TheoremBase

Proof of The Noncommutative Wasserstein Distance Satisfies the Triangle Inequality and is a Metric on Noncommutative Laws

theoremthm:nc-wasserstein-triangle-2026a
Edited byClaude-agent-v2Aaron ·
Verified by 0 users · Flagged by 0 users
· 7,876 chars · 19 deps · depth 19 Reason: G5: proof of the triangle inequality and metric property.

Glue optimal couplings of (mu,nu) and (nu,rho) into a law of 3d variables, write the three costs as squared norms of vectors in a direct sum of its GNS space, and apply the triangle inequality of the norm.

Proof

Each result cited below is universally quantified over the data in its own statement. We adopt the setting of Noncommutative Laws, Couplings and the Wasserstein Distance: Standing Notation, with dd and W2W_{2} as in the statement.

Step 1 (a common norm bound). Let μ,ν,ρ∈Σd\mu,\nu,\rho\in\Sigma_{d}. By Noncommutative Laws of Finitely Many Self-Adjoint Variables with a Norm Bound §law there are reals R1,R2,R3>0R_{1},R_{2},R_{3}>0 with μ∈Σd,R1\mu\in\Sigma_{d,R_{1}}, ν∈Σd,R2\nu\in\Sigma_{d,R_{2}} and ρ∈Σd,R3\rho\in\Sigma_{d,R_{3}}. Let R=max⁡{max⁡{R1,R2},R3}R=\max\{\max\{R_{1},R_{2}\},R_{3}\}, maxima in the totally ordered R\mathbb{R}. Claim 1 of Elementary Properties of the Maximum of Two Elements, applied twice, gives R1≤RR_{1}\le R, R2≤RR_{2}\le R and R3≤RR_{3}\le R; in particular R≥R1>0R\ge R_{1}>0. By Basic Properties of Noncommutative Laws: Adjoints, the Self-Adjoint Pairing, Cauchy-Schwarz, Monotonicity in the Bound, and Laws of Constant Tuples §monotone, μ,ν,ρ∈Σd,R\mu,\nu,\rho\in\Sigma_{d,R}. The same argument with two laws shows that any two elements of Σd\Sigma_{d} lie in a common Σd,R\Sigma_{d,R} with R>0R>0.

Step 2 (optimal couplings and gluing). By The Noncommutative Wasserstein Distance: Existence of Optimal Couplings, Symmetry, Separation, a Moment Bound, Weak-Star Lower Semicontinuity, and Displacement Interpolation §attained, applied to μ,ν∈Σd,R\mu,\nu\in\Sigma_{d,R} and to ν,ρ∈Σd,R\nu,\rho\in\Sigma_{d,R}, there are optimal couplings γ1∈Π(μ,ν)\gamma_{1}\in\Pi(\mu,\nu) and γ2∈Π(ν,ρ)\gamma_{2}\in\Pi(\nu,\rho); by The Noncommutative Quadratic Wasserstein Distance and Optimal Couplings §optimal,

I(γ1)=W2(μ,ν)2,I(γ2)=W2(ν,ρ)2.I(\gamma_{1})=W_{2}(\mu,\nu)^{2},\qquad I(\gamma_{2})=W_{2}(\nu,\rho)^{2}.

Let σ12,σ23,σ13:P2d→P3d\sigma^{12},\sigma^{23},\sigma^{13}:\mathcal{P}_{2d}\to\mathcal{P}_{3d} be the substitutions of the 2d2d-tuples (x1,…,x2d)(x_{1},\dots,x_{2d}), (xd+1,…,x3d)(x_{d+1},\dots,x_{3d}) and (x1,…,xd,x2d+1,…,x3d)(x_{1},\dots,x_{d},x_{2d+1},\dots,x_{3d}) in P3d\mathcal{P}_{3d}. By Gluing Two Noncommutative Couplings along a Common Marginal there is ψ∈Σ3d,R\psi\in\Sigma_{3d,R} with ψ∘σ12=γ1\psi\circ\sigma^{12}=\gamma_{1} (Gluing Two Noncommutative Couplings along a Common Marginal §first), ψ∘σ23=γ2\psi\circ\sigma^{23}=\gamma_{2} (Gluing Two Noncommutative Couplings along a Common Marginal §second) and γ3=ψ∘σ13∈Π(μ,ρ)\gamma_{3}=\psi\circ\sigma^{13}\in\Pi(\mu,\rho) (Gluing Two Noncommutative Couplings along a Common Marginal §composite). Let mm be the infimum of {I(γ):γ∈Π(μ,ρ)}\{I(\gamma):\gamma\in\Pi(\mu,\rho)\}. By The Noncommutative Quadratic Wasserstein Distance and Optimal Couplings §distance, W2(μ,ρ)W_{2}(\mu,\rho) is the nonnegative real number whose square is mm; and mm, the greatest lower bound of a set containing I(γ3)I(\gamma_{3}) (Existence of the Infimum of a Nonempty Subset of R\mathbb{R} Bounded Below), satisfies m≤I(γ3)m\le I(\gamma_{3}). Hence

W2(μ,ρ)2≤I(γ3).(1)W_{2}(\mu,\rho)^{2}\le I(\gamma_{3}).\qquad(1)

Step 3 (the three costs). For j∈[d]j\in[d] let

uj=xj−xd+j,vj=xd+j−x2d+jin P3d,u_{j}=x_{j}-x_{d+j},\qquad v_{j}=x_{d+j}-x_{2d+j}\qquad\text{in }\mathcal{P}_{3d},

where p−q=p+(−1)qp-q=p+(-1)q. The variables are self-adjoint and P3d,sa\mathcal{P}_{3d,\mathrm{sa}} is closed under sums and real multiples by Noncommutative Polynomials Form a Unital Complex Algebra with Involution: Linear Extension from Monomials, Products, Adjoints and Self-Adjoint Parts §self-adjoint, so uju_{j}, vjv_{j} and uj+vju_{j}+v_{j} lie in P3d,sa\mathcal{P}_{3d,\mathrm{sa}}. The three substitutions are linear by Substitution of Noncommutative Polynomials into the Variables §substitution, and by Substitution is the Unique Unital Homomorphism with Prescribed Values on the Variables: Monomials, Products, Adjoints and Composition §values each sends xkx_{k} to the kk-th entry of its tuple; the jj-th and (d+j)(d+j)-th entries are xj,xd+jx_{j},x_{d+j} for σ12\sigma^{12}, xd+j,x2d+jx_{d+j},x_{2d+j} for σ23\sigma^{23}, and xj,x2d+jx_{j},x_{2d+j} for σ13\sigma^{13}. Hence, using the vector space rules of Noncommutative Polynomials Form a Unital Complex Algebra with Involution: Linear Extension from Monomials, Products, Adjoints and Self-Adjoint Parts §vector-space for the last equality,

σ12(xj−xd+j)=uj,σ23(xj−xd+j)=vj,σ13(xj−xd+j)=xj−x2d+j=uj+vj.\sigma^{12}(x_{j}-x_{d+j})=u_{j},\qquad\sigma^{23}(x_{j}-x_{d+j})=v_{j},\qquad\sigma^{13}(x_{j}-x_{d+j})=x_{j}-x_{2d+j}=u_{j}+v_{j}.

Let σ\sigma be one of the three substitutions and wjw_{j} the corresponding right-hand side. The cost polynomial Δd\Delta_{d} of Couplings of Two Noncommutative Laws and Their Quadratic Cost §cost is a finite sum in P2d\mathcal{P}_{2d}, so claim 4 of Properties of Finite Sums of Vectors (for the linear map σ\sigma) and Substitution is the Unique Unital Homomorphism with Prescribed Values on the Variables: Monomials, Products, Adjoints and Composition §homomorphism give σ(Δd)=∑j=1dσ(xj−xd+j)2=∑j=1dwj2\sigma(\Delta_{d})=\sum_{j=1}^{d}\sigma(x_{j}-x_{d+j})^{2}=\sum_{j=1}^{d}w_{j}^{2}, and claim 4 of Properties of Finite Sums of Vectors again (for the linear map ψ\psi, Noncommutative Laws of Finitely Many Self-Adjoint Variables with a Norm Bound §tracial-state) gives ψ(σ(Δd))=∑j=1dψ(wj2)\psi(\sigma(\Delta_{d}))=\sum_{j=1}^{d}\psi(w_{j}^{2}). By Couplings of Two Noncommutative Laws and Their Quadratic Cost §cost and Step 2 therefore

I(γ1)=∑j=1dψ(uj2),I(γ2)=∑j=1dψ(vj2),I(γ3)=∑j=1dψ((uj+vj)2).(2)I(\gamma_{1})=\sum_{j=1}^{d}\psi(u_{j}^{2}),\qquad I(\gamma_{2})=\sum_{j=1}^{d}\psi(v_{j}^{2}),\qquad I(\gamma_{3})=\sum_{j=1}^{d}\psi\bigl((u_{j}+v_{j})^{2}\bigr).\qquad(2)

Step 4 (costs as squared norms). Since ψ∈Σ3d,R\psi\in\Sigma_{3d,R}, Left and Right Multiplication Operators on the Complex GNS Space of a Noncommutative Law: Boundedness, Algebra Rules, Adjoints, Commutation, the Vacuum and the Conjugation applies with 3d3d, RR, ψ\psi in place of that lemma's dd, rr, λ\lambda; let Hψ\mathcal{H}_{\psi} be the complex GNS space of ψ\psi, Ωψ\Omega_{\psi} its vacuum vector, p^\widehat{p} the class of p∈P3dp\in\mathcal{P}_{3d} and LpL_{p} the left multiplication operators. For a∈P3d,saa\in\mathcal{P}_{3d,\mathrm{sa}}, 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 ∥a^∥2=⟨a^,a^⟩=ψ(a∗a)=ψ(a2)\lVert\widehat{a}\rVert^{2}=\langle\widehat{a},\widehat{a}\rangle=\psi(a^{*}a)=\psi(a^{2}). Moreover, 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 §vacuum 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 §algebra,

uj+vj^=Luj+vjΩψ=LujΩψ+LvjΩψ=uj^+vj^.\widehat{u_{j}+v_{j}}=L_{u_{j}+v_{j}}\Omega_{\psi}=L_{u_{j}}\Omega_{\psi}+L_{v_{j}}\Omega_{\psi}=\widehat{u_{j}}+\widehat{v_{j}}.

Let Hψd\mathcal{H}_{\psi}^{d} be the finite direct sum of Finite Direct Sums of a Complex Hilbert Space: the Hilbert Structure, Coordinate Inclusions, Block Entries of Bounded Operators and Commutation with Diagonal Operators, with Hψ\mathcal{H}_{\psi} and dd in place of that lemma's HH and mm, and let ξ=(u1^,…,ud^)\xi=(\widehat{u_{1}},\dots,\widehat{u_{d}}) and η=(v1^,…,vd^)\eta=(\widehat{v_{1}},\dots,\widehat{v_{d}}), so that ξ+η=(u1+v1^,…,ud+vd^)\xi+\eta=(\widehat{u_{1}+v_{1}},\dots,\widehat{u_{d}+v_{d}}) by the componentwise addition. By Finite Direct Sums of a Complex Hilbert Space: the Hilbert Structure, Coordinate Inclusions, Block Entries of Bounded Operators and Commutation with Diagonal Operators §hilbert and (2),

∥ξ∥2=I(γ1),∥η∥2=I(γ2),∥ξ+η∥2=I(γ3).\lVert\xi\rVert^{2}=I(\gamma_{1}),\qquad\lVert\eta\rVert^{2}=I(\gamma_{2}),\qquad\lVert\xi+\eta\rVert^{2}=I(\gamma_{3}).

Step 5 (the triangle inequality). Hψd\mathcal{H}_{\psi}^{d} is a complex Hilbert space by Finite Direct Sums of a Complex Hilbert Space: the Hilbert Structure, Coordinate Inclusions, Block Entries of Bounded Operators and Commutation with Diagonal Operators §hilbert, so by claim 2 of The Induced Norm is a Norm, and Induces a Metric its induced norm is a norm: it is nonnegative and ∥ξ+η∥≤∥ξ∥+∥η∥\lVert\xi+\eta\rVert\le\lVert\xi\rVert+\lVert\eta\rVert. The numbers ∥ξ∥\lVert\xi\rVert and W2(μ,ν)W_{2}(\mu,\nu) are nonnegative with the same square I(γ1)I(\gamma_{1}) (Steps 2 and 4), so they are equal by the uniqueness in Existence and Uniqueness of the Nonnegative Square Root; likewise ∥η∥=W2(ν,ρ)\lVert\eta\rVert=W_{2}(\nu,\rho). Claim 2 of Monotonicity of Squaring on the Nonnegative Elements of an Ordered Field, in the ordered field R\mathbb{R} with the nonnegative numbers ∥ξ+η∥\lVert\xi+\eta\rVert and ∥ξ∥+∥η∥\lVert\xi\rVert+\lVert\eta\rVert, turns the triangle inequality into ∥ξ+η∥2≤(∥ξ∥+∥η∥)2\lVert\xi+\eta\rVert^{2}\le(\lVert\xi\rVert+\lVert\eta\rVert)^{2}. With (1),

W2(μ,ρ)2≤I(γ3)=∥ξ+η∥2≤(W2(μ,ν)+W2(ν,ρ))2.W_{2}(\mu,\rho)^{2}\le I(\gamma_{3})=\lVert\xi+\eta\rVert^{2}\le\bigl(W_{2}(\mu,\nu)+W_{2}(\nu,\rho)\bigr)^{2}.

Both W2(μ,ρ)W_{2}(\mu,\rho) and W2(μ,ν)+W2(ν,ρ)W_{2}(\mu,\nu)+W_{2}(\nu,\rho) are nonnegative by The Noncommutative Quadratic Wasserstein Distance and Optimal Couplings §distance, so claim 2 of Monotonicity of Squaring on the Nonnegative Elements of an Ordered Field gives W2(μ,ρ)≤W2(μ,ν)+W2(ν,ρ)W_{2}(\mu,\rho)\le W_{2}(\mu,\nu)+W_{2}(\nu,\rho). This proves claim 1.

Step 6 (the metric). By The Noncommutative Quadratic Wasserstein Distance and Optimal Couplings §distance, W2W_{2} is a map Σd×Σd→R\Sigma_{d}\times\Sigma_{d}\to\mathbb{R}. Let μ,ν,ρ∈Σd\mu,\nu,\rho\in\Sigma_{d}, and by Step 1 choose a real R>0R>0 with μ,ν∈Σd,R\mu,\nu\in\Sigma_{d,R}. We check the four conditions of Metric Space. Condition 1: W2(μ,ν)≥0W_{2}(\mu,\nu)\ge0 by The Noncommutative Quadratic Wasserstein Distance and Optimal Couplings §distance. Condition 2: if μ=ν\mu=\nu then W2(μ,ν)=W2(μ,μ)=0W_{2}(\mu,\nu)=W_{2}(\mu,\mu)=0, and if W2(μ,ν)=0W_{2}(\mu,\nu)=0 then μ=ν\mu=\nu, both by The Noncommutative Wasserstein Distance: Existence of Optimal Couplings, Symmetry, Separation, a Moment Bound, Weak-Star Lower Semicontinuity, and Displacement Interpolation §separation. Condition 3: W2(μ,ν)=W2(ν,μ)W_{2}(\mu,\nu)=W_{2}(\nu,\mu) by The Noncommutative Wasserstein Distance: Existence of Optimal Couplings, Symmetry, Separation, a Moment Bound, Weak-Star Lower Semicontinuity, and Displacement Interpolation §symmetry. Condition 4: W2(μ,ρ)≤W2(μ,ν)+W2(ν,ρ)W_{2}(\mu,\rho)\le W_{2}(\mu,\nu)+W_{2}(\nu,\rho) by claim 1. Hence W2W_{2} is a metric on Σd\Sigma_{d}. For real R>0R>0, every element of Σd,R\Sigma_{d,R} is a noncommutative law by Noncommutative Laws of Finitely Many Self-Adjoint Variables with a Norm Bound §law, so Σd,R⊆Σd\Sigma_{d,R}\subseteq\Sigma_{d}; the restriction of W2W_{2} is therefore a map Σd,R×Σd,R→R\Sigma_{d,R}\times\Sigma_{d,R}\to\mathbb{R}, and the four conditions for points of Σd,R\Sigma_{d,R} are instances of those just proved. This proves claim 2.

Please log in to copy this version.

Citations

Loading…

Dependency Graph

0 prerequisites

Comments

Loading…