Separable versus Entangled States

The deepest structural question you can ask about a shared quantum state is embarrassingly simple to state and notoriously hard to answer: could two distant parties have built it with nothing but a telephone? Werner's 1989 definition turns that question into geometry. A bipartite state $\rho_{AB}$ is separable if it is a convex mixture of product states, $\rho_{AB}=\sum_i p_i\,\rho_i^A\otimes\rho_i^B$ — exactly the states preparable by local operations and classical communication (LOCC), whose only correlations are the classical kind carried by a shared random label $i$ . Everything else is entangled. The definition is a negation — "not separable" — and that innocent-looking not is where all the difficulty lives. It is the boundary between correlations a coin flip can fake and correlations that demand a genuine quantum channel between the labs. This dividing line organizes an enormous amount of physics. Entanglement is the resource that powers teleportation, superdense coding, device-independent cryptography, and quantum speedups; separable states are the "free" states, useless as a teleportation channel. For pure states the frontier is fully charted — the Schmidt decomposition reduces separability to a single number, the Schmidt rank. For mixed states it turns forbidding: deciding whether a generic $\rho$ is separable is NP-hard, and the separable set has a curved, high-dimensional convex boundary no finite catalogue of tests can trace. Yet that convexity is a gift. Because the separable set is closed and convex, every entangled state can be sliced off by a flat plane — an entanglement witness — and because separability survives transposing one party (a positive but not completely positive map), a single negative eigenvalue of the partial transpose betrays entanglement. This article assembles the definition, the pure-state criterion, the convex geometry, and the two workhorse detectors, and marks the line between entanglement and the rarer property of Bell nonlocality.

Key takeaways

A bipartite state is separable if it is a convex mixture of product states, $\rho_{AB}=\sum_i p_i\,\rho_i^A\otimes\rho_i^B$ ; entangled means exactly "not of this form."
Separable states are precisely those preparable by LOCC; their correlations are classical, indexed by a shared random variable — no quantum channel required.
Pure states are easy: $|\psi\rangle_{AB}$ is separable iff its Schmidt rank is $1$ , equivalently iff its reduced state is pure, equivalently iff the entanglement entropy vanishes.
Mixed states are hard: deciding separability is NP-hard in general; in practice one uses the PPT criterion and entanglement witnesses, which exist because the separable set is closed and convex. The canonical test-bed, the two-qubit Werner state, is separable for $p\le\tfrac13$ and entangled for $p>\tfrac13$ .
Entanglement is necessary but not sufficient for Bell nonlocality: some entangled mixed states admit a local hidden-variable model, so entangled $\neq$ nonlocal.
Key idea: Separable = convex mixture of product states = preparable by LOCC = only classical correlations; entangled is the complement. The separable set is closed and convex, which is exactly why hyperplane separation (witnesses) and positive-but-not-completely-positive maps (PPT) can detect entanglement.

The core insight

Separable = convex mixture of product states = preparable by LOCC = carries only classical correlations, and entangled is everything else. The separable set is closed and convex, which is precisely why a separating hyperplane (an entanglement witness) exists for every entangled state and why a positive-but-not-completely-positive map (the partial transpose) can expose entanglement through a negative eigenvalue.

Where this lives today

The separable–entangled boundary is not a piece of abstract convex geometry that stays on the page — it is measured, daily, in every quantum-technology laboratory. Because full tomography scales exponentially, experimentalists certify entanglement in photonic, trapped-ion, and superconducting platforms by measuring a single tailored witness $W$ and reading off $\langle W\rangle<0$ with only a handful of local settings; GHZ- and W-state witnesses of this kind now certify entanglement across tens of qubits. The Werner threshold $p=1/3$ (and its cousins for other noise models) sets concrete hardware targets: a Bell pair must emerge robustly outside SEP to be useful, so fidelity budgets are written directly against these boundaries. In dimensions $\ge 3\times3$ the frontier splits again — bound-entangled states are PPT yet entangled, undistillable into Bell pairs by LOCC, with real consequences for the achievable key rates in entanglement-based quantum key distribution. And the deepest open problem is exactly the shape of the set in the figure: deciding membership in SEP is NP-hard, so despite the PPT test, entanglement witnesses, symmetric-extension hierarchies (the Doherty–Parrilo–Spedalieri semidefinite programs), and decades of effort, the separable body's high-dimensional geometry remains only partly charted — a rare case where a foundational definition still hides genuinely hard mathematics.

The physical picture

Picture two laboratories, Alice's and Bob's, joined only by a telephone line — classical bits, no qubits. They share a random seed $i$ drawn with probability $p_i$ ; conditioned on the seed, Alice builds a local state $\rho_i^A$ and Bob builds $\rho_i^B$ , independently, in their own labs. Average over the seed and you get $\rho_{AB}=\sum_i p_i\,\rho_i^A\otimes\rho_i^B$ . Every state buildable this way is separable, and its correlations are exactly as classical as the shared coin — Alice and Bob know $i$ , an eavesdropper reading only $\rho_{AB}$ does not, but nothing quantum is being shared. Entanglement is defined by what the telephone cannot build: it requires an actual quantum channel connecting the labs at some point in the past. That operational story has a clean geometric shadow. Density operators form a convex body; the pure product states $|\alpha\rangle\langle\alpha|\otimes|\beta\rangle\langle\beta|$ are its extreme "corners," and the separable states are their convex hull — a smaller convex body sitting inside the full state space. Entangled states are the points that fall outside this inner body but still inside the outer one. The maximally mixed state $I/(d_Ad_B)$ sits deep in the interior of the separable set, wrapped in a whole ball of separable states, while the maximally entangled Bell states sit as far outside as geometry allows. The first figure is this picture in the abstract, together with the two tools convexity hands us for free.

Figure

The convex body of all states with the separable set SEP nested inside. Product states are the corners of SEP, the maximally mixed state I/4 sits deep in its interior, and a maximally entangled Bell state lies far outside. Because SEP is closed and convex, any entangled ρ is cut off by a flat plane — an entanglement witness W with tr(Wσ) ≥ 0 for every separable σ but negative tr(Wρ) on ρ

Two features of that picture do the real work. First, because SEP is closed and convex and $\rho$ is a point outside it, the Hahn–Banach separation theorem promises a hyperplane between them — written as a Hermitian operator $W$ , that plane is an entanglement witness, and a single expectation value $\langle W\rangle<0$ certifies entanglement with a handful of local measurements. Second, transposing Bob's subsystem is a positive map (it sends density operators to density operators) but not a completely positive one, so it can act benignly on every separable term yet drag an entangled state to a matrix with a negative eigenvalue. The cleanest laboratory of both effects is a one-parameter family: a Bell state blended with white noise. The second figure follows that family across the separable–entangled boundary and shows exactly where, and why, the partial transpose goes negative.

Figure

The two-qubit Werner family ρ_W = p|Φ⁻⟩⟨Φ⁻| + (1−p)/4 · I₄ as p runs from 0 to 1. The smallest eigenvalue of the partial transpose is λ_min = (1−3p)/4, plotted here: it starts at +1/4 (maximally mixed, separable), crosses zero at the threshold p = 1/3, and reaches −1/2 at the pure Bell state. Separable for p ≤ 1/3, entangled for p > 1/3; at p = 1/2 the value is −1/8

The core equations

Separability is defined not by a formula for a single state but by membership in a set — the convex hull of product states. The mixed-state definition is the master equation from which everything else descends.

\rho_{AB}=\sum_{i} p_i\,\rho_i^{A}\otimes\rho_i^{B},\qquad p_i\ge0,\ \sum_i p_i=1 \qquad (1)

where: $\rho_{AB}$ — a density operator on $\mathcal{H}_A\otimes\mathcal{H}_B$ ; $p_i$ — classical probabilities of the shared label $i$ ; $\rho_i^{A},\rho_i^{B}$ — local density operators held by Alice and Bob. Reading: a state is separable exactly when such a decomposition exists; one may always take the local states to be pure, $\rho_i^A=|\alpha_i\rangle\langle\alpha_i|$ and $\rho_i^B=|\beta_i\rangle\langle\beta_i|$ , so separability is membership in the convex hull of pure product states. For pure global states the whole question collapses to a single integer supplied by the Schmidt decomposition. This is the one regime where separability is trivially decidable.

|\psi\rangle_{AB}=|\alpha\rangle_A\otimes|\beta\rangle_B\ \Longleftrightarrow\ \mathrm{Schmidt\ rank}=1 \qquad (2)

where: $|\alpha\rangle_A,|\beta\rangle_B$ — unit vectors in the two factors; $\mathrm{Schmidt\ rank}$ — the number of nonzero coefficients $\sqrt{\lambda_k}$ in $|\psi\rangle_{AB}=\sum_k\sqrt{\lambda_k}\,|k\rangle_A|k\rangle_B$ . Reading: rank $1$ means a single term, which factorizes; rank $\ge2$ is entanglement, and there is nothing in between. Equivalently, one can read entanglement off the marginal alone, because tracing out Bob leaves a state whose eigenvalues are the squared Schmidt coefficients.

S(\rho_A)=-\mathrm{Tr}(\rho_A\log_2\rho_A)=0\ \Longleftrightarrow\ |\psi\rangle_{AB}\ \text{separable} \qquad (3)

where: $S(\rho_A)$ — the von Neumann entropy of $\rho_A=\mathrm{Tr}_B|\psi\rangle\langle\psi|$ , measured in bits (the entanglement entropy). Reading: a pure global state is a product exactly when its reduced state is pure, i.e. when the entanglement entropy is zero; any nonzero $S(\rho_A)$ signals — and quantifies — pure-state entanglement. The first detector for mixed states is the entanglement witness, the operator form of the separating hyperplane in the figure above.

W=W^{\dagger},\quad \mathrm{Tr}(W\sigma)\ge0\ \ \forall\,\sigma\in\mathrm{SEP},\quad \mathrm{Tr}(W\rho)<0\ \ \text{for some entangled }\rho \qquad (4)

where: $W$ — a Hermitian observable defining the hyperplane; $\mathrm{SEP}$ — the convex set of separable states; $\sigma$ — an arbitrary separable state. Reading: a single negative expectation value $\langle W\rangle<0$ is a certificate of entanglement, requiring no full tomography — only the few local settings that build $W$ . The second detector is the partial transpose, the concrete positive-but-not-completely-positive map that separability cannot survive going negative.

\big(\rho^{T_B}\big)_{m\mu,\,n\nu}=\rho_{m\nu,\,n\mu}\,,\qquad \rho\in\mathrm{SEP}\ \Rightarrow\ \rho^{T_B}\ge0 \qquad (5)

where: $\rho^{T_B}$ — the partial transpose on subsystem $B$ ; $m,n$ — basis indices of $\mathcal{H}_A$ ; $\mu,\nu$ — basis indices of $\mathcal{H}_B$ . Reading: separability forces $\rho^{T_B}\ge0$ , so a negative eigenvalue is an unambiguous entanglement certificate (the Peres necessary condition); in $2\times2$ and $2\times3$ it is also sufficient. The canonical test-bed for both detectors is the two-qubit Werner state, a Bell state diluted with white noise, whose entanglement can be dialed continuously.

\rho_W=p\,|\Phi^-\rangle\langle\Phi^-|+\frac{1-p}{4}\,I_4,\qquad |\Phi^-\rangle=\tfrac{1}{\sqrt2}\big(|00\rangle-|11\rangle\big) \qquad (6)

where: $p$ — the mixing parameter, $0\le p\le1$ ; $I_4$ — the $4\times4$ identity (the maximally mixed two-qubit state); $|\Phi^-\rangle$ — a Bell state. Reading: the reduced states are $I/2$ for every $p$ , so all correlations are nonlocal in origin; the PPT criterion locates the entanglement boundary sharply at $p=1/3$ .

Limiting cases

The character of the separability problem changes completely as you move between purity and dimension — it is trivial in one corner and intractable in another.

Case	Result	Meaning
Pure states	$\vert\psi\rangle_{AB}=\sum_k\sqrt{\lambda_k}\,\vert k\rangle_A\vert k\rangle_B$	Separable iff Schmidt rank $=1$ ; entanglement fully quantified by $S(\rho_A)=-\sum_k\lambda_k\log_2\lambda_k$ .
Two qubits / qubit–qutrit ( $2\times2$ , $2\times3$ )	$\rho^{T_B}\ge0\iff\rho$ separable	Peres–Horodecki PPT is necessary and sufficient — but only in these lowest dimensions.
Higher dimensions ( $\ge 3\times3$ , $2\times4$ )	$\rho^{T_B}\ge0\not\Rightarrow\rho$ separable	PPT-entangled (bound-entangled) states exist; deciding separability is NP-hard (Gurvits).
Near the center	ball around $I/(d_Ad_B)$ is separable	The maximally mixed state is robustly separable; enough white noise destroys any entanglement.
Entanglement vs nonlocality	SEP $\subsetneq$ LHV $\subsetneq$ all states	Some entangled states admit a local hidden-variable model, so entangled $\neq$ Bell-nonlocal.

Common trapConfusing separable with product, and entangled with nonlocal

Two conflations trip up almost everyone. First, "separable" is not "product." A product state $\rho_A\otimes\rho_B$ is genuinely uncorrelated, but a separable state $\sum_i p_i\,\rho_i^A\otimes\rho_i^B$ can be strongly correlated — the classically correlated state $\tfrac12(|00\rangle\langle00|+|11\rangle\langle11|)$ has perfectly correlated outcomes yet is manifestly separable. Separable means "no more than classical correlation," not "no correlation." Second, "entangled" is not "nonlocal." Every entangled pure state violates some Bell inequality, but Werner's own 1989 construction exhibits entangled mixed states that admit a local hidden-variable model and violate no Bell inequality at all. Entanglement is necessary for Bell nonlocality but not sufficient, and steering sits strictly between them: nonlocal $\subsetneq$ steerable $\subsetneq$ entangled.

Derivation APure-state separability equals Schmidt rank one

The goal is to prove that a pure bipartite state is separable if and only if exactly one Schmidt coefficient is nonzero, and to show that this single criterion wears three equivalent faces — Schmidt rank one, a pure reduced state, and zero entanglement entropy — so that the entire pure-state problem reduces to reading a single number off the marginal. We start from the Schmidt form guaranteed by the singular-value decomposition, trace out one party to expose the reduced spectrum, and analyze when factorization is possible.

Step 1. Every pure bipartite state can be written in Schmidt form $|\psi\rangle_{AB}=\sum_{k=1}^{r}\sqrt{\lambda_k}\,|k\rangle_A|k\rangle_B$ with $\lambda_k>0$ , $\sum_k\lambda_k=1$ , and orthonormal Schmidt bases $\{|k\rangle_A\},\{|k\rangle_B\}$ ; the integer $r$ is the Schmidt rank. Why: applying the singular-value decomposition to the coefficient matrix of $|\psi\rangle$ diagonalizes it simultaneously in both factors, and $r$ is the number of nonzero singular values — a basis-independent invariant.

Step 2. Trace out Bob: $\rho_A=\mathrm{Tr}_B|\psi\rangle\langle\psi|=\sum_k\lambda_k\,|k\rangle_A\langle k|$ . The eigenvalues of $\rho_A$ are exactly the squared Schmidt coefficients $\lambda_k$ , so $\rho_A$ is pure ( $\mathrm{Tr}\,\rho_A^2=1$ ) precisely when $r=1$ . Why: purity of the marginal is equivalent to a single nonzero eigenvalue, which is the same statement as Schmidt rank one — the marginal remembers exactly the entanglement, nothing else.

Step 3. If $r=1$ then normalization forces $\lambda_1=1$ , and $|\psi\rangle=\sqrt{\lambda_1}\,|1\rangle_A|1\rangle_B=|1\rangle_A\otimes|1\rangle_B$ , a product. Conversely a product state $|\alpha\rangle\otimes|\beta\rangle$ already is a one-term Schmidt decomposition, so it has $r=1$ . Why: a single Schmidt term with unit weight is literally a tensor product, closing the equivalence in both directions.

Step 4. Assemble the three faces: Schmidt rank $1$ $\iff$ reduced state pure $\iff$ entanglement entropy $S(\rho_A)=-\sum_k\lambda_k\log_2\lambda_k=0$ . Any one of them detects pure-state separability. Why: the entropy vanishes iff a single $\lambda_k=1$ , which is rank one again — the same fact stated informationally, and the reason $S(\rho_A)$ both detects and measures entanglement for pure states. *Verified: for the Bell state $|\Phi^+\rangle=(|00\rangle+|11\rangle)/\sqrt2$ one gets $\lambda_1=\lambda_2=\tfrac12$ , so $r=2$ , $\rho_A=I/2$ , and $S(\rho_A)=1$ bit — maximally entangled, consistent with Nielsen & Chuang §2.5.*

Go deeper · Derivation B · 1 minThe partial-transpose threshold of the two-qubit Werner family

Derivation BThe partial-transpose threshold of the two-qubit Werner family

The goal is to compute exactly where a Bell state, progressively diluted with white noise, stops being entangled, by building the partial transpose of the Werner state, diagonalizing it, and locating the sign change of its smallest eigenvalue — a calculation that both produces the sharp threshold $p=1/3$ and showcases the PPT criterion as a complete test in the $2\times2$ world. We compute the transpose of the pure Bell projector first, then add the noise floor, then impose positivity.

Step 1. Partial-transpose the Bell projector on Bob. Since $|\Phi^-\rangle\langle\Phi^-|=\tfrac12(|00\rangle\langle00|-|00\rangle\langle11|-|11\rangle\langle00|+|11\rangle\langle11|)$ and $|ij\rangle\langle kl|\mapsto|il\rangle\langle kj|$ under $T_B$ , the two coherences move: $|00\rangle\langle11|\mapsto|01\rangle\langle10|$ and $|11\rangle\langle00|\mapsto|10\rangle\langle01|$ . Why: the partial transpose swaps only Bob's indices, converting the $\{00,11\}$ coherence of the Bell state into a $\{01,10\}$ coherence, which is where the negativity will hide.

Step 2. The transposed projector is $\tfrac12(|00\rangle\langle00|+|11\rangle\langle11|-|01\rangle\langle10|-|10\rangle\langle01|)$ ; on the $\{|01\rangle,|10\rangle\}$ block this is $\tfrac12\!\left(\begin{smallmatrix}0&-1\\-1&0\end{smallmatrix}\right)$ with eigenvalues $\pm\tfrac12$ , so $(|\Phi^-\rangle\langle\Phi^-|)^{T_B}$ has spectrum $\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ . Why: transposition preserves the trace ( $=1$ ) but not positivity — the lone $-\tfrac12$ is the fingerprint of maximal entanglement.

Step 3. Add the noise floor. Because $(I_4/4)^{T_B}=I_4/4$ , the Werner transpose is $\rho_W^{T_B}=p\,(|\Phi^-\rangle\langle\Phi^-|)^{T_B}+\tfrac{1-p}{4}I_4$ , with eigenvalues $p(\tfrac12)+\tfrac{1-p}{4}=\tfrac{1+p}{4}$ (three times) and $p(-\tfrac12)+\tfrac{1-p}{4}=\tfrac{1-3p}{4}$ (once). Why: the identity is invariant under transposition and merely shifts every eigenvalue by the same noise amount, so the spectrum is the Bell spectrum rescaled by $p$ and lifted by $\tfrac{1-p}{4}$ .

Step 4. Impose positivity. Three eigenvalues $\tfrac{1+p}{4}$ are positive for all $p\in[0,1]$ ; the fourth obeys $\tfrac{1-3p}{4}\ge0\iff p\le\tfrac13$ . So $\rho_W^{T_B}\ge0$ exactly for $p\le\tfrac13$ . Since the Horodecki theorem makes PPT necessary and sufficient in $2\times2$ , the boundary is sharp: separable for $p\le\tfrac13$ , entangled for $p>\tfrac13$ . Why: in the lowest dimensions no bound entanglement exists, so a single scalar sign flip decides separability outright — the threshold is exact, not merely a bound. *Verified: the four eigenvalues sum to $1$ ; at $p=1$ they are $\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ , the known Bell transpose; at $p=\tfrac12$ they are $\{\tfrac38,\tfrac38,\tfrac38,-\tfrac18\}$ . Consistent with Horodecki RMP 81, 865 (2009).*

Go deeper · Derivation C · 2 minEvery entangled state has an entanglement witness

Derivation CEvery entangled state has an entanglement witness

The goal is to prove that the closedness and convexity of the separable set force a separating hyperplane between it and any entangled state, so that a Hermitian witness operator exists for every entangled $\rho$ — a purely geometric guarantee that underlies all experimental entanglement certification. We establish the two topological properties of SEP, invoke the Hahn–Banach separation theorem in the real space of Hermitian operators, and normalize the resulting functional into a witness.

Step 1. SEP is the closed convex hull of the compact set of pure product states $\{|\alpha\rangle\langle\alpha|\otimes|\beta\rangle\langle\beta|\}$ ; both closedness and convexity are inherited by the hull, so SEP is a closed, bounded, convex body in the finite-dimensional real vector space of Hermitian operators. Why: finite dimension makes the hull of a compact set compact, hence closed — the separating-hyperplane theorem needs exactly this.

Step 2. Equip that real space with the Hilbert–Schmidt inner product $\langle X,Y\rangle=\mathrm{Tr}(XY)$ . For an entangled $\rho\notin\mathrm{SEP}$ , Hahn–Banach supplies a Hermitian $W$ and a real constant $c$ with $\mathrm{Tr}(W\sigma)\ge c$ for all $\sigma\in\mathrm{SEP}$ and $\mathrm{Tr}(W\rho)<c$ . Why: a point outside a closed convex set can always be strictly separated from it by a hyperplane, and in this inner-product space every hyperplane is "measure the observable $W$ ."

Step 3. Normalize away the offset. Replacing $W\to W-cI$ and using $\mathrm{Tr}\,\sigma=1$ shifts the threshold to zero: $\mathrm{Tr}(W\sigma)\ge0$ for all separable $\sigma$ , while $\mathrm{Tr}(W\rho)<0$ . This is the canonical witness of equation (4). Why: subtracting a multiple of the identity is free because every state has unit trace, letting us fix the separating value at zero without loss.

Step 4. Read off the scope and the limit. Every entangled state is caught by some witness — but because the boundary of SEP is curved, no single $W$ catches them all; each witness is one tangent plane to a rounded body that needs infinitely many. Why: a convex body with a smooth boundary is the intersection of all its tangent half-spaces, so witnesses are individually incomplete yet collectively exhaustive. *Verified: this is the Horodecki characterization (Phys. Lett. A 223, 1 (1996)); the argument uses only finite-dimensionality and convexity, matching Watrous ch. 6.*

Worked examples

1 · The Werner state at p = 1/2 (fully worked)

Take the two-qubit Werner state $\rho_W=\tfrac12|\Phi^-\rangle\langle\Phi^-|+\tfrac14\cdot\tfrac12 I_4$ (that is, $p=\tfrac12$ ). Decide whether it is entangled, and quantify the entanglement by its negativity.

Step 1. Write the density matrix in the computational basis $\{00,01,10,11\}$ . The Bell projector contributes $\tfrac12$ on the $\{00,11\}$ diagonal and $-\tfrac12$ on the $\{00,11\}$ off-diagonal; the noise term adds $\tfrac{1-p}{4}=\tfrac18$ to every diagonal entry. With $p=\tfrac12$ the diagonal entries are $\tfrac12\cdot\tfrac12+\tfrac18=\tfrac38$ (for $00,11$ ) and $\tfrac18$ (for $01,10$ ), and the off-diagonal $\{00,11\}$ entry is $-\tfrac12\cdot\tfrac12=-\tfrac14$ .

Step 2. Apply the PPT spectrum from the derivation with $p=\tfrac12$ : the eigenvalues of $\rho_W^{T_B}$ are $\tfrac{1+p}{4}=\tfrac{3}{8}$ (three times) and $\tfrac{1-3p}{4}=\tfrac{1-3/2}{4}=-\tfrac18$ (once), i.e. $\{\tfrac38,\tfrac38,\tfrac38,-\tfrac18\}$ . As a check they sum to $\tfrac{9}{8}-\tfrac18=1$ , as any transposed density operator must.

Step 3. Read the verdict. The partial transpose has a negative eigenvalue $-\tfrac18$ , so by the Peres criterion $\rho_W$ is entangled; because this is a $2\times2$ system, PPT is also sufficient, so there is no ambiguity. Consistently, $p=\tfrac12>\tfrac13$ , on the entangled side of the threshold.

Step 4. Quantify it. The negativity $\mathcal{N}=\sum_{\lambda_k<0}|\lambda_k|=\tfrac18$ , and in general for this family $\mathcal{N}(p)=\tfrac{3p-1}{4}$ for $p>\tfrac13$ (zero below). So at $p=\tfrac12$ the state carries a modest but nonzero $\mathcal{N}=\tfrac18$ , rising to $\mathcal{N}=\tfrac12$ at the pure Bell state $p=1$ . $\rho_W$ at $p=\tfrac12$ is entangled, with PPT spectrum $\{\tfrac38,\tfrac38,\tfrac38,-\tfrac18\}$ and negativity $\tfrac18$ .

2 · Classical correlation versus entanglement

Compare two two-qubit states: the classically correlated mixture $\rho_{\mathrm{cc}}=\tfrac12\big(|00\rangle\langle00|+|11\rangle\langle11|\big)$ and a pure state whose reduced density matrix is $\rho_A=\mathrm{diag}(0.8,\,0.2)$ . Which are entangled?

Step 1. Test $\rho_{\mathrm{cc}}$ against the definition. It is already written as a convex combination of the product projectors $|00\rangle\langle00|=|0\rangle\langle0|\otimes|0\rangle\langle0|$ and $|11\rangle\langle11|=|1\rangle\langle1|\otimes|1\rangle\langle1|$ , each with weight $\tfrac12$ . That is exactly the separable form of equation (1) with $p_1=p_2=\tfrac12$ . So $\rho_{\mathrm{cc}}$ is separable — despite having perfectly correlated measurement outcomes, its correlation is purely classical.

Step 2. Cross-check with PPT. The only off-diagonal structure in $\rho_{\mathrm{cc}}$ is absent (it is diagonal), so $\rho_{\mathrm{cc}}^{T_B}=\rho_{\mathrm{cc}}\ge0$ ; no negative eigenvalue, consistent with separability. A diagonal state can never be entangled.

Step 3. Now the pure state. Its reduced state $\rho_A=\mathrm{diag}(0.8,0.2)$ is mixed (two nonzero eigenvalues), so the Schmidt rank is $2$ and the global pure state cannot factorize — it is entangled. Quantify by the entanglement entropy: $S(\rho_A)=-0.8\log_2 0.8-0.2\log_2 0.2$ .

Step 4. Evaluate. $-0.8\log_2 0.8=0.8\times0.3219=0.2575$ and $-0.2\log_2 0.2=0.2\times2.3219=0.4644$ , so $S(\rho_A)=\mathbf{0.72}$ bits — partial entanglement, between the $0$ of a product state and the $1$ bit of a Bell pair. The classically correlated $\rho_{\mathrm{cc}}$ is separable; the pure state with mixed marginal is entangled with $0.72$ bits of entanglement entropy.

Physics of the portrait

The modern definition of entanglement — the convex, operational one this whole article rests on — is not a hundred years old. It was written down in 1989 by the person who also proved that entanglement and nonlocality are not the same thing.

Origins · 1 minReinhard Werner and how this was discovered

Reinhard Werner

born 1954 · the operational definition of separable and entangled states

Reinhard F. Werner, a German mathematical physicist, gave entanglement the definition it still carries. Before 1989 "entangled" was used loosely for correlated quantum states, and it was widely assumed that entanglement and Bell nonlocality were one and the same. Werner's landmark paper "Quantum states with Einstein–Podolsky–Rosen correlations admitting a hidden-variable model" (Phys. Rev. A 40, 4277) severed both confusions at once. He defined a state to be separable precisely when it is a convex mixture of product states — the operationally clean statement that it can be prepared by local operations and shared classical randomness — and called everything else entangled. Then, in the same paper, he constructed an explicit family of entangled states (now the Werner states, symmetric under $U\otimes U$ ) that nonetheless admit a local hidden-variable model for all projective measurements, proving that an entangled state can be perfectly local. This single stroke created the modern hierarchy — product $\subsetneq$ separable, and nonlocal $\subsetneq$ steerable $\subsetneq$ entangled — and set the agenda for two decades of work on detection (the Peres and Horodecki criteria, entanglement witnesses) and quantification (entanglement measures) that his convex-geometric framing made possible. Werner, long at the Technical University of Braunschweig and Hannover, remained a central figure in the mathematical foundations of quantum information, but it is this one definition — SEP as a closed convex set — that every article on entanglement now begins from.

History

1935

— Einstein, Podolsky, and Rosen question completeness; Schrödinger names the phenomenon entanglement

1964

— Bell derives an inequality separating quantum predictions from local hidden variables

1989

— Werner defines separable states and builds entangled states with a local hidden-variable model

1996

— Peres proposes the partial-transpose test; the Horodeckis prove it exact for 2×2 and 2×3 and characterize witnesses

2003

— Gurvits proves that deciding separability is NP-hard

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §2.2.8 and §12.5 (density operators, entanglement) · Watrous, The Theory of Quantum Information, ch. 6 (separability, witnesses, PPT) · Wilde, Quantum Information Theory, ch. 4 and 11 · Preskill, Lecture Notes on Quantum Computation, ch. 4 (entanglement) · R. Horodecki, P. Horodecki, M. Horodecki, K. Horodecki, "Quantum entanglement," Rev. Mod. Phys. 81, 865 (2009) · Werner, Phys. Rev. A 40, 4277 (1989) · Peres, Phys. Rev. Lett. 77, 1413 (1996); Horodecki³, Phys. Lett. A 223, 1 (1996) · Gurvits, STOC 2003 (NP-hardness of separability).

Provenance. A verifier should recompute: (1) the Schmidt-rank/entropy criterion and the Bell values $\lambda_k=\tfrac12$ , $S=1$ bit (Nielsen & Chuang §2.5, §12.5); (2) the Werner partial-transpose spectrum $\{\tfrac{1+p}{4}\times3,\ \tfrac{1-3p}{4}\}$ , its $p=\tfrac13$ threshold, and the $p=\tfrac12$ values $\{\tfrac38,\tfrac38,\tfrac38,-\tfrac18\}$ with negativity $\tfrac18$ (Horodecki RMP 81, 865; Werner PRA 40, 4277); (3) the PPT positivity-preservation argument and its sufficiency in $2\times2$ , $2\times3$ (Peres PRL 77, 1413; Horodecki³ PLA 223, 1); (4) the witness-existence argument via Hahn–Banach (Watrous ch. 6; Horodecki PLA 223); (5) $S(\mathrm{diag}(0.8,0.2))\approx0.72$ bit. NP-hardness of separability per Gurvits, STOC 2003; all numbers are standard.

Prerequisites. Density Operators and Mixed States · Tensor Products and Bipartite Systems · Schmidt Decomposition. Leads to. Peres–Horodecki / PPT Criterion · Entanglement Witnesses · Concurrence and Wootters' Formula · Entanglement of Formation · CHSH Inequality and Its Quantum Violation · Bell States and EPR Pairs.

Ask the tutor: Ask me to run the PPT test on any two-qubit state you have in mind, or to show why your favorite Werner state sits on the separable or the entangled side of $p=\tfrac13$ .

What exactly is the difference between a product state and a separable state?

A product state $\rho_A\otimes\rho_B$ has no correlation at all: measurement outcomes on the two halves are statistically independent. A separable state $\sum_i p_i\,\rho_i^A\otimes\rho_i^B$ is more general — it can be strongly correlated, but only classically, through the shared random label $i$ . The textbook example is $\tfrac12(|00\rangle\langle00|+|11\rangle\langle11|)$ : measure both qubits in the computational basis and you always get matching results, a perfect correlation, yet the state is separable because that correlation is just a shared coin flip. So every product state is separable, but most separable states are not products. "Separable" means "no more than classical correlation," never "no correlation."

Why is deciding separability hard when the definition looks so simple?

The definition asks whether any decomposition into product states exists — an existential quantifier over an infinite, unstructured set of possible ensembles. There is no bound on how many product terms you might need (up to $(d_Ad_B)^2$ by Carathéodory), and the product states can be non-orthogonal and point anywhere. Searching that space is genuinely intractable: Gurvits proved in 2003 that deciding separability is NP-hard. The PPT criterion and witnesses are tractable one-sided tests — they can prove entanglement (or, in low dimensions, separability) — but no efficient algorithm certifies separability in general dimension, which is why the field relies on a toolbox of partial tests rather than a single decision procedure.

If the partial transpose is positive, is the state definitely separable?

Only in the smallest systems. In $2\times2$ and $2\times3$ dimensions the Horodecki theorem makes PPT necessary and sufficient, so $\rho^{T_B}\ge0$ does guarantee separability there. But from $3\times3$ and $2\times4$ onward, PPT-entangled states exist: they have a positive partial transpose yet are entangled. This "bound entanglement" cannot be distilled into pure Bell pairs by LOCC, which is why it hid from the earliest tests. So in general $\rho^{T_B}\ge0$ is only a necessary condition for separability — it rules out one kind of entanglement (the distillable, NPT kind) but not all of it.

How can an entangled state fail to be nonlocal?

Bell nonlocality means a state produces measurement statistics that no local hidden-variable (LHV) model can reproduce. Werner's 1989 construction gave explicit entangled mixed states — the Werner states below a certain noise level — that do admit an LHV model for all projective measurements, so they violate no Bell inequality despite being entangled. The resolution is that entanglement is a property of the state (it cannot be made by LOCC), while nonlocality is a property of the measurement statistics the state can generate. Mixing with noise can wash out the nonlocal correlations while leaving a trace of entanglement that a witness or the PPT test still detects. Hence the strict hierarchy: nonlocal states are a proper subset of entangled ones.

Why does the separable set being convex matter so much?

Convexity is what makes entanglement detectable. Because SEP is a closed convex set, the Hahn–Banach separation theorem guarantees that any entangled state — a point outside the set — can be separated from it by a flat hyperplane. Expressed as a Hermitian operator, that hyperplane is an entanglement witness $W$ : an observable whose expectation is nonnegative on every separable state but negative on the target. This is why a laboratory can certify entanglement with a few measurements instead of full tomography. Convexity also underlies the PPT test (a positive map maps the convex set into positive operators) and the semidefinite-programming hierarchies used to approximate SEP from outside. Without convexity, none of these tools would exist.

Does adding noise always destroy entanglement eventually?

Yes. The maximally mixed state $I/(d_Ad_B)$ sits deep in the interior of the separable set, surrounded by a ball of separable states of finite radius (the Gurvits–Barnum result). So any state, mixed with enough white noise, is eventually dragged into that ball and becomes separable — entanglement is not robust to unlimited noise. The Werner family makes this quantitative: $\rho_W$ is separable once $p\le\tfrac13$ , i.e. once the Bell component is diluted to one-third weight or less. This is exactly why noise thresholds like $p=\tfrac13$ function as hardware targets: a device must keep its shared states robustly outside the separable ball to deliver any entanglement-based advantage.