Introduction
The asymmetry between space and time gives rise to conceptual difficulties known as the problem of time when one attempts to reconcile quantum mechanics with relativity, and it becomes especially significant in quantum field theory and quantum gravity [1 , 2 ] . One approach for this is the spatiotemporal quantum mechanics, where a local Hilbert space is assigned to each spacetime point rather than just consider the vector bundle over spatial manifold, namely H t o t = ⨂ ( x , t ) H x , t \mathcal{H}_{\mathrm{tot}} = \bigotimes_{(x,t)} \mathcal{H}_{x,t} H tot = ⨂ ( x , t ) H x , t and the notion of a spatial joint density operator is extended to that of a spatiotemporal state. Examples include consistent histories [3 ] , the Feynman-Vernon influence functional [4 ] and the influence matrix [5 ] , multiple-time states [6 , 7 , 8 ] , pseudo-density operators [9 ] (equivalently, states over time [10 , 11 ] ), quantum combs [12 ] (equivalently quantum-classical games [13 ] , process tensors [14 ] ), process matrices [15 ] which can capture the indefinite causal order, superdensity operators [16 ] , doubled density operators and spatiotemporal states [17 , 18 , 19 , 20 ] , and spacetime quantum mechanics [21 , 22 , 23 ] , also see [24 , 25 , 26 , 27 ] . Clarifying how these approaches connect to and differ from one another remains an important open direction [10 , 11 , 18 , 20 , 22 , 28 ] . In [18 , 20 ] , we use spatiotemporal Kirkwood-Dirac phase-space distribution to give a unification of some existing frameworks. This unification provides a natural language in which to formulate quantum transport directly in spacetime. There have been attempts to connect optimal transport with spatiotemporal physics [29 ] , mainly in the framework and gauge/gravity duality [30 , 31 , 32 , 33 ] . In this work, we apply our Kirkwood-Dirac phase space and spatiotemporal state framework to define quantum optimal transport in spacetime.
The Monge-Kantorovich optimal transport problem, originally proposed by Gaspard Monge in 1781 [34 ] , concerns the optimal transportation of resources with respect to a prescribed cost function. A transport map y = T ( x ) y = T(x) y = T ( x ) induces the push-forward T # p T_{\#}p T # p of a probability measure p ( x ) p(x) p ( x ) on X \mathcal{X} X , with the constraint T # p = q T_{\#}p = q T # p = q on Y \mathcal{Y} Y . Given input and output probability measures p p p and q q q , together with a cost function c ( x , T ( x ) ) c(x,T(x)) c ( x , T ( x )) , Monge’s problem is formulated as
M P = inf T { ∫ X c ( x , T ( x ) ) p ( x ) d x | T # p = q } . (1) \mathrm{MP} = \inf_{T} \left\{ \int_{\mathcal{X}} c(x,T(x))p(x)\,\mathrm{d}x \;\middle|\; T_{\#}p = q \right\}. \tag*{(1)} MP = T inf { ∫ X c ( x , T ( x )) p ( x ) d x T # p = q } . (1)
A major advance was later made by Leonid Kantorovich [35 ] , leading to the formulation now known under both of their names. In Kantorovich’s formulation, one is given input and output probability measures p ( x ) p(x) p ( x ) and q ( y ) q(y) q ( y ) , along with a cost function c ( x , y ) c(x,y) c ( x , y ) defined on the product polish space X × Y \mathcal{X} \times\mathcal{Y} X × Y . Imagine X \mathcal{X} X as the producers and Y \mathcal{Y} Y as the consumers; then c ( x , y ) c(x,y) c ( x , y ) can be regarded as the cost of transporting goods from x x x to y y y . Transport plans are represented by joint probability distributions γ ( x , y ) \gamma(x,y) γ ( x , y ) whose marginals coincide with p ( x ) p(x) p ( x ) and q ( y ) q(y) q ( y ) , γ ( x , y ) \gamma(x,y) γ ( x , y ) is also called coupling, the set of of coupling can be denoted as C ( p , q ) \mathcal{C}(p,q) C ( p , q ) . The objective is to minimize the expected cost over all admissible couplings:
K P = inf γ ∈ C ( p , q ) ∫ X × Y c ( x , y ) γ ( x , y ) d x d y , (2) \mathrm{KP} = \inf_{\gamma\in\mathcal{C}(p,q)} \int_{\mathcal{X}\times\mathcal{Y}} c(x,y)\gamma(x,y)\,\mathrm{d}x\,\mathrm{d}y, \tag*{(2)} KP = γ ∈ C ( p , q ) inf ∫ X × Y c ( x , y ) γ ( x , y ) d x d y , (2)
For k ∈ [ 1 , ∞ ] k \in[1,\infty] k ∈ [ 1 , ∞ ] , Wasserstein metric is defined as [36 ]
W k ( p , q ) = inf γ ∈ C ( p , q ) [ E γ ( x , y ) d ( x , y ) k ] 1 / k = inf γ ∈ C ( p , q ) ( ∫ X × Y d ( x , y ) k d γ ( x , y ) ) 1 / k (3) \begin{aligned}
W_k(p,q) &= \inf_{\gamma\in\mathcal{C}(p,q)} \left[\mathbb{E}_{\gamma(x,y)}d(x,y)^k\right]^{1/k} \\
&= \inf_{\gamma\in\mathcal{C}(p,q)} \left(\int_{\mathcal{X}\times\mathcal{Y}}d(x,y)^k\mathrm{d}\gamma(x,y)\right)^{1/k} \tag*{(3)}
\end{aligned} W k ( p , q ) = γ ∈ C ( p , q ) inf [ E γ ( x , y ) d ( x , y ) k ] 1/ k = γ ∈ C ( p , q ) inf ( ∫ X × Y d ( x , y ) k d γ ( x , y ) ) 1/ k (3)
The optimal transport problem has many applications in economics, fluid dynamics, meteorology and oceanography, surface reconstruction, etc., see [37 ] .
Quantum generalizations of the Monge-Kantorovich problem have been studied from various perspectives since the 1990s. Early work defined a distance between two quantum states as the Monge distance between their corresponding Husimi Q-functions [38 , 39 ] . Replacing classical marginal distributions with density operators and classical couplings with joint density operators leads to a widely studied formulation of quantum optimal transport [40 , 41 , 42 , 43 , 44 , 45 , 46 , 47 , 48 , 49 , 50 ] . In this standard spatial formulation , the prescribed marginals are local states ρ A \rho_A ρ A and ρ B \rho_B ρ B , and a spatial quantum coupling is a spatial joint quantum state Υ A B \Upsilon_{AB} Υ A B with reduced states ρ A \rho_A ρ A and ρ B \rho_B ρ B . The objective is to minimize a prescribed quantum transport cost Tr [ C Υ A B ] \operatorname{Tr}[C\Upsilon_{AB}] Tr [ C Υ A B ] over all such plans. The quantum optimal transport also found many applications, like measuring distance between quantum states [43 , 45 , 46 , 48 , 51 ] and quantum machine learning [42 , 52 , 53 ] .
Recently, some temporal generalizations have also been explored. In Ref. [54 ] , temporal quantum couplings are characterized by quantum channels, while in Ref. [55 ] , pseudo-density operators are employed as temporal quantum transport plans. In this work, we propose a general framework for quantum optimal transport in spacetime using spatiotemporal quasiprobability distribution in spatiotemporal phase space and spatiotemporal states, which are systematically developed in [18 , 20 ] , also see [22 , 24 , 25 , 26 ] for related discussions on spatiotemporal formalism of quantum mechanics.
Our main tool is the spatiotemporal Kirkwood-Dirac distribution [18 , 20 , 56 , 57 ] , which can be regarded as a spatiotemporal state in the framework of generalized probabilistic theory. In the spatiotemporal phase space, the local marginal states are characterized by quasiprobability distributions Q z i ( α i ) Q_{z_i}(\alpha_i) Q z i ( α i ) over the phase space Ω z i \Omega_{z_i} Ω z i for a given spacetime point z i = ( x i , t i ) z_i=(x_i,t_i) z i = ( x i , t i ) , the spatiotemporal coupling is given by the joint phase-space distribution Q A ( α ) Q_A(\alpha) Q A ( α ) over the joint phase space Ω A = Ω z 1 × ⋯ × Ω z N \Omega_A=\Omega_{z_1}\times\cdots\times\Omega_{z_N} Ω A = Ω z 1 × ⋯ × Ω z N for the given spacetime set A = { z 1 , ⋯ , z N } A=\{z_1,\cdots,z_N\} A = { z 1 , ⋯ , z N } . The cost function is a generally complex-valued function C : Ω A → C C:\Omega_A\to\mathbb{C} C : Ω A → C . The objective function is ∣ ∑ α C ( α ) Q ( α ) ∣ \left|\sum_{\alpha} C(\alpha)Q(\alpha)\right| ∣ ∑ α C ( α ) Q ( α ) ∣ . The phase-space optimal transport is to find the infimum of objective function. This is a very natural generalization of classical optimal transport, since one simply replaces the classical probability distribution with a quantum quasiprobability distribution; the quantumness is reflected in the deviation of the quasiprobability distribution from satisfying the Kolmogorov axioms.
Using frames F α F_\alpha F α and their dual frames D α D_\alpha D α , we can transform the phase-space formulation into the spatiotemporal state formulation: the cost function becomes a cost operator C = ∑ α C ( α ) F α C=\sum_\alpha C(\alpha)F_\alpha C = ∑ α C ( α ) F α , and the Kirkwood-Dirac distribution yields the spatiotemporal state Υ = ∑ α Q ( α ) D α \Upsilon=\sum_\alpha Q(\alpha)D_\alpha Υ = ∑ α Q ( α ) D α . In general, both the cost operator and the spatiotemporal state can be non-Hermitian. The objective function becomes ∣ Tr C Υ ∣ \left|\operatorname{Tr}C\Upsilon\right| ∣ Tr C Υ ∣ . The standard spatial quantum optimal transport naturally becomes a special case of this problem if we choose A A A to be an equal-time set A = { ( x 1 , t ) , ⋯ , ( x N , t ) } A=\{(x_1,t),\cdots,(x_N,t)\} A = {( x 1 , t ) , ⋯ , ( x N , t )} . Unlike the usual choice of a positive semidefinite cost operator, in this work we consider general non-Hermitian cost operators.
When the phase-space distribution is chosen as the Margenau-Hill distribution [18 , 20 , 58 ] , the corresponding spatiotemporal state becomes Hermitian; in this case, we obtain a Hermitian optimal transport problem.
The duality of optimal transport is established carefully; we discuss both Kantorovich duality and geometric duality. Defining the quantum Wasserstein metric is a crucial topic in quantum optimal transport. For the spatiotemporal case, directly defining a Wasserstein metric is challenging; instead, using the phase-space formalism, we can introduce the Kantorovich-Rubinstein metric, which can be regarded as a Wasserstein metric for p = 1 p = 1 p = 1 .
The rest of the paper is organized as follows. § 2 reviews the fundamentals of the spatiotemporal phase space and Kirkwood-Dirac distributions over the phase space, and explains how it relates to the spatiotemporal state. In § 3 , we define the quantum optimal transport using both spatiotemporal Kirkwood-Dirac distribution and spatiotemporal state. The phase-space definition can be transformed into the spatiotemporal-state definition via a frame and its dual frame; the reverse direction is given by the spatiotemporal Born rule. We show that the spatiotemporal quantum optimal transport is independent of the choice of frames. In § 4 , we apply our framework to the spatial setting, we especially carefully discuss the spatial optimal transport for Hermitian and non-Hermitian operators (both are beyond the standard positive semidefinite cost operator formalism). For Hermitian cost operator, we establish the quantum Kantorovich duality by showing that it is a duality of semidefinite programming (SDP). Using the phase-space distribution, this quantum Kantorovich duality has a very similar form as that for classical optimal transport. For non-Hermitian operator, we show that the duality can be established via introducing an extra unit disk optimization in the complex plane. § 5 applies the framework to the temporal setting. We carefully discuss the fixed-branch case and show that, in two-time case, it is equivalent to the quantum optimal transport for channels. For the multi-time case, it is a multi-convex problem. Then we introduce the convex temporal coupling set for fixed-branch, and mixed-branch cases, and discuss how to establish the geometric duality for these problems. We also discuss how to use the dual branch and Hermitianization to give a Margenau-Hill temporal quantum optimal transport. In § 6 , we briefly discuss how to extend our temporal optimal transport framework to the spatiotemporal setting. Finally, in § 7 , we use the phase-space formalism to give a Kantorovich-Rubinstein metric by comparing it with the definition for classical optimal transport, we prove the metric satisfies the axioms of a distance function. In Appendix, we collection some results from convex analysis and matrix analysis that is helpful for the discussion in the main text.
Preliminaries on spatiotemporal quasiprobability distribution and its frame representation
In this section, we will review the basics of spatiotemporal quasiprobability distributions and operator frames. The standard Kirkwood-Dirac distribution is a complex-valued quasiprobabilistic representation of a spatial density operator with respect to a specified operator frame. Similarly, the spatiotemporal Kirkwood-Dirac distribution is a complex-valued quasiprobabilistic representation of a spatiotemporal density operator with respect to a specified operator frame. For a fixed frame, the spatiotemporal Kirkwood-Dirac distribution and the corresponding spatiotemporal density operator uniquely determine one another.
Standard spatial Kirkwood-Dirac distribution
To begin with, we recall several basic definitions. For a Hilbert space H \mathcal{H} H , we denote by B ( H ) \mathbf{B}(\mathcal{H}) B ( H ) the space of all bounded linear operators and by Herm ( H ) \operatorname{Herm}(\mathcal{H}) Herm ( H ) the space of all Hermitian operators. For an operator A A A , we denote its Schatten p p p -norm by ∥ A ∥ p \lVert A\rVert_p ∥ A ∥ p . The Hilbert-Schmidt inner product is defined as ⟨ A , B ⟩ : = Tr [ A † B ] \langle A,B\rangle:= \operatorname{Tr}[A^\dagger B] ⟨ A , B ⟩ := Tr [ A † B ] . An operator frame [59 ] is a family of operators { F α } α \{F_\alpha\}_\alpha { F α } α for which there exist constants 0 < a ≤ b < ∞ 0<a\le b<\infty 0 < a ≤ b < ∞ such that, for every A ∈ B ( H ) A\in\mathbf{B}(\mathcal{H}) A ∈ B ( H ) ,
a ∥ A ∥ 2 2 ≤ ∑ α ∣ Tr [ F α † A ] ∣ 2 ≤ b ∥ A ∥ 2 2 . (4) a\lVert A\rVert_2^2\le\sum_\alpha\left\lvert\operatorname{Tr}[F_\alpha^\dagger A]\right\rvert^2\le b\lVert A\rVert_2^2. \tag*{(4)} a ∥ A ∥ 2 2 ≤ α ∑ Tr [ F α † A ] 2 ≤ b ∥ A ∥ 2 2 . (4)
A dual frame of { F α } α \{F_\alpha\}_\alpha { F α } α is a family of operators { D β } β \{D_\beta\}_\beta { D β } β satisfying the reconstruction formula
A = ∑ α Tr [ F α † A ] D α , (5) A=\sum_\alpha\operatorname{Tr}[F_\alpha^\dagger A]D_\alpha, \tag*{(5)} A = α ∑ Tr [ F α † A ] D α , (5)
for every A ∈ B ( H ) A\in\mathbf{B}(\mathcal{H}) A ∈ B ( H ) . Operator frames play a central role in quasiprobabilistic descriptions of quantum theory; see Ref. [60 ] for a review. Using an operator frame, a density operator ρ \rho ρ can be represented by
Q ρ ( α ) = Tr [ F α † ρ ] , (6) Q_\rho(\alpha)=\operatorname{Tr}[F_\alpha^\dagger\rho], \tag*{(6)} Q ρ ( α ) = Tr [ F α † ρ ] , (6)
and a measurement effect E E E by
η E ( β ) = Tr [ D β E ] . (7) \eta_E(\beta)=\operatorname{Tr}[D_\beta E]. \tag*{(7)} η E ( β ) = Tr [ D β E ] . (7)
The Born rule then takes the form Tr [ ρ E ] = ∑ α Q ρ ( α ) η E ( α ) \operatorname{Tr}[\rho E]=\sum_\alpha Q_\rho(\alpha)\eta_E(\alpha) Tr [ ρE ] = ∑ α Q ρ ( α ) η E ( α ) . In quantum theory, states and measurement effects cannot both admit classical probabilistic representations within a single frame representation; at least one must be represented quasiprobabilistically [61 ] .
The Kirkwood-Dirac distribution admits a representation in terms of a non-Hermitian operator frame. Let { ∣ a ⟩ } a = 1 d \{|a\rangle\}_{a=1}^d { ∣ a ⟩ } a = 1 d and { ∣ b ⟩ } b = 1 d \{|b\rangle\}_{b=1}^d { ∣ b ⟩ } b = 1 d be orthonormal bases of H \mathcal{H} H , with P a = ∣ a ⟩ ⟨ a ∣ P_a=|a\rangle\langle a| P a = ∣ a ⟩ ⟨ a ∣ and Q b = ∣ b ⟩ ⟨ b ∣ Q_b=|b\rangle\langle b| Q b = ∣ b ⟩ ⟨ b ∣ , the standard Kirkwood-Dirac distribution for a density operator ρ \rho ρ is given by [56 , 57 ]
Q ( a , b ) = Tr [ Q b P a ρ ] . (8) Q(a,b)=\operatorname{Tr}[Q_bP_a\rho]. \tag*{(8)} Q ( a , b ) = Tr [ Q b P a ρ ] . (8)
Assume that ⟨ b ∣ a ⟩ ≠ 0 \langle b|a\rangle\ne0 ⟨ b ∣ a ⟩ = 0 for every pair ( a , b ) (a,b) ( a , b ) . The operators
F a b = P a Q b = ⟨ a ∣ b ⟩ ∣ a ⟩ ⟨ b ∣ (9) F_{ab}=P_aQ_b=\langle a|b\rangle|a\rangle\langle b| \tag*{(9)} F ab = P a Q b = ⟨ a ∣ b ⟩ ∣ a ⟩ ⟨ b ∣ (9)
form an orthogonal basis of B ( H ) \mathbf{B}(\mathcal{H}) B ( H ) , since
Tr [ F a b † F a ′ b ′ ] = δ a a ′ δ b b ′ ∣ ⟨ b ∣ a ⟩ ∣ 2 . (10) \operatorname{Tr}[F_{ab}^\dagger F_{a'b'}]=\delta_{aa'}\delta_{bb'}\lvert\langle b|a\rangle\rvert^2. \tag*{(10)} Tr [ F ab † F a ′ b ′ ] = δ a a ′ δ b b ′ ∣⟨ b ∣ a ⟩ ∣ 2 . (10)
To verify the frame condition explicitly, note that the operators ∣ a ⟩ ⟨ b ∣ |a\rangle\langle b| ∣ a ⟩ ⟨ b ∣ form an orthonormal basis with respect to the Hilbert-Schmidt inner product. Therefore, for every A ∈ B ( H ) A\in\mathbf{B}(\mathcal{H}) A ∈ B ( H ) ,
∑ a , b ∣ Tr [ F a b † A ] ∣ 2 = ∑ a , b ∣ ⟨ b ∣ a ⟩ ∣ 2 ∣ ⟨ a ∣ A ∣ b ⟩ ∣ 2 , (11) \sum_{a,b}\left\lvert\operatorname{Tr}[F_{ab}^\dagger A]\right\rvert^2=\sum_{a,b}\lvert\langle b|a\rangle\rvert^2\lvert\langle a|A|b\rangle\rvert^2, \tag*{(11)} a , b ∑ Tr [ F ab † A ] 2 = a , b ∑ ∣⟨ b ∣ a ⟩ ∣ 2 ∣⟨ a ∣ A ∣ b ⟩ ∣ 2 , (11)
∥ A ∥ 2 2 = ∑ a , b ∣ ⟨ a ∣ A ∣ b ⟩ ∣ 2 . (12) \lVert A\rVert_2^2=\sum_{a,b}\lvert\langle a|A|b\rangle\rvert^2. \tag*{(12)} ∥ A ∥ 2 2 = a , b ∑ ∣⟨ a ∣ A ∣ b ⟩ ∣ 2 . (12)
Defining c min = min a , b ∣ ⟨ b ∣ a ⟩ ∣ 2 > 0 c_{\min}=\min_{a,b}|\langle b|a\rangle|^2>0 c m i n = min a , b ∣ ⟨ b ∣ a ⟩ ∣ 2 > 0 and c max = max a , b ∣ ⟨ b ∣ a ⟩ ∣ 2 c_{\max}=\max_{a,b}|\langle b|a\rangle|^2 c m a x = max a , b ∣ ⟨ b ∣ a ⟩ ∣ 2 , we obtain
c min ∥ A ∥ 2 2 ≤ ∑ a , b ∣ Tr [ F a b † A ] ∣ 2 ≤ c max ∥ A ∥ 2 2 . (13) c_{\min}\lVert A\rVert_2^2\leq\sum_{a,b}\left|\operatorname{Tr}[F_{ab}^{\dagger}A]\right|^2\leq c_{\max}\lVert A\rVert_2^2. \tag*{(13)} c m i n ∥ A ∥ 2 2 ≤ a , b ∑ Tr [ F ab † A ] 2 ≤ c m a x ∥ A ∥ 2 2 . (13)
Thus { F a b } \{F_{ab}\} { F ab } is an operator frame. It also satisfies ∑ a , b F a b = 1 \sum_{a,b}F_{ab}=\mathbb{1} ∑ a , b F ab = 1 . Its unique dual frame is
G a b = F a b ∣ ⟨ b ∣ a ⟩ ∣ 2 = ∣ a ⟩ ⟨ b ∣ ⟨ b ∣ a ⟩ , (14) G_{ab}=\frac{F_{ab}}{|\langle b|a\rangle|^2}=\frac{|a\rangle\langle b|}{\langle b|a\rangle}, \tag*{(14)} G ab = ∣ ⟨ b ∣ a ⟩ ∣ 2 F ab = ⟨ b ∣ a ⟩ ∣ a ⟩ ⟨ b ∣ , (14)
which satisfies Tr [ F a b † G a ′ b ′ ] = δ a a ′ δ b b ′ \operatorname{Tr}[F_{ab}^{\dagger}G_{a'b'}]=\delta_{aa'}\delta_{bb'} Tr [ F ab † G a ′ b ′ ] = δ a a ′ δ b b ′ .
The Kirkwood-Dirac distribution is the corresponding frame representation of a density operator:
Q ρ ( a , b ) = Tr [ F a b † ρ ] = Tr [ Q b P a ρ ] = ⟨ b ∣ a ⟩ ⟨ a ∣ ρ ∣ b ⟩ . (15) Q_{\rho}(a,b)=\operatorname{Tr}[F_{ab}^{\dagger}\rho]=\operatorname{Tr}[Q_bP_a\rho]=\langle b|a\rangle\langle a|\rho|b\rangle. \tag*{(15)} Q ρ ( a , b ) = Tr [ F ab † ρ ] = Tr [ Q b P a ρ ] = ⟨ b ∣ a ⟩ ⟨ a ∣ ρ ∣ b ⟩ . (15)
The dual frame yields the reconstruction formula
ρ = ∑ a , b Q ρ ( a , b ) G a b . (16) \rho=\sum_{a,b}Q_{\rho}(a,b)G_{ab}. \tag*{(16)} ρ = a , b ∑ Q ρ ( a , b ) G ab . (16)
A measurement effect E E E is represented by
η E ( a , b ) = Tr [ G a b E ] = ⟨ b ∣ E ∣ a ⟩ ⟨ b ∣ a ⟩ , (17) \eta_E(a,b)=\operatorname{Tr}[G_{ab}E]=\frac{\langle b|E|a\rangle}{\langle b|a\rangle}, \tag*{(17)} η E ( a , b ) = Tr [ G ab E ] = ⟨ b ∣ a ⟩ ⟨ b ∣ E ∣ a ⟩ , (17)
so that the Born rule takes the form
Tr [ ρ E ] = ∑ a , b Q ρ ( a , b ) η E ( a , b ) . (18) \operatorname{Tr}[\rho E]=\sum_{a,b}Q_{\rho}(a,b)\eta_E(a,b). \tag*{(18)} Tr [ ρE ] = a , b ∑ Q ρ ( a , b ) η E ( a , b ) . (18)
Although the state and measurement representations can be complex, their pairing reproduces the physical measurement probability. Moreover,
∑ b Q ρ ( a , b ) = Tr [ P a ρ ] , ∑ a Q ρ ( a , b ) = Tr [ Q b ρ ] , (19) \sum_b Q_{\rho}(a,b)=\operatorname{Tr}[P_a\rho],\qquad\sum_a Q_{\rho}(a,b)=\operatorname{Tr}[Q_b\rho], \tag*{(19)} b ∑ Q ρ ( a , b ) = Tr [ P a ρ ] , a ∑ Q ρ ( a , b ) = Tr [ Q b ρ ] , (19)
the Kirkwood-Dirac distribution has the measurement probabilities in the two bases as its marginals.
Spatiotemporal Kirkwood-Dirac distribution
A spatiotemporal Kirkwood-Dirac distribution is a complex-valued quasiprobability representation of a spatiotemporal density operator with respect to a chosen spatiotemporal operator frame [18 , 20 ] . We focus primarily on the temporal setting; the extension to the spatiotemporal setting is straightforward.
Following Refs. [18 , 20 ] , we define the temporal Kirkwood-Dirac distribution for memoryless dynamics as1
Q ( α N , … , α 1 ) = Tr [ P α N ( t N ) ∘ E t N ← t N − 1 ∘ ⋯ ∘ P α 1 ( t 1 ) ∘ E t 1 ← t 0 ∘ P α 0 t 0 ( ρ t 0 ) ] . (20) Q(\alpha_N,\ldots,\alpha_1)=\operatorname{Tr}\left[\mathcal{P}_{\alpha_N}^{(t_N)}\circ\mathcal{E}_{t_N\leftarrow t_{N-1}}\circ\cdots\circ\mathcal{P}_{\alpha_1}^{(t_1)}\circ\mathcal{E}_{t_1\leftarrow t_0}\circ\mathcal{P}_{\alpha_0}^{t_0}(\rho_{t_0})\right]. \tag*{(20)} Q ( α N , … , α 1 ) = Tr [ P α N ( t N ) ∘ E t N ← t N − 1 ∘ ⋯ ∘ P α 1 ( t 1 ) ∘ E t 1 ← t 0 ∘ P α 0 t 0 ( ρ t 0 ) ] . (20)
where E t k + 1 ← t k \mathcal{E}_{t_{k+1}\leftarrow t_k} E t k + 1 ← t k is the completely positive trace-preserving (CPTP) evolution map from t k t_k t k to t k + 1 t_{k+1} t k + 1 , α k \alpha_k α k labels a phase-space point at time t k t_k t k , and P α k ( t k ) \mathcal{P}_{\alpha_k}^{(t_k)} P α k ( t k ) is the corresponding phase-space operation, which need not be completely positive. Typical choices [18 , 20 ] include right multiplication, P α k ( t k ) ( ∙ ) = ∙ F α k ( t k ) \mathcal{P}_{\alpha_k}^{(t_k)}(\bullet)=\bullet F_{\alpha_k}^{(t_k)} P α k ( t k ) ( ∙ ) = ∙ F α k ( t k ) ; left multiplication, P α k ( t k ) ( ∙ ) = F α k ( t k ) ∙ \mathcal{P}_{\alpha_k}^{(t_k)}(\bullet)=F_{\alpha_k}^{(t_k)}\bullet P α k ( t k ) ( ∙ ) = F α k ( t k ) ∙ ; and the doubled construction, P α k , β k t k ( ∙ ) = F α k ( t k ) ∙ F β k ( t k ) \mathcal{P}_{\alpha_k,\beta_k}^{t_k}(\bullet)=F_{\alpha_k}^{(t_k)}\bullet F_{\beta_k}^{(t_k)} P α k , β k t k ( ∙ ) = F α k ( t k ) ∙ F β k ( t k ) , where { F α k ( t k ) } \{F_{\alpha_k}^{(t_k)}\} { F α k ( t k ) } and { F β k ( t k ) } \{F_{\beta_k}^{(t_k)}\} { F β k ( t k ) } are operator frames. Following the convention introduced in [20 ] , we refer to the choice made at each time slot as a branch and denote B = ( t n s n , … , t 0 s 0 ) \mathrm{B}=(t_n^{s_n},\ldots,t_0^{s_0}) B = ( t n s n , … , t 0 s 0 ) with s k = L , R s_k=L,R s k = L , R . See [18 , 20 ] for comprehensive discussion.
For spacetime lattice A = { ( t 1 , x 1 ) , … , ( t N , x N ) } \mathcal{A}=\{(t_1,x_1),\ldots,(t_N,x_N)\} A = {( t 1 , x 1 ) , … , ( t N , x N )} with the associated spatiotemporal Hilbert space H A = ⨂ k H ( t k , x k ) \mathcal{H}_{\mathcal{A}}=\bigotimes_k\mathcal{H}_{(t_k,x_k)} H A = ⨂ k H ( t k , x k ) , a spatiotemporal state is an operator of the form
Υ A = 1 d N ∑ μ 1 , … , μ N T μ 1 , … , μ N σ μ 1 ⊗ ⋯ ⊗ σ μ N . (21) \Upsilon_{\mathcal{A}}=\frac{1}{d^N}\sum_{\mu_1,\ldots,\mu_N}T^{\mu_1,\ldots,\mu_N}\sigma_{\mu_1}\otimes\cdots\otimes\sigma_{\mu_N}. \tag*{(21)} Υ A = d N 1 μ 1 , … , μ N ∑ T μ 1 , … , μ N σ μ 1 ⊗ ⋯ ⊗ σ μ N . (21)
Here, σ k \sigma_k σ k are generalized Gell-Mann matrices, and T μ 1 , … , μ N T^{\mu_1,\ldots,\mu_N} T μ 1 , … , μ N is the spatiotemporal joint expectation value of σ μ 1 ( t 1 , x 1 ) , … , σ μ N ( t N , x N ) \sigma_{\mu_1}(t_1,x_1),\ldots,\sigma_{\mu_N}(t_N,x_N) σ μ 1 ( t 1 , x 1 ) , … , σ μ N ( t N , x N ) . Notice that Υ A \Upsilon_{\mathcal{A}} Υ A is generally non-Hermitian and non-normal, but all reduced state Υ ( t k , x k ) \Upsilon_{(t_k,x_k)} Υ ( t k , x k ) is a density operator. The spatiotemporal Kirkwood-Dirac distribution provides a quasiprobability representation of Υ \Upsilon Υ with respect to a chosen operator frame [18 , 20 ] ,
Q ( α 1 , … , α N ) = Tr [ ( F α 1 ( 1 ) ⊗ ⋯ ⊗ F α N ( N ) ) Υ ] (22) Q(\alpha_1,\ldots,\alpha_N)=\operatorname{Tr}[(F_{\alpha_1}^{(1)}\otimes\cdots\otimes F_{\alpha_N}^{(N)})\Upsilon] \tag*{(22)} Q ( α 1 , … , α N ) = Tr [( F α 1 ( 1 ) ⊗ ⋯ ⊗ F α N ( N ) ) Υ ] (22)
with F α k F_{\alpha_k} F α k ’s frames. Thus, the spatiotemporal Kirkwood-Dirac distribution can be viewed as a quasiprobabilistic spatiotemporal state within the framework of generalized probabilistic theory. Write z k = ( t k , x k ) z_k=(t_k,x_k) z k = ( t k , x k ) and let H k \mathcal{H}_k H k be the Hilbert space associated with z k z_k z k . We use the notation
α = ( α 1 , … , α N ) , F α = F α 1 ( 1 ) ⊗ ⋯ ⊗ F α N ( N ) . (23) \alpha=(\alpha_1,\ldots,\alpha_N),\quad F_\alpha=F_{\alpha_1}^{(1)}\otimes\cdots\otimes F_{\alpha_N}^{(N)}. \tag*{(23)} α = ( α 1 , … , α N ) , F α = F α 1 ( 1 ) ⊗ ⋯ ⊗ F α N ( N ) . (23)
The superscript on F α k ( k ) F_{\alpha_k}^{(k)} F α k ( k ) indicates the spacetime site on which the operator acts. We call α \alpha α a spatiotemporal phase-space point, the set of them is the spatiotemporal phase space. See Figure 1 for an illustration. For the standard Kirkwood-Dirac distribution, each local phase-space point is a pair α i = ( a i , b i ) \alpha_i=(a_i,b_i) α i = ( a i , b i ) . Characterizing the full set of Kirkwood-Dirac distributions arising from density operators is generally challenging; partial results have been obtained, see [62 , 63 ] and references therein. For spatiotemporal case, the problem largely remain open.
Figure 1. Illustration of spatiotemporal phase space and spatiotemporal Kirkwood-Dirac distribution. The real part of spatiotemporal Kirkwood-Dirac distribution is a real-valued quasiprobability distribution called Margenau-Hill quasiprobability distribution.
Recall that a positive operator-valued measure (POVM) is a family of positive semidefinite operators { F α } α ⊂ H e r m ≥ 0 ( H ) \{F_\alpha\}_\alpha\subset\mathbf{Herm}_{\geq0}(\mathcal{H}) { F α } α ⊂ Herm ≥ 0 ( H ) satisfying ∑ α F α = 1 \sum_\alpha F_\alpha= 1 ∑ α F α = 1 . A POVM is informationally complete (IC) if its elements span the real vector space of Hermitian operators, span R { F α } = H e r m ( H ) \operatorname{span}_{\mathbb{R}}\{F_\alpha\} = \mathbf{Herm}(\mathcal{H}) span R { F α } = Herm ( H ) . Since every linear operator is a complex linear combination of Hermitian operators, an IC-POVM also spans the complex vector space of all linear operators, span C { F α } = B ( H ) \operatorname{span}_{\mathbb{C}}\{F_\alpha\} = \mathbf{B}(\mathcal{H}) span C { F α } = B ( H ) . If dim H = d \dim\mathcal{H} = d dim H = d , an IC-POVM with d 2 d^2 d 2 elements is called minimal. In finite dimensions, a POVM forms an operator frame if and only if it is IC. Throughout this work, we restrict attention to finite-dimensional Hilbert spaces and take our frames to be minimal IC-POVMs. In this case the dual frame G β G_\beta G β is given by
D β = ∑ α G α β − 1 F α , (24) D_\beta= \sum_\alpha G^{-1}_{\alpha\beta} F_\alpha, \tag*{(24)} D β = α ∑ G α β − 1 F α , (24)
where G α β = Tr [ F α F β ] G_{\alpha\beta} = \operatorname{Tr}[F_\alpha F_\beta] G α β = Tr [ F α F β ] is Gram matrix and we have Tr [ F α D β ] = δ α β \operatorname{Tr}[F_\alpha D_\beta] = \delta_{\alpha\beta} Tr [ F α D β ] = δ α β . For over-complete POVM frame, we need replace the inverse in Eq. (24) with Moore-Penrose pseudoinverse.
For two minimal IC-POVM frames { F α } \{F_\alpha\} { F α } and { F α ′ } \{F'_\alpha\} { F α ′ } with dual frames { D α } \{D_\alpha\} { D α } and { D α ′ } \{D'_\alpha\} { D α ′ } , suppose that they are related by an invertible transformation M M M :
F α ′ = ∑ β M α β F β . (25) F'_\alpha= \sum_\beta M_{\alpha\beta} F_\beta. \tag*{(25)} F α ′ = β ∑ M α β F β . (25)
Then the dual frames transform as
D α ′ = ∑ β ( M − 1 ) β α D β . (26) D'_\alpha= \sum_\beta(M^{-1})_{\beta\alpha} D_\beta. \tag*{(26)} D α ′ = β ∑ ( M − 1 ) β α D β . (26)
At each spacetime point z k z_k z k , we can choose an IC-POVM as the operator frame. The resulting single-time distribution Q ( α k ) = Tr [ F α k ρ t k ] Q(\alpha_k) = \operatorname{Tr}[F_{\alpha_k}\rho_{t_k}] Q ( α k ) = Tr [ F α k ρ t k ] is then a probability distribution. As we will see, this choice is particularly convenient for formulating the spatiotemporal quantum optimal transport problem in phase space, since it is a natural generalization of that for classical phase space.
Defining quantum optimal transport in spacetime
In this section, we define the quantum Monge-Kantorovich optimal transport problem in spacetime in the general setting, which corresponds to the multi-marginal optimal transport problem in classical theory. We refer the reader to [37 ] for basic definitions and results of optimal transport.
Consider a spacetime set A = { ( t 1 , x 1 ) , ⋯ , ( t N , x N ) } \mathcal{A} = \{(t_1, x_1), \cdots, (t_N, x_N)\} A = {( t 1 , x 1 ) , ⋯ , ( t N , x N )} and the corresponding local minimal IC-POVM frames { F α 1 } , ⋯ , { F α N } \{F_{\alpha_1}\}, \cdots, \{F_{\alpha_N}\} { F α 1 } , ⋯ , { F α N } . Set z k = ( t k , x k ) z_k = (t_k, x_k) z k = ( t k , x k ) , suppose that we are given local density operators ρ z k \rho_{z_k} ρ z k and the associated local quantum distributions Q z k ( α k ) = Tr [ F α k ρ z k ] Q_{z_k}(\alpha_k) = \operatorname{Tr}[F_{\alpha_k}\rho_{z_k}] Q z k ( α k ) = Tr [ F α k ρ z k ] . We denote the collection of these local marginal distributions by Q A Q_{\mathcal{A}} Q A . A joint2 spatiotemporal Kirkwood-Dirac distribution Q A ( α 1 , ⋯ , α N ) Q_A(\alpha_1,\cdots,\alpha_N) Q A ( α 1 , ⋯ , α N ) with marginals Q z k ( α k ) Q_{z_k}(\alpha_k) Q z k ( α k ) can be regarded as a spatiotemporal coupling . We denote the set of all such Kirkwood-Dirac spatiotemporal couplings by K D ( Q A ) \mathrm{KD}(Q_A) KD ( Q A ) . The Margenau-Hill spatiotemporal couplings is the real part of Kirkwood-Dirac spatiotemporal couplings, we similarly denote their set as M H ( Q A ) \mathrm{MH}(Q_A) MH ( Q A ) .
For a given cost function C ( α 1 , ⋯ , α N ) C(\alpha_1,\cdots,\alpha_N) C ( α 1 , ⋯ , α N ) , the quantum Monge-Kantorovich optimal transport problem in spatiotemporal phase space is defined as follows
Q M K K D = inf Q A ∈ K D ( Q A ) ∣ ∑ α 1 , … , α N C ( α 1 , ⋯ , α N ) Q A ( α 1 , ⋯ , α N ) ∣ . (27) \mathrm{QMK}_{\mathrm{KD}}=\inf_{Q_A\in\mathrm{KD}(Q_A)}\left|\sum_{\alpha_1,\ldots,\alpha_N}C(\alpha_1,\cdots,\alpha_N)Q_A(\alpha_1,\cdots,\alpha_N)\right|. \tag*{(27)} QMK KD = Q A ∈ KD ( Q A ) inf α 1 , … , α N ∑ C ( α 1 , ⋯ , α N ) Q A ( α 1 , ⋯ , α N ) . (27)
Note that “all such couplings” requires further clarification; specifically, one must specify the physical Kirkwood-Dirac distributions in a chosen frame. As we will see, they are equivalent to the quantum optimal transport based on spatiotemporal state, since Kirkwood-Dirac distribution is obtained from spatiotemporal state via spatiotemporal Born rule [18 , 20 ] . We will give more details later about the spatiotemporal distributions and spatiotemporal states where the optimization is taken over.
For a given set of density operators { ρ z k } z k ∈ A \{\rho_{z_k}\}_{z_k\in A} { ρ z k } z k ∈ A , the set of multi-marginal quantum spatiotemporal state couplings is
C ( ρ z 1 , ⋯ , ρ z N ) = { Υ A ∈ S t a t e S T ( H A ) : Tr z k ‾ Υ A = ρ z k ∀ k } . (28) \mathcal{C}(\rho_{z_1},\cdots,\rho_{z_N})=\left\{\Upsilon_A\in\mathrm{State}_{\mathrm{ST}}(\mathcal{H}_A):\operatorname{Tr}_{\overline{z_k}}\Upsilon_A=\rho_{z_k}\ \forall k\right\}. \tag*{(28)} C ( ρ z 1 , ⋯ , ρ z N ) = { Υ A ∈ State ST ( H A ) : Tr z k Υ A = ρ z k ∀ k } . (28)
where Tr z k ‾ \operatorname{Tr}_{\overline{z_k}} Tr z k denotes the partial trace over all tensor factors except the k k k -th one and S t a t e S T ( H A ) \mathrm{State}_{\mathrm{ST}}(\mathcal{H}_A) State ST ( H A ) denote the set of all spatiotemporal density operators. This set is nonempty since
ρ z 1 ⊗ ⋯ ⊗ ρ z N ∈ C ( ρ z 1 , … , ρ z N ) , (29) \rho_{z_1}\otimes\cdots\otimes\rho_{z_N}\in\mathcal{C}(\rho_{z_1},\ldots,\rho_{z_N}), \tag*{(29)} ρ z 1 ⊗ ⋯ ⊗ ρ z N ∈ C ( ρ z 1 , … , ρ z N ) , (29)
recall the we can use replacement channel as evolution to obtain the spatiotemporal product state.
Every spatiotemporal state Υ A \Upsilon_A Υ A defines a joint spatiotemporal phase-space distribution via the Born rule [18 , 20 ]
Q Υ A ( α ) = Tr [ F α Υ A ] , (30) Q_{\Upsilon_A}(\alpha)=\operatorname{Tr}\left[F_\alpha\Upsilon_A\right], \tag*{(30)} Q Υ A ( α ) = Tr [ F α Υ A ] , (30)
which is the quasiprobabilistic representation of the spatiotemporal state in the chosen frame. Its k k k -th marginal is given by
∑ α 1 , … , α k ^ , … , α N Q Υ A ( α ) = Tr [ F α k ( k ) Tr z k ‾ Υ A ] = Tr [ F α k ( k ) ρ z k ] = Q z k ( α k ) . (31) \begin{aligned}
\sum_{\alpha_1,\ldots,\widehat{\alpha_k},\ldots,\alpha_N}Q_{\Upsilon_A}(\alpha)=\operatorname{Tr}\left[F_{\alpha_k}^{(k)}\operatorname{Tr}_{\overline{z_k}}\Upsilon_A\right] \\
&=\operatorname{Tr}\left[F_{\alpha_k}^{(k)}\rho_{z_k}\right]=Q_{z_k}(\alpha_k). \tag*{(31)}
\end{aligned} α 1 , … , α k , … , α N ∑ Q Υ A ( α ) = Tr [ F α k ( k ) Tr z k Υ A ] = Tr [ F α k ( k ) ρ z k ] = Q z k ( α k ) . (31)
Here the hat over α k \alpha_k α k indicates that this index is not summed over. For each local IC-POVM, we choose a dual operator frame { D α k ( k ) } \{D_{\alpha_k}^{(k)}\} { D α k ( k ) } satisfying
X = ∑ α k Tr [ F α k ( k ) X ] D α k ( k ) (32) X=\sum_{\alpha_k}\operatorname{Tr}\left[F_{\alpha_k}^{(k)}X\right]D_{\alpha_k}^{(k)} \tag*{(32)} X = α k ∑ Tr [ F α k ( k ) X ] D α k ( k ) (32)
for every operator X X X on H z k \mathcal{H}_{z_k} H z k . Given a spatiotemporal Kirkwood-Dirac distribution Q A ( α ) Q_A(\alpha) Q A ( α ) , we define its reconstructed spatiotemporal state by
Υ Q A = ∑ α 1 , … , α N Q A ( α ) D α 1 ( 1 ) ⊗ ⋯ ⊗ D α N ( N ) . (33) \Upsilon_{Q_A}=\sum_{\alpha_1,\ldots,\alpha_N}Q_A(\alpha)D_{\alpha_1}^{(1)}\otimes\cdots\otimes D_{\alpha_N}^{(N)}. \tag*{(33)} Υ Q A = α 1 , … , α N ∑ Q A ( α ) D α 1 ( 1 ) ⊗ ⋯ ⊗ D α N ( N ) . (33)
The set of physical Kirkwood-Dirac spatiotemporal couplings is then
K D ( Q A ) = { Q : Υ ∈ S t a t e S T ( H A ) , Q ( α ) = Tr [ F α Υ Q ] , ∑ α 1 , … , α k ^ , … , α N Q ( α ) = Q z k ( α k ) for every k } . (34) \mathrm{KD}(Q_A)=\left\{Q:\Upsilon\in\mathrm{State}_{\mathrm{ST}}(\mathcal{H}_A),\quad Q(\alpha)=\operatorname{Tr}[F_\alpha\Upsilon_Q],\quad\sum_{\alpha_1,\ldots,\widehat{\alpha_k},\ldots,\alpha_N}Q(\alpha)=Q_{z_k}(\alpha_k)\text{ for every }k\right\}. \tag*{(34)} KD ( Q A ) = ⎩ ⎨ ⎧ Q : Υ ∈ State ST ( H A ) , Q ( α ) = Tr [ F α Υ Q ] , α 1 , … , α k , … , α N ∑ Q ( α ) = Q z k ( α k ) for every k ⎭ ⎬ ⎫ . (34)
When each local IC-POVM is minimal, fixing the local Q z k ( α k ) Q_{z_k}(\alpha_k) Q z k ( α k ) distribution is equivalent to fixing the local density operators ρ z k \rho_{z_k} ρ z k .
From the perspective of generalized probabilistic theory, we can also define the optimal transport beyond quantum theory. Let Q D r C ( Q A ) \mathrm{QD}_{r}^{\mathbb{C}}(Q_A) QD r C ( Q A ) denote the set of all complex-valued quasiprobability distribution with fixed marginals that take values in disk D = { w ∈ C : ∣ w ∣ ≤ r } \mathbb{D}=\{w\in\mathbb{C}:|w|\le r\} D = { w ∈ C : ∣ w ∣ ≤ r } , for sufficiently large r r r ,
K D ( Q A ) ⊂ Q D r C ( Q A ) (35) \mathrm{KD}(Q_A)\subset\mathrm{QD}_{r}^{\mathbb{C}}(Q_A) \tag*{(35)} KD ( Q A ) ⊂ QD r C ( Q A ) (35)
and this inclusion is generally strict. If we allow the coupling to be any elements in Q D r C ( Q A ) \mathrm{QD}_{r}^{\mathbb{C}}(Q_A) QD r C ( Q A ) , we will obtain a optimal transport problem
Q M K Q D r C = inf Q A ∈ Q D r C ( Q A ) ∣ ∑ α 1 , … , α N C ( α 1 , … , α N ) Q A ( α 1 , … , α N ) ∣ . (36) \mathrm{QMK}_{\mathrm{QD}_{r}^{\mathbb{C}}}=\inf_{Q_A\in\mathrm{QD}_{r}^{\mathbb{C}}(Q_A)}\left|\sum_{\alpha_1,\ldots,\alpha_N}C(\alpha_1,\ldots,\alpha_N)Q_A(\alpha_1,\ldots,\alpha_N)\right|. \tag*{(36)} QMK QD r C = Q A ∈ QD r C ( Q A ) inf α 1 , … , α N ∑ C ( α 1 , … , α N ) Q A ( α 1 , … , α N ) . (36)
We have Q M K Q D r C ≤ Q M K K D \mathrm{QMK}_{\mathrm{QD}_{r}^{\mathbb{C}}}\le\mathrm{QMK}_{\mathrm{KD}} QMK QD r C ≤ QMK KD for sufficiently large r r r . When restricting to the real quasiprobability distributions and real cost functions, the problem reduce to the classical optimal transport problem for signed measures. The real quasiprobability distributions coupling also has physical subset of Margenau-Hill distributions M H ( Q A ) \mathrm{MH}(Q_A) MH ( Q A ) , there will also be a corresponding spatiotemporal optimal transport problem Q M K M H \mathrm{QMK}_{\mathrm{MH}} QMK MH . Similarly, Q M K M H ≥ Q M K Q D r R \mathrm{QMK}_{\mathrm{MH}}\ge\mathrm{QMK}_{\mathrm{QD}_{r}^{\mathbb{R}}} QMK MH ≥ QMK QD r R for sufficiently large r r r .
The phase-space cost function determines the cost operator and the cost operator has a unique frame expansion in minimal IC-POVM frame
C = ∑ α C C ( α ) ⨂ k = 1 N F α k ( k ) , C C ( α ) = Tr [ ( ⨂ k = 1 N D α k ( k ) ) C ] . (37) C=\sum_{\alpha}C_C(\alpha)\bigotimes_{k=1}^{N}F_{\alpha_k}^{(k)},\qquad C_C(\alpha)=\operatorname{Tr}\left[\left(\bigotimes_{k=1}^{N}D_{\alpha_k}^{(k)}\right)C\right]. \tag*{(37)} C = α ∑ C C ( α ) k = 1 ⨂ N F α k ( k ) , C C ( α ) = Tr [ ( k = 1 ⨂ N D α k ( k ) ) C ] . (37)
Using the spatiotemporal state, we can also define the spatiotemporal state optimal transport as
inf Υ ∈ C ( ρ z 1 , … , ρ z N ) ∣ Tr [ C Υ ] ∣ , (38) \inf_{\Upsilon\in\mathcal{C}(\rho_{z_1},\ldots,\rho_{z_N})}\left|\operatorname{Tr}[C\Upsilon]\right|, \tag*{(38)} Υ ∈ C ( ρ z 1 , … , ρ z N ) inf ∣ Tr [ C Υ ] ∣ , (38)
which is equivalent to the phase-space formulation in Eq. (27) .
Theorem 3.1 (Equivalence of spatiotemporal phase-space optimal transport and spatiotemporal state optimal transport). For a spacetime set A \mathcal{A} A and minimal IC-POVM frames, the map given by Eq. (30)
Υ A ⟼ Q Υ A (39) \Upsilon_{\mathcal{A}} \longmapsto Q_{\Upsilon_{\mathcal{A}}} \tag*{(39)} Υ A ⟼ Q Υ A (39) is a bijection from C ( ρ z 1 , … , ρ z N ) \mathcal{C}(\rho_{z_1},\ldots,\rho_{z_N}) C ( ρ z 1 , … , ρ z N ) onto K D ( Q A ) \mathsf{KD}(\mathcal{Q}_{\mathcal{A}}) KD ( Q A ) , with inverse given in Eq. (33) . In the chosen frame, we have
Tr [ C Υ ] = ∑ α 1 , … , α N C ( α ) Q ( α ) (40) \operatorname{Tr}[C\Upsilon]=\sum_{\alpha_1,\ldots,\alpha_N} C(\alpha)Q(\alpha) \tag*{(40)} Tr [ C Υ ] = α 1 , … , α N ∑ C ( α ) Q ( α ) (40) where Q ( α ) Q(\alpha) Q ( α ) and C ( α ) C(\alpha) C ( α ) are coefficients in the frame and dual frame. Thus the phase-space optimal transport is equivalent to the spatiotemporal state optimal transport.
Proof. Informational completeness of the local frames implies informational completeness of their tensor product. Consequently,
X = ∑ α Tr [ F α X ] D α 1 ( 1 ) ⊗ ⋯ ⊗ D α N ( N ) (41) X=\sum_{\alpha}\operatorname{Tr}[F_{\alpha}X]D_{\alpha_1}^{(1)}\otimes\cdots\otimes D_{\alpha_N}^{(N)} \tag*{(41)} X = α ∑ Tr [ F α X ] D α 1 ( 1 ) ⊗ ⋯ ⊗ D α N ( N ) (41) for every operator X X X on H z 1 ⊗ ⋯ ⊗ H z N \mathcal{H}_{z_1}\otimes\cdots\otimes\mathcal{H}_{z_N} H z 1 ⊗ ⋯ ⊗ H z N . This proves injectivity and shows that Υ Q Υ A = Υ A \Upsilon_{Q_{\Upsilon_{\mathcal{A}}}}=\Upsilon_{\mathcal{A}} Υ Q Υ A = Υ A . If Υ A \Upsilon_{\mathcal{A}} Υ A has the required reduced states, then Eq. (31) gives the prescribed marginals. Conversely, let Q Q Q satisfy Eq. (34) . For every k k k ,
Tr [ F α k ( k ) Tr k ˉ Υ Q ] = Tr [ F α k ( k ) ρ z k ] . (42) \operatorname{Tr}\left[F_{\alpha_k}^{(k)}\operatorname{Tr}_{\bar{k}}\Upsilon_Q\right]=\operatorname{Tr}\left[F_{\alpha_k}^{(k)}\rho_{z_k}\right]. \tag*{(42)} Tr [ F α k ( k ) Tr k ˉ Υ Q ] = Tr [ F α k ( k ) ρ z k ] . (42) Informational completeness gives Tr k ˉ Υ Q = ρ z k \operatorname{Tr}_{\bar{k}}\Upsilon_Q=\rho_{z_k} Tr k ˉ Υ Q = ρ z k , it follows that Υ Q ∈ C ( ρ z 1 , … , ρ z N ) \Upsilon_Q\in\mathcal{C}(\rho_{z_1},\ldots,\rho_{z_N}) Υ Q ∈ C ( ρ z 1 , … , ρ z N ) . ■
Note that in Eq. (27) we choose the absolute value as the objective function; however, there are many other choices we can consider:
Absolute value r C ( Q ) : = ∣ ∑ α C ( α ) Q ( α ) ∣ = ∣ Tr [ C Υ ] ∣ = : r C ( Υ ) r_C(Q):=\left|\sum_{\alpha}C(\alpha)Q(\alpha)\right|=\left|\operatorname{Tr}[C\Upsilon]\right|=:r_C(\Upsilon) r C ( Q ) := ∣ ∑ α C ( α ) Q ( α ) ∣ = ∣ Tr [ C Υ ] ∣ =: r C ( Υ ) .
Phase value θ C ( Q ) : = − i Log ( ∑ α C ( α ) Q ( α ) ∣ ∑ α C ( α ) Q ( α ) ∣ ) = − i Log ( Tr [ C Υ ] ∣ Tr [ C Υ ] ∣ ) = : θ C ( Υ ) \theta_C(Q):=-i\operatorname{Log}\left(\frac{\sum_{\alpha}C(\alpha)Q(\alpha)}{\left|\sum_{\alpha}C(\alpha)Q(\alpha)\right|}\right)=-i\operatorname{Log}\left(\frac{\operatorname{Tr}[C\Upsilon]}{|\operatorname{Tr}[C\Upsilon]|}\right)=:\theta_C(\Upsilon) θ C ( Q ) := − i Log ( ∣ ∑ α C ( α ) Q ( α ) ∣ ∑ α C ( α ) Q ( α ) ) = − i Log ( ∣ Tr [ C Υ ] ∣ Tr [ C Υ ] ) =: θ C ( Υ ) .
Real part u C ( Q ) = Re ∑ α C ( α ) Q ( α ) = Re Tr [ C Υ ] = : u C ( Υ ) u_C(Q)=\operatorname{Re}\sum_{\alpha}C(\alpha)Q(\alpha)=\operatorname{Re}\operatorname{Tr}[C\Upsilon]=:u_C(\Upsilon) u C ( Q ) = Re ∑ α C ( α ) Q ( α ) = Re Tr [ C Υ ] =: u C ( Υ ) .
Imaginary part v C ( Q ) = Im ∑ α C ( α ) Q ( α ) = Im Tr [ C Υ ] = : v C ( Υ ) v_C(Q)=\operatorname{Im}\sum_{\alpha}C(\alpha)Q(\alpha)=\operatorname{Im}\operatorname{Tr}[C\Upsilon]=:v_C(\Upsilon) v C ( Q ) = Im ∑ α C ( α ) Q ( α ) = Im Tr [ C Υ ] =: v C ( Υ ) .
We will mainly focus on the absolute value and real part cases, since they are directly related the those considered in spatial quantum optimal transport.
For a general operator X X X , define its Hermitian components by
X R = X + X † 2 , X I = X − X † 2 i . (43) X_{\mathrm{R}}=\frac{X+X^\dagger}{2},\qquad X_{\mathrm{I}}=\frac{X-X^\dagger}{2i}. \tag*{(43)} X R = 2 X + X † , X I = 2 i X − X † . (43)
When X ∈ Herm ( H ) X\in\operatorname{Herm}(\mathcal{H}) X ∈ Herm ( H ) , we have X R = X X_{\mathrm{R}}=X X R = X and X I = 0 X_{\mathrm{I}}=0 X I = 0 . For every Υ ∈ State S T \Upsilon\in\operatorname{State}_{\mathrm{ST}} Υ ∈ State ST and cost operator C C C , we have
Tr [ C Υ ] = Tr [ C R Υ R − C I Υ I ] + i Tr [ C I Υ R + C R Υ I ] . (44) \operatorname{Tr}[C\Upsilon]=\operatorname{Tr}[C_{\mathrm{R}}\Upsilon_{\mathrm{R}}-C_{\mathrm{I}}\Upsilon_{\mathrm{I}}]+i\operatorname{Tr}[C_{\mathrm{I}}\Upsilon_{\mathrm{R}}+C_{\mathrm{R}}\Upsilon_{\mathrm{I}}]. \tag*{(44)} Tr [ C Υ ] = Tr [ C R Υ R − C I Υ I ] + i Tr [ C I Υ R + C R Υ I ] . (44)
When C ∈ Herm ( H ) C \in\operatorname{Herm}(\mathcal{H}) C ∈ Herm ( H ) , Re Tr [ C Υ ] = Tr [ C Υ R ] \operatorname{Re}\operatorname{Tr}[C\Upsilon] = \operatorname{Tr}[C\Upsilon_{\mathrm{R}}] Re Tr [ C Υ ] = Tr [ C Υ R ] . Thus, for a Hermitian cost operator and a real-part objective function, the Kirkwood-Dirac quantum optimal transport reduces to the Margenau-Hill quantum optimal transport. When Υ ∈ Herm ( H ) \Upsilon\in\operatorname{Herm}(\mathcal{H}) Υ ∈ Herm ( H ) , namely in the Margenau-Hill phase space, we have Re Tr [ C Υ ] = Tr [ C R Υ ] \operatorname{Re}\operatorname{Tr}[C\Upsilon] = \operatorname{Tr}[C_{\mathrm{R}}\Upsilon] Re Tr [ C Υ ] = Tr [ C R Υ ] . When both C C C and Υ \Upsilon Υ are Hermitian, we have Re Tr [ C Υ ] = Tr [ C Υ ] \operatorname{Re}\operatorname{Tr}[C\Upsilon] = \operatorname{Tr}[C\Upsilon] Re Tr [ C Υ ] = Tr [ C Υ ] . This means that in the above three cases we always obtain the Margenau-Hill quantum optimal transport. The real-part objective function also gives the correct classical limit, whereas the absolute value function introduces an extra absolute value sign.
Optimal transport in projector-product Kirkwood-Dirac phase space
The previous discussion mainly use the IC-POVM frames, we can also define the optimal transport using projector-product frame. For H z k \mathcal{H}_{z_k} H z k , let { ∣ a k ⟩ } \{|a_k\rangle\} { ∣ a k ⟩} and { ∣ a k ′ ⟩ } \{|a'_k\rangle\} { ∣ a k ′ ⟩} be two orthonormal bases with nonzero mutual overlaps, and write
P a k = ∣ a k ⟩ ⟨ a k ∣ , P ~ a k ′ = ∣ a k ′ ⟩ ⟨ a k ′ ∣ , K a k a k ′ A = P a k P ~ a k ′ . (45) P_{a_k} = |a_k\rangle\langle a_k|,\qquad\widetilde{P}_{a'_k} = |a'_k\rangle\langle a'_k|,\qquad K^{A}_{a_k a'_k} = P_{a_k}\widetilde{P}_{a'_k}. \tag*{(45)} P a k = ∣ a k ⟩ ⟨ a k ∣ , P a k ′ = ∣ a k ′ ⟩ ⟨ a k ′ ∣ , K a k a k ′ A = P a k P a k ′ . (45)
The local Kirkwood-Dirac distributions are
Q ρ z k ( a k , a k ′ ) = Tr [ ( K a k a k ′ ) † ρ z k ] , (46) Q_{\rho_{z_k}}(a_k,a'_k) = \operatorname{Tr}\left[(K_{a_k a'_k})^\dagger\rho_{z_k}\right], \tag*{(46)} Q ρ z k ( a k , a k ′ ) = Tr [ ( K a k a k ′ ) † ρ z k ] , (46)
which, unlike those for IC-POVM frames, are complex-valued quasiprobability distributions. The projector-product Kirkwood-Dirac couplings are K D p r o j ( Q A ) \mathrm{KD}^{\mathrm{proj}}(\mathcal{Q}_A) KD proj ( Q A ) . For cost function C ( a 1 , a 1 ′ ; ⋯ ; a N , a N ′ ) C(a_1,a'_1;\cdots;a_N,a'_N) C ( a 1 , a 1 ′ ; ⋯ ; a N , a N ′ ) , define
C = ∑ a 1 , a 1 ′ ; ⋯ ; a N , a N ′ C ( a 1 , a 1 ′ ; ⋯ ; a N , a N ′ ) ( K a 1 , a 1 ′ A ) † ⊗ ⋯ ⊗ ( K a N , a N ′ B ) † . (47) C = \sum_{a_1,a'_1;\cdots;a_N,a'_N} C(a_1,a'_1;\cdots;a_N,a'_N)(K^{A}_{a_1,a'_1})^\dagger\otimes\cdots\otimes(K^{B}_{a_N,a'_N})^\dagger. \tag*{(47)} C = a 1 , a 1 ′ ; ⋯ ; a N , a N ′ ∑ C ( a 1 , a 1 ′ ; ⋯ ; a N , a N ′ ) ( K a 1 , a 1 ′ A ) † ⊗ ⋯ ⊗ ( K a N , a N ′ B ) † . (47)
Then
inf Q ∈ K D p r o j ( Q A ) ∣ ∑ a 1 , a 1 ′ ; ⋯ ; a N , a N ′ C ( a 1 , a 1 ′ ; ⋯ ; a N , a N ′ ) Q ( a 1 , a 1 ′ ; ⋯ ; a N , a N ′ ) ∣ = inf Υ ∈ C ( ρ z 1 , ⋯ , ρ z N ) ∣ Tr ( C Υ ) ∣ . (48) \inf_{Q\in\mathrm{KD}^{\mathrm{proj}}(\mathcal{Q}_A)}
\left|
\sum_{a_1,a'_1;\cdots;a_N,a'_N}
C(a_1,a'_1;\cdots;a_N,a'_N)Q(a_1,a'_1;\cdots;a_N,a'_N)
\right|
=
\inf_{\Upsilon\in\mathcal{C}(\rho_{z_1},\cdots,\rho_{z_N})}
\left|\operatorname{Tr}(C\Upsilon)\right|. \tag*{(48)} Q ∈ KD proj ( Q A ) inf a 1 , a 1 ′ ; ⋯ ; a N , a N ′ ∑ C ( a 1 , a 1 ′ ; ⋯ ; a N , a N ′ ) Q ( a 1 , a 1 ′ ; ⋯ ; a N , a N ′ ) = Υ ∈ C ( ρ z 1 , ⋯ , ρ z N ) inf ∣ Tr ( C Υ ) ∣ . (48)
Spatial optimal transport in spatial phase space
We now specialize the preceding spatiotemporal construction to a purely spatial setting A = ( x 1 , ⋯ , x N ) \mathcal{A}=(x_1,\cdots,x_N) A = ( x 1 , ⋯ , x N ) . This specialization identifies the phase-space formulation with the density-operator formulation of quantum optimal transport and clarifies which features are specifically spatiotemporal. The main difference is that a spatial transport plan is a positive multipartite density operator, whereas a general spatiotemporal state need not be positive.
Let H x 1 , … , H x N \mathcal{H}_{x_1},\ldots,\mathcal{H}_{x_N} H x 1 , … , H x N be finite-dimensional Hilbert spaces, let H A = ⨂ k = 1 N H x k \mathcal{H}_A = \bigotimes_{k=1}^{N}\mathcal{H}_{x_k} H A = ⨂ k = 1 N H x k , and fix density operators ρ k \rho_k ρ k on each H x k \mathcal{H}_{x_k} H x k . Write x ^ k = { x 1 , … , x N } ∖ { x k } \widehat{x}_k = \{x_1,\ldots,x_N\}\setminus\{x_k\} x k = { x 1 , … , x N } ∖ { x k } , so that Tr x ^ k \operatorname{Tr}_{\widehat{x}_k} Tr x k traces out every subsystem except k k k . The set of spatial quantum couplings with these marginals is
C ( ρ 1 , … , ρ N ) = { Υ A ≥ 0 : Tr Υ A = 1 , Tr x ^ k Υ A = ρ k for every k } . (49) \mathcal{C}(\rho_1,\ldots,\rho_N)=\left\{\Upsilon_A\geq0:\operatorname{Tr}\Upsilon_A=1,\ \operatorname{Tr}_{\widehat{x}_k}\Upsilon_A=\rho_k\ \text{for every }k\right\}. \tag*{(49)} C ( ρ 1 , … , ρ N ) = { Υ A ≥ 0 : Tr Υ A = 1 , Tr x k Υ A = ρ k for every k } . (49)
This set is nonempty, compact, and convex; in particular, ⨂ k = 1 N ρ k ∈ C ( ρ 1 , … , ρ N ) \bigotimes_{k=1}^{N}\rho_k \in\mathcal{C}(\rho_1,\ldots,\rho_N) ⨂ k = 1 N ρ k ∈ C ( ρ 1 , … , ρ N ) .
Fix minimal IC-POVM frame and phase space Ω A = ∏ k = 1 N Ω x k \Omega_A=\prod_{k=1}^{N}\Omega_{x_k} Ω A = ∏ k = 1 N Ω x k , every spatial coupling Υ A \Upsilon_A Υ A has the joint phase-space representation
Q Υ ( α ) = Tr [ F α Υ A ] . (50) Q_{\Upsilon}(\alpha)=\operatorname{Tr}[F_{\alpha}\Upsilon_A]. \tag*{(50)} Q Υ ( α ) = Tr [ F α Υ A ] . (50)
It satisfies
Q Υ ( α ) ≥ 0 , ∑ α Q Υ ( α ) = 1 , (51) Q_{\Upsilon}(\alpha)\geq0,\qquad\sum_{\alpha}Q_{\Upsilon}(\alpha)=1, \tag*{(51)} Q Υ ( α ) ≥ 0 , α ∑ Q Υ ( α ) = 1 , (51)
and its single-subsystem marginals are
∑ α k ˉ Q Υ ( α ) = Q x k ( α k ) = Tr [ F α k ρ k ] , k = 1 , … , N , (52) \sum_{\alpha_{\bar{k}}}Q_{\Upsilon}(\alpha)=Q_{x_k}(\alpha_k)=\operatorname{Tr}[F_{\alpha_k}\rho_k],\qquad k=1,\ldots,N, \tag*{(52)} α k ˉ ∑ Q Υ ( α ) = Q x k ( α k ) = Tr [ F α k ρ k ] , k = 1 , … , N , (52)
where the sum runs over all outcome labels except α k \alpha_k α k .
An arbitrary classical coupling of Q x 1 , … , Q x N Q_{x_1},\ldots,Q_{x_N} Q x 1 , … , Q x N need not arise from a spatial quantum coupling. Accordingly, we define the physical spatial phase-space transport coupling as
K D s p ( p 1 , … , p N ) = { Q : Ω A ⟶ R : ∑ α k ˉ Q ( α ) = p k ( α k ) ∀ k , Υ Q ∈ S t a t e s } . (53) \mathsf{KD}_{\mathrm{sp}}(p_1,\ldots,p_N)=\left\{Q:\Omega_A\longrightarrow\mathbb{R}:\sum_{\alpha_{\bar{k}}}Q(\alpha)=p_k(\alpha_k)\ \forall k,\ \Upsilon_Q\in\mathbf{States}\right\}. \tag*{(53)} KD sp ( p 1 , … , p N ) = ⎩ ⎨ ⎧ Q : Ω A ⟶ R : α k ˉ ∑ Q ( α ) = p k ( α k ) ∀ k , Υ Q ∈ States ⎭ ⎬ ⎫ . (53)
We can equivalently replace condition Υ Q ∈ S t a t e s \Upsilon_Q\in\mathbf{States} Υ Q ∈ States with Υ Q ≥ 0 \Upsilon_Q\geq0 Υ Q ≥ 0 . The marginal conditions imply Tr Υ Q = 1 \operatorname{Tr}\Upsilon_Q=1 Tr Υ Q = 1 . Thus, within the set of classical couplings, quantum realizability is characterized by the semidefinite constraint Υ Q ≥ 0 \Upsilon_Q\geq0 Υ Q ≥ 0 .
In particular, for any real-valued cost function C : Ω A → R C:\Omega_A\to\mathbb{R} C : Ω A → R , define cost operator as that in Eq. (37) . The Theorem 3.1 shows that spatial phase-space optimal transport is equivalent to joint density operator optimal transport
inf Q ∈ K D s ( Q x 1 , … , Q x N ) ∑ α C ( α ) Q ( α ) = inf Υ A ∈ C ( ρ 1 , … , ρ N ) Tr ( C Υ A ) . (54) \inf_{Q\in\mathsf{KD}_{\mathrm{s}}(Q_{x_1},\ldots,Q_{x_N})}\sum_{\alpha}C(\alpha)Q(\alpha)=\inf_{\Upsilon_A\in\mathcal{C}(\rho_1,\ldots,\rho_N)}\operatorname{Tr}(C\Upsilon_A). \tag*{(54)} Q ∈ KD s ( Q x 1 , … , Q x N ) inf α ∑ C ( α ) Q ( α ) = Υ A ∈ C ( ρ 1 , … , ρ N ) inf Tr ( C Υ A ) . (54)
For a nonnegative cost function, C ≥ 0 C\geq0 C ≥ 0 , it reduces to the quantum optimal transport discussed in [40 , 41 , 42 , 43 , 44 , 45 , 46 , 47 , 48 , 49 , 50 ] .
Conversely, every cost operator C C C on H A \mathcal{H}_A H A has the unique frame expansion given in Eq. (37) . If C C C is Hermitian, this kernel is real. However, C ≥ 0 C\geq0 C ≥ 0 does not generally imply C ( α ) = Tr D α C ≥ 0 C(\alpha)=\operatorname{Tr}D_{\alpha}C\geq0 C ( α ) = Tr D α C ≥ 0 , because the dual-frame elements need not be positive. Nevertheless, positivity of C C C ensures Tr ( C Υ A ) ≥ 0 \operatorname{Tr}(C\Upsilon_A)\geq0 Tr ( C Υ A ) ≥ 0 for every physical coupling.
Proposition 4.1 (Frame independence). Fix a set of local density operators ρ 1 , … , ρ N \rho_1,\ldots,\rho_N ρ 1 , … , ρ N , consider two collections of local minimal IC-POVM frames, the corresponding joint phase-space distributions are related by an invertible linear transformation, while the cost function transform contragrediently. The optimal Monge-Kantorovich value is unchanged.
Proof. Write the change of local frames as
F ~ μ k ( k ) = ∑ α k M μ k α k ( k ) F α k ( k ) , (55) \widetilde{F}^{(k)}_{\mu_k}=\sum_{\alpha_k}M^{(k)}_{\mu_k\alpha_k}F^{(k)}_{\alpha_k}, \tag*{(55)} F μ k ( k ) = α k ∑ M μ k α k ( k ) F α k ( k ) , (55)
where each M ( k ) M^{(k)} M ( k ) is a real invertible matrix, and set M = ⨂ k = 1 N M ( k ) M = \bigotimes_{k=1}^{N} M^{(k)} M = ⨂ k = 1 N M ( k ) . Viewing the distributions and cost functions as column vectors, the joint distributions satisfy
Q ~ Υ ~ = M Q Υ . (56) \widetilde{Q}_{\widetilde{\Upsilon}} = M Q_{\Upsilon}. \tag*{(56)} Q Υ = M Q Υ . (56)
To represent the same cost operator, the kernels must satisfy
C ~ C = ( M − 1 ) T C C . (57) \widetilde{C}_{\mathcal C} = (M^{-1})^{T} C_{\mathcal C}. \tag*{(57)} C C = ( M − 1 ) T C C . (57)
Consequently,
∑ μ C ~ C ( μ ) Q ~ Υ ~ ( μ ) = ∑ α C C ( α ) Q Υ ( α ) = Tr ( C Υ A ) . (58) \sum_{\mu} \widetilde{C}_{\mathcal C}(\mu)\widetilde{Q}_{\widetilde{\Upsilon}}(\mu) = \sum_{\alpha} C_{\mathcal C}(\alpha)Q_{\Upsilon}(\alpha) = \operatorname{Tr}(C\Upsilon_A). \tag*{(58)} μ ∑ C C ( μ ) Q Υ ( μ ) = α ∑ C C ( α ) Q Υ ( α ) = Tr ( C Υ A ) . (58)
In either collection of frames, the admissible phase-space couplings are precisely the coordinate images of C ( ρ 1 , … , ρ N ) \mathcal C(\rho_1,\ldots,\rho_N) C ( ρ 1 , … , ρ N ) . The frame transformation therefore gives a bijection between the two admissible sets and preserves the objective function. Hence the optimal values coincide. ■
For a fixed operator C C C , changing the local frames changes only its coordinates and those of the couplings. By contrast, keeping cost function fixed while changing the frames generally changes the associated cost operator and hence the transport problem.
Quantum Kantorovich duality
We now establish the Kantorovich duality for spatial quantum optimal transport and discuss how to represent it in phase space by generalizing the discussing in [42 , 48 ] for positive semidefinite cost operator to general case.
The key observation is that, for Hermitian cost operator, if we consider the objective function Tr C Υ \operatorname{Tr} C\Upsilon Tr C Υ (without modulus), the Kantorovich duality is actually the dual problem of SDP of quantum optimal transport. We consider the primal problem of an SDP of the following form [64 , 65 , 66 ] :
inf X ⟨ C , X ⟩ = Tr ( C X ) subject to X ≥ 0 , E ( X ) = B , (59) \begin{aligned}
\inf_X\quad&\langle C,X\rangle= \operatorname{Tr}(CX)\\
\text{subject to}\quad&X \ge0,\\
&\mathcal E(X)=B, \tag*{(59)}
\end{aligned} X inf subject to ⟨ C , X ⟩ = Tr ( C X ) X ≥ 0 , E ( X ) = B , (59)
where E : Herm ( H ) → Herm ( V ) \mathcal E:\operatorname{Herm}(\mathcal H)\to\operatorname{Herm}(\mathcal V) E : Herm ( H ) → Herm ( V ) is linear and Hermiticity preserving. The dual problem is
sup Y ⟨ B , Y ⟩ = Tr ( B Y ) subject to E † ( Y ) ≤ C , (60) \begin{aligned}
\sup_Y\quad&\langle B,Y\rangle= \operatorname{Tr}(BY)\\
\text{subject to}\quad&\mathcal E^\dagger(Y)\le C, \tag*{(60)}
\end{aligned} Y sup subject to ⟨ B , Y ⟩ = Tr ( B Y ) E † ( Y ) ≤ C , (60)
where the inner product is the Hilbert–Schmidt inner product and E † \mathcal E^\dagger E † is the adjoint of E \mathcal E E with respect to this inner product.
Assume that the cost operator is Hermitian (not necessarily positive semidefinite), C = C † C=C^\dagger C = C † . Then spatial optimal transport can be recast as an SDP
QMK C ( ρ A , ρ B ) = inf Υ Tr ( C Υ ) subject to Υ ≥ 0 , Tr B Υ = ρ A , Tr A Υ = ρ B . (61) \begin{aligned}
\operatorname{QMK}_{C}(\rho_A,\rho_B)&=\inf_{\Upsilon}\quad\operatorname{Tr}(C\Upsilon)\\
\text{subject to}\quad&\Upsilon\ge0,\\
&\operatorname{Tr}_{B}\Upsilon=\rho_A,\\
&\operatorname{Tr}_{A}\Upsilon=\rho_B. \tag*{(61)}
\end{aligned} QMK C ( ρ A , ρ B ) subject to = Υ inf Tr ( C Υ ) Υ ≥ 0 , Tr B Υ = ρ A , Tr A Υ = ρ B . (61)
The marginal constraints imply Tr Υ = 1 \operatorname{Tr}\Upsilon= 1 Tr Υ = 1 . Thus the feasible set is nonempty and compact, and the minimum is attained. In particular, the eigenvalues of C C C give the following bounds
λ min ( C ) ≤ QMK C ( ρ A , ρ B ) ≤ λ max ( C ) , (62) \lambda_{\min}(C) \leq\operatorname{QMK}_{C}(\rho_A,\rho_B) \leq\lambda_{\max}(C), \tag*{(62)} λ m i n ( C ) ≤ QMK C ( ρ A , ρ B ) ≤ λ m a x ( C ) , (62)
since λ min ( C ) ≤ Tr C Υ ≤ λ max ( C ) \lambda_{\min}(C) \leq\operatorname{Tr} C\Upsilon\leq\lambda_{\max}(C) λ m i n ( C ) ≤ Tr C Υ ≤ λ m a x ( C ) for every density operator Υ \Upsilon Υ .
Notice that for SDP in Eq. (61)
E : Herm ( H A ⊗ H B ) ⟶ Herm ( H A ) ⊕ Herm ( H B ) , \mathcal{E}:\operatorname{Herm}(\mathcal{H}_A \otimes\mathcal{H}_B) \longrightarrow\operatorname{Herm}(\mathcal{H}_A) \oplus\operatorname{Herm}(\mathcal{H}_B), E : Herm ( H A ⊗ H B ) ⟶ Herm ( H A ) ⊕ Herm ( H B ) ,
X ⟼ ( Tr B X , Tr A X ) . X \longmapsto(\operatorname{Tr}_B X,\operatorname{Tr}_A X). X ⟼ ( Tr B X , Tr A X ) .
the adjoint is given by adjoint is
E † : Herm ( H A ) ⊕ Herm ( H B ) ⟶ Herm ( H A ⊗ H B ) , \mathcal{E}^{\dagger}:\operatorname{Herm}(\mathcal{H}_A) \oplus\operatorname{Herm}(\mathcal{H}_B) \longrightarrow\operatorname{Herm}(\mathcal{H}_A \otimes\mathcal{H}_B), E † : Herm ( H A ) ⊕ Herm ( H B ) ⟶ Herm ( H A ⊗ H B ) ,
( Y A , Y B ) ⟼ Y A ⊗ I B + I A ⊗ Y B . (Y_A,Y_B) \longmapsto Y_A \otimes I_B + I_A \otimes Y_B. ( Y A , Y B ) ⟼ Y A ⊗ I B + I A ⊗ Y B .
the dual problem in this setting is called quantum Kantorovich duality:
QMK ^ C ( ρ A , ρ B ) = sup X = X † , Y = Y † { Tr ( X ρ A ) + Tr ( Y ρ B ) : X ⊗ I B + I A ⊗ Y ≤ C } . (63) \widehat{\operatorname{QMK}}_{C}(\rho_A,\rho_B) =
\sup_{X=X^\dagger,\;Y=Y^\dagger}
\left\{\operatorname{Tr}(X\rho_A) + \operatorname{Tr}(Y\rho_B) : X \otimes I_B + I_A \otimes Y \leq C\right\}. \tag*{(63)} QMK C ( ρ A , ρ B ) = X = X † , Y = Y † sup { Tr ( X ρ A ) + Tr ( Y ρ B ) : X ⊗ I B + I A ⊗ Y ≤ C } . (63)
where X , Y X,Y X , Y can be regarded as quantum generalization of Kantorovich potential.
Theorem 4.2 (Quantum Kantorovich duality for Hermitian cost operator). For every Hermitian cost operator C C C , quantum Kantorovich duality is given in Eq. (63) . The weak duality of SDP implies that QMK ^ C ( ρ A , ρ B ) ≤ QMK C ( ρ A , ρ B ) \widehat{\operatorname{QMK}}_{C}(\rho_A,\rho_B) \leq\operatorname{QMK}_{C}(\rho_A,\rho_B) QMK C ( ρ A , ρ B ) ≤ QMK C ( ρ A , ρ B ) . Since the primal is feasible and the dual is strictly feasible, so Slater’s theorem applies (see Theorem A.2 in Appendix), we in fact have strong duality:
QMK ^ C ( ρ A , ρ B ) = QMK C ( ρ A , ρ B ) (64) \widehat{\operatorname{QMK}}_{C}(\rho_A,\rho_B) = \operatorname{QMK}_{C}(\rho_A,\rho_B) \tag*{(64)} QMK C ( ρ A , ρ B ) = QMK C ( ρ A , ρ B ) (64)
The primal infimum is always attained. If ρ A \rho_A ρ A and ρ B \rho_B ρ B are strictly positive, the dual supremum is also attained. For arbitrary marginals, both optima are attained after restricting the Hilbert spaces to their supports and compressing the cost operator to the resulting tensor-product space.
Proof. Introduce Hermitian Lagrange multipliers X X X and Y Y Y for the marginal constraints. The Lagrangian is
L ( Υ , X , Y ) = Tr ( C Υ ) + Tr [ X ( ρ A − Tr B Υ ) ] + Tr [ Y ( ρ B − Tr A Υ ) ] = Tr ( X ρ A ) + Tr ( Y ρ B ) + Tr [ ( C − X ⊗ I B − I A ⊗ Y ) Υ ] . (65) \begin{aligned}
L(\Upsilon,X,Y) &= \operatorname{Tr}(C\Upsilon) + \operatorname{Tr}\left[X(\rho_A-\operatorname{Tr}_B\Upsilon)\right] + \operatorname{Tr}\left[Y(\rho_B-\operatorname{Tr}_A\Upsilon)\right] \\
&= \operatorname{Tr}(X\rho_A) + \operatorname{Tr}(Y\rho_B) \\
&\quad+ \operatorname{Tr}\left[(C-X\otimes I_B-I_A\otimes Y)\Upsilon\right]. \tag*{(65)}
\end{aligned} L ( Υ , X , Y ) = Tr ( C Υ ) + Tr [ X ( ρ A − Tr B Υ ) ] + Tr [ Y ( ρ B − Tr A Υ ) ] = Tr ( X ρ A ) + Tr ( Y ρ B ) + Tr [ ( C − X ⊗ I B − I A ⊗ Y ) Υ ] . (65)
The infimum over Υ ≥ 0 \Upsilon\geq0 Υ ≥ 0 is finite precisely when
C − X ⊗ I B − I A ⊗ Y ≥ 0. (66) C-X\otimes I_B-I_A\otimes Y \geq0. \tag*{(66)} C − X ⊗ I B − I A ⊗ Y ≥ 0. (66)
When the operator is positive, the infimum of the last term is zero. This gives the stated dual problem.
By Slater’s condition [66 Theorem 1.1] , since the primal feasible set is non-empty and for every Hermitian C C C : choose X = − t I A X=-tI_A X = − t I A and Y = − t I B Y=-tI_B Y = − t I B , with 2 t > − λ min ( C ) 2t>-\lambda_{\min}(C) 2 t > − λ m i n ( C ) . Then
C − X ⊗ I B − I A ⊗ Y = C + 2 t I A ⊗ I B > 0. (67) C-X\otimes I_B-I_A\otimes Y=C+2tI_A\otimes I_B>0. \tag*{(67)} C − X ⊗ I B − I A ⊗ Y = C + 2 t I A ⊗ I B > 0. (67)
Then we have strong duality as in Eq. (64) and there exists a primal feasible spatial state Υ ⋆ \Upsilon_\star Υ ⋆ for which QMK C ( ρ A , ρ B ) = Tr [ C Υ ⋆ ] \operatorname{QMK}_C(\rho_A,\rho_B)=\operatorname{Tr}[C\Upsilon_\star] QMK C ( ρ A , ρ B ) = Tr [ C Υ ⋆ ] .
Similarly, when ρ A \rho_A ρ A and ρ B \rho_B ρ B are strictly positive, since the dual feasible set is non-empty and there exists ρ A ⊗ ρ B > 0 \rho_A\otimes\rho_B>0 ρ A ⊗ ρ B > 0 that is primally feasible, we have strong duality as in Eq. (64) and there exist X ⋆ X_\star X ⋆ and Y ⋆ Y_\star Y ⋆ such that Tr [ X ⋆ ρ A ] + Tr [ Y ⋆ ρ B ] = QMK ^ C ( ρ A , ρ B ) \operatorname{Tr}[X_\star\rho_A]+\operatorname{Tr}[Y_\star\rho_B]=\widehat{\operatorname{QMK}}_C(\rho_A,\rho_B) Tr [ X ⋆ ρ A ] + Tr [ Y ⋆ ρ B ] = QMK C ( ρ A , ρ B ) .
The above argument directly implies that, for general ρ A \rho_A ρ A and ρ B \rho_B ρ B when restricted to tensor product space of their supports, both optima are attained. ■
The dual variables also possess phase-space representations. Define
φ X ( α ) = Tr ( D α X ) , ψ Y ( β ) = Tr ( E β Y ) . (68) \varphi_X(\alpha)=\operatorname{Tr}(D_\alpha X),\qquad\psi_Y(\beta)=\operatorname{Tr}(E_\beta Y). \tag*{(68)} φ X ( α ) = Tr ( D α X ) , ψ Y ( β ) = Tr ( E β Y ) . (68)
Then
Tr ( X ρ A ) + Tr ( Y ρ B ) = ∑ α φ X ( α ) p ρ A ( α ) + ∑ β ψ Y ( β ) p ρ B ( β ) . (69) \operatorname{Tr}(X\rho_A)+\operatorname{Tr}(Y\rho_B)=\sum_\alpha\varphi_X(\alpha)p_{\rho_A}(\alpha)+\sum_\beta\psi_Y(\beta)p_{\rho_B}(\beta). \tag*{(69)} Tr ( X ρ A ) + Tr ( Y ρ B ) = α ∑ φ X ( α ) p ρ A ( α ) + β ∑ ψ Y ( β ) p ρ B ( β ) . (69)
These functions are the quantum analogues of Kantorovich potentials in phase space. Unlike classical Kantorovich potentials, however, they are constrained by the operator inequality (66) , rather than only by pointwise inequalities.
For general non-Hermitian cost operator C C C , we need consider the objective function with modulus ∣ Tr C Υ ∣ \lvert\operatorname{Tr}C\Upsilon\rvert ∣ Tr C Υ ∣ , in this case, we have the following duality:
Theorem 4.3 (Quantum Kantorovich duality for non-Hermitian cost operator). For general cost operator C ∈ B ( H ) C\in\mathcal{B}(\mathcal{H}) C ∈ B ( H ) , the Kantorovich duality is given by
QKD ^ C ( ρ A , ρ B ) = sup u 1 , u 2 ∈ R u 1 2 + u 2 2 ≤ 1 sup X = X † , Y = Y † { Tr ( X ρ A ) + Tr ( Y ρ B ) : X ⊗ I B + I A ⊗ Y ≤ u 1 C R + u 2 C I } . (70) \widehat{\operatorname{QKD}}_C(\rho_A,\rho_B)=\sup_{\substack{u_1,u_2\in\mathbb{R}\\u_1^2+u_2^2\leq1}}\ \sup_{\substack{X=X^\dagger,Y=Y^\dagger}}\left\{\operatorname{Tr}(X\rho_A)+\operatorname{Tr}(Y\rho_B):X\otimes I_B+I_A\otimes Y\leq u_1C_R+u_2C_I\right\}. \tag*{(70)} QKD C ( ρ A , ρ B ) = u 1 , u 2 ∈ R u 1 2 + u 2 2 ≤ 1 sup X = X † , Y = Y † sup { Tr ( X ρ A ) + Tr ( Y ρ B ) : X ⊗ I B + I A ⊗ Y ≤ u 1 C R + u 2 C I } . (70) The strong duality holds as in Eq. (64) . The primal infimum is always attained. If ρ A \rho_A ρ A and ρ B \rho_B ρ B are strictly positive, the dual supremum is also attained. For arbitrary marginals, both optima are attained after restricting the Hilbert spaces to their supports and compressing the cost operator to the resulting tensor-product space.
Proof. For every complex number z z z , when treated as two-dimensional real vectors, Cauchy–Schwarz inequality implies
∣ z ∣ = sup u 1 2 + u 2 2 ≤ 1 ( u 1 Re z + u 2 Im z ) = sup ∣ u ∣ ≤ 1 Re u ‾ z . (71) \lvert z\rvert=\sup_{u_1^2+u_2^2\leq1}(u_1\operatorname{Re}z+u_2\operatorname{Im}z)=\sup_{\lvert u\rvert\leq1}\operatorname{Re}\overline{u}z. \tag*{(71)} ∣ z ∣ = u 1 2 + u 2 2 ≤ 1 sup ( u 1 Re z + u 2 Im z ) = ∣ u ∣ ≤ 1 sup Re u z . (71)
Consequently,
inf Υ ∈ C ( ρ A , ρ B ) ∣ Tr ( C Υ ) ∣ = inf Υ ∈ C ( ρ A , ρ B ) sup u 1 2 + u 2 2 ≤ 1 Tr [ ( u 1 C R + u 2 C I ) Υ ] . (72) \inf_{\Upsilon\in\mathcal{C}(\rho_A,\rho_B)}\lvert\operatorname{Tr}(C\Upsilon)\rvert
=\inf_{\Upsilon\in\mathcal{C}(\rho_A,\rho_B)}\sup_{u_1^2+u_2^2\leq1}\operatorname{Tr}\left[(u_1C_R+u_2C_I)\Upsilon\right]. \tag*{(72)} Υ ∈ C ( ρ A , ρ B ) inf ∣ Tr ( C Υ )∣ = Υ ∈ C ( ρ A , ρ B ) inf u 1 2 + u 2 2 ≤ 1 sup Tr [ ( u 1 C R + u 2 C I ) Υ ] . (72)
The spatial coupling set C ( ρ A , ρ B ) \mathcal{C}(\rho_A,\rho_B) C ( ρ A , ρ B ) and the closed unit disk are nonempty, compact, and convex, and the objective is real-valued, continuous, and bilinear. The minimax theorem (see Theorem A.3 in Appendix) therefore gives
inf Υ ∈ C ( ρ A , ρ B ) ∣ Tr ( C Υ ) ∣ = sup u 1 2 + u 2 2 ≤ 1 inf Υ ∈ C ( ρ A , ρ B ) Tr [ ( u 1 C R + u 2 C I ) Υ ] . (73) \inf_{\Upsilon\in\mathcal{C}(\rho_A,\rho_B)} \lvert\operatorname{Tr}(C\Upsilon)\rvert
= \sup_{u_1^2+u_2^2\leq1}\inf_{\Upsilon\in\mathcal{C}(\rho_A,\rho_B)} \operatorname{Tr}\left[(u_1C_{\mathrm{R}}+u_2C_{\mathrm{I}})\Upsilon\right]. \tag*{(73)} Υ ∈ C ( ρ A , ρ B ) inf ∣ Tr ( C Υ )∣ = u 1 2 + u 2 2 ≤ 1 sup Υ ∈ C ( ρ A , ρ B ) inf Tr [ ( u 1 C R + u 2 C I ) Υ ] . (73)
For each fixed ( u 1 , u 2 ) (u_1,u_2) ( u 1 , u 2 ) , the operator u 1 C R + u 2 C I u_1C_{\mathrm{R}}+u_2C_{\mathrm{I}} u 1 C R + u 2 C I is Hermitian. Applying Theorem 4.2 to the inner problem proves (70) .
The primal infimum is attained by compactness and continuity. Moreover, the inner optimal value in (73) is a continuous function of ( u 1 , u 2 ) (u_1,u_2) ( u 1 , u 2 ) , so its supremum over the closed unit disk is attained. On the support space of ρ A , ρ B \rho_A,\rho_B ρ A , ρ B , Theorem 4.2 guarantees dual attainment for the corresponding Hermitian cost. Since unit disk is compact, the joint dual supremum is attained on that space. ■
Note that for the real-part objective function Re Tr [ C Υ ] = Tr [ C R Υ ] \operatorname{Re}\operatorname{Tr}[C\Upsilon]=\operatorname{Tr}[C_{\mathrm{R}}\Upsilon] Re Tr [ C Υ ] = Tr [ C R Υ ] , it reduces to the standard setting by replacing the cost operator with C R C_{\mathrm{R}} C R . Theorem 4.2 can thus be applied to give the Kantorovich duality.
The complex-cost problem has the conic formulation
inf Υ , t t subject to Υ ≥ 0 , Tr B Υ = ρ A , Tr A Υ = ρ B , Tr ( C R Υ ) 2 + Tr ( C I Υ ) 2 ≤ t . (74) \begin{aligned}
\inf_{\Upsilon,t}\quad&t\\
\text{subject to}\quad&\Upsilon\geq0,\quad\operatorname{Tr}_{B}\Upsilon=\rho_A,\quad\operatorname{Tr}_{A}\Upsilon=\rho_B,\\
&\sqrt{\operatorname{Tr}(C_{\mathrm{R}}\Upsilon)^2+\operatorname{Tr}(C_{\mathrm{I}}\Upsilon)^2}\leq t. \tag*{(74)}
\end{aligned} Υ , t inf subject to t Υ ≥ 0 , Tr B Υ = ρ A , Tr A Υ = ρ B , Tr ( C R Υ ) 2 + Tr ( C I Υ ) 2 ≤ t . (74)
The last constraint is a three-dimensional second-order cone constraint. Equivalently, writing a = Tr ( C R Υ ) a=\operatorname{Tr}(C_{\mathrm{R}}\Upsilon) a = Tr ( C R Υ ) and b = Tr ( C I Υ ) b=\operatorname{Tr}(C_{\mathrm{I}}\Upsilon) b = Tr ( C I Υ ) , it is the linear matrix inequality
\begin{pmatrix}
t+a & b\\
b & t-a
\end{pmatrix}\geq0. \tag*{(75)}\end{pmatrix}
Hence the complex-cost problem can also be expressed entirely as a semidefinite program.
For a Hermitian but not necessarily positive cost C C C , define
m C = inf Υ ∈ C ( ρ A , ρ B ) Tr ( C Υ ) , M C = sup Υ ∈ C ( ρ A , ρ B ) Tr ( C Υ ) . (76) m_C=\inf_{\Upsilon\in\mathcal{C}(\rho_A,\rho_B)}\operatorname{Tr}(C\Upsilon),\qquad
M_C=\sup_{\Upsilon\in\mathcal{C}(\rho_A,\rho_B)}\operatorname{Tr}(C\Upsilon). \tag*{(76)} m C = Υ ∈ C ( ρ A , ρ B ) inf Tr ( C Υ ) , M C = Υ ∈ C ( ρ A , ρ B ) sup Tr ( C Υ ) . (76)
Compactness and convexity of the coupling set imply that the attainable expectation values form the closed interval [ m C , M C ] [m_C,M_C] [ m C , M C ] . Therefore,
inf Υ ∈ C ( ρ A , ρ B ) ∣ Tr ( C Υ ) ∣ = { m C , m C > 0 , 0 , m C ≤ 0 ≤ M C , − M C , M C < 0. (77) \inf_{\Upsilon\in\mathcal{C}(\rho_A,\rho_B)}\lvert\operatorname{Tr}(C\Upsilon)\rvert
=
\begin{cases}
m_C, & m_C>0,\\
0, & m_C\leq0\leq M_C,\\
-M_C, & M_C<0.
\end{cases} \tag*{(77)} Υ ∈ C ( ρ A , ρ B ) inf ∣ Tr ( C Υ )∣ = ⎩ ⎨ ⎧ m C , 0 , − M C , m C > 0 , m C ≤ 0 ≤ M C , M C < 0. (77)
Thus the modulus problem for a Hermitian cost can be solved by two ordinary quantum Kantorovich semidefinite programs, with cost operators C C C and − C -C − C , respectively.
Temporal optimal transport in temporal phase space
We now apply our framework to optimal transport in temporal phase space. The essential difference from the spatial problem is that a temporal coupling is not, in general, a positive semidefinite density operator. Instead, it is a generally non-Hermitian and non-normal temporal state obtained by temporally linking an initial state with the Jamiołkowski operators of the intervening quantum evolutions. Its Kirkwood-Dirac representation is consequently a generally complex quasiprobability distribution.
Let t 0 < t 1 < ⋯ < t n t_0 < t_1 < \cdots< t_n t 0 < t 1 < ⋯ < t n be a discrete set of times, and let ρ t k ∈ State ( H t k ) \rho_{t_k} \in\operatorname{State}(\mathcal{H}_{t_k}) ρ t k ∈ State ( H t k ) be prescribed single-time density operators. We write ρ = ( ρ t 0 , ρ t 1 , … , ρ t n ) \boldsymbol{\rho} = (\rho_{t_0}, \rho_{t_1}, \ldots, \rho_{t_n}) ρ = ( ρ t 0 , ρ t 1 , … , ρ t n ) . At each time t k t_k t k , let { F α k ( k ) } \{F_{\alpha_k}^{(k)}\} { F α k ( k ) } be a minimal IC-POVM frame, with dual frame { D α k ( k ) } α k ∈ I k \{D_{\alpha_k}^{(k)}\}_{\alpha_k \in I_k} { D α k ( k ) } α k ∈ I k . For an operator Υ \Upsilon Υ on H t n ⊗ ⋯ ⊗ H t 0 \mathcal{H}_{t_n} \otimes\cdots\otimes\mathcal{H}_{t_0} H t n ⊗ ⋯ ⊗ H t 0 , its temporal Kirkwood-Dirac representation is
Q Υ ( α n , … , α 0 ) = Tr [ F α Υ ] = Tr [ ( F α n ( n ) ⊗ ⋯ ⊗ F α 0 ( 0 ) ) Υ ] . (78) Q_{\Upsilon}(\alpha_n,\ldots,\alpha_0) = \operatorname{Tr}[F_{\boldsymbol{\alpha}}\Upsilon] = \operatorname{Tr}\left[\left(F_{\alpha_n}^{(n)} \otimes\cdots\otimes F_{\alpha_0}^{(0)}\right)\Upsilon\right]. \tag*{(78)} Q Υ ( α n , … , α 0 ) = Tr [ F α Υ ] = Tr [ ( F α n ( n ) ⊗ ⋯ ⊗ F α 0 ( 0 ) ) Υ ] . (78)
The inverse relation reads
Υ = ∑ α Q Υ ( α ) D α = ∑ α n , … , α 0 Q Υ ( α n , … , α 0 ) D α n ( n ) ⊗ ⋯ ⊗ D α 0 ( 0 ) . (79) \Upsilon= \sum_{\boldsymbol{\alpha}} Q_{\Upsilon}(\boldsymbol{\alpha})D_{\boldsymbol{\alpha}} = \sum_{\alpha_n,\ldots,\alpha_0} Q_{\Upsilon}(\alpha_n,\ldots,\alpha_0)D_{\alpha_n}^{(n)} \otimes\cdots\otimes D_{\alpha_0}^{(0)}. \tag*{(79)} Υ = α ∑ Q Υ ( α ) D α = α n , … , α 0 ∑ Q Υ ( α n , … , α 0 ) D α n ( n ) ⊗ ⋯ ⊗ D α 0 ( 0 ) . (79)
Thus a temporal Kirkwood-Dirac coupling and its temporal-state representation contain exactly the same information.
For a temporal phase-space cost C ( α n , … , α 0 ) C(\alpha_n,\ldots,\alpha_0) C ( α n , … , α 0 ) , define the corresponding cost operator by
C = ∑ α C ( α ) F α = ∑ α n , … , α 0 C ( α n , … , α 0 ) F α n ( n ) ⊗ ⋯ ⊗ F α 0 ( 0 ) . (80) C = \sum_{\boldsymbol{\alpha}} C(\boldsymbol{\alpha})F_{\boldsymbol{\alpha}} = \sum_{\alpha_n,\ldots,\alpha_0} C(\alpha_n,\ldots,\alpha_0)F_{\alpha_n}^{(n)} \otimes\cdots\otimes F_{\alpha_0}^{(0)}. \tag*{(80)} C = α ∑ C ( α ) F α = α n , … , α 0 ∑ C ( α n , … , α 0 ) F α n ( n ) ⊗ ⋯ ⊗ F α 0 ( 0 ) . (80)
Then
∑ α n , … , α 0 C ( α n , … , α 0 ) Q Υ ( α n , … , α 0 ) = Tr ( C Υ ) . (81) \sum_{\alpha_n,\ldots,\alpha_0} C(\alpha_n,\ldots,\alpha_0)Q_{\Upsilon}(\alpha_n,\ldots,\alpha_0) = \operatorname{Tr}(C\Upsilon). \tag*{(81)} α n , … , α 0 ∑ C ( α n , … , α 0 ) Q Υ ( α n , … , α 0 ) = Tr ( C Υ ) . (81)
The temporal quantum optimal transport can therefore be formulated either in temporal phase space or directly in terms of temporal states.
Fixed-branch case
We first consider the fixed-branch temporal state [20 ] , where the temporal link product is given in a fixed order. As we will see, in two-time case, this is equivalent to quantum channel optimal transport (see [54 ] for the discussion of quantum channel optimal transport).
Consider the branch
B = ( t n s n , … , t 0 s 0 ) , s k ∈ { L , R } . (82) B = (t_n^{s_n},\ldots,t_0^{s_0}), \qquad s_k \in\{L,R\}. \tag*{(82)} B = ( t n s n , … , t 0 s 0 ) , s k ∈ { L , R } . (82)
The branch determines the ordering of the temporal link products and, equivalently, which copy of each doubled temporal slot remains uncontracted. For the temporal Kirkwood–Dirac distribution, this determines the position where we insert the IC-POVMs, e.g., Q ( α n , … , α 0 ) = Tr [ F α n ( n ) E n ( ⋯ E 1 ( F α 0 ( 0 ) ρ t 0 ) F α 1 ( 1 ) ) ] Q(\alpha_n,\ldots,\alpha_0) = \operatorname{Tr}[F_{\alpha_n}^{(n)}\mathcal{E}_n(\cdots\mathcal{E}_1(F_{\alpha_0}^{(0)}\rho_{t_0})F_{\alpha_1}^{(1)})] Q ( α n , … , α 0 ) = Tr [ F α n ( n ) E n ( ⋯ E 1 ( F α 0 ( 0 ) ρ t 0 ) F α 1 ( 1 ) )] for which B = ( t n L , … , t 1 R , t 0 L ) B = (t_n^L,\ldots,t_1^R,t_0^L) B = ( t n L , … , t 1 R , t 0 L ) .
For every k = 1 , … , n k = 1,\ldots,n k = 1 , … , n , let E k : B ( H t k − 1 ) → B ( H t k ) \mathcal{E}_k : \mathcal{B}(\mathcal{H}_{t_{k-1}}) \to\mathcal{B}(\mathcal{H}_{t_k}) E k : B ( H t k − 1 ) → B ( H t k ) be quantum channels, its Choi operator is defined by
J k ′ = C [ E k ] = d t k − 1 E k ⊗ id ( ∣ G H Z ⟩ ⟨ G H Z ∣ ) (83) J_k' = C[\mathcal{E}_k] = d_{t_{k-1}} \mathcal{E}_k \otimes\operatorname{id}(|\mathrm{GHZ}\rangle\langle\mathrm{GHZ}|) \tag*{(83)} J k ′ = C [ E k ] = d t k − 1 E k ⊗ id ( ∣ GHZ ⟩ ⟨ GHZ ∣ ) (83)
where ∣ G H Z ⟩ = 1 d t k − 1 ∑ i = 1 d t k − 1 ∣ i i ⟩ |\mathrm{GHZ}\rangle= \frac{1}{\sqrt{d_{t_{k-1}}}}\sum_{i=1}^{d_{t_{k-1}}}|ii\rangle ∣ GHZ ⟩ = d t k − 1 1 ∑ i = 1 d t k − 1 ∣ ii ⟩ is Greenberger-Horne-Zeilinger (GHZ) state. It’s clear that Choi operator is positive semidefinite and Tr t k C [ E k ] = I t k − 1 \operatorname{Tr}_{t_k} C[\mathcal{E}_k] = I_{t_{k-1}} Tr t k C [ E k ] = I t k − 1 . The Jamiołkowski operator is partial transpose of Choi operator
J k = J [ E k ] = ∑ i , j E k ( ∣ i ⟩ ⟨ j ∣ ) ⊗ ∣ j ⟩ ⟨ i ∣ = C [ E k ] T t k − 1 . (84) J_k = J[\mathcal{E}_k] = \sum_{i,j}\mathcal{E}_k(|i\rangle\langle j|) \otimes|j\rangle\langle i| = C[\mathcal{E}_k]^{T_{t_{k-1}}}. \tag*{(84)} J k = J [ E k ] = i , j ∑ E k ( ∣ i ⟩ ⟨ j ∣ ) ⊗ ∣ j ⟩ ⟨ i ∣ = C [ E k ] T t k − 1 . (84)
This convention is adapted to ensure Tr t k − 1 [ ( A ⊗ B ) J k ] = A E k ( B ) \operatorname{Tr}_{t_{k-1}}[(A \otimes B)J_k] = A\mathcal{E}_k(B) Tr t k − 1 [( A ⊗ B ) J k ] = A E k ( B ) .
In Ref. [20 ] , we show that the temporal state for branch B is given by temporal link product
Υ t n , … , t 0 B = J n ⋆ B J n − 1 ⋆ B ⋯ ⋆ B J 1 ⋆ B ρ t 0 . (85) \Upsilon^{\mathrm{B}}_{t_n,\ldots,t_0} = J_n \star_{\mathrm{B}} J_{n-1} \star_{\mathrm{B}} \cdots\star_{\mathrm{B}} J_1 \star_{\mathrm{B}} \rho_{t_0}. \tag*{(85)} Υ t n , … , t 0 B = J n ⋆ B J n − 1 ⋆ B ⋯ ⋆ B J 1 ⋆ B ρ t 0 . (85)
At time t k t_k t k , the reduced state Tr t k Υ t n , … , t 0 B = ρ t k = E k ∘ ⋯ ∘ E 1 ( ρ t 0 ) \operatorname{Tr}_{t_k}\Upsilon^{\mathrm{B}}_{t_n,\ldots,t_0} = \rho_{t_k} = \mathcal{E}_k \circ\cdots\circ\mathcal{E}_1(\rho_{t_0}) Tr t k Υ t n , … , t 0 B = ρ t k = E k ∘ ⋯ ∘ E 1 ( ρ t 0 ) . We denote the set of all fixed-branch temporal state with marginals ρ t 0 , … , ρ t n \rho_{t_0},\ldots,\rho_{t_n} ρ t 0 , … , ρ t n as C B ( ρ t 0 , … , ρ t n ) \mathcal{C}^{\mathrm{B}}(\rho_{t_0},\ldots,\rho_{t_n}) C B ( ρ t 0 , … , ρ t n ) . The condition that E k \mathcal{E}_k E k transports the prescribed state ρ t k − 1 \rho_{t_{k-1}} ρ t k − 1 to ρ t k \rho_{t_k} ρ t k is
E k ( ρ t k − 1 ) = ρ t k . (86) \mathcal{E}_k(\rho_{t_{k-1}}) = \rho_{t_k}. \tag*{(86)} E k ( ρ t k − 1 ) = ρ t k . (86)
In terms of Jamiołkowski operator J k J_k J k , this becomes
Tr t k − 1 [ ( I t k ⊗ ρ t k − 1 ) J k ] = ρ t k , (87) \operatorname{Tr}_{t_{k-1}}\left[(I_{t_k} \otimes\rho_{t_{k-1}})J_k\right] = \rho_{t_k}, \tag*{(87)} Tr t k − 1 [ ( I t k ⊗ ρ t k − 1 ) J k ] = ρ t k , (87)
Define the set of admissible one-step transport Choi operator by
Choi k ( ρ t k − 1 → ρ t k ) = { J k : J k ≥ 0 , Tr t k J k = I t k − 1 , Tr t k − 1 [ J k ( I t k ⊗ ρ t k − 1 ) ] = ρ t k } . (88) \operatorname{Choi}_k(\rho_{t_{k-1}} \to\rho_{t_k}) = \left\{J_k : J_k \ge0,\quad\operatorname{Tr}_{t_k}J_k = I_{t_{k-1}},\quad\operatorname{Tr}_{t_{k-1}}\left[J_k(I_{t_k} \otimes\rho_{t_{k-1}})\right] = \rho_{t_k}\right\}. \tag*{(88)} Choi k ( ρ t k − 1 → ρ t k ) = { J k : J k ≥ 0 , Tr t k J k = I t k − 1 , Tr t k − 1 [ J k ( I t k ⊗ ρ t k − 1 ) ] = ρ t k } . (88)
This set is a compact convex set. It is always nonempty, since the replacement channel
E k ( X ) = Tr ( X ) ρ t k (89) \mathcal{E}_k(X) = \operatorname{Tr}(X)\rho_{t_k} \tag*{(89)} E k ( X ) = Tr ( X ) ρ t k (89)
belongs to it. The set of admissible Jamiołkowski operator is
Jami k ( ρ t k − 1 → ρ t k ) = { J : J T t k − 1 ∈ Choi k ( ρ t k − 1 → ρ t k ) } . (90) \operatorname{Jami}_k(\rho_{t_{k-1}} \to\rho_{t_k}) = \left\{J : J^{T_{t_{k-1}}} \in\operatorname{Choi}_k(\rho_{t_{k-1}} \to\rho_{t_k})\right\}. \tag*{(90)} Jami k ( ρ t k − 1 → ρ t k ) = { J : J T t k − 1 ∈ Choi k ( ρ t k − 1 → ρ t k ) } . (90)
The fixed-branch temporal-state coupling set is thus of the form
C B ( ρ t 0 , … , ρ t n ) = { Υ B ( ρ t 0 ; J 1 , … , J n ) : J k ∈ Jami k ( ρ t k − 1 → ρ t k ) , k = 1 , … , n } . \mathcal{C}^{\mathrm{B}}(\rho_{t_0},\ldots,\rho_{t_n}) = \left\{\Upsilon^{\mathrm{B}}(\rho_{t_0};J_1,\ldots,J_n) : J_k \in\operatorname{Jami}_k(\rho_{t_{k-1}} \to\rho_{t_k}),\quad k=1,\ldots,n\right\}. C B ( ρ t 0 , … , ρ t n ) = { Υ B ( ρ t 0 ; J 1 , … , J n ) : J k ∈ Jami k ( ρ t k − 1 → ρ t k ) , k = 1 , … , n } .
Each element of this set satisfies
Tr t n , … , t k ^ , … , t 0 Υ B = ρ t k , k = 0 , … , n . (91) \operatorname{Tr}_{t_n,\ldots,\widehat{t_k},\ldots,t_0}\Upsilon^{\mathrm{B}} = \rho_{t_k},\quad k=0,\ldots,n. \tag*{(91)} Tr t n , … , t k , … , t 0 Υ B = ρ t k , k = 0 , … , n . (91)
Fix a minimal IC-POVM frame, The corresponding set of fixed-branch temporal Kirkwood-Dirac coupling is
K D B ( Q t 0 , … , Q t n ) = { Q Υ : Υ ∈ C B ( ρ t 0 , … , ρ t n ) } . (92) \mathrm{KD}^{\mathrm{B}}(Q_{t_0},\ldots,Q_{t_n}) = \left\{Q\Upsilon: \Upsilon\in\mathcal{C}^{\mathrm{B}}(\rho_{t_0},\ldots,\rho_{t_n})\right\}. \tag*{(92)} KD B ( Q t 0 , … , Q t n ) = { Q Υ : Υ ∈ C B ( ρ t 0 , … , ρ t n ) } . (92)
the map
Υ ⟼ Q Υ (93) \Upsilon\longmapsto Q_{\Upsilon} \tag*{(93)} Υ ⟼ Q Υ (93)
is a bijection from C B ( ρ ) \mathcal{C}^{\mathrm{B}}(\rho) C B ( ρ ) onto K D B ( ρ ) \mathbb{K}\mathbb{D}^{\mathrm{B}}(\rho) K D B ( ρ ) . Its inverse is given by (79) .
The fixed-branch temporal Monge-Kantorovich optimal transport problem is defined as
Q M K B , C ( ρ ) : = inf Q ∈ K D B ( ρ t 0 , … , ρ t n ) ∣ ∑ α n , … , α 0 C ( α n , … , α 0 ) Q ( α n , … , α 0 ) ∣ = inf Υ ∈ C B ( ρ ) ∣ Tr ( C Υ ) ∣ = : Q M K B , C ( ρ ) . \begin{aligned}
\mathrm{QMK}_{\mathrm{B},\mathcal{C}}(\rho)
&:=\inf_{Q\in\mathbb{K}\mathbb{D}^{\mathrm{B}}(\rho_{t_0},\ldots,\rho_{t_n})}
\left|\sum_{\alpha_n,\ldots,\alpha_0}\mathcal{C}(\alpha_n,\ldots,\alpha_0)Q(\alpha_n,\ldots,\alpha_0)\right|\\
&=\inf_{\Upsilon\in\mathcal{C}^{\mathrm{B}}(\rho)}|\operatorname{Tr}(\mathcal{C}\Upsilon)|=:\mathrm{QMK}_{\mathrm{B},\mathcal{C}}(\rho).
\end{aligned} QMK B , C ( ρ ) := Q ∈ K D B ( ρ t 0 , … , ρ t n ) inf α n , … , α 0 ∑ C ( α n , … , α 0 ) Q ( α n , … , α 0 ) = Υ ∈ C B ( ρ ) inf ∣ Tr ( C Υ ) ∣ =: QMK B , C ( ρ ) .
Equivalently,
Q M K B , C ( ρ ) = inf J k ∈ Jami k ( ρ t k − 1 → ρ t k ) k = 1 , … , n ∣ Tr [ C Υ B ( ρ t 0 ; J 1 , … , J n ) ] ∣ . (94) \mathrm{QMK}_{\mathrm{B},\mathcal{C}}(\rho)
=\inf_{\substack{J_k\in\operatorname{Jami}_k(\rho_{t_{k-1}}\to\rho_{t_k})\\ k=1,\ldots,n}}
\left|\operatorname{Tr}\left[\mathcal{C}\,\Upsilon^{\mathrm{B}}(\rho_{t_0};J_1,\ldots,J_n)\right]\right|. \tag*{(94)} QMK B , C ( ρ ) = J k ∈ Jami k ( ρ t k − 1 → ρ t k ) k = 1 , … , n inf Tr [ C Υ B ( ρ t 0 ; J 1 , … , J n ) ] . (94)
Note that ρ t 0 \rho_{t_0} ρ t 0 is one of the local marginals, thus it is fixed fixed-branch temporal optimal transport, we can only vary the quantum channels. In this sense, the problem is equivalent to the quantum channel optimal transport. However, the formalism in [54 ] can only work in two-time case, our formalism gives a natural multi-marginal generalization. The quantum channel optimal transport in [54 ] in fact considers temporal couplings of the form Ξ = E ⊗ id ( ∣ ρ t 0 ⟩ ⟩ ⟨ ⟨ ρ t 0 ∣ ) \Xi=\mathcal{E}\otimes\operatorname{id}(|\sqrt{\rho_{t_0}}\rangle\!\rangle\langle\!\langle\sqrt{\rho_{t_0}}|) Ξ = E ⊗ id ( ∣ ρ t 0 ⟩ ⟩ ⟨ ⟨ ρ t 0 ∣ ) , where E \mathcal{E} E is required to give a fixed output E ( ρ t 0 ) = ρ t 1 \mathcal{E}(\rho_{t_0})=\rho_{t_1} E ( ρ t 0 ) = ρ t 1 . Let F E , ρ t 0 = E ( ρ t 0 ∙ ρ t 0 ) \mathcal{F}_{\mathcal{E},\rho_{t_0}}=\mathcal{E}(\sqrt{\rho_{t_0}}\bullet\sqrt{\rho_{t_0}}) F E , ρ t 0 = E ( ρ t 0 ∙ ρ t 0 ) , using Eq. (A.2) in Appendix, we see that Ξ \Xi Ξ is indeed the Choi operator Ξ = C [ F E , ρ t 0 ] \Xi=C[\mathcal{F}_{\mathcal{E},\rho_{t_0}}] Ξ = C [ F E , ρ t 0 ] , from which it is clear that Tr t 0 Ξ = ρ t 1 \operatorname{Tr}_{t_0}\Xi=\rho_{t_1} Tr t 0 Ξ = ρ t 1 . It can also be expressed as Ξ = E ⊗ id ( ( 1 ⊗ ρ T ) ∣ Ω ⟩ ⟩ ⟨ ⟨ Ω ∣ ( 1 ⊗ ρ T ) ) \Xi=\mathcal{E}\otimes\operatorname{id}((1\otimes\sqrt{\rho^{T}})|\Omega\rangle\!\rangle\langle\!\langle\Omega|(1\otimes\sqrt{\rho^{T}})) Ξ = E ⊗ id (( 1 ⊗ ρ T ) ∣Ω ⟩ ⟩ ⟨ ⟨ Ω∣ ( 1 ⊗ ρ T )) , from which we have Tr t 1 Ξ = ρ t 0 T \operatorname{Tr}_{t_1}\Xi=\rho_{t_0}^{T} Tr t 1 Ξ = ρ t 0 T . For t 0 t_0 t 0 , there is a transpose, which makes the marginals for the two times not on an equal footing. In our setting via setting Υ = J [ E ] ( 1 ⊗ ρ t 0 ) \Upsilon=\mathcal{J}[\mathcal{E}](1\otimes\rho_{t_0}) Υ = J [ E ] ( 1 ⊗ ρ t 0 ) , which gives Tr t 0 Υ = ρ t 1 \operatorname{Tr}_{t_0}\Upsilon=\rho_{t_1} Tr t 0 Υ = ρ t 1 and Tr t 1 Υ = ρ t 0 \operatorname{Tr}_{t_1}\Upsilon=\rho_{t_0} Tr t 1 Υ = ρ t 0 , which is more symmetric. In two-time case, they are essentially equivalent. However, the expression for Ξ \Xi Ξ in [54 ] cannot be generalized straightforwardly to the multi-marginal setting, our setting give a more natural definition of temporal quantum optimal transport. For instance, for three-time with B = ( t 2 R , t 1 R , t 0 R ) \mathrm{B}=(t_2^{R},t_1^{R},t_0^{R}) B = ( t 2 R , t 1 R , t 0 R ) , we have Υ = J [ E 2 ] J [ E 1 ] ( 1 ⊗ ρ t 0 ) \Upsilon=\mathcal{J}[\mathcal{E}_2]\mathcal{J}[\mathcal{E}_1](1\otimes\rho_{t_0}) Υ = J [ E 2 ] J [ E 1 ] ( 1 ⊗ ρ t 0 ) , which gives correct marginals ρ t 0 \rho_{t_0} ρ t 0 , ρ t 1 \rho_{t_1} ρ t 1 and ρ t 2 \rho_{t_2} ρ t 2 . A more detailed discussion will be given in a forthcoming work [67 ] .
Proposition 5.1 (Existence and multiaffine structure). The infimum of temporal optimal transport in Eq. (94) is attained. The map
( J 1 , … , J n ) ⟼ Υ B ( ρ t 0 ; J 1 , … , J n ) (95) (J_1,\ldots,J_n)\longmapsto\Upsilon^{\mathrm{B}}(\rho_{t_0};J_1,\ldots,J_n) \tag*{(95)} ( J 1 , … , J n ) ⟼ Υ B ( ρ t 0 ; J 1 , … , J n ) (95) is linear in each J k J_k J k when the remaining Jamiolkowski operators are held fixed. It is not, in general, jointly affine3 when n ≥ 2 n\geq2 n ≥ 2 . Consequently, the fixed-branch temporal quantum optimal transport problem is a multi-convex optimization problem, but not generally a convex optimization problem.
Proof. Each set Jami ( ρ t k − 1 → ρ t k ) \operatorname{Jami}(\rho_{t_{k-1}}\to\rho_{t_k}) Jami ( ρ t k − 1 → ρ t k ) is compact. Their Cartesian product is therefore compact. Fixed-branch temporal link products ⋆ B \star_{\mathrm{B}} ⋆ B is continuous, thus the objective function in (94) is continuous. The infimum is therefore attained.
Every Jamiolkowski operator occurs exactly once in the temporal link-product expression. Holding all other factors fixed makes the temporal state linear in that Jamiolkowski operator. When two or more Jamiolkowski operators are varied simultaneously, products such as J k + 1 J k J_{k+1}J_k J k + 1 J k occur. The resulting map is therefore multiaffine rather than jointly affine.
This distinction of two-time (n = 1 n=1 n = 1 ) and multi-time (n ≥ 2 n\geq2 n ≥ 2 ) is important. For two times, there is only one variable Choi/Jamiłkowski operator, and the fixed-branch problem is convex. For three or more times, independent optimization over all intervening channels generally produces a nonconvex factorized-process problem.
Quantum Kantorovich duality of the two-time temporal quantum optimal transport
Consider two-time case t 0 < t 1 t_0<t_1 t 0 < t 1 . There are only two independent branches4 B = ( t 1 L , t 0 L ) \mathrm{B}=(t_1^{L},t_0^{L}) B = ( t 1 L , t 0 L ) and B c = ( t 1 R , t 0 R ) \mathrm{B}^{c}=(t_1^{R},t_0^{R}) B c = ( t 1 R , t 0 R ) which gives two temporal states
Υ B = ( 1 ⊗ ρ t 0 ) J 1 , Υ B c = J 1 ( 1 ⊗ ρ t 0 ) . (96) \Upsilon^{\mathrm{B}}=(1\otimes\rho_{t_0})J_1,\qquad\Upsilon^{\mathrm{B}^{c}}=J_1(1\otimes\rho_{t_0}). \tag*{(96)} Υ B = ( 1 ⊗ ρ t 0 ) J 1 , Υ B c = J 1 ( 1 ⊗ ρ t 0 ) . (96)
Set A B , C = C ( 1 ⊗ ρ t 0 ) A_{\mathrm{B},\mathrm{C}}=C(1\otimes\rho_{t_0}) A B , C = C ( 1 ⊗ ρ t 0 ) , we have
Tr [ C Υ B ( ρ t 0 ; J 1 ) ] = Tr ( A B , C J 1 ) (97) \operatorname{Tr}\left[C\Upsilon^{\mathrm{B}}(\rho_{t_0};J_1)\right]=\operatorname{Tr}(A_{\mathrm{B},\mathrm{C}}J_1) \tag*{(97)} Tr [ C Υ B ( ρ t 0 ; J 1 ) ] = Tr ( A B , C J 1 ) (97)
for every J 1 J_1 J 1 . Thus the temporal optimal transport becomes a optimization over all J 1 ∈ Jami ( ρ t 0 → ρ t 1 ) J_1\in\operatorname{Jami}(\rho_{t_0}\to\rho_{t_1}) J 1 ∈ Jami ( ρ t 0 → ρ t 1 ) with cost operator A B , C A_{\mathrm{B},\mathrm{C}} A B , C .
To establish the quantum Kantorovich duality, it is more convenient to work with Choi operator. We can use Lemma A.4 in Appendix to transform the temporal optimal transport problem into
inf J ′ ∈ Choi ( ρ t 0 → ρ t 1 ) ∣ Tr ( A B , C T t 0 J ′ ) ∣ (98) \inf_{J'\in\operatorname{Choi}(\rho_{t_0}\to\rho_{t_1})}\left|\operatorname{Tr}\left(A_{\mathrm{B},\mathrm{C}}^{T_{t_0}}J'\right)\right| \tag*{(98)} J ′ ∈ Choi ( ρ t 0 → ρ t 1 ) inf Tr ( A B , C T t 0 J ′ ) (98)
In this case, the optimization is taken over positive semidefinite operators.
Theorem 5.2 (Quantum Kantorovich duality of two-time fixed-branch temporal quantum optimal transport). For u ∈ C u\in\mathbb{C} u ∈ C , define the Hermitian operator
H B , C ( u ) = 1 2 ( u ‾ A B , C T t 0 + u ( A B , C T t 0 ) † ) ∈ Herm ( H t 1 ⊗ H t 0 ) . (99) H_{\mathrm{B},\mathrm{C}}(u)=\frac{1}{2}\left(\overline{u}A_{\mathrm{B},\mathrm{C}}^{T_{t_0}}+u\left(A_{\mathrm{B},\mathrm{C}}^{T_{t_0}}\right)^\dagger\right)\in\operatorname{Herm}(\mathcal{H}_{t_1}\otimes\mathcal{H}_{t_0}). \tag*{(99)} H B , C ( u ) = 2 1 ( u A B , C T t 0 + u ( A B , C T t 0 ) † ) ∈ Herm ( H t 1 ⊗ H t 0 ) . (99) The quantum Kantorovich duality of temporal quantum optimal transport is given by
QMK ^ B , C ( ρ t 0 , ρ t 1 ) = sup u ∈ C , ∣ u ∣ ≤ 1 X = X † , Y = Y † Tr ( X ) + Tr ( Y ρ t 1 ) \widehat{\operatorname{QMK}}_{\mathrm{B},\mathrm{C}}(\rho_{t_0},\rho_{t_1})=
\sup_{\substack{u\in\mathbb{C},\ |u|\leq1\\X=X^\dagger,\ Y=Y^\dagger}}
\operatorname{Tr}(X)+\operatorname{Tr}(Y\rho_{t_1}) QMK B , C ( ρ t 0 , ρ t 1 ) = u ∈ C , ∣ u ∣ ≤ 1 X = X † , Y = Y † sup Tr ( X ) + Tr ( Y ρ t 1 ) subject to I t 1 ⊗ X + Y ⊗ ρ t 0 T ≤ H B , C ( u ) . (100) \text{subject to}\quad I_{t_1}\otimes X+Y\otimes\rho_{t_0}^{T}\leq H_{\mathrm{B},\mathrm{C}}(u). \tag*{(100)} subject to I t 1 ⊗ X + Y ⊗ ρ t 0 T ≤ H B , C ( u ) . (100)
Here X X X acts on H t 0 \mathcal{H}_{t_0} H t 0 and Y Y Y acts on H t 1 \mathcal{H}_{t_1} H t 1 . We have
QMK B , C ( ρ t 0 , ρ t 1 ) = QMK ‾ B , C ( ρ t 0 , ρ t 1 ) (101) \operatorname{QMK}_{B,C}(\rho_{t_0},\rho_{t_1})=\overline{\operatorname{QMK}}_{B,C}(\rho_{t_0},\rho_{t_1}) \tag*{(101)} QMK B , C ( ρ t 0 , ρ t 1 ) = QMK B , C ( ρ t 0 , ρ t 1 ) (101)
The primal infimum is always attained. If ρ t 1 \rho_{t_1} ρ t 1 is full rank, the dual supremum is always attained.
Proof. For every z ∈ C z \in\mathbb{C} z ∈ C ,
∣ z ∣ = sup ∣ u ∣ ≤ 1 Re ( u ‾ z ) . (102) |z|=\sup_{|u|\leq1}\operatorname{Re}(\overline{u}z). \tag*{(102)} ∣ z ∣ = ∣ u ∣ ≤ 1 sup Re ( u z ) . (102)
This implies
inf J ′ ∈ Choi ( ρ t 0 → ρ t 1 ) ∣ Tr ( A B , C T t 0 J ′ ) ∣ = inf J ′ ∈ Choi ( ρ t 0 → ρ t 1 ) sup ∣ u ∣ ≤ 1 Tr [ H B , C ( u ) J ′ ] . (103) \begin{aligned}
\inf_{J'\in\operatorname{Choi}(\rho_{t_0}\to\rho_{t_1})}\left|\operatorname{Tr}\left(A_{B,C}^{T_{t_0}}J'\right)\right|
&=\inf_{J'\in\operatorname{Choi}(\rho_{t_0}\to\rho_{t_1})}\sup_{|u|\leq1}\operatorname{Tr}\left[H_{B,C}(u)J'\right]. \tag*{(103)}
\end{aligned} J ′ ∈ Choi ( ρ t 0 → ρ t 1 ) inf Tr ( A B , C T t 0 J ′ ) = J ′ ∈ Choi ( ρ t 0 → ρ t 1 ) inf ∣ u ∣ ≤ 1 sup Tr [ H B , C ( u ) J ′ ] . (103)
The Choi ( ρ t 0 → ρ t 1 ) \operatorname{Choi}(\rho_{t_0}\to\rho_{t_1}) Choi ( ρ t 0 → ρ t 1 ) and the unit disk are compact and convex, and the objective is bilinear. The minimax theorem therefore gives
sup ∣ u ∣ ≤ 1 inf J ′ ∈ Choi ( ρ t 0 → ρ t 1 ) Tr [ H B , C ( u ) J ′ ] . (104) \sup_{|u|\leq1}\inf_{J'\in\operatorname{Choi}(\rho_{t_0}\to\rho_{t_1})}\operatorname{Tr}\left[H_{B,C}(u)J'\right]. \tag*{(104)} ∣ u ∣ ≤ 1 sup J ′ ∈ Choi ( ρ t 0 → ρ t 1 ) inf Tr [ H B , C ( u ) J ′ ] . (104)
For fixed u u u , the inner problem of Eq. (104) is a SDP
inf J ′ ∈ Choi ( ρ t 0 → ρ t 1 ) Tr [ H B , C ( u ) J ′ ] subject to J ′ ≥ 0 , Tr t 1 J ′ = I t 0 , Tr t 0 [ J ′ ( I t 1 ⊗ ρ t 0 T ) ] = ρ t 1 . (105) \begin{aligned}
\inf_{J'\in\operatorname{Choi}(\rho_{t_0}\to\rho_{t_1})}\quad&\operatorname{Tr}\left[H_{B,C}(u)J'\right]\\
\text{subject to}\quad&J'\geq0,\\
&\operatorname{Tr}_{t_1}J'=I_{t_0},\\
&\operatorname{Tr}_{t_0}\left[J'(I_{t_1}\otimes\rho_{t_0}^{T})\right]=\rho_{t_1}. \tag*{(105)}
\end{aligned} J ′ ∈ Choi ( ρ t 0 → ρ t 1 ) inf subject to Tr [ H B , C ( u ) J ′ ] J ′ ≥ 0 , Tr t 1 J ′ = I t 0 , Tr t 0 [ J ′ ( I t 1 ⊗ ρ t 0 T ) ] = ρ t 1 . (105)
Introducing Hermitian multipliers X X X and Y Y Y for the last two equalities gives the dual
sup X = X † , Y = Y † Tr ( X ) + Tr ( Y ρ t 1 ) subject to I t 1 ⊗ X + Y ⊗ ρ t 0 T ≤ H B , C ( u ) . (106) \begin{aligned}
\sup_{X=X^\dagger,\;Y=Y^\dagger}\quad&\operatorname{Tr}(X)+\operatorname{Tr}(Y\rho_{t_1})\\
\text{subject to}\quad&I_{t_1}\otimes X+Y\otimes\rho_{t_0}^{T}\leq H_{B,C}(u). \tag*{(106)}
\end{aligned} X = X † , Y = Y † sup subject to Tr ( X ) + Tr ( Y ρ t 1 ) I t 1 ⊗ X + Y ⊗ ρ t 0 T ≤ H B , C ( u ) . (106)
Since primal feasible set is non-empty since replacement channel gives a feasible Choi operator, and we define
E : J ′ ↦ ( Tr t 1 J ′ , Tr ( J ′ ρ t 0 T ) ) (107) \mathcal{E}:J'\mapsto\left(\operatorname{Tr}_{t_1}J',\operatorname{Tr}(J'\rho_{t_0}^{T})\right) \tag*{(107)} E : J ′ ↦ ( Tr t 1 J ′ , Tr ( J ′ ρ t 0 T ) ) (107)
the adjoint is
E † : ( X , Y ) ↦ I t 1 ⊗ X + Y ⊗ ρ t 0 T . (108) \mathcal{E}^{\dagger}:(X,Y)\mapsto I_{t_1}\otimes X+Y\otimes\rho_{t_0}^{T}. \tag*{(108)} E † : ( X , Y ) ↦ I t 1 ⊗ X + Y ⊗ ρ t 0 T . (108)
Set X = − t I t 0 X=-tI_{t_0} X = − t I t 0 and Y = 0 Y=0 Y = 0 , E † ( X , Y ) = − t I t 1 t 0 \mathcal{E}^{\dagger}(X,Y)=-tI_{t_1t_0} E † ( X , Y ) = − t I t 1 t 0 , when t > − λ min ( H B , C ( u ) ) t>-\lambda_{\min}(H_{B,C}(u)) t > − λ m i n ( H B , C ( u )) , we have E † ( X , Y ) < H B , C ( u ) \mathcal{E}^{\dagger}(X,Y)<H_{B,C}(u) E † ( X , Y ) < H B , C ( u ) , they satisfy Slater’s condition, thus the primal infimum is always attained and Eq. (101) holds.
If ρ t 1 \rho_{t_1} ρ t 1 is full rank, the replacement channel has the strictly positive Choi matrix
J ′ = ρ t 1 ⊗ I t 0 , (109) J'=\rho_{t_1}\otimes I_{t_0}, \tag*{(109)} J ′ = ρ t 1 ⊗ I t 0 , (109)
so Slater’s condition holds, thus the dual supremum is always attained. ■
Convex temporal-state optimal transport
As proven in Ref. [20 ] , we can take the probabilistic mixture of temporal state with difference branches to obtain a new temporal state, and this type of construction unifies many existing frameworks of temporal states, including pseudo-density operator [9 ] (or state over time [10 , 25 ] ).
For the initial state ρ t 0 \rho_{t_0} ρ t 0 and n n n quantum channels with Choi matrices J 1 ′ , … , J n ′ J'_1,\ldots,J'_n J 1 ′ , … , J n ′ , let Branch n \operatorname{Branch}_n Branch n denote the set of allowed 2 n 2^n 2 n branch assignments. The convex set of temporal states is defined as
State T ( ρ t 0 ; J 1 ′ , … , J n ′ ) = conv { Υ B ( ρ t 0 ; J 1 ′ , … , J n ′ ) : B ∈ Branch n } . (110) \operatorname{State}_{T}(\rho_{t_0};J'_1,\ldots,J'_n)=\operatorname{conv}\left\{\Upsilon^{\mathrm{B}}(\rho_{t_0};J'_1,\ldots,J'_n):\mathrm{B}\in\operatorname{Branch}_n\right\}. \tag*{(110)} State T ( ρ t 0 ; J 1 ′ , … , J n ′ ) = conv { Υ B ( ρ t 0 ; J 1 ′ , … , J n ′ ) : B ∈ Branch n } . (110)
An element of this set has the form
Υ = ∑ B ∈ Branch n p B Υ B ( ρ t 0 ; J 1 ′ , … , J n ′ ) , (111) \Upsilon=\sum_{\mathrm{B}\in\operatorname{Branch}_n}p_{\mathrm{B}}\Upsilon^{\mathrm{B}}(\rho_{t_0};J'_1,\ldots,J'_n), \tag*{(111)} Υ = B ∈ Branch n ∑ p B Υ B ( ρ t 0 ; J 1 ′ , … , J n ′ ) , (111)
where
p B ≥ 0 , ∑ B p B = 1. (112) p_{\mathrm{B}}\geq0,\qquad\sum_{\mathrm{B}}p_{\mathrm{B}}=1. \tag*{(112)} p B ≥ 0 , B ∑ p B = 1. (112)
In particular, the left, right, mixed-branch, Margenau-Hill, and Lüders-von Neumann temporal states arise from distinguished choices of { p B } \{p_{\mathrm{B}}\} { p B } [20 ] . Since there are only 2 n 2^n 2 n branches for t 0 , … , t n t_0,\ldots,t_n t 0 , … , t n , State T ( ρ t 0 ; J 1 ′ , … , J n ′ ) \operatorname{State}_{T}(\rho_{t_0};J'_1,\ldots,J'_n) State T ( ρ t 0 ; J 1 ′ , … , J n ′ ) is indeed a convex polytope.
However, for temporal quantum optimal transport, we need to vary the dynamics, the above convex polytope of temporal state is thus not sufficient.
Fixed-branch convex temporal coupling set
As we have see, fixed-branch temporal quantum optimal transport can be transformed into a multi-convex optimization problem over Choi operators J k ′ ∈ Choi k ( ρ t k − 1 , ρ t k ) J'_k\in\operatorname{Choi}_k(\rho_{t_{k-1}},\rho_{t_k}) J k ′ ∈ Choi k ( ρ t k − 1 , ρ t k ) . For two-time setting, this is equivalent to the quantum channel optimal transport. From temporal state perspective, we in fact consider the temporal coupling set as
C B ( ρ t 0 , … , ρ t n ) = { Υ B ( ρ t 0 ; J 1 ′ , … , J n ′ ) : J k ′ ∈ Choi k ( ρ t k − 1 , ρ t k ) , k = 1 , … , n } . (113) \mathcal{C}^{\mathrm{B}}(\rho_{t_0},\ldots,\rho_{t_n})=\left\{\Upsilon^{\mathrm{B}}(\rho_{t_0};J'_1,\ldots,J'_n):J'_k\in\operatorname{Choi}_k(\rho_{t_{k-1}},\rho_{t_k}),\ k=1,\ldots,n\right\}. \tag*{(113)} C B ( ρ t 0 , … , ρ t n ) = { Υ B ( ρ t 0 ; J 1 ′ , … , J n ′ ) : J k ′ ∈ Choi k ( ρ t k − 1 , ρ t k ) , k = 1 , … , n } . (113)
We can enlarge the coupling set by considering the convex hull
C c o n v B ( ρ t 0 , … , ρ t n ) = conv C B ( ρ t 0 , … , ρ t n ) . (114) \mathcal{C}^{\mathrm{B}}_{\mathrm{conv}}(\rho_{t_0},\ldots,\rho_{t_n})=\operatorname{conv}\mathcal{C}^{\mathrm{B}}(\rho_{t_0},\ldots,\rho_{t_n}). \tag*{(114)} C conv B ( ρ t 0 , … , ρ t n ) = conv C B ( ρ t 0 , … , ρ t n ) . (114)
Each state in C c o n v B ( ρ t 0 , … , ρ t n ) \mathcal{C}^{\mathrm{B}}_{\mathrm{conv}}(\rho_{t_0},\ldots,\rho_{t_n}) C conv B ( ρ t 0 , … , ρ t n ) gives correct marginals. Take optimization over this set naturally give a convex temporal quantum optimal transport problem. Also note that for two-time case, C c o n v B ( ρ t 0 , … , ρ t 1 ) = C B ( ρ t 0 , … , ρ t 1 ) \mathcal{C}^{\mathrm{B}}_{\mathrm{conv}}(\rho_{t_0},\ldots,\rho_{t_1})=\mathcal{C}^{\mathrm{B}}(\rho_{t_0},\ldots,\rho_{t_1}) C conv B ( ρ t 0 , … , ρ t 1 ) = C B ( ρ t 0 , … , ρ t 1 ) , since the set Choi 1 ( ρ t 0 , ρ t 1 ) \operatorname{Choi}_1(\rho_{t_0},\rho_{t_1}) Choi 1 ( ρ t 0 , ρ t 1 ) is convex and the initial state is fixed.
General convex temporal coupling set
The more general convex temporal-state coupling set is the following:
C c o n v ( ρ t 0 , ⋯ , ρ t n ) = conv ( ⋃ B ∈ B r a n c h n C B ( ρ t 0 , ⋯ , ρ t n ) ) . (115) \mathcal{C}_{\mathrm{conv}}(\rho_{t_0},\cdots,\rho_{t_n})=\operatorname{conv}\left(\bigcup_{B\in\mathrm{Branch}_n}\mathcal{C}^{B}(\rho_{t_0},\cdots,\rho_{t_n})\right). \tag*{(115)} C conv ( ρ t 0 , ⋯ , ρ t n ) = conv ( B ∈ Branch n ⋃ C B ( ρ t 0 , ⋯ , ρ t n ) ) . (115)
Thus, temporal coupling in C c o n v ( ρ t 0 , ⋯ , ρ t n ) \mathcal{C}_{\mathrm{conv}}(\rho_{t_0},\cdots,\rho_{t_n}) C conv ( ρ t 0 , ⋯ , ρ t n ) can be a probabilistic mixture of temporal states with different branches and with different Choi operators. Elements in C c o n v ( ρ t 0 , ⋯ , ρ t n ) \mathcal{C}_{\mathrm{conv}}(\rho_{t_0},\cdots,\rho_{t_n}) C conv ( ρ t 0 , ⋯ , ρ t n ) gives correct marginals. It’s clear that C c o n v B ( ρ t 0 , ⋯ , ρ t n ) ⊂ C c o n v ( ρ t 0 , ⋯ , ρ t n ) \mathcal{C}^{B}_{\mathrm{conv}}(\rho_{t_0},\cdots,\rho_{t_n})\subset\mathcal{C}_{\mathrm{conv}}(\rho_{t_0},\cdots,\rho_{t_n}) C conv B ( ρ t 0 , ⋯ , ρ t n ) ⊂ C conv ( ρ t 0 , ⋯ , ρ t n ) , which implies that
Q M K C B , C ≥ Q M K C c o n v B , C ≥ Q M K C c o n v , C . (116) \mathrm{QMK}_{\mathcal{C}^{B},\mathrm{C}}\geq\mathrm{QMK}_{\mathcal{C}^{B}_{\mathrm{conv}},\mathrm{C}}\geq\mathrm{QMK}_{\mathcal{C}_{\mathrm{conv}},\mathrm{C}}. \tag*{(116)} QMK C B , C ≥ QMK C conv B , C ≥ QMK C conv , C . (116)
The convex set of temporal Kirkwood-Dirac coupling is
K D c o n v ( Q ρ t 0 , ⋯ , Q ρ t n ) = { Q Γ ( α ) = Tr [ F α Γ ] : Γ ∈ C c o n v ( ρ t 0 , ⋯ , ρ t n ) } . (117) \mathrm{KD}_{\mathrm{conv}}(Q_{\rho_{t_0}},\cdots,Q_{\rho_{t_n}})=\{Q_{\Gamma}(\alpha)=\operatorname{Tr}[F_{\alpha}\Gamma]:\Gamma\in\mathcal{C}_{\mathrm{conv}}(\rho_{t_0},\cdots,\rho_{t_n})\}. \tag*{(117)} KD conv ( Q ρ t 0 , ⋯ , Q ρ t n ) = { Q Γ ( α ) = Tr [ F α Γ ] : Γ ∈ C conv ( ρ t 0 , ⋯ , ρ t n )} . (117)
In temporal phase space, the problem can be formulated as
Q M K K D c o n v ( Q ρ t 0 , ⋯ , Q ρ t n ) = inf Q ∈ K D c o n v ( Q ρ t 0 , ⋯ , Q ρ t n ) ∣ ∑ α n , … , α 0 C ( α n , … , α 0 ) Q ( α n , … , α 0 ) ∣ = inf Γ ∈ C c o n v ( ρ t 0 , ⋯ , ρ t n ) ∣ Tr ( C Γ ) ∣ . \begin{aligned}
\mathrm{QMK}_{\mathrm{KD}_{\mathrm{conv}}}(Q_{\rho_{t_0}},\cdots,Q_{\rho_{t_n}})
&=\inf_{Q\in\mathrm{KD}_{\mathrm{conv}}(Q_{\rho_{t_0}},\cdots,Q_{\rho_{t_n}})}
\left|\sum_{\alpha_n,\ldots,\alpha_0}C(\alpha_n,\ldots,\alpha_0)Q(\alpha_n,\ldots,\alpha_0)\right|\\
&=\inf_{\Gamma\in\mathcal{C}_{\mathrm{conv}}(\rho_{t_0},\cdots,\rho_{t_n})}|\operatorname{Tr}(C\Gamma)|.
\end{aligned} QMK KD conv ( Q ρ t 0 , ⋯ , Q ρ t n ) = Q ∈ KD conv ( Q ρ t 0 , ⋯ , Q ρ t n ) inf α n , … , α 0 ∑ C ( α n , … , α 0 ) Q ( α n , … , α 0 ) = Γ ∈ C conv ( ρ t 0 , ⋯ , ρ t n ) inf ∣ Tr ( C Γ ) ∣.
Cost set in the complex plane and geometric duality
The convex temporal quantum optimal transport can be transformed into a convex problem in complex plane. Fix cost operator C, for each branch, define the set of attainable complex costs
C B , C ( ρ t 0 , ⋯ , ρ t n ) = { Tr ( C Γ ) : Γ ∈ C B ( ρ t 0 , ⋯ , ρ t n ) } ⊂ C , (118) \mathcal{C}_{B,\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})=\{\operatorname{Tr}(C\Gamma):\Gamma\in\mathcal{C}^{B}(\rho_{t_0},\cdots,\rho_{t_n})\}\subset\mathbb{C}, \tag*{(118)} C B , C ( ρ t 0 , ⋯ , ρ t n ) = { Tr ( C Γ ) : Γ ∈ C B ( ρ t 0 , ⋯ , ρ t n )} ⊂ C , (118)
and define
C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) = conv ( ⋃ B ∈ B r a n c h n C B , C ) . (119) \mathcal{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})=\operatorname{conv}\left(\bigcup_{B\in\mathrm{Branch}_n}\mathcal{C}_{B,\mathrm{C}}\right). \tag*{(119)} C conv , C ( ρ t 0 , ⋯ , ρ t n ) = conv ( B ∈ Branch n ⋃ C B , C ) . (119)
Theorem 5.3 (Geometric duality). The convex temporal transport value is the Euclidean distance from the origin to the convex attainable-cost set:
Q M K c o n v , C ( ρ t 0 , ⋯ , ρ t n ) = dist ( 0 , C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) ) . (120) \mathrm{QMK}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})=\operatorname{dist}(0,\mathcal{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})). \tag*{(120)} QMK conv , C ( ρ t 0 , ⋯ , ρ t n ) = dist ( 0 , C conv , C ( ρ t 0 , ⋯ , ρ t n )) . (120) Equivalently,
Q M K c o n v , C ( ρ t 0 , ⋯ , ρ t n ) = sup ∣ u ∣ ≤ 1 inf B ∈ B r a n c h n inf Γ ∈ C B , C ( ρ t 0 , ⋯ , ρ t n ) Re [ u ˉ Tr ( C Γ ) ] . (121) \mathrm{QMK}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})=\sup_{|u|\leq1}\inf_{B\in\mathrm{Branch}_n}\inf_{\Gamma\in\mathcal{C}_{B,\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})}\operatorname{Re}[\bar{u}\operatorname{Tr}(C\Gamma)]. \tag*{(121)} QMK conv , C ( ρ t 0 , ⋯ , ρ t n ) = ∣ u ∣ ≤ 1 sup B ∈ Branch n inf Γ ∈ C B , C ( ρ t 0 , ⋯ , ρ t n ) inf Re [ u ˉ Tr ( C Γ )] . (121) This provide a geometric duality for the original temporal optimal transport, and we can use it to simplify the optimization process.
Proof. By linearity of the cost functional, the set of attainable costs over convex temporal couplings is precisely C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) \mathcal{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n}) C conv , C ( ρ t 0 , ⋯ , ρ t n ) . Consequently,
Q M K c o n v , C ( ρ t 0 , ⋯ , ρ t n ) = min z ∈ C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) ∣ z ∣ = dist ( 0 , C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) ) , (122) \mathrm{QMK}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})
= \min_{z\in\mathcal{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})}|z|
= \operatorname{dist}\left(0,\mathcal{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})\right), \tag*{(122)} QMK conv , C ( ρ t 0 , ⋯ , ρ t n ) = z ∈ C conv , C ( ρ t 0 , ⋯ , ρ t n ) min ∣ z ∣ = dist ( 0 , C conv , C ( ρ t 0 , ⋯ , ρ t n ) ) , (122)
which proves (120) .
To establish the dual characterization, we use Eq. (71) . Since C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) \mathcal{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n}) C conv , C ( ρ t 0 , ⋯ , ρ t n ) and the closed unit disk { u ∈ C : ∣ u ∣ ≤ 1 } \{u\in\mathbb{C}:|u|\leq1\} { u ∈ C : ∣ u ∣ ≤ 1 } are nonempty compact convex sets, and Re ( u ‾ z ) \operatorname{Re}(\overline{u}z) Re ( u z ) is continuous and real-bilinear in ( u , z ) (u,z) ( u , z ) , the minimax theorem yields
inf z ∈ C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) ∣ z ∣ = sup ∣ u ∣ ≤ 1 inf z ∈ C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) Re ( u ‾ z ) . (123) \inf_{z\in\mathcal{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})}|z|
= \sup_{|u|\leq1}\inf_{z\in\mathcal{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})}\operatorname{Re}(\overline{u}z). \tag*{(123)} z ∈ C conv , C ( ρ t 0 , ⋯ , ρ t n ) inf ∣ z ∣ = ∣ u ∣ ≤ 1 sup z ∈ C conv , C ( ρ t 0 , ⋯ , ρ t n ) inf Re ( u z ) . (123)
For each fixed u u u , the functional z ↦ Re ( u ‾ z ) z\mapsto\operatorname{Re}(\overline{u}z) z ↦ Re ( u z ) is real-linear. Its infimum over a set therefore coincides with its infimum over the convex hull of that set. By the definition of the convex attainable-cost set, we obtain
inf z ∈ C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) Re ( u ‾ z ) = inf B ∈ Branch n inf Υ ∈ C B , C ( ρ t 0 , ⋯ , ρ t n ) Re [ u ‾ Tr ( C Υ ) ] . (124) \inf_{z\in\mathcal{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})}\operatorname{Re}(\overline{u}z)
= \inf_{B\in\operatorname{Branch}_n}\inf_{\Upsilon\in\mathcal{C}_{B,\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})}\operatorname{Re}\left[\overline{u}\operatorname{Tr}(\mathrm{C}\Upsilon)\right]. \tag*{(124)} z ∈ C conv , C ( ρ t 0 , ⋯ , ρ t n ) inf Re ( u z ) = B ∈ Branch n inf Υ ∈ C B , C ( ρ t 0 , ⋯ , ρ t n ) inf Re [ u Tr ( C Υ ) ] . (124)
Combining these identities establishes (121) . ■
Corollary 5.4 (Criterion for a zero temporal transport cost). Suppose that the convex attainable cost set is nonempty and compact. Then
Q M K c o n v , C ( ρ t 0 , ⋯ , ρ t n ) = 0 (125) \mathrm{QMK}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})=0 \tag*{(125)} QMK conv , C ( ρ t 0 , ⋯ , ρ t n ) = 0 (125) if and only if
0 ∈ C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) . (126) 0\in\mathcal{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n}). \tag*{(126)} 0 ∈ C conv , C ( ρ t 0 , ⋯ , ρ t n ) . (126) Consequently, the convex temporal transport cost may vanish even when the cost of every individual fixed branch is strictly positive. This occurs when attainable complex costs from different branches cancel under convex mixing.
Proof. By Theorem 5.3 ,
Q M K c o n v , C ( ρ t 0 , ⋯ , ρ t n ) = dist ( 0 , C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) ) . (127) \mathrm{QMK}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})
= \operatorname{dist}\left(0,\mathcal{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})\right). \tag*{(127)} QMK conv , C ( ρ t 0 , ⋯ , ρ t n ) = dist ( 0 , C conv , C ( ρ t 0 , ⋯ , ρ t n ) ) . (127) Since the convex attainable-cost set is compact and therefore closed, its distance from the origin vanishes if and only if it contains the origin. ■
Corollary 5.5 (Three temporal couplings suffice). Suppose that the convex attainable cost set is nonempty and compact, and that it is the convex hull of the fixed-branch attainable costs. For any fixed cost operator C \mathrm{C} C , the optimal value of Q M K c o n v , C ( ρ t 0 , ⋯ , ρ t n ) \mathrm{QMK}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n}) QMK conv , C ( ρ t 0 , ⋯ , ρ t n ) can be attained by a convex combination of at most three fixed-branch temporal states. If all attainable cost values are real, at most two fixed-branch temporal states suffice.
Proof. Since C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) \mathrm{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n}) C conv , C ( ρ t 0 , ⋯ , ρ t n ) is nonempty and compact, there exists an optimal cost value z ∗ z_* z ∗ such that
∣ z ∗ ∣ = inf z ∈ C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) ∣ z ∣ = Q M K c o n v , C ( ρ t 0 , ⋯ , ρ t n ) . (128) |z_*|=\inf_{z\in\mathrm{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})}|z|
=\mathrm{QMK}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n}). \tag*{(128)} ∣ z ∗ ∣ = z ∈ C conv , C ( ρ t 0 , ⋯ , ρ t n ) inf ∣ z ∣ = QMK conv , C ( ρ t 0 , ⋯ , ρ t n ) . (128) By the definition of the convex attainable cost set, z ∗ z_* z ∗ belongs to the convex hull of the fixed-branch attainable costs. Since these cost values lie in C ≅ R 2 \mathbb{C}\cong\mathbb{R}^{2} C ≅ R 2 , Carathéodory’s theorem implies that there exist at most three fixed-branch temporal states Υ 1 , Υ 2 , Υ 3 \Upsilon_1,\Upsilon_2,\Upsilon_3 Υ 1 , Υ 2 , Υ 3 and nonnegative coefficients λ 1 , λ 2 , λ 3 \lambda_1,\lambda_2,\lambda_3 λ 1 , λ 2 , λ 3 satisfying ∑ j = 1 3 λ j = 1 \sum_{j=1}^{3}\lambda_j=1 ∑ j = 1 3 λ j = 1 , such that
z ∗ = ∑ j = 1 3 λ j Tr ( C Υ j ) . (129) z_*=\sum_{j=1}^{3}\lambda_j\operatorname{Tr}(\mathrm{C}\Upsilon_j). \tag*{(129)} z ∗ = j = 1 ∑ 3 λ j Tr ( C Υ j ) . (129) Thus, the convex combination
Υ ∗ = ∑ j = 1 3 λ j Υ j (130) \Upsilon_*=\sum_{j=1}^{3}\lambda_j\Upsilon_j \tag*{(130)} Υ ∗ = j = 1 ∑ 3 λ j Υ j (130) attains the optimal convex temporal transport cost. If all attainable cost values are real, the relevant space has real dimension at most one, and Carathéodory’s theorem reduces the required number of temporal states to two. ■
The bound of three temporal states in Corollary 5.5 concerns the attainment of the optimal value for a fixed complex cost operator. Representing an arbitrary temporal state in the full operator space may require more components.
Corollary 5.6 (When convexification improves the optimum). For any fixed cost operator C \mathrm{C} C , the convex temporal transport cost satisfies
Q M K c o n v , C ( ρ t 0 , ⋯ , ρ t n ) ≤ inf B ∈ B r a n c h n Q M K B , C ( ρ t 0 , ⋯ , ρ t n ) . (131) \mathrm{QMK}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})
\leq\inf_{B\in\mathrm{Branch}_n}\mathrm{QMK}_{B,\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n}). \tag*{(131)} QMK conv , C ( ρ t 0 , ⋯ , ρ t n ) ≤ B ∈ Branch n inf QMK B , C ( ρ t 0 , ⋯ , ρ t n ) . (131) The inequality can be strict due to cancellation between complex cost values attained by different branches. By contrast, for the real-linear cost functional, convexification does not change the optimal value:
inf z ∈ C c o n v , C ( ρ t 0 , ⋯ , ρ t n ) Re ( z ) = inf B ∈ B r a n c h n inf Υ ∈ C B , C ( ρ t 0 , ⋯ , ρ t n ) Re [ Tr ( C Υ ) ] . (132) \inf_{z\in\mathrm{C}_{\mathrm{conv},\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})}\operatorname{Re}(z)
=
\inf_{B\in\mathrm{Branch}_n}\inf_{\Upsilon\in\mathrm{C}_{B,\mathrm{C}}(\rho_{t_0},\cdots,\rho_{t_n})}\operatorname{Re}[\operatorname{Tr}(\mathrm{C}\Upsilon)]. \tag*{(132)} z ∈ C conv , C ( ρ t 0 , ⋯ , ρ t n ) inf Re ( z ) = B ∈ Branch n inf Υ ∈ C B , C ( ρ t 0 , ⋯ , ρ t n ) inf Re [ Tr ( C Υ )] . (132) Thus, any improvement resulting from convexification arises from the nonlinearity of the modulus cost.
Dual branches, Hermitianization and Margenau-Hill temporal quantum optimal transport
Let B c \mathrm{B}^{\mathrm{c}} B c be the complementary branch obtained by interchanging L L L and R R R at every temporal slot. From the temporal state construction,
Υ B c = ( Υ B ) † . (133) \Upsilon^{\mathrm{B}^{\mathrm{c}}} = \left(\Upsilon^{\mathrm{B}}\right)^\dagger. \tag*{(133)} Υ B c = ( Υ B ) † . (133)
Proposition 5.7 (Dual-branch symmetry). For every cost operator C \mathrm{C} C ,
T M K B , C ( ρ ) = T M K B c , C † ( ρ ) . (134) \mathrm{TMK}_{\mathrm{B},\mathrm{C}}(\rho) = \mathrm{TMK}_{\mathrm{B}^{\mathrm{c}},\mathrm{C}^\dagger}(\rho). \tag*{(134)} TMK B , C ( ρ ) = TMK B c , C † ( ρ ) . (134) In particular, if C = C † \mathrm{C} = \mathrm{C}^\dagger C = C † , then
T M K B , C ( ρ ) = T M K B c , C ( ρ ) . (135) \mathrm{TMK}_{\mathrm{B},\mathrm{C}}(\rho) = \mathrm{TMK}_{\mathrm{B}^{\mathrm{c}},\mathrm{C}}(\rho). \tag*{(135)} TMK B , C ( ρ ) = TMK B c , C ( ρ ) . (135)
Proof. For corresponding temporal states on the two branches,
Tr [ C † Υ B c ] = Tr [ C † ( Υ B ) † ] = Tr [ C Υ B ] . (136) \operatorname{Tr}\left[\mathrm{C}^\dagger\Upsilon^{\mathrm{B}^{\mathrm{c}}}\right]
= \operatorname{Tr}\left[\mathrm{C}^\dagger\left(\Upsilon^{\mathrm{B}}\right)^\dagger\right]
= \operatorname{Tr}\left[\mathrm{C}\Upsilon^{\mathrm{B}}\right]. \tag*{(136)} Tr [ C † Υ B c ] = Tr [ C † ( Υ B ) † ] = Tr [ C Υ B ] . (136)
Taking absolute values and then optimizing proves the result. ■
A branch-symmetric mixture,
p B = p B c , (137) p_{\mathrm{B}} = p_{\mathrm{B}^{\mathrm{c}}}, \tag*{(137)} p B = p B c , (137)
produces a Hermitian temporal state, called Margenau-Hill state [18 , 20 ] . For example
1 2 ( Υ B + Υ B c ) = 1 2 ( Υ B + ( Υ B ) † ) (138) \frac{1}{2}\left(\Upsilon^{\mathrm{B}} + \Upsilon^{\mathrm{B}^{\mathrm{c}}}\right) = \frac{1}{2}\left(\Upsilon^{\mathrm{B}} + \left(\Upsilon^{\mathrm{B}}\right)^\dagger\right) \tag*{(138)} 2 1 ( Υ B + Υ B c ) = 2 1 ( Υ B + ( Υ B ) † ) (138)
is a Margenau-Hill temporal state. For a Hermitian cost, its cost value is
Tr [ C Υ B + Υ B c 2 ] = Re Tr [ C Υ B ] . (139) \operatorname{Tr}\left[\mathrm{C}\frac{\Upsilon^{\mathrm{B}}+\Upsilon^{\mathrm{B}^{\mathrm{c}}}}{2}\right] = \operatorname{Re}\operatorname{Tr}\left[\mathrm{C}\Upsilon^{\mathrm{B}}\right]. \tag*{(139)} Tr [ C 2 Υ B + Υ B c ] = Re Tr [ C Υ B ] . (139)
The Lüders-von Neumann state (pseudo-density operator) is obtained by a distinguished branch-symmetric mixture over all admissible orderings,
Υ = 1 2 n ∑ B ∈ B r a n c h n Υ B . (140) \Upsilon= \frac{1}{2^n}\sum_{\mathrm{B}\in\mathrm{Branch}_n}\Upsilon^{\mathrm{B}}. \tag*{(140)} Υ = 2 n 1 B ∈ Branch n ∑ Υ B . (140)
It has a recursive expression
Υ t k , … , t 0 = 1 2 { J [ E k ] , Υ t k − 1 , … , t 0 } , Υ t 0 = ρ t 0 , (141) \Upsilon_{t_k,\ldots,t_0} = \frac{1}{2}\left\{\mathcal{J}[\mathcal{E}_k],\Upsilon_{t_{k-1},\ldots,t_0}\right\},\quad\Upsilon_{t_0}=\rho_{t_0}, \tag*{(141)} Υ t k , … , t 0 = 2 1 { J [ E k ] , Υ t k − 1 , … , t 0 } , Υ t 0 = ρ t 0 , (141)
where { ∙ , ∙ } \{\bullet,\bullet\} { ∙ , ∙ } is anti-commutator. The optimal transport for pseudo-density operator are recently discussed in [55 ] , which can be regarded as a special case of our setting. From our previous discussion and combining the Hermitization, we will obtain more temporal optimal transport with Hermitian temporal states.
Spatiotemporal optimal transport with a fixed background structure
The general spatiotemporal optimal transport requires further structure. Since, given an arbitrary spacetime set A \mathcal{A} A , the set of all spatiotemporal states is not well defined, we need to introduce a background structure [68 ] .
We will consider the following setting, given a background spacetime lattice
L = { ( t 0 , x 1 ) , ⋯ , ( t 0 , x m ) ; ⋯ ; ( t n , x 1 ) , ⋯ , ( t n , x m ) } . \mathcal{L} = \{(t_0,x_1),\cdots,(t_0,x_m);\cdots;(t_n,x_1),\cdots,(t_n,x_m)\}. L = {( t 0 , x 1 ) , ⋯ , ( t 0 , x m ) ; ⋯ ; ( t n , x 1 ) , ⋯ , ( t n , x m )} .
A spacetime set A ⊂ L \mathcal{A} \subset\mathcal{L} A ⊂ L , see Figure 2 for illustration, where the spacetime set is shown as red dots. For the spacetime lattice L \mathcal{L} L , the spatiotemporal state is well-defined, which are determined by initial state ρ x 1 , ⋯ , x m ( t 0 ) \rho_{x_1,\cdots,x_m}(t_0) ρ x 1 , ⋯ , x m ( t 0 ) and evolutions E t k → t k + 1 \mathcal{E}_{t_k\to t_{k+1}} E t k → t k + 1 , k = 0 , ⋯ , n − 1 k=0,\cdots,n-1 k = 0 , ⋯ , n − 1 . For instance, for fixed branch B, it is of the form
Figure 2. The background spacetime lattice and spacetime set.
Υ L B = J n ⋆ B J n − 1 ⋆ B ⋯ ⋆ B J 1 ⋆ B ρ x 1 , ⋯ , x m ( t 0 ) , (142) \Upsilon_{\mathcal{L}}^{\mathrm{B}}=J_n\star_{\mathrm{B}}J_{n-1}\star_{\mathrm{B}}\cdots\star_{\mathrm{B}}J_1\star_{\mathrm{B}}\rho_{x_1,\cdots,x_m}(t_0), \tag*{(142)} Υ L B = J n ⋆ B J n − 1 ⋆ B ⋯ ⋆ B J 1 ⋆ B ρ x 1 , ⋯ , x m ( t 0 ) , (142)
where J k J_k J k ’s are Jamiłkowski operators. The spatiotemporal state for A \mathcal{A} A is thus obtained by taking reduced state Υ A = Tr L ∖ A Υ L \Upsilon_{\mathcal{A}}=\operatorname{Tr}_{\mathcal{L}\setminus\mathcal{A}}\Upsilon_{\mathcal{L}} Υ A = Tr L ∖ A Υ L . Since Υ L \Upsilon_{\mathcal{L}} Υ L is a temporal state, the discussion in the previous section can be applied, we can first consider the temporal couplings for L \mathcal{L} L and then take the reduced state to define the quantum optimal transport for the general spatiotemporal setting. The optimization will taken over the initial state ρ x 1 , ⋯ , x m ( t 0 ) \rho_{x_1,\cdots,x_m}(t_0) ρ x 1 , ⋯ , x m ( t 0 ) and all quantum channels E t k → t k + 1 \mathcal{E}_{t_k\to t_{k+1}} E t k → t k + 1 , k = 0 , ⋯ , n − 1 k=0,\cdots,n-1 k = 0 , ⋯ , n − 1 that gives the correct marginals.
Take fixed branch case as an example, the set of all spatiotemporal state for L \mathcal{L} L is
State B ( L ) = { Υ = J n ⋆ B J n − 1 ⋆ B ⋯ ⋆ B J 1 ⋆ B ρ x 1 , ⋯ , x m ( t 0 ) : J k ∈ J a m i , ρ x 1 , ⋯ , x m ( t 0 ) ≥ 0 , Tr ρ x 1 , ⋯ , x m ( t 0 ) = 1 } \operatorname{State}_{\mathrm{B}}(\mathcal{L})=\{\Upsilon=J_n\star_{\mathrm{B}}J_{n-1}\star_{\mathrm{B}}\cdots\star_{\mathrm{B}}J_1\star_{\mathrm{B}}\rho_{x_1,\cdots,x_m}(t_0): J_k\in\mathbf{Jami},\ \rho_{x_1,\cdots,x_m}(t_0)\geq0,\ \operatorname{Tr}\rho_{x_1,\cdots,x_m}(t_0)=1\} State B ( L ) = { Υ = J n ⋆ B J n − 1 ⋆ B ⋯ ⋆ B J 1 ⋆ B ρ x 1 , ⋯ , x m ( t 0 ) : J k ∈ Jami , ρ x 1 , ⋯ , x m ( t 0 ) ≥ 0 , Tr ρ x 1 , ⋯ , x m ( t 0 ) = 1 }
The set of all spatiotemporal state for A \mathcal{A} A is State B ( A ) = { Υ A = Tr L ∖ A Υ L : Υ ∈ State B ( L ) } \operatorname{State}_{\mathrm{B}}(\mathcal{A})=\{\Upsilon_{\mathcal{A}}=\operatorname{Tr}_{\mathcal{L}\setminus\mathcal{A}}\Upsilon_{\mathcal{L}}:\Upsilon\in\operatorname{State}_{\mathrm{B}}(\mathcal{L})\} State B ( A ) = { Υ A = Tr L ∖ A Υ L : Υ ∈ State B ( L )}
For a set of given marginal states { ρ z k } \{\rho_{z_k}\} { ρ z k } , the spatiotemporal couplings are thus
C B ( ρ z 1 , ⋯ , ρ z N ) = { Υ ∈ State B ( A ) : Tr z k Υ = ρ z k } (143) \mathcal{C}^{\mathrm{B}}(\rho_{z_1},\cdots,\rho_{z_N})=\{\Upsilon\in\operatorname{State}_{\mathrm{B}}(\mathcal{A}):\operatorname{Tr}_{z_k}\Upsilon=\rho_{z_k}\} \tag*{(143)} C B ( ρ z 1 , ⋯ , ρ z N ) = { Υ ∈ State B ( A ) : Tr z k Υ = ρ z k } (143)
All other types of couplings in the previous section can be defined similarly.
Kantorovich-Rubinstein metric for spatiotemporal states
The phase-space description makes it more natural to define the quantum Wasserstein distance, since it allows for a direct comparison with the classical definition. See Ref. [37 Chapter 6] for basics of Wasserstein distance. Recall that for classical probability distributions p p p and q q q over the same polish space X \mathcal{X} X , the Kantorovich-Rubinstein theorem states that the Wasserstein-1 distance can be expressed in the Kantorovich-Rubinstein form using 1-Lipschitz functions: W 1 ( p , q ) = KR ( p , q ) W_1(p,q)=\operatorname{KR}(p,q) W 1 ( p , q ) = KR ( p , q ) . In this section, we generalize this to spatiotemporal states. We emphasize that our approach is quite different from those discussed in spatial quantum optimal transport, e.g., [42 , 46 , 47 , 51 ] , where they use a generalization of Wasserstein’s definition.
For spatiotemporal causal set A = { z 1 , ⋯ , z N } \mathcal{A}=\{z_1,\cdots,z_N\} A = { z 1 , ⋯ , z N } with z k = ( t k , x k ) z_k=(t_k,x_k) z k = ( t k , x k ) ’s spacetime points, let Υ , Ξ ∈ S t a t e s t ( H A ) \Upsilon,\Xi\in\mathbf{State}_{\mathrm{st}}(\mathcal{H}_{\mathcal{A}}) Υ , Ξ ∈ State st ( H A ) be two spatiotemporal states on the same lattice. Their IC-POVM phase-space representations are
Q Υ ( α ) = Tr [ F α Υ ] , Q Ξ ( α ) = Tr [ F α Ξ ] . (144) Q_{\Upsilon}(\alpha)=\operatorname{Tr}[F_{\alpha}\Upsilon],\qquad Q_{\Xi}(\alpha)=\operatorname{Tr}[F_{\alpha}\Xi]. \tag*{(144)} Q Υ ( α ) = Tr [ F α Υ ] , Q Ξ ( α ) = Tr [ F α Ξ ] . (144)
Both are normalized:
∑ α Q Υ ( α ) = ∑ α Q Ξ ( α ) = 1. (145) \sum_{\alpha}Q_{\Upsilon}(\alpha)=\sum_{\alpha}Q_{\Xi}(\alpha)=1. \tag*{(145)} α ∑ Q Υ ( α ) = α ∑ Q Ξ ( α ) = 1. (145)
Their difference therefore has total mass zero, although it can be complex.
The phase-space is Ω A = Ω z 1 × ⋯ × Ω z N \Omega_{\mathcal{A}}=\Omega_{z_1}\times\cdots\times\Omega_{z_N} Ω A = Ω z 1 × ⋯ × Ω z N , we can introduce equip a ground distance d A ( α , β ) d_{\mathcal{A}}(\alpha,\beta) d A ( α , β ) on it. For a function f : Ω A → C f:\Omega_{\mathcal{A}}\to\mathbb{C} f : Ω A → C , define Lipschitz seminorm
Lip d A ( f ) = ∥ f ∥ L i p = sup α ≠ β ∣ f ( α ) − f ( β ) ∣ d A ( α , β ) . (146) \operatorname{Lip}_{d_{\mathcal{A}}}(f)=\lVert f\rVert_{\mathrm{Lip}}=\sup_{\alpha\ne\beta}\frac{\lvert f(\alpha)-f(\beta)\rvert}{d_{\mathcal{A}}(\alpha,\beta)}. \tag*{(146)} Lip d A ( f ) = ∥ f ∥ Lip = α = β sup d A ( α , β ) ∣ f ( α ) − f ( β )∣ . (146)
Equivalently, Lip ( f ) \operatorname{Lip}(f) Lip ( f ) is the smallest number L ≥ 0 L\ge0 L ≥ 0 such that
∣ f ( x ) − f ( y ) ∣ ≤ L d ( x , y ) for all x , y ∈ X . \lvert f(x)-f(y)\rvert\le Ld(x,y)\quad\text{for all }x,y\in\mathcal{X}. ∣ f ( x ) − f ( y )∣ ≤ L d ( x , y ) for all x , y ∈ X .
If Lip ( f ) ≤ 1 \operatorname{Lip}(f)\le1 Lip ( f ) ≤ 1 , then f f f is called 1-Lipschitz. It is called a seminorm since ∥ f ∥ L i p = 0 \lVert f\rVert_{\mathrm{Lip}}=0 ∥ f ∥ Lip = 0 does not imply f = 0 f=0 f = 0 , but all other conditions of norm are satisfied.
Definition 7.1 (Spatiotemporal Kantorovich–Rubinstein distance). For two spatiotemporal state Υ , Ξ ∈ S t a t e ( H A ) \Upsilon,\Xi\in\mathbf{State}(\mathcal{H}_{\mathcal{A}}) Υ , Ξ ∈ State ( H A ) , and a chosen IC-POVM spatiotemporal frame { F α } \{F_{\alpha}\} { F α } , fix a reference point α 0 ∈ Ω A \alpha_0\in\Omega_{\mathcal{A}} α 0 ∈ Ω A . The spatiotemporal Kantorovich–Rubinstein distance induced by spatiotemporal Kirkwood-Dirac distribution is
KR ( Υ , Ξ ) = sup { ∣ ∑ α ∈ Ω A f ( α ) [ Q Υ ( α ) − Q Ξ ( α ) ] ∣ : f : Ω A → C , Lip d A ( f ) ≤ 1 , f ( α 0 ) = 0 } . (147) \operatorname{KR}(\Upsilon,\Xi)=\sup\left\{\left\lvert\sum_{\alpha\in\Omega_{\mathcal{A}}}f(\alpha)\left[Q_{\Upsilon}(\alpha)-Q_{\Xi}(\alpha)\right]\right\rvert:f:\Omega_{\mathcal{A}}\to\mathbb{C},\ \operatorname{Lip}_{d_{\mathcal{A}}}(f)\le1,\ f(\alpha_0)=0\right\}. \tag*{(147)} KR ( Υ , Ξ ) = sup { α ∈ Ω A ∑ f ( α ) [ Q Υ ( α ) − Q Ξ ( α ) ] : f : Ω A → C , Lip d A ( f ) ≤ 1 , f ( α 0 ) = 0 } . (147) The condition f ( α 0 ) = 0 f(\alpha_0)=0 f ( α 0 ) = 0 removes the irrelevant freedom to add a constant, as given constant c c c , f c ( ∙ ) : = f + c f_c(\bullet):=f+c f c ( ∙ ) := f + c has the same Lipschitz norm as f f f , f ( α 0 ) = 0 f(\alpha_0)=0 f ( α 0 ) = 0 chooses exactly
one representative from each equivalence class { f + c } \{f+c\} { f + c } . For the Margenau-Hill distribution, we may restrict f f f to real-valued functions, thereby obtaining K R M H ( Υ , Ξ ) \mathrm{KR}_{\mathrm{MH}}(\Upsilon,\Xi) KR MH ( Υ , Ξ ) . We can also define the Kantorovich–Rubinstein distance in Eq. (147) by taking the supremum over all real-valued functions. To distinguish the two cases, we denote them by K R K D R ( Υ , Ξ ) \mathrm{KR}^{\mathbb{R}}_{\mathrm{KD}}(\Upsilon,\Xi) KR KD R ( Υ , Ξ ) and K R K D C ( Υ , Ξ ) \mathrm{KR}^{\mathbb{C}}_{\mathrm{KD}}(\Upsilon,\Xi) KR KD C ( Υ , Ξ ) . Note that K R K D R ( Υ , Ξ ) ≥ K R M H ( Υ , Ξ ) \mathrm{KR}^{\mathbb{R}}_{\mathrm{KD}}(\Upsilon,\Xi) \ge\mathrm{KR}_{\mathrm{MH}}(\Upsilon,\Xi) KR KD R ( Υ , Ξ ) ≥ KR MH ( Υ , Ξ ) .
The metric also has the operator representation
K R ( Υ , Ξ ) = sup Lip d A ( f ) ≤ 1 f ( α 0 ) = 0 ∣ Tr [ A f ( Υ − Ξ ) ] ∣ (148) \mathrm{KR}(\Upsilon,\Xi)=\sup_{\substack{\operatorname{Lip}_{d}^{A}(f)\le1\\f(\alpha_{0})=0}}\left|\operatorname{Tr}\left[A_f(\Upsilon-\Xi)\right]\right| \tag*{(148)} KR ( Υ , Ξ ) = Lip d A ( f ) ≤ 1 f ( α 0 ) = 0 sup ∣ Tr [ A f ( Υ − Ξ ) ] ∣ (148)
where
A f = ∑ α f ( α ) F α . (149) A_f=\sum_{\alpha}f(\alpha)F_{\alpha}. \tag*{(149)} A f = α ∑ f ( α ) F α . (149)
For f ∈ R Ω A f\in\mathbb{R}^{\Omega_A} f ∈ R Ω A , A f A_f A f is Hermitian. The Kantorovich–Rubinstein distance measures the largest distinguishability of the two spatiotemporal states by a phase-space observable whose coefficients is 1-Lipschitz.
Theorem 7.2 (Metric properties). For a finite metric phase space and a tensor product of local IC-POVM frames , K R ( Υ , Ξ ) \mathrm{KR}(\Upsilon,\Xi) KR ( Υ , Ξ ) is a metric on State ( H A ) \operatorname{State}(\mathcal{H}_A) State ( H A ) and a metric on KD ( A ) \operatorname{KD}(\mathcal{A}) KD ( A ) if regard as function of Q Υ Q_{\Upsilon} Q Υ and Q Ξ Q_{\Xi} Q Ξ .
Proof. Nonnegativity is immediate. Exchanging Υ \Upsilon Υ and Ξ \Xi Ξ changes the sign of the expression inside the absolute value, so symmetry holds.
For three spatiotemporal states Υ \Upsilon Υ , Ξ \Xi Ξ , Θ \Theta Θ , every admissible f f f satisfies
∣ ∑ α f ( α ) [ Q Υ ( α ) − Q Θ ( α ) ] ∣ ≤ ∣ ∑ α f ( α ) [ Q Υ ( α ) − Q Ξ ( α ) ] ∣ + ∣ ∑ α f ( α ) [ Q Ξ ( α ) − Q Θ ( α ) ] ∣ . (150) \left|\sum_{\alpha}f(\alpha)\left[Q_{\Upsilon}(\alpha)-Q_{\Theta}(\alpha)\right]\right|
\le\left|\sum_{\alpha}f(\alpha)\left[Q_{\Upsilon}(\alpha)-Q_{\Xi}(\alpha)\right]\right|+\left|\sum_{\alpha}f(\alpha)\left[Q_{\Xi}(\alpha)-Q_{\Theta}(\alpha)\right]\right|. \tag*{(150)} α ∑ f ( α ) [ Q Υ ( α ) − Q Θ ( α ) ] ≤ α ∑ f ( α ) [ Q Υ ( α ) − Q Ξ ( α ) ] + α ∑ f ( α ) [ Q Ξ ( α ) − Q Θ ( α ) ] . (150)
Taking the supremum gives the triangle inequality. It remains to prove definiteness. If
K R ( Υ , Ξ ) = 0 , (151) \mathrm{KR}(\Upsilon,\Xi)=0, \tag*{(151)} KR ( Υ , Ξ ) = 0 , (151)
then the difference Q Υ − Q Ξ Q_{\Upsilon}-Q_{\Xi} Q Υ − Q Ξ pairs to zero with every Lipschitz function. On a finite metric space (like Ω A \Omega_A Ω A ) every function is Lipschitz after multiplication by a sufficiently small nonzero constant. Hence
Q Υ ( α ) = Q Ξ ( α ) (152) Q_{\Upsilon}(\alpha)=Q_{\Xi}(\alpha) \tag*{(152)} Q Υ ( α ) = Q Ξ ( α ) (152)
for every α \alpha α . Informational completeness of the tensor-product frame implies Υ = Ξ \Upsilon=\Xi Υ = Ξ . ■
When Q Υ Q_{\Upsilon} Q Υ and Q Ξ Q_{\Xi} Q Ξ are positive probability distributions, the complex dual formulation (147) reduces to the classical Kantorovich-Rubinstein metric which coincides with classical 1-Wasserstein metric.
For spatial density operators ρ A , ρ B ∈ State ( H ) \rho_A,\rho_B\in\operatorname{State}(\mathcal{H}) ρ A , ρ B ∈ State ( H ) , and IC-POVM frame { F α , F β } \{F_{\alpha},F_{\beta}\} { F α , F β } Q ρ A ( α ) Q_{\rho_A}(\alpha) Q ρ A ( α ) and Q ρ B ( β ) Q_{\rho_B}(\beta) Q ρ B ( β ) are probability distributions, give a ground distance d ( α , β ) d(\alpha,\beta) d ( α , β ) over phase space and take supremum over real 1-Lipschitz functions, we obtain a distance function of density operators KR ( ρ A , ρ B ) : = KR K D ( ρ A , ρ B ) = KR M H ( ρ A , ρ B ) \operatorname{KR}(\rho_A,\rho_B) := \operatorname{KR}_{\mathrm{KD}}(\rho_A,\rho_B) = \operatorname{KR}_{\mathrm{MH}}(\rho_A,\rho_B) KR ( ρ A , ρ B ) := KR KD ( ρ A , ρ B ) = KR MH ( ρ A , ρ B ) . Recall that classical p p p -Wasserstein metric is defined by
W p C ( ρ A , ρ B ) = [ inf { ∑ α , β d ( α , β ) p P ( α , β ) : ∑ β P ( α , β ) = Q ρ A ( α ) , ∑ α P ( α , β ) = Q ρ B ( β ) } ] 1 / p (153) W_p^{\mathrm{C}}(\rho_A,\rho_B)=\left[\inf\left\{\sum_{\alpha,\beta}d(\alpha,\beta)^pP(\alpha,\beta):\sum_\beta P(\alpha,\beta)=Q_{\rho_A}(\alpha),\ \sum_\alpha P(\alpha,\beta)=Q_{\rho_B}(\beta)\right\}\right]^{1/p} \tag*{(153)} W p C ( ρ A , ρ B ) = inf ⎩ ⎨ ⎧ α , β ∑ d ( α , β ) p P ( α , β ) : β ∑ P ( α , β ) = Q ρ A ( α ) , α ∑ P ( α , β ) = Q ρ B ( β ) ⎭ ⎬ ⎫ 1/ p (153)
We have KR ( ρ A , ρ B ) = W 1 C ( ρ A , ρ B ) \operatorname{KR}(\rho_A,\rho_B)=W_1^{\mathrm{C}}(\rho_A,\rho_B) KR ( ρ A , ρ B ) = W 1 C ( ρ A , ρ B ) , Theorem 7.2 implies that this is a metric. Assume that the joint frame is the product frame { F α ⊗ F β } \{F_\alpha\otimes F_\beta\} { F α ⊗ F β } , so that
Q ρ A B ( α , β ) = Tr [ ( F α ⊗ F β ) ρ A B ] . (154) Q_{\rho_{AB}}(\alpha,\beta)=\operatorname{Tr}[(F_\alpha\otimes F_\beta)\rho_{AB}]. \tag*{(154)} Q ρ A B ( α , β ) = Tr [( F α ⊗ F β ) ρ A B ] . (154)
Every such distribution is a classical coupling of Q ρ A Q_{\rho_A} Q ρ A and Q ρ B Q_{\rho_B} Q ρ B . However, not every classical coupling admits this quantum realization. We can also define
W p Q ( ρ A , ρ B ) = [ inf { ∑ α , β d ( α , β ) p Q ρ A B ( α , β ) : Tr B ρ A B = ρ A , Tr A ρ A B = ρ B } ] 1 / p . (155) W_p^{\mathrm{Q}}(\rho_A,\rho_B)=\left[\inf\left\{\sum_{\alpha,\beta}d(\alpha,\beta)^pQ_{\rho_{AB}}(\alpha,\beta):\operatorname{Tr}_B\rho_{AB}=\rho_A,\ \operatorname{Tr}_A\rho_{AB}=\rho_B\right\}\right]^{1/p}. \tag*{(155)} W p Q ( ρ A , ρ B ) = inf ⎩ ⎨ ⎧ α , β ∑ d ( α , β ) p Q ρ A B ( α , β ) : Tr B ρ A B = ρ A , Tr A ρ A B = ρ B ⎭ ⎬ ⎫ 1/ p . (155)
It is generally not a metric. To see this, consider the qubit tetrahedral informationally complete positive operator-valued measure
F i = 1 4 ( 1 + v i ⋅ σ ) , i = 1 , … , 4 , (156) F_i=\frac{1}{4}(1+\mathbf{v}_i\cdot\boldsymbol{\sigma}),\qquad i=1,\ldots,4, \tag*{(156)} F i = 4 1 ( 1 + v i ⋅ σ ) , i = 1 , … , 4 , (156)
and the discrete metric d ( i , j ) = 1 − δ i j d(i,j)=1-\delta_{ij} d ( i , j ) = 1 − δ ij . For every p > 0 p>0 p > 0 , we have d ( i , j ) p = d ( i , j ) d(i,j)^p=d(i,j) d ( i , j ) p = d ( i , j ) . Consequently,
W p Q ( 1 / 2 , 1 / 2 ) = [ 1 − sup ρ A B ≥ 0 Tr B ρ A B = Tr A ρ A B = 1 / 2 Tr [ ρ A B ∑ i F i ⊗ F i ] ] 1 / p = ( 1 − 1 3 ) 1 / p = ( 2 3 ) 1 / p , (157) \begin{aligned}
W_p^{\mathrm{Q}}(1/2,1/2)=\left[1-\sup_{\substack{\rho_{AB}\geq0\\ \operatorname{Tr}_B\rho_{AB}=\operatorname{Tr}_A\rho_{AB}=1/2}}\operatorname{Tr}\left[\rho_{AB}\sum_iF_i\otimes F_i\right]\right]^{1/p} \\
&=\left(1-\frac{1}{3}\right)^{1/p}=\left(\frac{2}{3}\right)^{1/p}, \tag*{(157)}
\end{aligned} W p Q ( 1/2 , 1/2 ) = 1 − ρ A B ≥ 0 Tr B ρ A B = Tr A ρ A B = 1/2 sup Tr [ ρ A B i ∑ F i ⊗ F i ] 1/ p = ( 1 − 3 1 ) 1/ p = ( 3 2 ) 1/ p , (157)
where we used5
∑ i F i ⊗ F i = 1 3 P s y m . (158) \sum_iF_i\otimes F_i=\frac{1}{3}P_{\mathrm{sym}}. \tag*{(158)} i ∑ F i ⊗ F i = 3 1 P sym . (158)
v 1 = ( 1 , 1 , 1 ) 3 , v 2 = ( 1 , − 1 , − 1 ) 3 , v 3 = ( − 1 , 1 , − 1 ) 3 , v 4 = ( − 1 , − 1 , 1 ) 3 . (159) \mathbf{v}_1=\frac{(1,1,1)}{\sqrt{3}},\qquad\mathbf{v}_2=\frac{(1,-1,-1)}{\sqrt{3}},\qquad\mathbf{v}_3=\frac{(-1,1,-1)}{\sqrt{3}},\qquad\mathbf{v}_4=\frac{(-1,-1,1)}{\sqrt{3}}. \tag*{(159)} v 1 = 3 ( 1 , 1 , 1 ) , v 2 = 3 ( 1 , − 1 , − 1 ) , v 3 = 3 ( − 1 , 1 , − 1 ) , v 4 = 3 ( − 1 , − 1 , 1 ) . (159)
they satisfy
∑ i v i = 0 , ∑ i v i v i T = 4 3 1 3 . (160) \sum_i\mathbf{v}_i=0,\qquad\sum_i\mathbf{v}_i\mathbf{v}_i^{\mathsf{T}}=\frac{4}{3}1_3. \tag*{(160)} i ∑ v i = 0 , i ∑ v i v i T = 3 4 1 3 . (160)
Each coordinate sums to zero, each diagonal second moment equals 4 / 3 4/3 4/3 , and each off-diagonal second moment vanishes. Expanding the tensor products therefore gives
∑ i = 1 4 F i ⊗ F i = 1 16 ∑ i = 1 4 [ 1 ⊗ 1 + ( v i ⋅ σ ) ⊗ 1 + 1 ⊗ ( v i ⋅ σ ) (161) \sum_{i=1}^{4}F_i\otimes F_i=\frac{1}{16}\sum_{i=1}^{4}[1\otimes1+(\mathbf{v}_i\cdot\boldsymbol{\sigma})\otimes1+1\otimes(\mathbf{v}_i\cdot\boldsymbol{\sigma}) \tag*{(161)} i = 1 ∑ 4 F i ⊗ F i = 16 1 i = 1 ∑ 4 [ 1 ⊗ 1 + ( v i ⋅ σ ) ⊗ 1 + 1 ⊗ ( v i ⋅ σ ) (161)
with P s y m = ( 1 ⊗ 1 + S ) / 2 P_{\mathrm{sym}}=(1\otimes1+S)/2 P sym = ( 1 ⊗ 1 + S ) /2 orthogonal projector onto the symmetric subspace, where S S S is swap operator. The supremum is attained by ρ A B = P s y m / 3 \rho_{AB}=P_{\mathrm{sym}}/3 ρ A B = P sym /3 , whose two marginals are 1 / 2 1/2 1/2 . Thus W p Q ( ρ , ρ ) W^Q_p(\rho,\rho) W p Q ( ρ , ρ ) need not vanish, violating a necessary metric axiom.
Proposition 7.3. For a finite metric outcome space and a product informationally complete positive operator-valued measure frame,
KR ( ρ A , ρ B ) = W 1 C ( ρ A , ρ B ) ≤ W 1 Q ( ρ A , ρ B ) . (162) \operatorname{KR}(\rho_A,\rho_B)=W^C_1(\rho_A,\rho_B)\le W^Q_1(\rho_A,\rho_B). \tag*{(162)} KR ( ρ A , ρ B ) = W 1 C ( ρ A , ρ B ) ≤ W 1 Q ( ρ A , ρ B ) . (162) Equality holds if and only if at least one optimal classical transport plan is induced by a bipartite density operator with marginals ρ A \rho_A ρ A and ρ B \rho_B ρ B .
Proof. Let ρ A B \rho_{AB} ρ A B be any admissible bipartite density operator and write π ( α , β ) = Q ρ A B ( α , β ) \pi(\alpha,\beta)=Q_{\rho_{AB}}(\alpha,\beta) π ( α , β ) = Q ρ A B ( α , β ) . For every real-valued 1-Lipschitz function f f f ,
∣ ∑ α f ( α ) [ Q ρ A ( α ) − Q ρ B ( α ) ] ∣ = ∣ ∑ α , β [ f ( α ) − f ( β ) ] π ( α , β ) ∣ ≤ ∑ α , β d ( α , β ) π ( α , β ) . (163) \left|\sum_{\alpha} f(\alpha)[Q_{\rho_A}(\alpha)-Q_{\rho_B}(\alpha)]\right|
=\left|\sum_{\alpha,\beta}[f(\alpha)-f(\beta)]\pi(\alpha,\beta)\right|
\le\sum_{\alpha,\beta}d(\alpha,\beta)\pi(\alpha,\beta). \tag*{(163)} α ∑ f ( α ) [ Q ρ A ( α ) − Q ρ B ( α )] = α , β ∑ [ f ( α ) − f ( β )] π ( α , β ) ≤ α , β ∑ d ( α , β ) π ( α , β ) . (163)
Taking the supremum over f f f and then the infimum over ρ A B \rho_{AB} ρ A B proves the inequality. Classical Kantorovich–Rubinstein duality identifies KR ( ρ A , ρ B ) \operatorname{KR}(\rho_A,\rho_B) KR ( ρ A , ρ B ) with the minimum cost over all classical couplings of Q ρ A Q_{\rho_A} Q ρ A and Q ρ B Q_{\rho_B} Q ρ B . The admissible quantum states form a compact set, and their induced distributions form a subset of these classical couplings. Thus equality holds precisely when this subset contains a classical optimal coupling. ■
For example, consider a qubit with the tetrahedral informationally complete positive operator-valued measure, we have
KR ( 1 / 2 , 1 / 2 ) = 0. (164) \operatorname{KR}(1/2,1/2)=0. \tag*{(164)} KR ( 1/2 , 1/2 ) = 0. (164)
1 4 1 ⊗ 1 + 1 12 ∑ a = 1 3 σ a ⊗ σ a (165) \frac{1}{4}1\otimes1+\frac{1}{12}\sum_{a=1}^{3}\sigma_a\otimes\sigma_a \tag*{(165)} 4 1 1 ⊗ 1 + 12 1 a = 1 ∑ 3 σ a ⊗ σ a (165)
For two qubits, the swap operator satisfies
S = 1 2 ( 1 ⊗ 1 + ∑ a = 1 3 σ a ⊗ σ a ) . (166) S=\frac{1}{2}\left(1\otimes1+\sum_{a=1}^{3}\sigma_a\otimes\sigma_a\right). \tag*{(166)} S = 2 1 ( 1 ⊗ 1 + a = 1 ∑ 3 σ a ⊗ σ a ) . (166)
Hence
∑ i = 1 4 F i ⊗ F i = 1 4 1 ⊗ 1 + 1 12 ( 2 S − 1 ⊗ 1 ) = 1 6 ( 1 ⊗ 1 + S ) = 1 3 P s y m , (167) \sum_{i=1}^{4}F_i\otimes F_i=\frac{1}{4}1\otimes1+\frac{1}{12}(2S-1\otimes1)
=\frac{1}{6}(1\otimes1+S)=\frac{1}{3}P_{\mathrm{sym}}, \tag*{(167)} i = 1 ∑ 4 F i ⊗ F i = 4 1 1 ⊗ 1 + 12 1 ( 2 S − 1 ⊗ 1 ) = 6 1 ( 1 ⊗ 1 + S ) = 3 1 P sym , (167)
where P s y m = ( 1 ⊗ 1 + S ) / 2 P_{\mathrm{sym}}=(1\otimes1+S)/2 P sym = ( 1 ⊗ 1 + S ) /2 .
But
W 1 Q ( 1 / 2 , 1 / 2 ) = 2 3 > KR ( 1 / 2 , 1 / 2 ) = 0. (168) W^{Q}_{1}\left(1/2,1/2\right)=\frac{2}{3}>\operatorname{KR}\left(1/2,1/2\right)=0. \tag*{(168)} W 1 Q ( 1/2 , 1/2 ) = 3 2 > KR ( 1/2 , 1/2 ) = 0. (168)
Conclusions and outlook
In this work, we establish a framework of quantum optimal transport in spacetime using the Kirkwood-Dirac phase space and spatiotemporal state. We carefully discussed the duality and Kantorovich-Rubinstein metric.
Several problems are worth further investigation. First, a general definition of the spatiotemporal Wasserstein p p p -metric for p > 1 p>1 p > 1 has not yet been established and thus requires further study. Second, the continuous-variable case remains largely unexplored: the corresponding phase-space formalism can be developed by directly extending our finite-dimensional discussion to infinite dimensions, but a spatiotemporal-state formalism requires further work. This is because that the Choi and Jamiolkowski operators are not well-defined in this case, so a different approach needs to be developed. Third, one can also consider subsets of couplings that give rise to different correlations, which would yield a hierarchy of spatiotemporal optimal transport. For instance, consider the spatial correlation hierarchy: quantum discord, quantum entanglement, quantum steering, and Bell nonlocality (see [47 ] for application in entanglement theory and [69 ] for applications in quantum coherence). As the coupling set shrinks, the optimal transport value increases; extending this to general spatiotemporal correlations may shed new light on the study of different types of correlation in the case where only local data are available. Finally, the connection of our formalism with those discussed in gauge/gravity duality [29 ] and in quantum chaos [70 ] is also a crucial open direction. These will be addressed in our future work.
Acknowledgments
Z.J. is supported by by the National Natural Science Foundation of China (Grant No. 12605015), Start-up Grant of Central South University (Grant No. 502045031) and by the Frontier Interdisciplinary Direction “Quantum Technologies” Project of Central South University (Grant No. 506010805).
Some technical results from convex analysis and matrix analysis
In this appendix, we recall some technical results from convex analysis and matrix analysis that will be useful for the discussion in the main text.
Theorem A.1 (Carathéodory theorem [71 ] ). Let A ⊆ R d A \subseteq\mathbb{R}^{d} A ⊆ R d . Then every point in the convex hull conv ( A ) \operatorname{conv}(A) conv ( A ) can be written as a convex combination of at most d + 1 d+1 d + 1 points of A A A .
Theorem A.2 (Slater’s theorem for semidefinite programming [66 ] ). The following implications hold for every semidefinite programming ( E , C , D ) (\mathcal{E},\mathcal{C},\mathcal{D}) ( E , C , D ) . Let
P = { X ∈ Pos ( X ) : E ( X ) = D } and D = { Y ∈ Herm ( Y ) : E † ( Y ) ≥ C } \mathcal{P}=\{X\in\operatorname{Pos}(\mathcal{X}):\mathcal{E}(X)=D\}\quad\text{and}\quad\mathcal{D}=\{Y\in\operatorname{Herm}(\mathcal{Y}):\mathcal{E}^{\dagger}(Y)\geq C\} P = { X ∈ Pos ( X ) : E ( X ) = D } and D = { Y ∈ Herm ( Y ) : E † ( Y ) ≥ C }
denote the sets of primal and dual feasible operators, and α \alpha α and β \beta β denote the primal and dual optimal values respectively.
If P ≠ ∅ \mathcal{P} \ne\emptyset P = ∅ and there exists a Hermitian operator Y Y Y for which E † ( Y ) > C \mathcal{E}^{\dagger}(Y) > C E † ( Y ) > C , then the primal and dual optimal values coincide and there exists a primal feasible operator X ∈ P X \in\mathcal{P} X ∈ P for which ⟨ C , X ⟩ = α \langle C, X\rangle= \alpha ⟨ C , X ⟩ = α .
If D ≠ ∅ \mathcal{D} \ne\emptyset D = ∅ and there exists a positive definite operator X > 0 X > 0 X > 0 for which E ( X ) = D \mathcal{E}(X) = D E ( X ) = D , then the primal and dual optimal values coincide and there exists a dual feasible operator Y ∈ D Y \in\mathcal{D} Y ∈ D for which ⟨ D , Y ⟩ = β \langle D, Y\rangle= \beta ⟨ D , Y ⟩ = β .
Theorem A.3 (Sion’s minimax theorem [72 ] ). Let X \mathcal{X} X be a compact convex subset of a topological vector space, and let Y \mathcal{Y} Y be a convex subset of a topological vector space. If f : X × Y → R f : \mathcal{X} \times\mathcal{Y} \to\mathbb{R} f : X × Y → R is quasi-concave and upper semicontinuous in x x x for each fixed y ∈ Y y \in\mathcal{Y} y ∈ Y , and quasi-convex and lower semicontinuous in y y y for each fixed x ∈ X x \in\mathcal{X} x ∈ X , then
sup x ∈ X inf y ∈ Y f ( x , y ) = inf y ∈ Y sup x ∈ X f ( x , y ) . \sup_{x \in\mathcal{X}} \inf_{y \in\mathcal{Y}} f(x,y) = \inf_{y \in\mathcal{Y}} \sup_{x \in\mathcal{X}} f(x,y). x ∈ X sup y ∈ Y inf f ( x , y ) = y ∈ Y inf x ∈ X sup f ( x , y ) .
Proposition A.4 (Cyclicity of the partial transpose). Let H = H A ⊗ H B \mathcal{H} = \mathcal{H}_{A} \otimes\mathcal{H}_{B} H = H A ⊗ H B and let X , Y ∈ B ( H ) X,Y \in\mathcal{B}(\mathcal{H}) X , Y ∈ B ( H ) . Denote by T B T_{B} T B the partial transpose on subsystem B B B . Then
Tr ( X T B Y ) = Tr ( X Y T B ) . (169) \operatorname{Tr}(X^{T_{B}}Y) = \operatorname{Tr}(XY^{T_{B}}). \tag*{(169)} Tr ( X T B Y ) = Tr ( X Y T B ) . (169) Proof. Choose product bases { ∣ i ⟩ A } \{\lvert i\rangle_{A}\} {∣ i ⟩ A } and { ∣ α ⟩ B } \{\lvert\alpha\rangle_{B}\} {∣ α ⟩ B } and write X = ∑ i , j , α , β X i α , j β ∣ i ⟩ ⟨ j ∣ ⊗ ∣ α ⟩ ⟨ β ∣ X = \sum_{i,j,\alpha,\beta} X_{i\alpha,j\beta} \lvert i\rangle\langle j\rvert\otimes\lvert\alpha\rangle\langle\beta\rvert X = ∑ i , j , α , β X i α , j β ∣ i ⟩ ⟨ j ∣ ⊗ ∣ α ⟩ ⟨ β ∣ . By definition of the partial transpose, ( X T B ) i α , j β = X i β , j α (X^{T_{B}})_{i\alpha,j\beta} = X_{i\beta,j\alpha} ( X T B ) i α , j β = X i β , j α . Hence
Tr ( X T B Y ) = ∑ i , α , j , β ( X T B ) i α , j β Y j β , i α = ∑ i , α , j , β X i β , j α Y j β , i α , \operatorname{Tr}(X^{T_{B}}Y) = \sum_{i,\alpha,j,\beta} (X^{T_{B}})_{i\alpha,j\beta}Y_{j\beta,i\alpha} = \sum_{i,\alpha,j,\beta} X_{i\beta,j\alpha}Y_{j\beta,i\alpha}, Tr ( X T B Y ) = i , α , j , β ∑ ( X T B ) i α , j β Y j β , i α = i , α , j , β ∑ X i β , j α Y j β , i α , Tr ( X Y T B ) = ∑ i , α , j , β X i α , j β ( Y T B ) j β , i α = ∑ i , α , j , β X i α , j β Y j α , i β . \operatorname{Tr}(XY^{T_{B}}) = \sum_{i,\alpha,j,\beta} X_{i\alpha,j\beta}(Y^{T_{B}})_{j\beta,i\alpha} = \sum_{i,\alpha,j,\beta} X_{i\alpha,j\beta}Y_{j\alpha,i\beta}. Tr ( X Y T B ) = i , α , j , β ∑ X i α , j β ( Y T B ) j β , i α = i , α , j , β ∑ X i α , j β Y j α , i β . The two expressions coincide after relabelling the dummy indices α ↔ β \alpha\leftrightarrow\beta α ↔ β in the first line. ■
Proposition A.5. For operator X = ∑ i j X i j ∣ i ⟩ ⟨ j ∣ X = \sum_{ij} X_{ij}\lvert i\rangle\langle j\rvert X = ∑ ij X ij ∣ i ⟩ ⟨ j ∣ , let ∣ X ⟩ ⟩ = ∑ i j X i j ∣ i ⟩ ∣ j ⟩ \lvert X\rangle\rangle= \sum_{ij} X_{ij}\lvert i\rangle\lvert j\rangle ∣ X ⟩⟩ = ∑ ij X ij ∣ i ⟩ ∣ j ⟩ and ∣ Ω ⟩ = ∑ i ∣ i i ⟩ \lvert\Omega\rangle= \sum_i \lvert ii\rangle ∣ Ω ⟩ = ∑ i ∣ ii ⟩ , we have
∣ X ⟩ ⟩ = ( X ⊗ 1 ) ∣ Ω ⟩ = ( 1 ⊗ X T ) ∣ Ω ⟩ , \lvert X\rangle\rangle= (X \otimes1)\lvert\Omega\rangle= (1 \otimes X^{T})\lvert\Omega\rangle, ∣ X ⟩⟩ = ( X ⊗ 1 ) ∣ Ω ⟩ = ( 1 ⊗ X T ) ∣ Ω ⟩ , ⟨ ⟨ X ∣ = ⟨ ⟨ Ω ∣ ( X † ⊗ 1 ) = ⟨ ⟨ Ω ∣ ( 1 ⊗ X ∗ ) . (170) \langle\langle X\rvert= \langle\langle\Omega\rvert(X^{\dagger} \otimes1) = \langle\langle\Omega\rvert(1 \otimes X^{*}). \tag*{(170)} ⟨⟨ X ∣ = ⟨⟨ Ω ∣ ( X † ⊗ 1 ) = ⟨⟨ Ω ∣ ( 1 ⊗ X ∗ ) . (170)