Simple Algorithm
Quantum Computing · Rydberg State · 30 min

Multipartite Entanglement: GHZ, W & SLOCC

Three qubits can be entangled in inequivalent ways.

With two parties, entanglement is a single ladder: more or less of the same thing. With three qubits the ladder splits. GHZ and W are both fully entangled, yet no local operation, not even one that succeeds with tiny probability, can turn one into the other.

What you’ll learn

  • Define SLOCC equivalence and relate it to invertible local operators.
  • Prove that GHZ and W are SLOCC-inequivalent, and state the six-class classification of three-qubit pure states.
  • Compute the three-tangle and apply the Coffman–Kundu–Wootters monogamy inequality.
  • Explain why four or more qubits have infinitely many SLOCC classes, and how genuine multipartite entanglement is certified in experiments.

Two kinds of "fully entangled"

∣GHZ⟩=12(∣000⟩+∣111⟩),∣W⟩=13(∣001⟩+∣010⟩+∣100⟩)\ket{\mathrm{GHZ}} = \tfrac{1}{\sqrt2}\big(\ket{000}+\ket{111}\big),\qquad \ket{W} = \tfrac{1}{\sqrt3}\big(\ket{001}+\ket{010}+\ket{100}\big)
Both are symmetric under permutations of the qubits, and both are entangled across every bipartition.
PropertyGHZW
One-qubit marginalI/2I/2 (maximally mixed)23∣0⟩ ⁣⟨0∣+13∣1⟩ ⁣⟨1∣\tfrac23\dyad0+\tfrac13\dyad1
Two-qubit marginal12(∣00⟩ ⁣⟨00∣+∣11⟩ ⁣⟨11∣)\tfrac12(\dyad{00}+\dyad{11}): separable13∣00⟩ ⁣⟨00∣+23∣Ψ+⟩ ⁣⟨Ψ+∣\tfrac13\dyad{00}+\tfrac23\dyad{\Psi^+}: concurrence 23\tfrac23
After losing one qubitno entanglement leftthe remaining pair is still entangled
Entropy of one qubit11 ebith(13)≈0.918h(\tfrac13)\approx0.918 ebit
Three-tangle τ\tau1100
Tensor rank2233 (border rank 22)
Predict before you look

What is the length of each qubit's Bloch vector in the W state?

Show the answer

0.33

ρ1=23∣0⟩ ⁣⟨0∣+13∣1⟩ ⁣⟨1∣\rho_1 = \tfrac23\dyad0+\tfrac13\dyad1, so r⃗=(0,0,23−13)\vec r = (0,0,\tfrac23-\tfrac13) and ∣r⃗∣=13\abs{\vec r} = \tfrac13. For GHZ it is 00.

SLOCC and the Dür–Vidal–Cirac theorem

For two qubits this recovers the Schmidt rank: there are just two classes, product and entangled, since any two states of Schmidt rank 2 are related by an invertible A⊗BA\otimes B. For three qubits, invertible operators preserve the local ranks (rA,rB,rC)(r_A,r_B,r_C) of the one-qubit marginals, which already separates the product class, the three biseparable classes and the genuinely entangled states with ranks (2,2,2)(2,2,2). The surprise is that (2,2,2)(2,2,2) splits in two.

DerivationProof that GHZ and W are SLOCC-inequivalent
  1. range ρBCϕ=(B⊗C) range ρBCψ\mathrm{range}\,\rho^{\phi}_{BC} = (B\otimes C)\,\mathrm{range}\,\rho^{\psi}_{BC}

    For any state, range ρBC\mathrm{range}\,\rho_{BC} is spanned by the vectors (⟨i∣⊗I)∣ψ⟩(\bra{i}\otimes I)\ket\psi. If ∣ϕ⟩=(A⊗B⊗C)∣ψ⟩\ket\phi = (A\otimes B\otimes C)\ket\psi with AA invertible, these vectors transform as follows:

  2. GHZ: span{∣00⟩,∣11⟩},W: span{∣00⟩, ∣01⟩+∣10⟩}\mathrm{GHZ}:\ \mathrm{span}\{\ket{00},\ket{11}\},\qquad W:\ \mathrm{span}\{\ket{00},\ \ket{01}+\ket{10}\}

    An invertible B⊗CB\otimes C maps product vectors to product vectors and back, so the number of product vectors (up to scalars) in range ρBC\mathrm{range}\,\rho_{BC} is an SLOCC invariant. Now compute the ranges:

  3. #prod(GHZ)=2 ≠ 1=#prod(W)\#_{\text{prod}}(\mathrm{GHZ}) = 2\ \neq\ 1 = \#_{\text{prod}}(W)

    A vector a∣00⟩+b∣11⟩a\ket{00}+b\ket{11} is a product iff its coefficient determinant abab vanishes: exactly two product vectors. A vector a∣00⟩+b(∣01⟩+∣10⟩)a\ket{00}+b(\ket{01}+\ket{10}) has determinant −b2-b^2: exactly one product vector, ∣00⟩\ket{00}.

  4. So no invertible local operators connect the two states. Non-invertible operators would lower a local rank, but both states have local ranks (2,2,2)(2,2,2). Hence neither can be converted into the other, in either direction, with any nonzero probability.

The three-tangle and monogamy

The three-tangle has a closed form: τ=4∣Det(ψ)∣\tau = 4\abs{\mathrm{Det}(\psi)}, where Det\mathrm{Det} is Cayley's hyperdeterminant of the 2×2×22\times2\times2 array of amplitudes:

Det(ψ)=ψ0002ψ1112+ψ0012ψ1102+ψ0102ψ1012+ψ1002ψ0112−2(ψ000ψ111ψ001ψ110+ψ000ψ111ψ010ψ101+ψ000ψ111ψ100ψ011+ψ001ψ110ψ010ψ101+ψ001ψ110ψ100ψ011+ψ010ψ101ψ100ψ011)+4(ψ000ψ011ψ101ψ110+ψ001ψ010ψ100ψ111)\begin{aligned}\mathrm{Det}(\psi) ={}& \psi_{000}^2\psi_{111}^2+\psi_{001}^2\psi_{110}^2+\psi_{010}^2\psi_{101}^2+\psi_{100}^2\psi_{011}^2\\ &-2\big(\psi_{000}\psi_{111}\psi_{001}\psi_{110}+\psi_{000}\psi_{111}\psi_{010}\psi_{101}+\psi_{000}\psi_{111}\psi_{100}\psi_{011}\\ &\qquad+\psi_{001}\psi_{110}\psi_{010}\psi_{101}+\psi_{001}\psi_{110}\psi_{100}\psi_{011}+\psi_{010}\psi_{101}\psi_{100}\psi_{011}\big)\\ &+4\big(\psi_{000}\psi_{011}\psi_{101}\psi_{110}+\psi_{001}\psi_{010}\psi_{100}\psi_{111}\big)\end{aligned}
Det\mathrm{Det} is invariant under SL(2,C)×3SL(2,\C)^{\times3} and is multiplied by (det⁡A det⁡B det⁡C)2(\det A\,\det B\,\det C)^2 under general invertible local operators. So whether τ\tau vanishes is an SLOCC invariant: a second proof that GHZ (τ=1\tau = 1) and W (τ=0\tau = 0) are inequivalent.
Check your understanding

In the W state, CAB=CAC=23\mathcal C_{AB} = \mathcal C_{AC} = \tfrac23. What is its three-tangle τ\tau?

  1. 11
  2. 13\tfrac13
  3. 00
  4. 49\tfrac49
Show the answer

C. 00

ρA=diag⁡(23,13)\rho_A = \diag(\tfrac23,\tfrac13), so CA∣BC2=4⋅29=89=CAB2+CAC2\mathcal C^2_{A|BC} = 4\cdot\tfrac29 = \tfrac89 = \mathcal C^2_{AB}+\mathcal C^2_{AC}. W saturates the CKW inequality: all of its entanglement sits in pairs, none of it is three-way.

DerivationW is a limit of GHZ-class states
  1. ∣ϕε⟩=1εAε⊗3(∣000⟩−∣111⟩)=∣vε⟩⊗3−∣000⟩ε\ket{\phi_\varepsilon} = \frac{1}{\varepsilon}A_\varepsilon^{\otimes3}\big(\ket{000}-\ket{111}\big) = \frac{\ket{v_\varepsilon}^{\otimes3}-\ket{000}}{\varepsilon}

    Let ∣vε⟩=∣0⟩+ε∣1⟩\ket{v_\varepsilon} = \ket0+\varepsilon\ket1. For ε≠0\varepsilon\neq0, the operator AεA_\varepsilon with Aε∣0⟩=∣vε⟩A_\varepsilon\ket0 = \ket{v_\varepsilon} and Aε∣1⟩=∣0⟩A_\varepsilon\ket1 = \ket0 is invertible, so the following state lies in the GHZ class (∣000⟩−∣111⟩\ket{000}-\ket{111} is GHZ up to a local ZZ):

  2. ∣ϕε⟩=(∣001⟩+∣010⟩+∣100⟩)+ε(∣011⟩+∣101⟩+∣110⟩)+ε2∣111⟩ → ε→0  3 ∣W⟩\ket{\phi_\varepsilon} = \big(\ket{001}+\ket{010}+\ket{100}\big)+\varepsilon\big(\ket{011}+\ket{101}+\ket{110}\big)+\varepsilon^2\ket{111}\ \xrightarrow{\ \varepsilon\to0\ }\ \sqrt3\,\ket W

    Expand the cube:

  3. So W lies in the closure of the GHZ orbit but not in the orbit: in algebraic-complexity language, the W tensor has tensor rank 3 but border rank 2. Under closure the classes form a hierarchy, GHZ above W above the biseparable classes above product.

More qubits, and certifying it in the lab

DerivationWhy there are infinitely many classes for n ≥ 4
  1. dim⁡CCP2n−1=2n−1\dim_{\C}\mathbb{CP}^{2^n-1} = 2^n-1

    SLOCC acts on states up to normalization, that is, on projective space:

  2. dim⁡C(orbit)≤4n−n=3n\dim_{\C}(\text{orbit})\le4n-n = 3n

    GL(2,C)×nGL(2,\C)^{\times n} has complex dimension 4n4n, but the nn local scalings act only through their product, which is invisible on rays. So every orbit has dimension at most 3n3n:

  3. n=4: 15−12=3,n=10: 1023−30=993n = 4:\ 15-12 = 3,\qquad n = 10:\ 1023-30 = 993

    For n=3n = 3: 7≤97\le9, consistent with finitely many orbits (there are six). For n=4n = 4: 15>1215>12, so the orbits come in continuous families with at least three complex parameters; in general at least 2n−1−3n2^n-1-3n.

  4. Verstraete, Dehaene, De Moor and Verschelde (2002) organized four-qubit states into nine families, several with continuous parameters. Beyond that, complete classification is hopeless, and research turns to invariants, entanglement polytopes and operational tasks.

In experiments one usually wants to certify genuine multipartite entanglement (GME): a state that is not a mixture of states each separable across some bipartition. Fidelity witnesses do this with few measurement settings.

Key takeaways

  • SLOCC classes are orbits of invertible local operators; for two parties they are labelled by Schmidt rank alone.
  • Three qubits have six SLOCC classes. GHZ (generic, τ>0\tau>0) and W (τ=0\tau = 0, measure zero, in the closure of GHZ) are the two genuinely tripartite ones.
  • CKW monogamy: CAB2+CAC2≤4det⁡ρA\mathcal C^2_{AB}+\mathcal C^2_{AC}\le4\det\rho_A. GHZ puts all of its entanglement into the three-tangle, W puts all of it into pairs.
  • For n≥4n\ge4 there are infinitely many classes; experiments certify genuine multipartite entanglement with witnesses such as GHZ fidelity above 12\tfrac12.

Check yourself

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

Question 1

Why can GHZ not be converted into W by SLOCC, even with tiny probability?

  1. GHZ has more entanglement entropy across each cut than W
  2. SLOCC can only reduce the number of qubits
  3. An invariant of invertible local operators differs: the number of product vectors in the range of a two-qubit marginal (equivalently, whether τ=0\tau = 0)
  4. W is not genuinely tripartite entangled
Show the answer

C. An invariant of invertible local operators differs: the number of product vectors in the range of a two-qubit marginal (equivalently, whether τ=0\tau = 0)

Non-invertible operators would lower a local rank, and invertible ones preserve both the count of product vectors in range ρBC\mathrm{range}\,\rho_{BC} (2 for GHZ, 1 for W) and the vanishing of the hyperdeterminant.

Question 2

Tracing out one qubit of ∣W⟩\ket W leaves the other two:

  1. entangled, with concurrence 23\tfrac23
  2. in a separable mixture
  3. in a Bell state
  4. maximally mixed
Show the answer

A. entangled, with concurrence 23\tfrac23

ρAB=13∣00⟩ ⁣⟨00∣+23∣Ψ+⟩ ⁣⟨Ψ+∣\rho_{AB} = \tfrac13\dyad{00}+\tfrac23\dyad{\Psi^+}, an X-shaped state with C=2max⁡(0, ∣ρ01,10∣−ρ00,00 ρ11,11)=2⋅13=23\mathcal C = 2\max\big(0,\ \abs{\rho_{01,10}}-\sqrt{\rho_{00,00}\,\rho_{11,11}}\big) = 2\cdot\tfrac13 = \tfrac23.

Question 3

A lab prepares a noisy 10-qubit GHZ state and measures its fidelity FF with the ideal state. Which condition certifies genuine 10-partite entanglement?

  1. Any F>110F>\tfrac1{10}
  2. F>12F>\tfrac12
  3. F>910F>\tfrac9{10}
  4. Only F=1F = 1
Show the answer

B. F>12F>\tfrac12

Every biseparable state has GHZ fidelity at most 12\tfrac12, the largest squared Schmidt coefficient of GHZ across any cut. Anything above that certifies GME.

Question 4

How many SLOCC classes do pure four-qubit states form?

  1. 6, as for three qubits
  2. 9
  3. 24=162^4 = 16
  4. Infinitely many
Show the answer

D. Infinitely many

Orbits have dimension at most 1212 in a 1515-dimensional projective space, so the classes come in continuous families. The "nine families" of Verstraete et al. are families, several depending on continuous parameters, not nine classes.

Further reading

  • W. Dür, G. Vidal & J. I. Cirac, "Three qubits can be entangled in two inequivalent ways", Phys. Rev. A 62, 062314 (2000).
  • V. Coffman, J. Kundu & W. K. Wootters, "Distributed entanglement", Phys. Rev. A 61, 052306 (2000).
  • F. Verstraete, J. Dehaene, B. De Moor & H. Verschelde, "Four qubits can be entangled in nine different ways", Phys. Rev. A 65, 052112 (2002).
  • O. Gühne & G. Tóth, "Entanglement detection", Phys. Rep. 474, 1 (2009).
  • M. Walter, B. Doran, D. Gross & M. Christandl, "Entanglement polytopes: multiparticle entanglement from single-particle information", Science 340, 1205 (2013).