Simple Algorithm
Quantum Computing · Rydberg State · 30 min

Quantifying Entanglement

Entanglement entropy, concurrence and negativity.

How entangled is a state? For pure states there is one right answer, the entanglement entropy. For mixed states there are many inequivalent answers, each with its own operational meaning, and only a few of them, such as concurrence and negativity, can actually be computed.

What you’ll learn

  • State the axioms of an entanglement measure and explain why LOCC monotonicity is the central one.
  • Compute the entanglement entropy and concurrence of pure states, and apply Wootters' formula to two-qubit mixed states.
  • Define negativity and logarithmic negativity via the partial transpose, and compute them for Werner states.
  • Compare entanglement of formation, entanglement cost, distillable entanglement, relative entropy of entanglement and squashed entanglement.

What a measure must do

An entanglement measure is a function EE on bipartite states such that:

  1. E(ρ)=0E(\rho) = 0 whenever ρ\rho is separable;
  2. LOCC monotonicity: EE does not increase under LOCC, even on average over the outcomes of a protocol, ∑kpkE(ρk)≤E(ρ)\sum_kp_kE(\rho_k)\le E(\rho);
  3. EE is invariant under local unitaries (implied by 2, since they are reversible LOCC);
  4. normalization: E(Φd)=log⁡2dE(\Phi_d) = \log_2d for a maximally entangled state of local dimension dd.

Convexity, additivity and continuity are often added. No single function has every desirable property, which is why there is a zoo.

Pure states: entropy and concurrence

E(ψ)=S(ρA)=−∑iλi2log⁡2λi2 ∈ [0,log⁡2d]E(\psi) = S(\rho_A) = -\sum_i\lambda_i^2\log_2\lambda_i^2\ \in\ [0,\log_2d]
The entanglement entropy in ebits, from the Schmidt coefficients λi\lambda_i: the von Neumann entropy of either marginal. The Rényi versions Sα=11−αlog⁡2∑iλi2αS_\alpha = \tfrac{1}{1-\alpha}\log_2\sum_i\lambda_i^{2\alpha} are Schur-concave in the Schmidt probabilities, hence also non-increasing under LOCC on pure states.
DerivationConcurrence of a pure two-qubit state
  1. (Y⊗Y)∣ψ∗⟩=−δ‾∣00⟩+γ‾∣01⟩+β‾∣10⟩−α‾∣11⟩(Y\otimes Y)\ket{\psi^*} = -\overline\delta\ket{00}+\overline\gamma\ket{01}+\overline\beta\ket{10}-\overline\alpha\ket{11}

    Define the spin flip ∣ψ~⟩=(Y⊗Y)∣ψ∗⟩\ket{\tilde\psi} = (Y\otimes Y)\ket{\psi^*} (complex conjugation in the computational basis) and C(ψ)=∣⟨ψ∣ψ~⟩∣\mathcal{C}(\psi) = \abs{\braket{\psi|\tilde\psi}}. For ∣ψ⟩=α∣00⟩+β∣01⟩+γ∣10⟩+δ∣11⟩\ket\psi = \alpha\ket{00}+\beta\ket{01}+\gamma\ket{10}+\delta\ket{11}:

  2. ⟨ψ∣ψ~⟩=−2 (αδ−βγ)‾⟹C(ψ)=2∣αδ−βγ∣=2∣det⁡C∣\braket{\psi|\tilde\psi} = -2\,\overline{(\alpha\delta-\beta\gamma)}\quad\Longrightarrow\quad\mathcal{C}(\psi) = 2\abs{\alpha\delta-\beta\gamma} = 2\abs{\det C}

    Take the overlap with ∣ψ⟩\ket\psi: the spin flip turns the determinant test into a number.

  3. C=2λ1λ2=2det⁡ρA,E=h(1+1−C22)\mathcal{C} = 2\lambda_1\lambda_2 = 2\sqrt{\det\rho_A},\qquad E = h\Big(\tfrac{1+\sqrt{1-\mathcal{C}^2}}{2}\Big)

    In Schmidt form ∣det⁡C∣=λ1λ2\abs{\det C} = \lambda_1\lambda_2. Since EE is a monotonic function of C\mathcal C (hh is the binary entropy), the two carry the same information for pure two-qubit states.

Mixed states: the convex roof and Wootters' formula

For mixed states S(ρA)S(\rho_A) fails as a measure: I2⊗I2\tfrac I2\otimes\tfrac I2 is a product state with S(ρA)=1S(\rho_A) = 1. The entanglement of formation extends a pure-state measure by the cheapest decomposition, a "convex roof":

EF(ρ)=min⁡{∑kpk E(ψk) : ∑kpk∣ψk⟩ ⁣⟨ψk∣=ρ}.E_F(\rho) = \min\Big\{\sum_kp_k\,E(\psi_k)\ :\ \sum_kp_k\dyad{\psi_k} = \rho\Big\}.

Minimizing over all decompositions is generally intractable, which makes the next theorem remarkable.

DerivationWerner states
  1. ρp=1+3p4∣Ψ−⟩ ⁣⟨Ψ−∣+1−p4(∣Ψ+⟩ ⁣⟨Ψ+∣+∣Φ+⟩ ⁣⟨Φ+∣+∣Φ−⟩ ⁣⟨Φ−∣)\rho_p = \tfrac{1+3p}{4}\dyad{\Psi^-}+\tfrac{1-p}{4}\big(\dyad{\Psi^+}+\dyad{\Phi^+}+\dyad{\Phi^-}\big)

    Let ρp=p ∣Ψ−⟩ ⁣⟨Ψ−∣+(1−p) I/4\rho_p = p\,\dyad{\Psi^-}+(1-p)\,I/4. It is diagonal in the Bell basis:

  2. C(ρp)=max⁡(0, 1+3p2−1)=max⁡(0, 3p−12)\mathcal{C}(\rho_p) = \max\Big(0,\ \frac{1+3p}{2}-1\Big) = \max\Big(0,\ \frac{3p-1}{2}\Big)

    Bell states are real eigenvectors of Y⊗YY\otimes Y, so ρ~=ρ\tilde\rho = \rho for any Bell-diagonal state; the μi\mu_i are the Bell weights and C=max⁡(0,2wmax⁡−1)\mathcal C = \max(0,2w_{\max}-1):

  3. ρp entangled  ⟺  p>13,CHSH violated  ⟺  p>12\rho_p\ \text{entangled}\iff p>\tfrac13,\qquad \text{CHSH violated}\iff p>\tfrac{1}{\sqrt2}

    So ρp\rho_p is entangled exactly when p>13p>\tfrac13, yet it violates the CHSH inequality only for p>12p>\tfrac{1}{\sqrt2}: entanglement and Bell nonlocality are different resources.

Predict before you look

What is the concurrence of the Werner state with p=0.6p = 0.6?

Show the answer

0.4

C=3p−12=1.8−12=0.4\mathcal C = \tfrac{3p-1}{2} = \tfrac{1.8-1}{2} = 0.4. Its entanglement of formation is h(1+0.842)≈h(0.958)≈0.25h\big(\tfrac{1+\sqrt{0.84}}{2}\big)\approx h(0.958)\approx0.25 ebit.

Negativity: entanglement you can compute

The partial transpose transposes only B's indices: (∣i⟩ ⁣⟨k∣⊗∣j⟩ ⁣⟨l∣)TB=∣i⟩ ⁣⟨k∣⊗∣l⟩ ⁣⟨j∣\big(\ketbra{i}{k}\otimes\ketbra{j}{l}\big)^{T_B} = \ketbra{i}{k}\otimes\ketbra{l}{j}. For a separable state, ρTB=∑kpk ρkA⊗(ρkB)T\rho^{T_B} = \sum_kp_k\,\rho^A_k\otimes(\rho^B_k)^{T} is again a valid state, which gives a powerful test:

N(ρ)=∥ρTB∥1−12=∑ν<0∣ν∣,EN(ρ)=log⁡2∥ρTB∥1\mathcal{N}(\rho) = \frac{\norm{\rho^{T_B}}_1-1}{2} = \sum_{\nu<0}\abs{\nu},\qquad E_{\mathcal N}(\rho) = \log_2\norm{\rho^{T_B}}_1
Negativity (the total weight of the negative eigenvalues ν\nu of ρTB\rho^{T_B}) and logarithmic negativity. Both are entanglement monotones (Vidal & Werner 2002) and cost one eigenvalue decomposition. ENE_{\mathcal N} upper-bounds distillable entanglement, is additive, and is not convex (Plenio 2005).
DerivationNegativity of Werner states and pure states
  1. ∣Ψ−⟩ ⁣⟨Ψ−∣TB=12I−∣Φ+⟩ ⁣⟨Φ+∣\dyad{\Psi^-}^{T_B} = \tfrac12I-\dyad{\Phi^+}

    Partial transposition moves the singlet's coherence ∣01⟩ ⁣⟨10∣\ketbra{01}{10} to ∣00⟩ ⁣⟨11∣\ketbra{00}{11}:

  2. N(ρp)=max⁡(0, 3p−14),EN(ρp)=log⁡21+3p2(p>13)\mathcal{N}(\rho_p) = \max\Big(0,\ \frac{3p-1}{4}\Big),\qquad E_{\mathcal N}(\rho_p) = \log_2\frac{1+3p}{2}\quad(p>\tfrac13)

    So ρpTB=p(12I−∣Φ+⟩ ⁣⟨Φ+∣)+1−p4I\rho_p^{T_B} = p\big(\tfrac12I-\dyad{\Phi^+}\big)+\tfrac{1-p}{4}I has eigenvalue 1−3p4\tfrac{1-3p}{4} on ∣Φ+⟩\ket{\Phi^+} and 1+p4\tfrac{1+p}{4} (three times) elsewhere:

  3. EN(ψ)=2log⁡2∑iλiE_{\mathcal N}(\psi) = 2\log_2\sum_i\lambda_i

    For pure states, ∥ψTB∥1=(∑iλi)2\norm{\psi^{T_B}}_1 = \big(\sum_i\lambda_i\big)^2, so ENE_{\mathcal N} is the Rényi-½ entropy of the Schmidt probabilities. For two qubits 2N=C2\mathcal N = \mathcal C on pure states (and on Werner states), but in general 2N≤C2\mathcal N\le\mathcal C, often strictly, for mixed states.

Check your understanding

A two-qubit state has negativity N(ρ)=0\mathcal N(\rho) = 0. What can you conclude?

  1. It is separable
  2. It is pure
  3. It may still be bound entangled
  4. Its concurrence may still be positive
Show the answer

A. It is separable

For 2×22\times2 systems PPT is equivalent to separability (Horodecki 1996), so N=0\mathcal N = 0 means separable, and then C=0\mathcal C = 0 too. Bound entangled PPT states exist only in larger dimensions, starting with 2×42\times4 and 3×33\times3.

The zoo, and what is still open

MeasureIdeaComputable?On pure states
Distillable entanglement EDE_Dbest rate of Bell pairs extractable by LOCCno formula, even for two-qubit Werner statesS(ρA)S(\rho_A)
Entanglement cost ECE_Cbest rate of Bell pairs needed to prepare ρ\rho; equals regularized EFE_F (Hayden, Horodecki & Terhal 2001)generally unknownS(ρA)S(\rho_A)
Entanglement of formation EFE_Fconvex roof of S(ρA)S(\rho_A)closed form for two qubits (Wootters)S(ρA)S(\rho_A)
Relative entropy of entanglement ERE_Rmin⁡σ separableS(ρ∥σ)\min_{\sigma\ \text{separable}}S(\rho\Vert\sigma)hard in general: the separable set has no efficient descriptionS(ρA)S(\rho_A)
Squashed entanglement EsqE_{sq}12inf⁡I(A;B∣E)\tfrac12\inf I(A;B\vert E) over extensions ρABE\rho_{ABE}hard: the extension is unboundedS(ρA)S(\rho_A)
Log-negativity ENE_{\mathcal N}log⁡2∥ρTB∥1\log_2\norm{\rho^{T_B}}_1yes, one eigendecomposition2log⁡2∑iλi ≥S(ρA)2\log_2\sum_i\lambda_i\ \ge S(\rho_A)

General inequalities: ED≤Esq≤EC≤EFE_D\le E_{sq}\le E_C\le E_F and ED≤ER≤EFE_D\le E_R\le E_F, as well as ED≤ENE_D\le E_{\mathcal N}. The asymptotic measures all coincide on pure states but separate dramatically on mixed ones: bound entangled states have ED=0<ECE_D = 0<E_C, an irreversibility with no pure-state analogue.

Key takeaways

  • Entanglement measures vanish on separable states and never increase under LOCC; on pure states the asymptotically unique measure is S(ρA)S(\rho_A).
  • Pure two-qubit concurrence is 2∣αδ−βγ∣2\abs{\alpha\delta-\beta\gamma}; Wootters' formula C=max⁡(0,μ1−μ2−μ3−μ4)\mathcal C = \max(0,\mu_1-\mu_2-\mu_3-\mu_4) gives EFE_F for every two-qubit state.
  • Negativity counts the negative eigenvalues of ρTB\rho^{T_B}: efficiently computable and an upper bound on distillable entanglement, but blind to PPT entanglement.
  • Werner states are entangled for p>13p>\tfrac13 but violate CHSH only for p>12p>\tfrac{1}{\sqrt2}.
  • Mixed-state measures form a hierarchy (ED≤ER≤EFE_D\le E_R\le E_F, ED≤Esq≤ECE_D\le E_{sq}\le E_C) and differ on bound entangled states; NPT bound entanglement remains open.

Check yourself

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

Question 1

What is the concurrence of cos⁡θ∣00⟩+sin⁡θ∣11⟩\cos\theta\ket{00}+\sin\theta\ket{11}?

  1. sin⁡2θ\sin^2\theta
  2. cos⁡2θ\cos2\theta
  3. ∣sin⁡2θ∣\abs{\sin2\theta}
  4. 2sin⁡θ2\sin\theta
Show the answer

C. ∣sin⁡2θ∣\abs{\sin2\theta}

C=2∣αδ−βγ∣=2∣cos⁡θsin⁡θ∣=∣sin⁡2θ∣\mathcal C = 2\abs{\alpha\delta-\beta\gamma} = 2\abs{\cos\theta\sin\theta} = \abs{\sin2\theta}: 00 at θ=0\theta = 0 and 11 at θ=π/4\theta = \pi/4.

Question 2

The Werner state p∣Ψ−⟩ ⁣⟨Ψ−∣+(1−p)I/4p\dyad{\Psi^-}+(1-p)I/4 is entangled if and only if:

  1. p>0p>0
  2. p>12p>\tfrac{1}{\sqrt2}
  3. p>12p>\tfrac12
  4. p>13p>\tfrac13
Show the answer

D. p>13p>\tfrac13

C=max⁡(0,3p−12)\mathcal C = \max(0,\tfrac{3p-1}{2}), and the partial transpose has the eigenvalue 1−3p4\tfrac{1-3p}{4}: both switch on at p=13p = \tfrac13.

Question 3

Which measure can be computed efficiently for every bipartite state, in any dimension?

  1. Entanglement of formation
  2. Logarithmic negativity
  3. Distillable entanglement
  4. Squashed entanglement
Show the answer

B. Logarithmic negativity

EN=log⁡2∥ρTB∥1E_{\mathcal N} = \log_2\norm{\rho^{T_B}}_1 needs a single eigenvalue decomposition. The price: it vanishes on PPT entangled states.

Question 4

Why can't S(ρA)S(\rho_A) serve as an entanglement measure for mixed states?

  1. It can be positive for product states such as I2⊗I2\tfrac I2\otimes\tfrac I2
  2. It is not invariant under local unitaries
  3. It can be negative
  4. It is not defined for mixed states
Show the answer

A. It can be positive for product states such as I2⊗I2\tfrac I2\otimes\tfrac I2

Classical uncertainty also makes ρA\rho_A mixed. For a pure global state, mixedness of ρA\rho_A signals entanglement; for a mixed global state it need not.

Further reading

  • W. K. Wootters, "Entanglement of formation of an arbitrary state of two qubits", Phys. Rev. Lett. 80, 2245 (1998).
  • 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).
  • G. Vidal & R. F. Werner, "Computable measure of entanglement", Phys. Rev. A 65, 032314 (2002).
  • M. B. Plenio & S. Virmani, "An introduction to entanglement measures", Quantum Inf. Comput. 7, 1 (2007).
  • R. Horodecki, P. Horodecki, M. Horodecki & K. Horodecki, "Quantum entanglement", Rev. Mod. Phys. 81, 865 (2009).