Simple Algorithm
Quantum Computing · Rydberg State · 28 min

Entanglement Distillation

Purifying many noisy pairs into fewer perfect ones.

Every real channel degrades entanglement. Distillation is the remedy: Alice and Bob sacrifice some noisy pairs, using only local operations and classical talk, to purify others. How many perfect Bell pairs can nn noisy ones yield? The answer defines the distillable entanglement, and comparing it with the cost of creating the state reveals an irreversibility at the heart of entanglement theory.

What you’ll learn

  • Define the distillable entanglement EDE_D and the entanglement cost ECE_C, and explain why ED≤ECE_D\le E_C, with equality for pure states.
  • Derive the bilateral-CNOT recurrence map on Bell-diagonal states, including the BBPSSW fidelity formula and the DEJMPS improvement.
  • Explain the hashing protocol, its yield 1−H(p)1-H(p), and the coherent-information lower bound on EDE_D.
  • Prove that PPT states are undistillable, and explain bound entanglement and irreversibility.

Two numbers: distillable entanglement and cost

Alice and Bob share nn copies of a mixed state ρAB\rho_{AB}. They want as many Bell pairs as possible, with an error that vanishes as n→∞n\to\infty, using LOCC. Conversely, how many Bell pairs does it take to create ρ⊗n\rho^{\otimes n}? Write Φ2=∣Φ+⟩ ⁣⟨Φ+∣\Phi_2=\dyad{\Phi^+}.

For pure states both equal the entropy of entanglement S(ρA)S(\rho_A), by concentration and dilution. For mixed states, Hayden, Horodecki & Terhal (2001) showed EC(ρ)=lim⁡n1nEF(ρ⊗n)E_C(\rho)=\lim_n\tfrac1nE_F(\rho^{\otimes n}), the regularized entanglement of formation EF(ρ)=min⁡∑ipi S(Tr⁡B∣ψi⟩ ⁣⟨ψi∣)E_F(\rho)=\min\sum_ip_i\,S(\Tr_B\dyad{\psi_i}), minimized over decompositions ρ=∑ipi∣ψi⟩ ⁣⟨ψi∣\rho=\sum_ip_i\dyad{\psi_i}. Additivity of EFE_F fails in general (Hastings 2009, via Shor's equivalence of the additivity conjectures), one reason ECE_C is hard to compute.

Recurrence: BBPSSW and DEJMPS

Label the Bell states by two bits: a phase bit (00 for Φ+,Ψ+\Phi^+,\Psi^+; 11 for Φ−,Ψ−\Phi^-,\Psi^-) and a parity bit (00 for Φ±\Phi^\pm; 11 for Ψ±\Psi^\pm). The basic step acts on two pairs, a source and a target: Alice applies a CNOT from her half of the source to her half of the target, and Bob does the same on his side. This bilateral CNOT permutes Bell states:

∣Bp1q1⟩src∣Bp2q2⟩tgt ⟼ ∣Bp1⊕p2, q1⟩src∣Bp2, q1⊕q2⟩tgt\ket{B_{p_1q_1}}_{\text{src}}\ket{B_{p_2q_2}}_{\text{tgt}}\ \longmapsto\ \ket{B_{p_1\oplus p_2,\ q_1}}_{\text{src}}\ket{B_{p_2,\ q_1\oplus q_2}}_{\text{tgt}}
Parity bits flow forward (source to target); phase bits flow backward (target to source).

Both then measure their halves of the target in the ZZ basis and compare results by phone. Their outcomes agree iff q1⊕q2=0q_1\oplus q_2=0. They keep the source pair if the outcomes agree, and discard it otherwise.

DerivationOne round on Bell-diagonal inputs
  1. Pkeep=(pΦ++pΦ−)2+(pΨ++pΨ−)2P_{\text{keep}}=\big(p_{\Phi^+}+p_{\Phi^-}\big)^2+\big(p_{\Psi^+}+p_{\Psi^-}\big)^2

    Inputs are independent copies of ρ=∑bpb∣Bb⟩ ⁣⟨Bb∣\rho=\sum_bp_b\dyad{B_b}. Keep iff the parity bits agree: both pairs Φ\Phi-type or both Ψ\Psi-type.

  2. pΦ+′=pΦ+2+pΦ−2Pkeep,pΦ−′=2 pΦ+pΦ−Pkeepp'_{\Phi^+}=\frac{p_{\Phi^+}^2+p_{\Phi^-}^2}{P_{\text{keep}}},\qquad p'_{\Phi^-}=\frac{2\,p_{\Phi^+}p_{\Phi^-}}{P_{\text{keep}}}

    A kept source carries phase p1⊕p2p_1\oplus p_2: it is Φ+\Phi^+ if both pairs were Φ+\Phi^+ or both were Φ−\Phi^-.

  3. pΨ+′=pΨ+2+pΨ−2Pkeep,pΨ−′=2 pΨ+pΨ−Pkeepp'_{\Psi^+}=\frac{p_{\Psi^+}^2+p_{\Psi^-}^2}{P_{\text{keep}}},\qquad p'_{\Psi^-}=\frac{2\,p_{\Psi^+}p_{\Psi^-}}{P_{\text{keep}}}

    Likewise in the Ψ\Psi sector. Parity errors (the Ψ\Psi weight) are suppressed quadratically; phase errors are not detected at all.

  4. F′=F2+19(1−F)2F2+23F(1−F)+59(1−F)2F'=\frac{F^2+\frac19(1-F)^2}{F^2+\frac23F(1-F)+\frac59(1-F)^2}

    BBPSSW (1996): feed in Werner-form pairs p=(F,1−F3,1−F3,1−F3)p=\big(F,\tfrac{1-F}{3},\tfrac{1-F}{3},\tfrac{1-F}{3}\big) and twirl the output back to Werner form. Then F′>FF'>F exactly when F>12F>\tfrac12; F=12F=\tfrac12 is a fixed point.

  5. DEJMPS (1996): instead of twirling, Alice applies Rx(π/2)R_x(\pi/2) and Bob Rx(−π/2)R_x(-\pi/2) to each pair before every round. This swaps the Φ−\Phi^- and Ψ−\Psi^- weights, turning undetected phase errors into detectable parity errors for the next round. It converges much faster and drives every Bell-diagonal state with pΦ+>12p_{\Phi^+}>\tfrac12 toward ∣Φ+⟩\ket{\Phi^+}.

Predict before you look

Start with Werner pairs of fidelity F=0.75F=0.75. After one BBPSSW round, what is the fidelity of the pairs that are kept?

Show the answer

0.79

F′=0.5625+0.00690.5625+0.125+0.0347=0.56940.7222≈0.788F'=\dfrac{0.5625+0.0069}{0.5625+0.125+0.0347}=\dfrac{0.5694}{0.7222}\approx0.788, and the round succeeds with probability ≈0.72\approx0.72. Each success turns two pairs into one, so after kk rounds the yield is at most 2−k2^{-k}: recurrence alone has zero asymptotic rate as F→1F\to1.

Hashing: a positive yield

Recurrence protocols discard at least half of the pairs in every round. The hashing protocol (Bennett, DiVincenzo, Smolin & Wootters, 1996) achieves a constant rate by treating the unknown Bell labels as a classical string to be learned.

DerivationThe hashing yield
  1. ρ⊗n=∑bp(b) ∣Bb⟩ ⁣⟨Bb∣,p(b)=∏i=1npbi\rho^{\otimes n}=\sum_{\mathbf b}p(\mathbf b)\,\dyad{B_{\mathbf b}},\qquad p(\mathbf b)=\prod_{i=1}^{n}p_{b_i}

    nn Bell-diagonal pairs form a classical mixture over strings b\mathbf b of Bell labels: 2n2n unknown bits carrying entropy H(p)H(p) per pair.

  2. With high probability b\mathbf b lies in a typical set of about 2nH(p)2^{nH(p)} strings, so identifying it takes about nH(p)nH(p) bits of information.

  3. Bilateral local Clifford operations (CNOTs and local rotations) act on the Bell labels as invertible linear maps over F2\mathbb F_2. With them Alice and Bob collect a chosen parity s⋅b\mathbf s\cdot\mathbf b of the whole string onto one pair, measure that pair and compare: one pair sacrificed, one bit learned. Random parities roughly halve the set of candidate strings each time.

  4. DH=1−H(p)perfect pairs per input pairD_H=1-H(p)\quad\text{perfect pairs per input pair}

    After about nH(p)nH(p) measurements the string is known, local Pauli corrections fix every remaining pair, and about n(1−H(p))n\big(1-H(p)\big) perfect Bell pairs remain.

  5. ED(ρAB) ≥ S(ρB)−S(ρAB)E_D(\rho_{AB})\ \ge\ S(\rho_B)-S(\rho_{AB})

    The general hashing inequality (Devetak & Winter, 2005): the coherent information is achievable with one-way LOCC from Alice to Bob. For Bell-diagonal states S(ρB)=1S(\rho_B)=1 and S(ρAB)=H(p)S(\rho_{AB})=H(p), recovering the hashing yield.

For Werner-form states H(p)=h(F)+(1−F)log⁡23H(p)=h(F)+(1-F)\log_23, so hashing alone works only for F≳0.811F\gtrsim0.811. Combining the methods, recurrence first to push FF above that threshold and hashing afterwards, gives a positive yield for every F>12F>\tfrac12. Since local filtering and twirling bring any entangled two-qubit state to Werner form with F>12F>\tfrac12, every entangled two-qubit state is distillable (Horodecki, Horodecki & Horodecki, 1997).

Quantity (isotropic two-qubit state, F=0.9F=0.9)ValueRole
Hashing yield 1−H(p)1-H(p)≈0.37\approx0.37achievable: lower bound on EDE_D
Relative entropy of entanglement 1−h(F)1-h(F)≈0.53\approx0.53upper bound on EDE_D
Logarithmic negativity log⁡2(2F)\log_2(2F)≈0.85\approx0.85upper bound on EDE_D
Entanglement of formation h(1+1−C22)h\big(\tfrac{1+\sqrt{1-C^2}}{2}\big) with C=2F−1C=2F-1≈0.72\approx0.72upper bound on ECE_C (Wootters' formula)
hh is the binary entropy. No closed formula for EDE_D of two-qubit Werner states is known; here it lies somewhere in [0.37, 0.53][0.37,\,0.53].

Bound entanglement and irreversibility

DerivationProof
  1. σΓ⪰0 ⇒ (σ⊗n)Γ=(σΓ)⊗n⪰0\sigma^\Gamma\succeq0\ \Rightarrow\ \big(\sigma^{\otimes n}\big)^\Gamma=\big(\sigma^\Gamma\big)^{\otimes n}\succeq0

    Write Γ\Gamma for the partial transpose on Bob's side. Tensor powers of a PPT state are PPT.

  2. [(A⊗B) σ (A⊗B)†]Γ=(A⊗Bˉ) σΓ (A⊗Bˉ)†⪰0\Big[(A\otimes B)\,\sigma\,(A\otimes B)^\dagger\Big]^\Gamma=(A\otimes\bar B)\,\sigma^\Gamma\,(A\otimes\bar B)^\dagger\succeq0

    Every separable operation, hence every LOCC protocol, preserves PPT: the partial transpose turns Bob's Kraus operator into its complex conjugate.

  3. Tr⁡(σ Φd)=Tr⁡(σΓ ΦdΓ)=1dTr⁡(σΓ F)≤1d ∥σΓ∥1∥F∥∞=1d\Tr(\sigma\,\Phi_d)=\Tr\big(\sigma^\Gamma\,\Phi_d^\Gamma\big)=\tfrac1d\Tr\big(\sigma^\Gamma\,\mathbb F\big)\le\tfrac1d\,\norm{\sigma^\Gamma}_1\norm{\mathbb F}_\infty=\tfrac1d

    For the maximally entangled state Φd\Phi_d of local dimension dd, ΦdΓ=F/d\Phi_d^\Gamma=\mathbb F/d with F\mathbb F the swap operator. A PPT state has ∥σΓ∥1=Tr⁡σΓ=1\norm{\sigma^\Gamma}_1=\Tr\sigma^\Gamma=1, and ∥F∥∞=1\norm{\mathbb F}_\infty=1.

  4. So every state reachable from σ⊗n\sigma^{\otimes n} by LOCC has overlap at most 2−m2^{-m} with mm Bell pairs (at most 12\tfrac12 with a single one), whereas distillation needs overlap close to 11. Hence ED(σ)=0E_D(\sigma)=0. ■\blacksquare

Entangled PPT states exist in every dimension beyond 2×32\times3: in 2×42\times4, 3×33\times3 and larger (P. Horodecki, 1997; see Separability, PPT & Entanglement Witnesses). Their entanglement is bound: it took entanglement to create them (every bound entangled state has EC>0E_C>0; Yang, Horodecki, Horodecki & Synak-Radtke, 2005), yet none of it can be distilled back. Entanglement theory is therefore irreversible: ED<ECE_D<E_C.

Check your understanding

Why does the PPT argument rule out distillation even with unlimited copies and unlimited rounds of communication?

  1. Because PPT states are separable
  2. Because PPT is preserved by tensor powers and by every LOCC map, while states close to Bell pairs are NPT
  3. Because PPT states have zero entanglement cost
  4. Because the hashing yield of a PPT state is negative
Show the answer

B. Because PPT is preserved by tensor powers and by every LOCC map, while states close to Bell pairs are NPT

The argument is an invariant: the PPT property survives tensor powers and every separable (hence LOCC) operation, and any PPT state has overlap at most 1/d1/d with Φd\Phi_d. No protocol, however long, can leave the PPT set and approach Bell pairs.

Key takeaways

  • EDE_D (Bell pairs out) ≤EC\le E_C (Bell pairs in); both equal S(ρA)S(\rho_A) for pure states, and ECE_C is the regularized entanglement of formation.
  • The bilateral CNOT moves parity forward and phase backward. BBPSSW improves Werner fidelity iff F>12F>\tfrac12; DEJMPS rotates between rounds so that phase errors become detectable.
  • Hashing yields 1−H(p)1-H(p) (positive for Werner states with F≳0.811F\gtrsim0.811); recurrence followed by hashing gives a positive rate for every F>12F>\tfrac12; in general ED≥S(ρB)−S(ρAB)E_D\ge S(\rho_B)-S(\rho_{AB}).
  • PPT states are undistillable: PPT is closed under LOCC and tensor powers, and PPT states have overlap at most 1/d1/d with Φd\Phi_d. Bound entanglement makes the theory irreversible.
  • Whether NPT bound entanglement exists is open, and EDE_D is unknown even for two-qubit Werner states.

Check yourself

In the interactive lesson these are scored with your confidence, and what you miss comes back for review.

Question 1

A Bell-diagonal state has weights 0.90.9 on Φ+\Phi^+ and 0.10.1 on Φ−\Phi^- (and zero on Ψ±\Psi^\pm). What yield does the hashing protocol achieve?

  1. 0.90.9
  2. 0.470.47
  3. 0.530.53
  4. 0.80.8
Show the answer

C. 0.530.53

H(p)=h(0.1)≈0.469H(p)=h(0.1)\approx0.469, so DH=1−0.469≈0.53D_H=1-0.469\approx0.53 Bell pairs per input pair. For this rank-2 state the relative entropy of entanglement gives the same value, so hashing is optimal here: ED=0.53E_D=0.53.

Question 2

Why does recurrence-only distillation (BBPSSW or DEJMPS, iterated) have zero asymptotic yield?

  1. Each round turns at least two pairs into at most one, so kk rounds yield at most 2−k2^{-k}, and reaching F→1F\to1 needs k→∞k\to\infty
  2. It works only for F>0.81F>0.81
  3. It requires sending qubits between the labs
  4. It converges to a fixed point strictly below F=1F=1
Show the answer

A. Each round turns at least two pairs into at most one, so kk rounds yield at most 2−k2^{-k}, and reaching F→1F\to1 needs k→∞k\to\infty

Recurrence improves fidelity but halves (at least) the number of pairs per round. Constant-rate distillation needs hashing-type protocols, which learn the error string collectively.

Question 3

Which statement is always true?

  1. EC≤EDE_C\le E_D for every state
  2. ED≤ECE_D\le E_C, with equality for pure states
  3. ED=ECE_D=E_C for every two-qubit state
  4. EDE_D is at least the logarithmic negativity
Show the answer

B. ED≤ECE_D\le E_C, with equality for pure states

You cannot distill more Bell pairs than it took to create the state, since LOCC cannot create entanglement. For pure states both equal S(ρA)S(\rho_A).

Question 4

A 3×33\times3 state is PPT but entangled. What follows?

  1. It can be distilled, given enough copies
  2. It is separable after all
  3. It can be converted into an exact Bell pair with small but nonzero probability
  4. Its distillable entanglement is zero, but its entanglement cost is positive
Show the answer

D. Its distillable entanglement is zero, but its entanglement cost is positive

PPT implies ED=0E_D=0, while every entangled state has EC>0E_C>0: this is bound entanglement, the clearest instance of irreversibility.

Further reading

  • C. H. Bennett, G. Brassard, S. Popescu, B. Schumacher, J. A. Smolin & W. K. Wootters, "Purification of noisy entanglement and faithful teleportation via noisy channels", Phys. Rev. Lett. 76, 722 (1996).
  • D. Deutsch, A. Ekert, R. Jozsa, C. Macchiavello, S. Popescu & A. Sanpera, "Quantum privacy amplification and the security of quantum cryptography over noisy channels", Phys. Rev. Lett. 77, 2818 (1996).
  • C. H. Bennett, D. P. DiVincenzo, J. A. Smolin & W. K. Wootters, "Mixed-state entanglement and quantum error correction", Phys. Rev. A 54, 3824 (1996).
  • M. Horodecki, P. Horodecki & R. Horodecki, "Mixed-state entanglement and distillation: is there a 'bound' entanglement in nature?", Phys. Rev. Lett. 80, 5239 (1998).
  • R. Horodecki, P. Horodecki, M. Horodecki & K. Horodecki, "Quantum entanglement", Rev. Mod. Phys. 81, 865 (2009).