Entanglement Distillation

Entanglement distillation is the conversion of several imperfect entangled states shared by distant parties into fewer states closer to a chosen maximally entangled target, using local quantum operations and classical communication (LOCC). It trades quantity for quality. A successful branch can have higher Bell-state fidelity than any one input pair, but the inputs consumed, failed branches, communication rounds, and local imperfections remain part of the protocol cost.

The task is also called entanglement purification. Neither name means purifying a density operator by adjoining an abstract reference system. Here Alice and Bob physically process distributed pairs and aim to produce usable shared entanglement.

Required background. Bell states fixes the target basis; LOCC explains local processing and classical messages. Density operators, entanglement measures and multi-qubit gates supply mixed-state fidelity, entanglement monotonicity and the controlled-NOT operation.

Conditional fidelity and pair yield

Alice holds registers $A_1,\ldots,A_n$ and Bob holds $B_1,\ldots,B_n$. Ideally, the input is an independent and identically distributed supply

$$\rho_{AB}^{\otimes n}.$$

An LOCC instrument produces a classical transcript, an acceptance flag and, when it accepts, some output pairs. The output state conditioned on acceptance and the probability of acceptance are separate quantities. The direction of classical communication also matters: one-way and two-way LOCC define different resource constraints.

For the target

$$\begin{aligned} \Phi^+&=|\Phi^+\rangle\langle\Phi^+|,\\ |\Phi^+\rangle&=\frac{|00\rangle+|11\rangle}{\sqrt2}, \end{aligned}$$

the accepted-state fidelity is

$$F_{\rm out} = \langle\Phi^+|\rho_{\rm out}|\Phi^+\rangle.$$

If $p_{\rm acc}$ is the acceptance probability, a two-to-one round has raw-pair yield

$$Y_{\rm pair} = \frac{p_{\rm acc}}{2}$$

before source failures, memory expiration, readout loss, verification samples, and scheduling overhead are included. Reporting only $F_{\rm out}>F_{\rm in}$ hides the central tradeoff.

The asymptotic resource-theory statement has a different scale. A sequence of protocols distills at rate $R$ when

$$\begin{aligned} \rho_{AB}^{\otimes n}&\xrightarrow{\ \mathrm{LOCC}\ } \text{approximately }\Phi_2^{\otimes m_n},\\ \frac{m_n}{n}&\longrightarrow R, \end{aligned}$$

while the output error tends to zero for the full protocol. If an abort flag is used, its probability must also tend to zero; conditioning on increasingly rare successes does not define the same distillation rate. Optimizing $R$ gives $E_D$ for the specified class of one-way or two-way LOCC. A finite recurrence experiment demonstrates a particular transformation. Determining the optimal asymptotic rate requires a stronger argument. Here $\Phi_2$ denotes one standard two-qubit ebit.

Bell-Diagonal Noise Model

Label the Bell basis by two error bits,

$$|\beta_{ab}\rangle=(I\otimes X^bZ^a)|\Phi^+\rangle.$$

Here $a,b\in\{0,1\}$, and $a$ is a phase-error label and $b$ is a bit-parity label. In the order used below,

Label	Bell state	Probability
$(0,0)$	$\lvert\Phi^+\rangle$	$A$
$(1,0)$	$\lvert\Phi^-\rangle$	$B$
$(0,1)$	$\lvert\Psi^+\rangle$	$C$
$(1,1)$	$\lvert\Psi^-\rangle$	$D$

A Bell-diagonal input is

$$\begin{aligned} \rho_{\rm Bell}&=\sum_{a,b}p_{ab}|\beta_{ab}\rangle\langle\beta_{ab}|,\\ (A,B,C,D)&=(p_{00},p_{10},p_{01},p_{11}). \end{aligned}$$

The particularly symmetric family used for the first recurrence calculation is

$$\rho_F = F\Phi^+ + \frac{1-F}{3} \left(I_4-\Phi^+\right).$$

This is often called a Werner state in purification discussions, although “Werner” and “isotropic” denote distinct invariant families in general dimension. Bell-isotropic state is the unambiguous name for the formula above. For two qubits it is entangled exactly when $F>1/2$.

Twirling and what it costs

A coordinated random bilateral unitary, followed by forgetting the random choice, can map an arbitrary two-qubit state to $\rho_F$ while preserving its overlap $F$ with the chosen Bell target. For $|\Phi^+\rangle$, the continuous version averages operations of the form $U\otimes U^*$; a finite shared Clifford ensemble is enough in practice.

Twirling is LOCC, but it is not a theorem that it helps every protocol. It erases Bell-basis asymmetries and coherences that a better-matched protocol might exploit. It also assumes Alice and Bob share the same target frame and coordinate the random choices. The symmetric model is valuable because it makes the recurrence transparent, not because all laboratory noise is depolarizing.

One BBPSSW Recurrence Round

The Bennett–Brassard–Popescu–Schumacher–Smolin–Wootters (BBPSSW) recurrence protocol of Bennett et al. (1996a) consumes two shared pairs. Call pair 1 the source pair and pair 2 the target pair.

1. Alice applies $\operatorname{CNOT}_{A_1\to A_2}$ and Bob applies $\operatorname{CNOT}_{B_1\to B_2}$.
2. Both parties measure their target qubit in the $Z$ basis.
3. They compare the two classical outcomes.
4. They retain the source pair only when the outcomes agree.
5. If another scalar recurrence round is wanted, they twirl the retained pair back to the Bell-isotropic family.

One recurrence round. The source pair is retained when the two target measurement bits agree. Under the ideal Bell-isotropic model and $1/2<F<1$, its conditional fidelity is larger, while the raw-pair yield is $p_{\rm acc}/2$.

Error-label propagation

Let the two input Bell labels be $(a_1,b_1)$ and $(a_2,b_2)$. Conjugating the Pauli labels through the bilateral CNOT gives

$$\begin{aligned} (a_1,b_1) &\longmapsto (a_1\oplus a_2,b_1), \\ (a_2,b_2) &\longmapsto (a_2,b_1\oplus b_2). \end{aligned}$$

Computational-basis outcomes on a Bell state agree exactly when its parity label is zero. Acceptance therefore tests

$$b_1\oplus b_2=0,$$

without revealing the individual source parity. This is a distributed error check: the target pair is consumed to learn one syndrome bit about the pair labels.

For independent, identical Bell-diagonal inputs, the acceptance probability is

$$N = (A+B)^2+(C+D)^2.$$

Conditioned on acceptance, the retained source has probabilities

$$\begin{aligned} A'&=\frac{A^2+B^2}{N},\\ B'&=\frac{2AB}{N},\\ C'&=\frac{C^2+D^2}{N},\\ D'&=\frac{2CD}{N}. \end{aligned}$$

These equations expose what the parity check does. Cases with mismatched bit labels are rejected. Within an accepted parity sector, equal phase labels map to phase label zero and unequal phase labels map to phase label one. A later round with local basis changes can exchange the roles of bit and phase errors.

Scalar fidelity map

For the Bell-isotropic input, set

$$\begin{aligned} q&=\frac{1-F}{3},\\ (A,B,C,D)&=(F,q,q,q). \end{aligned}$$

Then

$$\begin{aligned} p_{\rm acc}(F) &= F^2+2Fq+5q^2 \\ &= \frac{8F^2-4F+5}{9}, \end{aligned}$$

and the retained pair has target fidelity

$$\begin{aligned} F'&=\frac{F^2+q^2}{p_{\rm acc}(F)}\\ &=\frac{10F^2-2F+1}{8F^2-4F+5}. \end{aligned}$$

The gain factors as

$$F'-F = \frac{ 8(1-F)(F-\tfrac12)(F-\tfrac14) }{8F^2-4F+5}.$$

Once the target Bell component has been chosen as the largest component, the relevant regime has $F\geq1/4$. Thus the ideal scalar protocol improves the target fidelity for

$$\frac12<F<1.$$

The threshold $F=1/2$ is not merely an artifact of the recurrence formula: $\rho_F$ is separable at and below that value. LOCC cannot distill a Bell pair from separable inputs.

Numerical example

For $F=0.8$,

$$p_{\rm acc}\approx0.7689,\qquad F'\approx0.8382.$$

The accepted pair is better by about $0.0382$, but one output attempt consumed two inputs. The expected raw-pair yield is therefore

$$Y_{\rm pair}=\frac{p_{\rm acc}}2\approx0.3844.$$

The higher fidelity comes with fewer available pairs.

Iteration and Yield

With an ideal twirl after every accepted round, define

$$\begin{aligned} F_{j+1}&=f(F_j),\\ f(F)&=\frac{10F^2-2F+1}{8F^2-4F+5}. \end{aligned}$$

For a large pool with enough surviving pairs to form new groups at every stage, the expected number after $k$ rounds is approximately

$$N_k \simeq N_0\,2^{-k} \prod_{j=0}^{k-1}p_{\rm acc}(F_j).$$

For $F_0=0.8$, ideal iteration gives the following conditional fidelities and expected outputs per raw pair:

Rounds	Fidelity	Yield
0	$0.8000$	$1$
1	$0.8382$	$0.3844$
2	$0.8736$	$0.1552$
3	$0.9045$	$0.06563$
4	$0.9300$	$0.02890$

This idealized calculation assumes independent inputs, perfect local gates and measurements, no waiting-time decoherence, and a large pool of outputs from each round. Approaching unit fidelity by recurrence alone consumes an increasing number of raw pairs per output.

In a finite batch, pairing and postselection are stochastic. An odd surviving pair cannot be processed alone: the next round uses the integer part of half the surviving count. The product above therefore need not equal the exact finite-batch expectation. If one branch finishes early, its surviving pair may wait in memory for a partner. Nested rounds therefore need a scheduling policy: maximum pair age, timeout, re-pairing rule, and whether pairs from different calibration epochs may be combined.

DEJMPS and Other Recurrence Protocols

The Deutsch–Ekert–Jozsa–Macchiavello–Popescu–Sanpera (DEJMPS) protocol (Deutsch et al., 1996) is a closely related two-to-one recurrence scheme. Local basis rotations are chosen before the bilateral CNOT so that the accepted branch suppresses both relevant error classes efficiently. For Bell-diagonal inputs, DEJMPS keeps the four Bell probabilities rather than forcing the state through a full isotropic twirl after every round.

That distinction matters when $(B,C,D)$ are unequal. Two states with the same target fidelity $A$ can have different recurrence performance because their remaining probability is distributed among different Bell errors. A scalar $F$ is then not a sufficient state description. For the unrotated parity check above, $(A,B,C,D)=(0.6,0.4,0,0)$ is accepted with probability one but gives $A'=0.52$, below its input fidelity. In contrast, $(0.6,0,0.4,0)$ gives acceptance $0.52$ and $A'\approx0.6923$. The same target overlap can lead to different results because the check detects one type of error while combining another.

Choose the Bell convention, local basis rotations and acceptance rule together. Relabeling a target singlet as $|\Phi^+\rangle$ or exchanging phase and bit errors changes the written probability map, although the underlying local operations may be equivalent.

Hashing: An Asymptotic Protocol

Recurrence tests a small number of parity relations and discards aggressively. The hashing protocol of Bennett et al. (1996b) acts jointly on a large block of independent, identically distributed Bell-diagonal pairs. Alice and Bob use random bilateral parity checks to learn the typical bit-and-phase error string, consuming measured pairs as syndrome carriers. Once the error string is known with vanishing failure probability, they correct the unmeasured pairs.

For Bell probabilities $\mathbf p=(A,B,C,D)$, define the Shannon entropy

$$H(\mathbf p) = -\sum_{a,b}p_{ab}\log_2 p_{ab}.$$

The one-way hashing protocol achieves the asymptotic rate

$$R_{\rm hash} = 1-H(\mathbf p)$$

when the right-hand side is positive. More carefully, it proves the achievable lower bound

$$E_D^{\to}(\rho_{\rm Bell}) \geq \max\{0,1-H(\mathbf p)\};$$

it does not assert that hashing is optimal for every Bell-diagonal state.

For the Bell-isotropic distribution, with $0\log_20$ defined by continuity,

$$H_F={}-F\log_2F -(1-F)\log_2\!\left(\frac{1-F}{3}\right).$$

The hashing rate becomes positive at

$$F>0.81071.$$

This does not conflict with the recurrence threshold $F>1/2$. Two-way recurrence can first raise fidelity above the hashing threshold, after which a large surviving block can be hashed. Such recurrence-plus-hashing schemes trade several low-yield purification rounds for a positive asymptotic finishing rate.

Hashing is an asymptotic coding theorem. At finite blocklength, the number of checks must include statistical margins, target failure probability, and protocol overhead. Simply substituting an estimated distribution into $1-H(\mathbf p)$ is not a certified finite-batch yield.

Relation to Neighboring Tasks

Pure-state concentration

Entanglement concentration starts from pure but nonmaximally entangled pairs, for example

$$|\psi\rangle = \sqrt{\lambda}|00\rangle + \sqrt{1-\lambda}|11\rangle.$$

In the many-copy limit, Bell pairs can be extracted reversibly at the entropy of entanglement. Mixed-state distillation includes noise and is generally irreversible. The words are sometimes used broadly enough to overlap, so the input state class should always be named.

Quantum error correction

Distillation and quantum error correction extract the same syndrome logic in different operational arrangements. A one-way entanglement-purification protocol for Bell pairs sent through a channel can be converted into a quantum error-correcting code for that channel, and conversely (Bennett et al., 1996b). Stabilizer hashing checks are the distributed counterpart of code syndromes.

The protocols are not interchangeable at the hardware level. Recurrence uses two-way messages and postselection on already shared pairs. A quantum code encodes an unknown state and normally aims for deterministic logical recovery. Why Quantum Error Correction Is Possible develops that protection task.

Magic-state distillation

Magic State Distillation purifies nonstabilizer resource states inside a fault-tolerant computer. Entanglement distillation purifies a nonlocal resource shared across a bipartition. Both consume many noisy resources to produce fewer better ones, but their free operations, targets, thresholds, and applications differ.

QKD privacy amplification

Entanglement-based security proofs can imagine Alice and Bob distilling nearly perfect private Bell pairs before measuring them for a key. This viewpoint connects phase-error correction to privacy amplification. An implemented QKD system usually performs classical error correction and privacy amplification rather than building a universal quantum computer at each endpoint. Quantum Key Distribution explains composable security and finite-key corrections. BB84 makes the Shor–Preskill bridge concrete by relating virtual phase-error correction to privacy amplification in a prepare-and-measure protocol.

Fundamental Boundaries

Separability and PPT obstruction

LOCC maps separable states to separable states, even conditionally. No distillation protocol can create entanglement from an actually separable input. For bipartite states, positive partial transpose is a stronger obstruction: PPT states remain PPT under LOCC, whereas a two-qubit Bell pair is NPT. Therefore every PPT entangled state has zero ordinary LOCC distillable entanglement (Horodecki et al., 1998). Exercise 6 supplies the quantitative fidelity obstruction, including the approximate limit.

For two qubits, PPT is equivalent to separability, and every entangled state is distillable with two-way LOCC in the asymptotic sense. In larger local dimensions, PPT entangled states exist and provide established examples of bound entanglement. See Negativity and PPT Criterion for the partial-transpose calculation.

In higher dimensions, an NPT test does not by itself supply a general distillation protocol or a rate. A failure to distill one or two copies also does not establish failure for arbitrarily many copies. Entanglement measures discusses the distinction between finite-copy tests and asymptotic distillability.

Imperfect local operations

The ideal recurrence assumes its local operations are cleaner than the links being purified. With a noisy local instrument $\mathcal N_{\rm loc}$ and acceptance map $\mathcal M_{\rm acc}$, the physical output is

$$\begin{aligned} \sigma_{\rm acc}&=(\mathcal M_{\rm acc}\circ\mathcal N_{\rm loc})(\rho^{\otimes2}),\\ \tau_{\rm out}&=\operatorname{Tr}_{A_2B_2}\sigma_{\rm acc},\\ p_{\rm acc}&=\operatorname{Tr}\tau_{\rm out},\\ \rho_{\rm out}&=\tau_{\rm out}/p_{\rm acc}. \end{aligned}$$

Gate, readout, reset, leakage, and memory errors can reduce the gain, move the input threshold, and produce a fidelity ceiling below one. No universal “purification threshold” exists without a specified noise channel and protocol.

Correlated and drifting inputs

The tensor-power model $\rho^{\otimes n}$ excludes temporal drift, common-mode phase noise, source afterpulsing, crosstalk, and adversarial correlations. A protocol may still work under weaker exchangeability or security assumptions, but the proof must say so. Random permutation can enforce useful symmetry; it does not magically make physical trials independent.

Distillation Inside a Network

A quantum repeater combines link generation, storage, swapping and distillation (Dür et al., 1999).

Swapping and distillation solve different problems:

• entanglement swapping extends the endpoints of a link and normally degrades quality;
• distillation consumes parallel links to improve the conditional quality of a surviving link.

A first-generation repeater nests the two operations. Elementary segments are generated and possibly distilled, neighboring segments are swapped, and the longer links may be distilled again. Every level introduces waiting-time and classical-latency costs.

For a two-to-one round, $p_{\rm acc}/2$ is the expected output per already available raw pair. Converting that yield to outputs per second requires the preparation and processing time.

For a simple model, let each cycle start with empty memories, prepare two compatible pairs, perform one round, and then reset after either success or failure. Suppose these cycles are independent and identically distributed, with finite mean duration $\mathbb E[T_{\rm cycle}]$. If at most one usable pair can be produced per cycle, the long-run rate is

$$R_{\rm usable}=\frac{p_{\rm acc}p_{\rm usable|acc}}{\mathbb E[T_{\rm cycle}]}.$$

This is the renewal-reward ratio: average outputs per cycle divided by average seconds per cycle. Duration and outcome may be correlated within a cycle; the averages must describe the same complete cycle, including failed attempts and timeouts. A protocol that retains some pairs across failures needs a different cycle definition or a state-based scheduling model. Exercise 8 uses the simple reset-after-each-round model.

Experimental verification and delivery also differ. A pair destroyed during tomography cannot then be used for a remote gate, while testing a sample of an output stream consumes part of the yield. Compare conditional fidelity and its uncertainty with the full number of raw attempts and the time needed to deliver surviving pairs.

Exercises

1. Derive the acceptance probability

For independent Bell-diagonal inputs with probabilities $(A,B,C,D)$, show that equal target measurement outcomes occur with probability

$$N=(A+B)^2+(C+D)^2.$$

Solution

The target Bell parity after the bilateral CNOT is $b_1\oplus b_2$. Equal $Z$ outcomes require this bit to vanish, so the two input parity labels must agree. The probability that both have $b=0$ is $(A+B)^2$; the probability that both have $b=1$ is $(C+D)^2$. These events are disjoint, giving

$$p_{\rm acc} = (A+B)^2+(C+D)^2.$$

2. Check the Bell-label map

Conditioned on acceptance, derive $A'$ and $B'$ for the retained source pair.

Solution

For the accepted $b=0$ sector, the source phase label is $a_1\oplus a_2$. It vanishes for the two input combinations $(a_1,a_2)=(0,0)$ and $(1,1)$, whose probabilities are $A^2$ and $B^2$. It equals one for $(0,1)$ and $(1,0)$, with total probability $2AB$. Normalizing by $N$ gives

$$A'=\frac{A^2+B^2}{N}, \qquad B'=\frac{2AB}{N}.$$

The $b=1$ calculation similarly gives $C'$ and $D'$.

3. Fidelity and yield at one round

Evaluate $F'$, $p_{\rm acc}$, and expected outputs per raw input pair for a Bell-isotropic input with $F=0.7$.

Solution

Here $q=(1-0.7)/3=0.1$, so

$$\begin{aligned} p_{\rm acc}&=0.7^2+2(0.7)(0.1)+5(0.1)^2=0.68,\\ F'&=\frac{0.7^2+0.1^2}{0.68}\approx0.7353. \end{aligned}$$

The expected output count per raw pair is

$$\frac{p_{\rm acc}}{2}=0.34.$$

4. Fixed points and stability

Find the fixed points of the scalar recurrence in $0\leq F\leq1$ and determine the direction of flow in the physically targeted regime $F\geq1/4$.

Solution

The factorization

$$F'-F = \frac{ 8(1-F)(F-\tfrac12)(F-\tfrac14) }{8F^2-4F+5}$$

shows fixed points at $F=1/4$, $1/2$, and $1$. The denominator is positive. For $1/4<F<1/2$, the map decreases $F$ toward $1/4$; for $1/2<F<1$, it increases $F$ toward $1$. The point $F=1/2$ is the unstable threshold separating the two behaviors under the ideal twirled model.

5. Hashing threshold

Show numerically that Bell-isotropic hashing first has positive rate near $F=0.8107$. What is the rate at $F=0.9$?

Solution

Solve $H_F=1$ using the Bell-isotropic entropy above.

The solution in $1/2<F<1$ is

$$F=0.81071.$$

At $F=0.9$,

$$\begin{aligned} H_F &= -0.9\log_2 0.9-0.1\log_2(0.1/3) \\ &\approx 0.62749, \end{aligned}$$

so the ideal asymptotic hashing rate is

$$R_{\rm hash}\approx0.37251$$

Bell pairs per input pair.

6. Why PPT states cannot be distilled

Give a short partial-transpose argument that an LOCC protocol cannot distill a PPT input into a two-qubit Bell pair.

Solution

Every LOCC branch is a separable operation with product Kraus operators. Such operations preserve positivity of the partial transpose, including after tensoring copies and conditioning on a nonzero-probability branch. A Bell pair has a negative partial transpose. For a normalized two-qubit PPT output $\sigma$, let $S$ be the swap operator. Since $(\Phi^+)^{T_B}=S/2$ and the largest eigenvalue of $S$ is one,

$$\begin{aligned} \operatorname{Tr}(\Phi^+\sigma) &=\tfrac12\operatorname{Tr}(S\sigma^{T_B})\\ &\leq\tfrac12. \end{aligned}$$

Thus PPT outputs cannot even approach unit Bell fidelity. Discarding all but one target pair of a proposed many-pair output preserves PPT, so arbitrarily many copies do not remove this obstruction. PPT entangled states are consequently bound entangled under ordinary LOCC distillation.

7. Monotonicity and postselection

Why does an accepted branch with more entanglement than one input pair not contradict LOCC monotonicity?

Solution

For an entanglement measure $E$ that is monotone on average under selective LOCC,

$$E(\rho_{AB}^{\otimes2})\geq\sum_r p_r E(\rho_r).$$

The sum includes every accepted and rejected branch. Its accepted contribution is $p_{\rm acc}E(\rho_{\rm out})$ when the accepted output is described by one state.

It does not forbid one conditional branch from having greater entanglement than one input copy. The protocol started with two copies, and the favorable branch occurs only probabilistically; rejected resources are consumed or degraded. Ignoring those branches is the apparent paradox.

8. From pairs per round to pairs per second

A node obtains the first raw pair after $4\ \mathrm{ms}$ on average and then waits another $6\ \mathrm{ms}$ for a compatible second pair. Local operations and the classical comparison take $1\ \mathrm{ms}$. If $p_{\rm acc}=0.75$ and $90\%$ of accepted pairs pass the age policy, find the long-run usable-pair rate when every round is followed by a reset and the cycles are independent and identically distributed.

Solution

The mean two-pair preparation time is $10\ \mathrm{ms}$ and the LOCC step adds $1\ \mathrm{ms}$. Thus

$$\begin{aligned} R_{\rm usable}&=\frac{(0.75)(0.90)}{0.011\ \mathrm{s}}\\ &\approx61.4\ \mathrm{s}^{-1}. \end{aligned}$$

The product $0.75\times0.90$ is the probability that a cycle yields a usable pair, with the second factor conditional on acceptance. The denominator includes all cycle time. Correlations within a cycle do not invalidate the ratio; a different timeout or retry policy can change both its numerator and denominator.

References

• C. H. Bennett, G. Brassard, S. Popescu, B. Schumacher, J. A. Smolin, and W. K. Wootters, “Purification of Noisy Entanglement and Faithful Teleportation via Noisy Channels,” Physical Review Letters 76, 722–725 (1996a), doi:10.1103/PhysRevLett.76.722.
• C. H. Bennett, D. P. DiVincenzo, J. A. Smolin, and W. K. Wootters, “Mixed-State Entanglement and Quantum Error Correction,” Physical Review A 54, 3824–3851 (1996b), doi:10.1103/PhysRevA.54.3824.
• D. Deutsch, A. Ekert, R. Jozsa, C. Macchiavello, S. Popescu, and A. Sanpera, “Quantum Privacy Amplification and the Security of Quantum Cryptography over Noisy Channels,” Physical Review Letters 77, 2818–2821 (1996), doi:10.1103/PhysRevLett.77.2818.
• W. Dür, H.-J. Briegel, J. I. Cirac, and P. Zoller, “Quantum Repeaters Based on Entanglement Purification,” Physical Review A 59, 169–181 (1999), doi:10.1103/PhysRevA.59.169.
• M. Horodecki, P. Horodecki, and R. Horodecki, “Mixed-State Entanglement and Distillation: Is There a ‘Bound’ Entanglement in Nature?” Physical Review Letters 80, 5239–5242 (1998), doi:10.1103/PhysRevLett.80.5239.