Entanglement witnesses and separability

For specified factors A and B, a separable density operator is a convex sum of product density operators. The local factors and their probabilities may vary from term to term. This permits ordinary shared randomness: a source may choose a classical label and distribute corresponding local states. A product state is a special case with no correlation. Entanglement means that no such product-mixture representation exists, not simply that one chosen decomposition happens to contain entangled vectors.

For a pure joint state, the Schmidt rank gives a complete answer: rank one is product and greater rank is entangled. For mixed states, local entropy and correlation in a single basis are insufficient. A classical zero-zero/one-one mixture has strong agreement but is separable. Testing a witness instead specifies an observable on the joint space and proves a bound for every allowed classical mixture. The proof must cover pure product states and then extend by linearity to all mixtures.

Take the Bell vector formed by equal positive zero-zero and one-one amplitudes. Its squared overlap with any normalized product vector is at most one half. For local amplitudes $a_i,b_i$, the overlap is $|a_0b_0+a_1b_1|^2/2\le(\sum_i|a_i|^2)(\sum_i|b_i|^2)/2=1/2$, by Cauchy–Schwarz applied to the conjugated first vector and the second vector. Convexity extends the same overlap bound to every separable mixed state. Subtracting the Bell projector from half the identity therefore produces a valid witness.

The witness can also be written as a combination of identity and three Pauli products. Each product expectation is measured using a corresponding pair of local bases on independently prepared copies, then combined with the fixed signs. These are correlations, not merely products of independently estimated local means. The minus sign associated with the y-product in the Bell projector is essential. Ignoring it can change the certificate and falsely classify a state.

A nonnegative expectation does not generally prove separability because a witness is aimed at particular directions outside the convex separable set. An entangled state orthogonal to the targeted Bell vector can escape this test. Rotating the witness or using other criteria can expose different states. For a finite data set, the experimentally estimated expectation needs an uncertainty treatment and trusted measurement calibration; a tiny negative point estimate alone is not a confidence statement.

Partial transposition offers another criterion. Transposing only B maps every separable state to a positive operator, because transposition preserves positivity on each local factor. A negative eigenvalue therefore certifies entanglement. Positivity of the partial transpose is sufficient for separability in two-by-two and two-by-three dimensions, but not in arbitrary dimensions. These dimension restrictions matter: one cannot promote an accessible two-qubit test into a universal characterization of all multipartite or high-dimensional entanglement.