W States: Entanglement That Can Survive Losing a Qubit

W States: Entanglement That Can Survive Losing a Qubit

Discard one qubit from a W state and the other two remain entangled, but not as a perfect Bell pair. Trace out the missing qubit and verify the surviving state in NumPy.

A three-qubit state can be genuinely entangled without storing all its useful correlations in an all-or-nothing package. The W state gives a particularly clean example: discard any one qubit, and the other two are still entangled.

That does not mean they are left in a perfect Bell pair. The surviving state is mixed. Understanding that distinction is the interesting part.

One excitation, shared coherently

Label the qubits A, B, and C, with basis strings written in that order. The three-qubit W state is

|W> = (|001> + |010> + |100>)/sqrt(3).

Every computational-basis outcome contains exactly one 1. Each of the three strings appears with probability 1/3.

Those probabilities alone do not define the state. A classical source that randomly prepares one of the same three strings has the same computational-basis histogram, but no entanglement. The W state also has coherence between the three alternatives. It is a sum of amplitudes, not a lottery over product states.

The contrast with GHZ states and the survival of pairwise entanglement were established in Dür, Vidal, and Cirac's original analysis of three-qubit entanglement. We can see the mechanism directly, without needing the full classification theory.

Losing C means tracing it out

Here, losing a qubit means that C becomes inaccessible and no measurement result from it is available. Mathematically, we take a partial trace. We are not selecting a favorable measurement outcome, and we are not modeling arbitrary noise on A and B.

Group the W-state terms by the value of C:

|W> = sqrt(2/3)|Psi+>_AB |0>_C + sqrt(1/3)|00>_AB |1>_C,

where |Psi+> = (|01> + |10>)/sqrt(2).

The two C states are orthogonal. When we discard C, cross terms connecting its 0 and 1 branches vanish. The coherence between |01> and |10> remains, because those two alternatives share the same C state.

The result is

rho_AB = (2/3)|Psi+><Psi+| + (1/3)|00><00|.

This is exactly the reduced state given in Section IV of the original paper. Its Bell component and vacuum component form a mixture, not a coherent superposition. The numbers 2/3 and 1/3 are mixture weights, not the probability that a loss event happens.

In the basis |00>, |01>, |10>, |11>, the matrix is

rho_AB = (1/3) * [[1, 0, 0, 0],
                  [0, 1, 1, 0],
                  [0, 1, 1, 0],
                  [0, 0, 0, 0]]

It has purity 5/9, so it is not a pure Bell state. By permutation symmetry, discarding A or B instead leaves the same kind of state on the remaining pair.

Prove the surviving pair is entangled

A mixture containing an entangled component is not automatically entangled. Other components can wash the entanglement out. We need an actual test.

One convenient test is partial transposition. For matrix entries indexed by the two qubits, transposing only B means

(rho_AB^TB)[a,b,a',b'] = rho_AB[a,b',a',b].

Every separable state remains positive semidefinite under this operation. Therefore a negative eigenvalue certifies entanglement. This is the Peres separability criterion. Partial transposition is a mathematical diagnostic, not a physical gate we propose applying to an unknown state.

For the surviving W-state pair, the eigenvalues are

(1 - sqrt(5))/6, 1/3, 1/3, (1 + sqrt(5))/6.

The first is approximately -0.206011. That negative value proves the pair is entangled even though it is mixed.

For this conclusion, we only need the implication from a negative eigenvalue to entanglement. We are not assuming that a nonnegative partial transpose settles separability in arbitrary dimensions.

Reproduce the loss in NumPy

The code below constructs both W and GHZ states, discards the final qubit, and checks the surviving pair. No random measurement sampling is needed.

The reshape convention matters. With strings ordered A, B, C, the row index AB has dimension 4 and the column index C has dimension 2. Multiplying the reshaped statevector by its conjugate transpose sums over C.

import numpy as np

def projector(ket):
    return np.outer(ket, ket.conj())

def discard_C(ket):
    # Basis order: |ABC>, with C the fastest-changing index.
    amplitudes = ket.reshape(4, 2)
    return amplitudes @ amplitudes.conj().T

def partial_transpose_B(rho):
    # Axes initially mean a, b, a_prime, b_prime.
    return rho.reshape(2, 2, 2, 2).transpose(0, 3, 2, 1).reshape(4, 4)

W = np.zeros(8, dtype=complex)
W[[1, 2, 4]] = 1 / np.sqrt(3)
GHZ = np.zeros(8, dtype=complex)
GHZ[[0, 7]] = 1 / np.sqrt(2)
psi_plus = np.array([0, 1, 1, 0], dtype=complex) / np.sqrt(2)
zero_zero = np.array([1, 0, 0, 0], dtype=complex)
rho_W, rho_GHZ = discard_C(W), discard_C(GHZ)
expected = (2/3) * projector(psi_plus) + (1/3) * projector(zero_zero)

assert np.allclose(rho_W, expected)
assert np.isclose(np.trace(rho_W), 1)
assert np.linalg.eigvalsh(rho_W).min() >= -1e-12
for name, rho in [("W", rho_W), ("GHZ", rho_GHZ)]:
    minimum = np.linalg.eigvalsh(partial_transpose_B(rho)).min()
    print(f"{name} pair, minimum PT eigenvalue: {minimum:.6f}")
print(f"W pair purity: {np.trace(rho_W @ rho_W).real:.6f}")

# What if C was measured in Z and its outcome was retained?
branches = W.reshape(4, 2)
for c in [0, 1]:
    branch = branches[:, c]
    probability = np.vdot(branch, branch).real
    conditional = projector(branch) / probability
    bell_overlap = np.vdot(psi_plus, conditional @ psi_plus).real
    print(f"C={c}: probability={probability:.6f}, Bell overlap={bell_overlap:.6f}")

The output is

W pair, minimum PT eigenvalue: -0.206011
GHZ pair, minimum PT eigenvalue: 0.000000
W pair purity: 0.555556
C=0: probability=0.666667, Bell overlap=1.000000
C=1: probability=0.333333, Bell overlap=0.000000

The last two lines describe an alternative experiment where C is accessible, measured in Z, and its outcome retained. They do not turn an inaccessible qubit into a heralding signal.

Why GHZ behaves differently

For comparison,

|GHZ> = (|000> + |111>)/sqrt(2).

Discarding C gives

rho_AB = (1/2)|00><00| + (1/2)|11><11|.

The two surviving qubits have perfectly correlated Z outcomes, but this state is explicitly a mixture of product states. It is separable. The GHZ coherence connected branches with different C values, so it disappears from the reduced pair when C is traced out.

This does not make GHZ entanglement useless or mean it cannot yield a Bell pair. If C is available, measuring it in the X basis and communicating the result prepares A and B in one of two Bell states; a conditional local phase correction can make the output fixed. That protocol uses information which is unavailable in the discard scenario.

Likewise, a Z measurement on C in the W state yields a perfect |Psi+> pair for outcome 0, with probability 2/3. Outcome 1 leaves |00>. Forgetting that outcome reproduces the mixed state above. Keeping an outcome and throwing a qubit away are different operations on the information available to us.

What the robustness does and does not buy

The W state retains some pairwise entanglement after any one qubit is discarded. That is a specific and useful robustness property, not blanket protection against noise. It does not preserve an arbitrary encoded logical qubit, and it does not guarantee a deterministic perfect Bell pair after loss.

There is also a limit to the slogan: discard two qubits from this three-qubit state and only one qubit remains, so there is no pair left to share entanglement.

The practical lesson is to inspect the reduced state, not just the full register's entangled label or its measurement histogram. For W and GHZ states, the same operation of losing one participant exposes very different resources in the surviving pair.