Entanglement Witnesses

An entanglement witness is an observable whose expectation value proves that a state is entangled. It is designed so that all separable states give nonnegative expectation value, while at least one entangled state gives a negative value.

With the sign convention used here, a Hermitian operator $W$ is an entanglement witness if

$$\operatorname{Tr}(W\sigma)\ge0 \quad \text{for every separable state }\sigma,$$

but there exists an entangled state $\rho$ such that

$$\operatorname{Tr}(W\rho)<0.$$

Other sign conventions are common. Some authors define witnesses with separable states below a bound and entangled states above it. The physics is the same: a measured expectation value crosses a threshold that no separable state can cross.

Helpful background. Separable mixtures gives the set whose bound must be proved; expectation values translates an observable into a measured average. We work in finite-dimensional subsystem spaces.

Entanglement detection by an observable

Separability is a global property of a density operator. Checking it exactly is hard in general. Entanglement witnesses trade completeness for practicality:

• a negative witness expectation proves entanglement;
• a nonnegative value for one witness does not prove separability;
• the witness can often be written in terms of experimentally accessible observables.

Thus witnesses are detection tools. They answer the question “Can this particular observable certify entanglement for this family of states?” rather than “Have we solved separability in full generality?”

Geometry of the Definition

In finite dimensions, the set of density operators is convex. The subset of separable states is compact and convex:

$$\sigma = \sum_k p_k\, \sigma_A^{(k)}\otimes\sigma_B^{(k)}.$$

If $\sigma_1$ and $\sigma_2$ are separable, then

$$q\sigma_1+(1-q)\sigma_2$$

is separable for $0\le q\le1$. Geometrically, mixtures stay inside the separable set.

An entangled state lies outside that convex set. In finite dimensions, separating-hyperplane theorems imply that for every entangled state there exists a Hermitian operator $W$ such that the affine functional

$$\rho \longmapsto \operatorname{Tr}(W\rho)$$

is nonnegative on all separable states but negative on that entangled state.

This is the conceptual reason witnesses exist. The practical difficulty is finding a useful $W$ and measuring it with sufficient precision.

Bell-State Witness

For the Bell state

$$\lvert\Phi^+\rangle = \frac{1}{\sqrt2} \left( \lvert00\rangle+\lvert11\rangle \right),$$

define

$$W_{\Phi^+} = \frac12 I - \lvert\Phi^+\rangle\langle\Phi^+\rvert.$$

For a product pure state $\lvert a\rangle\otimes\lvert b\rangle$,

$$\left\lvert \langle\Phi^+\vert a,b\rangle \right\rvert^2 \le \frac12.$$

Therefore

$$\langle a,b\vert W_{\Phi^+}\vert a,b\rangle = \frac12 - \left\lvert \langle\Phi^+\vert a,b\rangle \right\rvert^2 \ge0.$$

By convexity, the same nonnegative bound holds for every separable mixed state.

On the Bell state itself,

$$\operatorname{Tr} \left( W_{\Phi^+} \lvert\Phi^+\rangle\langle\Phi^+\rvert \right) = \frac12-1 = -\frac12.$$

Thus $W_{\Phi^+}$ detects $\lvert\Phi^+\rangle$ as entangled.

The same construction works for any entangled pure target $|\psi\rangle$ with largest Schmidt coefficient $s_{\max}$:

$$W_\psi=s_{\max}^2 I-|\psi\rangle\langle\psi|.$$

The largest squared overlap with a normalized product vector is $s_{\max}^2$, by the operator norm of the coefficient matrix. Hence a measured target overlap $F_\psi=\langle\psi|\rho|\psi\rangle>s_{\max}^2$ certifies entanglement. A target close to a product state has $s_{\max}^2$ close to one, so this particular witness demands a very high overlap.

Pauli-Correlator Form

The Bell projector has the Pauli expansion

$$\lvert\Phi^+\rangle\langle\Phi^+\rvert = \frac14 \left( I\otimes I + \sigma_x\otimes\sigma_x - \sigma_y\otimes\sigma_y + \sigma_z\otimes\sigma_z \right).$$

Therefore the witness can be written

$$W_{\Phi^+} = \frac14 \left( I\otimes I - \sigma_x\otimes\sigma_x + \sigma_y\otimes\sigma_y - \sigma_z\otimes\sigma_z \right).$$

Its expectation value is

$$\langle W_{\Phi^+}\rangle = \frac14 \left( 1 - \langle\sigma_x\otimes\sigma_x\rangle + \langle\sigma_y\otimes\sigma_y\rangle - \langle\sigma_z\otimes\sigma_z\rangle \right).$$

This form shows why witnesses are experimentally useful: one may certify entanglement by measuring a small set of correlations rather than reconstructing the full density matrix.

Bell-Mixed Example

Consider

$$\rho(p) = p\lvert\Phi^+\rangle\langle\Phi^+\rvert + \frac{1-p}{4}I, \qquad 0\le p\le1.$$

Then

$$\operatorname{Tr}(W_{\Phi^+}\rho(p)) = \frac12 - \operatorname{Tr} \left( \lvert\Phi^+\rangle\langle\Phi^+\rvert \rho(p) \right).$$

The Bell-state fidelity is

$$F_{\Phi^+} = \operatorname{Tr} \left( \lvert\Phi^+\rangle\langle\Phi^+\rvert \rho(p) \right) = p+\frac{1-p}{4} = \frac{1+3p}{4}.$$

Thus

$$\operatorname{Tr}(W_{\Phi^+}\rho(p)) = \frac{1-3p}{4}.$$

The witness detects entanglement when

$$p>\frac13.$$

For this symmetric two-qubit family, that threshold matches the PPT and concurrence thresholds. For a general state, one witness detects only the states it was designed to detect.

Building a witness from a negative partial-transpose eigenvector

Suppose $\rho^{T_B}|v\rangle=\lambda|v\rangle$ with $\lambda<0$. Define

$$W=(|v\rangle\langle v|)^{T_B}.$$

Partial transpose is self-adjoint with respect to the trace pairing, so

$$\operatorname{Tr}(W\rho)=\langle v|\rho^{T_B}|v\rangle=\lambda<0.$$

For every separable $\sigma$, $\sigma^{T_B}\ge0$, so the same expression is nonnegative. This proves that $W$ is a witness. For $\rho=|\Phi^+\rangle\langle\Phi^+|$, the negative eigenvector is $|\Psi^-\rangle$, and this construction gives $W_{\Phi^+}$ above.

Such witnesses cannot detect PPT entanglement: their expectation on any PPT state is nonnegative. More general positive-map constructions can do so, but finding a useful separable bound remains the essential step.

From correlators to a justified conclusion

The Pauli witness can be estimated using local $X$, $Y$, and $Z$ measurements on separately prepared copies of the same two-qubit state. Each correlator is the mean product of the two $\pm1$ outcomes. This assumes the subsystem dimensions and the measured observables are calibrated as stated. It is a trusted-observable test; a Bell test uses different assumptions.

Uncertainty must be small enough to exclude the separable threshold. If simultaneous valid error bounds for the three correlators are $\delta_x,\delta_y,\delta_z$, then

$$|\widehat W-\langle W\rangle|\le(\delta_x+\delta_y+\delta_z)/4.$$

An upper bound on $\langle W\rangle$ below zero certifies entanglement at the confidence level of those simultaneous bounds. Separate one-standard-deviation errors are not automatically such bounds. Entanglement certification explains finite-data and calibration uncertainties in more detail.

Witnesses can also use graph-state stabilizers, collective spin, or optical quadratures. The threshold must be proved for the intended class of states; for example, excluding fully separable multipartite states is weaker than excluding all biseparable states.

A witness can miss an entangled state

$W_{\Phi^+}$ has expectation $+1/2$ on the orthogonal Bell state $|\Psi^-\rangle$, even though the latter is maximally entangled. Rotating or redesigning the witness can detect it. A nonnegative result therefore means this test is inconclusive, not that the state is separable.

References

• O. Gühne and G. Tóth, “Entanglement detection”, Physics Reports 474, 1–75 (2009).
• M. Horodecki, P. Horodecki, and R. Horodecki, “Separability of mixed states: necessary and sufficient conditions”, Physics Letters A 223, 1–8 (1996), DOI: 10.1016/S0375-9601(96)00706-2.
• M. Lewenstein, B. Kraus, J. I. Cirac, and P. Horodecki, “Optimization of entanglement witnesses”, Physical Review A 62, 052310 (2000).

Exercises

1. Product-state bound. Let

$$\lvert a\rangle = a_0\lvert0\rangle+a_1\lvert1\rangle, \qquad \lvert b\rangle = b_0\lvert0\rangle+b_1\lvert1\rangle$$

be normalized qubit states. Prove

$$\left\lvert \langle\Phi^+\vert a,b\rangle \right\rvert^2 \le \frac12.$$

Solution

The overlap is

$$\langle\Phi^+\vert a,b\rangle = \frac{1}{\sqrt2} \left( a_0b_0+a_1b_1 \right)$$

exactly with the ket coefficients defined above. By Cauchy–Schwarz applied to the vectors $a^*$ and $b$,

$$\lvert a_0b_0+a_1b_1\rvert^2 \le \left( \lvert a_0\rvert^2+\lvert a_1\rvert^2 \right) \left( \lvert b_0\rvert^2+\lvert b_1\rvert^2 \right) =1.$$

Therefore the squared Bell overlap is at most $1/2$.

2. Witness expectation for a Bell-mixed state. For

$$\rho(p) = p\lvert\Phi^+\rangle\langle\Phi^+\rvert + \frac{1-p}{4}I,$$

compute $\operatorname{Tr}(W_{\Phi^+}\rho(p))$.

Solution

The Bell fidelity is

$$F_{\Phi^+} = p+\frac{1-p}{4} = \frac{1+3p}{4}.$$

Since $W_{\Phi^+}=I/2-\lvert\Phi^+\rangle\langle\Phi^+\rvert$,

$$\operatorname{Tr}(W_{\Phi^+}\rho(p)) = \frac12-F_{\Phi^+} = \frac{1-3p}{4}.$$

The witness is negative exactly when $p>1/3$.

3. Pauli measurements. Suppose the measured correlators are

$$\langle\sigma_x\otimes\sigma_x\rangle=0.82, \qquad \langle\sigma_y\otimes\sigma_y\rangle=-0.76, \qquad \langle\sigma_z\otimes\sigma_z\rangle=0.80.$$

Assume simultaneous absolute error bounds of $0.03$ on each correlator. Does $W_{\Phi^+}$ detect entanglement even with those errors?

Solution

Use

$$\langle W_{\Phi^+}\rangle = \frac14 \left( 1 - \langle\sigma_x\otimes\sigma_x\rangle + \langle\sigma_y\otimes\sigma_y\rangle - \langle\sigma_z\otimes\sigma_z\rangle \right).$$

Substitution gives

$$\langle W_{\Phi^+}\rangle = \frac14 \left( 1-0.82-0.76-0.80 \right) = -0.345.$$

The absolute error bound is $(0.03+0.03+0.03)/4=0.0225$. Even the upper bound $-0.345+0.0225=-0.3225$ is negative, so the witness detects entanglement under the stated bounds and measurement assumptions.

4. Failure to detect is not separability. Explain why $\operatorname{Tr}(W\rho)\ge0$ for one witness $W$ does not prove that $\rho$ is separable.

Solution

A witness defines one separating test. It is guaranteed to be nonnegative on all separable states, but it need not be negative on every entangled state. Geometrically, one hyperplane can separate some outside points from the separable set while leaving other outside points on the nonnegative side. To prove separability one needs a complete criterion, not a single failed witness.