TheoremBase

The barycentric tuple z = E(Lxd+j)E(L_{x_{d+j}}) splits the cost of any coupling orthogonally into the cost of (LxL_x, z) plus M(nu) - M(zeta); gluing an optimal coupling of zeta and mu freely over zeta with the coupling (pi(z), Lxd+j)L_{x_{d+j}}) gives a coupling of mu and nu whose cross term vanishes, which yields the upper bound, and comparing the two estimates at an optimal coupling gives the optimality and the variational formula.

Proof

Each result cited below is universally quantified over the data in its own statement. We work in the setting of Noncommutative Laws, Couplings and the Wasserstein Distance: Standing Notation, with μ,ν∈Σd\mu,\nu\in\Sigma_{d} as in the statement, and use the following elementary facts throughout. Every variable satisfies xi∗=xix_{i}^{*}=x_{i} by Noncommutative Polynomials Form a Unital Complex Algebra with Involution: Linear Extension from Monomials, Products, Adjoints and Self-Adjoint Parts §adjoint, so the variables lie in the self-adjoint part by Noncommutative Polynomials Form a Unital Complex Algebra with Involution: Linear Extension from Monomials, Products, Adjoints and Self-Adjoint Parts §self-adjoint. The inner product is linear in its second and conjugate-linear in its first argument by Complex Hilbert Spaces and Bounded Linear Maps: Standing Notation §spaces; for a bounded linear map TT with adjoint T∗T^{*} we have ⟨T∗ξ,ξ′⟩=⟨ξ,Tξ′⟩\langle T^{*}\xi,\xi'\rangle=\langle\xi,T\xi'\rangle, hence also ⟨Tξ′,ξ⟩=⟨ξ′,T∗ξ⟩\langle T\xi',\xi\rangle=\langle\xi',T^{*}\xi\rangle by conjugate symmetry. For k∈Nk\in\mathbb{N}, a real r>0r>0, λ∈Σk,r\lambda\in\Sigma_{k,r} and p,q∈Pkp,q\in\mathcal{P}_{k}, Left and Right Multiplication Operators on the Complex GNS Space of a Noncommutative Law: Boundedness, Algebra Rules, Adjoints, Commutation, the Vacuum and the Conjugation §vacuum gives

LpΩλ=p^ λ,⟨p^ λ,q^ λ⟩=λ(p∗q),⟨Ωλ,LpΩλ⟩=λ(p);L_{p}\Omega_{\lambda}=\widehat{p}^{\,\lambda},\qquad\langle\widehat{p}^{\,\lambda},\widehat{q}^{\,\lambda}\rangle=\lambda(p^{*}q),\qquad\langle\Omega_{\lambda},L_{p}\Omega_{\lambda}\rangle=\lambda(p);

for T∈MλT\in\mathcal{M}_{\lambda}, τλ(T)=⟨Ωλ,TΩλ⟩\tau_{\lambda}(T)=\langle\Omega_{\lambda},T\Omega_{\lambda}\rangle by The Tracial Algebra of a Noncommutative Law and Its Trace §trace; τλ\tau_{\lambda} is linear with τλ(Lp)=λ(p)\tau_{\lambda}(L_{p})=\lambda(p) by The Tracial Algebra of a Noncommutative Law: a Norm-Closed Unital *-Algebra with a Faithful Positive Trace, Determined by Vacuum Vectors, Closed under Square Roots §trace; and Mλ\mathcal{M}_{\lambda} contains II and every LpL_{p} and is closed under sums, complex multiples, composition and adjoints by The Tracial Algebra of a Noncommutative Law: a Norm-Closed Unital *-Algebra with a Faithful Positive Trace, Determined by Vacuum Vectors, Closed under Square Roots §star-algebra. For a tuple TT of bounded operators, evaluation p↦p(T)p\mapsto p(T) is linear by Evaluation of Noncommutative Polynomials at a Tuple of Bounded Operators §evaluation, multiplicative by Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §homomorphism, and 1(T)=I1(T)=I, xi(T)=Tix_{i}(T)=T_{i} by Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §values.

Step 0 (Norm bounds, the marginal isometry and the barycentric vectors). By Noncommutative Laws of Finitely Many Self-Adjoint Variables with a Norm Bound §law there are reals R1,R2>0R_{1},R_{2}>0 with μ∈Σd,R1\mu\in\Sigma_{d,R_{1}} and ν∈Σd,R2\nu\in\Sigma_{d,R_{2}}; with R=max⁡{R1,R2}R=\max\{R_{1},R_{2}\}, Basic Properties of Noncommutative Laws: Adjoints, the Self-Adjoint Pairing, Cauchy-Schwarz, Monotonicity in the Bound, and Laws of Constant Tuples §monotone gives μ,ν∈Σd,R\mu,\nu\in\Sigma_{d,R}. We fix this RR.

Fix γ∈Π(μ,ν)\gamma\in\Pi(\mu,\nu) and abbreviate zj=zjγz_{j}=z^{\gamma}_{j}, z=(z1,…,zd)z=(z_{1},\dots,z_{d}), ζ=ζγ\zeta=\zeta_{\gamma}, β=βγ\beta=\beta_{\gamma} and E=EγE=E_{\gamma}; the operators LpL_{p} act on Hγ\mathcal{H}_{\gamma} for p∈P2dp\in\mathcal{P}_{2d} and on Hμ\mathcal{H}_{\mu} for p∈Pdp\in\mathcal{P}_{d}, as in the statement. By Couplings of Noncommutative Laws: the Norm Bound, the Cost Identity, the Tensor, Diagonal and Swapped Couplings, Weak-Star Closedness, and Displacement Interpolants §bound, γ∈Σ2d,R\gamma\in\Sigma_{2d,R}, so γ∈Σ2d\gamma\in\Sigma_{2d} by Noncommutative Laws of Finitely Many Self-Adjoint Variables with a Norm Bound §law. Let a=(x1,…,xd)a=(x_{1},\dots,x_{d}) and a′=(xd+1,…,x2d)a'=(x_{d+1},\dots,x_{2d}), dd-tuples in P2d,sa\mathcal{P}_{2d,\mathrm{sa}}. By Couplings of Two Noncommutative Laws and Their Quadratic Cost §marginals, σa=ι1\sigma_{a}=\iota^{1} and σa′=ι2\sigma_{a'}=\iota^{2}, so Couplings of Two Noncommutative Laws and Their Quadratic Cost §coupling gives γ∘σa=μ\gamma\circ\sigma_{a}=\mu and γ∘σa′=ν\gamma\circ\sigma_{a'}=\nu; and by Substitution is the Unique Unital Homomorphism with Prescribed Values on the Variables: Monomials, Products, Adjoints and Composition §values, σa(xj)=xj\sigma_{a}(x_{j})=x_{j} and σa′(xj)=xd+j\sigma_{a'}(x_{j})=x_{d+j} for j∈[d]j\in[d].

Let V∈L(Hμ,Hγ)V\in\mathcal{L}(\mathcal{H}_{\mu},\mathcal{H}_{\gamma}) and π:Mμ→Mγ\pi:\mathcal{M}_{\mu}\to\mathcal{M}_{\gamma} be the isometry and the embedding of Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation for γ\gamma and aa. By Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation §isometry, V∗V=IV^{*}V=I, VΩμ=ΩγV\Omega_{\mu}=\Omega_{\gamma} and Vp^ μ=ι1(p)^ γV\widehat{p}^{\,\mu}=\widehat{\iota^{1}(p)}^{\,\gamma} for p∈Pdp\in\mathcal{P}_{d}; by Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation §embedding, π(T)∈Mγ\pi(T)\in\mathcal{M}_{\gamma}, π(T)Ωγ=VTΩμ\pi(T)\Omega_{\gamma}=VT\Omega_{\mu} and ∥π(T)∥op≤∥T∥op\lVert\pi(T)\rVert_{\mathrm{op}}\le\lVert T\rVert_{\mathrm{op}} for T∈MμT\in\mathcal{M}_{\mu}; by Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation §homomorphism, π\pi is linear, π(I)=I\pi(I)=I, π(ST)=π(S)π(T)\pi(ST)=\pi(S)\pi(T), π(T∗)=π(T)∗\pi(T^{*})=\pi(T)^{*} and τγ(π(T))=τμ(T)\tau_{\gamma}(\pi(T))=\tau_{\mu}(T); and by Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation §expectation, E(b)=V∗bVE(b)=V^{*}bV and ∥E(b)∥op≤∥b∥op\lVert E(b)\rVert_{\mathrm{op}}\le\lVert b\rVert_{\mathrm{op}} for b∈Mγb\in\mathcal{M}_{\gamma}. For j∈[d]j\in[d] put yj=xd+j^ γ=Lxd+jΩγ∈Hγy_{j}=\widehat{x_{d+j}}^{\,\gamma}=L_{x_{d+j}}\Omega_{\gamma}\in\mathcal{H}_{\gamma}. Then

zjΩμ=V∗Lxd+jVΩμ=V∗Lxd+jΩγ=V∗yj.(0.1)z_{j}\Omega_{\mu}=V^{*}L_{x_{d+j}}V\Omega_{\mu}=V^{*}L_{x_{d+j}}\Omega_{\gamma}=V^{*}y_{j}.\qquad(0.1)

Moreover, for every y∈Hγy\in\mathcal{H}_{\gamma} we have ∥V∗y∥≤∥y∥\lVert V^{*}y\rVert\le\lVert y\rVert: indeed ∥VV∗y∥2=⟨V∗y,V∗VV∗y⟩=∥V∗y∥2\lVert VV^{*}y\rVert^{2}=\langle V^{*}y,V^{*}VV^{*}y\rangle=\lVert V^{*}y\rVert^{2} because V∗V=IV^{*}V=I, and by Cauchy-Schwarz Inequality in a Complex Inner Product Space

∥V∗y∥2=⟨V∗y,V∗y⟩=∣⟨VV∗y,y⟩∣≤∥VV∗y∥ ∥y∥=∥V∗y∥ ∥y∥,(0.2)\lVert V^{*}y\rVert^{2}=\langle V^{*}y,V^{*}y\rangle=\bigl|\langle VV^{*}y,y\rangle\bigr|\le\lVert VV^{*}y\rVert\,\lVert y\rVert=\lVert V^{*}y\rVert\,\lVert y\rVert,\qquad(0.2)

so ∥V∗y∥≤∥y∥\lVert V^{*}y\rVert\le\lVert y\rVert (trivially if V∗y=0V^{*}y=0, and after division by ∥V∗y∥\lVert V^{*}y\rVert otherwise).

Step 1 (Claim 1, coupling). Let S>0S>0 be real with ν∈Σd,S\nu\in\Sigma_{d,S}. Let π′:Mν→Mγ\pi':\mathcal{M}_{\nu}\to\mathcal{M}_{\gamma} be the embedding of Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation for γ\gamma and a′a' (that lemma's μ\mu being our ν=γ∘σa′\nu=\gamma\circ\sigma_{a'}), and write LpνL^{\nu}_{p} for the multiplication operators on Hν\mathcal{H}_{\nu}. By Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation §homomorphism, π′(Lxjν)=Lσa′(xj)=Lxd+j\pi'(L^{\nu}_{x_{j}})=L_{\sigma_{a'}(x_{j})}=L_{x_{d+j}}; by Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation §embedding, π′\pi' does not increase operator norms; and by Left and Right Multiplication Operators on the Complex GNS Space of a Noncommutative Law: Boundedness, Algebra Rules, Adjoints, Commutation, the Vacuum and the Conjugation §multiplication for ν∈Σd,S\nu\in\Sigma_{d,S}, ∥Lxjν∥op≤S\lVert L^{\nu}_{x_{j}}\rVert_{\mathrm{op}}\le S. With the contraction property of EE from Step 0,

∥zj∥op=∥E(Lxd+j)∥op≤∥Lxd+j∥op=∥π′(Lxjν)∥op≤∥Lxjν∥op≤S(j∈[d]).(1.1)\lVert z_{j}\rVert_{\mathrm{op}}=\lVert E(L_{x_{d+j}})\rVert_{\mathrm{op}}\le\lVert L_{x_{d+j}}\rVert_{\mathrm{op}}=\lVert\pi'(L^{\nu}_{x_{j}})\rVert_{\mathrm{op}}\le\lVert L^{\nu}_{x_{j}}\rVert_{\mathrm{op}}\le S\qquad(j\in[d]).\qquad(1.1)

By the statement, zz is a self-adjoint dd-tuple in Mμ\mathcal{M}_{\mu}, and (Hμ,Mμ,Ωμ)(\mathcal{H}_{\mu},\mathcal{M}_{\mu},\Omega_{\mu}) is a tracial W*-probability space by The Tracial Algebra of a Noncommutative Law is a Tracial W*-Probability Space: the W*-Closure of the Left Multiplications §w-star. Hence Laws of Self-Adjoint Tuples in a Tracial W*-Probability Space: Moments, Affine Images, Couplings, Embeddings and L^2 Approximation §law and (1.1) give ζ=law(z)∈Σd,S\zeta=\mathrm{law}(z)\in\Sigma_{d,S}. This is the last sentence of claim 1; with S=RS=R it gives ζ∈Σd,R\zeta\in\Sigma_{d,R}, and so ζ∈Σd\zeta\in\Sigma_{d} by Noncommutative Laws of Finitely Many Self-Adjoint Variables with a Norm Bound §law.

Let Lx=(Lx1,…,Lxd)L_{x}=(L_{x_{1}},\dots,L_{x_{d}}) on Hμ\mathcal{H}_{\mu}, a self-adjoint dd-tuple in Mμ\mathcal{M}_{\mu} (statement). By The Tracial Algebra of a Noncommutative Law is a Tracial W*-Probability Space: the W*-Closure of the Left Multiplications §law, ⟨Ωμ,p(Lx)Ωμ⟩=μ(p)\langle\Omega_{\mu},p(L_{x})\Omega_{\mu}\rangle=\mu(p) for p∈Pdp\in\mathcal{P}_{d}, so by Self-Adjoint Tuples in a Tracial W*-Probability Space and Their Laws §law the law of LxL_{x} is μ\mu. By Laws of Self-Adjoint Tuples in a Tracial W*-Probability Space: Moments, Affine Images, Couplings, Embeddings and L^2 Approximation §coupling, applied to the tuples LxL_{x} and zz, the law β\beta of s=(Lx1,…,Lxd,z1,…,zd)s=(L_{x_{1}},\dots,L_{x_{d}},z_{1},\dots,z_{d}) belongs to Π(μ,ζ)\Pi(\mu,\zeta). This proves claim 1.

Step 2 (Claim 2, orthogonal splitting). Let j∈[d]j\in[d]. Since xj∗=xjx_{j}^{*}=x_{j} and xj^ γ=ι1(xj)^ γ=Vxj^ μ\widehat{x_{j}}^{\,\gamma}=\widehat{\iota^{1}(x_{j})}^{\,\gamma}=V\widehat{x_{j}}^{\,\mu} (Step 0), the vacuum identities for γ∈Σ2d,R\gamma\in\Sigma_{2d,R} and μ∈Σd,R\mu\in\Sigma_{d,R} and (0.1) give

γ(xjxd+j)=⟨xj^ γ,xd+j^ γ⟩=⟨Vxj^ μ,yj⟩=⟨xj^ μ,V∗yj⟩=⟨LxjΩμ,zjΩμ⟩.\gamma(x_{j}x_{d+j})=\langle\widehat{x_{j}}^{\,\gamma},\widehat{x_{d+j}}^{\,\gamma}\rangle=\langle V\widehat{x_{j}}^{\,\mu},y_{j}\rangle=\langle\widehat{x_{j}}^{\,\mu},V^{*}y_{j}\rangle=\langle L_{x_{j}}\Omega_{\mu},z_{j}\Omega_{\mu}\rangle.

By Laws of Self-Adjoint Tuples in a Tracial W*-Probability Space: Moments, Affine Images, Couplings, Embeddings and L^2 Approximation §moments for the 2d2d-tuple ss of Step 1 (with indices jj and d+jd+j), β(xjxd+j)=⟨LxjΩμ,zjΩμ⟩\beta(x_{j}x_{d+j})=\langle L_{x_{j}}\Omega_{\mu},z_{j}\Omega_{\mu}\rangle. Hence

γ(xjxd+j)=β(xjxd+j)(j∈[d]).(2.1)\gamma(x_{j}x_{d+j})=\beta(x_{j}x_{d+j})\qquad(j\in[d]).\qquad(2.1)

Next, ι2(xjxj)=xd+jxd+j\iota^{2}(x_{j}x_{j})=x_{d+j}x_{d+j} by Substitution is the Unique Unital Homomorphism with Prescribed Values on the Variables: Monomials, Products, Adjoints and Composition §homomorphism and Substitution is the Unique Unital Homomorphism with Prescribed Values on the Variables: Monomials, Products, Adjoints and Composition §values, so with γ∘ι2=ν\gamma\circ\iota^{2}=\nu and the vacuum identities, ν(xj2)=γ(xd+jxd+j)=∥yj∥2\nu(x_{j}^{2})=\gamma(x_{d+j}x_{d+j})=\lVert y_{j}\rVert^{2}. By Laws of Self-Adjoint Tuples in a Tracial W*-Probability Space: Moments, Affine Images, Couplings, Embeddings and L^2 Approximation §moments for the tuple zz and (0.1), ζ(xj2)=∥zjΩμ∥2=∥V∗yj∥2\zeta(x_{j}^{2})=\lVert z_{j}\Omega_{\mu}\rVert^{2}=\lVert V^{*}y_{j}\rVert^{2}. Summing over jj, with MM as in Noncommutative Laws, Couplings and the Wasserstein Distance: Standing Notation §laws,

M(ν)=∑j=1d∥yj∥2,M(ζ)=∑j=1d∥V∗yj∥2,(2.2)M(\nu)=\sum_{j=1}^{d}\lVert y_{j}\rVert^{2},\qquad M(\zeta)=\sum_{j=1}^{d}\lVert V^{*}y_{j}\rVert^{2},\qquad(2.2)

and (0.2) gives M(ζ)≤M(ν)M(\zeta)\le M(\nu). Finally, Couplings of Noncommutative Laws: the Norm Bound, the Cost Identity, the Tensor, Diagonal and Swapped Couplings, Weak-Star Closedness, and Displacement Interpolants §cost, applied to γ∈Π(μ,ν)\gamma\in\Pi(\mu,\nu) and to β∈Π(μ,ζ)\beta\in\Pi(\mu,\zeta) (with μ,ζ∈Σd\mu,\zeta\in\Sigma_{d} by Step 1), gives

I(γ)=M(μ)+M(ν)−2∑j=1dγ(xjxd+j),I(β)=M(μ)+M(ζ)−2∑j=1dβ(xjxd+j),I(\gamma)=M(\mu)+M(\nu)-2\sum_{j=1}^{d}\gamma(x_{j}x_{d+j}),\qquad I(\beta)=M(\mu)+M(\zeta)-2\sum_{j=1}^{d}\beta(x_{j}x_{d+j}),

and subtracting, by (2.1), I(γ)=I(β)+M(ν)−M(ζ)I(\gamma)=I(\beta)+M(\nu)-M(\zeta). This proves claim 2.

Step 3 (A second coupling of ζ\zeta and ν\nu). For j∈[d]j\in[d] let gj=π(zj)g_{j}=\pi(z_{j}), and let s′=(g1,…,gd,Lxd+1,…,Lx2d)s'=(g_{1},\dots,g_{d},L_{x_{d+1}},\dots,L_{x_{2d}}), a 2d2d-tuple of operators on Hγ\mathcal{H}_{\gamma}. By Step 0, gj∈Mγg_{j}\in\mathcal{M}_{\gamma}, gj∗=π(zj∗)=π(zj)=gjg_{j}^{*}=\pi(z_{j}^{*})=\pi(z_{j})=g_{j} and ∥gj∥op≤∥zj∥op≤R\lVert g_{j}\rVert_{\mathrm{op}}\le\lVert z_{j}\rVert_{\mathrm{op}}\le R by (1.1) with S=RS=R. The operators Lxd+jL_{x_{d+j}} are self-adjoint elements of Mγ\mathcal{M}_{\gamma} (statement) with ∥Lxd+j∥op≤R\lVert L_{x_{d+j}}\rVert_{\mathrm{op}}\le R by (1.1). So s′s' is a self-adjoint 2d2d-tuple in the tracial W*-probability space (Hγ,Mγ,Ωγ)(\mathcal{H}_{\gamma},\mathcal{M}_{\gamma},\Omega_{\gamma}) of The Tracial Algebra of a Noncommutative Law is a Tracial W*-Probability Space: the W*-Closure of the Left Multiplications §w-star; let η\eta be its law, η(p)=⟨Ωγ,p(s′)Ωγ⟩\eta(p)=\langle\Omega_{\gamma},p(s')\Omega_{\gamma}\rangle. By Laws of Self-Adjoint Tuples in a Tracial W*-Probability Space: Moments, Affine Images, Couplings, Embeddings and L^2 Approximation §law, η∈Σ2d,R\eta\in\Sigma_{2d,R}.

Let g=(g1,…,gd)g=(g_{1},\dots,g_{d}) and h=(Lxd+1,…,Lx2d)h=(L_{x_{d+1}},\dots,L_{x_{2d}}). For q∈Pdq\in\mathcal{P}_{d}, Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §transport, applied with Hμ\mathcal{H}_{\mu}, Hγ\mathcal{H}_{\gamma}, dd, zz, Mμ\mathcal{M}_{\mu}, π\pi in place of its HH, KK, nn, TT, A\mathcal{A}, Φ\Phi (the set Mμ\mathcal{M}_{\mu} contains II and every zjz_{j} and is closed under sums, complex multiples and composition, and π\pi is linear, multiplicative and unital, by Step 0), gives

q(z)∈Mμandq(g)=π(q(z)).(3.1)q(z)\in\mathcal{M}_{\mu}\qquad\text{and}\qquad q(g)=\pi\bigl(q(z)\bigr).\qquad(3.1)

Hence ⟨Ωγ,q(g)Ωγ⟩=τγ(π(q(z)))=τμ(q(z))=⟨Ωμ,q(z)Ωμ⟩=ζ(q)\langle\Omega_{\gamma},q(g)\Omega_{\gamma}\rangle=\tau_{\gamma}(\pi(q(z)))=\tau_{\mu}(q(z))=\langle\Omega_{\mu},q(z)\Omega_{\mu}\rangle=\zeta(q), so the law of gg is ζ\zeta. For the second block, write Lxγ=(Lx1,…,Lx2d)L^{\gamma}_{x}=(L_{x_{1}},\dots,L_{x_{2d}}); then a′(Lxγ)=ha'(L^{\gamma}_{x})=h by Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §values, so Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §substitution and Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §gns (for γ∈Σ2d\gamma\in\Sigma_{2d}) give q(h)=(ι2(q))(Lxγ)=Lι2(q)q(h)=(\iota^{2}(q))(L^{\gamma}_{x})=L_{\iota^{2}(q)}, and ⟨Ωγ,q(h)Ωγ⟩=γ(ι2(q))=ν(q)\langle\Omega_{\gamma},q(h)\Omega_{\gamma}\rangle=\gamma(\iota^{2}(q))=\nu(q). So the law of hh is ν\nu, and Laws of Self-Adjoint Tuples in a Tracial W*-Probability Space: Moments, Affine Images, Couplings, Embeddings and L^2 Approximation §coupling, applied to the tuples gg and hh, gives η∈Π(ζ,ν)\eta\in\Pi(\zeta,\nu).

Step 4 (A martingale identity for η\eta). We show

η(ι1(q) xd+j)=η(ι1(q) xj)(q∈Pd, j∈[d]).(4.1)\eta\bigl(\iota^{1}(q)\,x_{d+j}\bigr)=\eta\bigl(\iota^{1}(q)\,x_{j}\bigr)\qquad(q\in\mathcal{P}_{d},\ j\in[d]).\qquad(4.1)

Let k∈[2d]k\in[2d]. By the homomorphism and value rules of evaluation, (ι1(q)xk)(s′)=(ι1(q))(s′) sk′(\iota^{1}(q)x_{k})(s')=(\iota^{1}(q))(s')\,s'_{k}; and a(s′)=ga(s')=g by Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §values, so Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §substitution and (3.1) give (ι1(q))(s′)=q(g)=π(q(z))(\iota^{1}(q))(s')=q(g)=\pi(q(z)). With π(q(z))∗=π(q(z)∗)\pi(q(z))^{*}=\pi(q(z)^{*}) and π(q(z)∗)Ωγ=Vq(z)∗Ωμ\pi(q(z)^{*})\Omega_{\gamma}=Vq(z)^{*}\Omega_{\mu} (Step 0; q(z)∗∈Mμq(z)^{*}\in\mathcal{M}_{\mu}), we get

η(ι1(q)xk)=⟨Ωγ,π(q(z)) sk′Ωγ⟩=⟨Vq(z)∗Ωμ,sk′Ωγ⟩.\eta\bigl(\iota^{1}(q)x_{k}\bigr)=\langle\Omega_{\gamma},\pi(q(z))\,s'_{k}\Omega_{\gamma}\rangle=\langle Vq(z)^{*}\Omega_{\mu},s'_{k}\Omega_{\gamma}\rangle.

For k=d+jk=d+j, sk′Ωγ=yjs'_{k}\Omega_{\gamma}=y_{j} and, by (0.1), ⟨Vq(z)∗Ωμ,yj⟩=⟨q(z)∗Ωμ,V∗yj⟩=⟨q(z)∗Ωμ,zjΩμ⟩\langle Vq(z)^{*}\Omega_{\mu},y_{j}\rangle=\langle q(z)^{*}\Omega_{\mu},V^{*}y_{j}\rangle=\langle q(z)^{*}\Omega_{\mu},z_{j}\Omega_{\mu}\rangle. For k=jk=j, sj′Ωγ=π(zj)Ωγ=VzjΩμs'_{j}\Omega_{\gamma}=\pi(z_{j})\Omega_{\gamma}=Vz_{j}\Omega_{\mu} and ⟨Vq(z)∗Ωμ,VzjΩμ⟩=⟨q(z)∗Ωμ,V∗VzjΩμ⟩=⟨q(z)∗Ωμ,zjΩμ⟩\langle Vq(z)^{*}\Omega_{\mu},Vz_{j}\Omega_{\mu}\rangle=\langle q(z)^{*}\Omega_{\mu},V^{*}Vz_{j}\Omega_{\mu}\rangle=\langle q(z)^{*}\Omega_{\mu},z_{j}\Omega_{\mu}\rangle. The two values agree, proving (4.1).

Step 5 (Gluing freely over ζ\zeta: the data). By Step 1, ζ,μ∈Σd,R\zeta,\mu\in\Sigma_{d,R}, so The Noncommutative Wasserstein Distance: Existence of Optimal Couplings, Symmetry, Separation, a Moment Bound, Weak-Star Lower Semicontinuity, and Displacement Interpolation §attained gives a coupling α∈Π(ζ,μ)\alpha\in\Pi(\zeta,\mu) that is optimal, i.e. by The Noncommutative Quadratic Wasserstein Distance and Optimal Couplings §optimal and The Noncommutative Wasserstein Distance: Existence of Optimal Couplings, Symmetry, Separation, a Moment Bound, Weak-Star Lower Semicontinuity, and Displacement Interpolation §symmetry,

I(α)=W2(ζ,μ)2=W2(μ,ζ)2.(5.1)I(\alpha)=W_{2}(\zeta,\mu)^{2}=W_{2}(\mu,\zeta)^{2}.\qquad(5.1)

By Couplings of Noncommutative Laws: the Norm Bound, the Cost Identity, the Tensor, Diagonal and Swapped Couplings, Weak-Star Closedness, and Displacement Interpolants §bound, α∈Σ2d,R\alpha\in\Sigma_{2d,R}; also η∈Σ2d,R\eta\in\Sigma_{2d,R} (Step 3). Put m1=m2=2dm_{1}=m_{2}=2d, n=dn=d, γ1=α\gamma_{1}=\alpha, γ2=η\gamma_{2}=\eta and a1=a2=aa^{1}=a^{2}=a. Since σa=ι1\sigma_{a}=\iota^{1}, Couplings of Two Noncommutative Laws and Their Quadratic Cost §coupling gives α∘σa=ζ=η∘σa\alpha\circ\sigma_{a}=\zeta=\eta\circ\sigma_{a}, so the data of Two Noncommutative Laws with a Common Marginal: Standing Notation for Their Amalgamated Free Product §data hold. The letters clash: the common marginal, which that setting (and every result stated in it) calls μ\mu, is here ζ\zeta; our μ\mu plays no role in the amalgamation. With this renaming we use the notation of its algebras and its embeddings and expectations: N=MζN=\mathcal{M}_{\zeta} on Hζ\mathcal{H}_{\zeta}, A1=MαA_{1}=\mathcal{M}_{\alpha} on Hα\mathcal{H}_{\alpha}, A2=MηA_{2}=\mathcal{M}_{\eta} on Hη\mathcal{H}_{\eta}, πε:N→Aε\pi_{\varepsilon}:N\to A_{\varepsilon} and Eε:Aε→NE_{\varepsilon}:A_{\varepsilon}\to N; let VεV_{\varepsilon} be the isometry of Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation §isometry for γε\gamma_{\varepsilon} and aa. Let H\mathcal{H}, Ω\Omega, φ\varphi and Λε\Lambda_{\varepsilon} be the amalgamated free product space, its vacuum vector, its vacuum state and the actions, as in The Amalgamated Free Product of Two Noncommutative Laws over a Common Marginal: Representations, Amalgamation, Freeness and the Tracial Vacuum State; by The Amalgamated Free Product Space of Two Noncommutative Laws over a Common Marginal, Its Vacuum Vector and Vacuum State §state, φ(X)=⟨Ω,XΩ⟩H\varphi(X)=\langle\Omega,X\Omega\rangle_{\mathcal{H}}, so φ\varphi is linear. Write Lq1L^{1}_{q}, Lq2L^{2}_{q} for the multiplication operators on Hα\mathcal{H}_{\alpha}, Hη\mathcal{H}_{\eta} (q∈P2dq\in\mathcal{P}_{2d}), and LqζL^{\zeta}_{q} for those on Hζ\mathcal{H}_{\zeta} (q∈Pdq\in\mathcal{P}_{d}). They lie in A1A_{1}, A2A_{2}, NN respectively (preliminaries); by Left and Right Multiplication Operators on the Complex GNS Space of a Noncommutative Law: Boundedness, Algebra Rules, Adjoints, Commutation, the Vacuum and the Conjugation §multiplication and Left and Right Multiplication Operators on the Complex GNS Space of a Noncommutative Law: Boundedness, Algebra Rules, Adjoints, Commutation, the Vacuum and the Conjugation §adjoint for α,η∈Σ2d,R\alpha,\eta\in\Sigma_{2d,R}, ∥Lxiε∥op≤R\lVert L^{\varepsilon}_{x_{i}}\rVert_{\mathrm{op}}\le R and (Lxiε)∗=Lxiε(L^{\varepsilon}_{x_{i}})^{*}=L^{\varepsilon}_{x_{i}} for i∈[2d]i\in[2d]; and by Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation §homomorphism and σa(xj)=xj\sigma_{a}(x_{j})=x_{j},

πε(Lxjζ)=Lxjε(ε∈{1,2}, j∈[d]).(5.2)\pi_{\varepsilon}(L^{\zeta}_{x_{j}})=L^{\varepsilon}_{x_{j}}\qquad(\varepsilon\in\{1,2\},\ j\in[d]).\qquad(5.2)

We use repeatedly, for ε∈{1,2}\varepsilon\in\{1,2\}, b∈Aεb\in A_{\varepsilon} and X,Y∈NX,Y\in N: πε\pi_{\varepsilon} is linear and multiplicative with πε(I)=I\pi_{\varepsilon}(I)=I and τγε(πε(X))=τζ(X)\tau_{\gamma_{\varepsilon}}(\pi_{\varepsilon}(X))=\tau_{\zeta}(X) (Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation §homomorphism); Eε(πε(X) b πε(Y))=X Eε(b) YE_{\varepsilon}(\pi_{\varepsilon}(X)\,b\,\pi_{\varepsilon}(Y))=X\,E_{\varepsilon}(b)\,Y, τζ(Eε(b))=τγε(b)\tau_{\zeta}(E_{\varepsilon}(b))=\tau_{\gamma_{\varepsilon}}(b) and ⟨ξ,Eε(b)ξ′⟩=⟨Vεξ,bVεξ′⟩\langle\xi,E_{\varepsilon}(b)\xi'\rangle=\langle V_{\varepsilon}\xi,bV_{\varepsilon}\xi'\rangle for ξ,ξ′∈Hζ\xi,\xi'\in\mathcal{H}_{\zeta} (Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation §expectation); Λε\Lambda_{\varepsilon} is linear and multiplicative with Λε(b∗)=Λε(b)∗\Lambda_{\varepsilon}(b^{*})=\Lambda_{\varepsilon}(b)^{*} and ∥Λε(b)∥op≤∥b∥op\lVert\Lambda_{\varepsilon}(b)\rVert_{\mathrm{op}}\le\lVert b\rVert_{\mathrm{op}} (The Amalgamated Free Product of Two Noncommutative Laws over a Common Marginal: Representations, Amalgamation, Freeness and the Tracial Vacuum State §representation); Λ1(π1(X))=Λ2(π2(X))\Lambda_{1}(\pi_{1}(X))=\Lambda_{2}(\pi_{2}(X)) (The Amalgamated Free Product of Two Noncommutative Laws over a Common Marginal: Representations, Amalgamation, Freeness and the Tracial Vacuum State §amalgamation); and φ(Λε(b))=τγε(b)\varphi(\Lambda_{\varepsilon}(b))=\tau_{\gamma_{\varepsilon}}(b), ∥Ω∥H=1\lVert\Omega\rVert_{\mathcal{H}}=1 (The Amalgamated Free Product of Two Noncommutative Laws over a Common Marginal: Representations, Amalgamation, Freeness and the Tracial Vacuum State §marginals).

Step 6 (The expectation of the ν\nu-variables of η\eta). We show E2(Lxd+j2)=LxjζE_{2}(L^{2}_{x_{d+j}})=L^{\zeta}_{x_{j}} for j∈[d]j\in[d]. Let q∈Pdq\in\mathcal{P}_{d}. By Marginals of a Noncommutative Law: the Isometry of GNS Spaces, the Trace-Preserving Embedding of Tracial Algebras and the Conditional Expectation §isometry, V2q^ ζ=ι1(q)^ ηV_{2}\widehat{q}^{\,\zeta}=\widehat{\iota^{1}(q)}^{\,\eta} and V2Ωζ=ΩηV_{2}\Omega_{\zeta}=\Omega_{\eta}; by Substitution is the Unique Unital Homomorphism with Prescribed Values on the Variables: Monomials, Products, Adjoints and Composition §adjoint, Substitution is the Unique Unital Homomorphism with Prescribed Values on the Variables: Monomials, Products, Adjoints and Composition §homomorphism and Substitution is the Unique Unital Homomorphism with Prescribed Values on the Variables: Monomials, Products, Adjoints and Composition §values, ι1(q)∗=ι1(q∗)\iota^{1}(q)^{*}=\iota^{1}(q^{*}) and ι1(q∗)xj=ι1(q∗xj)\iota^{1}(q^{*})x_{j}=\iota^{1}(q^{*}x_{j}). With the inner-product rule for E2E_{2} of Step 5, the vacuum identities for η∈Σ2d,R\eta\in\Sigma_{2d,R} and ζ∈Σd,R\zeta\in\Sigma_{d,R}, (4.1) for q∗q^{*}, and η∘ι1=ζ\eta\circ\iota^{1}=\zeta (Step 3),

⟨q^ ζ,E2(Lxd+j2)Ωζ⟩=⟨ι1(q)^ η,xd+j^ η⟩=η(ι1(q∗)xd+j)=η(ι1(q∗)xj)=ζ(q∗xj)=⟨q^ ζ,LxjζΩζ⟩.\langle\widehat{q}^{\,\zeta},E_{2}(L^{2}_{x_{d+j}})\Omega_{\zeta}\rangle=\langle\widehat{\iota^{1}(q)}^{\,\eta},\widehat{x_{d+j}}^{\,\eta}\rangle=\eta\bigl(\iota^{1}(q^{*})x_{d+j}\bigr)=\eta\bigl(\iota^{1}(q^{*})x_{j}\bigr)=\zeta(q^{*}x_{j})=\langle\widehat{q}^{\,\zeta},L^{\zeta}_{x_{j}}\Omega_{\zeta}\rangle.

So D=E2(Lxd+j2)−Lxjζ∈ND=E_{2}(L^{2}_{x_{d+j}})-L^{\zeta}_{x_{j}}\in N satisfies ⟨q^ ζ,DΩζ⟩=0\langle\widehat{q}^{\,\zeta},D\Omega_{\zeta}\rangle=0 for every q∈Pdq\in\mathcal{P}_{d}. The classes are the image of Pd\mathcal{P}_{d} under the canonical map of the completion Hζ\mathcal{H}_{\zeta} (The Complex GNS Space of a Tracial State on Noncommutative Polynomials §gns and The Complex GNS Space of a Tracial State on Noncommutative Polynomials §classes), which is dense by The Complex Hilbert Completion is a Complex Hilbert Space Containing a Dense Isometric Image, and Bounded Complex-Linear Maps Extend to It §dense; choosing classes qk^ ζ\widehat{q_{k}}^{\,\zeta} converging to DΩζD\Omega_{\zeta}, continuity of the inner product gives ∥DΩζ∥2=lim⁡k⟨qk^ ζ,DΩζ⟩=0\lVert D\Omega_{\zeta}\rVert^{2}=\lim_{k}\langle\widehat{q_{k}}^{\,\zeta},D\Omega_{\zeta}\rangle=0. Thus DΩζ=0D\Omega_{\zeta}=0, and D=0D=0 by The Tracial Algebra of a Noncommutative Law: a Norm-Closed Unital *-Algebra with a Faithful Positive Trace, Determined by Vacuum Vectors, Closed under Square Roots §vacuum.

Step 7 (Mixed moments across the free product). We show φ(Λ1(b)Λ2(c))=τζ(E1(b)E2(c))\varphi(\Lambda_{1}(b)\Lambda_{2}(c))=\tau_{\zeta}(E_{1}(b)E_{2}(c)) for b∈A1b\in A_{1} and c∈A2c\in A_{2}. Let X=E1(b)X=E_{1}(b), Y=E2(c)Y=E_{2}(c), and let b∘=b−π1(X)b^{\circ}=b-\pi_{1}(X) and c∘=c−π2(Y)c^{\circ}=c-\pi_{2}(Y) be the centred parts, so b∘∈A1∘b^{\circ}\in A_{1}^{\circ}, c∘∈A2∘c^{\circ}\in A_{2}^{\circ}, i.e. E1(b∘)=0E_{1}(b^{\circ})=0 and E2(c∘)=0E_{2}(c^{\circ})=0. As Λ1,Λ2,φ\Lambda_{1},\Lambda_{2},\varphi are linear and composition is bilinear, φ(Λ1(b)Λ2(c))\varphi(\Lambda_{1}(b)\Lambda_{2}(c)) is the sum of the four terms below, evaluated with the rules of Step 5. First,

φ(Λ1(π1(X))Λ2(π2(Y)))=φ(Λ2(π2(XY)))=τη(π2(XY))=τζ(XY).\varphi\bigl(\Lambda_{1}(\pi_{1}(X))\Lambda_{2}(\pi_{2}(Y))\bigr)=\varphi\bigl(\Lambda_{2}(\pi_{2}(XY))\bigr)=\tau_{\eta}(\pi_{2}(XY))=\tau_{\zeta}(XY).

Second,

φ(Λ1(π1(X))Λ2(c∘))=φ(Λ2(π2(X)c∘))=τη(π2(X)c∘π2(I))=τζ(X E2(c∘))=0.\varphi\bigl(\Lambda_{1}(\pi_{1}(X))\Lambda_{2}(c^{\circ})\bigr)=\varphi\bigl(\Lambda_{2}(\pi_{2}(X)c^{\circ})\bigr)=\tau_{\eta}(\pi_{2}(X)c^{\circ}\pi_{2}(I))=\tau_{\zeta}\bigl(X\,E_{2}(c^{\circ})\bigr)=0.

Third,

φ(Λ1(b∘)Λ2(π2(Y)))=φ(Λ1(b∘π1(Y)))=τα(π1(I)b∘π1(Y))=τζ(E1(b∘) Y)=0.\varphi\bigl(\Lambda_{1}(b^{\circ})\Lambda_{2}(\pi_{2}(Y))\bigr)=\varphi\bigl(\Lambda_{1}(b^{\circ}\pi_{1}(Y))\bigr)=\tau_{\alpha}(\pi_{1}(I)b^{\circ}\pi_{1}(Y))=\tau_{\zeta}\bigl(E_{1}(b^{\circ})\,Y\bigr)=0.

Fourth, ((1,b∘),(2,c∘))((1,b^{\circ}),(2,c^{\circ})) is an alternating tuple of type (1,2)(1,2), so by The Amalgamated Free Product of Two Noncommutative Laws over a Common Marginal: Representations, Amalgamation, Freeness and the Tracial Vacuum State §freeness, Λ1(b∘)Λ2(c∘)Ω=Ξ(b∘,c∘)\Lambda_{1}(b^{\circ})\Lambda_{2}(c^{\circ})\Omega=\Xi(b^{\circ},c^{\circ}) and φ(Λ1(b∘)Λ2(c∘))=⟨Ω,Ξ(b∘,c∘)⟩H=0\varphi(\Lambda_{1}(b^{\circ})\Lambda_{2}(c^{\circ}))=\langle\Omega,\Xi(b^{\circ},c^{\circ})\rangle_{\mathcal{H}}=0. Summing, φ(Λ1(b)Λ2(c))=τζ(XY)\varphi(\Lambda_{1}(b)\Lambda_{2}(c))=\tau_{\zeta}(XY).

Step 8 (The glued coupling ψ\psi). Define f:[2d]→{1,2}f:[2d]\to\{1,2\} and ci∈Af(i)c_{i}\in A_{f(i)} by f(i)=1f(i)=1, ci=Lxd+i1c_{i}=L^{1}_{x_{d+i}} and f(d+i)=2f(d+i)=2, cd+i=Lxd+i2c_{d+i}=L^{2}_{x_{d+i}} for i∈[d]i\in[d], and let T=(T1,…,T2d)T=(T_{1},\dots,T_{2d}) with Ti=Λf(i)(ci)∈L(H)T_{i}=\Lambda_{f(i)}(c_{i})\in\mathcal{L}(\mathcal{H}). By Step 5, Ti∗=Λf(i)(ci∗)=TiT_{i}^{*}=\Lambda_{f(i)}(c_{i}^{*})=T_{i} and ∥Ti∥op≤∥ci∥op≤R\lVert T_{i}\rVert_{\mathrm{op}}\le\lVert c_{i}\rVert_{\mathrm{op}}\le R, and ∥Ω∥H=1\lVert\Omega\rVert_{\mathcal{H}}=1; so The Law of a Tuple of Bounded Self-Adjoint Operators in a Tracial Vector State applies with 2d2d, H\mathcal{H}, TT in place of its nn, HH, TT. Let ψ=λT\psi=\lambda_{T}, i.e. ψ(p)=⟨Ω,p(T)Ω⟩=φ(p(T))\psi(p)=\langle\Omega,p(T)\Omega\rangle=\varphi(p(T)) for p∈P2dp\in\mathcal{P}_{2d}.

To apply The Law of a Tuple of Bounded Self-Adjoint Operators in a Tracial Vector State §law we show φ(TuTv)=φ(TvTu)\varphi(T_{u}T_{v})=\varphi(T_{v}T_{u}) for all words u,v∈W2du,v\in W_{2d}. If u=∅u=\varnothing, then Tu=IT_{u}=I by Evaluation of Noncommutative Polynomials at a Tuple of Bounded Operators §word-products and both sides equal φ(Tv)\varphi(T_{v}); likewise if v=∅v=\varnothing. For nonempty u,vu,v, TuTv=TuvT_{u}T_{v}=T_{uv} and TvTu=TvuT_{v}T_{u}=T_{vu} by Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §values, so it suffices to show φ(Tuv)=φ(Tvu)\varphi(T_{uv})=\varphi(T_{vu}) for all nonempty u,vu,v. First, for a letter (i)(i) and a nonempty word ww of length ll with letters w1,…,wlw_{1},\dots,w_{l}, we have φ(T(i)w)=φ(Tw(i))\varphi(T_{(i)w})=\varphi(T_{w(i)}): indeed T(i)w=TiTwT_{(i)w}=T_{i}T_{w}, Tw(i)=TwTiT_{w(i)}=T_{w}T_{i} and Tw=Tw1⋯TwlT_{w}=T_{w_{1}}\cdots T_{w_{l}} by Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §values and Evaluation of Noncommutative Polynomials at a Tuple of Bounded Operators §word-products, so by associativity of composition T(i)wT_{(i)w} is the product Λf(i)(ci)Λf(w1)(cw1)⋯Λf(wl)(cwl)\Lambda_{f(i)}(c_{i})\Lambda_{f(w_{1})}(c_{w_{1}})\cdots\Lambda_{f(w_{l})}(c_{w_{l}}) of l+1≥2l+1\ge2 factors and Tw(i)T_{w(i)} is the product of the same factors with the first one moved to the end, and the claim is The Amalgamated Free Product of Two Noncommutative Laws over a Common Marginal: Representations, Amalgamation, Freeness and the Tracial Vacuum State §trace with r=l+1r=l+1. Now induct on the length kk of uu, for all nonempty vv at once. If k=1k=1, uu is a letter by Basic Properties of Words: Associativity, Reversal, Finitely Many Factorisations, and Countability §last-letter and the claim is the rotation just shown. If the claim holds for words of length kk and uu has length k+1k+1, then u=u′(i)u=u'(i) with u′u' of length kk by Basic Properties of Words: Associativity, Reversal, Finitely Many Factorisations, and Countability §last-letter; by Basic Properties of Words: Associativity, Reversal, Finitely Many Factorisations, and Countability §monoid, uv=u′((i)v)uv=u'((i)v), ((i)v)u′=(i)(vu′)((i)v)u'=(i)(vu') with (i)v(i)v and vu′vu' nonempty, and (vu′)(i)=v(u′(i))=vu(vu')(i)=v(u'(i))=vu; so the induction hypothesis (for u′u' and (i)v(i)v) and the rotation (for ii and vu′vu') give φ(Tuv)=φ(T(i)(vu′))=φ(T(vu′)(i))=φ(Tvu)\varphi(T_{uv})=\varphi(T_{(i)(vu')})=\varphi(T_{(vu')(i)})=\varphi(T_{vu}). Hence ψ∈Σ2d,R\psi\in\Sigma_{2d,R} by The Law of a Tuple of Bounded Self-Adjoint Operators in a Tracial Vector State §law; in particular ψ\psi is a tracial state by Noncommutative Laws of Finitely Many Self-Adjoint Variables with a Norm Bound §norm-bound.

Marginals of ψ\psi. Let p∈Pdp\in\mathcal{P}_{d}. As a(T)=(T1,…,Td)a(T)=(T_{1},\dots,T_{d}) by Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §values, Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §substitution gives (ι1(p))(T)=p(T1,…,Td)(\iota^{1}(p))(T)=p(T_{1},\dots,T_{d}). Apply Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §transport with Hα\mathcal{H}_{\alpha}, H\mathcal{H}, dd, (Lxd+11,…,Lx2d1)(L^{1}_{x_{d+1}},\dots,L^{1}_{x_{2d}}), A1A_{1}, Λ1\Lambda_{1} in place of its HH, KK, nn, TT, A\mathcal{A}, Φ\Phi (the set A1A_{1} has the required closure properties by the preliminaries, and Λ1\Lambda_{1} is linear, multiplicative and unital by Step 5): p(T1,…,Td)=Λ1(p(Lxd+11,…,Lx2d1))p(T_{1},\dots,T_{d})=\Lambda_{1}\bigl(p(L^{1}_{x_{d+1}},\dots,L^{1}_{x_{2d}})\bigr). With Lx1=(Lx11,…,Lx2d1)L^{1}_{x}=(L^{1}_{x_{1}},\dots,L^{1}_{x_{2d}}) we have a′(Lx1)=(Lxd+11,…,Lx2d1)a'(L^{1}_{x})=(L^{1}_{x_{d+1}},\dots,L^{1}_{x_{2d}}), so Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §substitution and Evaluation of Noncommutative Polynomials at Bounded Operators: a Unital Homomorphism Compatible with Adjoints, Substitution, Representations and the GNS Multiplication Operators §gns (for α∈Σ2d\alpha\in\Sigma_{2d}) give p(Lxd+11,…,Lx2d1)=(ι2(p))(Lx1)=Lι2(p)1p(L^{1}_{x_{d+1}},\dots,L^{1}_{x_{2d}})=(\iota^{2}(p))(L^{1}_{x})=L^{1}_{\iota^{2}(p)}. Hence, as α∈Π(ζ,μ)\alpha\in\Pi(\zeta,\mu),

ψ(ι1(p))=φ(Λ1(Lι2(p)1))=τα(Lι2(p)1)=α(ι2(p))=μ(p).\psi(\iota^{1}(p))=\varphi\bigl(\Lambda_{1}(L^{1}_{\iota^{2}(p)})\bigr)=\tau_{\alpha}(L^{1}_{\iota^{2}(p)})=\alpha(\iota^{2}(p))=\mu(p).

In the same way, with a′(T)=(Td+1,…,T2d)=(Λ2(Lxd+12),…,Λ2(Lx2d2))a'(T)=(T_{d+1},\dots,T_{2d})=(\Lambda_{2}(L^{2}_{x_{d+1}}),\dots,\Lambda_{2}(L^{2}_{x_{2d}})) and Hη\mathcal{H}_{\eta}, A2A_{2}, Λ2\Lambda_{2}, η\eta in place of Hα\mathcal{H}_{\alpha}, A1A_{1}, Λ1\Lambda_{1}, α\alpha, we get ψ(ι2(p))=τη(Lι2(p)2)=η(ι2(p))=ν(p)\psi(\iota^{2}(p))=\tau_{\eta}(L^{2}_{\iota^{2}(p)})=\eta(\iota^{2}(p))=\nu(p), as η∈Π(ζ,ν)\eta\in\Pi(\zeta,\nu). By Couplings of Two Noncommutative Laws and Their Quadratic Cost §coupling, ψ∈Π(μ,ν)\psi\in\Pi(\mu,\nu).

Step 9 (Claim 3, upper bound). Let j∈[d]j\in[d]. By the homomorphism and value rules of evaluation, Step 7, Step 6, the expectation and trace rules of Step 5 (with π1(I)=I\pi_{1}(I)=I) and (5.2), then Left and Right Multiplication Operators on the Complex GNS Space of a Noncommutative Law: Boundedness, Algebra Rules, Adjoints, Commutation, the Vacuum and the Conjugation §algebra, the identity τα(Lp1)=α(p)\tau_{\alpha}(L^{1}_{p})=\alpha(p) of the preliminaries, and condition (c) of Noncommutative Laws of Finitely Many Self-Adjoint Variables with a Norm Bound §tracial-state for α\alpha,

ψ(xjxd+j)=φ(TjTd+j)=τζ(E1(Lxd+j1) Lxjζ)=τζ(E1(Lxd+j1π1(Lxjζ)))=τα(Lxd+j1Lxj1)=α(xd+jxj)=α(xjxd+j).\psi(x_{j}x_{d+j})=\varphi(T_{j}T_{d+j})=\tau_{\zeta}\bigl(E_{1}(L^{1}_{x_{d+j}})\,L^{\zeta}_{x_{j}}\bigr)=\tau_{\zeta}\bigl(E_{1}(L^{1}_{x_{d+j}}\pi_{1}(L^{\zeta}_{x_{j}}))\bigr)=\tau_{\alpha}(L^{1}_{x_{d+j}}L^{1}_{x_{j}})=\alpha(x_{d+j}x_{j})=\alpha(x_{j}x_{d+j}).

By Couplings of Noncommutative Laws: the Norm Bound, the Cost Identity, the Tensor, Diagonal and Swapped Couplings, Weak-Star Closedness, and Displacement Interpolants §cost, applied to ψ∈Π(μ,ν)\psi\in\Pi(\mu,\nu) and to α∈Π(ζ,μ)\alpha\in\Pi(\zeta,\mu) (with ζ,μ∈Σd\zeta,\mu\in\Sigma_{d}),

I(ψ)=M(μ)+M(ν)−2∑j=1dα(xjxd+j),I(α)=M(ζ)+M(μ)−2∑j=1dα(xjxd+j),I(\psi)=M(\mu)+M(\nu)-2\sum_{j=1}^{d}\alpha(x_{j}x_{d+j}),\qquad I(\alpha)=M(\zeta)+M(\mu)-2\sum_{j=1}^{d}\alpha(x_{j}x_{d+j}),

so I(ψ)=I(α)+M(ν)−M(ζ)I(\psi)=I(\alpha)+M(\nu)-M(\zeta). By The Noncommutative Quadratic Wasserstein Distance and Optimal Couplings §distance, W2(μ,ν)2W_{2}(\mu,\nu)^{2} is the infimum of the costs of couplings of μ\mu and ν\nu, so W2(μ,ν)2≤I(ψ)W_{2}(\mu,\nu)^{2}\le I(\psi), and with (5.1)

W2(μ,ν)2≤W2(μ,ζ)2+M(ν)−M(ζ).(9.1)W_{2}(\mu,\nu)^{2}\le W_{2}(\mu,\zeta)^{2}+M(\nu)-M(\zeta).\qquad(9.1)

As γ∈Π(μ,ν)\gamma\in\Pi(\mu,\nu) was arbitrary, this proves claim 3.

Step 10 (Claim 4, optimal couplings). Let γ\gamma be optimal, so I(γ)=W2(μ,ν)2I(\gamma)=W_{2}(\mu,\nu)^{2} by The Noncommutative Quadratic Wasserstein Distance and Optimal Couplings §optimal. By Laws of Self-Adjoint Tuples in a Tracial W*-Probability Space: Moments, Affine Images, Couplings, Embeddings and L^2 Approximation §coupling for the tuples LxL_{x} and zz of Step 1 (whose laws are μ\mu and ζ\zeta), I(β)≥W2(μ,ζ)2I(\beta)\ge W_{2}(\mu,\zeta)^{2}. With claim 2 and (9.1),

W2(μ,ν)2=I(γ)=I(β)+M(ν)−M(ζ)≥W2(μ,ζ)2+M(ν)−M(ζ)≥W2(μ,ν)2.W_{2}(\mu,\nu)^{2}=I(\gamma)=I(\beta)+M(\nu)-M(\zeta)\ge W_{2}(\mu,\zeta)^{2}+M(\nu)-M(\zeta)\ge W_{2}(\mu,\nu)^{2}.

Hence equality holds throughout: W2(μ,ν)2=W2(μ,ζ)2+M(ν)−M(ζ)W_{2}(\mu,\nu)^{2}=W_{2}(\mu,\zeta)^{2}+M(\nu)-M(\zeta) and I(β)=W2(μ,ζ)2I(\beta)=W_{2}(\mu,\zeta)^{2}, so β∈Π(μ,ζ)\beta\in\Pi(\mu,\zeta) is optimal by The Noncommutative Quadratic Wasserstein Distance and Optimal Couplings §optimal. This proves claim 4.

Step 11 (Claim 5, variational formula). By claim 3, W2(μ,ν)2W_{2}(\mu,\nu)^{2} is a lower bound of the set {W2(μ,ζγ)2+M(ν)−M(ζγ): γ∈Π(μ,ν)}\{W_{2}(\mu,\zeta_{\gamma})^{2}+M(\nu)-M(\zeta_{\gamma}):\ \gamma\in\Pi(\mu,\nu)\}, and by claim 4 it is the value of this expression at every optimal γ\gamma. Since μ,ν∈Σd,R\mu,\nu\in\Sigma_{d,R}, The Noncommutative Wasserstein Distance: Existence of Optimal Couplings, Symmetry, Separation, a Moment Bound, Weak-Star Lower Semicontinuity, and Displacement Interpolation §attained provides an optimal γ∈Π(μ,ν)\gamma\in\Pi(\mu,\nu), so W2(μ,ν)2W_{2}(\mu,\nu)^{2} belongs to the set and is its least element. This proves claim 5.

Citations

Loading…

Dependencies

Uses0

Loading…

Comments

Log in to comment.

Loading…