Entanglement-Assisted Quantum Rate-Distortion

• The paper establishes a single-letter formula for the entanglement-assisted quantum rate-distortion function by minimizing quantum mutual information subject to distortion constraints.
• It leverages unlimited shared entanglement to simulate quantum channels via the reverse Shannon theorem, unifying lossless and lossy quantum compression frameworks.
• It extends the theory to one-shot, finite-blocklength, and distributed settings, addressing mixed-state ensembles and full resource trade-offs in practical scenarios.

The entanglement-assisted quantum rate-distortion function is the optimal asymptotic qubit rate for lossy compression of a memoryless quantum information source when sender and receiver share unlimited prior entanglement. In the canonical setting, the source is a density operator $\rho_A$ with purification $|\psi^\rho_{RA}\rangle$, the distortion constraint is imposed on the joint reference–output state, and the optimum is given by a single-letter minimization of a quantum mutual information over channels obeying the distortion threshold. Subsequent work extended this structure to one-shot coding, mixed-state ensemble sources with encoder side information, and distributed redistribution-type scenarios, while preserving the central role of mutual-information quantities in the entanglement-assisted regime (Datta et al., 2011, Khanian et al., 2022, Khanian et al., 2021).

1. Canonical formulation

A standard memoryless source model takes

$\rho_A \in \mathcal D(\mathcal H_A),$

with purification

$$|\psi_{RA}^{\rho}\rangle \in \mathcal H_R \otimes \mathcal H_A, \qquad \rho_A = \operatorname{Tr}_R \psi_{RA}^{\rho}.$$

For $n$ uses of the source, the state is $\rho^{\otimes n}$. In the entanglement-assisted setting, Alice and Bob share unlimited prior entanglement, Alice encodes the source together with her entanglement share, sends qubits to Bob, and Bob decodes using the received qubits and his entanglement share. One formulation employs Barnum’s entanglement-fidelity distortion,

$d(\rho,\mathcal{N}) = 1 - F_e(\rho,\mathcal{N}),$

where

$$F_{e}(\rho,\mathcal{N}) \equiv \langle \psi_{RA}^{\rho}| (\mathrm{id}_{R}\otimes\mathcal{N}^{A\rightarrow B})(\psi_{RA}^{\rho}) |\psi_{RA}^{\rho}\rangle.$$

A more general formulation uses a positive semidefinite distortion observable $\Delta_{RB}\ge 0$ and

$$d(\rho,\mathcal{N}) = \operatorname{Tr}\!\left[\Delta_{RB}\,(\operatorname{id}_R\otimes \mathcal{N}_{A\to B})(\psi^\rho_{RA})\right].$$

The entanglement-assisted rate-distortion function is the infimum achievable qubit rate in the asymptotic sense (Datta et al., 2011, Wilde et al., 2012).

The central single-letter theorem for noiseless quantum communication is

$|\psi^\rho_{RA}\rangle$0

with

$|\psi^\rho_{RA}\rangle$1

or equivalently

$|\psi^\rho_{RA}\rangle$2

For the corresponding noiseless classical-communication entanglement-assisted setting, the rate is

$|\psi^\rho_{RA}\rangle$3

The quantum version is exactly one-half the classical-communication entanglement-assisted rate, matching the teleportation intuition. The paper presenting this formula emphasizes that it is the closest quantum analogue of Shannon’s classical rate-distortion theorem, with $|\psi^\rho_{RA}\rangle$4 replaced by the quantum mutual information $|\psi^\rho_{RA}\rangle$5 (Datta et al., 2011).

2. Operational derivation, converse bounds, and comparison with unassisted coding

The operational content of the single-letter formula is that one chooses a CPTP map $|\psi^\rho_{RA}\rangle$6 that satisfies the distortion threshold and then simulates that map with free entanglement. In the classical-communication version, the key resource inequality is

$|\psi^\rho_{RA}\rangle$7

while in the quantum-communication version it is

$|\psi^\rho_{RA}\rangle$8

This is why the entanglement-assisted case is single-letter: unlimited shared entanglement lets the reverse Shannon theorem be used “as is.” The direct part follows by channel simulation, and the converse uses entropy and data-processing inequalities together with superadditivity and convexity arguments to reduce blocklength expressions to the same single-letter minimum (Datta et al., 2011).

The entanglement-assisted expression also gives a sharp comparison with the unassisted theory. A basic bound is

$|\psi^\rho_{RA}\rangle$9

so the assisted function lower-bounds the unassisted one. Two consequences are singled out. First, the unassisted quantum rate-distortion function is non-negative because the mutual information is nonnegative. Second, Barnum’s coherent-information conjecture fails in general, because

$\rho_A \in \mathcal D(\mathcal H_A),$0

whereas Barnum’s candidate can become negative. The mutual-information obstruction is therefore the correct single-letter lower bound in the entanglement-assisted setting, not the coherent information (Datta et al., 2011).

The same framework yields entanglement-assisted source-channel separation statements. For source transmission up to distortion $\rho_A \in \mathcal D(\mathcal H_A),$1,

$\rho_A \in \mathcal D(\mathcal H_A),$2

is necessary and sufficient, where

$\rho_A \in \mathcal D(\mathcal H_A),$3

In the exact-transmission case $\rho_A \in \mathcal D(\mathcal H_A),$4, this reduces to

$\rho_A \in \mathcal D(\mathcal H_A),$5

A related trade-off formulation, with qubit communication rate $\rho_A \in \mathcal D(\mathcal H_A),$6 and entanglement consumption $\rho_A \in \mathcal D(\mathcal H_A),$7, is

$\rho_A \in \mathcal D(\mathcal H_A),$8

and unrestricted $\rho_A \in \mathcal D(\mathcal H_A),$9 collapses the region to the single-letter entanglement-assisted expression (Datta et al., 2011, Wilde et al., 2012).

3. One-shot, finite-blocklength, and distributed versions

One-shot entanglement-assisted quantum rate-distortion coding replaces asymptotic average-rate statements by a minimum qubit compression size $$|\psi_{RA}^{\rho}\rangle \in \mathcal H_R \otimes \mathcal H_A, \qquad \rho_A = \operatorname{Tr}_R \psi_{RA}^{\rho}.$$0, defined relative to an excess-distortion criterion. For a distortion observable

$$|\psi_{RA}^{\rho}\rangle \in \mathcal H_R \otimes \mathcal H_A, \qquad \rho_A = \operatorname{Tr}_R \psi_{RA}^{\rho}.$$1

the excess-distortion projector is

$$|\psi_{RA}^{\rho}\rangle \in \mathcal H_R \otimes \mathcal H_A, \qquad \rho_A = \operatorname{Tr}_R \psi_{RA}^{\rho}.$$2

and a code is admissible when the probability of exceeding $$|\psi_{RA}^{\rho}\rangle \in \mathcal H_R \otimes \mathcal H_A, \qquad \rho_A = \operatorname{Tr}_R \psi_{RA}^{\rho}.$$3 is at most $$|\psi_{RA}^{\rho}\rangle \in \mathcal H_R \otimes \mathcal H_A, \qquad \rho_A = \operatorname{Tr}_R \psi_{RA}^{\rho}.$$4. In this setting, the minimum qubit compression size is characterized, up to explicit logarithmic correction terms, by the smooth max-information

$$|\psi_{RA}^{\rho}\rangle \in \mathcal H_R \otimes \mathcal H_A, \qquad \rho_A = \operatorname{Tr}_R \psi_{RA}^{\rho}.$$5

with

$$|\psi_{RA}^{\rho}\rangle \in \mathcal H_R \otimes \mathcal H_A, \qquad \rho_A = \operatorname{Tr}_R \psi_{RA}^{\rho}.$$6

The one-shot achievability uses one-shot quantum state splitting or channel simulation, and the asymptotic equipartition properties show that

$$|\psi_{RA}^{\rho}\rangle \in \mathcal H_R \otimes \mathcal H_A, \qquad \rho_A = \operatorname{Tr}_R \psi_{RA}^{\rho}.$$7

so the smooth max-information is the one-shot analogue of the mutual-information rate-distortion formula (Datta et al., 2013).

For the isotropic qubit source

$$|\psi_{RA}^{\rho}\rangle \in \mathcal H_R \otimes \mathcal H_A, \qquad \rho_A = \operatorname{Tr}_R \psi_{RA}^{\rho}.$$8

the entanglement-assisted asymptotic rate-distortion function under entanglement-fidelity distortion is

$$|\psi_{RA}^{\rho}\rangle \in \mathcal H_R \otimes \mathcal H_A, \qquad \rho_A = \operatorname{Tr}_R \psi_{RA}^{\rho}.$$9

The same source is used for a tight finite-blocklength characterization, obtained through a converse adapted from Kostina–Verdú and a teleportation-based achievability argument (Wilde et al., 2012, Datta et al., 2013).

A broader distributed version is formulated as a rate-distortion form of quantum state redistribution. The source is a pure tripartite state $n$0, with Alice holding $n$1, Bob holding $n$2, and $n$3 the reference. Distortion is a general convex continuous functional on $n$4, extended additively across copies. The single-letter optimization is

$n$5

over CPTP maps

$n$6

that satisfy the distortion bound after decoding. The exact entanglement-assisted rate-distortion function is

$n$7

where

$n$8

This formula is exact for all $n$9, but one must take the right-continuous extension at the minimal feasible distortion $\rho^{\otimes n}$0. A peculiarity is that the auxiliary register $\rho^{\otimes n}$1 is not a priori bounded in dimension, so the formula is single-letter but not automatically computable (Khanian et al., 2021).

4. Mixed-state ensemble sources, encoder side information, and full resource trade-offs

A later generalization treats compression of an i.i.d. ensemble of mixed quantum states with encoder side information. The source is

$\rho^{\otimes n}$2

equivalently

$\rho^{\otimes n}$3

Here $\rho^{\otimes n}$4 is a classical reference system identifying the source label, $\rho^{\otimes n}$5 is side information available to the encoder, and a purification is taken as

$\rho^{\otimes n}$6

The distortion function is local: $\rho^{\otimes n}$7 subject to continuity, convexity, and

$\rho^{\otimes n}$8

A canonical example is

$\rho^{\otimes n}$9

The theory considers both worst-case and average local distortions,

$d(\rho,\mathcal{N}) = 1 - F_e(\rho,\mathcal{N}),$0

For both criteria, the entanglement-assisted rate-distortion function is

$d(\rho,\mathcal{N}) = 1 - F_e(\rho,\mathcal{N}),$1

where

$d(\rho,\mathcal{N}) = 1 - F_e(\rho,\mathcal{N}),$2

Thus the assisted theory remains single-letter, whereas the unassisted average-local-error function is

$d(\rho,\mathcal{N}) = 1 - F_e(\rho,\mathcal{N}),$3

so the unassisted setting is governed by a regularized entanglement of purification rather than a single-letter mutual information (Khanian et al., 2022).

The same work gives the full qubit-entanglement rate region. Rate-distortion coding with distortion $d(\rho,\mathcal{N}) = 1 - F_e(\rho,\mathcal{N}),$4 is achievable iff for some $d(\rho,\mathcal{N}) = 1 - F_e(\rho,\mathcal{N}),$5,

$d(\rho,\mathcal{N}) = 1 - F_e(\rho,\mathcal{N}),$6

and

$d(\rho,\mathcal{N}) = 1 - F_e(\rho,\mathcal{N}),$7

for some channel $d(\rho,\mathcal{N}) = 1 - F_e(\rho,\mathcal{N}),$8 satisfying the distortion constraint and some channel $d(\rho,\mathcal{N}) = 1 - F_e(\rho,\mathcal{N}),$9. The framework covers blind compression by taking $$F_{e}(\rho,\mathcal{N}) \equiv \langle \psi_{RA}^{\rho}| (\mathrm{id}_{R}\otimes\mathcal{N}^{A\rightarrow B})(\psi_{RA}^{\rho}) |\psi_{RA}^{\rho}\rangle.$$0, visible compression by taking $$F_{e}(\rho,\mathcal{N}) \equiv \langle \psi_{RA}^{\rho}| (\mathrm{id}_{R}\otimes\mathcal{N}^{A\rightarrow B})(\psi_{RA}^{\rho}) |\psi_{RA}^{\rho}\rangle.$$1, and intermediate side-information models. It also uses the Koashi-Imoto decomposition

$$F_{e}(\rho,\mathcal{N}) \equiv \langle \psi_{RA}^{\rho}| (\mathrm{id}_{R}\otimes\mathcal{N}^{A\rightarrow B})(\psi_{RA}^{\rho}) |\psi_{RA}^{\rho}\rangle.$$2

to show that redundant parts do not affect the rate-distortion function in either the assisted or unassisted setting (Khanian et al., 2022).

5. Zero-distortion limits and neighboring compression theories

Entanglement-assisted rate-distortion theory meets lossless quantum compression at zero distortion. In the distributed framework, choosing a pure source $$F_{e}(\rho,\mathcal{N}) \equiv \langle \psi_{RA}^{\rho}| (\mathrm{id}_{R}\otimes\mathcal{N}^{A\rightarrow B})(\psi_{RA}^{\rho}) |\psi_{RA}^{\rho}\rangle.$$3 with no decoder side information and distortion

$$F_{e}(\rho,\mathcal{N}) \equiv \langle \psi_{RA}^{\rho}| (\mathrm{id}_{R}\otimes\mathcal{N}^{A\rightarrow B})(\psi_{RA}^{\rho}) |\psi_{RA}^{\rho}\rangle.$$4

or equivalently a distortion observable $$F_{e}(\rho,\mathcal{N}) \equiv \langle \psi_{RA}^{\rho}| (\mathrm{id}_{R}\otimes\mathcal{N}^{A\rightarrow B})(\psi_{RA}^{\rho}) |\psi_{RA}^{\rho}\rangle.$$5, one obtains

$$F_{e}(\rho,\mathcal{N}) \equiv \langle \psi_{RA}^{\rho}| (\mathrm{id}_{R}\otimes\mathcal{N}^{A\rightarrow B})(\psi_{RA}^{\rho}) |\psi_{RA}^{\rho}\rangle.$$6

recovering the Schumacher rate. Under the same formalism, the zero-distortion limit of quantum state redistribution is

$$F_{e}(\rho,\mathcal{N}) \equiv \langle \psi_{RA}^{\rho}| (\mathrm{id}_{R}\otimes\mathcal{N}^{A\rightarrow B})(\psi_{RA}^{\rho}) |\psi_{RA}^{\rho}\rangle.$$7

and state merging is the special case with trivial Alice-side information, yielding

$$F_{e}(\rho,\mathcal{N}) \equiv \langle \psi_{RA}^{\rho}| (\mathrm{id}_{R}\otimes\mathcal{N}^{A\rightarrow B})(\psi_{RA}^{\rho}) |\psi_{RA}^{\rho}\rangle.$$8

This places Schumacher compression, state merging, and quantum state redistribution inside a common entanglement-assisted rate-distortion envelope (Khanian et al., 2021).

Pure-state ensemble compression with encoder side information provides a closely related zero-distortion boundary case. The source ensemble is

$$F_{e}(\rho,\mathcal{N}) \equiv \langle \psi_{RA}^{\rho}| (\mathrm{id}_{R}\otimes\mathcal{N}^{A\rightarrow B})(\psi_{RA}^{\rho}) |\psi_{RA}^{\rho}\rangle.$$9

equivalently

$\Delta_{RB}\ge 0$0

Although this theory is not framed as a distortion-based rate-distortion theorem, it is very much in the spirit of quantum rate tradeoff problems. With unlimited entanglement, the optimal quantum communication rate is

$\Delta_{RB}\ge 0$1

the entanglement consumption is

$\Delta_{RB}\ge 0$2

and the full asymptotic resource tradeoff is

$\Delta_{RB}\ge 0$3

Relative to Schumacher compression,

$\Delta_{RB}\ge 0$4

so the improvement is exactly half the mutual information between the source system and the accessible side information. The converse formalizes this through the quantity

$\Delta_{RB}\ge 0$5

which measures how much classical information can be extracted by an isometry while keeping the overall state close to the original. In the visible case $\Delta_{RB}\ge 0$6,

$\Delta_{RB}\ge 0$7

matching remote state preparation, while for an irreducible blind source entanglement does not help and $\Delta_{RB}\ge 0$8 remains optimal (Khanian et al., 2019).

A common misconception is that entanglement assistance necessarily lowers every zero-distortion compression rate. The blind irreducible-source result just cited shows otherwise. Another is that lossless compression and rate distortion are disjoint subjects. The zero-distortion limits above indicate that, in the entanglement-assisted regime, they are contiguous parts of a single resource-theoretic landscape (Khanian et al., 2019, Khanian et al., 2021).

6. Structure of optimizers and numerical computation

The entanglement-assisted quantum rate-distortion problem admits a convex reformulation over joint output states. For input $\Delta_{RB}\ge 0$9, purification $$d(\rho,\mathcal{N}) = \operatorname{Tr}\!\left[\Delta_{RB}\,(\operatorname{id}_R\otimes \mathcal{N}_{A\to B})(\psi^\rho_{RA})\right].$$0, and distortion operator $$d(\rho,\mathcal{N}) = \operatorname{Tr}\!\left[\Delta_{RB}\,(\operatorname{id}_R\otimes \mathcal{N}_{A\to B})(\psi^\rho_{RA})\right].$$1,

$$d(\rho,\mathcal{N}) = \operatorname{Tr}\!\left[\Delta_{RB}\,(\operatorname{id}_R\otimes \mathcal{N}_{A\to B})(\psi^\rho_{RA})\right].$$2

where

$$d(\rho,\mathcal{N}) = \operatorname{Tr}\!\left[\Delta_{RB}\,(\operatorname{id}_R\otimes \mathcal{N}_{A\to B})(\psi^\rho_{RA})\right].$$3

For entanglement-fidelity distortion,

$$d(\rho,\mathcal{N}) = \operatorname{Tr}\!\left[\Delta_{RB}\,(\operatorname{id}_R\otimes \mathcal{N}_{A\to B})(\psi^\rho_{RA})\right].$$4

the problem is invariant under a symmetry group fixing $$d(\rho,\mathcal{N}) = \operatorname{Tr}\!\left[\Delta_{RB}\,(\operatorname{id}_R\otimes \mathcal{N}_{A\to B})(\psi^\rho_{RA})\right].$$5. If $$d(\rho,\mathcal{N}) = \operatorname{Tr}\!\left[\Delta_{RB}\,(\operatorname{id}_R\otimes \mathcal{N}_{A\to B})(\psi^\rho_{RA})\right].$$6 is diagonal, an optimizer exists in a fixed-point subspace of real dimension

$$d(\rho,\mathcal{N}) = \operatorname{Tr}\!\left[\Delta_{RB}\,(\operatorname{id}_R\otimes \mathcal{N}_{A\to B})(\psi^\rho_{RA})\right].$$7

If $$d(\rho,\mathcal{N}) = \operatorname{Tr}\!\left[\Delta_{RB}\,(\operatorname{id}_R\otimes \mathcal{N}_{A\to B})(\psi^\rho_{RA})\right].$$8, the fixed-point subspace collapses to

$$d(\rho,\mathcal{N}) = \operatorname{Tr}\!\left[\Delta_{RB}\,(\operatorname{id}_R\otimes \mathcal{N}_{A\to B})(\psi^\rho_{RA})\right].$$9

and the rate-distortion function becomes

$|\psi^\rho_{RA}\rangle$00

An achieving channel is the depolarizing channel

$|\psi^\rho_{RA}\rangle$01

The same work develops an inexact mirror descent method, relates it to Blahut-Arimoto and expectation-maximization, and reports experiments up to 9 qubits, with symmetry reduction reducing a $|\psi^\rho_{RA}\rangle$02-qubit problem from an effectively $|\psi^\rho_{RA}\rangle$03 billion-dimensional optimization variable to dimension $|\psi^\rho_{RA}\rangle$04 (He et al., 2023).

A later algorithmic development writes the entanglement-assisted function as

$|\psi^\rho_{RA}\rangle$05

subject to

$|\psi^\rho_{RA}\rangle$06

The paper emphasizes that the problem is convex, but it has a difficult combination of positive semidefiniteness constraints, linear marginal constraints, and a nonlinear distortion constraint. Its alternating minimization method updates the distortion multiplier $|\psi^\rho_{RA}\rangle$07 at each iteration so that the distortion constraint is enforced, rather than treating $|\psi^\rho_{RA}\rangle$08 as a fixed penalty parameter. The resulting updates are closed form for $|\psi^\rho_{RA}\rangle$09, $|\psi^\rho_{RA}\rangle$10, and $|\psi^\rho_{RA}\rangle$11, and require only a one-dimensional root-finding step for $|\psi^\rho_{RA}\rangle$12. For entanglement fidelity distortion, symmetry reduction can reduce per-iteration cost from

$|\psi^\rho_{RA}\rangle$13

The reported optimality residual reaches about $|\psi^\rho_{RA}\rangle$14, and representative speedups over mirror descent include $|\psi^\rho_{RA}\rangle$15 and $|\psi^\rho_{RA}\rangle$16 for $|\psi^\rho_{RA}\rangle$17 uniform input, $|\psi^\rho_{RA}\rangle$18 and $|\psi^\rho_{RA}\rangle$19 for $|\psi^\rho_{RA}\rangle$20 uniform input, and up to about $|\psi^\rho_{RA}\rangle$21 speedup for $|\psi^\rho_{RA}\rangle$22 random input (Chen et al., 26 Jul 2025).

The computational literature clarifies an important conceptual distinction. The entanglement-assisted theory is often described as “single-letter,” but single-letter does not always mean readily computable. In the basic memoryless source model the objective is a convex mutual-information minimization, and recent symmetry-reduction and first-order or alternating-minimization methods make moderate dimensions accessible. In the distributed redistribution setting, by contrast, the exact single-letter formula may still require optimization over an unbounded auxiliary register $|\psi^\rho_{RA}\rangle$23, so the obstruction is not regularization but auxiliary-system cardinality (He et al., 2023, Chen et al., 26 Jul 2025, Khanian et al., 2021).