Introduction
Gaussian states and operations are central to continuous-variable quantum information, in which information is encoded in the quadratures of light and other bosonic systems [1]. In quantum optics, this framework encompasses coherent and squeezed light, optical interferometers, and Gaussian channels modeling loss, amplification, and thermal noise [2]. Gaussian states are completely specified by their first moments and covariance matrices, while Gaussian channels are specified by their action on these moments, giving a compact description even when many modes are correlated [3, 4].
This framework brings together a wealth of fundamental questions: on one hand, how much entanglement is required to prepare a shared state or simulate a channel, and on the other hand, what rates of classical communication can a channel support under an energy constraint? Answering these questions would require in principle a formidable optimization over arbitrary quantum protocols, including non-Gaussian states and correlations across many copies or channel uses.
For a shared Gaussian state, a simple preparation starts with a pure Gaussian state and adds correlated random displacements. The least entanglement required by such an ensemble is the Gaussian entanglement of formation E F G E_{\mathrm{F}}^{\mathrm{G}} E F G : a minimum over pure covariance matrices lying below the target covariance [5]. The entanglement of formation E F E_{\mathrm{F}} E F instead minimizes the average entanglement over all pure-state decompositions [6], including possibly non-Gaussian ensembles, while the entanglement cost E C E_{\mathrm{C}} E C is the asymptotic entanglement consumed per copy under local operations and classical communication [7, 8]. The central questions are whether Gaussian ensembles are optimal for the entanglement of formation (Gaussian optimality) and whether preparing many copies jointly can or cannot reduce the cost per copy (additivity).
Gaussian optimality was established first for symmetric two-mode states [9]. The general two-mode problem subsequently prompted analytic proposals, bounds, and numerical methods [10–16]. Wilde established Gaussian optimality and additivity for states obtained by sending one mode of a two-mode squeezed vacuum through a quantum-limited attenuator or amplifier [17]. More recently, Adesso established both Gaussian optimality for every two-mode Gaussian state and additivity on tensor products of such states, using a many-pair entropy inequality valid for arbitrary correlated states [18]. Those results also extend to bisymmetric multimode states, whose entanglement can be concentrated onto a single pair of modes by local Gaussian unitaries [19].
In this work, we solve the problem in full generality for arbitrary bipartitions comprising any finite number of modes. For every finite-energy Gaussian state on such a multimode bipartition, we prove
E C ( ρ V ) = E F ( ρ V ) = E F G ( ρ V ) = G ( V ) . (1) E_{\mathrm{C}}(\rho_{V}) = E_{\mathrm{F}}(\rho_{V}) = E_{\mathrm{F}}^{\mathrm{G}}(\rho_{V}) = G(V). \tag*{(1)} E C ( ρ V ) = E F ( ρ V ) = E F G ( ρ V ) = G ( V ) . (1)
Here ρ V \rho_{V} ρ V denotes the centered bipartite Gaussian state with covariance V V V , and G ( V ) G(V) G ( V ) is the attained minimum over pure Gaussian preparations described above. We also prove strong superadditivity: the entanglement of formation and entanglement cost across the full bipartition are at least the sum of their values for reduced states on disjoint groups of Alice’s and Bob’s modes. Both quantities are additive on tensor products.
The result also provides entanglement bounds beyond Gaussian states. Every finite-energy state with covariance V V V has entanglement of formation and entanglement cost at least G ( V ) G(V) G ( V ) . Measured quadrature moments therefore give lower bounds on both quantities without any assumption of Gaussianity. We establish this fixed-covariance bound through a supporting entropy inequality, complementing the general Gaussian-extremality framework for strongly superadditive measures [20].
The theorem captured by Eq. (1) for states has two closely connected applications to channels. First, it determines the entanglement cost of every multimode Gaussian channel in the sequential setting of Wilde [17], where later inputs may depend on earlier outputs. Second, it determines the classical capacity of every multimode Gaussian channel under a mean energy constraint. This extends the recent single-mode result by Adesso, Lami, and Giovannetti [21 ] . For the capacity problem, the connection to state entanglement is provided by the identity between the minimum average output entropy at a fixed average input state and the entanglement of formation of the corresponding output–environment state in a dilation of the channel [22 ] .
A common bound on pure-state entanglement underlies all three results. Geodesic convexity supplies the bound within the Gaussian family, expressed through the expectation of a quadratic Hamiltonian. A relative-entropy inequality establishes it for arbitrary pure states. After averaging over any preparation ensemble, the bound depends only on the covariance matrix, connecting the geometry to optimal state preparation, channel simulation, and classical communication.
The paper is organized as follows. Section II proves Gaussian optimality, strong superadditivity, and covariance bounds on the entanglement of arbitrary states. Sections III and IV determine the entanglement cost and classical capacity of Gaussian channels. Section V develops analytic state and channel examples and introduces the numerical toolkit, while Sec. VI summarizes our results and their broader implications. The Appendices contain auxiliary lemmas and technical parts of the proofs.
Entanglement of formation and cost of multimode Gaussian states
Preliminaries
We consider bosonic continuous-variable systems containing finitely many modes. For a system of k k k modes, we collect the canonical quadrature operators in the vector R = ( x 1 , p 1 , … , x k , p k ) T R=(x_1,p_1,\ldots,x_k,p_k)^{\mathsf T} R = ( x 1 , p 1 , … , x k , p k ) T , where T \mathsf T T denotes matrix transposition and † \dagger † denotes the adjoint. Setting ℏ = 1 \hbar=1 ℏ = 1 , their commutation relations read
[ R j , R l ] = i ( Ω k ) j l , Ω k = ⨁ j = 1 k ( 0 1 − 1 0 ) . (2) [R_j,R_l]=\mathrm{i}(\Omega_k)_{jl},\qquad\Omega_k=\bigoplus_{j=1}^{k}\begin{pmatrix}0&1\\-1&0\end{pmatrix}. \tag*{(2)} [ R j , R l ] = i ( Ω k ) j l , Ω k = j = 1 ⨁ k ( 0 − 1 1 0 ) . (2)
where Ω k \Omega_k Ω k is the standard symplectic form on the 2 k 2k 2 k -dimensional phase space; we omit its subscript when the number of modes is clear.
For a state ρ \rho ρ with finite second moments, we write ⟨ O ⟩ = Tr ( ρ O ) \langle O\rangle=\operatorname{Tr}(\rho O) ⟨ O ⟩ = Tr ( ρO ) and define its first-moment vector d d d and covariance matrix V V V by
d = ⟨ R ⟩ , V j l = ⟨ { R j − d j , R l − d l } ⟩ . (3) d=\langle R\rangle,\qquad V_{jl}=\langle\{R_j-d_j,R_l-d_l\}\rangle. \tag*{(3)} d = ⟨ R ⟩ , V j l = ⟨{ R j − d j , R l − d l }⟩ . (3)
where { A , B } = A B + B A \{A,B\}=AB+BA { A , B } = A B + B A denotes the anticommutator. This normalization assigns to the vacuum state the identity covariance I 2 k I_{2k} I 2 k . Note that, relative to the convention in Ref. [18], here all quantum covariance matrices are doubled; the quadrature operators and first moments are unchanged.
A real symmetric matrix V V V is a physical covariance matrix if and only if it satisfies the bona fide condition [23 ] , which expresses the uncertainty principle:
V + i Ω k ≥ 0. (4) V+\mathrm{i}\Omega_k\ge0. \tag*{(4)} V + i Ω k ≥ 0. (4)
A physical covariance matrix Γ \Gamma Γ describes a pure Gaussian state precisely when
Γ Ω k Γ = Ω k . (5) \Gamma\Omega_k\Gamma=\Omega_k. \tag*{(5)} Γ Ω k Γ = Ω k . (5)
Equivalently, det Γ = 1 \det\Gamma=1 det Γ = 1 .
Every real symmetric positive-definite 2 k × 2 k 2k\times2k 2 k × 2 k matrix V V V admits a Williamson decomposition [23 ] [3],
V = Ξ D Ξ T , D = ⨁ j = 1 k ν j I 2 . (6) V=\Xi D\Xi^{\mathsf T},\qquad D=\bigoplus_{j=1}^{k}\nu_j I_2. \tag*{(6)} V = Ξ D Ξ T , D = j = 1 ⨁ k ν j I 2 . (6)
where Ξ \Xi Ξ is real and symplectic, meaning Ξ Ω k Ξ T = Ω k \Xi\Omega_k\Xi^{\mathsf T}=\Omega_k Ξ Ω k Ξ T = Ω k , and the numbers ν j > 0 \nu_j>0 ν j > 0 are the symplectic eigenvalues of V V V , equivalently the positive eigenvalues of i Ω k V \mathrm{i}\Omega_kV i Ω k V . For a physical covariance matrix, the uncertainty relation (4) requires ν j ≥ 1 \nu_j\ge1 ν j ≥ 1 . For a Gaussian state, the normal modes with quadratures Ξ − 1 R \Xi^{-1}R Ξ − 1 R are independent thermal modes up to displacement, with mean occupations ( ν j − 1 ) / 2 (\nu_j-1)/2 ( ν j − 1 ) /2 .
With natural logarithms, the von Neumann entropy is S ( ρ ) = − Tr ρ log ρ S(\rho)=-\operatorname{Tr}\rho\log\rho S ( ρ ) = − Tr ρ log ρ and the Gaussian entropy is S G ( V ) = ∑ j g ( ( ν j − 1 ) / 2 ) S_G(V)=\sum_j g((\nu_j-1)/2) S G ( V ) = ∑ j g (( ν j − 1 ) /2 ) , where
g ( u ) = ( u + 1 ) log ( u + 1 ) − u log u , g ( 0 ) = 0. (7) g(u)=(u+1)\log(u+1)-u\log u,\qquad g(0)=0. \tag*{(7)} g ( u ) = ( u + 1 ) log ( u + 1 ) − u log u , g ( 0 ) = 0. (7)
Let N = ∑ j a j † a j N=\sum_j a_j^\dagger a_j N = ∑ j a j † a j , with a j = ( x j + i p j ) / 2 a_j=(x_j+\mathrm{i}p_j)/\sqrt{2} a j = ( x j + i p j ) / 2 , be the total photon-number operator. Its expectation is Tr ρ N = Tr V / 4 + ∥ d ∥ 2 / 2 − k / 2 \operatorname{Tr}\rho N=\operatorname{Tr}V/4+\lVert d\rVert^2/2-k/2 Tr ρN = Tr V /4 + ∥ d ∥ 2 /2 − k /2 ; finite energy means that this expectation value is finite. For the finite number of modes considered here, this also guarantees finite local entropies, which are continuous in trace norm on sets with a common mean energy bound [24 , 25 ] . Throughout the paper, entanglement measures and rates are expressed in nats; division by log 2 \log2 log 2 gives bits or ebits. The displacement operator D ( ξ ) = e − i ξ T Ω R D(\xi)=e^{-\mathrm{i}\xi^{\mathsf T}\Omega R} D ( ξ ) = e − i ξ T Ω R obeys D ( ξ ) † R D ( ξ ) = R + ξ D(\xi)^\dagger R D(\xi)=R+\xi D ( ξ ) † R D ( ξ ) = R + ξ . A classical random shift ξ \xi ξ of ordinary covariance K ξ K_\xi K ξ therefore adds 2 K ξ 2K_\xi 2 K ξ to V V V .
The Gaussian entanglement of formation quantifies the least average entanglement required to prepare a bipartite Gaussian state from an ensemble of pure Gaussian states [5]. For a fixed bipartition A : B A:B A : B , it admits the covariance characterization
E F G ( ρ V ) = G ( V ) , E_F^G(\rho_V)=G(V), E F G ( ρ V ) = G ( V ) ,
where
G ( V ) = min Γ ≤ V Γ p u r e G a u s s i a n E ( Γ ) , E ( Γ ) = S G ( Γ A ) . (8) G(V)=\min_{\Gamma\le V\atop\Gamma\ \mathrm{pure\ Gaussian}}E(\Gamma),\qquad E(\Gamma)=S_G(\Gamma_A). \tag*{(8)} G ( V ) = Γ pure Gaussian Γ ≤ V min E ( Γ ) , E ( Γ ) = S G ( Γ A ) . (8)
Here the entanglement of a pure Gaussian state is the entropy of its reduced state on A A A , with subscripts on states denoting reductions by partial trace and those on covariance matrices denoting the corresponding principal blocks.
The pure Gaussian covariance matrices satisfying Γ ≤ V \Gamma\le V Γ ≤ V form a compact set, so the minimum is attained by a covariance Γ ∗ \Gamma_* Γ ∗ . The corresponding pure Gaussian state provides an explicit preparation of ρ V \rho_V ρ V : applying displacements drawn from a centered Gaussian distribution with ordinary covariance $(V-\Gamma_*)/2 produces the target state. Since these displacements are local unitaries, every member of the ensemble has the same entanglement E ( Γ ∗ ) E(\Gamma_*) E ( Γ ∗ ) , while a further fixed local displacement sets any desired first moments without changing G G G .
For a general bipartite state, the entanglement of formation is the least average pure-state entanglement over all ensemble decompositions [6, 8],
E F ( ρ ) = inf μ : ∫ ∣ ψ ⟩ ⟨ ψ ∣ d μ = ρ ∫ S ( ψ A ) d μ ( ψ ) . (9) E_{\mathrm{F}}(\rho)=\inf_{\mu:\int\lvert\psi\rangle\langle\psi\rvert d\mu=\rho}\int S(\psi_A)d\mu(\psi). \tag*{(9)} E F ( ρ ) = μ : ∫ ∣ ψ ⟩ ⟨ ψ ∣ d μ = ρ inf ∫ S ( ψ A ) d μ ( ψ ) . (9)
Here ψ A = Tr B ∣ ψ ⟩ ⟨ ψ ∣ \psi_A=\operatorname{Tr}_B\lvert\psi\rangle\langle\psi\rvert ψ A = Tr B ∣ ψ ⟩ ⟨ ψ ∣ , and the infimum ranges over probability measures on pure states whose average is ρ \rho ρ , allowing both discrete and continuous ensembles [26 ] . When the bipartition needs to be explicit, we write E F A : B ( ρ ) E_{\mathrm{F}}^{A:B}(\rho) E F A : B ( ρ ) and G A : B ( V ) G^{A:B}(V) G A : B ( V ) .
The entanglement cost E C E_{\mathrm{C}} E C is the least asymptotic entanglement required per copy to prepare a state by local operations and classical communication (LOCC), allowing collective protocols with vanishing trace-distance error. For the finite-energy states considered here, it equals the regularized entanglement of formation [7, 8],
E C ( ρ ) = lim n → ∞ E F ( ρ ⊗ n ) n . (10) E_{\mathrm{C}}(\rho)=\lim_{n\to\infty}\frac{E_{\mathrm{F}}(\rho^{\otimes n})}{n}. \tag*{(10)} E C ( ρ ) = n → ∞ lim n E F ( ρ ⊗ n ) . (10)
Main result
Theorem 1 (Optimality of multimode Gaussian entanglement). (a) For every finite-energy Gaussian state ρ V \rho_V ρ V on an arbitrary bipartition A : B A:B A : B ,
E C ( ρ V ) = E F ( ρ V ) = E F G ( ρ V ) = G ( V ) . E_{\mathrm{C}}(\rho_V)=E_{\mathrm{F}}(\rho_V)=E_{\mathrm{F}}^{\mathrm{G}}(\rho_V)=G(V). E C ( ρ V ) = E F ( ρ V ) = E F G ( ρ V ) = G ( V ) . (b) For any disjoint mode groups A j ⊆ A A_j\subseteq A A j ⊆ A and B j ⊆ B B_j\subseteq B B j ⊆ B , the common value satisfies strong superadditivity,
G A : B ( V ) ≥ ∑ j G A j : B j ( V A j B j ) . (11) G^{A:B}(V)\ge\sum_j G^{A_j:B_j}(V_{A_jB_j}). \tag*{(11)} G A : B ( V ) ≥ j ∑ G A j : B j ( V A j B j ) . (11) and it is additive on finite tensor products of bipartite Gaussian states.
(c) Gaussian states also minimize both entanglement of formation and entanglement cost at fixed covariance: every finite-energy state ρ \rho ρ with covariance V V V satisfies
E F ( ρ ) ≥ G ( V ) , E C ( ρ ) ≥ G ( V ) . (12) E_{\mathrm{F}}(\rho)\ge G(V),\qquad E_{\mathrm{C}}(\rho)\ge G(V). \tag*{(12)} E F ( ρ ) ≥ G ( V ) , E C ( ρ ) ≥ G ( V ) . (12)
Proof of Theorem 1. The following four subsections develop the proof, beginning with strong superadditivity and additivity for the covariance minimum G G G in part (b). Its convexity then provides a quadratic lower bound on pure Gaussian entanglement, which we extend to arbitrary pure states through a relative-entropy inequality. Averaging over ensembles gives Gaussian optimality and the fixed-covariance bound for entanglement of formation, while regularization gives the corresponding statements for entanglement cost in parts (a) and (c).
Geometry of pure Gaussian preparations
To compare pure Gaussian preparations, we use paths that preserve both purity and the covariance constraint. For positive-definite matrices X X X , Y Y Y and 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , the weighted geometric mean is
X # t Y = X 1 / 2 ( X − 1 / 2 Y X − 1 / 2 ) t X 1 / 2 . (13) X\#_tY=X^{1/2}(X^{-1/2}YX^{-1/2})^tX^{1/2}. \tag*{(13)} X # t Y = X 1/2 ( X − 1/2 Y X − 1/2 ) t X 1/2 . (13)
We abbreviate the midpoint as X # Y : = X # 1 / 2 Y X\#Y:=X\#_{1/2}Y X # Y := X # 1/2 Y . The mean is covariant under congruences, jointly monotone in its arguments, and bounded above by the arithmetic mean [27 , 28 ] . Pure covariances are positive symplectic matrices: their positive powers and symplectic congruences preserve purity. Consequently Γ 0 # t Γ 1 \Gamma_0\#_t\Gamma_1 Γ 0 # t Γ 1 is pure whenever its endpoints are pure. Monotonicity also preserves the constraint Γ ≤ V \Gamma\le V Γ ≤ V .
The Williamson decomposition (6) supplies a pure Gaussian covariance Ξ Ξ T ≤ V \Xi\Xi^{\mathsf{T}}\le V Ξ Ξ T ≤ V , since Ξ \Xi Ξ is symplectic and D ≥ I 2 k D\ge I_{2k} D ≥ I 2 k . If L = λ max ( V ) L=\lambda_{\max}(V) L = λ m a x ( V ) , every feasible pure covariance obeys
L − 1 I ≤ Γ ≤ L I , Γ − 1 ≤ Ω Γ Ω T . (14) L^{-1}I\le\Gamma\le LI,\qquad\Gamma^{-1}\le\Omega\Gamma\Omega^{\mathsf{T}}. \tag*{(14)} L − 1 I ≤ Γ ≤ L I , Γ − 1 ≤ ΩΓ Ω T . (14)
The purity equation and upper constraint are closed on this compact positive interval, proving attainment of G G G . The same argument gives the one-sided continuity
G ( V + D ) ⟶ G ( V ) as D ≥ 0 , ∥ D ∥ ⟶ 0. (15) G(V+D)\longrightarrow G(V)\quad\text{as }D\ge0,\ \lVert D\rVert\longrightarrow0. \tag*{(15)} G ( V + D ) ⟶ G ( V ) as D ≥ 0 , ∥ D ∥ ⟶ 0. (15)
Indeed, the minimizing pure covariances have convergent subsequences with feasible limits below V V V , while G ( V + D ) ≤ G ( V ) G(V+D)\le G(V) G ( V + D ) ≤ G ( V ) by monotonicity.
To prove convexity along these paths, we express E E E through the relative spectrum of Γ \Gamma Γ and a locally transformed copy. Let
Π = I A ⊕ ( − I B ) , R Γ = Γ − 1 / 2 Π Γ Π Γ − 1 / 2 . \Pi=\mathbb{I}_A\oplus(-\mathbb{I}_B),\qquad R_\Gamma=\Gamma^{-1/2}\Pi\Gamma\Pi\Gamma^{-1/2}. Π = I A ⊕ ( − I B ) , R Γ = Γ − 1/2 ΠΓΠ Γ − 1/2 .
The matrix Π \Pi Π implements a local phase rotation by π \pi π on Bob’s modes, changing the signs of both quadratures. Local symplectic transformations commute with Π \Pi Π and preserve the relative spectrum of ( Γ , Π Γ Π ) (\Gamma,\Pi\Gamma\Pi) ( Γ , ΠΓΠ ) . These transformations bring any pure Gaussian state to Gaussian Schmidt form: a product of two-mode squeezed pairs and local vacua [29 ] . A pair with squeezing r r r contributes e 4 r , e 4 r , e − 4 r , e − 4 r e^{4r},e^{4r},e^{-4r},e^{-4r} e 4 r , e 4 r , e − 4 r , e − 4 r to the relative spectrum, whereas each local pure mode contributes two unit eigenvalues. Therefore
E ( Γ ) = 1 4 ∑ i F ( log λ i ( R Γ ) ) , F ( x ) = g ( sinh 2 ( x / 4 ) ) . (16) E(\Gamma)=\frac{1}{4}\sum_i F(\log\lambda_i(R_\Gamma)),\qquad F(x)=g(\sinh^2(x/4)). \tag*{(16)} E ( Γ ) = 4 1 i ∑ F ( log λ i ( R Γ )) , F ( x ) = g ( sinh 2 ( x /4 )) . (16)
The scalar function F F F is even, convex, and continuously differentiable, including at zero. Appendix A proves the required spectral comparison: the ordered log-eigenvalues of a relative matrix under geometric interpolation are majorized by the interpolated endpoint vectors. Convexity of F F F then gives
E ( Γ 0 # t Γ 1 ) ≤ ( 1 − t ) E ( Γ 0 ) + t E ( Γ 1 ) . (17) E(\Gamma_0\#_t\Gamma_1)\le(1-t)E(\Gamma_0)+tE(\Gamma_1). \tag*{(17)} E ( Γ 0 # t Γ 1 ) ≤ ( 1 − t ) E ( Γ 0 ) + tE ( Γ 1 ) . (17)
Thus E E E is convex along these geometric paths of pure Gaussian covariances, a property we refer to as geodesic convexity .
Strong superadditivity follows by using a local symmetry to remove correlations between groups without increasing entanglement. Group the coordinates as ( A 1 B 1 ) , ( A 2 B 2 ) (A_1B_1),(A_2B_2) ( A 1 B 1 ) , ( A 2 B 2 ) , write V i = V A i B i V_i=V_{A_iB_i} V i = V A i B i , and let Q Q Q change the signs of all quadratures in the first group. This transformation is local across the total Alice–Bob cut. For any pure Γ ≤ V \Gamma\le V Γ ≤ V , the midpoint
Γ ˉ = Γ # 1 / 2 ( Q Γ Q ) \bar{\Gamma}=\Gamma\#_{1/2}(Q\Gamma Q) Γ ˉ = Γ # 1/2 ( Q Γ Q )
is pure, has no greater entanglement, and is invariant under Q Q Q . It therefore splits into pure blocks Γ ˉ = γ 1 ⊕ γ 2 \bar{\Gamma}=\gamma_1\oplus\gamma_2 Γ ˉ = γ 1 ⊕ γ 2 . The arithmetic–geometric inequality gives
Γ ˉ ≤ 1 2 ( Γ + Q Γ Q ) ≤ 1 2 ( V + Q V Q ) = V 1 ⊕ V 2 . \bar{\Gamma}\le\frac{1}{2}(\Gamma+Q\Gamma Q)\le\frac{1}{2}(V+QVQ)=V_1\oplus V_2. Γ ˉ ≤ 2 1 ( Γ + Q Γ Q ) ≤ 2 1 ( V + Q V Q ) = V 1 ⊕ V 2 .
Each block is feasible for its marginal, so
E ( Γ ) ≥ E ( γ 1 ) + E ( γ 2 ) ≥ G ( V 1 ) + G ( V 2 ) . E(\Gamma)\ge E(\gamma_1)+E(\gamma_2)\ge G(V_1)+G(V_2). E ( Γ ) ≥ E ( γ 1 ) + E ( γ 2 ) ≥ G ( V 1 ) + G ( V 2 ) .
Minimizing proves Eq. (11) for two groups, and iteration handles any finite partition of the modes. If the chosen groups do not exhaust the system, we include the remaining modes as one further group and discard its nonnegative contribution. For product covariances, product minimizers give the matching upper bound:
G ( V 1 ⊕ V 2 ) = G ( V 1 ) + G ( V 2 ) . (18) G(V_1\oplus V_2)=G(V_1)+G(V_2). \tag*{(18)} G ( V 1 ⊕ V 2 ) = G ( V 1 ) + G ( V 2 ) . (18)
The same interpolation makes G G G convex under ordinary mixing of covariance matrices. If Γ j \Gamma_j Γ j minimizes G ( V j ) G(V_j) G ( V j ) , then
Γ 0 # t Γ 1 ≤ V 0 # t V 1 ≤ ( 1 − t ) V 0 + t V 1 . \Gamma_0\#_t\Gamma_1\le V_0\#_tV_1\le(1-t)V_0+tV_1. Γ 0 # t Γ 1 ≤ V 0 # t V 1 ≤ ( 1 − t ) V 0 + t V 1 .
Using Eq. (17) , we obtain
G ( ( 1 − t ) V 0 + t V 1 ) ≤ ( 1 − t ) G ( V 0 ) + t G ( V 1 ) . (19) G((1-t)V_0+tV_1)\le(1-t)G(V_0)+tG(V_1). \tag*{(19)} G (( 1 − t ) V 0 + t V 1 ) ≤ ( 1 − t ) G ( V 0 ) + tG ( V 1 ) . (19)
A supporting quadratic
At a strictly physical covariance V + i Ω > 0 V+i\Omega>0 V + i Ω > 0 , convexity of G G G gives a supporting hyperplane: a global affine lower bound that is tight at V V V . Thus there is a real symmetric matrix B V B_V B V such that
G ( U ) ≥ G ( V ) + Tr [ B V ( U − V ) ] . G(U)\ge G(V)+\operatorname{Tr}[B_V(U-V)]. G ( U ) ≥ G ( V ) + Tr [ B V ( U − V )] .
Adding a positive covariance can only enlarge the feasible set, so B V ≤ 0 B_V\le0 B V ≤ 0 . Set W = − 4 B V ≥ 0 W=-4B_V\ge0 W = − 4 B V ≥ 0 . The support extends to boundary covariances by applying it to ( 1 − ϵ ) U + ϵ V (1-\epsilon)U+\epsilon V ( 1 − ϵ ) U + ϵ V and using Eq. (19) . For an optimal pure Γ ∗ \Gamma_* Γ ∗ , we have G ( Γ ∗ ) = E ( Γ ∗ ) = G ( V ) G(\Gamma_*)=E(\Gamma_*)=G(V) G ( Γ ∗ ) = E ( Γ ∗ ) = G ( V ) , whence
0 ≤ 1 4 Tr [ W ( V − Γ ∗ ) ] ≥ 0. 0\le\frac{1}{4}\operatorname{Tr}[W(V-\Gamma_*)]\ge0. 0 ≤ 4 1 Tr [ W ( V − Γ ∗ )] ≥ 0.
It follows that
Tr [ W ( V − Γ ∗ ) ] = 0 , (20) \operatorname{Tr}[W(V-\Gamma_*)]=0, \tag*{(20)} Tr [ W ( V − Γ ∗ )] = 0 , (20)
E ( Γ ) + 1 4 Tr ( W Γ ) ≥ E ( Γ ∗ ) + 1 4 Tr ( W Γ ∗ ) (21) E(\Gamma)+\frac{1}{4}\operatorname{Tr}(W\Gamma)\ge E(\Gamma_*)+\frac{1}{4}\operatorname{Tr}(W\Gamma_*) \tag*{(21)} E ( Γ ) + 4 1 Tr ( W Γ ) ≥ E ( Γ ∗ ) + 4 1 Tr ( W Γ ∗ ) (21)
for every pure Gaussian Γ \Gamma Γ .
The physical quadratic H W = R T W R / 2 H_W=R^{\mathsf T}WR/2 H W = R T W R /2 has centered expectation Tr ( W V ) / 4 \operatorname{Tr}(WV)/4 Tr ( W V ) /4 . A covariance subscript on an expectation denotes the centered Gaussian state with that covariance. Equation (21) therefore says that Γ ∗ \Gamma_* Γ ∗ minimizes the sum of entanglement and quadratic energy among pure Gaussian states. Both terms are geodesically convex: the entropy by Eq. (17) , and the positive quadratic by arithmetic–geometric order. Equation (16) also shows that the functional is continuously differentiable on the pure manifold. Consequently, every stationary point is a global pure Gaussian minimum.
Gaussian optimality follows if this supporting inequality holds for every finite-energy pure state:
S ( ψ A ) + ⟨ H W ⟩ ψ ≥ E ( Γ ∗ ) + ⟨ H W ⟩ Γ ∗ . (22) S(\psi_A)+\langle H_W\rangle_\psi\ge E(\Gamma_*)+\langle H_W\rangle_{\Gamma_*}. \tag*{(22)} S ( ψ A ) + ⟨ H W ⟩ ψ ≥ E ( Γ ∗ ) + ⟨ H W ⟩ Γ ∗ . (22)
Once established, this inequality can be averaged over any preparation ensemble. Complementarity (20) then cancels the quadratic term at the target covariance.
From a Gaussian support to an unrestricted entropy bound
The required entropy bound follows from a comparison of input and output quantum relative entropies, defined by D ( ρ ∥ σ ) = Tr ( ρ log ρ − log σ ) D(\rho\Vert\sigma)=\operatorname{Tr}(\rho\log\rho-\log\sigma) D ( ρ ∥ σ ) = Tr ( ρ log ρ − log σ ) . For a centered Gaussian state σ \sigma σ that is faithful, meaning it has no zero eigenvalues, we write σ = e − H σ / Tr e − H σ \sigma=e^{-H_\sigma}/\operatorname{Tr}e^{-H_\sigma} σ = e − H σ / Tr e − H σ and choose the quadratic Hamiltonian to have ground energy zero. Using the adjoint Φ ∗ \Phi^* Φ ∗ of a channel Φ \Phi Φ defined by Tr [ Φ ( ρ ) Z ] = Tr [ ρ Φ ∗ ( Z ) ] \operatorname{Tr}[\Phi(\rho)Z]=\operatorname{Tr}[\rho\Phi^*(Z)] Tr [ Φ ( ρ ) Z ] = Tr [ ρ Φ ∗ ( Z )] , the following proposition provides the comparison, with its proof given in Appendix C.
Proposition 2 (Joint relative-entropy inequality). Let Φ \Phi Φ be a Gaussian unitary followed by a partial trace. Let σ \sigma σ and τ = Φ ( σ ) \tau=\Phi(\sigma) τ = Φ ( σ ) be faithful centered Gaussian states. If λ > 0 \lambda>0 λ > 0 and the real coefficient matrix of
B = λ H σ − Φ ∗ H τ (23) B=\lambda H_\sigma-\Phi^*H_\tau \tag*{(23)} B = λ H σ − Φ ∗ H τ (23)
in symmetric ordering is positive definite, then every finite-energy state obeys
D ( Φ ( ρ ) ∥ τ ) ≤ λ D ( ρ ∥ σ ) . (24) D(\Phi(\rho)\Vert\tau)\le\lambda D(\rho\Vert\sigma). \tag*{(24)} D ( Φ ( ρ ) ∥ τ ) ≤ λ D ( ρ ∥ σ ) . (24)
The comparison becomes an entropy minimization principle through the Gibbs identities. For
F λ ( ρ ) = S ( Φ ( ρ ) ) − λ S ( ρ ) + Tr ( B ρ ) , F_\lambda(\rho)=S(\Phi(\rho))-\lambda S(\rho)+\operatorname{Tr}(B\rho), F λ ( ρ ) = S ( Φ ( ρ )) − λ S ( ρ ) + Tr ( B ρ ) ,
we have
F λ ( ρ ) − F λ ( σ ) = λ D ( ρ ∥ σ ) − D ( Φ ( ρ ) ∥ τ ) . (25) F_\lambda(\rho)-F_\lambda(\sigma)=\lambda D(\rho\Vert\sigma)-D(\Phi(\rho)\Vert\tau). \tag*{(25)} F λ ( ρ ) − F λ ( σ ) = λ D ( ρ ∥ σ ) − D ( Φ ( ρ ) ∥ τ ) . (25)
Proposition 2 makes the right-hand side nonnegative, so σ \sigma σ minimizes F λ F_\lambda F λ over all finite-energy states. The proof of Proposition 2 shows that the limiting output of a regularized minimizing sequence for F λ F_\lambda F λ is the reference Gaussian state τ = Φ ( σ ) \tau=\Phi(\sigma) τ = Φ ( σ ) . The strictly positive quadratic term bounds the energy of the sequence and supplies a strict remainder in the comparison of relative-entropy second variations. Gaussian covariance identities transfer this comparison between input and output; stationarity under Gaussian thermal relaxation then enforces the thermal balance equations characterizing τ \tau τ . At this limit, the output relative entropy in Eq. (25) vanishes, leaving the nonnegative input term λ D ( ρ ∥ σ ) \lambda D(\rho\Vert\sigma) λ D ( ρ ∥ σ ) . The proof uses a weighted interpolation estimate with a strict remainder, established in Appendix D from [30 ] .
To apply this comparison to the supporting quadratic, we first separate the local pure factors and then approximate the remaining paired state by faithful Gaussian states. Write Γ ∗ \Gamma_* Γ ∗ in Gaussian Schmidt form, with n p n_p n p entangled mode pairs and local pure modes. If n p = 0 n_p=0 n p = 0 , Gaussian stationarity makes the state with covariance Γ ∗ \Gamma_* Γ ∗ a ground state of H W H_W H W , so Eq. (22) follows from the nonnegativity of entropy. For n p > 0 n_p>0 n p > 0 , completing a positive quadratic square gives
H W = H + H r e m , H r e m ≥ 0 , ⟨ H r e m ⟩ Γ ∗ = 0. (26) H_W=H+H_{\mathrm{rem}},\qquad H_{\mathrm{rem}}\ge0,\qquad\langle H_{\mathrm{rem}}\rangle_{\Gamma_*}=0. \tag*{(26)} H W = H + H rem , H rem ≥ 0 , ⟨ H rem ⟩ Γ ∗ = 0. (26)
Here H H H acts on the paired modes and has a positive-semidefinite real quadratic coefficient; its additive constant is fixed by the displayed operator identity. Appendix B gives this reduction in annihilation coordinates and verifies the stationarity of the paired seed.
The paired part of the optimal seed is a pure Gaussian state ω c \omega_c ω c with entanglement E ∗ : = S ( ( ω c ) A ) = E ( Γ ∗ ) E_*:=S((\omega_c)_A)=E(\Gamma_*) E ∗ := S (( ω c ) A ) = E ( Γ ∗ ) , since the discarded factors are locally pure. Choose annihilation operators c 1 , … , c 2 n p c_1,\ldots,c_{2n_p} c 1 , … , c 2 n p , generally mixing Alice’s and Bob’s modes, for which ω c \omega_c ω c is the joint vacuum, and define N C = ∑ j = 1 2 n p c j † c j N_C=\sum_{j=1}^{2n_p}c_j^\dagger c_j N C = ∑ j = 1 2 n p c j † c j . The stationary-point criterion above makes ω c \omega_c ω c a global pure Gaussian minimizer of entropy plus H H H . Displaced pure Gaussian states obey the same bound because the quadratic coefficient of H H H is positive semidefinite. Since every mixed Gaussian state admits a displacement ensemble of pure Gaussian states, concavity of the reduced entropy gives, for every Gaussian state ρ \rho ρ on the paired modes,
S ( ρ A ) + ⟨ H ⟩ ρ ≥ C ∗ , C ∗ = E ∗ + ⟨ H ⟩ ω c . (27) S(\rho_A)+\langle H\rangle_\rho\ge C_*,\qquad C_*=E_*+\langle H\rangle_{\omega_c}. \tag*{(27)} S ( ρ A ) + ⟨ H ⟩ ρ ≥ C ∗ , C ∗ = E ∗ + ⟨ H ⟩ ω c . (27)
Two small perturbations provide the strict positivity and faithfulness required by Proposition 2. For fixed η > 0 \eta>0 η > 0 , adding η N C \eta N_C η N C makes the quadratic strictly positive and selects ω c \omega_c ω c uniquely, since equality in the preceding bound then requires ⟨ N C ⟩ = 0 \langle N_C\rangle=0 ⟨ N C ⟩ = 0 . We then introduce a small entropy penalty term and minimize over Gaussian states:
F λ , η ( ρ ) = S ( ρ A ) + ⟨ H + η N C ⟩ ρ − λ S ( ρ ) . (28) F_{\lambda,\eta}(\rho)=S(\rho_A)+\langle H+\eta N_C\rangle_\rho-\lambda S(\rho). \tag*{(28)} F λ , η ( ρ ) = S ( ρ A ) + ⟨ H + η N C ⟩ ρ − λ S ( ρ ) . (28)
The linear energy penalty dominates the entropy bound
S ( ρ ) ≤ 2 n p g ( ⟨ N C ⟩ 2 n p ) . S(\rho)\le2n_p g\left(\frac{\langle N_C\rangle}{2n_p}\right). S ( ρ ) ≤ 2 n p g ( 2 n p ⟨ N C ⟩ ) .
so a minimum exists in a bounded covariance set, uniformly for small λ > 0 \lambda>0 λ > 0 at fixed η \eta η . Its value is at most C ∗ C_* C ∗ , by evaluation at ω c \omega_c ω c . Equation Eq. (27) then implies that the minimum tends to C ∗ C_* C ∗ and η ⟨ N C ⟩ \eta\langle N_C\rangle η ⟨ N C ⟩ tends to zero as λ ↓ 0 \lambda\downarrow0 λ ↓ 0 . Hence the minimizing covariances converge to that of ω c \omega_c ω c .
The entropy term makes the minimizer faithful. Its marginal on Alice’s subsystem is already faithful for small λ \lambda λ , since every limiting Schmidt occupation is positive. If a global symplectic eigenvalue were 1 1 1 , increasing the corresponding thermal occupation from zero to u u u would change the quadratic energy and the marginal entropy by O ( u ) O(u) O ( u ) , while − λ g ( u ) = − λ u log ( 1 / u ) + O ( u ) -\lambda g(u)=-\lambda u\log(1/u)+O(u) − λ g ( u ) = − λ u log ( 1/ u ) + O ( u ) would decrease the functional. Thus the full minimizer σ \sigma σ is faithful as well; positivity of the quadratic sets its first moments to zero.
At this interior Gaussian minimum, covariance differentiation gives
H + η N c = λ H σ − Φ ∗ H τ + constant , τ = Φ ( σ ) , (29) H+\eta N_c=\lambda H_\sigma-\Phi^*H_\tau+\text{constant},\qquad\tau=\Phi(\sigma), \tag*{(29)} H + η N c = λ H σ − Φ ∗ H τ + constant , τ = Φ ( σ ) , (29)
where Φ \Phi Φ traces out Bob’s paired modes. Since the coefficient on the left is strictly positive, Proposition 2 and Eq. (25) make σ \sigma σ a minimum over all finite-energy states. For a pure test state the global entropy term vanishes, so letting first λ ↓ 0 \lambda\downarrow0 λ ↓ 0 at fixed η \eta η , and then η ↓ 0 \eta\downarrow0 η ↓ 0 , yields the paired-mode bound
S ( ψ A ) + ⟨ H ⟩ ψ ≥ C ∗ . (30) S(\psi_A)+\langle H\rangle_\psi\ge C_*. \tag*{(30)} S ( ψ A ) + ⟨ H ⟩ ψ ≥ C ∗ . (30)
Averaging this bound over pure-state ensembles gives E F ( ρ p ) + Tr ( H ρ p ) ≥ C ∗ E_F(\rho_p)+\operatorname{Tr}(H\rho_p)\ge C_* E F ( ρ p ) + Tr ( H ρ p ) ≥ C ∗ for every finite-energy state on the paired modes, since finite mean photon number bounds all second moments and makes the quadratic term integrable. For a pure state ∣ ψ ⟩ |\psi\rangle ∣ ψ ⟩ of all modes, let ρ p \rho_p ρ p be its reduction to the paired modes. Local tracing of the unpaired modes gives S ( ψ A ) ≥ E F ( ρ p ) S(\psi_A)\ge E_F(\rho_p) S ( ψ A ) ≥ E F ( ρ p ) by LOCC monotonicity. Restoring the non-negative H r e m H_{\mathrm{rem}} H rem in Eq. (26) proves Eq. (22) , including the local pure modes.
The geometric interpolation and the supporting entropy bound are illustrated in Figure 1 .
Figure 1. Geometric structure of Gaussian optimality. (a) Purity-preserving interpolation: in the two-mode covariance slice U ( u , v ) = Ξ diag ( u , v , v , u ) Ξ T U(u,v)=\Xi\operatorname{diag}(u,v,v,u)\Xi^{\mathsf{T}} U ( u , v ) = Ξ diag ( u , v , v , u ) Ξ T , with fixed symplectic Ξ \Xi Ξ , the shaded region consists of physical covariances U ≤ V U\le V U ≤ V . Its curved boundary u v = 1 uv=1 uv = 1 consists of feasible pure covariances Γ ≤ V \Gamma\le V Γ ≤ V , with the geometric path from Γ 0 \Gamma_0 Γ 0 to Γ 1 \Gamma_1 Γ 1 highlighted. The arithmetic and geometric midpoints remain below the target covariance V V V , while the geometric midpoint also remains pure; entanglement is convex along the geometric path, as in Eq. (17) . (b) The supporting entropy bound: for a strictly physical target covariance, the supporting inequality relates pure-state entanglement S ( ψ A ) S(\psi_A) S ( ψ A ) to quadratic energy ⟨ H W ⟩ ψ \langle H_W\rangle_\psi ⟨ H W ⟩ ψ , with equality at the optimal Gaussian seed Γ ∗ \Gamma_* Γ ∗ . The relative-entropy argument based on Proposition 2 establishes Eq. (22) for arbitrary finite-energy pure states, with the shaded half-plane indicating the bound. For any ensemble of a centered state with covariance V V V , complementarity fixes the average quadratic energy at Tr ( W V ) / 4 = Tr ( W Γ ∗ ) / 4 \operatorname{Tr}(WV)/4=\operatorname{Tr}(W\Gamma_*)/4 Tr ( W V ) /4 = Tr ( W Γ ∗ ) /4 . Averaging the entropy inequality therefore gives the lower bound G ( V ) G(V) G ( V ) on the average entanglement.
Completion of the proof
Since local displacements leave the entanglement measures unchanged, we may take ρ \rho ρ to be centered. For a strictly physical covariance V V V , averaging Eq. (22) over any pure-state ensemble replaces its quadratic term by Tr ( W V ) / 4 \operatorname{Tr}(WV)/4 Tr ( W V ) /4 , which cancels by complementarity and gives E F ( ρ ) ≥ G ( V ) E_F(\rho)\ge G(V) E F ( ρ ) ≥ G ( V ) .
For a boundary covariance, independent local Gaussian shifts of ordinary covariance ϵ I / 2 \epsilon I/2 ϵ I /2 give a state ρ ϵ \rho_\epsilon ρ ϵ with quantum covariance V + ϵ I V+\epsilon I V + ϵ I . LOCC monotonicity and Eq. (15) yield
E F ( ρ ) ≥ E F ( ρ ϵ ) ≥ G ( V + ϵ I ) ⟶ G ( V ) . E_F(\rho)\ge E_F(\rho_\epsilon)\ge G(V+\epsilon I)\longrightarrow G(V). E F ( ρ ) ≥ E F ( ρ ϵ ) ≥ G ( V + ϵ I ) ⟶ G ( V ) .
The entanglement of formation bound in part (c) therefore holds for every physical covariance, and the Gaussian displacement ensemble attains it for ρ V \rho_V ρ V .
For Gaussian states, additivity (18) removes the regularization in Eq. (10) , proving part (a). The common value consequently inherits the strong superadditivity and additivity established for G G G , proving part (b). For an arbitrary finite-energy state ρ \rho ρ with covariance V V V , applying the entanglement of formation bound to each tensor power gives E F ( ρ ⊗ k ) ≥ G ( V ⊕ k ) = k G ( V ) E_F(\rho^{\otimes k})\ge G(V^{\oplus k})=kG(V) E F ( ρ ⊗ k ) ≥ G ( V ⊕ k ) = k G ( V ) ; dividing by k k k and using Eq. (10) yields E C ( ρ ) ≥ G ( V ) E_C(\rho)\ge G(V) E C ( ρ ) ≥ G ( V ) . This proves part (c) and completes the proof of Theorem 1. □
Interestingly, the lower bound in Eq. (12) is also attained by a concrete family of non-Gaussian states: the Gaussian displacement distribution in an optimal preparation can be replaced by any non-Gaussian distribution with the same first and second moments, preserving both the entanglement of formation and the entanglement cost.
Corollary 3 (Non-Gaussian states with exact entanglement cost). Let Γ ∗ \Gamma_* Γ ∗ attain G ( V ) G(V) G ( V ) , let ∣ ψ Γ ∗ ⟩ |\psi_{\Gamma_*}\rangle ∣ ψ Γ ∗ ⟩ be its centered pure Gaussian state, and let P P P be any centered probability distribution on real displacements with ordinary covariance ( V − Γ ∗ ) / 2 (V-\Gamma_*)/2 ( V − Γ ∗ ) /2 . Then
ρ P = ∫ D ( ξ ) ∣ ψ Γ ∗ ⟩ ⟨ ψ Γ ∗ ∣ D ( ξ ) † d P ( ξ ) , \rho_P=\int D(\xi)|\psi_{\Gamma_*}\rangle\langle\psi_{\Gamma_*}|D(\xi)^\dagger\,dP(\xi), ρ P = ∫ D ( ξ ) ∣ ψ Γ ∗ ⟩ ⟨ ψ Γ ∗ ∣ D ( ξ ) † d P ( ξ ) , E F ( ρ P ) = E C ( ρ P ) = G ( V ) . (31) E_F(\rho_P)=E_C(\rho_P)=G(V). \tag*{(31)} E F ( ρ P ) = E C ( ρ P ) = G ( V ) . (31)
Proof. The displayed ensemble has covariance V V V and average entanglement E ( Γ ∗ ) = G ( V ) E(\Gamma_*)=G(V) E ( Γ ∗ ) = G ( V ) , because displacements are local unitaries. Thus E C ( ρ P ) ≤ E F ( ρ P ) ≤ G ( V ) E_C(\rho_P)\le E_F(\rho_P)\le G(V) E C ( ρ P ) ≤ E F ( ρ P ) ≤ G ( V ) , while Theorem 1 gives the reverse bounds. □
When V ≠ Γ ∗ V\ne\Gamma_* V = Γ ∗ , a finite-support displacement distribution gives a non-Gaussian state. The same covariance minimum therefore determines the exact preparation cost of an explicit family of Gaussian and non-Gaussian states.
Entanglement cost of Gaussian channels
In this Section we determine the entanglement cost of a Gaussian channel by comparing the entanglement of the states it distributes with the resources required for its teleportation simulation [31 , 32 ] .
A Gaussian channel Φ : A → B \Phi:A\to B Φ : A → B from m ≥ 1 m\ge1 m ≥ 1 input modes to ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 output modes acts on covariance matrices as [33 ] [4]
V ⟼ X V X T + Y , Y = Y T ∈ R 2 ℓ × 2 ℓ , Y + i ( Ω ℓ − X Ω m X T ) ≥ 0 , (32) V\longmapsto XVX^{\mathsf T}+Y,\qquad Y=Y^{\mathsf T}\in\mathbb{R}^{2\ell\times2\ell},\qquad Y+i(\Omega_\ell-X\Omega_mX^{\mathsf T})\ge0, \tag*{(32)} V ⟼ X V X T + Y , Y = Y T ∈ R 2 ℓ × 2 ℓ , Y + i ( Ω ℓ − X Ω m X T ) ≥ 0 , (32)
where X ∈ R 2 ℓ × 2 m X\in\mathbb{R}^{2\ell\times2m} X ∈ R 2 ℓ × 2 m and Y = Y T ∈ R 2 ℓ × 2 ℓ Y=Y^{\mathsf T}\in\mathbb{R}^{2\ell\times2\ell} Y = Y T ∈ R 2 ℓ × 2 ℓ , and the matrix inequality is necessary and sufficient for complete positivity. We take the first moments to transform as d ↦ X d d\mapsto Xd d ↦ X d , since a fixed output displacement is a local unitary and can be implemented at no additional entanglement cost.
We measure simulation errors using the trace distance between states and the normalized diamond distance between channels,
d t r ( ρ , σ ) = 1 2 ∥ ρ − σ ∥ 1 , d ⋄ ( Φ , Ψ ) = 1 2 ∥ Φ − Ψ ∥ ⋄ , (33) d_{\mathrm{tr}}(\rho,\sigma)=\frac{1}{2}\lVert\rho-\sigma\rVert_1,\qquad d_\diamond(\Phi,\Psi)=\frac{1}{2}\lVert\Phi-\Psi\rVert_\diamond, \tag*{(33)} d tr ( ρ , σ ) = 2 1 ∥ ρ − σ ∥ 1 , d ⋄ ( Φ , Ψ ) = 2 1 ∥ Φ − Ψ ∥ ⋄ , (33)
where ∥ ⋅ ∥ 1 \lVert\cdot\rVert_1 ∥ ⋅ ∥ 1 and ∥ ⋅ ∥ ⋄ \lVert\cdot\rVert_\diamond ∥ ⋅ ∥ ⋄ denote the trace norm and diamond norm, respectively; d ⋄ d_\diamond d ⋄ is the supremum of the output trace distance over all input states and arbitrary reference systems.
Following Wilde [17], we consider LOCC simulations of n n n sequential uses of Φ \Phi Φ , supplied with a maximally entangled resource of Schmidt rank M n M_n M n . A test may retain a quantum memory and choose later inputs using earlier outputs; we denote by ϵ n \epsilon_n ϵ n the supremum, over all such adaptive tests, of the trace distance between the final states produced by the channel and its simulation. The entanglement cost of the channel Φ \Phi Φ is
E C ( Φ ) = inf { lim sup n → ∞ log M n n : ϵ n → 0 } , (34) E_C(\Phi)=\inf\left\{\limsup_{n\to\infty}\frac{\log M_n}{n}:\epsilon_n\to0\right\}, \tag*{(34)} E C ( Φ ) = inf { n → ∞ lim sup n log M n : ϵ n → 0 } , (34)
where the infimum runs over sequences of these simulations, and rates are measured in nats per channel use.
The parallel entanglement cost E C P ( Φ ) E_{\mathrm{C}}^{\mathrm{P}}(\Phi) E C P ( Φ ) is defined analogously for joint LOCC simulations of Φ ⊗ n \Phi^{\otimes n} Φ ⊗ n whose diamond distance from the target channel tends to zero, extending the finite-dimensional formulation of Ref. [34 ] . In both settings, local operations may have continuous classical outcomes, and input energy is unrestricted.
We probe the channel with a product Ψ s R A \Psi_{s}^{RA} Ψ s R A of m m m two-mode squeezed vacua, pairing each input mode in A A A with a mode of an m m m -mode reference R R R . Both marginals have covariance s I 2 m sI_{2m} s I 2 m , with s ≥ 1 s \ge1 s ≥ 1 , and applying Φ \Phi Φ to A A A gives the joint reference–output state
ω Φ ( s ) = ( id R ⊗ Φ ) ( Ψ s R A ) . (35) \omega_{\Phi}(s) = (\operatorname{id}_{R} \otimes\Phi)(\Psi_{s}^{RA}). \tag*{(35)} ω Φ ( s ) = ( id R ⊗ Φ ) ( Ψ s R A ) . (35)
With Z m = ⨁ j = 1 m diag ( 1 , − 1 ) Z_m = \bigoplus_{j=1}^{m}\operatorname{diag}(1,-1) Z m = ⨁ j = 1 m diag ( 1 , − 1 ) and c s = s 2 − 1 c_s = \sqrt{s^2-1} c s = s 2 − 1 , the covariance matrix of ω Φ ( s ) \omega_{\Phi}(s) ω Φ ( s ) is
V Φ ( s ) = ( s I 2 m c s Z m X T c s X Z m s X X T + Y ) . (36) V_{\Phi}(s) =
\begin{pmatrix}
sI_{2m} & c_s Z_m X^{\mathsf{T}} \\
c_s X Z_m & sXX^{\mathsf{T}} + Y
\end{pmatrix}. \tag*{(36)} V Φ ( s ) = ( s I 2 m c s X Z m c s Z m X T s X X T + Y ) . (36)
Both the probe Ψ s R A \Psi_{s}^{RA} Ψ s R A and the output ω Φ ( s ) \omega_{\Phi}(s) ω Φ ( s ) have finite energy for every finite s s s , and the state entanglement measures in the following theorem are evaluated across the reference–output bipartition R : B R:B R : B .
Theorem 4 (Entanglement cost of multimode Gaussian channels). (a) For every multimode Gaussian channel Φ = ( X , Y ) \Phi=(X,Y) Φ = ( X , Y ) ,
E C ( Φ ) = E C P ( Φ ) = sup s ≥ 1 E F ( ω Φ ( s ) ) = lim s → ∞ G ( V Φ ( s ) ) . (37) E_{\mathrm{C}}(\Phi)=E_{\mathrm{C}}^{\mathrm{P}}(\Phi)=\sup_{s\geq1}E_{\mathrm{F}}(\omega_{\Phi}(s))
=\lim_{s\to\infty}G(V_{\Phi}(s)). \tag*{(37)} E C ( Φ ) = E C P ( Φ ) = s ≥ 1 sup E F ( ω Φ ( s )) = s → ∞ lim G ( V Φ ( s )) . (37) (b) The common value is finite if and only if the noise matrix Y Y Y is positive definite; if Y Y Y is singular, it is + ∞ +\infty + ∞ .
(c) The entanglement cost is additive on tensor products: for any two multimode Gaussian channels,
E C ( Φ 1 ⊗ Φ 2 ) = E C ( Φ 1 ) + E C ( Φ 2 ) , (38) E_{\mathrm{C}}(\Phi_1\otimes\Phi_2)=E_{\mathrm{C}}(\Phi_1)+E_{\mathrm{C}}(\Phi_2), \tag*{(38)} E C ( Φ 1 ⊗ Φ 2 ) = E C ( Φ 1 ) + E C ( Φ 2 ) , (38) including when either cost is infinite.
Proof of Theorem 4. We prove parts (a) and (b) by obtaining a lower bound from Gaussian probes and matching it by teleportation when Y > 0 Y>0 Y > 0 . For singular Y Y Y , a noiseless output quadrature forces the cost to diverge, while part (c) follows from additivity for Gaussian states.
Probe bounds and teleportation
A channel simulator must also reproduce its action on an entangled probe. Fix s < ∞ s<\infty s < ∞ while the number of simulated uses tends to infinity. Alice prepares ( Ψ s R A ) ⊗ n (\Psi_{s}^{RA})^{\otimes n} ( Ψ s R A ) ⊗ n locally and feeds the input halves into a parallel simulation, obtaining an approximation to ω Φ ( s ) ⊗ n \omega_{\Phi}(s)^{\otimes n} ω Φ ( s ) ⊗ n with the same entanglement consumption and no larger error. Theorem 1 therefore gives the lower bound
G ( V Φ ( s ) ) = E C ( ω Φ ( s ) ) ≤ E C P ( Φ ) ≤ E C ( Φ ) . (39) G(V_{\Phi}(s))=E_{\mathrm{C}}(\omega_{\Phi}(s))\leq E_{\mathrm{C}}^{\mathrm{P}}(\Phi)\leq E_{\mathrm{C}}(\Phi). \tag*{(39)} G ( V Φ ( s )) = E C ( ω Φ ( s )) ≤ E C P ( Φ ) ≤ E C ( Φ ) . (39)
The same probe outputs also serve as teleportation resources. Write T Q \mathcal{T}_{Q} T Q for the additive-noise channel that adds Q ≥ 0 Q\geq0 Q ≥ 0 to the quantum covariance; its displacement law has ordinary covariance Q / 2 Q/2 Q /2 . Braunstein–Kimble teleportation [35 ] with a resource state Ψ s \Psi_s Ψ s implements T ϵ s I 2 m \mathcal{T}_{\epsilon_s I_{2m}} T ϵ s I 2 m , where
ϵ s = 2 ( s − c s ) ↓ 0. (40) \epsilon_s=2(s-c_s)\downarrow0. \tag*{(40)} ϵ s = 2 ( s − c s ) ↓ 0. (40)
Since the channel maps an input displacement ξ \xi ξ to an output displacement X ξ X\xi X ξ , teleportation with a resource ω Φ ( s ) \omega_{\Phi}(s) ω Φ ( s ) and correction X ξ X\xi X ξ implements exactly [31 ]
Φ s = Φ ∘ T ϵ s I 2 m , ( X , Y ) s = ( X , Y + ϵ s X X T ) . (41) \Phi_s=\Phi\circ\mathcal{T}_{\epsilon_s I_{2m}},\qquad
(X,Y)_s=(X,Y+\epsilon_s XX^{\mathsf{T}}). \tag*{(41)} Φ s = Φ ∘ T ϵ s I 2 m , ( X , Y ) s = ( X , Y + ϵ s X X T ) . (41)
Writing A ′ A' A ′ for Alice’s resource modes, the protocol measures the commuting quadratures R A − Z m R A ′ R_A-Z_mR_{A'} R A − Z m R A ′ and displaces Bob’s output by X X X times the outcome. This description applies to arbitrary inputs and their reference extensions.
Teleportation also establishes the existence of the limit in Eq. (37) . Consider a fixed finite-energy Gaussian probe ψ R A \psi^{RA} ψ R A , prepared locally by Alice. Since its simulated output is obtained by LOCC from the shared resource ω Φ ( s ) \omega_{\Phi}(s) ω Φ ( s ) , the output entanglement is at most G ( V Φ ( s ) ) G(V_{\Phi}(s)) G ( V Φ ( s )) . Its covariance differs from that of the actual channel output by 0 R ⊕ ϵ s X X T 0_R\oplus\epsilon_sXX^{\mathsf{T}} 0 R ⊕ ϵ s X X T , which tends to zero. The covariance continuity in Eq. (15) thus gives
E F ( ( id ⊗ Φ ) ( ψ ) ) ≤ lim inf s → ∞ G ( V Φ ( s ) ) . (42) E_{\mathrm{F}}((\operatorname{id}\otimes\Phi)(\psi))
\leq\liminf_{s\to\infty}G(V_{\Phi}(s)). \tag*{(42)} E F (( id ⊗ Φ ) ( ψ )) ≤ s → ∞ lim inf G ( V Φ ( s )) . (42)
The same bound holds after fixed local Gaussian processing of the output. Taking ψ = Ψ t \psi=\Psi_t ψ = Ψ t and then the supremum over t t t yields
sup t ≥ 1 G ( V Φ ( t ) ) ≤ lim inf s → ∞ G ( V Φ ( s ) ) ≤ lim sup s → ∞ G ( V Φ ( s ) ) ≤ sup t ≥ 1 G ( V Φ ( t ) ) . (43) \sup_{t\geq1}G(V_{\Phi}(t))
\leq\liminf_{s\to\infty}G(V_{\Phi}(s))
\leq\limsup_{s\to\infty}G(V_{\Phi}(s))
\leq\sup_{t\geq1}G(V_{\Phi}(t)). \tag*{(43)} t ≥ 1 sup G ( V Φ ( t )) ≤ s → ∞ lim inf G ( V Φ ( s )) ≤ s → ∞ lim sup G ( V Φ ( s )) ≤ t ≥ 1 sup G ( V Φ ( t )) . (43)
The probe limit therefore exists and equals the supremum, with + ∞ +\infty + ∞ allowed.
Achievability with positive noise
When Y > 0 Y>0 Y > 0 , the additional teleportation noise can be produced by a small change of the environment in a fixed Gaussian dilation. Choose a dilation with k k k environmental modes in vacuum, whose retained output quadratures are
R B = X R A + B R E , Y = B B T , B Ω k B T = Δ , (44) R_B=XR_A+BR_E,\qquad
Y=BB^{\mathsf{T}},\qquad
B\Omega_kB^{\mathsf{T}}=\Delta, \tag*{(44)} R B = X R A + B R E , Y = B B T , B Ω k B T = Δ , (44)
where Δ = Ω ℓ − X Ω m X T \Delta=\Omega_{\ell}-X\Omega_mX^{\mathsf{T}} Δ = Ω ℓ − X Ω m X T . Such a dilation exists for arbitrary input and output mode counts, with extra vacuum modes and discarded outputs as needed [33 ] . When Y > 0 Y>0 Y > 0 , the environmental coupling B B B has the right inverse
L B = B T Y − 1 , B L B = I 2 ℓ . (45) L_B=B^{\mathsf{T}}Y^{-1},\qquad
BL_B=I_{2\ell}. \tag*{(45)} L B = B T Y − 1 , B L B = I 2 ℓ . (45)
This right inverse lets us realize the teleportation noise by changing only the environment. Keep the dilation unitary and replace the vacuum state ρ E , 0 \rho_{E,0} ρ E , 0 by a centered Gaussian environment ρ E , s \rho_{E,s} ρ E , s of covariance
V E ( s ) = I 2 k + ϵ s L B X X T L B T . (46) V_E(s)=I_{2k}+\epsilon_sL_BXX^{\mathsf{T}}L_B^{\mathsf{T}}. \tag*{(46)} V E ( s ) = I 2 k + ϵ s L B X X T L B T . (46)
The added noise is exactly that in Eq. (41) . Both channels now use the same unitary with different environment states, so trace-distance contractivity controls the simulation error uniformly over all inputs and reference systems. The overlap of ρ E , s \rho_{E,s} ρ E , s with the vacuum yields
d ⋄ ( Φ s , Φ ) ≤ d t r ( ρ E , s , ρ E , 0 ) ≤ Tr ( X T Y − 1 X ) 2 s . (47) d_{\diamond}(\Phi_s,\Phi) \le d_{\mathrm{tr}}(\rho_{E,s},\rho_{E,0}) \le\sqrt{\frac{\operatorname{Tr}(X^{\mathsf T}Y^{-1}X)}{2s}} . \tag*{(47)} d ⋄ ( Φ s , Φ ) ≤ d tr ( ρ E , s , ρ E , 0 ) ≤ 2 s Tr ( X T Y − 1 X ) . (47)
Appendix E gives the determinant estimate leading to this bound. Environment-comparison proofs of uniform convergence were developed in Refs. [36 , 37 ] ; the latter also characterizes full-rank noise as the criterion for single-mode Gaussian channels. The right inverse (45) gives the bound for every Y > 0 Y>0 Y > 0 , including channels with unequal mode counts.
To prove that the cost is finite, it suffices to construct one finite-energy resource that simulates Φ \Phi Φ exactly. For teleportation with gain X X X , the matrix selecting the resource contribution to the output noise is
M t e l = ( − X Z m I 2 ℓ ) . (48) M_{\mathrm{tel}}=\left(-XZ_m\ \mathbb{I}_{2\ell}\right). \tag*{(48)} M tel = ( − X Z m I 2 ℓ ) . (48)
Both M t e l M_{\mathrm{tel}} M tel and B B B have full row rank, and their rows have the same pairwise symplectic products:
M t e l Ω m + ℓ M t e l T = Δ = B Ω k B T . (49) M_{\mathrm{tel}}\Omega_{m+\ell}M_{\mathrm{tel}}^{\mathsf T}=\Delta=B\Omega_kB^{\mathsf T}. \tag*{(49)} M tel Ω m + ℓ M tel T = Δ = B Ω k B T . (49)
The map matching corresponding rows extends to a symplectic transformation after adding ancillary modes. Appendix E3 uses this extension to construct a finite-energy Gaussian resource with covariance V r e s V_{\mathrm{res}} V res satisfying
M t e l V r e s M t e l T = Y . (50) M_{\mathrm{tel}}V_{\mathrm{res}}M_{\mathrm{tel}}^{\mathsf T}=Y. \tag*{(50)} M tel V res M tel T = Y . (50)
This simulator condition also appears in Ref. [38 Proposition 8.8] , which proves the existence of a finite-energy resource when Y + i Δ > 0 Y+i\Delta>0 Y + i Δ > 0 . The symplectic extension covers every Y > 0 Y>0 Y > 0 , including cases where Y + i Δ Y+i\Delta Y + i Δ is singular. The characteristic-function calculation in Appendix E3 verifies exact simulation on arbitrary inputs and their reference extensions. Since the resource has finite local entropy, it has finite entanglement cost; preparing many copies and applying the simulator therefore proves E C ( Φ ) < ∞ E_{\mathrm C}(\Phi)<\infty E C ( Φ ) < ∞ . The lower bound in Eq. (39) then makes the probe limit finite.
We attain the probe limit by letting the resource squeezing increase with the number of simulated uses. The required pure Gaussian seeds are prepared from maximally entangled states by entanglement dilution [17]. The Gaussian Schmidt decomposition expresses the reduced state of each seed on Alice’s resource modes, up to a local unitary, as a product of at most m m m thermal states. For a thermal state with eigenvalues p j p_j p j and mean occupation u > 0 u>0 u > 0 , its information variance is the variance of − log p j -\log p_j − log p j when j j j is sampled with probability p j p_j p j . Writing β t h = log ( 1 + u − 1 ) \beta_{\mathrm{th}}=\log(1+u^{-1}) β th = log ( 1 + u − 1 ) , this variance is
u ( u + 1 ) log 2 ( 1 + u − 1 ) = β t h 2 4 sinh 2 ( β t h / 2 ) ≤ 1. (51) u(u+1)\log^2(1+u^{-1})=\frac{\beta_{\mathrm{th}}^2}{4\sinh^2(\beta_{\mathrm{th}}/2)}\le1. \tag*{(51)} u ( u + 1 ) log 2 ( 1 + u − 1 ) = 4 sinh 2 ( β th /2 ) β th 2 ≤ 1. (51)
Vacuum factors contribute zero and variances add, so the information variance of the seed’s Schmidt probabilities is at most m m m . Because this bound depends only on the number of modes, it also applies when the seed changes with n n n and its energy grows. A cutoff of the Schmidt expansion with excess entanglement rate a n > 0 a_n>0 a n > 0 consequently approximates n n n copies with trace-distance error at most m / ( n a n 2 ) \sqrt{m/(na_n^2)} m / ( n a n 2 ) ; Lemma 9 gives the precise rank bound and proof. Choose s n → ∞ s_n\to\infty s n → ∞ so that d ⋄ ( Φ s n , Φ ) ≤ n − 2 d_{\diamond}(\Phi_{s_n},\Phi)\le n^{-2} d ⋄ ( Φ s n , Φ ) ≤ n − 2 , and let Γ n \Gamma_n Γ n attain G ( V Φ ( s n ) ) G(\mathcal{V}_{\Phi}(s_n)) G ( V Φ ( s n )) . Approximate n n n copies of the corresponding pure Gaussian seed using a n = n − 1 / 4 a_n=n^{-1/4} a n = n − 1/4 , and apply independent Gaussian random displacement unitaries to the copies, with ordinary covariance [ V Φ ( s n ) − Γ n ] / 2 [V_{\Phi}(s_n)-\Gamma_n]/2 [ V Φ ( s n ) − Γ n ] /2 on each copy. The resulting state approximates ω Φ ( s n ) ⊗ n \omega_{\Phi(s_n)}^{\otimes n} ω Φ ( s n ) ⊗ n with no larger trace-distance error. Replacing channel calls one at a time in an adaptive test contributes at most n d ⋄ ( Φ s n , Φ ) n d_{\diamond}(\Phi_{s_n},\Phi) n d ⋄ ( Φ s n , Φ ) [31 , 37 ] . Contractivity bounds the additional error from resource preparation by the trace distance of the joint resource. Hence
ε n ≤ n − 1 + m n − 1 / 4 ⟶ 0 , (52) \varepsilon_n\le n^{-1}+\sqrt{m}\,n^{-1/4}\longrightarrow0, \tag*{(52)} ε n ≤ n − 1 + m n − 1/4 ⟶ 0 , (52)
log M n n ≤ G ( V Φ ( s n ) ) + n − 1 / 4 + log 2 n ⟶ lim s → ∞ G ( V Φ ( s ) ) . (53) \frac{\log M_n}{n}\le G(\mathcal{V}_{\Phi}(s_n))+n^{-1/4}+\frac{\log2}{n}\longrightarrow\lim_{s\to\infty}G(\mathcal{V}_{\Phi}(s)). \tag*{(53)} n log M n ≤ G ( V Φ ( s n )) + n − 1/4 + n log 2 ⟶ s → ∞ lim G ( V Φ ( s )) . (53)
The right-hand side of Eq. (53) converges to the probe limit, which is therefore achievable by sequential simulation. Together with Eq. (39) , this proves parts (a) and (b) when Y > 0 Y>0 Y > 0 .
Noiseless quadratures and additivity
If Y Y Y is singular, at least one output quadrature receives no added noise. We isolate a canonical input–output pair containing this quadrature and show that its probe entanglement diverges. Choose a nonzero real w ∈ ker Y w\in\ker Y w ∈ ker Y and z z z such that w T Ω 2 ℓ z = 1 w^{\mathsf T}\Omega_{2\ell}z=1 w T Ω 2 ℓ z = 1 . Positivity of Y + i Δ Y+i\Delta Y + i Δ implies Δ w = 0 \Delta w=0 Δ w = 0 , and therefore
( X T w ) T Ω m ( X T z ) = 1. (54) (X^{\mathsf T}w)^{\mathsf T}\Omega_m(X^{\mathsf T}z)=1. \tag*{(54)} ( X T w ) T Ω m ( X T z ) = 1. (54)
Thus the output quadratures w T R B , z T R B w^{\mathsf T}R_B,z^{\mathsf T}R_B w T R B , z T R B and their input counterparts w T X R A , z T X R A w^{\mathsf T}XR_A,z^{\mathsf T}XR_A w T X R A , z T X R A both form canonical pairs. Completing them to symplectic coordinates makes the selected output mode depend only on the selected input mode. Preparing the remaining inputs in vacuum and discarding the remaining outputs therefore gives an effective single-mode channel B ν \mathcal{B}_{\nu} B ν with
X B ν = I 2 , Y B ν = ( 0 0 0 ν ) , ν = z T Y z ≥ 0. (55) X_{\mathcal{B}_{\nu}}=\mathbb{I}_2,\qquad
Y_{\mathcal{B}_{\nu}}=
\begin{pmatrix}
0 & 0\\
0 & \nu
\end{pmatrix},
\qquad\nu=z^{\mathsf T}Yz\ge0. \tag*{(55)} X B ν = I 2 , Y B ν = ( 0 0 0 ν ) , ν = z T Y z ≥ 0. (55)
For a two-mode squeezed probe with marginal covariance t I 2 t\mathbb{I}_2 t I 2 , its output marginal has symplectic eigenvalue t ( t + ν ) \sqrt{t(t+\nu)} t ( t + ν ) , while the joint state has symplectic eigenvalues 1 1 1 and 1 + t ν \sqrt{1+t\nu} 1 + t ν . The coherent information of the joint state ω B ν ( t ) \omega_{\mathcal{B}_{\nu}}(t) ω B ν ( t ) , defined as its output entropy minus its joint entropy, is
I ν ( t ) = g ( t ( t + ν ) − 1 2 ) − g ( 1 + t ν − 1 2 ) . (56) I_{\nu}(t)=g\left(\frac{\sqrt{t(t+\nu)}-1}{2}\right)-g\left(\frac{\sqrt{1+t\nu}-1}{2}\right). \tag*{(56)} I ν ( t ) = g ( 2 t ( t + ν ) − 1 ) − g ( 2 1 + t ν − 1 ) . (56)
For ν = 0 \nu=0 ν = 0 , this equals g ( ( t − 1 ) / 2 ) g((t-1)/2) g (( t − 1 ) /2 ) ; for fixed ν > 0 \nu>0 ν > 0 , it grows as 1 2 log t + O ( 1 ) \frac{1}{2}\log t+O(1) 2 1 log t + O ( 1 ) . Concavity of conditional entropy gives I ν ( t ) ≤ E F ( ω B ν ( t ) ) I_{\nu}(t)\le E_{\mathrm F}(\omega_{\mathcal{B}_{\nu}}(t)) I ν ( t ) ≤ E F ( ω B ν ( t )) , which equals E C ( ω B ν ( t ) ) E_{\mathrm C}(\omega_{\mathcal{B}_{\nu}}(t)) E C ( ω B ν ( t )) by Theorem 1 . Every simulation of Φ \Phi Φ also simulates B ν \mathcal{B}_{\nu} B ν by the local Gaussian operations just described, so the probe lower bound forces both channel costs to be infinite. The limit in Eq. (37) diverges as well: for each fixed t t t , apply Eq. (42) to the Gaussian input preparation realizing this effective channel and its local output processing, and then let t → ∞ t \to\infty t → ∞ . This proves parts (a) and (b) for singular Y Y Y .
Finally, the probe covariance for Φ 1 ⊗ Φ 2 \Phi_1 \otimes\Phi_2 Φ 1 ⊗ Φ 2 is V Φ 1 ( s ) ⊕ V Φ 2 ( s ) V_{\Phi_1}(s) \oplus V_{\Phi_2}(s) V Φ 1 ( s ) ⊕ V Φ 2 ( s ) up to local mode ordering. State additivity and the established limits give Eq. (38) , proving part (c) and completing the proof of Theorem 4. □
Bounds from moments
The probe argument also applies to non-Gaussian channels, because Theorem 1 bounds the entanglement cost of an arbitrary output state using only its covariance matrix.
Corollary 5 (Channel bounds from moments). Let N : A → B \mathcal{N}: A \to B N : A → B be a channel whose probe outputs ω N ( s ) = ( id R ⊗ N ) ( Ψ s R A ) \omega_{\mathcal{N}}(s) = (\operatorname{id}_{R} \otimes\mathcal{N})(\Psi_{s}^{RA}) ω N ( s ) = ( id R ⊗ N ) ( Ψ s R A ) have finite energy for every finite s s s , and denote their covariance matrices by V N ( s ) V_{\mathcal{N}}(s) V N ( s ) . Then
E C ( N ) ≥ E C p ( N ) ≥ sup s ≥ 1 G ( V N ( s ) ) . (57) E_{\mathrm{C}}(\mathcal{N}) \ge E_{\mathrm{C}}^{\mathrm{p}}(\mathcal{N}) \ge\sup_{s \ge1} G(V_{\mathcal{N}}(s)). \tag*{(57)} E C ( N ) ≥ E C p ( N ) ≥ s ≥ 1 sup G ( V N ( s )) . (57) If N \mathcal{N} N has the same first- and second-moment action as a Gaussian channel Φ = ( X , Y ) \Phi= (X,Y) Φ = ( X , Y ) on every finite-energy input, the supremum in Eq. (57) equals E C ( Φ ) E_{\mathrm{C}}(\Phi) E C ( Φ ) . In particular, singular Y Y Y gives E C p ( N ) = E C ( N ) = + ∞ E_{\mathrm{C}}^{\mathrm{p}}(\mathcal{N}) = E_{\mathrm{C}}(\mathcal{N}) = +\infty E C p ( N ) = E C ( N ) = + ∞ .
Proof. For each fixed s s s , feeding locally prepared copies of Ψ s R A \Psi_{s}^{RA} Ψ s R A into a parallel simulator gives E C p ( N ) ≥ E C ( ω N ( s ) ) E_{\mathrm{C}}^{\mathrm{p}}(\mathcal{N}) \ge E_{\mathrm{C}}(\omega_{\mathcal{N}}(s)) E C p ( N ) ≥ E C ( ω N ( s )) . Theorem 1 bounds the latter by G ( V N ( s ) ) G(V_{\mathcal{N}}(s)) G ( V N ( s )) from below, and taking the supremum over s s s proves Eq. (57) .
If the first-moment actions agree on all finite-energy states, polarization gives N ∗ ( R B ) = X R A \mathcal{N}^{*}(R_{B}) = X R_{A} N ∗ ( R B ) = X R A as an identity of quadratic forms, where N ∗ \mathcal{N}^{*} N ∗ is the Heisenberg adjoint of N \mathcal{N} N . This identity also fixes the cross moments between the reference and output. Together with equality of the second-moment actions, it gives V N ( s ) = V Φ ( s ) V_{\mathcal{N}}(s) = V_{\Phi}(s) V N ( s ) = V Φ ( s ) , and Theorem 4 gives the remaining claims. □
Equation (37) expresses the entanglement cost of a channel as the squeezing limit of a covariance minimum on m + ℓ m+\ell m + ℓ modes. Finite-energy teleportation simulations for single-mode phase-insensitive channels were developed in Refs. [39 , 40 ] and used to bound private communication in Refs. [41 , 42 ] . Wilde determined the entanglement cost of pure-loss and quantum-limited amplifier channels [17]; Gaussian optimality and additivity for general two-mode states subsequently extended these results to thermal-loss, thermal-amplifier, and additive-noise channels [18]. The present work extends this characterization to arbitrary multimode Gaussian channels, including those with coupled modes and correlated noise.
Classical capacity of Gaussian channels
The state theorem also determines how much classical information can be sent through a Gaussian channel. The connection is the output entropy: a pure input in an isometric dilation produces a pure output–environment state whose entanglement equals that entropy. Minimizing the average output entropy at a fixed average input state is therefore an entanglement of formation problem, to which the covariance minimum applies.
Gaussian encodings attain the capacity of gauge-covariant channels, which respect global phase rotations, under phase-invariant quadratic energy constraints [43 , 44 ] . For general single-mode channels, Ji established Gaussian optimality in the single-use Holevo information [45 ] , and Adesso, Lami, and Giovannetti established additivity and the resulting classical capacity [21 ] . The result in this Section allows arbitrary coupling between the modes within each channel use.
We retain the channel notation of Eq. (32) , with m , ℓ ≥ 1 m,\ell\ge1 m , ℓ ≥ 1 , and constrain the input energy using the quadratic Hamiltonian H = 1 2 R T h R + c H = \frac{1}{2}R^{\mathsf{T}}hR + c H = 2 1 R T h R + c , where h = h T > 0 h = h^{\mathsf{T}} > 0 h = h T > 0 ; any linear term can be removed by an input displacement. Its ground energy is E 0 = c + 1 2 ∑ j = 1 m ν j ( h ) E_0 = c + \frac{1}{2}\sum_{j=1}^{m}\nu_j(h) E 0 = c + 2 1 ∑ j = 1 m ν j ( h ) , where ν j ( h ) \nu_j(h) ν j ( h ) are the symplectic eigenvalues of h h h , and its expectation in a state with first moments d d d and covariance V V V is 1 4 Tr ( h V ) + 1 2 d T h d + c \frac{1}{4}\operatorname{Tr}(hV) + \frac{1}{2}d^{\mathsf{T}}hd + c 4 1 Tr ( hV ) + 2 1 d T h d + c .
To communicate classical information through Φ \Phi Φ , an n n n -use code encodes one of K n K_n K n equiprobable messages into an input state ρ x ( n ) \rho_x^{(n)} ρ x ( n ) , and the receiver performs a joint measurement on the output of Φ ⊗ n \Phi^{\otimes n} Φ ⊗ n to identify the message. The rate is log K n / n \log K_n/n log K n / n nats per channel use, and we denote the average decoding error by ϵ n \epsilon_n ϵ n . The average input ρ ˉ n = K n − 1 ∑ x = 1 K n ρ x ( n ) \bar{\rho}_n = K_n^{-1}\sum_{x=1}^{K_n}\rho_x^{(n)} ρ ˉ n = K n − 1 ∑ x = 1 K n ρ x ( n ) must satisfy Tr ( ρ ˉ n H n ) ≤ n E \operatorname{Tr}(\bar{\rho}_n H_n) \le nE Tr ( ρ ˉ n H n ) ≤ n E , where H n = ∑ j = 1 n H ( j ) H_n = \sum_{j=1}^{n}H^{(j)} H n = ∑ j = 1 n H ( j ) . The unassisted classical capacity without feedback is the largest asymptotic rate achievable with vanishing error,
C ( Φ , H , E ) = sup { lim inf n → ∞ log K n n : ϵ n → 0 } , (58) C(\Phi,H,E) = \sup\left\{\liminf_{n \to\infty}\frac{\log K_n}{n} : \epsilon_n \to0\right\}, \tag*{(58)} C ( Φ , H , E ) = sup { n → ∞ lim inf n log K n : ϵ n → 0 } , (58)
where the supremum runs over sequences of codes obeying the average-energy constraint.
For an input ensemble E = { p x , ρ x } \mathcal{E} = \{p_x,\rho_x\} E = { p x , ρ x } with average ρ ˉ E = ∑ x p x ρ x \bar{\rho}_{\mathcal{E}} = \sum_x p_x\rho_x ρ ˉ E = ∑ x p x ρ x , the output Holevo information is [46 ]
χ ( Φ , E ) = S ( Φ ( ρ ˉ E ) ) − ∑ x p x S ( Φ ( ρ x ) ) . (59) \chi(\Phi,\mathcal{E}) = S(\Phi(\bar{\rho}_{\mathcal{E}})) - \sum_x p_x S(\Phi(\rho_x)). \tag*{(59)} χ ( Φ , E ) = S ( Φ ( ρ ˉ E )) − x ∑ p x S ( Φ ( ρ x )) . (59)
It is the entropy of the average output minus the average entropy of the individual signal outputs, and bounds the classical information recoverable by measuring the output. Optimizing over ensembles E \mathcal{E} E under the same average-energy constraint gives
χ n ( E ) = sup E n Tr ( ρ ˉ E n H n ) ≤ n E χ ( Φ ⊗ n , E n ) . (60) \chi_n(E) = \sup_{\substack{\mathcal{E}_n\\ \operatorname{Tr}(\bar{\rho}_{\mathcal{E}_n}H_n)\le nE}} \chi(\Phi^{\otimes n},\mathcal{E}_n). \tag*{(60)} χ n ( E ) = E n Tr ( ρ ˉ E n H n ) ≤ n E sup χ ( Φ ⊗ n , E n ) . (60)
The ensemble probabilities are arbitrary, and the signal states may be non-Gaussian and entangled across uses. Continuous ensembles are also included, with sums replaced by integrals [47 ] .
A common input displacement induces a common output displacement, so centering the average input preserves its Holevo information and removes the nonnegative contribution of its first moments to the energy. For a centered Gaussian average with covariance V V V , a pure Gaussian seed with covariance γ ≤ V \gamma\le V γ ≤ V can be modulated by centered Gaussian displacements with ordinary covariance ( V − γ ) / 2 (V-\gamma)/2 ( V − γ ) /2 . We write V H ( E ) V_H(E) V H ( E ) for the allowed average covariances and f Φ ( V ) f_\Phi(V) f Φ ( V ) for the largest Holevo information obtained from these displacement ensembles at fixed V V V :
V H ( E ) = { V : V + i Ω m ≥ 0 , 1 4 Tr ( h V ) + c ≤ E } , (61) V_H(E)=\left\{V: V+i\Omega_m\ge0,\quad\frac{1}{4}\operatorname{Tr}(hV)+c\le E\right\}, \tag*{(61)} V H ( E ) = { V : V + i Ω m ≥ 0 , 4 1 Tr ( hV ) + c ≤ E } , (61)
μ Φ ( V ) = min γ > 0 , γ Ω m γ = Ω m γ ≤ V S G ( X γ X T + Y ) , (62) \mu_\Phi(V)=\min_{\substack{\gamma>0,\,\gamma\Omega_m\gamma=\Omega_m\\ \gamma\le V}} S_G(X\gamma X^{\mathsf T}+Y), \tag*{(62)} μ Φ ( V ) = γ > 0 , γ Ω m γ = Ω m γ ≤ V min S G ( X γ X T + Y ) , (62)
f Φ ( V ) = S G ( X V X T + Y ) − μ Φ ( V ) . (63) f_\Phi(V)=S_G(XVX^{\mathsf T}+Y)-\mu_\Phi(V). \tag*{(63)} f Φ ( V ) = S G ( X V X T + Y ) − μ Φ ( V ) . (63)
Here μ Φ ( V ) \mu_\Phi(V) μ Φ ( V ) is the least output entropy of a compatible pure Gaussian seed, while the first term in f Φ ( V ) f_\Phi(V) f Φ ( V ) is the entropy of the Gaussian average output. For the usual constraint on the mean total photon number, the input Hamiltonian is H = ∑ j = 1 m a j † a j = 1 2 R T R − m 2 H=\sum_{j=1}^{m}a_j^\dagger a_j=\frac{1}{2}\mathbf{R}^{\mathsf T}R-\frac{m}{2} H = ∑ j = 1 m a j † a j = 2 1 R T R − 2 m . Thus h = I 2 m h=I_{2m} h = I 2 m and c = − m / 2 c=-m/2 c = − m /2 , and E E E is the allowed mean total photon number per channel use. For the centered average inputs considered above, the constraint becomes Tr V ≤ 4 E + 2 m \operatorname{Tr}V\le4E+2m Tr V ≤ 4 E + 2 m .
Theorem 6 (Classical capacity of multimode Gaussian channels). (a) For every multimode Gaussian channel Φ \Phi Φ , quadratic Hamiltonian H H H as above, and finite E ≥ E 0 E\ge E_0 E ≥ E 0 ,
C ( Φ , H , E ) = max V ∈ V H ( E ) f Φ ( V ) , C(\Phi,H,E)=\max_{V\in V_H(E)}f_\Phi(V), C ( Φ , H , E ) = V ∈ V H ( E ) max f Φ ( V ) , χ n ( E ) = n C ( Φ , H , E ) ( n ≥ 1 ) . (64) \chi_n(E)=nC(\Phi,H,E)\qquad(n\ge1). \tag*{(64)} χ n ( E ) = n C ( Φ , H , E ) ( n ≥ 1 ) . (64) (b) *The covariance maximum and the seed minimum are attained. An optimal ensemble consists of Gaussian-distributed displacements of a pure Gaussian seed, which may be entangled across the input modes within one use. Independent copies of this ensemble attain χ n ( E ) \chi_n(E) χ n ( E ) , and the capacity is achievable with product signal states across uses and collective decoding. (c) The capacity is concave and nondecreasing in E E E , with C ( Φ , H , E 0 ) = 0 C(\Phi,H,E_0)=0 C ( Φ , H , E 0 ) = 0 .
Proof of Theorem 6. We first express the minimum average output entropy as entanglement of formation. The fixed-covariance bound and strong superadditivity then control arbitrary encodings across multiple uses, while Gaussian modulation attains the resulting Holevo bound.
Choose a Gaussian isometry T : A → B F T:A\to BF T : A → B F realizing Φ ( ρ ) = Tr F ( T ρ T † ) \Phi(\rho)=\operatorname{Tr}_F(T\rho T^\dagger) Φ ( ρ ) = Tr F ( T ρ T † ) , obtained by adjoining q q q vacuum modes and applying a Gaussian unitary U U U with symplectic matrix Ξ \Xi Ξ [33 ] :
T ∣ ψ ⟩ = U ( ∣ ψ ⟩ ⊗ ∣ 0 ⟩ ⊗ q ) , Σ ( V ) = Ξ ( V ⊕ I 2 q ) Ξ T . (65) T|\psi\rangle=U(|\psi\rangle\otimes|0\rangle^{\otimes q}),\qquad\Sigma(V)=\Xi(V\oplus I_{2q})\Xi^{\mathsf T}. \tag*{(65)} T ∣ ψ ⟩ = U ( ∣ ψ ⟩ ⊗ ∣0 ⟩ ⊗ q ) , Σ ( V ) = Ξ ( V ⊕ I 2 q ) Ξ T . (65)
The output B B B contains ℓ \ell ℓ modes and has covariance X V X T + Y XVX^{\mathsf T}+Y X V X T + Y , while F F F contains the discarded modes, with an idle vacuum mode included if necessary. Such vacuum dilations also cover unequal input and output mode counts, as detailed in Appendix F.
Refining mixed signals into pure states preserves the average input and cannot increase the average output entropy, by concavity. The isometry maps pure ensembles of ρ \rho ρ to pure ensembles of T ρ T † T\rho T^\dagger T ρ T † . Conversely, the vectors in any pure ensemble of T ρ T † T\rho T^\dagger T ρ T † lie in the range of T T T almost surely and can be pulled back by T † T^\dagger T † . Writing ψ \psi ψ for a pure-state projector, each dilated signal has entanglement S ( Φ ( ψ ) ) S(\Phi(\psi)) S ( Φ ( ψ )) , which gives the Matsumoto–Shimono–Winter identity [22 ] ,
S ^ Φ ( ρ ) : = inf ρ = ∫ ψ d π ( ψ ) ∫ S ( Φ ( ψ ) ) d π ( ψ ) = E F B : F ( T ρ T † ) . (66) \widehat{S}_\Phi(\rho):=\inf_{\rho=\int\psi\,\mathrm{d}\pi(\psi)}\int S(\Phi(\psi))\,\mathrm{d}\pi(\psi)=E_F^{B:F}(T\rho T^\dagger). \tag*{(66)} S Φ ( ρ ) := ρ = ∫ ψ d π ( ψ ) inf ∫ S ( Φ ( ψ )) d π ( ψ ) = E F B : F ( T ρ T † ) . (66)
For finite-energy inputs, the output and environment entropies are finite, and the ensemble correspondence applies to continuous ensembles as well.
Every pure input covariance γ ≤ V \gamma\le V γ ≤ V gives a pure covariance Γ ≤ Σ ( V ) \Gamma\le\Sigma(V) Γ ≤ Σ ( V ) through
Γ = Ξ ( γ ⊕ I 2 q ) Ξ T . (67) \Gamma=\Xi(\gamma\oplus I_{2q})\Xi^{\mathsf T}. \tag*{(67)} Γ = Ξ ( γ ⊕ I 2 q ) Ξ T . (67)
Conversely, undoing the dilation gives Γ ′ = Ξ − 1 Γ Ξ − T ≤ V ⊕ I 2 q \Gamma'=\Xi^{-1}\Gamma\Xi^{-\mathsf T}\le V\oplus I_{2q} Γ ′ = Ξ − 1 Γ Ξ − T ≤ V ⊕ I 2 q . The centered Gaussian state with covariance Γ ′ \Gamma' Γ ′ has an ancillary marginal with covariance at most I 2 q I_{2q} I 2 q , hence nonpositive mean photon number. Positivity of the number operator forces this marginal to be the vacuum, which factors from the remaining state. Thus Γ ′ = γ ⊕ I 2 q \Gamma'=\gamma\oplus I_{2q} Γ ′ = γ ⊕ I 2 q with γ \gamma γ pure and γ ≤ V \gamma\le V γ ≤ V , establishing a bijection between the input seeds and the pure covariances below Σ ( V ) \Sigma(V) Σ ( V ) . Minimizing their entanglement therefore gives
G B : F ( Σ ( V ) ) = min γ ≤ V γ p u r e S G ( X γ X T + Y ) = μ Φ ( V ) . (68) G^{B:F}(\Sigma(V))=\min_{\substack{\gamma\le V\\ \gamma\ \mathrm{pure}}}S_G(X\gamma X^{\mathsf T}+Y)=\mu_\Phi(V). \tag*{(68)} G B : F ( Σ ( V )) = γ ≤ V γ pure min S G ( X γ X T + Y ) = μ Φ ( V ) . (68)
For an arbitrary input state ρ \rho ρ with covariance V ρ V_\rho V ρ , combining Eqs. (66) and (68) with Theorem 1 yields
S ^ Φ ( ρ ) ≥ μ Φ ( V ρ ) , S ^ Φ ( ρ V ) = μ Φ ( V ) . (69) \widehat{S}_\Phi(\rho)\ge\mu_\Phi(V_\rho),\qquad\widehat{S}_\Phi(\rho_V)=\mu_\Phi(V). \tag*{(69)} S Φ ( ρ ) ≥ μ Φ ( V ρ ) , S Φ ( ρ V ) = μ Φ ( V ) . (69)
Thus μ Φ ( V ) \mu_\Phi(V) μ Φ ( V ) bounds the average output entropy of every ensemble whose average input has covariance V V V , and Gaussian displacement modulation attains this bound for the Gaussian average ρ V \rho_V ρ V .
Additivity and optimal encoding
The identity μ Φ ( V ) = G B : F ( Σ ( V ) ) \mu_\Phi(V)=G^{B:F}(\Sigma(V)) μ Φ ( V ) = G B : F ( Σ ( V )) makes μ Φ \mu_\Phi μ Φ convex, by Eq. (19) and the affine dependence of Σ ( V ) \Sigma(V) Σ ( V ) on V V V . Gaussian entropy is concave in the covariance: a mixture of centered Gaussian states has at least their average entropy, and the Gaussian state with the mixture’s covariance has at least the entropy of that mixture. Consequently, S G ( X V X T + Y ) S_G(XVX^{\mathsf T}+Y) S G ( X V X T + Y ) is concave in V V V , and so is f Φ ( V ) f_\Phi(V) f Φ ( V ) .
For an arbitrary average input ρ ˉ \bar{\rho} ρ ˉ on n n n channel uses, let V j V_j V j be its j j j th marginal covariance and let Σ n \Sigma_n Σ n be the covariance after the dilation T ⊗ n T^{\otimes n} T ⊗ n . Applying the fixed-covariance bound and Gaussian strong superadditivity from Theorem 1 gives
S ^ Φ ⊗ n ( ρ ˉ ) ≥ G B n : F n ( Σ n ) ≥ ∑ j = 1 n μ Φ ( V j ) . (70) \widehat{S}_{\Phi^{\otimes n}}(\bar{\rho})\ge G^{B^n:F^n}(\Sigma_n)\ge\sum_{j=1}^{n}\mu_\Phi(V_j). \tag*{(70)} S Φ ⊗ n ( ρ ˉ ) ≥ G B n : F n ( Σ n ) ≥ j = 1 ∑ n μ Φ ( V j ) . (70)
Here the block of ∑ n \sum_n ∑ n on B j F j B_jF_j B j F j is Σ ( V j ) \Sigma(V_j) Σ ( V j ) , including when ρ ˉ \bar{\rho} ρ ˉ is correlated and non-Gaussian. Entropy subadditivity and Gaussian maximal entropy at fixed covariance also bound the entropy of the average output by ∑ j S G ( X V j X T + Y ) \sum_j S_G(XV_jX^{\mathsf T}+Y) ∑ j S G ( X V j X T + Y ) . The Holevo information of any allowed ensemble therefore satisfies
χ ≤ ∑ j = 1 n f Φ ( V j ) ≤ n f Φ ( 1 n ∑ j = 1 n V j ) ≤ n max V ∈ V H ( E ) f Φ ( V ) . (71) \begin{aligned}
\chi\le\sum_{j=1}^{n} f_{\Phi}(V_j) \le n f_{\Phi}\left(\frac{1}{n}\sum_{j=1}^{n}V_j\right) \\
&\le n\max_{V\in\mathcal{V}_H(E)} f_{\Phi}(V). \tag*{(71)}
\end{aligned} χ ≤ j = 1 ∑ n f Φ ( V j ) ≤ n f Φ ( n 1 j = 1 ∑ n V j ) ≤ n V ∈ V H ( E ) max f Φ ( V ) . (71)
The second inequality is Jensen’s inequality. For the last, the average covariance n − 1 ∑ j V j n^{-1}\sum_j V_j n − 1 ∑ j V j is physical and belongs to V H ( E ) \mathcal{V}_H(E) V H ( E ) , since the first moments make a nonnegative contribution to the total input energy.
Both covariance optimizations are attained on the joint domain
{ ( V , γ ) : V ∈ V H ( E ) , γ ≤ V , γ pure } . (72) \{(V,\gamma): V\in\mathcal{V}_H(E),\ \gamma\le V,\ \gamma\ \text{pure}\}. \tag*{(72)} {( V , γ ) : V ∈ V H ( E ) , γ ≤ V , γ pure } . (72)
This set is compact: the energy constraint and h > 0 h>0 h > 0 bound V V V and γ \gamma γ above, purity bounds the eigenvalues of γ \gamma γ away from zero, and the constraints are closed. Gaussian entropy is continuous there, so S G ( X V X T + Y ) − S G ( X γ X T + Y ) S_G(XVX^{\mathsf T}+Y)-S_G(X\gamma X^{\mathsf T}+Y) S G ( X V X T + Y ) − S G ( X γ X T + Y ) attains its maximum. The same argument at fixed V V V proves attainment of the seed minimum.
Let ( V ∗ , γ ∗ ) (V_*,\gamma_*) ( V ∗ , γ ∗ ) be an optimizing pair. Displacing the pure Gaussian state ρ γ ∗ \rho_{\gamma_*} ρ γ ∗ according to a centered Gaussian law with ordinary covariance ( V ∗ − γ ∗ ) / 2 (V_*-\gamma_*)/2 ( V ∗ − γ ∗ ) /2 produces the average state ρ V ∗ \rho_{V_*} ρ V ∗ . Every signal output has entropy S G ( X γ ∗ X T + Y ) S_G(X\gamma_*X^{\mathsf T}+Y) S G ( X γ ∗ X T + Y ) , so the ensemble has Holevo information f Φ ( V ∗ ) f_{\Phi}(V_*) f Φ ( V ∗ ) . Independent copies saturate Eq. (71) for every n n n .
Concavity of f Φ f_{\Phi} f Φ and the affine energy constraint make the optimal value concave in E E E , while inclusion of the feasible sets makes it nondecreasing. Being finite and concave, it is continuous for E > E 0 E>E_0 E > E 0 . At E = E 0 E=E_0 E = E 0 , the only allowed average input is the unique pure ground state of H H H , so every signal is that state and both the Holevo optimum and the capacity vanish.
To identify the Holevo optimum with operational capacity, observe that h > 0 h>0 h > 0 makes the input energy constraint bound the total input photon number. The channel’s affine action on moments then bounds output photon number and hence output entropy on each allowed energy set. The constrained coding theorem, together with the Holevo bound for the converse, identifies C ( Φ , H , E ) C(\Phi,H,E) C ( Φ , H , E ) with lim n → ∞ χ n ( E ) / n \lim_{n\to\infty}\chi_n(E)/n lim n → ∞ χ n ( E ) / n [46 Proposition 3] .
For E > E 0 E>E_0 E > E 0 , choose an optimal Gaussian ensemble at a slightly smaller energy. Truncating and discretizing its displacement modulation gives finite signal alphabets within the budget, with Holevo information approaching that of the continuous ensemble; Appendix § 1.1.6 supplies the energy and approximation estimates. Continuity of the optimal value allows the smaller energy to approach E E E , proving achievability with product Gaussian signal states and collective decoding. This establishes parts (a) and (b), while the energy dependence proved above gives part (c), completing the proof of Theorem 6. □
Examples and applications
We now apply the state and channel results from the previous Sections to two relevant examples.
The first example uses Theorem 1 to calculate the exact entanglement cost of a four-mode Gaussian family containing bound entangled states, quantifying how much entanglement is needed to prepare states from which no entanglement can be distilled back. The optimal preparation also yields non-Gaussian states with the same entanglement cost.
The second example applies Theorems 4 and 6 to a two-mode optical channel with unequal losses and a correlated environment, determining exactly the entanglement cost for simulating such channel and its classical capacity above a finite input-energy threshold.
We further introduce a numerical package [48 ] for evaluating the entanglement G ( V ) G(V) G ( V ) of general multimode covariance matrices V V V . The package provides upper and lower bounds, whose gap quantifies the remaining optimization error on the exact value.
Entanglement cost of four-mode bound entangled states
Bound entangled states are entangled states with zero distillable entanglement: by LOCC one cannot extract maximally entangled pairs from many copies of such states at a nonzero asymptotic rate [49 , 50 ] . In the Gaussian realm, bound entangled states with a positive partial transpose (PPT) can exist when both parties hold at least two modes [51 ] , as also demonstrated experimentally in optical systems [52 ] .
Here we consider a four-mode family in which every reduction to one mode of Alice and one mode of Bob is separable. The exact entanglement cost therefore quantifies the resources needed to prepare entanglement shared by the two multimode parties, even though no individual pair of modes is entangled.
Alice and Bob each hold two modes. In the reordered coordinates ( x A 1 , x A 2 , x B 1 , x B 2 , p A 1 , p A 2 , p B 1 , p B 2 ) (x_{A_1},x_{A_2},x_{B_1},x_{B_2},p_{A_1},p_{A_2},p_{B_1},p_{B_2}) ( x A 1 , x A 2 , x B 1 , x B 2 , p A 1 , p A 2 , p B 1 , p B 2 ) , consider the covariance matrices
V a , c = X a , c ⊕ P a , c , X a , c = ( a I 2 c Z c Z a I 2 ) , V_{a,c}=X_{a,c}\oplus P_{a,c},\qquad X_{a,c}=\begin{pmatrix}a\mathbb{I}_2 & cZ\\ cZ & a\mathbb{I}_2\end{pmatrix}, V a , c = X a , c ⊕ P a , c , X a , c = ( a I 2 c Z c Z a I 2 ) ,
P a , c = ( a I 2 c M c M a I 2 ) . (73) P_{a,c}=\begin{pmatrix}a\mathbb{I}_2 & cM\\ cM & a\mathbb{I}_2\end{pmatrix}. \tag*{(73)} P a , c = ( a I 2 c M c M a I 2 ) . (73)
where Z = diag ( 1 , − 1 ) Z=\operatorname{diag}(1,-1) Z = diag ( 1 , − 1 ) , M = ( 0 1 1 0 ) M=\begin{pmatrix}0&1\\1&0\end{pmatrix} M = ( 0 1 1 0 ) , and c ≥ 0 c\ge0 c ≥ 0 . Each mode has marginal covariance matrix a I 2 a\mathbb{I}_2 a I 2 , corresponding to position and momentum ordinary variances a / 2 a/2 a /2 , while c c c controls the strength of the correlations between the parties. Position correlations connect A 1 A_1 A 1 with B 1 B_1 B 1 and A 2 A_2 A 2 with B 2 B_2 B 2 , whereas momentum correlations connect A 1 A_1 A 1 with B 2 B_2 B 2 and A 2 A_2 A 2 with B 1 B_1 B 1 .
Positive definiteness of V a , c V_{a,c} V a , c requires a > c a>c a > c , and the anticommutation relation Z M + M Z = 0 ZM+MZ=0 Z M + M Z = 0 gives the squared symplectic eigenvalues of the four-mode states as
ν ± 2 = a 2 ± c 2 a 2 − c 2 , \nu_{\pm}^{2}=a^{2}\pm c\sqrt{2a^{2}-c^{2}}, ν ± 2 = a 2 ± c 2 a 2 − c 2 ,
each with multiplicity two. The uncertainty condition ν − ≥ 1 \nu_- \ge1 ν − ≥ 1 therefore amounts to the physical region
a ≥ a 0 ( c ) : = 1 + c 2 + 2 c . (74) a \ge a_0(c) := \sqrt{1+c^2+\sqrt{2}c}. \tag*{(74)} a ≥ a 0 ( c ) := 1 + c 2 + 2 c . (74)
Partial transposition on Bob’s modes replaces c M cM c M by − c M -cM − c M without changing the symplectic spectrum, so every physical state in the family is PPT.
Separability criteria and the effect of noise on Gaussian bound entanglement have been studied in Refs. [53 , 54 , 55 ] . We now use Theorem 1 to determine the entanglement of formation and entanglement cost throughout the physical region by evaluating G ( V a , c ) G(V_{a,c}) G ( V a , c ) . This calculation identifies the bound entangled members of the family and gives the exact entanglement required to prepare them.
Optimal preparation. An optimal pure Gaussian seed can be chosen from the family
Γ ( u ) = Q ( u ) ⊕ Q ( u ) − 1 , Q ( u ) = ( cosh u I 2 sinh u 2 ( Z − M ) sinh u 2 ( Z − M ) cosh u I 2 ) . (75) \Gamma(u)=Q(u)\oplus Q(u)^{-1},
\qquad
Q(u)=
\begin{pmatrix}
\cosh u\,I_2 & \dfrac{\sinh u}{\sqrt{2}}(Z-M)\\
\dfrac{\sinh u}{\sqrt{2}}(Z-M) & \cosh u\,I_2
\end{pmatrix}. \tag*{(75)} Γ ( u ) = Q ( u ) ⊕ Q ( u ) − 1 , Q ( u ) = cosh u I 2 2 sinh u ( Z − M ) 2 sinh u ( Z − M ) cosh u I 2 . (75)
The marginal on either party has two symplectic eigenvalues equal to cosh u \cosh u cosh u . Thus, up to local Gaussian unitaries, the seed consists of two equally squeezed pairs, each with squeezing u / 2 u/2 u /2 , and its entanglement is 2 g ( sinh 2 ( u / 2 ) ) 2g(\sinh^2(u/2)) 2 g ( sinh 2 ( u /2 )) . Adding Gaussian displacements of ordinary covariance [ V a , c − Γ ( u ) ] / 2 [V_{a,c}-\Gamma(u)]/2 [ V a , c − Γ ( u )] /2 prepares the target whenever Γ ( u ) ≤ V a , c \Gamma(u)\le V_{a,c} Γ ( u ) ≤ V a , c . Since the seed entanglement increases with u ≥ 0 u\ge0 u ≥ 0 , the task is to find the smallest admissible squeezing. The solution is
u ∗ = [ atanh c 2 a − arcosh a 2 − c 2 + 1 4 a 2 − 2 c 2 ] + , (76) u_*=\left[\operatorname{atanh}\frac{c}{\sqrt{2a}}-\operatorname{arcosh}\frac{a^2-c^2+1}{\sqrt{4a^2-2c^2}}\right]_+, \tag*{(76)} u ∗ = [ atanh 2 a c − arcosh 4 a 2 − 2 c 2 a 2 − c 2 + 1 ] + , (76)
G ( V a , c ) = 2 g ( sinh 2 u ∗ 2 ) , [ x ] + : = max { x , 0 } . (77) G(V_{a,c})=2g\left(\sinh^2\frac{u_*}{2}\right),
\qquad[x]_+:=\max\{x,0\}. \tag*{(77)} G ( V a , c ) = 2 g ( sinh 2 2 u ∗ ) , [ x ] + := max { x , 0 } . (77)
Theorem 1 identifies Eq. (77) with both E F E_F E F and E C E_C E C of the Gaussian target; Figure 2 (a) shows the resulting entanglement cost across the family.
Figure 2. Entanglement cost of the four-mode Gaussian family (73). Independent Gaussian displacement noise gives a = a 0 ( c ) + t a=a_0(c)+t a = a 0 ( c ) + t , with ordinary noise covariance t I 8 / 2 t\mathbb{I}_8/2 t I 8 /2 . (a) The gradient shows E C = E F = G E_{\mathrm{C}}=E_{\mathrm{F}}=G E C = E F = G in nats, and the dashed line t = c + 1 − a 0 ( c ) t=c+1-a_0(c) t = c + 1 − a 0 ( c ) marks the separability threshold. (b) The exact cost decreases to zero along three noise curves. Every state shown is PPT and has zero distillable entanglement and logarithmic negativity; below the threshold it nevertheless requires entanglement to prepare, while all shared two-mode reductions remain separable.
The reduction to Eq. (75) is exact because target symmetries can be imposed by geometric averaging while preserving purity, feasibility, and the entanglement bound. To see this, define
J = ( 0 1 − 1 0 ) , R 0 = J ⊕ ( − J ) , W 0 = ( 0 I 2 I 2 0 ) , J=\begin{pmatrix}0&1\\-1&0\end{pmatrix},
\qquad
R_0=J\oplus(-J),
\qquad
W_0=\begin{pmatrix}0&I_2\\I_2&0\end{pmatrix}, J = ( 0 − 1 1 0 ) , R 0 = J ⊕ ( − J ) , W 0 = ( 0 I 2 I 2 0 ) ,
U = Z + M 2 , K = U ⊕ U . U=\frac{Z+M}{\sqrt{2}},
\qquad
K=U\oplus U. U = 2 Z + M , K = U ⊕ U .
The target is invariant under congruence by
Θ = I 4 ⊕ ( − I 4 ) , R = R 0 ⊕ R 0 , \Theta=I_4\oplus(-I_4),
\qquad
R=R_0\oplus R_0, Θ = I 4 ⊕ ( − I 4 ) , R = R 0 ⊕ R 0 ,
W = W 0 ⊕ W 0 , F = ( 0 K − K 0 ) . W=W_0\oplus W_0,
\qquad
F=\begin{pmatrix}0&K\\-K&0\end{pmatrix}. W = W 0 ⊕ W 0 , F = ( 0 − K K 0 ) .
Here Θ \Theta Θ represents global complex conjugation in phase space, R R R and F F F are local symplectic transformations, and W W W interchanges the parties. All preserve pure-state entanglement, and their congruence actions commute and square to the identity. The successive geometric midpoints Γ ↦ Γ # ( L Γ L T ) \Gamma\mapsto\Gamma\#(L\Gamma L^{\mathsf T}) Γ ↦ Γ# ( L Γ L T ) , for L = Θ , R , W , F L=\Theta,R,W,F L = Θ , R , W , F , therefore impose all four symmetries without increasing the objective, by Eq. (17) .
Invariance under Θ \Theta Θ gives Γ = Q ⊕ Q − 1 \Gamma=Q\oplus Q^{-1} Γ = Q ⊕ Q − 1 , and the R R R and W W W symmetries restrict Q Q Q to
Q = ( α I 2 v Z + w M v Z + w M α I 2 ) . Q=\begin{pmatrix}
\alpha I_2 & vZ+wM\\
vZ+wM & \alpha I_2
\end{pmatrix}. Q = ( α I 2 v Z + w M v Z + w M α I 2 ) .
The remaining condition Q = K Q − 1 K Q=KQ^{-1}K Q = K Q − 1 K forces α 2 − v 2 − w 2 = 1 \alpha^2-v^2-w^2=1 α 2 − v 2 − w 2 = 1 and w = − v w=-v w = − v , giving Eq. (75) . Replacing u u u by ∣ u ∣ \lvert u\rvert ∣ u ∣ leaves the entropy unchanged and improves feasibility for c ≥ 0 c\ge0 c ≥ 0 , so it suffices to use u ≥ 0 u\ge0 u ≥ 0 .
The position and momentum constraints are equivalent by the F F F symmetry and reduce to the scalar inequality
a − cosh u ≥ c 2 − 2 c sinh u + sinh 2 u . (78) a-\cosh u\ge\sqrt{c^2-\sqrt{2}c\sinh u+\sinh^2u}. \tag*{(78)} a − cosh u ≥ c 2 − 2 c sinh u + sinh 2 u . (78)
Writing β = atanh ( c / 2 a ) \beta=\operatorname{atanh}(c/\sqrt{2a}) β = atanh ( c / 2 a ) , its square becomes
4 a 2 − 2 c 2 cosh ( u − β ) ≤ a 2 − c 2 + 1. \sqrt{4a^2-2c^2}\cosh(u-\beta)\le a^2-c^2+1. 4 a 2 − 2 c 2 cosh ( u − β ) ≤ a 2 − c 2 + 1.
Physicality ensures ( a 2 − c 2 + 1 ) / 4 a 2 − 2 c 2 ≥ 1 (a^2-c^2+1)/\sqrt{4a^2-2c^2}\ge1 ( a 2 − c 2 + 1 ) / 4 a 2 − 2 c 2 ≥ 1 , so this inequality specifies an interval of allowed u u u centered at β \beta β . Its smallest nonnegative point is u ∗ u_* u ∗ in Eq. (76) , which obeys u ∗ ≤ β u_*\le\beta u ∗ ≤ β and cosh u ∗ ≤ cosh β ≤ a \cosh u_*\le\cosh\beta\le a cosh u ∗ ≤ cosh β ≤ a , so it also satisfies the un squared constraint (78) .
Irreversibility and displacement noise. The entanglement cost is positive precisely for
a 0 ( c ) ≤ a < c + 1. a_0(c)\le a<c+1. a 0 ( c ) ≤ a < c + 1.
For a ≥ c + 1 a\ge c+1 a ≥ c + 1 , the product covariance I 8 \mathbb{I}_8 I 8 is feasible and the state is separable. In the entangled region, PPT implies zero distillable entanglement and zero logarithmic negativity. Every shared two-mode marginal has correlations in at most one quadrature; it is PPT and hence separable. Equation (77) therefore quantifies an irreversible preparation cost carried exclusively by the joint four-mode correlations.
Independent Gaussian displacement noise increases a a a and leaves c c c unchanged: ordinary displacement covariance t I 8 / 2 t\mathbb{I}_8/2 t I 8 /2 gives V a , c + t I 8 = V a + t , c V_{a,c}+t\mathbb{I}_8=V_{a+t,c} V a , c + t I 8 = V a + t , c . An initially entangled member becomes separable at t = c + 1 − a t=c+1-a t = c + 1 − a . Figure 2 (b) follows these noise curves from the uncertainty boundary, setting a = a 0 ( c ) + t a=a_0(c)+t a = a 0 ( c ) + t . The boundary value also has a simple expression. Parameterizing it by
a = 1 + κ 2 1 − κ , c = 2 κ 1 − κ , 0 ≤ κ < 1 , a=\frac{\sqrt{1+\kappa^2}}{1-\kappa},
\qquad
c=\frac{\sqrt{2}\kappa}{1-\kappa},
\qquad
0\le\kappa<1, a = 1 − κ 1 + κ 2 , c = 1 − κ 2 κ , 0 ≤ κ < 1 ,
gives u ∗ = arsinh κ u_*=\operatorname{arsinh}\kappa u ∗ = arsinh κ and
G ( V a , c ) = 2 g ( 1 + κ 2 − 1 2 ) . G(V_{a,c})=2g\left(\frac{\sqrt{1+\kappa^2}-1}{2}\right). G ( V a , c ) = 2 g ( 2 1 + κ 2 − 1 ) .
Non-Gaussian states with the same cost. The entanglement cost calculated in Eq. (77) also provides a lower bound on E F E_F E F and E C E_C E C for every non-Gaussian state with covariance V a , c V_{a,c} V a , c , by Theorem 1 . To construct states attaining this bound, we retain the optimal pure Gaussian seed Γ ( u ∗ ) \Gamma(u_*) Γ ( u ∗ ) from Eq. (75) and replace the Gaussian displacement distribution by a discrete one with the same first and second moments.
In the entangled region a 0 ( c ) ≤ a < c + 1 a_0(c)\leq a<c+1 a 0 ( c ) ≤ a < c + 1 , equality in Eq. (78) implies that V a , c − Γ ( u ∗ ) V_{a,c}-\Gamma(u_*) V a , c − Γ ( u ∗ ) has rank four. We can therefore choose a real 8 × 4 8\times4 8 × 4 matrix L ∗ L_* L ∗ and define the random displacement ξ \xi ξ by
L ∗ L ∗ T = V a , c − Γ ( u ∗ ) 2 , ξ = L ∗ z , (79) L_*L_*^{\mathsf{T}}=\frac{V_{a,c}-\Gamma(u_*)}{2},\qquad\xi=L_*z, \tag*{(79)} L ∗ L ∗ T = 2 V a , c − Γ ( u ∗ ) , ξ = L ∗ z , (79)
where the four components of z z z independently take the values + 1 +1 + 1 and − 1 -1 − 1 with equal probability. Averaging over these sixteen displacements produces a non-Gaussian mixture of sixteen equally weighted pure Gaussian states. Since the displacement has zero mean and ordinary covariance L ∗ L ∗ T L_*L_*^{\mathsf{T}} L ∗ L ∗ T , the mixture has covariance V a , c V_{a,c} V a , c . Corollary 3 therefore gives E F = E C = G ( V a , c ) E_{\mathrm{F}}=E_{\mathrm{C}}=G(V_{a,c}) E F = E C = G ( V a , c ) for the resulting non-Gaussian state.
These finite mixtures are distillable. Indeed, the mixture has rank at most sixteen, whereas each displaced seed has infinite Schmidt rank for u ∗ > 0 u_*>0 u ∗ > 0 . Local projection onto seventeen matching Schmidt vectors of one component succeeds with nonzero probability and produces a state of rank at most sixteen whose two marginals each have rank seventeen. The rank criterion for distillability [56 ] then applies to the projected state, establishing distillability of the original mixture. Thus the Gaussian and non-Gaussian preparations have the same covariance and entanglement cost, but the Gaussian state is bound entangled while the non-Gaussian mixture is distillable and has a negative partial transpose.
Our second example compares the entanglement needed to simulate a noisy optical device with the classical information that it can transmit. Each input mode is mixed at a beam splitter with one of two environmental modes, which are jointly prepared in a two-mode squeezed vacuum with squeezing r ≥ 0 r\geq0 r ≥ 0 ; the beam
splitter with one of two environmental modes, which are jointly prepared in a two-mode squeezed vacuum with squeezing r ≥ 0 r\geq0 r ≥ 0 ; the beam splitters have unequal transmissivities 0 < η 1 , η 2 < 1 0<\eta_1,\eta_2<1 0 < η 1 , η 2 < 1 . The resulting channel Φ r \Phi_r Φ r retains the two signal outputs and discards the two environmental outputs. When η 1 ≠ η 2 \eta_1\ne\eta_2 η 1 = η 2 and r > 0 r>0 r > 0 , the correlated noise and unequal losses prevent a separation into independent one-mode channels. Nevertheless, symmetry reduces the simulation cost and the minimum output entropy to one-dimensional optimizations, allowing us to obtain an explicit capacity formula above a finite energy threshold.
Using Z = diag ( 1 , − 1 ) Z=\operatorname{diag}(1,-1) Z = diag ( 1 , − 1 ) , we write the pure covariance matrix of a two-mode squeezed vacuum state with squeezing t ∈ R t\in\mathbb{R} t ∈ R as
γ t = ( cosh ( 2 t ) I 2 sinh ( 2 t ) Z sinh ( 2 t ) Z cosh ( 2 t ) I 2 ) . (80) \gamma_t=
\begin{pmatrix}
\cosh(2t)\mathbb{I}_2 & \sinh(2t)Z\\
\sinh(2t)Z & \cosh(2t)\mathbb{I}_2
\end{pmatrix}. \tag*{(80)} γ t = ( cosh ( 2 t ) I 2 sinh ( 2 t ) Z sinh ( 2 t ) Z cosh ( 2 t ) I 2 ) . (80)
The environmental state has covariance γ r \gamma_r γ r , so the channel Φ r \Phi_r Φ r is described by the covariance map V ↦ X V X T + Y V\mapsto XVX^{\mathsf{T}}+Y V ↦ X V X T + Y of Eq. (32) , with
X = ⨁ j = 1 2 η j I 2 , L = ⨁ j = 1 2 1 − η j I 2 , (81) X=\bigoplus_{j=1}^{2}\sqrt{\eta_j}\mathbb{I}_2,\qquad
L=\bigoplus_{j=1}^{2}\sqrt{1-\eta_j}\mathbb{I}_2, \tag*{(81)} X = j = 1 ⨁ 2 η j I 2 , L = j = 1 ⨁ 2 1 − η j I 2 , (81)
Y = L γ r L T . Y=L\gamma_rL^{\mathsf{T}}. Y = L γ r L T .
Here X X X describes attenuation of the input quadratures, while L L L gives the contribution of the environmental quadratures to the signal outputs. The noise matrix Y Y Y therefore incorporates the correlations between the environmental modes.
For η 1 ≠ η 2 \eta_1\ne\eta_2 η 1 = η 2 and r > 0 r>0 r > 0 , the channel Φ r \Phi_r Φ r cannot be transformed into a product of one-mode channels by Gaussian input and output unitaries. To establish this, consider the matrices X Ω 2 X T Ω 2 − 1 X\Omega_2X^{\mathsf{T}}\Omega_2^{-1} X Ω 2 X T Ω 2 − 1 and Y Ω 2 − 1 Y\Omega_2^{-1} Y Ω 2 − 1 . Both undergo the same similarity transformation associated with the output unitary, so whether they commute is unchanged by Gaussian input and output unitaries. For a product of one-mode channels, both matrices are block diagonal, and the first is proportional to the identity on each one-mode block; they must therefore commute. For the channel in Eq. (81) , however, the upper-right 2 × 2 2 \times2 2 × 2 block of their commutator is
− ( η 1 − η 2 ) ( 1 − η 1 ) ( 1 − η 2 ) sinh ( 2 r ) ( 0 1 1 0 ) ≠ 0 , -(\eta_1-\eta_2)\sqrt{(1-\eta_1)(1-\eta_2)}\sinh(2r)\begin{pmatrix}0 & 1 \\ 1 & 0\end{pmatrix}\ne0, − ( η 1 − η 2 ) ( 1 − η 1 ) ( 1 − η 2 ) sinh ( 2 r ) ( 0 1 1 0 ) = 0 ,
which rules out such a product representation.
For equal losses, a balanced mixing of the input and output modes separates the channel into two independent one-mode channels. This reduction and the advantages of entangled inputs were studied in Ref. [57 ] . With unequal losses, the nonzero commutator shows that no Gaussian input and output unitaries can make both the attenuation and the noise act independently on two modes, making the problem genuinely multimode.
A common symmetry reduction. Capacity depends on keeping the entropy of the individual signal outputs low. We first minimize this entropy over all finite-energy inputs; the same symmetry argument will then reduce the probe optimization that determines the simulation cost. Put J = ( 0 1 − 1 0 ) J=\begin{pmatrix}0 & 1 \\ -1 & 0\end{pmatrix} J = ( 0 − 1 1 0 ) , K = J ⊕ ( − J ) K=J\oplus(-J) K = J ⊕ ( − J ) , and Θ = Z ⊕ Z \Theta=Z\oplus Z Θ = Z ⊕ Z . The channel and its beam-splitter dilation are covariant under the opposite phase rotation K K K and under global transposition Θ \Theta Θ . For a pure Gaussian input γ \gamma γ , the dilation gives a pure output–environment state. Its entanglement equals the output entropy, so Eq. (17) makes that entropy convex along geometric means of γ \gamma γ .
The midpoint γ # ( K γ K T ) \gamma\#(K\gamma K^{\mathsf T}) γ # ( K γ K T ) has no larger output entropy and is fixed by K K K , since K 2 = − I 4 K^2=-I_4 K 2 = − I 4 . Taking its midpoint with its Θ \Theta Θ transform also imposes Θ \Theta Θ invariance, while retaining K K K invariance because Θ K Θ = − K \Theta K\Theta=-K Θ K Θ = − K . Global transposition preserves pure-state entanglement. A covariance fixed by both actions has local blocks a I 2 aI_2 a I 2 , b I 2 bI_2 b I 2 and correlation block c Z cZ c Z ; purity forces a = b a=b a = b and a 2 − c 2 = 1 a^2-c^2=1 a 2 − c 2 = 1 . The optimum thus belongs to the family γ t \gamma_t γ t .
Define
s r ( t ) = S G ( X γ t X T + Y ) . (82) s_r(t)=S_G(X\gamma_tX^{\mathsf T}+Y). \tag*{(82)} s r ( t ) = S G ( X γ t X T + Y ) . (82)
Since γ t 0 # u γ t 1 = γ ( 1 − u ) t 0 + u t 1 \gamma_{t_0}\#_u\gamma_{t_1}=\gamma_{(1-u)t_0+ut_1} γ t 0 # u γ t 1 = γ ( 1 − u ) t 0 + u t 1 , this is an ordinary convex function of t t t . Its minimum is attained: the output covariance is bounded below by Y > 0 Y>0 Y > 0 and has an eigenvalue tending to infinity as ∣ t ∣ → ∞ |t|\to\infty ∣ t ∣ → ∞ , so S G ( V ) ≥ 1 2 log det V S_G(V)\ge\frac{1}{2}\log\det V S G ( V ) ≥ 2 1 log det V gives diverging entropy. For every finite-energy input ρ \rho ρ , Eq. (69) gives
S ( Φ r ( ρ ) ) ≥ S ^ Φ r ( ρ ) ≥ μ Φ r ( V ρ ) ≥ min t s r ( t ) . S(\Phi_r(\rho))\ge\widehat{S}_{\Phi_r}(\rho)\ge\mu_{\Phi_r}(V_\rho)\ge\min_t s_r(t). S ( Φ r ( ρ )) ≥ S Φ r ( ρ ) ≥ μ Φ r ( V ρ ) ≥ t min s r ( t ) .
A minimizing γ t \gamma_t γ t attains the last value, which is therefore the minimum output entropy over all finite-energy inputs.
Exact entanglement cost. For simulation, the optimized entropy is that of one party’s share of a pure teleportation resource. This leads to a second scalar minimization. Define
W r ( t ) = L − 1 ( γ t + X γ r X T ) L − T . (83) W_r(t)=L^{-1}(\gamma_t+X\gamma_rX^{\mathsf T})L^{-\mathsf T}. \tag*{(83)} W r ( t ) = L − 1 ( γ t + X γ r X T ) L − T . (83)
Then
E C ( Φ r ) = min t ∈ R S G ( W r ( t ) ) . (84) E_C(\Phi_r)=\min_{t\in\mathbb{R}}S_G(W_r(t)). \tag*{(84)} E C ( Φ r ) = t ∈ R min S G ( W r ( t )) . (84)
For the upper bound, we construct a finite-energy resource that implements the channel exactly. Let Z 2 = Z ⊕ Z Z_2=Z\oplus Z Z 2 = Z ⊕ Z , A η = L − 1 A_\eta=L^{-1} A η = L − 1 , B η = L − 1 X B_\eta=L^{-1}X B η = L − 1 X , and
Ξ η = ( A η B η Z B η Z A η ) . \Xi_\eta=\begin{pmatrix}A_\eta& B_\eta Z\\ B_\eta Z & A_\eta\end{pmatrix}. Ξ η = ( A η B η Z B η Z A η ) .
The identity A η 2 − B η 2 = I 4 A_\eta^2-B_\eta^2=I_4 A η 2 − B η 2 = I 4 makes Ξ η \Xi_\eta Ξ η symplectic. The pure resource covariance matrix Ξ η ( γ t ⊕ γ r ) Ξ η T \Xi_\eta(\gamma_t\oplus\gamma_r)\Xi_\eta^{\mathsf T} Ξ η ( γ t ⊕ γ r ) Ξ η T has Alice’s marginal W r ( t ) W_r(t) W r ( t ) . Moreover, the teleportation noise matrix M t e l = ( − X Z 2 I 4 ) M_{\mathrm{tel}}=(-XZ_2\ \ I_4) M tel = ( − X Z 2 I 4 ) obeys
M t e l Ξ η = ( 0 L ) , M_{\mathrm{tel}}\Xi_\eta=(0\ \ L), M tel Ξ η = ( 0 L ) ,
so this resource implements Φ r \Phi_r Φ r exactly. Its entanglement is convex in t t t by Eq. (17) ; a uniform positive lower eigenvalue of W r ( t ) W_r(t) W r ( t ) and a diverging largest eigenvalue make the minimum attained.
For the matching converse, Gaussian conditioning relates the probe entanglement in Theorem 4 to a pure Gaussian measurement on its purifying environment. The same two symmetries restrict this measurement to the family γ t \gamma_t γ t . As the probe squeezing grows, the conditional entropy tends to S G ( W r ( t ) ) S_G(W_r(t)) S G ( W r ( t )) ; bounded-entropy minimizers remain in a compact interval of t t t , so the limit gives the reverse inequality in Eq. (84) . Appendix E4 provides the conditioning formula and the limit argument. The resulting entanglement cost is illustrated in Figure 3 (a).
Figure 3. Resources and communication rates of the correlated two-mode attenuator (81) , with η 1 = 0.55 \eta_1 = 0.55 η 1 = 0.55 and η 2 = 0.8 \eta_2 = 0.8 η 2 = 0.8 , as the environmental squeezing r r r increases. (a) Entanglement cost of simulating the channel, from Eq. (84) . (b) Classical capacity at total input photon budget E = 20 E = 20 E = 20 , compared with the optimal rate for product Gaussian signals across the physical input ports. The modulation conditions for both formulas hold throughout the plotted range. Both encoding classes allow collective decoding, and the shaded gap is the benefit of entangling the input ports within each use. Rates are in nats per use of the two-mode device. At r = 0 r = 0 r = 0 , the channel is a product of two pure-loss channels.
Capacity and entanglement between the input ports. The capacity calculation balances two entropy requirements: each signal should produce as little output entropy as possible, while the average output should have as much entropy as the energy budget permits. A signal seed minimizing s r ( t ) s_r(t) s r ( t ) meets the first requirement. At sufficiently large input energy, correlated classical displacements also make the average output a product of thermal states with the optimal occupations, so one ensemble attains both extrema.
Let t ∗ t_* t ∗ minimize s r ( t ) s_r(t) s r ( t ) and impose total input photon budget E > 0 E>0 E > 0 for H = N 1 + N 2 H=N_1+N_2 H = N 1 + N 2 , where N j = a j † a j N_j=a_j^\dagger a_j N j = a j † a j . Put n e = sinh 2 r n_e=\sinh^2r n e = sinh 2 r . Every output obeys
∑ j ⟨ N j ⟩ η j ≤ E + n e ∑ j 1 − η j η j . \sum_j\frac{\langle N_j\rangle}{\eta_j}\le E+n_e\sum_j\frac{1-\eta_j}{\eta_j}. j ∑ η j ⟨ N j ⟩ ≤ E + n e j ∑ η j 1 − η j .
The Gibbs maximum-entropy principle determines the optimal average output occupations from the unique λ > 0 \lambda>0 λ > 0 satisfying
n ˉ j = 1 e λ / η j − 1 , ∑ j n ˉ j η j = E + n e ∑ j 1 − η j η j . (85) \bar n_j=\frac{1}{e^{\lambda/\eta_j}-1},\qquad
\sum_j\frac{\bar n_j}{\eta_j}=E+n_e\sum_j\frac{1-\eta_j}{\eta_j}. \tag*{(85)} n ˉ j = e λ / η j − 1 1 , j ∑ η j n ˉ j = E + n e j ∑ η j 1 − η j . (85)
Its covariance is ⨁ j ( 2 n ˉ j + 1 ) I 2 \bigoplus_j(2\bar n_j+1)I_2 ⨁ j ( 2 n ˉ j + 1 ) I 2 , with inverse-image input covariance
V ∗ = X − 1 [ ⨁ j ( 2 n ˉ j + 1 ) I 2 − Y ] X − T . V_* = X^{-1}\left[\bigoplus_j(2\bar n_j+1)I_2-Y\right]X^{-\mathsf T}. V ∗ = X − 1 [ j ⨁ ( 2 n ˉ j + 1 ) I 2 − Y ] X − T .
The condition V ∗ ⪰ γ t ∗ V_*\succeq\gamma_{t_*} V ∗ ⪰ γ t ∗ is precisely the requirement that this average can be prepared by displacing the optimal seed. When it holds, modulation with ordinary covariance ( V ∗ − γ t ∗ ) / 2 (V_* - \gamma_{t_*})/2 ( V ∗ − γ t ∗ ) /2 attains both entropy extrema, giving
C ( Φ r , H , E ) = ∑ j g ( n ˉ j ) − s r ( t ∗ ) . (86) C(\Phi_r,H,E)=\sum_j g(\bar n_j)-s_r(t_*). \tag*{(86)} C ( Φ r , H , E ) = j ∑ g ( n ˉ j ) − s r ( t ∗ ) . (86)
The condition holds above a finite energy threshold: the offdiagonal block of V ∗ V_* V ∗ is independent of E E E , whereas its diagonal blocks diverge as E → ∞ E \to\infty E → ∞ . At lower energies, Theorem 6 still determines the capacity through the constrained covariance optimization.
We can now quantify the benefit of entangling the two physical input ports. As a benchmark, restrict each Gaussian signal to be a product across the ports, while allowing classical correlations in the modulation, correlations across uses within either port, and collective decoding. For a pure one-mode covariance A A A , J A J T = A − 1 JAJ^{\mathsf T}=A^{-1} J A J T = A − 1 , so the first symmetry midpoint of a product seed is vacuum. Its minimum output entropy is therefore s r ( 0 ) s_r(0) s r ( 0 ) . The same argument includes block codes: a signal has the form ρ 1 n ⊗ ρ 2 n \rho_1^n\otimes\rho_2^n ρ 1 n ⊗ ρ 2 n , where either Gaussian port state may correlate its n n n uses. After refining to pure Gaussian factors, the simultaneous opposite rotations send each port covariance A A A to Ω n A Ω n T = A − 1 \Omega_n A\Omega_n^{\mathsf T}=A^{-1} Ω n A Ω n T = A − 1 . The midpoint of the full product covariance and its transform is I 4 n I_{4n} I 4 n , giving output entropy at least n s r ( 0 ) ns_r(0) n s r ( 0 ) . Entropy concavity extends this bound to mixed factors and classical mixtures. Thus, when V ∗ ≥ I 4 V_* \ge I_4 V ∗ ≥ I 4 , the optimal rate for this class is
C p r o d , G = ∑ j g ( n ˉ j ) − s r ( 0 ) . (87) C_{\mathrm{prod,G}}=\sum_j g(\bar n_j)-s_r(0). \tag*{(87)} C prod , G = j ∑ g ( n ˉ j ) − s r ( 0 ) . (87)
Correlated modulation and time sharing are included in the same average-output entropy bound.
When both modulation conditions hold, the advantage is strictly positive for r > 0 r > 0 r > 0 . The symplectic eigenvalues of a covariance matrix with blocks a I 2 aI_2 a I 2 , b I 2 bI_2 b I 2 , c Z cZ c Z are [ ( a + b ) 2 − 4 c 2 ± [\sqrt{(a+b)^2-4c^2}\pm [ ( a + b ) 2 − 4 c 2 ±
( a − b ) ] / 2 (a-b)]/2 ( a − b )] /2 . Writing the blocks of X γ t X T + Y X\gamma_tX^{\mathsf T}+Y X γ t X T + Y as a ( t ) I 2 a(t)I_2 a ( t ) I 2 , b ( t ) I 2 b(t)I_2 b ( t ) I 2 , and c ( t ) Z c(t)Z c ( t ) Z , one has a ′ ( 0 ) = b ′ ( 0 ) = 0 a'(0)=b'(0)=0 a ′ ( 0 ) = b ′ ( 0 ) = 0 , c ′ ( 0 ) = 2 η 1 η 2 > 0 c'(0)=2\sqrt{\eta_1\eta_2}>0 c ′ ( 0 ) = 2 η 1 η 2 > 0 , and c ( 0 ) > 0 c(0)>0 c ( 0 ) > 0 . Both symplectic eigenvalues decrease at t = 0 t=0 t = 0 and have negative derivatives at t = 0 t=0 t = 0 . Hence s r ′ ( 0 ) < 0 s_r'(0)<0 s r ′ ( 0 ) < 0 and
C ( Φ r , H , E ) − C p r o d , G = s r ( 0 ) − s r ( t ∗ ) > 0. C(\Phi_r,H,E)-C_{\mathrm{prod,G}}=s_r(0)-s_r(t_*)>0. C ( Φ r , H , E ) − C prod , G = s r ( 0 ) − s r ( t ∗ ) > 0.
The gap is an advantage from entangling modes within one use of the device; independent uses of the optimal ensemble still attain the unrestricted capacity. Figure 3 compares this communication advantage with the entanglement cost of simulating the same channel.
Numerical evaluation with upper and lower bounds
The analytic examples presented earlier in this Section exploit special symmetries, whereas the evaluation of the quantities in Eq. (1) for arbitrary multimode covariance matrices generally requires numerical optimization. Here we provide a standalone Python toolkit [48 ] that takes a physical covariance matrix and a mode bipartition, searches for an optimal pure Gaussian preparation, and evaluates upper and lower bounds on G ( V ) G(V) G ( V ) . For a Gaussian state, these bounds apply to both its entanglement of formation and its entanglement cost. For any non-Gaussian state with the same covariance, the lower endpoint remains a lower bound on both quantities.
The covariance formulation replaces an optimization over ensembles by a finite matrix problem [5]. Existing approaches include optimization on pure Gaussian manifolds [58 ] and conic upper bounds [59 ] . Here geodesic convexity also controls the error of a candidate solution. A feasible pure seed Γ 0 ≤ V \Gamma_0 \le V Γ 0 ≤ V specifies a preparation and gives the upper bound G ( V ) ≤ E ( Γ 0 ) G(V) \le E(\Gamma_0) G ( V ) ≤ E ( Γ 0 ) , while its first variation bounds how much the entanglement can decrease elsewhere in the feasible set.
For k k k total modes, define the normalized target covariance M 0 M_0 M 0 and a radius d V d_V d V that bounds the distance of every feasible pure seed from the candidate:
M 0 = Γ 0 − 1 / 2 V Γ 0 − 1 / 2 , d V = 2 k log λ max ( M 0 ) . (88) M_0=\Gamma_0^{-1/2}V\Gamma_0^{-1/2},\qquad d_V=\sqrt{2k}\log\lambda_{\max}(M_0). \tag*{(88)} M 0 = Γ 0 − 1/2 V Γ 0 − 1/2 , d V = 2 k log λ m a x ( M 0 ) . (88)
Let K 0 K_0 K 0 represent the first variation along pure Gaussian paths,
d d t E ( Γ 0 1 / 2 e t Q Γ 0 1 / 2 ) ∣ t = 0 = Tr ( K 0 Q ) , \left.\frac{d}{dt}E\left(\Gamma_0^{1/2}e^{tQ}\Gamma_0^{1/2}\right)\right|_{t=0}=\operatorname{Tr}(K_0Q), d t d E ( Γ 0 1/2 e tQ Γ 0 1/2 ) t = 0 = Tr ( K 0 Q ) ,
where Q = Q T Q=Q^{\mathsf{T}} Q = Q T and Q Ω + Ω Q = 0 Q\Omega+\Omega Q=0 Q Ω + Ω Q = 0 . The orthogonal projection onto this tangent space is T Ω ( A ) = ( A + Ω A Ω ) / 2 T_{\Omega}(A)=(A+\Omega A\Omega)/2 T Ω ( A ) = ( A + Ω A Ω ) /2 ; Appendix A.2 gives the analytic derivative. A positive semidefinite matrix Λ \Lambda Λ acts as a multiplier for the covariance constraint in these normalized coordinates. It gives the error bound
ε ( Λ ) = Tr [ Λ ( M 0 − I ) ] + d V ∥ T Ω ( K 0 + Λ ) ∥ F , (89) \varepsilon(\Lambda)=\operatorname{Tr}[\Lambda(M_0-\mathbb{I})]+d_V\|T_{\Omega}(K_0+\Lambda)\|_{\mathrm{F}}, \tag*{(89)} ε ( Λ ) = Tr [ Λ ( M 0 − I )] + d V ∥ T Ω ( K 0 + Λ ) ∥ F , (89)
where ∥ ⋅ ∥ F \|\cdot\|_{\mathrm{F}} ∥ ⋅ ∥ F is the Frobenius norm, and hence the interval
max { 0 , E ( Γ 0 ) − ε ( Λ ) } ≤ G ( V ) ≤ E ( Γ 0 ) . (90) \max\{0,E(\Gamma_0)-\varepsilon(\Lambda)\}\leq G(V)\leq E(\Gamma_0). \tag*{(90)} max { 0 , E ( Γ 0 ) − ε ( Λ )} ≤ G ( V ) ≤ E ( Γ 0 ) . (90)
The first term of ε \varepsilon ε measures slack in the covariance constraint, while the second measures the remaining first-order variation. Their sum controls the possible improvement over the candidate throughout the feasible set.
Indeed, any feasible pure Γ \Gamma Γ gives a positive symplectic matrix C = Γ 0 − 1 / 2 Γ Γ 0 − 1 / 2 ≤ M 0 C=\Gamma_0^{-1/2}\Gamma\Gamma_0^{-1/2}\leq M_0 C = Γ 0 − 1/2 Γ Γ 0 − 1/2 ≤ M 0 . Its ordinary eigenvalues occur in reciprocal pairs, so ∥ log C ∥ F ≤ d V \|\log C\|_{\mathrm{F}}\leq d_V ∥ log C ∥ F ≤ d V . Apply the first-order geodesic-convexity inequality to E ( Γ ) + Tr ( W Γ ) / 4 E(\Gamma)+\operatorname{Tr}(W\Gamma)/4 E ( Γ ) + Tr ( W Γ ) /4 , with W = 4 Γ 0 − 1 / 2 Λ Γ 0 − 1 / 2 W=4\Gamma_0^{-1/2}\Lambda\Gamma_0^{-1/2} W = 4 Γ 0 − 1/2 Λ Γ 0 − 1/2 . Covariance order bounds the quadratic term, and Cauchy–Schwarz bounds the tangent term, giving Eq. (90) .
For a fixed candidate, minimizing ε ( Λ ) \varepsilon(\Lambda) ε ( Λ ) improves the lower endpoint through a semidefinite program. If v v v stacks the columns of T Ω ( K 0 + Λ ) T_{\Omega}(K_0+\Lambda) T Ω ( K 0 + Λ ) , introduce a scalar t t t and impose
Λ ≥ 0 , ( t v T v t I ) ≥ 0. \Lambda\geq0,\qquad
\begin{pmatrix}
t & v^{\mathsf{T}}\\
v & t\mathbb{I}
\end{pmatrix}\geq0. Λ ≥ 0 , ( t v v T t I ) ≥ 0.
Minimizing Tr [ Λ ( M 0 − I ) ] + d V t \operatorname{Tr}[\Lambda(M_0-\mathbb{I})]+d_Vt Tr [ Λ ( M 0 − I )] + d V t gives the best bound of this form, while any feasible weight already gives a valid lower bound. Thus the pure-seed search is a nonlinear matrix optimization, followed by a convex optimization that bounds how far its entanglement lies above the global minimum.
The implementation uses Williamson coordinates to separate mixed normal modes from pure ones. With the mixed modes ordered first, it writes a candidate as Ξ ( e Q ⊕ I ) Ξ T \Xi(e^Q\oplus\mathbb{I})\Xi^{\mathsf{T}} Ξ ( e Q ⊕ I ) Ξ T , where Q Q Q is symmetric and anticommutes with the corresponding symplectic form. This parameterization enforces purity, while the remaining normal modes are fixed to vacuum by the covariance constraint. The search then imposes covariance order and evaluates Eq. (90) . The package evaluates these bounds in double precision and returns the candidate seed, both endpoints, and their gap, together with purity and covariance-order residuals. The same covariance minimum can be used in the channel formulas of Secs. III and IV.
Conclusions and outlook
In this paper we have resolved the Gaussian optimality problem for entanglement of formation on arbitrary multimode bipartitions, listed as Problem 29 in the Open Quantum Problems collection [60 ] . Theorem 1 proves that Gaussian entanglement of formation, unrestricted entanglement of formation, and entanglement cost coincide for every finite-energy Gaussian state, extending the two-mode and bisymmetric results of Ref. [18] to general bipartitions. The supporting entropy inequality underpinning Theorem 1 also enables us to determine the entanglement cost of simulating multimode Gaussian channels (Theorem 4) and their energy-constrained classical capacity (Theorem 6). These channel results extend earlier single-mode solutions [21 ] [17, 18] to correlated multimode devices, including those that cannot be reduced to independent single-mode channels by Gaussian input and output unitaries.
The additivity established here contrasts with the general quantum setting, where entanglement of formation can be non-additive and entangled inputs across channel uses can increase classical communication rates [61 , 62 ] . Here the asymptotic quantities are determined by covariance optimizations for a single state or channel use, together with a squeezing limit for channel simulation. This reduction holds even though the underlying operational problems allow non-Gaussian states and arbitrary correlations across repetitions. The Gaussian description therefore determines the unrestricted preparation costs and communication rates, rather than merely the performance of Gaussian protocols.
The proofs reveal a useful and insightful geometry of pure Gaussian preparations. Their covariances can be interpolated along paths that preserve purity and the target covariance constraint, while the entanglement entropy is convex along those paths. This allows the symmetries used in the state and channel examples of § 1.4 A and V B—including the exact determination of the entanglement cost of a family of four-mode bound entangled states with all separable two-mode marginals—to reduce the optimization over pure Gaussian preparations. Supporting quadratic observables connect this geometry to entropy inequalities for arbitrary states and provide global lower bounds on the Gaussian entanglement of formation. Comparing these bounds with the entanglement of a feasible preparation quantifies the remaining optimization error, as implemented in the numerical toolkit of § 1.4 C. The same structure offers a route to further exactly solvable entanglement quantification in relevant multimode families and to quantitative resource estimates when no symmetry reduction is available.
Our results establish a general operational instance of Gaussian extremality [20] for entanglement preparation: among all states with the same second moments, the Gaussian state minimizes both entanglement of formation and entanglement cost. For non-Gaussian sources, Eq. (12) therefore gives lower bounds on both entanglement quantifiers from measured second moments alone. This makes it possible to estimate entanglement experimentally without full state tomography, extending the two-mode applications of Ref. [18] to arbitrary multimode bipartitions. The bounds apply in particular to the non-Gaussian optical states produced by conditional measurements, including photon addition and subtraction [63 ] . For larger systems, bounds obtained from disjoint groups of modes can be added to lower-bound the entanglement across the full bipartition, reducing the number of correlations that must be measured and the size of the covariance optimizations.
Corollary 3 also identifies non-Gaussian families for which the covariance bound is exact. Replacing the Gaussian displacement distribution of an optimal preparation by a non-Gaussian distribution with the same first and second moments produces states with exactly known entanglement of formation and entanglement cost.
For quantum communication, our multimode results address Gaussian devices whose loss and noise cannot be separated into independent one-mode channels. Theorem 6 reduces their classical capacity to a covariance optimization for one use of the full device. The unequal-loss example in § 1.4 B gives an advantage for an entangled Gaussian seed over Gaussian signals that are product across the physical ports, while independent uses of the optimal ensemble attain the full capacity with collective decoding. The optimization therefore identifies how squeezing, correlations among input modes, and classical modulation should be matched to the device’s noise and energy budget.
This description is increasingly accessible experimentally: recent work has reconstructed the full transformation and noise matrices of 16-mode Gaussian optical processes, including processes combining mode-dependent loss with parametric interactions [64 ] . Such characterizations provide precisely the matrices entering our capacity and simulation cost formulas. For non-Gaussian devices, Corollary 5 still gives lower bounds on the entanglement cost from the covariances of finite-energy probe outputs, extending the resource assessment beyond devices described by a Gaussian model.
A central next step is to approach these capacities with practical collective receivers, building on progress in joint optical detection [65 , 66 ] . Quantifying achievable rates at finite code length would connect the optimal encodings to realistic decoding resources, while characterizing the tradeoff among entanglement, squeezing, and simulation accuracy would extend existing finite-resource teleportation methods to coupled multimode devices [39 , 40 ] .
By determining optimal encodings and exact channel simulation costs, the results in this paper establish a foundation for designing communication protocols that optimally exploit correlations among modes. The challenge now is to translate this control of multimode quantum resources into practical devices that approach their fundamental transmission limits.
Note added. Parts of this work build on results first reported in preprints by one of the authors [67 , 68 ] . Toward completion of this manuscript, we became aware of He’s preprint [69 ] , which also gives a general multimode classical capacity formula as the zero-assistance case of an energy-entanglement tradeoff. We are also aware of related results in preparation by Shao-Hua Hu and Ludovico Lami.
Acknowledgments
Artificial intelligence tools based on OpenAI and Anthropic models assisted with mathematical development, literature searches, numerical methods, analytic checks, and manuscript preparation. The authors are responsible for the results and the final manuscript. GA acknowledges financial support from the Engineering and Physical Sciences Research Council (EPSRC Grant No. UKRI6536).
Data availability
The data that support the findings of this study are available from the authors upon reasonable request. The numerical toolkit, including usage instructions and example calculations, is provided as accompanying software [48 ] .
Spectral properties of the covariance minimum
Relative spectra under geometric means
The geodesic convexity in § 1.1 C follows from a comparison of relative spectra. Let A 0 , A 1 , B 0 , B 1 A_0,A_1,B_0,B_1 A 0 , A 1 , B 0 , B 1 be positive-definite matrices, and for 0 ≤ t ≤ 1 0 \le t \le1 0 ≤ t ≤ 1 set A t = A 0 # t A 1 A_t=A_0\#_t A_1 A t = A 0 # t A 1 and B t = B 0 # t B 1 B_t=B_0\#_t B_1 B t = B 0 # t B 1 . Let x ( t ) x(t) x ( t ) be the decreasing vector of log-eigenvalues of A t − 1 / 2 B t A t − 1 / 2 A_t^{-1/2}B_tA_t^{-1/2} A t − 1/2 B t A t − 1/2 . Then
x ( t ) ≺ ( 1 − t ) x ( 0 ) + t x ( 1 ) . (91) x(t) \prec(1-t)x(0)+tx(1). \tag*{(91)} x ( t ) ≺ ( 1 − t ) x ( 0 ) + t x ( 1 ) . (91)
Here x ≺ y x \prec y x ≺ y denotes vector majorization: the sum of the largest j j j entries of x x x is at most that for y y y for every j j j , with equality for the full sum.
To bound the largest k k k relative eigenvalues at once, apply the k k k th exterior-power representation. Its eigenvalues are products of k k k original eigenvalues, and it commutes with multiplication, inversion, positive powers, and hence the geometric mean. The endpoint inequalities
λ k B j ≤ exp ( ∑ i = 1 k χ i ( j ) ) λ k A j \lambda^k B_j \le\exp\left(\sum_{i=1}^{k} \chi_i(j)\right)\lambda^k A_j λ k B j ≤ exp ( i = 1 ∑ k χ i ( j ) ) λ k A j
interpolate by joint monotonicity and the homogeneity ( a X ) # t ( b Y ) = a 1 − t b t X # t Y (aX)\#_t(bY)=a^{1-t}b^tX\#_tY ( a X ) # t ( bY ) = a 1 − t b t X # t Y . They give every partial-sum inequality in Eq. (91) ; the identity det ( X # t Y ) = ( det X ) 1 − t ( det Y ) t \det(X\#_tY)=(\det X)^{1-t}(\det Y)^t det ( X # t Y ) = ( det X ) 1 − t ( det Y ) t gives equality for the full sum. Karamata’s inequality and convexity therefore imply, for every convex scalar function F F F ,
∑ i F ( x i ( t ) ) ≤ ( 1 − t ) ∑ i F ( x i ( 0 ) ) + t ∑ i F ( x i ( 1 ) ) . (92) \sum_i F(x_i(t)) \le(1-t)\sum_i F(x_i(0))+t\sum_i F(x_i(1)). \tag*{(92)} i ∑ F ( x i ( t )) ≤ ( 1 − t ) i ∑ F ( x i ( 0 )) + t i ∑ F ( x i ( 1 )) . (92)
For the entropy function in Eq. (16) , convexity follows from the scalar derivatives. Put f s q ( r ) = g ( sinh 2 r ) f_{\mathrm{sq}}(r)=g(\sinh^2 r) f sq ( r ) = g ( sinh 2 r ) for r > 0 r>0 r > 0 . Then
f s q ′ ( r ) = 2 sinh ( 2 r ) log coth r , f'_{\mathrm{sq}}(r)=2\sinh(2r)\log\coth r, f sq ′ ( r ) = 2 sinh ( 2 r ) log coth r ,
f s q ′ ′ ( r ) = 4 cosh ( 2 r ) log coth r − 4 > 0. f''_{\mathrm{sq}}(r)=4\cosh(2r)\log\coth r-4>0. f sq ′′ ( r ) = 4 cosh ( 2 r ) log coth r − 4 > 0.
The last inequality is equivalent to ( z + z − 1 ) atanh z > 1 (z+z^{-1})\operatorname{atanh}z>1 ( z + z − 1 ) atanh z > 1 for z = e − 2 r ∈ ( 0 , 1 ) z=e^{-2r}\in(0,1) z = e − 2 r ∈ ( 0 , 1 ) , which holds because atanh z > z \operatorname{atanh}z>z atanh z > z . Also f s q ′ ( 0 + ) = 0 f'_{\mathrm{sq}}(0+)=0 f sq ′ ( 0 + ) = 0 , so its even extension is continuously differentiable. Congruence by Π \Pi Π commutes with geometric means, so Π Γ t Π = ( Π Γ 0 Π ) # t ( Π Γ 1 Π ) \Pi\Gamma_t\Pi=(\Pi\Gamma_0\Pi)\#_t(\Pi\Gamma_1\Pi) Π Γ t Π = ( Π Γ 0 Π ) # t ( Π Γ 1 Π ) . Equation (A2), applied to ( A j , B j ) = ( Γ j , Π Γ j Π ) (A_j,B_j)=(\Gamma_j,\Pi\Gamma_j\Pi) ( A j , B j ) = ( Γ j , Π Γ j Π ) , therefore gives Eq. (17) .
Derivative of pure Gaussian entanglement
The spectral formula (16) gives the analytic derivative used in Sec. V C. Put A 0 = Γ 0 − 1 / 2 A_0=\Gamma_0^{-1/2} A 0 = Γ 0 − 1/2 , R Γ 0 = A 0 Π Γ 0 Π A 0 R_{\Gamma_0}=A_0\Pi\Gamma_0\Pi A_0 R Γ 0 = A 0 Π Γ 0 Π A 0 , and f ( x ) = F ( log x ) f(x)=F(\log x) f ( x ) = F ( log x ) . Differentiating the trace of a function of the generalized eigenvalues of ( Π Γ Π , Γ ) (\Pi\Gamma\Pi,\Gamma) ( ΠΓΠ , Γ ) gives a full-matrix gradient extending the pure-state differential:
D E = 1 4 [ Π A 0 f ′ ( R Γ 0 ) A 0 Π − A 0 R Γ 0 f ′ ( R Γ 0 ) A 0 ] , K 0 = Γ 0 1 / 2 D E Γ 0 1 / 2 . (93) D_E=\frac{1}{4}[\Pi A_0f'(R_{\Gamma_0})A_0\Pi-A_0R_{\Gamma_0}f'(R_{\Gamma_0})A_0],
\qquad
K_0=\Gamma_0^{1/2}D_E\Gamma_0^{1/2}. \tag*{(93)} D E = 4 1 [ Π A 0 f ′ ( R Γ 0 ) A 0 Π − A 0 R Γ 0 f ′ ( R Γ 0 ) A 0 ] , K 0 = Γ 0 1/2 D E Γ 0 1/2 . (93)
Here f ′ ( x ) = F ′ ( log x ) / x f'(x)=F'(\log x)/x f ′ ( x ) = F ′ ( log x ) / x , with
F ′ ( x ) = 1 2 sinh ( x / 2 ) log coth ( ∣ x ∣ / 4 ) , F ′ ( 0 ) = 0. F'(x)=\frac{1}{2}\sinh(x/2)\log\coth(|x|/4),
\qquad
F'(0)=0. F ′ ( x ) = 2 1 sinh ( x /2 ) log coth ( ∣ x ∣/4 ) , F ′ ( 0 ) = 0.
The value F ′ ( 0 ) = 0 F'(0)=0 F ′ ( 0 ) = 0 includes local pure modes. The tangent projection T Ω ( K 0 ) T_{\Omega}(K_0) T Ω ( K 0 ) determines the variation on pure covariances, independently of the chosen extension to the surrounding matrix space.
Quadratic reduction of local pure modes
The supporting quadratic may couple entangled Schmidt pairs to local pure modes. Completing its square separates these modes into a nonnegative remainder and leaves a stationary quadratic support on the paired modes. This gives the decomposition (26) used in Sec. II E, including its additive constant. Put the optimal pure covariance in Gaussian Schmidt form, with n p n_p n p entangled pairs. Let a A j , a B j a_{A j},a_{B j} a A j , a B j be the annihilation operators in these coordinates, and set q j = tanh r j ∈ ( 0 , 1 ) q_j=\tanh r_j\in(0,1) q j = tanh r j ∈ ( 0 , 1 ) for the pair squeezing r j r_j r j , j = 1 , … , n p j=1,\ldots,n_p j = 1 , … , n p . Define
c A j = a A j − q j a B j † 1 − q j 2 , c B j = a B j − q j a A j † 1 − q j 2 , (94) c_{A j}=\frac{a_{A j}-q_j a_{B j}^{\dagger}}{\sqrt{1-q_j^2}},
\qquad
c_{B j}=\frac{a_{B j}-q_j a_{A j}^{\dagger}}{\sqrt{1-q_j^2}}, \tag*{(94)} c A j = 1 − q j 2 a A j − q j a B j † , c B j = 1 − q j 2 a B j − q j a A j † , (94)
α j = − 2 log q j , n j = q j 2 1 − q j 2 . \alpha_j=-2\log q_j,
\qquad
n_j=\frac{q_j^2}{1-q_j^2}. α j = − 2 log q j , n j = 1 − q j 2 q j 2 .
Collect the paired annihilators in a vector c p c_p c p , with all Alice labels before all Bob labels, and the local vacuum annihilators in c u c_u c u ; write c = ( c p T , c u T ) T c=(c_p^{\mathsf T},c_u^{\mathsf T})^{\mathsf T} c = ( c p T , c u T ) T . The normalized full Gaussian seed ∣ ψ Γ ∗ ⟩ |\psi_{\Gamma_*}\rangle ∣ ψ Γ ∗ ⟩ is characterized by c j ∣ ψ Γ ∗ ⟩ = 0 c_j|\psi_{\Gamma_*}\rangle=0 c j ∣ ψ Γ ∗ ⟩ = 0 , and its reduction to the paired modes is the state ω c \omega_c ω c defined in Sec. II E. Thus E ∗ = S ( ( ω c ) A ) = ∑ j = 1 n p g ( n j ) E_* = S((\omega_c)_A)=\sum_{j=1}^{n_p}g(n_j) E ∗ = S (( ω c ) A ) = ∑ j = 1 n p g ( n j ) . Let H A = ∑ j α j a A j † a A j H_A=\sum_j\alpha_j a_{A j}^{\dagger}a_{A j} H A = ∑ j α j a A j † a A j be the thermal Hamiltonian of the paired Alice marginal, with zero coefficient on her unpaired modes; the sum is zero if n p = 0 n_p=0 n p = 0 .
The first variation of E E E at the minimizer equals that of ⟨ H A ⟩ \langle H_A\rangle ⟨ H A ⟩ . For positive marginal occupations this is the thermal entropy derivative. For a vacuum occupation, any analytic two-sided covariance path has occupation O ( t 2 ) O(t^2) O ( t 2 ) : the eigenvalue branches of i Γ A ( t ) 1 / 2 Ω A Γ A ( t ) 1 / 2 i\Gamma_A(t)^{1/2}\Omega_A\Gamma_A(t)^{1/2} i Γ A ( t ) 1/2 Ω A Γ A ( t ) 1/2 are analytic, and uncertainty for both signs of t t t forces their first derivatives at 1 1 1 to vanish. Their entropy contribution is O ( t 2 ∣ log ∣ t ∣ ∣ ) = o ( t ) O(t^2|\log|t||)=o(t) O ( t 2 ∣ log ∣ t ∣∣ ) = o ( t ) . The vacuum modes therefore contribute zero to the first variation.
Stationarity at Γ ∗ \Gamma_* Γ ∗ of the functional on the left-hand side of Eq. (21) under all squeezing variations eliminates terms with two creation or two annihilation operators. In the c c c coordinates,
H W + H A = c † K c + C 0 , K ≥ 0 , C 0 = Tr K − ∑ j α j 2 . (95) H_W+H_A=c^\dagger Kc+C_0,\qquad K\geq0,\qquad C_0=\frac{\operatorname{Tr}K-\sum_j\alpha_j}{2}. \tag*{(95)} H W + H A = c † K c + C 0 , K ≥ 0 , C 0 = 2 Tr K − ∑ j α j . (95)
The sum H W + H A H_W+H_A H W + H A is bounded below, so evaluation on coherent states forces K ≥ 0 K\geq0 K ≥ 0 ; the constant follows from symmetric ordering. In particular ⟨ H W ⟩ ψ Γ ∗ = C 0 − ∑ j α j n j \langle H_W\rangle_{\psi_{\Gamma_*}}=C_0-\sum_j\alpha_j n_j ⟨ H W ⟩ ψ Γ ∗ = C 0 − ∑ j α j n j .
If n p = 0 n_p=0 n p = 0 , then H A = 0 H_A=0 H A = 0 , so ∣ ψ Γ ∗ ⟩ |\psi_{\Gamma_*}\rangle ∣ ψ Γ ∗ ⟩ is a ground state of H W H_W H W , as used in Sec. II E. Henceforth assume n p > 0 n_p>0 n p > 0 .
Partition K K K according to the paired and unpaired components of c c c . Completing the square for this positive block matrix gives
c † K c = c p † K e f f c p + ( c u + K u u + K u p c p ) † K u u ( c u + K u u + K u p c p ) , c^\dagger Kc=c_p^\dagger K_{\mathrm{eff}}c_p+(c_u+K_{uu}^{+}K_{up}c_p)^\dagger K_{uu}(c_u+K_{uu}^{+}K_{up}c_p), c † K c = c p † K eff c p + ( c u + K uu + K u p c p ) † K uu ( c u + K uu + K u p c p ) ,
K e f f = K p p − K p u K u u + K u p ≥ 0. (96) K_{\mathrm{eff}}=K_{pp}-K_{pu}K_{uu}^{+}K_{up}\geq0. \tag*{(96)} K eff = K pp − K p u K uu + K u p ≥ 0. (96)
Here + + + denotes the Moore–Penrose inverse; the necessary range condition follows from K ≥ 0 K\geq0 K ≥ 0 . The operator c p † K e f f c p − H A c_p^\dagger K_{\mathrm{eff}}c_p-H_A c p † K eff c p − H A acts only on the paired modes and has a positive-semidefinite real quadratic coefficient. Indeed, in the classical quadratic forms, minimizing that of H W ≥ 0 H_W\geq0 H W ≥ 0 over the quadratures of the unpaired modes sets the square in Eq. (96) to zero. The coefficient of c p † K e f f c p − H A c_p^\dagger K_{\mathrm{eff}}c_p-H_A c p † K eff c p − H A is therefore the Schur complement of W W W on the paired quadratures. Expressed in creation and annihilation operators, this positivity condition is
\begin{pmatrix}
K_{\mathrm{eff}}-D & -P\\
-P & K_{\mathrm{eff}}^{\mathsf T}-D
\end{pmatrix}\geq0, \tag*{(97)}\end{pmatrix}
where Q = diag ( q j ) Q=\operatorname{diag}(q_j) Q = diag ( q j ) , D α = diag ( α j / ( 1 − q j 2 ) ) D_\alpha=\operatorname{diag}(\alpha_j/(1-q_j^2)) D α = diag ( α j / ( 1 − q j 2 )) , D = diag ( D α , Q D α Q ) > 0 D=\operatorname{diag}(D_\alpha,QD_\alpha Q)>0 D = diag ( D α , Q D α Q ) > 0 , and P = ( 0 D α Q D α Q 0 ) P=\begin{pmatrix}0&D_\alpha Q\\D_\alpha Q&0\end{pmatrix} P = ( 0 D α Q D α Q 0 ) . In particular K e f f ≥ D > 0 K_{\mathrm{eff}}\geq D>0 K eff ≥ D > 0 .
Define the reduced quadratic on the paired modes by
H = c p † K e f f c p − H A + C 0 , H=c_p^\dagger K_{\mathrm{eff}}c_p-H_A+C_0, H = c p † K eff c p − H A + C 0 ,
and let H r e m H_{\mathrm{rem}} H rem be the nonnegative square in Eq. (96) . Equations (B2) and (B3) give H W = H + H r e m H_W=H+H_{\mathrm{rem}} H W = H + H rem . Every annihilator in that square annihilates ∣ ψ Γ ∗ ⟩ |\psi_{\Gamma_*}\rangle ∣ ψ Γ ∗ ⟩ , so the remainder has zero expectation at the Gaussian seed. Equation (B4) proves that H H H has a positive-semidefinite real coefficient.
Finally, the first variation of the paired marginal entropy is that of H A H_A H A . The first variation of ⟨ c p † K e f f c p ⟩ \langle c_p^\dagger K_{\mathrm{eff}}c_p\rangle ⟨ c p † K eff c p ⟩ at its vacuum is zero. Hence the paired seed ω c \omega_c ω c is stationary for entropy plus H H H , and the geodesic-convex criterion in Sec. II D makes it a global pure Gaussian minimum. This completes the reduction.
Relative-entropy inequality for joint Gaussian references
Proof of Proposition 2. We argue by contradiction, ruling out a negative variational gap. Thermal replacement and a quartic number penalty make the minimization regular enough to use its stationarity equations. A weighted comparison then controls the departure of the output from thermal equilibrium and identifies its limit as the reference state. Subsection C 1 rewrites the hypotheses in thermal coordinates and introduces the balance defects. Subsection C 2 regularizes a putative counterexample and derives the Gibbs equation of its minimizer. Subsections C 3 and C 4 turn second-order optimality and the strict hypothesis into a quantitative bound on the defects. Subsection C 5 controls the error from thermal replacement, and Subsection C 6 evaluates stationarity along the thermal semigroup to reach the contradiction.
The reference states, Gaussian unitaries and thermal coordinates are fixed throughout this appendix. Constants may depend on their thermal occupations; unless stated otherwise, they are uniform in the quartic penalty introduced below. We use N = ∑ i = 1 n a i † a i N=\sum_{i=1}^{n}a_i^\dagger a_i N = ∑ i = 1 n a i † a i and K = I + N K=I+N K = I + N , where n n n is the input mode count. A fixed Gaussian change of coordinates preserves finite energy.
Reference identities and weak thermal balance
We first express the covariance relation τ = Φ ( σ ) \tau=\Phi(\sigma) τ = Φ ( σ ) and the strict quadratic hypothesis in thermal coordinates. The faithful input and output reference Hamiltonians are
H σ = ∑ i h i a i † a i , H τ = ∑ j h 0 j b j † b j H_\sigma=\sum_i h_i a_i^\dagger a_i,\qquad H_\tau=\sum_j h_{0j}b_j^\dagger b_j H σ = i ∑ h i a i † a i , H τ = j ∑ h 0 j b j † b j
h i = log n σ , i + 1 n σ , i , h 0 j = log n τ , j + 1 n τ , j , h_i=\log\frac{n_{\sigma,i}+1}{n_{\sigma,i}},\qquad h_{0j}=\log\frac{n_{\tau,j}+1}{n_{\tau,j}}, h i = log n σ , i n σ , i + 1 , h 0 j = log n τ , j n τ , j + 1 ,
where n σ , i = Tr σ a i † a i > 0 n_{\sigma,i}=\operatorname{Tr}\sigma a_i^\dagger a_i>0 n σ , i = Tr σ a i † a i > 0 and n τ , j = Tr τ b j † b j > 0 n_{\tau,j}=\operatorname{Tr}\tau b_j^\dagger b_j>0 n τ , j = Tr τ b j † b j > 0 are thermal occupations. These Hamiltonians have ground energy zero. In particular, their scalar ordering terms are included in B = λ H σ − Φ ∗ H τ B=\lambda H_\sigma-\Phi^*H_\tau B = λ H σ − Φ ∗ H τ . Write the pullbacks of the retained annihilators as Φ ∗ ( b ) = U a + V a † \Phi^*(b)=Ua+Va^\dagger Φ ∗ ( b ) = U a + V a † , and set
C = ( U V V ‾ U ‾ ) , H = diag ( h i , h i ) , H 0 = diag ( h 0 j , h 0 j ) . C=\begin{pmatrix}U&V\\ \overline{V}&\overline{U}\end{pmatrix},\qquad H=\operatorname{diag}(h_i,h_i),\qquad H_0=\operatorname{diag}(h_{0j},h_{0j}). C = ( U V V U ) , H = diag ( h i , h i ) , H 0 = diag ( h 0 j , h 0 j ) .
The canonical relations are U U † − V V † = I UU^\dagger-VV^\dagger=I U U † − V V † = I and U V T = V U T UV^{\mathsf T}=VU^{\mathsf T} U V T = V U T . With D + = diag ( n σ , i , n σ , i + 1 ) D_+=\operatorname{diag}(n_{\sigma,i},n_{\sigma,i}+1) D + = diag ( n σ , i , n σ , i + 1 ) , D − = diag ( n σ , i + 1 , n σ , i ) D_-=\operatorname{diag}(n_{\sigma,i}+1,n_{\sigma,i}) D − = diag ( n σ , i + 1 , n σ , i ) , and the analogous output matrices T ± T_\pm T ± , the unsymmetrized second-moment identity reads
C D ± C † = T ± . (98) CD_\pm C^\dagger=T_\pm. \tag*{(98)} C D ± C † = T ± . (98)
These matrices contain ⟨ a i † a i ⟩ \langle a_i^\dagger a_i\rangle ⟨ a i † a i ⟩ and ⟨ a i a i † ⟩ \langle a_i a_i^\dagger\rangle ⟨ a i a i † ⟩ , so their entries are unchanged by our doubled-covariance convention. The identity includes the vanishing anomalous moments ⟨ b j b k ⟩ τ \langle b_jb_k\rangle_\tau ⟨ b j b k ⟩ τ . The hypothesis of Proposition 2 is equivalent to λ H − C † H 0 C > 0 \lambda H-C^\dagger H_0C>0 λ H − C † H 0 C > 0 . Comparing the nonzero singular values of H 0 1 / 2 C H − 1 / 2 H_0^{1/2}CH^{-1/2} H 0 1/2 C H − 1/2 and its adjoint gives some δ ∈ ( 0 , λ ) \delta\in(0,\lambda) δ ∈ ( 0 , λ ) such that
C H − 1 C † ≤ ( λ − δ ) H 0 − 1 . (99) CH^{-1}C^\dagger\leq(\lambda-\delta)H_0^{-1}. \tag*{(99)} C H − 1 C † ≤ ( λ − δ ) H 0 − 1 . (99)
To express thermal balance for an arbitrary finite-energy joint state ρ \rho ρ , let μ i = Tr ρ a i \mu_i=\operatorname{Tr}\rho a_i μ i = Tr ρ a i and a i ′ = a i − μ i I a_i'=a_i-\mu_i I a i ′ = a i − μ i I . Its centered balance defects are the complex-linear functionals on bounded observables defined by
q i + ( Z ) = ( n σ , i + 1 ) Tr ρ ( a i ′ ) † Z − n σ , i Tr ρ Z ( a i ′ ) † , q_i^+(Z)=(n_{\sigma,i}+1)\operatorname{Tr}\rho(a_i')^\dagger Z-n_{\sigma,i}\operatorname{Tr}\rho Z(a_i')^\dagger, q i + ( Z ) = ( n σ , i + 1 ) Tr ρ ( a i ′ ) † Z − n σ , i Tr ρZ ( a i ′ ) † ,
q i − ( Z ) = ( n σ , i + 1 ) Tr ρ Z a i ′ − n σ , i Tr ρ a i ′ Z . (100) q_i^-(Z)=(n_{\sigma,i}+1)\operatorname{Tr}\rho Za_i'-n_{\sigma,i}\operatorname{Tr}\rho a_i'Z. \tag*{(100)} q i − ( Z ) = ( n σ , i + 1 ) Tr ρZ a i ′ − n σ , i Tr ρ a i ′ Z . (100)
These weak traces are well defined: a i ρ 1 / 2 a_i\rho^{1/2} a i ρ 1/2 and a i † ρ 1 / 2 a_i^\dagger\rho^{1/2} a i † ρ 1/2 are Hilbert–Schmidt, so multiplication by ρ 1 / 2 \rho^{1/2} ρ 1/2 makes each required state–ladder product trace class. Thus the functionals act on every bounded Z Z Z . The functionals vanish on I I I , and q i − ( Z ) = q i + ( Z † ) q_i^-(Z)=q_i^+(Z^\dagger) q i − ( Z ) = q i + ( Z † ) .
Put ρ o u t = Φ ( ρ ) \rho_{\mathrm{out}}=\Phi(\rho) ρ out = Φ ( ρ ) , μ o u t , j = Tr ρ o u t b j \mu_{\mathrm{out},j}=\operatorname{Tr}\rho_{\mathrm{out}}b_j μ out , j = Tr ρ out b j , and define its functionals using the reference occupations n τ , j n_{\tau,j} n τ , j and these actual means. Then
\begin{pmatrix}q_{\mathrm{out},+}\\q_{\mathrm{out},-}\end{pmatrix}
=\overline{C}\begin{pmatrix}q_+\\q_-\end{pmatrix}\Phi^*, \tag*{(101)}\end{pmatrix}
where q Φ ∗ q\Phi^* q Φ ∗ denotes the functional Z ↦ q ( Φ ∗ Z ) Z\mapsto q(\Phi^*Z) Z ↦ q ( Φ ∗ Z ) . For example, expanding the plus functional as Tr ρ ( a i ′ ) † Z + n σ , i Tr ρ [ a i ′ † , Z ] \operatorname{Tr}\rho(a_i')^\dagger Z+n_{\sigma,i}\operatorname{Tr}\rho[a_i'^\dagger,Z] Tr ρ ( a i ′ ) † Z + n σ , i Tr ρ [ a i ′ † , Z ] and using
[ a i † , Φ ∗ Z ] = ∑ j ( U j i Φ ∗ [ b j † , Z ] − V ‾ j i Φ ∗ [ b j , Z ] ) [a_i^\dagger,\Phi^*Z]=\sum_j\left(U_{ji}\Phi^*[b_j^\dagger,Z]-\overline{V}_{ji}\Phi^*[b_j,Z]\right) [ a i † , Φ ∗ Z ] = j ∑ ( U j i Φ ∗ [ b j † , Z ] − V j i Φ ∗ [ b j , Z ] )
gives (C4); (C1) supplies the thermal coefficient and cancels the anomalous coefficient. These are weak commutator identities. They hold first on number-state cutoffs; Hilbert–Schmidt factorization passes every state–ladder product to the trace-norm limit, establishing the identity for arbitrary finite-energy joint states.
A regularized variational problem
We regularize a hypothetical violation in two stages: thermal replacement bounds the output logarithm, and a quartic number penalty gives the minimizer the moments needed in the variational argument. Suppose a finite-energy state strictly violates the contraction inequality. Let P L P_L P L project onto total photon number at most L L L , and set ρ L = P L ρ P L / Tr ( ρ P L ) \rho_L=P_L\rho P_L/\operatorname{Tr}(\rho P_L) ρ L = P L ρ P L / Tr ( ρ P L ) . These normalized compressions converge in the sandwiched trace norm ∥ K 1 / 2 ( ρ L − ρ ) K 1 / 2 ∥ 1 \|K^{1/2}(\rho_L-\rho)K^{1/2}\|_1 ∥ K 1/2 ( ρ L − ρ ) K 1/2 ∥ 1 . Thus their quadratic expectations converge. Input and output entropies converge by oscillator energy-constrained continuity [25 ] . Consequently a counterexample supported on finitely many number states suffices.
Choose a small θ > 0 \theta>0 θ > 0 , put w = 1 − θ w=1-\theta w = 1 − θ , and mix the channel output with its thermal reference: ρ 0 = w Φ ρ + θ τ \rho_0=w\Phi\rho+\theta\tau ρ 0 = w Φ ρ + θ τ . For ζ > 0 \zeta>0 ζ > 0 , minimize over states with finite fourth moment
J ζ ( ρ ) = S ( ρ 0 ) − S ( τ ) − w λ [ S ( ρ ) − S ( σ ) ] + w Tr B ( ρ − σ ) + ζ Tr ρ K 4 . (102) J_\zeta(\rho)=S(\rho_0)-S(\tau)-w\lambda[S(\rho)-S(\sigma)]+w\operatorname{Tr}B(\rho-\sigma)+\zeta\operatorname{Tr}\rho K^4. \tag*{(102)} J ζ ( ρ ) = S ( ρ 0 ) − S ( τ ) − w λ [ S ( ρ ) − S ( σ )] + w Tr B ( ρ − σ ) + ζ Tr ρ K 4 . (102)
Continuity at this fixed counterexample permits θ \theta θ to be chosen small enough that inf J ζ ≤ − β < 0 \inf J_\zeta\leq-\beta<0 inf J ζ ≤ − β < 0 for all sufficiently small ζ \zeta ζ . Additional restrictions on θ \theta θ below depend only on the fixed reference data and preserve this property.
Strict positivity of the real quadratic coefficient gives B ≥ c N − C I B\geq cN-CI B ≥ c N − C I . The input entropy satisfies S ( ρ ) ≤ n g ( Tr ρ N / n ) = S(\rho)\leq ng(\operatorname{Tr}\rho N/n)= S ( ρ ) ≤ n g ( Tr ρN / n ) = O ( log ( 1 + Tr ρ N ) ) O(\log(1+\operatorname{Tr}\rho N)) O ( log ( 1 + Tr ρN )) . Since S ( ρ 0 ) ≥ 0 S(\rho_0)\ge0 S ( ρ 0 ) ≥ 0 , a fixed sublevel of J ζ J_\zeta J ζ has a uniform energy bound, independent of ζ \zeta ζ ; for fixed ζ \zeta ζ it also has a fourth-moment bound. Compactness of oscillator energy sublevels, entropy continuity on those sublevels, and lower semicontinuity of positive energy forms give a minimizer. For fixed ζ \zeta ζ , write ρ = ρ ζ \rho=\rho_\zeta ρ = ρ ζ for such a minimizer.
Let H 0 = − log ρ 0 \mathcal{H}_0=-\log\rho_0 H 0 = − log ρ 0 . Operator monotonicity of the logarithm and ρ 0 ≥ θ τ \rho_0\ge\theta\tau ρ 0 ≥ θ τ imply the form bound
0 ≤ H 0 ≤ − log τ − log θ ≤ C θ K o u t , (103) 0\le\mathcal{H}_0\le-\log\tau-\log\theta\le C_\theta K_{\mathrm{out}}, \tag*{(103)} 0 ≤ H 0 ≤ − log τ − log θ ≤ C θ K out , (103)
where K o u t = I + ∑ j b j † b j K_{\mathrm{out}}=\mathbb{I}+\sum_j b_j^\dagger b_j K out = I + ∑ j b j † b j is the output number weight. Pullback of the positive form, defined by increasing bounded spectral truncations, obeys 0 ≤ Φ ∗ H 0 ≤ C θ K 0\le\Phi^*\mathcal{H}_0\le C_\theta K 0 ≤ Φ ∗ H 0 ≤ C θ K . For a competitor ρ ~ \widetilde{\rho} ρ , set ρ ~ 0 = w Φ ( ρ ~ ) + θ τ \widetilde{\rho}_0=w\Phi(\widetilde{\rho})+\theta\tau ρ 0 = w Φ ( ρ ) + θ τ , and replace its output entropy in (C5) by Tr ρ ~ 0 H 0 \operatorname{Tr}\widetilde{\rho}_0\mathcal{H}_0 Tr ρ 0 H 0 . Relative-entropy positivity makes this a majorant, with equality at the minimizer. Hence that minimizer also minimizes the resulting input Gibbs functional [70 ] .
Lemma 7 (Quartic Gibbs regularity). Let K = I + N K=\mathbb{I}+N K = I + N , c > 0 c>0 c > 0 , and let T ≥ 0 T\ge0 T ≥ 0 be bounded. The closed form
A = c K 4 + K 1 / 2 T K 1 / 2 , D ( A 1 / 2 ) = D ( K 2 ) , A=cK^4+K^{1/2}TK^{1/2},\qquad\mathcal{D}(A^{1/2})=\mathcal{D}(K^2), A = c K 4 + K 1/2 T K 1/2 , D ( A 1/2 ) = D ( K 2 ) ,
has compact resolvent and 0 < Tr e − A < ∞ 0<\operatorname{Tr}e^{-A}<\infty 0 < Tr e − A < ∞ . Its Gibbs state is faithful and has finite seventh number moment. It is the unique minimizer of Tr ω A − S ( ω ) \operatorname{Tr}\omega A-S(\omega) Tr ω A − S ( ω ) among states with finite fourth moment.
Proof of Lemma 7. The form is ∥ F 1 / 2 K 2 ψ ∥ 2 \lVert F^{1/2}K^2\psi\rVert^2 ∥ F 1/2 K 2 ψ ∥ 2 , where F = c I + K − 3 / 2 T K − 3 / 2 F=c\mathbb{I}+K^{-3/2}TK^{-3/2} F = c I + K − 3/2 T K − 3/2 is bounded and boundedly invertible. Thus it is closed and its inverse is the compact operator K − 2 F − 1 K − 2 K^{-2}F^{-1}K^{-2} K − 2 F − 1 K − 2 . The min–max principle applied to A ≥ c K 4 A\ge cK^4 A ≥ c K 4 bounds its eigenvalue counting function by that of c K 4 cK^4 c K 4 , which grows polynomially. For an eigenvector A ψ = u ψ A\psi=u\psi A ψ = u ψ , the number-coordinate form equation gives
c K 7 / 2 ψ = u K − 1 / 2 ψ − T K 1 / 2 ψ . cK^{7/2}\psi=uK^{-1/2}\psi-TK^{1/2}\psi. c K 7/2 ψ = u K − 1/2 ψ − T K 1/2 ψ .
The right side is square summable because ∥ K 2 ψ ∥ 2 ≤ u / c \lVert K^2\psi\rVert^2\le u/c ∥ K 2 ψ ∥ 2 ≤ u / c . It follows that ψ ∈ D ( K 7 / 2 ) \psi\in\mathcal{D}(K^{7/2}) ψ ∈ D ( K 7/2 ) , with its graph norm bounded by a polynomial in u u u . Summation against e − u e^{-u} e − u proves the trace and seventh-moment assertions. Finally, Tr ω A − S ( ω ) = D ( ω ∥ e − A / Tr e − A ) − log Tr e − A \operatorname{Tr}\omega A-S(\omega)=D(\omega\Vert e^{-A}/\operatorname{Tr}e^{-A})-\log\operatorname{Tr}e^{-A} Tr ω A − S ( ω ) = D ( ω ∥ e − A / Tr e − A ) − log Tr e − A ; all terms are finite on the stated domain. □
After a scalar shift, the Gibbs Hamiltonian of the preceding majorant has exactly the form in Lemma 7. Therefore the minimizer is faithful and satisfies the exact form identity
w λ H ρ = w Φ ∗ H 0 + w B + ζ K 4 + c ζ I , H ρ = − log ρ , (104) w\lambda\mathcal{H}_\rho=w\Phi^*\mathcal{H}_0+wB+\zeta K^4+c_\zeta\mathbb{I},\qquad\mathcal{H}_\rho=-\log\rho, \tag*{(104)} w λ H ρ = w Φ ∗ H 0 + w B + ζ K 4 + c ζ I , H ρ = − log ρ , (104)
together with Tr ρ K 7 < ∞ \operatorname{Tr}\rho K^7<\infty Tr ρ K 7 < ∞ and 0 ≤ H ρ ≤ C ζ K 4 0\le\mathcal{H}_\rho\le C_\zeta K^4 0 ≤ H ρ ≤ C ζ K 4 . These moment bounds hold at each fixed ζ \zeta ζ ; only the first moment is uniform as ζ ↓ 0 \zeta\downarrow0 ζ ↓ 0 .
The output and the replaced output also have finite seventh moments. Let U \mathcal{U} U be the Gaussian unitary in Φ \Phi Φ , and let K t o t K_{\mathrm{tot}} K tot denote one plus total photon number before discarding modes. Conjugating a number power by this unitary expands it into finitely many ladder monomials. Splitting each monomial into two factors and using the number-shift bounds proves
U ∗ K t o t k U ≤ C k K t o t k \mathcal{U}^*K_{\mathrm{tot}}^k\mathcal{U}\le C_kK_{\mathrm{tot}}^k U ∗ K tot k U ≤ C k K tot k
as forms for every integer k ≥ 1 k\ge1 k ≥ 1 . Partial trace then transfers the moment bound. The same shift estimates show that K t o t s U K − s K_{\mathrm{tot}}^s\mathcal{U}K^{-s} K tot s U K − s is bounded for integer s s s , which will transfer sandwiched trace convergence below.
The Hessian bound on a common centered core
The second variation at a minimizer is nonnegative. We express this condition as a norm bound for output observables pulled back to the input. For a faithful state ω = diag ( p l ) \omega=\operatorname{diag}(p_l) ω = diag ( p l ) , define the logarithmic mean and its associated weighted norm by
ℓ ( p , q ) = p − q log p − log q , ℓ ( p , p ) = p , \ell(p,q)=\frac{p-q}{\log p-\log q},\qquad\ell(p,p)=p, ℓ ( p , q ) = log p − log q p − q , ℓ ( p , p ) = p ,
∥ Z ∥ ℓ , ω 2 = ∑ l , k ℓ ( p l , p k ) ∣ Z l k ∣ 2 . \lVert Z\rVert_{\ell,\omega}^2=\sum_{l,k}\ell(p_l,p_k)|Z_{lk}|^2. ∥ Z ∥ ℓ , ω 2 = l , k ∑ ℓ ( p l , p k ) ∣ Z l k ∣ 2 .
We use the centered core consisting of operators Z − Tr ( ω Z ) I Z-\operatorname{Tr}(\omega Z)\mathbb{I} Z − Tr ( ω Z ) I , where Z Z Z is supported on finitely many eigenvectors of ω \omega ω . The subtraction leaves a constant diagonal tail outside that finite block.
The Gibbs identity gives, for finite-fourth-moment competitors,
J ζ ( ρ ~ ) − J ζ ( ρ ) = w λ D ( ρ ~ ∥ ρ ) − D ( ρ ~ 0 ∥ ρ 0 ) ≥ 0. J_\zeta(\widetilde{\rho})-J_\zeta(\rho)=w\lambda D(\widetilde{\rho}\Vert\rho)-D(\widetilde{\rho}_0\Vert\rho_0)\ge0. J ζ ( ρ ) − J ζ ( ρ ) = w λ D ( ρ ∥ ρ ) − D ( ρ 0 ∥ ρ 0 ) ≥ 0.
We claim that every bounded output test obeys
w ∥ Φ ∗ Z − Tr ρ o u t Z I ∥ ℓ , ρ 2 ≤ λ ∥ Z ∥ ℓ , ρ 0 2 . (105) w\lVert\Phi^*Z-\operatorname{Tr}\rho_{\mathrm{out}}Z\mathbb{I}\rVert_{\ell,\rho}^2\le\lambda\lVert Z\rVert_{\ell,\rho_0}^2. \tag*{(105)} w ∥ Φ ∗ Z − Tr ρ out Z I ∥ ℓ , ρ 2 ≤ λ ∥ Z ∥ ℓ , ρ 0 2 . (105)
We derive this bound first on finite blocks and then extend it by duality [71 ] . Let X = X ∗ X=X^* X = X ∗ be trace zero and supported on finitely many input eigenvectors. Both signs of ρ + t X \rho+tX ρ + tX are admissible for small t t t , and
D ( ρ + t X ∥ ρ ) = 1 2 t 2 Q ρ ( X ) + o ( t 2 ) , Q ρ ( X ) = ∑ l , k ∣ X l k ∣ 2 ℓ ( p l , p k ) . D(\rho+tX\Vert\rho)=\tfrac12t^2Q_\rho(X)+o(t^2),\qquad Q_\rho(X)=\sum_{l,k}\frac{|X_{lk}|^2}{\ell(p_l,p_k)}. D ( ρ + tX ∥ ρ ) = 2 1 t 2 Q ρ ( X ) + o ( t 2 ) , Q ρ ( X ) = l , k ∑ ℓ ( p l , p k ) ∣ X l k ∣ 2 .
Let Z Z Z be a Hermitian output-core test, centered so that Tr ρ 0 Z = 0 \operatorname{Tr}\rho_0Z=0 Tr ρ 0 Z = 0 . The output of ρ ~ = ρ + t X \widetilde{\rho}=\rho+tX ρ = ρ + tX is ρ ~ 0 = ρ 0 + w t Φ ( X ) \widetilde{\rho}_0=\rho_0+wt\Phi(X) ρ 0 = ρ 0 + w t Φ ( X ) . The Gibbs variational inequality D ( ρ 0 ∥ ρ 0 ) ≥ Tr ρ 0 Y − log Tr e log ρ 0 + Y D(\rho_0\Vert\rho_0)\ge\operatorname{Tr}\rho_0Y-\log\operatorname{Tr}e^{\log\rho_0+Y} D ( ρ 0 ∥ ρ 0 ) ≥ Tr ρ 0 Y − log Tr e l o g ρ 0 + Y , with Y = s Z Y=sZ Y = s Z , and the expansion log Tr e log ρ 0 + s Z = s 2 ∥ Z ∥ ℓ , ρ 0 2 / 2 + o ( s 2 ) \log\operatorname{Tr}e^{\log\rho_0+sZ}=s^2\lVert Z\rVert_{\ell,\rho_0}^2/2+o(s^2) log Tr e l o g ρ 0 + s Z = s 2 ∥ Z ∥ ℓ , ρ 0 2 /2 + o ( s 2 ) , combined with the preceding display, give
1 2 w λ t 2 Q ρ ( X ) + o ( t 2 ) ≥ s w t Tr ( X Φ ∗ Z ) − 1 2 s 2 ∥ Z ∥ ℓ , ρ 0 2 + o ( s 2 ) . \tfrac12w\lambda t^2Q_\rho(X)+o(t^2)\ge swt\operatorname{Tr}(X\Phi^*Z)-\tfrac12s^2\lVert Z\rVert_{\ell,\rho_0}^2+o(s^2). 2 1 w λ t 2 Q ρ ( X ) + o ( t 2 ) ≥ s w t Tr ( X Φ ∗ Z ) − 2 1 s 2 ∥ Z ∥ ℓ , ρ 0 2 + o ( s 2 ) .
Choosing s = w t Tr ( X Φ ∗ Z ) / ∥ Z ∥ ℓ , ρ 0 2 s=wt\operatorname{Tr}(X\Phi^*Z)/\lVert Z\rVert_{\ell,\rho_0}^2 s = w t Tr ( X Φ ∗ Z ) / ∥ Z ∥ ℓ , ρ 0 2 and letting t → 0 t\to0 t → 0 yields
w ∣ Tr X Φ ∗ Z ∣ 2 ≤ λ Q ρ ( X ) ∥ Z ∥ ℓ , ρ 0 2 . w|\operatorname{Tr}X\Phi^*Z|^2\le\lambda Q_\rho(X)\lVert Z\rVert_{\ell,\rho_0}^2. w ∣ Tr X Φ ∗ Z ∣ 2 ≤ λ Q ρ ( X ) ∥ Z ∥ ℓ , ρ 0 2 .
Finite trace-zero eigenblocks are dense in the trace-zero Q ρ Q_\rho Q ρ -space: compress the weighted coordinate sum and correct the residual trace at one fixed diagonal entry with p l > 0 p_l>0 p l > 0 . Trace is continuous there by Cauchy–Schwarz and ∑ l p l = 1 \sum_l p_l=1 ∑ l p l = 1 . Taking the supremum over X X X therefore gives the centered logarithmic-mean norm on the left of (C8). Real and imaginary parts give complex tests.
For later extensions use the arithmetic norm
∥ Z ∥ ℓ , ω 2 = ∑ l , k ( p l + p k ) ∣ Z l k ∣ 2 = Tr ω ( Z Z † + Z † Z ) . (106) \lVert Z\rVert^2_{\ell,\omega}=\sum_{l,k}(p_l+p_k)|Z_{lk}|^2=\operatorname{Tr}\omega(ZZ^\dagger+Z^\dagger Z). \tag*{(106)} ∥ Z ∥ ℓ , ω 2 = l , k ∑ ( p l + p k ) ∣ Z l k ∣ 2 = Tr ω ( Z Z † + Z † Z ) . (106)
For every bounded Z Z Z , its centered eigenblock compressions converge to Z − Tr ( ω Z ) I Z-\operatorname{Tr}(\omega Z)I Z − Tr ( ω Z ) I in this norm. The Schwarz inequalities for Φ ∗ \Phi^* Φ ∗ , followed by the variance identity for ρ 0 = w ρ o u t + θ τ \rho_0=w\rho_{\mathrm{out}}+\theta\tau ρ 0 = w ρ out + θ τ , give the same convergence for their actual centered pullbacks. Since ℓ ( p , q ) ≤ ( p + q ) / 2 \ell(p,q)\le(p+q)/2 ℓ ( p , q ) ≤ ( p + q ) /2 , this proves (C8) for every bounded Z Z Z , and identifies its continuous extension with the physical pullback. Each additional metric below is bounded above by a fixed multiple of the arithmetic metric. The same core is dense in each centered weighted space: truncate the weighted coordinate sum, then subtract the residual mean. This correction converges because every diagonal weight is a fixed positive multiple of p l p_l p l . Thus all endpoint comparisons use the same operator on the same core.
Four comparisons and a strict dual remainder
We combine the strict matrix bound (C2) with the Hessian bound (C8). These give four weighted norm comparisons, whose interpolation yields the norms of the thermal balance defects. The strict gap will remain as a quantitative bound on those defects.
Associate plus and minus components with each mode. For an output vector Z Z Z , centered componentwise against ρ 0 \rho_0 ρ 0 , define
L Z = w Φ ∗ ( Φ ∗ Z − Tr ρ o u t Z I ) . LZ=\sqrt{w}\,\Phi^*(\Phi^*Z-\operatorname{Tr}\rho_{\mathrm{out}}ZI). L Z = w Φ ∗ ( Φ ∗ Z − Tr ρ out Z I ) .
Every image component is centered against ρ \rho ρ . For each input or output mode, write h h h for its reference inverse temperature and n ˉ = ( e h − 1 ) − 1 \bar n=(e^h-1)^{-1} n ˉ = ( e h − 1 ) − 1 for its occupation. For a coordinate ( l , k ) (l,k) ( l , k ) in the eigenbasis of ρ \rho ρ or ρ 0 \rho_0 ρ 0 , set
x = 1 2 log ( p l / p k ) , m = 1 / h , y = ∓ h / 2 x=\frac{1}{2}\log(p_l/p_k),\qquad m=1/h,\qquad y=\mp h/2 x = 2 1 log ( p l / p k ) , m = 1/ h , y = ∓ h /2
on the plus/minus mode, respectively, and use base measure ℓ ( p l , p k ) \ell(p_l,p_k) ℓ ( p l , p k ) . Let f ( x ) = x / sinh x f(x)=x/\sinh x f ( x ) = x / sinh x , with f ( 0 ) = 1 f(0)=1 f ( 0 ) = 1 . The first pair of four weights
m f ( x ) e ± x , m f ( y ) e ± y (107) m f(x)e^{\pm x},\qquad m f(y)e^{\pm y} \tag*{(107)} m f ( x ) e ± x , m f ( y ) e ± y (107)
becomes p l / h p_l/h p l / h , p k / h p_k/h p k / h after multiplication by the base measure. These two metrics compare under L L L with constant λ − δ \lambda-\delta λ − δ . Indeed, put Y β = Φ ∗ Z β − Tr ( ρ o u t Z β ) I Y_\beta=\Phi^*Z_\beta-\operatorname{Tr}(\rho_{\mathrm{out}}Z_\beta)I Y β = Φ ∗ Z β − Tr ( ρ out Z β ) I for the output components β \beta β . The Gram matrix G β β ′ = Tr ρ Y β Y β ′ † G_{\beta\beta'}=\operatorname{Tr}\rho\,Y_\beta Y_{\beta'}^\dagger G β β ′ = Tr ρ Y β Y β ′ † is positive semidefinite, so (C2) gives
∑ α h α − 1 Tr ρ ( L Z ) α ( L Z ) α † ≤ ( λ − δ ) w ∑ β h β − 1 G β β . \sum_\alpha h_{\alpha}^{-1}\operatorname{Tr}\rho\,(LZ)_\alpha(LZ)_\alpha^\dagger\le(\lambda-\delta)w\sum_\beta h_{\beta}^{-1}G_{\beta\beta}. α ∑ h α − 1 Tr ρ ( L Z ) α ( L Z ) α † ≤ ( λ − δ ) w β ∑ h β − 1 G β β .
With Var ω ( Z ) = Tr ω Z Z † − ∣ Tr ω Z ∣ 2 \operatorname{Var}_\omega(Z)=\operatorname{Tr}_\omega ZZ^\dagger-|\operatorname{Tr}_\omega Z|^2 Var ω ( Z ) = Tr ω Z Z † − ∣ Tr ω Z ∣ 2 , the Kadison–Schwarz inequality gives G β β ≤ Var ρ o u t ( Z β ) G_{\beta\beta}\le\operatorname{Var}_{\rho_{\mathrm{out}}}(Z_\beta) G β β ≤ Var ρ out ( Z β ) , and concavity of the variance under ρ 0 = w ρ o u t + θ τ \rho_0=w\rho_{\mathrm{out}}+\theta\tau ρ 0 = w ρ out + θ τ gives w Var ρ o u t ( Z β ) ≤ Var ρ 0 ( Z β ) w\operatorname{Var}_{\rho_{\mathrm{out}}}(Z_\beta)\le\operatorname{Var}_{\rho_0}(Z_\beta) w Var ρ out ( Z β ) ≤ Var ρ 0 ( Z β ) . The same argument applies with Z Z † ZZ^\dagger Z Z † replaced by Z † Z Z^\dagger Z Z † Z . The second pair becomes n ˉ ℓ , ( n ˉ + 1 ) ℓ \bar n\ell,(\bar n+1)\ell n ˉ ℓ , ( n ˉ + 1 ) ℓ , with the order exchanged on the minus sign.
The pair becomes n ˉ ℓ , ( n ˉ + 1 ) ℓ \bar n\ell,(\bar n+1)\ell n ˉ ℓ , ( n ˉ + 1 ) ℓ , with the order exchanged on the minus sign. Since C D ± C † = T ± CD_\pm C^\dagger=T_\pm C D ± C † = T ± is diagonal,
∑ α ( D ± ) α ∥ ( L Z ) α ∥ ℓ , ρ 2 = w ∑ β ( T ± ) β ∥ Y β ∥ ℓ , ρ 2 . \sum_\alpha(D_\pm)_\alpha\lVert(LZ)_\alpha\rVert^2_{\ell,\rho}
=w\sum_\beta(T_\pm)_\beta\lVert Y_\beta\rVert^2_{\ell,\rho}. α ∑ ( D ± ) α ∥( L Z ) α ∥ ℓ , ρ 2 = w β ∑ ( T ± ) β ∥ Y β ∥ ℓ , ρ 2 .
and (C8), applied to each output component, gives comparison constant λ \lambda λ for each of these two metrics. These comparisons use the Hessian of the full joint input state.
To combine these four comparisons, we apply Lemma 8. Its test-space hypothesis holds because each source-core test has finitely many values of ( m , x , y ) (m,x,y) ( m , x , y ) , and a function of this tuple multiplies all diagonal entries, including the identity tail, by the same constant within each mode/sign. It therefore preserves the core and its centering. The resulting metrics, including their base measures, are
M + ( p l , p k ) = ℓ ( n ˉ p l , ( n ˉ + 1 ) p k ) , M_+(p_l,p_k)=\ell(\bar n p_l,(\bar n+1)p_k), M + ( p l , p k ) = ℓ ( n ˉ p l , ( n ˉ + 1 ) p k ) ,
M − ( p l , p k ) = ℓ ( ( n ˉ + 1 ) p l , n ˉ p k ) . (108) M_-(p_l,p_k)=\ell((\bar n+1)p_l,\bar n p_k). \tag*{(108)} M − ( p l , p k ) = ℓ (( n ˉ + 1 ) p l , n ˉ p k ) . (108)
The subscripts M i , ± M_{i,\pm} M i , ± and M 0 i , ± M_{0i,\pm} M 0 i , ± below use the input occupations n σ , i n_{\sigma,i} n σ , i and output occupations n τ , j n_{\tau,j} n τ , j , respectively. Input and output metric weights use the eigenvalues of ρ \rho ρ and ρ 0 \rho_0 ρ 0 , respectively. Vector norms sum over modes, signs, and state-eigenbasis coordinates. The lemma also retains the strict source remainder
∥ L Z ∥ M , ρ 2 ≤ λ ∥ L Z ∥ M , ρ 0 2 − ∥ Z ∥ Δ 2 , 0 < Δ ≤ δ M . (109) \lVert LZ\rVert^2_{M,\rho}\le\lambda\lVert LZ\rVert^2_{M,\rho_0}-\lVert Z\rVert^2_{\Delta},
\qquad0<\Delta\le\delta M. \tag*{(109)} ∥ L Z ∥ M , ρ 2 ≤ λ ∥ L Z ∥ M , ρ 0 2 − ∥ Z ∥ Δ 2 , 0 < Δ ≤ δ M . (109)
where Δ \Delta Δ is the output weight in (D5), multiplied by ℓ ( p l , p k ) \ell(p_l,p_k) ℓ ( p l , p k ) . For each fixed reference temperature, its explicit formula gives
Δ ≥ c ( p l + p k ) 1 + log 2 ( p l / p k ) , Δ M 2 ≥ c ′ p l + p k . (110) \Delta\ge\frac{c(p_l+p_k)}{1+\log^2(p_l/p_k)},\qquad\frac{\Delta}{M^2}\ge\frac{c'}{p_l+p_k}. \tag*{(110)} Δ ≥ 1 + log 2 ( p l / p k ) c ( p l + p k ) , M 2 Δ ≥ p l + p k c ′ . (110)
To obtain both bounds, set b ( x ) = x coth x b(x)=x\coth x b ( x ) = x coth x , with b ( 0 ) = 1 b(0)=1 b ( 0 ) = 1 , and use the minimizer x x x in (D4). With y y y fixed, the auxiliary weight E E E in (D2) satisfies E / m = O ( ∣ x ∣ − 1 ) E/m=O(|x|^{-1}) E / m = O ( ∣ x ∣ − 1 ) , and
κ = 1 − [ b ( y ) + y ] / x + O ( x − 2 ) , x ⟶ + ∞ . \kappa=1-[b(y)+y]/x+O(x^{-2}),\qquad x\longrightarrow+\infty. κ = 1 − [ b ( y ) + y ] / x + O ( x − 2 ) , x ⟶ + ∞.
κ = − 1 + [ b ( y ) − y ] / ∣ x ∣ + O ( x − 2 ) , x ⟶ − ∞ . \kappa=-1+[b(y)-y]/|x|+O(x^{-2}),\qquad x\longrightarrow-\infty. κ = − 1 + [ b ( y ) − y ] /∣ x ∣ + O ( x − 2 ) , x ⟶ − ∞.
Since b ( y ) ± y > 0 b(y)\pm y>0 b ( y ) ± y > 0 , substitution gives the claimed positive endpoint bounds; continuity supplies the bounds on compact intervals. Finitely many modes allow common constants. The ratio Δ / M \Delta/M Δ/ M may approach zero.
We next justify applying the dual inequality to the weak defects. Write ω = ∑ k p k ∣ ψ k ⟩ ⟨ ψ k ∣ \omega=\sum_k p_k|\psi_k\rangle\langle\psi_k| ω = ∑ k p k ∣ ψ k ⟩ ⟨ ψ k ∣ , H ω = − log ω \mathcal H_\omega=-\log\omega H ω = − log ω , and let a ′ = a − Tr ( ω a ) I a'=a-\operatorname{Tr}(\omega a)I a ′ = a − Tr ( ω a ) I be the centered operator under consideration. Let K ω K_\omega K ω be one plus total photon number on the system of ω \omega ω . For a state satisfying Tr ω K ω 7 < ∞ \operatorname{Tr}\omega K_\omega^7<\infty Tr ω K ω 7 < ∞ and H ω ≤ C K ω 4 \mathcal H_\omega\le CK_\omega^4 H ω ≤ C K ω 4 , a plus-defect has coordinate coefficients of modulus ∣ ( n ˉ + 1 ) p k − n ˉ p l ∣ ∣ a l k ′ ∣ \lvert(\bar n+1)p_k-\bar n p_l\rvert|a'_{lk}| ∣( n ˉ + 1 ) p k − n ˉ p l ∣ ∣ a l k ′ ∣ . Hence
∣ q l k ∣ 2 M + , l k ≤ C n ˉ ( p l + p k ) ( 1 − log p l − log p k ) ∣ a l k ′ ∣ 2 . \frac{|q_{lk}|^2}{M_{+,lk}}\le C\bar n(p_l+p_k)(1-\log p_l-\log p_k)|a'_{lk}|^2. M + , l k ∣ q l k ∣ 2 ≤ C n ˉ ( p l + p k ) ( 1 − log p l − log p k ) ∣ a l k ′ ∣ 2 .
The crossed logarithmic sum, for example, is bounded by
∑ l , k p k ( − log p l ) ∣ a l k ′ ∣ 2 = ∑ k p k ⟨ a ′ ψ k , H ω a ′ ψ k ⟩ ≤ C Tr ω K ω 5 . \sum_{l,k}p_k(-\log p_l)|a'_{lk}|^2
=\sum_k p_k\langle a'\psi_k,\mathcal H_\omega a'\psi_k\rangle
\le C\operatorname{Tr}\omega K_\omega^5. l , k ∑ p k ( − log p l ) ∣ a l k ′ ∣ 2 = k ∑ p k ⟨ a ′ ψ k , H ω a ′ ψ k ⟩ ≤ C Tr ω K ω 5 .
For the same-index sum, use
( − log p k ) ∥ a ′ ψ k ∥ 2 ≤ C ⟨ K 4 ⟩ ψ k ⟨ K ω ⟩ ψ k ≤ C ⟨ K 5 ⟩ ψ k . (-\log p_k)\lVert a'\psi_k\rVert^2 \le C\langle K^4\rangle_{\psi_k}\langle K_\omega\rangle_{\psi_k} \le C\langle K^5\rangle_{\psi_k}. ( − log p k ) ∥ a ′ ψ k ∥ 2 ≤ C ⟨ K 4 ⟩ ψ k ⟨ K ω ⟩ ψ k ≤ C ⟨ K 5 ⟩ ψ k .
The last inequality is the nonnegative covariance of the increasing functions u 4 u^4 u 4 and u u u in the spectral distribution of K ω K_\omega K ω . Interchanging indices gives the creation-operator terms. Thus all defects have finite M − 1 M^{-1} M − 1 norm. They also have finite arithmetic dual norm, already from finite energy, because
∣ q l k ∣ 2 ( p l + p k ) ≤ C n ~ ( p l + p k ) ∣ a l k ′ ∣ 2 . \frac{|q_{lk}|^2}{(p_l+p_k)} \le C_{\tilde n}(p_l+p_k)|a'_{lk}|^2. ( p l + p k ) ∣ q l k ∣ 2 ≤ C n ~ ( p l + p k ) ∣ a l k ′ ∣ 2 .
Consequently their pairings with a bounded physical pullback agree with their completed M M M -pairings: approximate that pullback by centered eigenblocks in the arithmetic norm, which also gives M M M -convergence, and use both dual continuities. This identifies the functional transpose L ′ q L'q L ′ q , defined by ( L ′ q ) ( Z ) = q ( L Z ) (L'q)(Z)=q(LZ) ( L ′ q ) ( Z ) = q ( L Z ) , with the actual weak functional in (C4). Moreover, M ± ( p , p ) = p / h M_{\pm}(p,p)=p/h M ± ( p , p ) = p / h , so the identity is orthogonal to centered tests. Since q ( I ) = 0 q(\mathbb{I})=0 q ( I ) = 0 , restriction to the centered space does not alter the displayed dual norms.
Put
Q i n = ∑ i ∥ q i + ∥ M i , + − 1 2 , Q m a i n = ∑ j ∥ q m a i n , j + ∥ M 0 j , + − 1 2 , Q_{\mathrm{in}}=\sum_i\lVert q_i^+\rVert^2_{M^{-1}_{i,+}},\qquad Q_{\mathrm{main}}=\sum_j\lVert q_{\mathrm{main},j}^+\rVert^2_{M^{-1}_{0j,+}}, Q in = i ∑ ∥ q i + ∥ M i , + − 1 2 , Q main = j ∑ ∥ q main , j + ∥ M 0 j , + − 1 2 ,
q m a i n = w L ′ q = w q o u t . q_{\mathrm{main}}=\sqrt{w}\,L'q=wq_{\mathrm{out}}. q main = w L ′ q = w q out .
Transpose of the eigenvalue indices makes the minus and plus norms equal. Applying the dual part of Lemma 8 to both signs therefore yields
Q g a p : = w λ Q i n − Q m a i n ≥ 1 2 λ ∑ α i n b o t h s i g n s ∣ q m a i n , α ∣ 2 Δ α M α 2 ≥ 0. (111) Q_{\mathrm{gap}}:=w\lambda Q_{\mathrm{in}}-Q_{\mathrm{main}}\ge\frac{1}{2\lambda}\sum_{\alpha\ \mathrm{in\ both\ signs}}|q_{\mathrm{main},\alpha}|^2\frac{\Delta_\alpha}{M_\alpha^2}\ge0. \tag*{(111)} Q gap := w λ Q in − Q main ≥ 2 λ 1 α in both signs ∑ ∣ q main , α ∣ 2 M α 2 Δ α ≥ 0. (111)
In particular, for every bounded componentwise output test,
∣ q m a i n , j ± ( Z ) ∣ 2 ≤ C Q g a p Tr ρ 0 ( Z Z † + Z † Z ) ≤ 2 C Q g a p ∥ Z ∥ 2 . (112) |q_{\mathrm{main},j}^{\pm}(Z)|^2\le CQ_{\mathrm{gap}}\operatorname{Tr}\rho_0(ZZ^\dagger+Z^\dagger Z)\le2CQ_{\mathrm{gap}}\lVert Z\rVert^2. \tag*{(112)} ∣ q main , j ± ( Z ) ∣ 2 ≤ C Q gap Tr ρ 0 ( Z Z † + Z † Z ) ≤ 2 C Q gap ∥ Z ∥ 2 . (112)
The constants are uniform in the minimizing state and in ζ \zeta ζ .
The replacement error and its logarithmic tail
We now relate the balance defects of ρ 0 \rho_0 ρ 0 , which enters the variational functional, to those of the actual channel output ρ o u t \rho_{\mathrm{out}} ρ out . Let q 0 q_0 q 0 denote the functionals of ρ 0 \rho_0 ρ 0 with reference occupations n τ , j n_{\tau,j} n τ , j . Centering the mixture gives the exact identity
q 0 = q m a i n + q t a i l , q_0=q_{\mathrm{main}}+q_{\mathrm{tail}}, q 0 = q main + q tail ,
q t a i l , j + ( Z ) = θ w μ o u t , j Tr [ ( ρ o u t − τ ) Z ] . (113) q_{\mathrm{tail},j}^{+}(Z)=\theta w\mu_{\mathrm{out},j}\operatorname{Tr}[(\rho_{\mathrm{out}}-\tau)Z]. \tag*{(113)} q tail , j + ( Z ) = θ w μ out , j Tr [( ρ out − τ ) Z ] . (113)
The minus expression has the conjugate mean. In the ρ 0 \rho_0 ρ 0 eigenbasis, the absolute values of these coefficients are θ ∣ μ o u t , j ∣ ∣ ( ρ 0 − τ ) l k ∣ \theta|\mu_{\mathrm{out},j}||(\rho_0-\tau)_{lk}| θ ∣ μ out , j ∣∣ ( ρ 0 − τ ) l k ∣ . By the first inequality in (C13),
∥ q t a i l ∥ Δ − 1 2 ≤ C θ 2 ∥ μ o u t ∥ 2 ∑ l , k ∣ ( ρ 0 − τ ) l k ∣ 2 [ 1 + log 2 ( p l / p k ) ] p l + p k . (114) \lVert q_{\mathrm{tail}}\rVert^2_{\Delta^{-1}}\le C\theta^2\lVert\mu_{\mathrm{out}}\rVert^2\sum_{l,k}\frac{|(\rho_0-\tau)_{lk}|^2[1+\log^2(p_l/p_k)]}{p_l+p_k}. \tag*{(114)} ∥ q tail ∥ Δ − 1 2 ≤ C θ 2 ∥ μ out ∥ 2 l , k ∑ p l + p k ∣ ( ρ 0 − τ ) l k ∣ 2 [ 1 + log 2 ( p l / p k )] . (114)
Order gives τ ρ 0 − 1 τ ≤ τ / θ \tau\rho_0^{-1}\tau\le\tau/\theta τ ρ 0 − 1 τ ≤ τ / θ as a form. It bounds the arithmetic sum by O ( θ − 1 ) O(\theta^{-1}) O ( θ − 1 ) . The logarithmic sum contains only the off-diagonal entries of τ \tau τ ; using ( u − v ) 2 ≤ 2 u 2 + 2 v 2 (u-v)^2\le2u^2+2v^2 ( u − v ) 2 ≤ 2 u 2 + 2 v 2 bounds it by 4 Tr τ H 0 2 / θ 4\operatorname{Tr}\tau\mathcal{H}_0^2/\theta 4 Tr τ H 0 2 / θ . We control this logarithmic second moment separately as follows.
Let ( τ i , u i ) (\tau_i,u_i) ( τ i , u i ) and ( p k , v k ) (p_k,v_k) ( p k , v k ) be spectral data of τ \tau τ and ρ 0 \rho_0 ρ 0 , respectively, and use the probability distribution τ i ∣ ⟨ u i , v k ⟩ ∣ 2 \tau_i|\langle u_i,v_k\rangle|^2 τ i ∣ ⟨ u i , v k ⟩ ∣ 2 , as in [72 ] . For Y = log τ i − log p k Y=\log\tau_i-\log p_k Y = log τ i − log p k and 0 < s < 1 0<s<1 0 < s < 1 , inverse order and Schatten Hölder give
E e s Y = Tr τ 0 1 + s ρ 0 − s ≤ θ − s , E e − s Y = Tr τ 0 1 − s ρ 0 s ≤ 1. (115) \mathbb{E}e^{sY}=\operatorname{Tr}\tau_0^{1+s}\rho_0^{-s}\le\theta^{-s},\qquad\mathbb{E}e^{-sY}=\operatorname{Tr}\tau_0^{1-s}\rho_0^s\le1. \tag*{(115)} E e s Y = Tr τ 0 1 + s ρ 0 − s ≤ θ − s , E e − s Y = Tr τ 0 1 − s ρ 0 s ≤ 1. (115)
For the negative power first add a positive scalar to the operators and then take the monotone limit; all spectral sums are nonnegative. Since Y 2 ≤ C s ( e s Y + e − s Y ) Y^2\le C_s(e^{sY}+e^{-sY}) Y 2 ≤ C s ( e s Y + e − s Y ) ,
Tr τ H 0 2 ≤ 2 Tr τ ( log τ ) 2 + 2 C s ( θ − s + 1 ) . (116) \operatorname{Tr}\tau\mathcal{H}_0^2\le2\operatorname{Tr}\tau(\log\tau)^2+2C_s(\theta^{-s}+1). \tag*{(116)} Tr τ H 0 2 ≤ 2 Tr τ ( log τ ) 2 + 2 C s ( θ − s + 1 ) . (116)
The reference has finite logarithmic second moment. Taking s = 1 / 2 s=1/2 s = 1/2 in (C17)–(C19) proves the uniform bound
∥ q t a i l ∥ Δ − 1 2 ≤ C θ 1 / 2 ∥ μ o u t ∥ 2 . (117) \lVert q_{\mathrm{tail}}\rVert^2_{\Delta^{-1}}\le C\theta^{1/2}\lVert\mu_{\mathrm{out}}\rVert^2. \tag*{(117)} ∥ q tail ∥ Δ − 1 2 ≤ C θ 1/2 ∥ μ out ∥ 2 . (117)
Here C C C is independent of both 0 < θ < 1 / 2 0<\theta<1/2 0 < θ < 1/2 and ζ \zeta ζ .
Stationarity and the thermal limit
We evaluate stationarity along thermal Ornstein–Uhlenbeck evolution with invariant state σ \sigma σ . Its generator is
L σ ρ = ∑ i [ ( n σ , i + 1 ) D [ a i ] ρ + n σ , i D [ a i † ] ρ ] , \mathcal{L}_{\sigma}\rho=\sum_i[(n_{\sigma,i}+1)\mathcal{D}[a_i]\rho+n_{\sigma,i}\mathcal{D}[a_i^\dagger]\rho], L σ ρ = i ∑ [( n σ , i + 1 ) D [ a i ] ρ + n σ , i D [ a i † ] ρ ] ,
D [ a ] ρ = a ρ a † − 1 2 { a † a , ρ } . \mathcal{D}[a]\rho=a\rho a^\dagger-\frac{1}{2}\{a^\dagger a,\rho\}. D [ a ] ρ = a ρ a † − 2 1 { a † a , ρ } .
The output generator L τ \mathcal{L}_{\tau} L τ is defined analogously. Their linear drifts agree, and their diffusion covariances transform by (C1); hence
Φ L σ ρ = L τ Φ ρ . (118) \Phi\mathcal{L}_{\sigma}\rho=\mathcal{L}_{\tau}\Phi\rho. \tag*{(118)} Φ L σ ρ = L τ Φ ρ . (118)
For a finite number cutoff, this follows by applying the annihilator transformation to the characteristic function χ ~ ρ ( z ) = Tr ρ exp [ ∑ i ( z i a i † − z ˉ i a i ) ] \widetilde{\chi}_{\rho}(z)=\operatorname{Tr}\rho\exp[\sum_i(z_i a_i^\dagger-\bar z_i a_i)] χ ρ ( z ) = Tr ρ exp [ ∑ i ( z i a i † − z ˉ i a i )] , where z z z is a vector of complex amplitudes:
χ ^ L σ ρ ( z ) = − 1 2 ( z ⋅ ∂ z + z ˉ ⋅ ∂ z ˉ ) χ ^ ρ ( z ) − ∑ i ( n σ , i + 1 2 ) ∣ z i ∣ 2 χ ^ ρ ( z ) . \widehat{\chi}_{\mathcal{L}_{\sigma}\rho}(z)=-\frac{1}{2}(z\cdot\partial_z+\bar z\cdot\partial_{\bar z})\widehat{\chi}_{\rho}(z)-\sum_i\left(n_{\sigma,i}+\frac{1}{2}\right)|z_i|^2\widehat{\chi}_{\rho}(z). χ L σ ρ ( z ) = − 2 1 ( z ⋅ ∂ z + z ˉ ⋅ ∂ z ˉ ) χ ρ ( z ) − i ∑ ( n σ , i + 2 1 ) ∣ z i ∣ 2 χ ρ ( z ) .
The seventh-moment bound extends this identity from number cutoffs to the minimizing state. With A ρ = K 3 ρ K 3 A_\rho=K^3\rho K^3 A ρ = K 3 ρ K 3 , every term of K 2 L σ ρ K 2 K^2\mathcal{L}_{\sigma}\rho K^2 K 2 L σ ρ K 2 is a bounded-factor product around the trace-class operator A ρ A_\rho A ρ . For example,
K 2 a ρ a † K 2 = ( K 2 a K − 3 ) A ρ ( K − 3 a † K 2 ) . K^2a\rho a^\dagger K^2=(K^2aK^{-3})A_\rho(K^{-3}a^\dagger K^2). K 2 a ρ a † K 2 = ( K 2 a K − 3 ) A ρ ( K − 3 a † K 2 ) .
The anticommutator terms satisfy the same bound. Number compressions give K 3 P ρ L P ρ K 3 = P ρ A ρ P ρ → A ρ K^3P_\rho L P_\rho K^3=P_\rho A_\rho P_\rho\to A_\rho K 3 P ρ L P ρ K 3 = P ρ A ρ P ρ → A ρ in trace norm. The fixed-unitary bounds established above and partial trace transfer this weighted convergence to the output. Equality of the Weyl characteristic functions then gives (C21) for the minimizing state, with trace-zero generator expressions.
The bounded form K − 2 H ρ K − 2 K^{-2}\mathcal{H}_\rho K^{-2} K − 2 H ρ K − 2 defines its logarithmic pairing with K 2 L σ ρ K 2 K^2\mathcal{L}_\sigma\rho K^2 K 2 L σ ρ K 2 . The same input pairing applies to $\Phi^*\mathcal{H}_0,B, and K 4 K^{4} K 4 , while the output pairing with H 0 \mathcal{H}_{0} H 0 uses K o u t K_{\mathrm{out}} K out . To pull back the output logarithm, take the positive and negative parts of the sandwiched input generator and conjugate them by K − 2 K^{-2} K − 2 . Each resulting positive operator has finite fourth moment, so the positive form pullback identity applies. Subtraction gives
Tr L τ ρ H 0 = w Tr L σ ρ Φ ∗ H 0 . (119) \operatorname{Tr}\mathcal{L}_{\tau}\rho\mathcal{H}_{0}=w\operatorname{Tr}\mathcal{L}_{\sigma}\rho\Phi^{*}\mathcal{H}_{0}. \tag*{(119)} Tr L τ ρ H 0 = w Tr L σ ρ Φ ∗ H 0 . (119)
This identity uses the weighted trace pairing defined above. The input logarithmic pairing admits the identity
Tr L σ ρ H σ = Q i n − ∑ i h i ( E i − ∣ μ i ∣ 2 − n σ , i ) , E i = Tr ρ a i † a i . (120) \operatorname{Tr}\mathcal{L}_{\sigma}\rho\mathcal{H}_{\sigma}=Q_{\mathrm{in}}-\sum_i h_i(E_i-\lvert\mu_i\rvert^2-n_{\sigma,i}),\qquad E_i=\operatorname{Tr}\rho a_i^{\dagger}a_i. \tag*{(120)} Tr L σ ρ H σ = Q in − i ∑ h i ( E i − ∣ μ i ∣ 2 − n σ , i ) , E i = Tr ρ a i † a i . (120)
To derive Eq. (120) , work in the eigenbasis of the joint state and put k l k = ( n σ , i + 1 ) p k − n σ , i p l k_{lk}=(n_{\sigma,i}+1)p_k-n_{\sigma,i}p_l k l k = ( n σ , i + 1 ) p k − n σ , i p l and v l k = log ( p l / p k ) / h i v_{lk}=\log(p_l/p_k)/h_i v l k = log ( p l / p k ) / h i . The defect norm divided by h i h_i h i has summands ( 1 − v l k ) k l k ∣ a i , l k i ∣ 2 (1-v_{lk})k_{lk}\lvert a^i_{i,lk}\rvert^2 ( 1 − v l k ) k l k ∣ a i , l k i ∣ 2 , whereas the logarithmic pairing divided by h i h_i h i has summands − v l k k l k ∣ a i , l k i ∣ 2 -v_{lk}k_{lk}\lvert a^i_{i,lk}\rvert^2 − v l k k l k ∣ a i , l k i ∣ 2 . Their difference sums to E i − ∣ μ i ∣ 2 − n σ , i E_i-\lvert\mu_i\rvert^2-n_{\sigma,i} E i − ∣ μ i ∣ 2 − n σ , i by the canonical commutation relation. Centering changes only diagonal ladder entries, where v l l = 0 v_{ll}=0 v l l = 0 . The logarithmic sums are absolutely summable by the estimates above. For the anticommutator expansion, every state eigenvector belongs to D ( K 7 / 2 ) \mathcal{D}(K^{7/2}) D ( K 7/2 ) , so a i † a i ψ ∈ D ( K 2 ) a_i^{\dagger}a_i\psi\in\mathcal{D}(K^2) a i † a i ψ ∈ D ( K 2 ) ; finite eigenstate compressions converge in the K 3 / 2 K^{3/2} K 3/2 -sandwiched trace norm. Thus the same weighted pairing justifies the displayed spectral calculation.
Pair (C7) with L σ ρ \mathcal{L}_{\sigma}\rho L σ ρ , and use (C21), (C22), and (C23). The centered energy terms cancel because B = λ H σ − Φ ∗ H τ B=\lambda\mathcal{H}_{\sigma}-\Phi^{*}\mathcal{H}_{\tau} B = λ H σ − Φ ∗ H τ . The result is
0 = Q 0 − w λ Q i n − M m e a n + ζ P 4 , 0=Q_0-w\lambda Q_{\mathrm{in}}-M_{\mathrm{mean}}+\zeta P_4, 0 = Q 0 − w λ Q in − M mean + ζ P 4 ,
M m e a n = w [ λ ∑ i h i ∣ μ i ∣ 2 − w ∑ j h 0 j ∣ μ o u t , j ∣ 2 ] . (121) M_{\mathrm{mean}}=w\left[\lambda\sum_i h_i\lvert\mu_i\rvert^2-w\sum_j h_{0j}\lvert\mu_{\mathrm{out},j}\rvert^2\right]. \tag*{(121)} M mean = w [ λ i ∑ h i ∣ μ i ∣ 2 − w j ∑ h 0 j ∣ μ out , j ∣ 2 ] . (121)
where Q 0 = ∑ j ∣ q 0 j + ∣ M 0 , j + 2 Q_0=\sum_j\lvert q_{0j}^{+}\rvert^2_{M_{0,j}^{+}} Q 0 = ∑ j ∣ q 0 j + ∣ M 0 , j + 2 and P 4 = Tr L σ ρ K 4 P_4=\operatorname{Tr}\mathcal{L}_{\sigma}\rho K^4 P 4 = Tr L σ ρ K 4 . Strict quadratic positivity implies M m e a n ≥ w c ∥ μ ∥ 2 M_{\mathrm{mean}}\ge wc\lVert\mu\rVert^2 M mean ≥ w c ∥ μ ∥ 2 , while ∥ μ o u t ∥ ≤ C ∥ μ ∥ \lVert\mu_{\mathrm{out}}\rVert\le C\lVert\mu\rVert ∥ μ out ∥ ≤ C ∥ μ ∥ . Also P 4 ≤ C P_4\le C P 4 ≤ C , uniformly in the state: the number-coordinate formula for L ∗ K 4 \mathcal{L}_{*}K^4 L ∗ K 4 has leading term − 4 N 4 -4N^4 − 4 N 4 and remainder bounded above by C ( 1 + N ) 3 C(1+N)^3 C ( 1 + N ) 3 .
Expand Q 0 Q_0 Q 0 by (C16). The strict dual bound (C14), Cauchy–Schwarz with weights Δ / M 2 \Delta/M^2 Δ/ M 2 and Δ − 1 \Delta^{-1} Δ − 1 , and (C20) give
∣ cross term ∣ ≤ C θ 1 / 4 Q g a p ∥ μ ∥ , ∥ q t a i l ∥ M − 1 2 ≤ C θ 1 / 2 ∥ μ ∥ 2 . \lvert\text{cross term}\rvert\le C\theta^{1/4}\sqrt{Q_{\mathrm{gap}}}\lVert\mu\rVert,\qquad\lVert q_{\mathrm{tail}}\rVert_{M^{-1}}^2\le C\theta^{1/2}\lVert\mu\rVert^2. ∣ cross term ∣ ≤ C θ 1/4 Q gap ∥ μ ∥ , ∥ q tail ∥ M − 1 2 ≤ C θ 1/2 ∥ μ ∥ 2 .
The second bound uses Δ ≤ λ M \Delta\le\lambda M Δ ≤ λ M . The constants in these bounds are independent of 0 < θ < 1 / 2 0<\theta<1/2 0 < θ < 1/2 and ζ \zeta ζ . Choose θ < 1 / 2 \theta<1/2 θ < 1/2 sufficiently small once, then apply Young’s inequality in (C24). This yields
1 2 Q g a p + 1 2 w c ∥ μ ∥ 2 ≤ C ζ . (122) \frac{1}{2}Q_{\mathrm{gap}}+\frac{1}{2}wc\lVert\mu\rVert^2\le C\zeta. \tag*{(122)} 2 1 Q gap + 2 1 w c ∥ μ ∥ 2 ≤ C ζ . (122)
With the references, strict gap and replacement parameter θ \theta θ fixed, the uniform energy bound supplies a sequence ρ ζ → ρ ^ \rho_{\zeta}\to\widehat{\rho} ρ ζ → ρ in trace norm as ζ ↓ 0 \zeta\downarrow0 ζ ↓ 0 , with Φ ρ ζ → ρ ^ o u t : = Φ ( ρ ^ ) \Phi\rho_{\zeta}\to\widehat{\rho}_{\mathrm{out}}:=\Phi(\widehat{\rho}) Φ ρ ζ → ρ out := Φ ( ρ ) . Equation (C25) gives vanishing means. Equation (C15), applied to every fixed finite-number output matrix, gives q m a i n , j ± ( Z ) → 0 q_{\mathrm{main},j}^{\pm}(Z)\to0 q main , j ± ( Z ) → 0 . For such a test both ladder products are bounded finite-rank operators; hence their weak traces pass to the trace-norm limit. Using
q m a i n = w q o u t q_{\mathrm{main}}=wq_{\mathrm{out}} q main = w q out and the vanishing means, the limit satisfies the thermal balance relation on all number-state matrix elements:
( n τ , j + 1 ) b ρ ^ o u t = n τ , j ρ ^ o u t b j . (n_{\tau,j}+1)b\widehat{\rho}_{\mathrm{out}}=\ n_{\tau,j}\widehat{\rho}_{\mathrm{out}}b_j. ( n τ , j + 1 ) b ρ out = n τ , j ρ out b j .
Let n , m \boldsymbol{n},\boldsymbol{m} n , m be output number multiindices and e j e_j e j the j j j th unit vector. With s j = n τ , j / ( n τ , j + 1 ) s_j=n_{\tau,j}/(n_{\tau,j}+1) s j = n τ , j / ( n τ , j + 1 ) , the balance relation reads n j + 1 ⟨ ρ ^ o u t ∣ n + e j , m = m j ⟨ ρ ^ o u t ∣ n , m − e j \sqrt{n_j+1}\langle\widehat{\rho}_{\mathrm{out}}\rvert_{n+e_j,m}=\sqrt{m_j}\langle\widehat{\rho}_{\mathrm{out}}\rvert_{n,m-e_j} n j + 1 ⟨ ρ out ∣ n + e j , m = m j ⟨ ρ out ∣ n , m − e j . Iteration reaches a zero boundary element whenever n j > m j n_j>m_j n j > m j ; self-adjointness handles n j < m j n_j<m_j n j < m j . Thus the state is diagonal, and the diagonal recurrence and trace one give ρ ^ o u t = τ \widehat{\rho}_{\mathrm{out}}=\tau ρ out = τ .
Let J 0 J_0 J 0 denote the objective in Eq. (102) with ζ = 0 \zeta=0 ζ = 0 . Entropy continuity on the common energy sublevel and lower semicontinuity of B + C I B+\mathrm{CI} B + CI give J 0 ( ρ ) ≤ lim inf J ζ ( ρ ζ ) ≤ − β J_0(\rho)\le\liminf J_{\zeta}(\rho_{\zeta})\le-\beta J 0 ( ρ ) ≤ lim inf J ζ ( ρ ζ ) ≤ − β . But its output is τ \tau τ , so the exact Gibbs-reference identities give J 0 ( ρ ) = w D ( ρ ∥ ρ ^ ) ≥ 0 J_0(\rho)=wD(\rho\Vert\widehat{\rho})\ge0 J 0 ( ρ ) = w D ( ρ ∥ ρ ) ≥ 0 . This contradiction completes the proof of Proposition 2.
Four-weight interpolation with a strict remainder
The following lemma strengthens the diagonal-form interpolation theorem of Ref. [30 Theorem 3.1] by retaining a quantitative gap from one pair of endpoint bounds. After duality, this gap controls the thermal balance defects in Appendix C.
Lemma 8 (Strict four-weight comparison). Put f ( x ) = x / sinh x f(x)=x/\sinh x f ( x ) = x / sinh x , with f ( 0 ) = 1 f(0)=1 f ( 0 ) = 1 . At each source and target coordinate fix m > 0 m>0 m > 0 and real x , y x,y x , y , with any fixed positive coordinate measure. In coordinate formulas for dual norms, that measure is absorbed into the weights. Let T \mathcal{T} T be a source test space such that each element is supported on coordinate sets at which ( m , x , y ) (m,x,y) ( m , x , y ) takes finitely many values, and suppose T \mathcal{T} T is invariant under multiplication by functions of this tuple. The subscripts “out” and “in” label the source and target, respectively, as in the observable pullback in Appendix C. We say that a linear map L L L from T \mathcal{T} T into the target coordinate space compares a positive diagonal weight ω \omega ω with constant Λ \Lambda Λ if ∥ L Z ∥ ω , i n 2 ≤ Λ ∥ Z ∥ ω , o u t 2 \lVert LZ\rVert^2_{\omega,\mathrm{in}}\le\Lambda\lVert Z\rVert^2_{\omega,\mathrm{out}} ∥ L Z ∥ ω , in 2 ≤ Λ ∥ Z ∥ ω , out 2 for every Z ∈ T Z\in\mathcal{T} Z ∈ T . Suppose L L L compares m f ( x ) e ± x mf(x)e^{\pm x} m f ( x ) e ± x with constant λ − δ \lambda-\delta λ − δ , and m f ( y ) e ± y mf(y)e^{\pm y} m f ( y ) e ± y with constant λ \lambda λ , where 0 < δ < λ 0<\delta<\lambda 0 < δ < λ . Then, for M = m f ( x ) f ( y ) / f ( x + y ) M=mf(x)f(y)/f(x+y) M = m f ( x ) f ( y ) / f ( x + y ) ,
∥ L Z ∥ M , i n 2 ≤ λ ∥ Z ∥ M , o u t 2 − ∥ Z ∥ Δ 2 , \lVert LZ\rVert^2_{M,\mathrm{in}}\le\lambda\lVert Z\rVert^2_{M,\mathrm{out}}-\lVert Z\rVert^2_{\Delta}, ∥ L Z ∥ M , in 2 ≤ λ ∥ Z ∥ M , out 2 − ∥ Z ∥ Δ 2 , with a strictly positive source weight Δ ≤ δ M \Delta\le\delta M Δ ≤ δ M given in (D5). The continuous extension between the completed form spaces has the dual estimate (D6).
Proof. Positive sums, pointwise limits and parallel sums preserve a common comparison constant. The parallel sum X : Y = X Y / ( X + Y ) X:Y=XY/(X+Y) X : Y = X Y / ( X + Y ) satisfies
( X : Y ) ∣ z ∣ 2 = min u + v = z ( X ∣ u ∣ 2 + Y ∣ v ∣ 2 ) , ( u , v ) = ( Y , X ) X + Y z . (X:Y)\lvert z\rvert^2=\min_{u+v=z}(X\lvert u\rvert^2+Y\lvert v\rvert^2),\qquad(u,v)=\frac{(Y,X)}{X+Y}z. ( X : Y ) ∣ z ∣ 2 = u + v = z min ( X ∣ u ∣ 2 + Y ∣ v ∣ 2 ) , ( u , v ) = X + Y ( Y , X ) z .
The source minimizing split belongs to T \mathcal{T} T ; on the target one uses only the lower bound by the minimum. For pointwise limits, a source test has finitely many parameter values, and Fatou’s lemma applies to target coordinate sums. These arguments work with unbounded diagonal forms. Positive quadratic sums and parallel sums give the two logarithmic-mean operations
ℓ ( X , Y ) − 1 = ∫ 0 1 d u ( 1 − u ) X + u Y , \ell(X,Y)^{-1}=\int_0^1\frac{du}{(1-u)X+uY}, ℓ ( X , Y ) − 1 = ∫ 0 1 ( 1 − u ) X + u Y d u ,
X Y ℓ ( X , Y ) = ∫ 0 1 d u u / X + ( 1 − u ) / Y . \frac{XY}{\ell(X,Y)}=\int_0^1\frac{du}{u/X+(1-u)/Y}. ℓ ( X , Y ) X Y = ∫ 0 1 u / X + ( 1 − u ) / Y d u .
Applied to the opposite-exponent pairs, they give comparisons for m , m s , m t m,m_s,m_t m , m s , m t , where s = f ( x ) 2 s=f(x)^2 s = f ( x ) 2 , t = f ( y ) 2 t=f(y)^2 t = f ( y ) 2 .
The remaining scalar ingredient is the Stieltjes representation proved in Ref. [30, Lemma 3.2]. The function defined by B ( f ( x ) 2 ) = b ( x ) = x coth x \mathcal{B}(f(x)^2)=b(x)=x\coth x B ( f ( x ) 2 ) = b ( x ) = x coth x , with b ( 0 ) = 1 b(0)=1 b ( 0 ) = 1 , extends across the positive axis and satisfies
− B ′ ( s ) = c + ∫ [ 0 , ∞ ) d ν ( r ) s + r > 0 , c ≥ 0 , ∫ d ν ( r ) 1 + r < ∞ . (123) -\mathcal{B}'(s)=c+\int_{[0,\infty)}\frac{d\nu(r)}{s+r}>0,\qquad c\geq0,\qquad\int\frac{d\nu(r)}{1+r}<\infty. \tag*{(123)} − B ′ ( s ) = c + ∫ [ 0 , ∞ ) s + r d ν ( r ) > 0 , c ≥ 0 , ∫ 1 + r d ν ( r ) < ∞. (123)
This representation supplies the positive combinations needed below. Define
1 E = 2 m ∫ 0 1 [ − B ′ ( u s + ( 1 − u ) t ) ] d u , E = m ( t − s ) 2 [ B ( s ) − B ( t ) ] , (124) \frac{1}{E}=\frac{2}{m}\int_0^1[-\mathcal{B}'(us+(1-u)t)]\,du,\qquad E=\frac{m(t-s)}{2[\mathcal{B}(s)-\mathcal{B}(t)]}, \tag*{(124)} E 1 = m 2 ∫ 0 1 [ − B ′ ( u s + ( 1 − u ) t )] d u , E = 2 [ B ( s ) − B ( t )] m ( t − s ) , (124)
with E = − m / [ 2 B ′ ( s ) ] E=-m/[2\mathcal{B}'(s)] E = − m / [ 2 B ′ ( s )] at s = t s=t s = t . The Stieltjes formula constructs E E E from m , m s , m t m,m_s,m_t m , m s , m t by positive combinations and parallel sums: each reciprocal integrand has denominator u ( m s ) + ( 1 − u ) ( m t ) + r m u(m_s)+(1-u)(m_t)+rm u ( m s ) + ( 1 − u ) ( m t ) + r m , together with a possible multiple of 1 / m 1/m 1/ m . Truncate the positive measure and use positive quadrature sums, then the preceding limit argument. Thus E E E also compares with constant λ \lambda λ .
For scalars ξ , ϕ \xi,\phi ξ , ϕ , put
Q ( ξ , ϕ ) = 1 4 ∑ ± m f ( x ) e ± x ∣ ξ ∓ ϕ ∣ 2 + 1 4 ∑ ± m f ( y ) e ± y ∣ ξ ± ϕ ∣ 2 + E ∣ ϕ ∣ 2 . (125) Q(\xi,\phi)=\frac{1}{4}\sum_{\pm}mf(x)e^{\pm x}|\xi\mp\phi|^2+\frac{1}{4}\sum_{\pm}mf(y)e^{\pm y}|\xi\pm\phi|^2+E|\phi|^2. \tag*{(125)} Q ( ξ , ϕ ) = 4 1 ± ∑ m f ( x ) e ± x ∣ ξ ∓ ϕ ∣ 2 + 4 1 ± ∑ m f ( y ) e ± y ∣ ξ ± ϕ ∣ 2 + E ∣ ϕ ∣ 2 . (125)
With a = b ( x ) + b ( y ) a=b(x)+b(y) a = b ( x ) + b ( y ) , e = E / m e=E/m e = E / m , direct minimization gives
ϕ = κ ξ , κ = x − y a + 2 e , \phi=\kappa\xi,\qquad\kappa=\frac{x-y}{a+2e}, ϕ = κ ξ , κ = a + 2 e x − y ,
min ϕ Q = m 2 [ a − ( x − y ) 2 a + 2 e ] ∣ ξ ∣ 2 = M ∣ ξ ∣ 2 . (126) \min_{\phi}Q=\frac{m}{2}\left[a-\frac{(x-y)^2}{a+2e}\right]|\xi|^2=M|\xi|^2. \tag*{(126)} ϕ min Q = 2 m [ a − a + 2 e ( x − y ) 2 ] ∣ ξ ∣ 2 = M ∣ ξ ∣ 2 . (126)
The last equality follows from b ( x ) 2 = x 2 + f ( x ) 2 b(x)^2=x^2+f(x)^2 b ( x ) 2 = x 2 + f ( x ) 2 and the addition formula for sinh. For x 2 ≠ y 2 x^2\ne y^2 x 2 = y 2 , the same identity and (D2) give a + 2 e = ( x 2 − y 2 ) / [ b ( x ) − b ( y ) ] a+2e=(x^2-y^2)/[b(x)-b(y)] a + 2 e = ( x 2 − y 2 ) / [ b ( x ) − b ( y )] . Consequently, by continuity, κ = [ b ( x ) − b ( y ) ] / ( x + y ) \kappa=[b(x)-b(y)]/(x+y) κ = [ b ( x ) − b ( y )] / ( x + y ) whenever x + y ≠ 0 x+y\ne0 x + y = 0 . At x = y x=y x = y the minimum is m b ( x ) ∣ ξ ∣ 2 mb(x)|\xi|^2 mb ( x ) ∣ ξ ∣ 2 ; at x = − y x=-y x = − y it is m f ( x ) 2 ∣ ξ ∣ 2 mf(x)^2|\xi|^2 m f ( x ) 2 ∣ ξ ∣ 2 . The zero cases follow by continuity.
Use the source minimizing split ϕ = κ Z \phi=\kappa Z ϕ = κ Z , which is admissible by tuple invariance. Apply the first pair of comparisons with constant λ − δ \lambda-\delta λ − δ , and the remaining comparisons with constant λ \lambda λ . On the target minimize freely. The retained source contribution is
Δ = δ m f ( x ) 4 [ e x ( 1 − κ ) 2 + e − x ( 1 + κ ) 2 ] . (127) \Delta=\frac{\delta mf(x)}{4}\left[e^x(1-\kappa)^2+e^{-x}(1+\kappa)^2\right]. \tag*{(127)} Δ = 4 δ m f ( x ) [ e x ( 1 − κ ) 2 + e − x ( 1 + κ ) 2 ] . (127)
It is strictly positive, and Δ ≤ δ M \Delta\leq\delta M Δ ≤ δ M because it is δ \delta δ times the first positive contribution to the minimum in (D3). The minimizing multiplier is applied only on the source space.
To obtain the dual estimate, let A = M i n 1 / 2 L M o u t − 1 / 2 \mathcal{A}=M_{\mathrm{in}}^{1/2}LM_{\mathrm{out}}^{-1/2} A = M in 1/2 L M out − 1/2 on the completed spaces, and R = M o u t − 1 / 2 Δ M o u t − 1 / 2 R=M_{\mathrm{out}}^{-1/2}\Delta M_{\mathrm{out}}^{-1/2} R = M out − 1/2 Δ M out − 1/2 . The interpolated bound reads A † A ≤ λ I − R \mathcal{A}^{\dagger}\mathcal{A}\leq\lambda I-R A † A ≤ λ I − R , where 0 ≤ R ≤ δ I 0\leq R\leq\delta I 0 ≤ R ≤ δ I . Tuple invariance ensures that the multiplier preserves the closure of T \mathcal{T} T . For an input functional q q q , set v = M i n − 1 / 2 q v=M_{\mathrm{in}}^{-1/2}q v = M in − 1/2 q , u = A † v u=\mathcal{A}^{\dagger}v u = A † v . Expansion of λ ∥ v − A u / λ ∥ 2 ≥ 0 \lambda\lVert v-\mathcal{A}u/\lambda\rVert^2\geq0 λ ∥ v − A u / λ ∥ 2 ≥ 0 then gives
λ ∥ q ∥ M i n − 1 2 − ∥ L ′ q ∥ M o u t − 1 2 ≥ 1 λ ∑ α ∣ ( L ′ q ) α ∣ 2 Δ α M α 2 . (128) \lambda\lVert q\rVert_{M_{\mathrm{in}}^{-1}}^2-\lVert L'q\rVert_{M_{\mathrm{out}}^{-1}}^2\geq\frac{1}{\lambda}\sum_{\alpha}|(L'q)_\alpha|^2\frac{\Delta_\alpha}{M_\alpha^2}. \tag*{(128)} λ ∥ q ∥ M in − 1 2 − ∥ L ′ q ∥ M out − 1 2 ≥ λ 1 α ∑ ∣ ( L ′ q ) α ∣ 2 M α 2 Δ α . (128)
Here coefficients are identified by their Riesz representation; L ′ q L'q L ′ q is the functional Z ↦ q ( L Z ) Z\mapsto q(LZ) Z ↦ q ( L Z ) . The estimates extend to the completed spaces by boundedness of A \mathcal{A} A and R R R . □
Gaussian channel simulation estimates
This appendix proves the environment comparison, resource-preparation estimate, and symplectic extension used in Section III, together with the probe reduction for the correlated-loss example in Sec. V B.
Environment comparison
The vacuum environment ρ E , 0 \rho_{E,0} ρ E , 0 and the Gaussian state ρ E , s \rho_{E,s} ρ E , s in Eq. (46) pass through the same dilation unitary. Contractivity and the overlap with the vacuum give
d ⋄ ( Φ s , Φ ) ≤ d t r ( ρ E , s , ρ E , 0 ) ≤ 1 − det ( I 2 k + ϵ s 2 L B X X T L B T ) − 1 / 2 . (129) \begin{aligned}
d_{\diamond}(\Phi_s,\Phi)\leq d_{\mathrm{tr}}(\rho_{E,s},\rho_{E,0}) \\
&\leq\sqrt{1-\det\left(I_{2k}+\frac{\epsilon_s}{2}L_BXX^{\mathsf T}L_B^{\mathsf T}\right)^{-1/2}}. \tag*{(129)}
\end{aligned} d ⋄ ( Φ s , Φ ) ≤ d tr ( ρ E , s , ρ E , 0 ) ≤ 1 − det ( I 2 k + 2 ϵ s L B X X T L B T ) − 1/2 . (129)
Since L B T L B = Y − 1 L_B^{\mathsf T}L_B=Y^{-1} L B T L B = Y − 1 , Sylvester’s determinant identity reduces the determinant to det ( I 2 m + ϵ s X T Y − 1 X / 2 ) \det(I_{2m}+\epsilon_sX^{\mathsf T}Y^{-1}X/2) det ( I 2 m + ϵ s X T Y − 1 X /2 ) . Using log det ( I + A ) ≤ Tr A \log\det(I+A)\leq\operatorname{Tr}A log det ( I + A ) ≤ Tr A , 1 − e − u ≤ u 1-e^{-u}\leq u 1 − e − u ≤ u , and ϵ s ≤ 2 / s \epsilon_s\leq2/s ϵ s ≤ 2/ s proves Eq. (47) .
The following dilution bound depends on the number of modes and is uniform in the resource energy.
Lemma 9 (Uniform Schmidt information variance). For a pure Gaussian state φ A B \varphi^{AB} φ A B with at most m m m modes on Alice’s side, the reduced state satisfies
Tr ρ A [ − log ρ A − S ( ρ A ) ] 2 ≤ m . (130) \operatorname{Tr}\rho_A[-\log\rho_A-S(\rho_A)]^2\leq m. \tag*{(130)} Tr ρ A [ − log ρ A − S ( ρ A ) ] 2 ≤ m . (130)
Here log φ A \log\varphi_A log φ A is restricted to the support of φ A \varphi_A φ A . For any sequence of such states φ n \varphi_n φ n and any a n > 0 a_n>0 a n > 0 , n n n copies of φ n \varphi_n φ n can consequently be approximated by LOCC from a maximally entangled state of Schmidt rank
M n = ⌈ exp { n [ S ( ( φ n ) A ) + a n ] } ⌉ (131) M_n=\left\lceil\exp\{n[S((\varphi_n)_A)+a_n]\}\right\rceil \tag*{(131)} M n = ⌈ exp { n [ S (( φ n ) A ) + a n ]} ⌉ (131)
with trace-distance error at most m / ( n a n 2 ) \sqrt{m/(na_n^2)} m / ( n a n 2 ) , whenever that bound is less than one.
Proof. The Gaussian Schmidt decomposition expresses φ A \varphi_A φ A , up to a unitary, as a product of at most m m m thermal states. Their information variances are bounded by one in Eq. (51) ; vacuum factors contribute zero. Adding variances proves Eq. (130) .
For each n n n , let P n P_n P n project onto eigenvalues of ( φ n ) A ⊗ n (\varphi_n)_A^{\otimes n} ( φ n ) A ⊗ n at least exp { − n [ S ( ( φ n ) A ) + a n ] } \exp\{-n[S((\varphi_n)_A)+a_n]\} exp { − n [ S (( φ n ) A ) + a n ]} . Chebyshev’s inequality gives
p n : = Tr P n ( φ n ) A ⊗ n ≥ 1 − m n a n 2 , (132) p_n := \operatorname{Tr}P_n(\varphi_n)_A^{\otimes n} \ge1-\frac{m}{na_n^2}, \tag*{(132)} p n := Tr P n ( φ n ) A ⊗ n ≥ 1 − n a n 2 m , (132)
rank P n ≤ exp { n [ S ( ( φ n ) A ) + a n ] } . \operatorname{rank}P_n \le\exp\{n[S((\varphi_n)_A)+a_n]\}. rank P n ≤ exp { n [ S (( φ n ) A ) + a n ]} .
The normalized projection of φ n ⊗ n \varphi_n^{\otimes n} φ n ⊗ n has Schmidt rank at most M n M_n M n and can be prepared by finite-dimensional teleportation. Its trace distance from φ n ⊗ n \varphi_n^{\otimes n} φ n ⊗ n is 1 − p n \sqrt{1-p_n} 1 − p n .
Exact simulation with finite-energy resources
For Y > 0 Y>0 Y > 0 , the full-row-rank matrices M t e l M_{\mathrm{tel}} M tel and B B B in Eq. (49) have the same symplectic Gram matrix. Pad them with zero columns to a common finite symplectic space. The correspondence between their rows is an isometry of the restricted alternating forms. By Witt’s extension theorem for symplectic spaces, it extends to a symplectic automorphism of the whole space. To construct the extension, choose canonical pairs in each row space, complete the nullspace of its restricted form by conjugate partners, and then complete to symplectic bases. Mapping the two bases gives a symplectic matrix Ξ \Xi Ξ such that
M ~ t e l Ξ = B ~ . (133) \widetilde{M}_{\mathrm{tel}}\Xi=\widetilde{B}. \tag*{(133)} M tel Ξ = B . (133)
The pure covariance Ξ Ξ T \Xi\Xi^{\mathsf T} Ξ Ξ T has a marginal V r e s V_{\mathrm{res}} V res on the original m + ℓ m+\ell m + ℓ resource modes satisfying M t e l V r e s M t e l T = B B T = Y M_{\mathrm{tel}}V_{\mathrm{res}}M_{\mathrm{tel}}^{\mathsf T}=BB^{\mathsf T}=Y M tel V res M tel T = B B T = Y .
Writing χ ρ ( z ) = Tr ρ e i z T R \chi_\rho(z)=\operatorname{Tr}\rho e^{iz^{\mathsf T}R} χ ρ ( z ) = Tr ρ e i z T R for the characteristic function, teleportation with gain X X X and the centered Gaussian resource ρ r e s \rho_{\mathrm{res}} ρ res of covariance V r e s V_{\mathrm{res}} V res gives, for an arbitrary input and every z ∈ R 2 ℓ z\in\mathbb{R}^{2\ell} z ∈ R 2 ℓ ,
χ o u t ( z ) = χ i n ( X T z ) χ ρ r e s ( − Z m X T z , z ) = χ i n ( X T z ) e − z T Y z / 4 . \begin{aligned}
\chi_{\mathrm{out}}(z)&=\chi_{\mathrm{in}}(X^{\mathsf T}z)\chi_{\rho_{\mathrm{res}}}(-Z_mX^{\mathsf T}z,z)\\
&=\chi_{\mathrm{in}}(X^{\mathsf T}z)e^{-z^{\mathsf T}Yz/4}.
\end{aligned} χ out ( z ) = χ in ( X T z ) χ ρ res ( − Z m X T z , z ) = χ in ( X T z ) e − z T Y z /4 .
The resulting map agrees with Φ \Phi Φ on all input characteristic functions, establishing exact simulation, including for inputs entangled with a reference.
We establish the lower bound in Eq. (84) for the channel of § 1.4 B. The pure resource constructed there gives the matching upper bound, so it remains to evaluate the large-squeezing limit of the reference–output probes in Theorem 4 . Pure covariances below a Gaussian probe covariance can be parameterized by rank-one Gaussian measurements on its purification [5, 13]. The inverse conditioning relation makes this correspondence explicit. A pure covariance below the probe covariance coincides with it on its pure normal modes, by the vacuum-factorization argument after Eq. (67) , so only the mixed normal modes matter.
In mixed Schmidt normal modes, write D = ⨁ j ν j I 2 D=\bigoplus_j\nu_j I_2 D = ⨁ j ν j I 2 , ν j > 1 \nu_j>1 ν j > 1 , and C = ⨁ j ν j 2 − 1 I C=\bigoplus_j\sqrt{\nu_j^2-1}I C = ⨁ j ν j 2 − 1 I . A pure covariance γ ≺ D \gamma\prec D γ ≺ D is obtained by measuring the purifying modes with covariance
M = C T ( D − γ ) − 1 C − D . M=C^{\mathsf T}(D-\gamma)^{-1}C-D. M = C T ( D − γ ) − 1 C − D .
Substitution of γ Ω γ = Ω \gamma\Omega\gamma=\Omega γ Ω γ = Ω shows that M M M is positive and pure. Boundary seeds follow by limits.
In the present dilation of the reference–output probe, the environment covariance V E , r ( s ) = s L 2 + X γ r X T V_{E,r}(s)=sL^2+X\gamma_rX^{\mathsf T} V E , r ( s ) = s L 2 + X γ r X T satisfies V E , r ( s ) + i Ω 2 ≥ ( s − 1 ) L 2 > 0 V_{E,r}(s)+i\Omega_2\ge(s-1)L^2>0 V E , r ( s ) + i Ω 2 ≥ ( s − 1 ) L 2 > 0 for s > 1 s>1 s > 1 . Both symplectic eigenvalues of the environment covariance therefore exceed one, so these two modes provide a minimal Gaussian purification of the probe output, related to the purification used above by a symplectic transformation of the environment. The reference–environment cross block − c s Z 2 L -c_sZ_2L − c s Z 2 L is invertible, so the conditional covariance determines the measurement covariance uniquely. The probe symmetries are ( − K ) R ⊕ K B (-K)R\oplus K_B ( − K ) R ⊕ K B and Θ R ⊕ Θ B \Theta_R\oplus\Theta_B Θ R ⊕ Θ B , with actions K E K_E K E and Θ E \Theta_E Θ E on the environment. Taking the two geometric midpoints imposes these symmetries on a minimizing pure probe seed, and uniqueness transfers them to M M M . A two-mode covariance invariant under K K K and Θ \Theta Θ has local blocks a I 2 aI_2 a I 2 , b I 2 bI_2 b I 2 and correlation block c Z cZ c Z , and purity forces a = b a=b a = b and a 2 − c 2 = 1 a^2-c^2=1 a 2 − c 2 = 1 . Consequently M = γ r M=\gamma_r M = γ r for some r ∈ R r\in\mathbb{R} r ∈ R . For boundary seeds, geometric interpolation toward the invariant seed obtained by measuring vacuum makes the covariance difference strictly positive on the mixed Schmidt modes. Applying the inverse relation and then taking the interpolation parameter to zero preserves the same reduction.
Gaussian conditioning now gives the reference marginal
A s ( t ) = s I 4 − ( s 2 − 1 ) Z 2 L 2 [ s L 2 + X γ r X T + γ t ] − 1 L Z 2 = Z 2 f s ( W r ( t ) ) Z 2 , f s ( W ) = ( s W + I 4 ) ( s I 4 + W ) − 1 . (134) \begin{aligned}
A_s(t)&=sI_4-(s^2-1)Z_2L_2[sL^2+X\gamma_rX^{\mathsf T}+\gamma_t]^{-1}LZ_2\\
&=Z_2f_s(W_r(t))Z_2,\qquad f_s(W)=(sW+I_4)(sI_4+W)^{-1}. \tag*{(134)}
\end{aligned} A s ( t ) = s I 4 − ( s 2 − 1 ) Z 2 L 2 [ s L 2 + X γ r X T + γ t ] − 1 L Z 2 = Z 2 f s ( W r ( t )) Z 2 , f s ( W ) = ( s W + I 4 ) ( s I 4 + W ) − 1 . (134)
Thus G ( Φ r ( s ) ) = inf t S G ( f s ( W r ( t ) ) ) G(\Phi_r(s))=\inf_t S_G(f_s(W_r(t))) G ( Φ r ( s )) = inf t S G ( f s ( W r ( t ))) . The remaining step is to identify its limit at large squeezing with the scalar minimum in Eq. (84) . For fixed t t t , f s ( W r ( t ) ) → W r ( t ) f_s(W_r(t))\to W_r(t) f s ( W r ( t )) → W r ( t ) as s → ∞ s\to\infty s → ∞ . To pass the limit through the infimum, observe that
W r ( t ) ≥ L − 1 X γ r X T L − 1 > 0 W_r(t)\ge L^{-1}X\gamma_rX^{\mathsf T}L^{-1}>0 W r ( t ) ≥ L − 1 X γ r X T L − 1 > 0
uniformly in t t t , while its largest eigenvalue diverges with ∣ t ∣ |t| ∣ t ∣ . For s ≥ 1 s\ge1 s ≥ 1 and w > 0 w>0 w > 0 ,
f s ( w ) ≥ min { 1 , w } , f s ( w ) ≥ 1 2 min { s , w } . f_s(w)\ge\min\{1,w\},\qquad f_s(w)\ge\frac{1}{2}\min\{s,w\}. f s ( w ) ≥ min { 1 , w } , f s ( w ) ≥ 2 1 min { s , w } .
Therefore det f s ( W r ( t ) ) → ∞ \det f_s(W_r(t))\to\infty det f s ( W r ( t )) → ∞ along every joint limit s → ∞ s\to\infty s → ∞ , ∣ t ∣ → ∞ |t|\to\infty ∣ t ∣ → ∞ . Bounded-entropy minimizing sequences stay in a compact interval of t t t , on which convergence is uniform. The channel cost theorem then gives Eq. (84) .
Dilations, ensembles, and constrained coding
We give the Gaussian dilation and ensemble details used in Section IV, followed by the approximation that turns continuous Gaussian modulation into codes with a finite signal alphabet.
Rectangular dilations and continuous ensembles
For equal input and output mode counts, every Gaussian channel has a Gaussian unitary dilation with a finite vacuum environment, including rank-deficient gains and noise [33 Corollary 1] . If ℓ > m \ell> m ℓ > m , extend the input by ℓ − m \ell-m ℓ − m modes discarded by the channel, replacing X X X by ( X 0 ) (X\ 0) ( X 0 ) . Prepare these modes in vacuum and count them as part of the environment. If ℓ < m \ell< m ℓ < m , append m − ℓ m-\ell m − ℓ vacuum outputs, replacing X X X by ( X 0 ) \begin{pmatrix}X\\0\end{pmatrix} ( X 0 ) and Y Y Y by Y ⊕ I 2 ( m − ℓ ) Y \oplus I_{2(m-\ell)} Y ⊕ I 2 ( m − ℓ ) , and include them among the discarded modes. In both constructions the extended complete-positivity condition contains the original condition as a block. This gives Eq. (65) ; an idle vacuum mode may also be added to F F F .
For the ensemble correspondence in Eq. (66) , let Q Q Q project onto the orthogonal complement of the range of T T T . Since Tr ( Q T ρ T † ) = 0 \operatorname{Tr}(Q T \rho T^\dagger)=0 Tr ( QT ρ T † ) = 0 and every ensemble expectation of Q Q Q is nonnegative, almost every vector of a pure ensemble lies in the range of T T T .
Energy bounds and finite-alphabet coding
A confining quadratic Hamiltonian bounds the total input photon number. Since a Gaussian channel acts affinely on first and second moments, its output photon number is bounded on every input energy-constrained set, and the thermal maximum-entropy bound gives
sup Tr ρ H ≤ E S ( Φ ( ρ ) ) < ∞ . (135) \sup_{\operatorname{Tr}\rho H\le E} S(\Phi(\rho)) < \infty. \tag*{(135)} Tr ρ H ≤ E sup S ( Φ ( ρ )) < ∞. (135)
The same holds for the pure dilation and every fixed number of uses. This is the hypothesis of the constrained coding theorem, applied to the nonnegative Hamiltonian H − E 0 I H-E_0 I H − E 0 I [46 Proposition 3] , which yields
C ( Φ , H , E ) = lim n → ∞ χ n ( E ) n . (136) C(\Phi,H,E)=\lim_{n\to\infty}\frac{\chi_n(E)}{n}. \tag*{(136)} C ( Φ , H , E ) = n → ∞ lim n χ n ( E ) . (136)
Its direct construction obeys the mean energy bound on each codeword; the converse also holds for energy averaged over messages.
A Gaussian displacement ensemble with mean energy strictly below E E E can be approximated by finite ensembles of pure Gaussian signals: truncate the displacement tail and discretize the remaining compact region within its support, which may be a proper subspace. The averages converge in trace norm with uniformly bounded output energy, so their entropies converge [25 ] . Every signal output has the same entropy, and sufficiently accurate approximations respect the energy constraint. Coding with this finite alphabet uses product displaced copies of the seed and collective measurements. For E > E 0 E>E_0 E > E 0 , continuity of the finite concave capacity expression allows the energy of these ensembles to approach E E E , giving the claimed rate.
References