Entanglement witness — a Hermitian W with Tr(Wσ)≥0 on all separable states but Tr(Wρ)<0 on some entangled ρ

Statement

Definition. An entanglement witness for an entangled state $\rho$ on $\mathcal H_A\otimes\mathcal H_B$ is a Hermitian operator $W=W^\dagger$ such that $$ \operatorname{Tr}(W\sigma)\ge0\ \ \text{for every separable state }\sigma,\qquad\text{but}\qquad \operatorname{Tr}(W\rho)<0. $$ The hyperplane $\{X:\operatorname{Tr}(WX)=0\}$ separates $\rho$ from the convex body of separable states; the negative expectation value $\operatorname{Tr}(W\rho)<0$ certifies that $\rho$ is entangled. (Source: en.wikipedia.org/wiki/Entanglement_witness, direct fetch.)

Existence (Hahn–Banach separation). The set $\mathcal S$ of separable states is closed and convex. If $\rho\notin\mathcal S$, the Hahn–Banach separation theorem (finite-dimensional geometric form) yields a real linear functional strictly separating $\rho$ from $\mathcal S$; since all states are Hermitian, that functional is $X\mapsto\operatorname{Tr}(WX)$ for some Hermitian $W$. Thus: $$ \rho \text{ entangled} \iff \exists\,\text{Hermitian }W:\ \operatorname{Tr}(W\rho)<0\ \text{and}\ \operatorname{Tr}(W\sigma)\ge0\ \forall\sigma\in\mathcal S. $$ This is a complete criterion — *every* entangled state has a witness — the practical converse of "separable states are hard to certify" being "entangled states always admit a cheap linear certificate." (Source: Wikipedia — "For every entangled state ρ, there exists a Hermitian operator A with Tr(Aρ)<0 and Tr(Aσ)≥0 for all separable σ"; Horodecki⁴ review arXiv:quant-ph/0702225.)

Block-positivity / dual to positive-but-not-CP maps. The witnesses are exactly the block-positive operators that are not PSD: $W$ is a witness iff $\langle a\otimes b|W|a\otimes b\rangle\ge0$ for all product vectors $|a\otimes b\rangle$ (so $\operatorname{Tr}(W\sigma)\ge0$ on separable $\sigma$) while $W\not\ge0$ (so some entangled $\rho$ dips negative). Under the Choi–Jamiołkowski isomorphism — state–channel duality: a map is CP iff its Choi matrix is PSD, block-positive operators correspond to positive maps, and non-PSD ones to positive maps that are not completely positive. This is the Horodecki criterion: $\rho$ is separable iff $(\mathrm{id}\otimes\Lambda)(\rho)\ge0$ for every positive map $\Lambda$, and each positive-but-not-CP $\Lambda$ gives, via its (partially transposed) Choi operator, a witness. The transpose map $\Lambda=T$ recovers the Peres–Horodecki (PPT) criterion — partial transpose positivity is necessary for separability, and sufficient only in 2×2 and 2×3 (PPT test) as the canonical special case. (Source: Wikipedia entanglement-witness "connection to positive maps"; Horodecki criterion.)

Facts

- **Its role in PPT² conjecture (Christandl 2012) — composition of two PPT channels is entanglement-breaking; open for dimension ≥ 4: the cheap refutation certificate. A PPT-squared counterexample** is a pair of PPT maps $\Phi_1,\Phi_2$ whose composition is *not* entanglement-breaking — equivalently, whose composite Choi matrix $J(\Phi_2\circ\Phi_1)$ is entangled (Entanglement-breaking channel — measure-and-prepare maps: separable Choi matrix ⇔ destroys all entanglement). To *prove* separability is the genuinely hard "separability problem"; to *disprove* it you only need to exhibit one witness $W$ with $\operatorname{Tr}\!\big(W\,J(\Phi_2\circ\Phi_1)\big)<0$. That is a single inner product — closed-form or one SDP — so verification of a candidate counterexample is computationally trivial even though *finding* the candidate is believed intractable (Jin 2020, arXiv:2010.15554: the direct search is "the genuine part of the separability problem"). - Witnesses turn an existence claim into a finite, checkable object. Because a witness is one Hermitian matrix, a claimed PPT² counterexample is fully auditable: publish $\Phi_1,\Phi_2$ (each PPT-checkable by two eigenvalue tests on their Choi matrices, see PPT map — a channel that is CP and completely co-positive: Choi matrix PSD with positive partial transpose) and the witness $W$; a verifier confirms (a) $J(\Phi_i)$ PPT, (b) $W$ block-positive, (c) $\operatorname{Tr}(WJ(\Phi_2\circ\Phi_1))<0$. No large computation, no trust required — the asymmetry "hard to find, cheap to check" is exactly the RLVR-style verification profile that makes this problem attractive. - Optimal / finer witnesses and the PPT ceiling. Witnesses arising from the transpose map (PPT criterion) detect only NPT (non-PPT) entanglement. Bound-entangled states (Bound entanglement — entangled states that are PPT, hence undistillable; the rare structured regime where a PPT² counterexample must live) are PPT-yet-entangled, so no transpose-based witness sees them — detecting them needs a witness from a genuinely non-decomposable positive map (one not of the form $P+Q\circ T$ with $P,Q$ CP). Since a PPT² counterexample's composite Choi matrix would be PPT (composition of PPT maps stays PPT) but entangled, any witness certifying it must be non-decomposable — this is precisely why the counterexample search is hard and why brute PPT-based screening cannot find it. - Geometric picture. Each witness $W$ defines a half-space $\{\rho:\operatorname{Tr}(W\rho)\ge0\}$ containing all separable states; the intersection of all such half-spaces reconstructs the separable body $\mathcal S$ exactly (a supporting-hyperplane / dual-cone description). "Optimal" witnesses touch $\partial\mathcal S$ along a face, detecting the largest set of entangled states; finding an optimal non-decomposable witness for a target bound-entangled region is itself a convex-optimization problem. - Measurable in the lab. Since $W$ is Hermitian it is an observable: $\operatorname{Tr}(W\rho)$ is estimated by decomposing $W$ into local measurements, so witnesses detect entanglement experimentally without full tomography — the original motivation (Terhal 2000; Gühne–Tóth review, Phys. Rep. 474 (2009) 1).

Related

- PPT² conjecture (Christandl 2012) — composition of two PPT channels is entanglement-breaking; open for dimension ≥ 4 — a witness is the cheap, auditable certificate that refutes a candidate: it detects entanglement in a composite Choi matrix, disproving the "entanglement-breaking" conclusion. - Entanglement-breaking channel — measure-and-prepare maps: separable Choi matrix ⇔ destroys all entanglement — EB ⇔ separable Choi matrix; a witness against that Choi matrix is exactly what breaks the EB claim. - Choi–Jamiołkowski isomorphism — state–channel duality: a map is CP iff its Choi matrix is PSD — the dual bridge: witnesses (block-positive, non-PSD operators) ↔ positive-but-not-completely-positive maps. - Peres–Horodecki (PPT) criterion — partial transpose positivity is necessary for separability, and sufficient only in 2×2 and 2×3 — the special case where the witness comes from the transpose map; detects all NPT but no bound entanglement. - Bound entanglement — entangled states that are PPT, hence undistillable; the rare structured regime where a PPT² counterexample must live — PPT-yet-entangled states; detecting them (and any PPT² counterexample) requires a non-decomposable witness, not a transpose-based one. - PPT map — a channel that is CP and completely co-positive: Choi matrix PSD with positive partial transpose — the hypothesis-class objects whose composite Choi matrix a witness would be aimed at.