-\frac{1}{4\varepsilon})D^{2}=\Phi(\hat{\gamma})+\frac{1}{4\varepsilon}D^{2} Φ ′ ( γ ^ ) = Φ ( γ ^ ) + ( 2 ε 1 − 4 ε 1 ) D 2 = Φ ( γ ^ ) + 4 ε 1 D 2 . Hence 1 4 ε D 2 = Φ ′ ( γ ^ ) − Φ ( γ ^ ) ≤ sup Φ ′ − ( sup Φ − κ ) \frac{1}{4\varepsilon}D^{2}=\Phi'(\hat{\gamma})-\Phi(\hat{\gamma})\le\sup\Phi'-(\sup\Phi-\kappa) 4 ε 1 D 2 = Φ ′ ( γ ^ ) − Φ ( γ ^ ) ≤ sup Φ ′ − ( sup Φ − κ ) .