Distributed Quantum State Purification

• Distributed quantum state purification is a set of protocols that improve the quality of shared noisy quantum states using local operations and classical communication.
• It encompasses techniques such as entanglement purification and local purity distillation, which are crucial for quantum repeaters, communication, and distributed computing.
• Recent advances integrate one-shot and asymptotic methods, employing smooth entropy tools and optimized recurrence protocols to finely balance ancilla cost and classical communication.

Distributed quantum state purification denotes a family of protocols in which remote parties holding noisy shared quantum states use local operations and classical communication, and in some settings ancillary pure states, dephasing channels, or additional flying qubits, to increase purity or fidelity. In one branch, entanglement purification distills higher-quality Bell or GHZ states from noisy distributed entanglement and plays a key role in quantum repeaters, quantum communication, and distributed quantum computing (2101.07441). In another, distributed purity distillation extracts local pure qubits such as $|0\rangle^{A'}|0\rangle^{B'}$ from a known bipartite mixed state $\rho^{AB}$, with net yield measured by $\log|A'|+\log|B'|$ minus any borrowed pure ancilla (Chakraborty et al., 2024). Taken together, these lines of work treat purification as a distributed resource-conversion problem constrained by LOCC, communication structure, and hardware noise.

1. Formal task definitions and resource models

A common formalization models a distributed purification protocol as a completely positive, trace-nonincreasing map $$\mathcal{E}_{A^nB^n\to A'B'}\in \mathrm{CPTN}\cap\mathrm{LOCC}$$, acting on $n$ noisy copies of a bipartite state and producing one purified output copy on $A'B'$. If the input copies are generated by a depolarizing channel $\mathcal{N}^\gamma(\rho)=(1-\gamma)\rho+\gamma I/d$, the normalized output $\sigma_\psi$ is evaluated by the fidelity condition $F(\sigma_\psi,\psi)\ge F(\mathcal{N}(\psi),\psi)$, where $F(\rho,\psi)=\langle\psi|\rho|\psi\rangle$ for pure $\rho^{AB}$0 (Zhao et al., 10 Sep 2025). In network settings, the fidelity of an actual density matrix $\rho^{AB}$1 relative to an ideal target $\rho^{AB}$2 is written $\rho^{AB}$3, with $\rho^{AB}$4 and $\rho^{AB}$5 exactly for $\rho^{AB}$6 (Fan et al., 18 Mar 2025).

Resource-theoretic distributed purity distillation uses a different target. The task is to extract a fixed pure product state $\rho^{AB}$7 from a known distributed mixed state $\rho^{AB}$8, allowing local unitaries, appending and discarding maximally mixed states, appending pure-state ancilla $\rho^{AB}$9 that must be “paid back,” and a one-way classical fully dephasing channel from Alice to Bob. The figure of merit is the distilled purity

$\log|A'|+\log|B'|$0

minus any borrowed pure ancilla (Chakraborty et al., 2024). In the three-party version, Alice and Bob communicate with Charlie through a one-way multiple-access dephasing channel of rates $\log|A'|+\log|B'|$1 and $\log|A'|+\log|B'|$2, while the objective is to distill a total of $\log|A'|+\log|B'|$3 pure qubits from $\log|A'|+\log|B'|$4 and return catalytic ancillas unmodified (Atif et al., 2022).

This suggests a useful division of the subject into two operational regimes. The first treats entanglement itself as the object being purified; the second treats local purity as the distilled resource. The mathematical tools differ accordingly: Bell-state recurrences, swapping fidelities, and success probabilities dominate entanglement-purification analyses, whereas smooth min-, max-, and hypothesis-testing entropies dominate one-shot purity distillation.

2. Bipartite entanglement purification protocols and laboratory realizations

The standard bipartite setting begins with two noisy Bell pairs or, more generally, Werner states. In the network-routing literature, a generic $\log|A'|+\log|B'|$5 recurrence step acting on two Werner pairs $\log|A'|+\log|B'|$6 updates the fidelity by

$\log|A'|+\log|B'|$7

with success probability

$\log|A'|+\log|B'|$8

and more selective $\log|A'|+\log|B'|$9 circuits are optimized numerically up to $$\mathcal{E}_{A^nB^n\to A'B'}\in \mathrm{CPTN}\cap\mathrm{LOCC}$$0 to maximize distillable entanglement under buffer-size constraints (Victora et al., 2020). Closely related recurrences also appear in purification-aware routing formulations based on BBPSSW and DEJMPS primitives (Peñas et al., 22 May 2026).

A distinct physical realization uses a single hyperentangled pair rather than two separate entangled pairs. In the long-distance protocol of Hu et al., the source prepares

$$\mathcal{E}_{A^nB^n\to A'B'}\in \mathrm{CPTN}\cap\mathrm{LOCC}$$1

with polarization and spatial-mode degrees of freedom, and after transmission through an 11 km multicore fiber plus controllable liquid-crystal noise the input state is modeled as $$\mathcal{E}_{A^nB^n\to A'B'}\in \mathrm{CPTN}\cap\mathrm{LOCC}$$2, where each degree of freedom is depolarized into a Werner-type mixture. The purification operation is a deterministic CNOT from spatial mode to polarization on each photon, followed by post-selection on “same” polarization outcomes. The output polarization fidelity is

$$\mathcal{E}_{A^nB^n\to A'B'}\in \mathrm{CPTN}\cap\mathrm{LOCC}$$3

which in the symmetric case reduces to

$$\mathcal{E}_{A^nB^n\to A'B'}\in \mathrm{CPTN}\cap\mathrm{LOCC}$$4

and satisfies $$\mathcal{E}_{A^nB^n\to A'B'}\in \mathrm{CPTN}\cap\mathrm{LOCC}$$5 whenever $$\mathcal{E}_{A^nB^n\to A'B'}\in \mathrm{CPTN}\cap\mathrm{LOCC}$$6. Experimentally, one pair of polarization spatial-mode hyperentanglement was distributed over 11 km multicore fiber, the fidelity of polarization entanglement rose from $$\mathcal{E}_{A^nB^n\to A'B'}\in \mathrm{CPTN}\cap\mathrm{LOCC}$$7 to $$\mathcal{E}_{A^nB^n\to A'B'}\in \mathrm{CPTN}\cap\mathrm{LOCC}$$8, the CHSH value rose from $$\mathcal{E}_{A^nB^n\to A'B'}\in \mathrm{CPTN}\cap\mathrm{LOCC}$$9 to $n$0, and the effective key rate in entanglement-based QKD increased from $n$1 to $n$2 (2101.07441).

Superconducting-network experiments realize a different noise regime, dominated by amplitude damping rather than depolarizing noise. In a 1-meter superconducting communication cable, two impure Bell pairs are generated, one is swapped into memory qubits, local CNOTs are applied, and the protocol post-selects on $n$3 outcomes on the sacrificial pair. The best final fidelity reported is $n$4, and the largest fractional increase in fidelity reaches approximately $n$5 at $n$6 ns. The same work combines purification with dynamical decoupling and Rabi driving, extending the effective dephasing time from $n$7 to $n$8 (Yan et al., 2022).

These experiments illustrate that distributed entanglement purification is not tied to a single microscopic error model. Hyperentanglement-assisted optical protocols target bit-flip and phase-flip mixtures in multiple degrees of freedom, while superconducting implementations explicitly target transmission loss and $n$9-type damping. The operational motif—consume noisy distributed entanglement plus local gates and classical communication to produce a higher-fidelity nonlocal resource—remains the same.

3. One-shot and asymptotic distributed purity distillation

In one-shot distributed purity distillation, the principal quantities are smooth entropies rather than Bell-basis fidelities. For any subnormalized state, the formalism uses the conditional smooth min-entropy $A'B'$0, the conditional smooth max-entropy $A'B'$1, and the hypothesis-testing conditional entropy $A'B'$2. Operationally, the local pure-qubit yield from $A'B'$3 is approximately $A'B'$4, while the distributed gain from correlations is governed by $A'B'$5, $A'B'$6, and $A'B'$7 (Chakraborty et al., 2024).

The main achievability result gives a one-way $A'B'$8-distillation protocol, denoted KD or FewQubits, that achieves

$A'B'$9

Here $\mathcal{N}^\gamma(\rho)=(1-\gamma)\rho+\gamma I/d$0 is the output of a POVM on Alice’s side constrained by the classical communication budget, and the ancilla cost is $\mathcal{N}^\gamma(\rho)=(1-\gamma)\rho+\gamma I/d$1 for KD or $\mathcal{N}^\gamma(\rho)=(1-\gamma)\rho+\gamma I/d$2 for FewQubits. The converse nearly matches this bound. In the asymptotic i.i.d. limit, the smooth entropies concentrate and the one-shot bounds recover the Devetak–Krovi–Devetak rate; FewQubits matches the best known one-shot and i.i.d. rates while borrowing only $\mathcal{N}^\gamma(\rho)=(1-\gamma)\rho+\gamma I/d$3 ancilla qubits in the one-shot regime and zero in the i.i.d. limit (Chakraborty et al., 2024).

The tripartite extension replaces a single one-way dephasing link by a one-way multiple-access dephasing channel from Alice and Bob to Charlie. The achievable region is expressed in terms of $\mathcal{N}^\gamma(\rho)=(1-\gamma)\rho+\gamma I/d$4, distributed measurement compression, and mutual-information quantities involving classical registers $\mathcal{N}^\gamma(\rho)=(1-\gamma)\rho+\gamma I/d$5 induced by product POVMs and algebraic binning. The protocol combines compressed sub-POVMs, coherent local unitaries implementing the compressed measurements, purity extraction on typical subspaces, and partitioned coset codes to characterize an inner bound on $\mathcal{N}^\gamma(\rho)=(1-\gamma)\rho+\gamma I/d$6 (Atif et al., 2022).

A plausible implication is that “distributed quantum state purification” is not restricted to entanglement distillation. In the purity-distillation formulation, purification converts shared mixedness and bounded classical communication into local pure qubits, and the central resource trade-off becomes ancilla cost versus classical communication rather than success probability versus Bell-pair consumption.

4. Distributed-computing architectures, multipartite purification, and carrier-based variants

In distributed quantum computing with small nodes, purification can target an operation rather than a static state. In the three-qubit-node architecture, each node contains a broker qubit, an intermediate qubit, and a client qubit. The protocol nests two layers: bit-flip entanglement pumping of broker pairs, followed by repeated parity projections on the client qubits. Because parity projection is idempotent, repeated measurements can suppress parity-misreporting errors. Mapping the resulting noise to the topologically protected cluster state gives a tolerance of up to $\mathcal{N}^\gamma(\rho)=(1-\gamma)\rho+\gamma I/d$7 total phase error and up to $\mathcal{N}^\gamma(\rho)=(1-\gamma)\rho+\gamma I/d$8 qubit loss. Numerical optimization shows that if $\mathcal{N}^\gamma(\rho)=(1-\gamma)\rho+\gamma I/d$9, network entangling-operation infidelity $\sigma_\psi$0 is tolerable, and for $\sigma_\psi$1, $\sigma_\psi$2 may approach approximately $\sigma_\psi$3 (Li et al., 2012).

Multipartite purification has also been formulated as a one-way, code-based distillation problem for GHZ states. Using an $\sigma_\psi$4 CSS QLDPC purification code and a normalized min-sum decoder with $\sigma_\psi$5, the GHZ protocol measures stabilizers on Alice’s block, broadcasts syndrome information, and lets Bob and Charlie decode correlated channel errors. For the lifted-product family “LP118,” the asymptotic rate is approximately $\sigma_\psi$6, the input threshold under i.i.d. single-qubit depolarizing noise is $\sigma_\psi$7, and the corresponding GHZ-fidelity threshold is

$\sigma_\psi$8

The same paper extends the construction to $\sigma_\psi$9-party GHZ states using a GHZ-map that transfers operators on one subsystem to structured operators on the remaining $F(\sigma_\psi,\psi)\ge F(\mathcal{N}(\psi),\psi)$0 parties (Rengaswamy et al., 2022).

A different reduction in hardware overhead is achieved by carrier-assisted entanglement purification. CAEPP uses exactly one stored noisy Bell-diagonal pair and one or more flying qubits (“carriers”) per purification round. In the single-carrier version, Alice applies $F(\sigma_\psi,\psi)\ge F(\mathcal{N}(\psi),\psi)$1, sends the carrier through a channel $F(\sigma_\psi,\psi)\ge F(\mathcal{N}(\psi),\psi)$2, Bob applies $F(\sigma_\psi,\psi)\ge F(\mathcal{N}(\psi),\psi)$3, measures the carrier in the $F(\sigma_\psi,\psi)\ge F(\mathcal{N}(\psi),\psi)$4-basis, and the parties keep the updated pair only if the measurement outcome is $F(\sigma_\psi,\psi)\ge F(\mathcal{N}(\psi),\psi)$5. For noiseless carrier transmission, the update sets $F(\sigma_\psi,\psi)\ge F(\mathcal{N}(\psi),\psi)$6, with

$F(\sigma_\psi,\psi)\ge F(\mathcal{N}(\psi),\psi)$7

so that a second successful round yields $F(\sigma_\psi,\psi)\ge F(\mathcal{N}(\psi),\psi)$8. For noisy carrier channels, purification requires that the channel not be entanglement-breaking, equivalently $F(\sigma_\psi,\psi)\ge F(\mathcal{N}(\psi),\psi)$9; in the multi-carrier version, sufficiently many parallel carriers drive the fixed point to $F(\rho,\psi)=\langle\psi|\rho|\psi\rangle$0 for any fixed $F(\rho,\psi)=\langle\psi|\rho|\psi\rangle$1 (Kim et al., 9 Sep 2025).

Virtual entanglement purification moves the projection step from the state level to the expectation-value level. The protocol samples stabilizer pairs, uses noisy Bell ancillae and local controlled operations, and estimates

$F(\rho,\psi)=\langle\psi|\rho|\psi\rangle$2

Replacing ideal Bell ancillae by Werner or other Bell-diagonal states rescales stabilizer correlations by $F(\rho,\psi)=\langle\psi|\rho|\psi\rangle$3 but does not bias the ratio. To first order in $F(\rho,\psi)=\langle\psi|\rho|\psi\rangle$4 and $F(\rho,\psi)=\langle\psi|\rho|\psi\rangle$5, the infidelity obeys

$F(\rho,\psi)=\langle\psi|\rho|\psi\rangle$6

and numerically $F(\rho,\psi)=\langle\psi|\rho|\psi\rangle$7 is achieved for $F(\rho,\psi)=\langle\psi|\rho|\psi\rangle$8 and $F(\rho,\psi)=\langle\psi|\rho|\psi\rangle$9. For $\rho^{AB}$00, the overhead per pair is approximately $\rho^{AB}$01, whereas optimal LOCC-only circuit knitting approaches $\rho^{AB}$02 for large $\rho^{AB}$03 (Yamamoto et al., 2024).

5. Quantum-network routing, scheduling, and distributed purification planning

When purification is embedded in a repeater network, the central question becomes not only whether purification improves fidelity, but where and when it should be performed. A dynamic-programming formulation models the network as an undirected graph whose links generate Bell pairs as Poisson processes with mean rate $\rho^{AB}$04 and fidelity $\rho^{AB}$05, and whose nodes perform Bell-state measurements, local purification gates, and classical messaging. The DP state $\rho^{AB}$06 is the minimum expected time to generate an entangled pair of fidelity at least $\rho^{AB}$07 between nodes $\rho^{AB}$08 and $\rho^{AB}$09, using link creation, swapping, and iterated purification recurrences. With $\rho^{AB}$10 and a fidelity grid of size $\rho^{AB}$11, the complexity is $\rho^{AB}$12, and in practice $\rho^{AB}$13 and $\rho^{AB}$14 give runtimes of order a few seconds–minutes. In NetSquid simulations on random Waxman graphs of 30–70 nodes, the single-tree EP-DP solution achieves approximately $\rho^{AB}$15 higher generation rate than prior end-to-end routing schemes with simple link-only purification, and the LP-based EP-LP solution adds $\rho^{AB}$16–$\rho^{AB}$17 throughput when multiple source–destination pairs are active (Fan et al., 18 Mar 2025).

Earlier routing work formulated the objective as end-to-end distillable entanglement

$\rho^{AB}$18

where $\rho^{AB}$19 is the one-way hashing bound per pair and $\rho^{AB}$20 is the bottleneck entanglement generation rate along the chain. Under imperfect channel fidelities, limited memory lifetime $\rho^{AB}$21, and a library of optimized $\rho^{AB}$22 purification circuits, Dijkstra routing with link costs $\rho^{AB}$23 for $\rho^{AB}$24 or $\rho^{AB}$25 achieves within $\rho^{AB}$26 of the optimum found by brute-force path search, whereas the hop-only cost $\rho^{AB}$27 performs $\rho^{AB}$28–$\rho^{AB}$29 worse (Victora et al., 2020).

A more explicitly distributed design appears in Q-GUARD. Time is slotted, nodes exchange realized link outcomes only within their $\rho^{AB}$30-hop neighborhoods, and the protocol builds per-link purification cost tables after link generation outcomes are known. For a request threshold $\rho^{AB}$31, Q-GUARD uses the Werner equal-split rule

$\rho^{AB}$32

with $\rho^{AB}$33, and scores candidate recovery segments by the expected-goodput metric

$\rho^{AB}$34

On synthetic 100-node topologies, Q-GUARD raises the qualified success rate from under $\rho^{AB}$35 to over $\rho^{AB}$36 on 4-hop paths and nearly doubles the qualified service radius in Euclidean distance relative to throughput-only and naive-purification baselines; Q-GUARD-WS adds further throughput gains under high hardware heterogeneity (Gatti et al., 30 Apr 2026).

Purification-strategy optimization on a single repeater chain leads to closely related conclusions. In a DP model with decoherence $\rho^{AB}$37, entanglement swapping

$\rho^{AB}$38

and a finite budget of purification rounds per link, numerical experiments identify three recurrent regimes: early purification is most effective; when $\rho^{AB}$39, purification gains are erased by storage loss and the optimal strategy becomes “no purification”; and near $\rho^{AB}$40, the mean purification-round curve “freezes,” indicating a shift from purification-enabled to purification-useless behavior (Peñas et al., 22 May 2026).

6. Fundamental limits, asymptotic optimality, and structured exceptions

A central limitation result is that no nontrivial $\rho^{AB}$41 LOCC purification protocol exists for several broad ensembles. Under depolarizing or local depolarizing noise, there is no $\rho^{AB}$42 LOCC map that strictly improves the fidelity of all pure two-qubit states, all four Bell states, or all maximally entangled states while never worsening fidelity on the same set. The proof strategy relaxes LOCC to PPT channels, converts the optimization to an SDP over Choi matrices, and uses dual feasible solutions to show that the maximum universal fidelity gain is zero. In contrast, single-state purification is achievable: after local Schmidt alignment to $\rho^{AB}$43, each party applies an $\rho^{AB}$44 rotation and local CNOT, measures the second qubits, and post-selects on outcome $\rho^{AB}$45; the analytical expression for the fidelity improvement is nonnegative for all $\rho^{AB}$46. For arbitrary finite state sets, a variational search over local unitaries and post-selection can nearly saturate a PPT upper bound (Zhao et al., 10 Sep 2025).

These no-go theorems do not preclude asymptotically optimal purification for structured noise models. For the Pauli dephasing channel, an iterative two-way protocol uses local CNOT gates, Hadamard-basis measurements, and recursive processing of both successful and failed branches. If $\rho^{AB}$47 noisy Bell pairs of fidelity $\rho^{AB}$48 are processed at a stage, the probability that Alice’s and Bob’s outcomes agree is

$\rho^{AB}$49

and in the special case $\rho^{AB}$50 the fidelity update reduces to

$\rho^{AB}$51

The protocol yields perfect Bell pairs in the limit of many recursive iterations, and the reverse coherent information approaches the two-way capacity $\rho^{AB}$52 of the dephasing channel. The per-stage yield is exactly $\rho^{AB}$53, and because both success and failure branches are recycled, no distillable entanglement is thrown away (Erkilic et al., 2024).

This contrast is conceptually important. Universal $\rho^{AB}$54 LOCC purification over large ensembles fails even probabilistically, but structured settings with known target families, known noise models, more rounds, or additional encoded resources can reach perfect fidelity or even channel capacity. The modern literature therefore places the “power” of distributed quantum state purification not in a single universal primitive, but in a hierarchy of tasks whose feasibility depends sharply on prior structure: known versus unknown target states, Bell-diagonal versus arbitrary noise, one-shot versus asymptotic regime, and local-only versus network-level coordination.