Quantum information · Theory (AI-assisted)

Entanglement cost and classical capacity of multimode Gaussian states and channels

Yuan Gao, Gerardo Adesso

arXiv:2611.22915v1 ↗PDF ↗Submitted 17 November 2026

Question
For a Gaussian state shared between two parties, a natural preparation starts from a pure Gaussian state and adds correlated random displacements, and the least entanglement such recipes need is the Gaussian entanglement of formation. Are these recipes optimal among all ensembles, Gaussian or not, and does preparing many copies together ever lower the cost per copy? The answers were known for two-mode states and a few symmetric families but not for arbitrary numbers of modes, and the same questions govern the entanglement cost and classical capacity of multimode Gaussian channels.
Result
For every finite-energy bipartite Gaussian state with any finite number of modes on each side, the entanglement cost, the entanglement of formation and the Gaussian entanglement of formation coincide, all equal to a covariance minimum \(G(V)\). The common value is strongly superadditive and additive on tensor products, and any state with covariance \(V\), Gaussian or not, has entanglement of formation and cost at least \(G(V)\). The entanglement cost of every multimode Gaussian channel is a limit of \(G\) over probe states, finite exactly when the noise matrix is positive definite; its energy-constrained classical capacity is single-letter and attained by Gaussian-modulated pure Gaussian signals. A family of four-mode bound entangled states gets a closed-form cost.

The paper in detail

A Gaussian state of bosonic modes is fixed by its first moments and its covariance matrix \(V\), normalised here so that the vacuum has \(V=\mathbb I\). Any bipartite Gaussian state can be prepared by taking a pure Gaussian state with covariance \(\Gamma\le V\) and adding Gaussian random displacements with covariance \((V-\Gamma)/2\). Displacements are local, so every member of the ensemble has the entanglement of the seed, and the cheapest such recipe defines \(G(V)\), the minimum of the seed's entanglement over pure \(\Gamma\le V\). The entanglement of formation allows any ensemble, and the entanglement cost is its regularisation over many copies. Proving \(E_{\mathrm C}=E_{\mathrm F}=G\) means showing that no non-Gaussian ensemble and no collective preparation does better.

The proof has two parts. Inside the Gaussian world, the geometric mean of two pure covariances is again pure and respects \(\Gamma\le V\), and the seed's entanglement is convex along these geometric paths. That gives strong superadditivity: averaging a seed with its image under a sign flip of one group of modes removes the cross-correlations without raising the entanglement. It also makes \(G\) convex in \(V\), so it has a supporting hyperplane at every target, which takes the form of a quadratic Hamiltonian \(H_W\) such that the optimal seed minimises entanglement plus energy among pure Gaussian states. A relative-entropy inequality extends that minimisation to all pure states, and averaging over any ensemble with covariance \(V\) cancels the energy term, leaving \(G(V)\) as a lower bound.

Channels follow from states. Probing a Gaussian channel with two-mode squeezed vacua and pushing the squeezing to infinity gives its entanglement cost as \(\lim_{s\to\infty}G(V_\Phi(s))\), matched by teleportation through the Gaussian resource and finite exactly when the noise matrix \(Y\) is positive definite. For capacity, the minimum average output entropy at fixed average input equals the entanglement of formation of an output–environment state in a dilation, so the state theorem makes the Holevo quantity additive, and the energy-constrained capacity is a maximum over input covariances of the Gaussian output entropy minus the least output entropy of a compatible pure seed. Section V works out a correlated two-mode lossy channel with unequal losses.

The four-mode example is a family of states with every mode's marginal \(a\mathbb I_2\) and correlations of strength \(c\), position correlations pairing \(A_1B_1\) and \(A_2B_2\) and momentum correlations pairing \(A_1B_2\) and \(A_2B_1\). Every physical member is PPT, so none is distillable, and every two-mode marginal is separable; yet the states with \(a<c+1\) are entangled. Imposing the family's symmetries by geometric averaging reduces the seed to a one-parameter family \(\Gamma(u)\) of two equal squeezed pairs, and the cost becomes \(2g(\sinh^2(u_*/2))\) with an explicit \(u_*\). Replacing the Gaussian displacements by sixteen discrete ones gives a non-Gaussian state with the same covariance and the same cost which, unlike the Gaussian state, is distillable.

Transferable methods

  1. Geodesic convexity of entanglement over pure Gaussian states

    The set of pure covariances below a target is not convex in the ordinary sense, since the average of two pure covariances is mixed. The matrix geometric mean keeps purity and the constraint, and along it the entanglement of the pure state is convex. That convexity drives strong superadditivity, convexity of \(G\) in the target, and the reduction of symmetric problems to small families.

    The weighted geometric mean is \(X\#_tY=X^{1/2}(X^{-1/2}YX^{-1/2})^tX^{1/2}\). Pure covariances are positive symplectic matrices, and powers and symplectic congruences preserve that, so \(\Gamma_0\#_t\Gamma_1\) is pure; monotonicity of the mean keeps \(\Gamma_0\#_t\Gamma_1\le V\) when both endpoints satisfy it, and \(X\#_tY\le(1-t)X+tY\). With \(\Pi=\mathbb I_A\oplus(-\mathbb I_B)\) and \(R_\Gamma=\Gamma^{-1/2}\Pi\Gamma\Pi\Gamma^{-1/2}\), \[E(\Gamma)=\frac14\sum_iF\bigl(\log\lambda_i(R_\Gamma)\bigr),\qquad F(x)=g\bigl(\sinh^2(x/4)\bigr),\] where \(g(x)=(x+1)\log(x+1)-x\log x\), because a squeezed pair with parameter \(r\) contributes the eigenvalues \(e^{\pm4r}\), each twice.

    The relative spectrum along a geometric path is majorised by the interpolated endpoint spectra (Appendix A), and \(F\) is even and convex, so \(E(\Gamma_0\#_t\Gamma_1)\le(1-t)E(\Gamma_0)+tE(\Gamma_1)\). For strong superadditivity, let \(Q\) flip the signs of all quadratures in group \(A_1B_1\), a local operation. The midpoint \(\bar\Gamma=\Gamma\#(Q\Gamma Q)\) is pure, \(Q\)-invariant and hence block diagonal, \(\bar\Gamma=\gamma_1\oplus\gamma_2\), and \[\bar\Gamma\le\tfrac12(\Gamma+Q\Gamma Q)\le\tfrac12(V+QVQ)=V_1\oplus V_2,\] so \(E(\Gamma)\ge E(\gamma_1)+E(\gamma_2)\ge G(V_1)+G(V_2)\).

    For the four-mode family, averaging with four commuting symmetries (conjugation, local rotations, party exchange and a local symplectic map) leaves \(\Gamma(u)=Q(u)\oplus Q(u)^{-1}\), and the constraint \(\Gamma(u)\le V_{a,c}\) becomes \(\sqrt{4a^2-2c^2}\,\cosh(u-\beta)\le a^2-c^2+1\) with \(\beta=\operatorname{atanh}(c/\sqrt2a)\). Its smallest nonnegative solution is \[u_*=\Bigl[\operatorname{atanh}\frac{c}{\sqrt2\,a}-\operatorname{arcosh}\frac{a^2-c^2+1}{\sqrt{4a^2-2c^2}}\Bigr]_+,\qquad G(V_{a,c})=2g\bigl(\sinh^2(u_*/2)\bigr).\] Editorial check: an unrestricted numerical search over all pure four-mode covariances below \(V_{a,c}\) finds no lower value at \((a,c)=(1.7,0.8)\) and \((2.4,1.5)\).

    How to reuse it

    When an optimisation runs over pure states with a matrix-order constraint, look for a mean that preserves both purity and the constraint, and check whether the objective is convex along it. Symmetries can then be imposed by averaging, which reduces many multimode problems to a handful of parameters.

  2. A supporting quadratic and a relative-entropy inequality to reach non-Gaussian ensembles

    Convexity of \(G\) yields a linear lower bound on it, which translates into the statement that the optimal Gaussian seed minimises entanglement plus the mean of a quadratic Hamiltonian among pure Gaussian states. Extending that inequality to every pure state, Gaussian or not, and averaging over an ensemble with the right covariance, makes the energy terms cancel and leaves \(G(V)\) as a lower bound on any ensemble.

    Convexity gives \(B_V\) with \(G(U)\ge G(V)+\operatorname{Tr}[B_V(U-V)]\), and since enlarging the covariance can only lower \(G\), \(W=-4B_V\ge0\). At an optimal seed \(\Gamma_*\), \(\operatorname{Tr}[W(V-\Gamma_*)]=0\), and every pure Gaussian \(\Gamma\) satisfies \[E(\Gamma)+\tfrac14\operatorname{Tr}(W\Gamma)\ge E(\Gamma_*)+\tfrac14\operatorname{Tr}(W\Gamma_*).\] Here \(\tfrac14\operatorname{Tr}(W\Gamma)\) is the mean of \(H_W=R^{\mathsf T}WR/2\) in the centred state.

    Proposition 2 compares relative entropies: if \(\Phi\) is a Gaussian unitary followed by a partial trace, \(\sigma\) and \(\tau=\Phi(\sigma)\) are faithful Gaussian states and \(\lambda H_\sigma-\Phi^*H_\tau\) has a positive-definite coefficient matrix, then \(D(\Phi(\rho)\|\tau)\le\lambda D(\rho\|\sigma)\) for every finite-energy \(\rho\). Via the Gibbs identity \[F_\lambda(\rho)-F_\lambda(\sigma)=\lambda D(\rho\|\sigma)-D(\Phi(\rho)\|\tau),\qquad F_\lambda(\rho)=S(\Phi(\rho))-\lambda S(\rho)+\operatorname{Tr}(B\rho),\] \(\sigma\) minimises \(F_\lambda\) over all states, and a limiting argument turns this into \(S(\psi_A)+\langle H_W\rangle_\psi\ge E(\Gamma_*)+\langle H_W\rangle_{\Gamma_*}\) for every pure \(\psi\).

    Editorial step: for any ensemble \(\{p_i,\psi_i\}\) of a centred state with covariance \(V\), the average of \(\langle H_W\rangle_{\psi_i}\), displacements included, is \(\operatorname{Tr}(WV)/4\). Averaging the pure-state inequality and using complementarity gives \[\sum_ip_iS(\psi_{i,A})\ge G(V)+\tfrac14\operatorname{Tr}\bigl[W(\Gamma_*-V)\bigr]=G(V).\] Since \(\rho^{\otimes n}\) has covariance \(V^{\oplus n}\) and \(G(V^{\oplus n})=nG(V)\) by additivity, the bound survives regularisation, giving \(E_{\mathrm C}\ge G(V)\) for any state with covariance \(V\).

    How to reuse it

    To show that a family of states is optimal for a convex quantity, find a supporting functional that is an expectation value, prove the corresponding pointwise inequality for all pure states, and average. The energy terms cancel by complementarity, and the same mechanism yields lower bounds from measured second moments alone.

Source: Section II.3 (geometry of pure Gaussian preparations), Section II.4 (supporting quadratic), Section II.5 (Proposition 2), Section V.1 (four-mode bound entangled states). Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. This settles Gaussian optimality and additivity of entanglement of formation in full multimode generality, extending the two-mode result and the single-mode capacity result from the same group. Because the bound depends only on second moments, measured quadrature covariances certify a minimum preparation cost without assuming that the state is Gaussian. The four-mode example quantifies an irreversible cost: states that are PPT, so no entanglement can be distilled from them, and whose two-mode marginals are all separable, still need entanglement to prepare.

Limitations. In general the answer is a covariance optimisation, not a closed formula; the authors provide a numerical package with upper and lower bounds. The channel cost is defined in Wilde's sequential setting, and the capacity theorem assumes a quadratic energy constraint at or above a threshold energy. The proof is long, with technical lemmas on Gibbs regularity and Schmidt information variance in the appendices that were not read in full here. A concurrent preprint by He gives a general multimode capacity formula. The acknowledgements state that AI tools based on OpenAI and Anthropic models assisted with mathematical development, literature searches, numerical methods, analytic checks and manuscript preparation.

Description based on the full arXiv HTML of v1, read 2026-11-19: Sections I–IV in the main text (Theorem 1 and its proof outline through geodesic convexity, the supporting quadratic and Proposition 2; Corollary 3; Theorems 4 and 6 with their setups), Section V.1 in full, Sections V.2–VI in outline, and the acknowledgements. Appendices A–F were consulted only for the relative-spectrum comparison. Editorial checks: for the four-mode family, the symplectic spectrum \(\nu_\pm^2=a^2\pm c\sqrt{2a^2-c^2}\), invariance under partial transposition, purity of the seed \(\Gamma(u)\), the formula for \(u_*\) (against a direct scan of \(\Gamma(u)\le V\)), the entanglement \(2g(\sinh^2(u_*/2))\), the boundary parametrisation in \(\kappa\) and separability at \(a\ge c+1\) all check numerically at five points; an unrestricted minimisation over the 20-dimensional manifold of pure four-mode covariances below \(V\) reproduces the closed form at two points to \(5\times10^{-6}\) and never goes below it; and for 200 random pure four-mode states the relative-spectrum formula (16) matches the symplectic-eigenvalue entropy to \(10^{-14}\), with no violation of geodesic convexity.