Product States and Separability

Quantum mechanics hands a composite system $AB$ the tensor-product Hilbert space $\mathcal{H}_A\otimes\mathcal{H}_B$ , and the first question one can ask of any joint state is whether it hides a correlation between the parts. The states that hide none are the product states $|\alpha\rangle_A\otimes|\beta\rangle_B$ and their classical mixtures, the separable states. These carry a crisp operational identity: they are exactly the states two distant experimenters, Alice and Bob, can build using local operations and a shared coin flip — no quantum channel ever runs between the labs. Everything they cannot build this way is entangled, and entanglement is the one structural feature of composite quantum systems with no classical shadow. That single dividing line organizes half of quantum information theory. Entanglement is the fuel of teleportation, dense coding, and device-independent cryptography, and it is the reason the state space of $AB$ is vastly larger than the product of the two local state spaces. For a pure state the boundary is sharp and computable: the Schmidt decomposition rewrites $|\psi\rangle_{AB}$ in a canonical bi-orthonormal form, and the count of nonzero terms — the Schmidt rank — is one exactly when the state factorizes. The reduced density operator $\rho_A=\operatorname{Tr}_B|\psi\rangle\langle\psi|$ is then pure, and its entropy measures the entanglement. For mixed states the geometry turns subtle. A state can carry classically correlated, genuinely mixed marginals and still be separable, so marginal mixedness no longer signals entanglement. Deciding whether an arbitrary $\rho_{AB}$ admits any product decomposition is NP-hard, so one leans on one-sided witnesses — above all positivity under partial transpose. This article builds the definitions, the LOCC meaning, the detection tools, and the small dimensions where the whole question closes.

Key takeaways

A bipartite pure state is a product state when it factorizes, $|\psi\rangle_{AB}=|\alpha\rangle_A\otimes|\beta\rangle_B$ ; if no such factorization exists it is entangled.
For mixed states the right notion is separability: $\rho_{AB}$ is separable iff it is a convex mixture of products $\sum_i p_i\,\rho_A^{(i)}\otimes\rho_B^{(i)}$ — precisely what Alice and Bob can prepare by local operations and shared randomness.
Pure-state product structure is detected by the Schmidt rank: a state is a product iff its Schmidt rank equals $1$ , equivalently iff its marginal is pure and $S(\rho_A)=0$ .
Product states have uncorrelated statistics: every joint expectation factorizes, $\langle A\otimes B\rangle=\langle A\rangle\langle B\rangle$ , which is why separable resources can never beat the classical bounds of Bell-type games.
Deciding separability is NP-hard in general, so the Peres–Horodecki (PPT) criterion is used as a one-way test: separable $\Rightarrow$ partial transpose positive, and it becomes iff only in $2\times2$ and $2\times3$ .
Key idea: separable = preparable by local operations and shared randomness, a convex mixture of products $\sum_i p_i\,\rho_A^{(i)}\otimes\rho_B^{(i)}$ ; everything else is entangled, and for pure states that collapses to the single sharp test Schmidt rank $=1$ .

The core insight

A bipartite state is separable exactly when it is a convex mixture of product states $\sum_i p_i\,\rho_A^{(i)}\otimes\rho_B^{(i)}$ — everything Alice and Bob can build with local operations and shared randomness — and everything outside this convex body is entangled; for pure states the criterion reduces to Schmidt rank $=1$ .

Where this lives today

The separable-versus-entangled boundary is now the daily currency of experimental quantum information. Because full state tomography scales exponentially, labs almost never reconstruct $\rho$ to certify entanglement; instead they measure a single entanglement witness $W$ — a Hermitian observable engineered so that $\operatorname{Tr}(W\rho_{\text{sep}})\ge0$ for every separable state while $\operatorname{Tr}(W\rho)<0$ for the target. The convexity of the separable body guarantees such a hyperplane exists, and a handful of local measurement settings suffices to evaluate it, which is how entanglement is verified in trapped-ion registers, superconducting processors, and photonic experiments. The Werner threshold reappears directly as a noise budget: an entanglement-distillation or teleportation protocol fed a depolarized Bell pair fails once the singlet fraction drops through the separability floor, so the number $p=\tfrac13$ sets how much depolarizing noise a link can tolerate before its resource is worthless. At the frontier sits bound entanglement — the PPT-entangled states the Horodeckis found in 1998 — which is entangled yet cannot be distilled to pure singlets by any LOCC protocol; some of these states nonetheless enable secret-key generation, and detecting them requires witnesses strictly stronger than PPT. Deciding separability in general remains provably NP-hard, so the field runs on a growing toolbox of witnesses, symmetric extensions, and semidefinite relaxations rather than any single exact test.

The physical picture

Two experimenters sit in distant labs. They may telephone each other and consult a shared table of random numbers, but no quantum system ever crosses between them. Whatever joint state they manage to prepare is separable — a menu of local recipes indexed by a broadcast classical label. Entanglement is by definition the correlation this procedure can never manufacture. Geometrically the separable states form a convex body sitting inside the full state space: mixing two separable states gives a separable state, and the extreme points that generate the body are exactly the pure products $\rho_A\otimes\rho_B$ . Every entangled state lives outside that body, and because the body is convex a flat hyperplane — an entanglement witness — can always be slipped between it and any chosen entangled point. The first figure is that geometry.

Figure

The state space is convex — the separable states form a convex body whose extreme points are the pure products ρ_A⊗ρ_B, the maximally mixed state 1/4 sits deep inside, and every entangled state lies outside, cut off by a witness hyperplane Tr(Wρ)=0 with a negative value Tr(Wρ) certifying entanglement

For pure states the whole question collapses to a count. Write the amplitudes of $|\psi\rangle_{AB}$ as a matrix $C$ with rows indexed by $A$ and columns by $B$ ; the Schmidt rank is just the matrix rank of $C$ . Rank one means the matrix is an outer product $C=\mathbf{a}\,\mathbf{b}^{\top}$ , which is precisely $|\alpha\rangle\otimes|\beta\rangle$ — a product. Any higher rank means entangled, and the singular values of $C$ are the Schmidt coefficients whose spread measures how much entanglement is present. The cleanest thermometer is a mixed marginal: trace out one side and read the entropy. Mixed states lose this simplicity because many distinct product ensembles average to the same $\rho$ , so a mixed marginal no longer implies entanglement. The sharpest practical handle for mixed states is the partial transpose, and the two-qubit Werner state shows exactly where its verdict flips from separable to entangled — the subject of the second figure.

Figure

The two-qubit Werner state ρ_W(p) mixes the singlet with white noise at weight p: its partial transpose has three eigenvalues (1+p)/4 and one smallest eigenvalue (1−3p)/4, which crosses zero at p = 1/3 — separable for p ≤ 1/3, entangled and NPT for p > 1/3

The core equations

The starting point is the simplest possible joint state: one where each subsystem carries a definite pure state of its own and the two are stapled together with a tensor product. Nothing about $A$ is correlated with anything about $B$ .

|\psi\rangle_{AB}=|\alpha\rangle_A\otimes|\beta\rangle_B \qquad (1)

where: $|\alpha\rangle_A$ — normalized state of subsystem $A$ ; $|\beta\rangle_B$ — normalized state of subsystem $B$ ; $\otimes$ — tensor (Kronecker) product joining the two factor spaces. Reading: a pure state is a product exactly when it splits into one ket per subsystem, and then a measurement on $A$ tells you nothing about $B$ . Mixed states demand the more general notion. Allow a shared classical label $i$ , drawn with probability $p_i$ and announced to both parties, and let each party prepare a local state conditioned on it. Averaging over the forgotten label gives the separable set.

\rho_{AB}=\sum_i p_i\,\rho_A^{(i)}\otimes\rho_B^{(i)},\qquad p_i\ge0,\ \sum_i p_i=1 \qquad (2)

where: $p_i$ — probabilities of the shared classical label $i$ ; $\rho_A^{(i)},\rho_B^{(i)}$ — local density operators conditioned on $i$ . Reading: this convex combination of products is, by construction, the most general state reachable with local operations and classical communication (LOCC); a state is separable iff at least one such decomposition exists. Any bipartite pure state can be rotated into a canonical diagonal form using a single pair of local orthonormal bases. This Schmidt decomposition is the workhorse that turns "is it a product?" into "how many terms survive?"

|\psi\rangle_{AB}=\sum_{k=1}^{r}\sqrt{\lambda_k}\,|u_k\rangle_A\otimes|v_k\rangle_B \qquad (3)

where: $r$ — the Schmidt rank, the number of nonzero terms; $\lambda_k>0$ with $\sum_k\lambda_k=1$ — the Schmidt coefficients; $|u_k\rangle,|v_k\rangle$ — orthonormal Schmidt bases of $A$ and $B$ . Reading: the state is a product exactly when $r=1$ ; any $r\ge2$ is entanglement, and the marginals $\rho_A,\rho_B$ share the spectrum $\{\lambda_k\}$ . Product structure has an immediate statistical fingerprint: joint expectation values come apart into a plain product of local averages. Correlations of every observable pair vanish.

\langle A\otimes B\rangle=\operatorname{Tr}\!\big[(A\otimes B)\,\rho_A\otimes\rho_B\big]=\langle A\rangle\langle B\rangle \qquad (4)

where: $A,B$ — local observables on the two subsystems; $\langle A\rangle=\operatorname{Tr}(A\rho_A)$ — the local expectation. Reading: for a product state the connected correlator $\langle A\otimes B\rangle-\langle A\rangle\langle B\rangle$ is identically zero, so a product state describes statistical independence. For a pure bipartite state the amount of entanglement is faithfully captured by the von Neumann entropy of either marginal — an entropy that is zero exactly for products and maximal for the flattest Schmidt spectrum.

S(\rho_A)=-\operatorname{Tr}\,\rho_A\log_2\rho_A=-\sum_{k}\lambda_k\log_2\lambda_k \qquad (5)

where: $S(\rho_A)$ — von Neumann entropy of the marginal, in bits; $\lambda_k$ — Schmidt coefficients of $|\psi\rangle_{AB}$ . Reading: $S=0$ iff a single $\lambda_k=1$ (a product), and $S=\log_2 d$ for the maximally entangled state of two $d$ -level systems. Detecting entanglement in a mixed state needs a computable test. Transpose only the $B$ block of $\rho$ in some product basis; a separable state stays a valid density operator, so any negative eigenvalue certifies entanglement.

\big(\rho^{T_B}\big)_{ij,kl}=\rho_{il,kj},\qquad \rho\ \text{separable}\ \Rightarrow\ \rho^{T_B}\ge0 \qquad (6)

where: $T_B$ — transpose applied only to $B$ 's indices; $i,k$ index $A$ , and $j,l$ index $B$ . Reading: this is the Peres–Horodecki (PPT) criterion — separability implies positivity of the partial transpose, and in $2\times2$ and $2\times3$ the converse holds too, making PPT a complete test there.

Derivation AWhy separability is exactly the set of LOCC-preparable states

The goal is to show that the algebraic definition of separability — a convex mixture of product density operators — is not an arbitrary formula but the precise mathematical image of what two spatially separated parties can build when they are allowed to act locally on their own systems and to talk on a classical telephone, but are forbidden to send any quantum system to each other. We want the operational story and the equation to line up term by term, so that "separable" and "preparable by local operations and shared randomness" become interchangeable names for one set.

Step 1. Give Alice and Bob a source of shared randomness: a referee samples an index $i$ with probability $p_i$ and broadcasts it to both labs over a classical channel. This is the only thing crossing between the parties, and it is classical. Why: shared randomness is free under LOCC and is exactly what a common classical label encodes.

Step 2. Conditioned on hearing $i$ , Alice prepares the local state $\rho_A^{(i)}$ on her system and Bob independently prepares $\rho_B^{(i)}$ on his, with no quantum link. The joint conditional state is the product $\rho_A^{(i)}\otimes\rho_B^{(i)}$ . Why: acting locally on disjoint systems produces a tensor product, never anything more correlated, because the two preparation maps commute and touch different factors.

Step 3. Now forget the label — either the referee discards it or it is never recorded. Averaging the conditional states over $i$ gives $\rho_{AB}=\sum_i p_i\,\rho_A^{(i)}\otimes\rho_B^{(i)}$ , which is exactly equation (2). Why: marginalizing a classical register turns a conditional ensemble into a convex mixture, so the reachable states are precisely the convex hull of product states.

Step 4. Finally, general LOCC — local quantum channels interleaved with rounds of classical messaging — cannot escape this set. A local channel maps each $\rho_A^{(i)}$ to $\mathcal{E}_A(\rho_A^{(i)})$ , still a valid local state, and classical messages only refine the label $i$ . Every operation preserves the convex-sum-of-products form. Why: this is why entanglement is a resource — it is monotone non-increasing under LOCC, so it can be consumed but never created locally.

Go deeper · Derivation B · 1 minA Bell state admits no product factorization

Derivation BA Bell state admits no product factorization

The goal is to prove, in two independent ways, that the maximally entangled state $|\Phi^+\rangle=\tfrac1{\sqrt2}(|00\rangle+|11\rangle)$ is genuinely entangled and not merely a product written in a clever basis. First we assume a factorization exists and drive the amplitude-matching conditions into a flat contradiction; then we cross-check by computing the reduced state and reading its Schmidt rank, so that the direct algebra and the rank criterion give the same verdict and reinforce each other.

Step 1. Write the most general two-qubit product, $(a|0\rangle+b|1\rangle)\otimes(c|0\rangle+d|1\rangle)=ac|00\rangle+ad|01\rangle+bc|10\rangle+bd|11\rangle$ , with $|a|^2+|b|^2=|c|^2+|d|^2=1$ . This is the target we must either match or rule out. Why: any product pure state, in any local bases, has exactly this bilinear amplitude structure.

Step 2. Match coefficients against $|\Phi^+\rangle$ : the $|01\rangle$ and $|10\rangle$ amplitudes are absent, forcing $ad=0$ and $bc=0$ , while $ac=bd=\tfrac1{\sqrt2}$ . Why: two states are equal only if every amplitude agrees, so the missing cross terms become hard algebraic constraints.

Step 3. Now $ad=0$ forces $a=0$ or $d=0$ . Either choice kills one of the required nonzero products — $a=0$ makes $ac=0$ , and $d=0$ makes $bd=0$ — contradicting $ac=bd=\tfrac1{\sqrt2}$ . No amplitudes work, so no factorization exists. Why: the vanishing cross terms and the nonvanishing diagonal terms cannot be satisfied at once, which is the algebraic signature of entanglement.

Step 4. Cross-check by Schmidt rank. Tracing out $B$ gives $\rho_A=\operatorname{Tr}_B|\Phi^+\rangle\langle\Phi^+|=\tfrac12\mathbb{1}$ , the maximally mixed qubit with eigenvalues $\{\tfrac12,\tfrac12\}$ , so the Schmidt rank is $2$ and $S(\rho_A)=-2\cdot\tfrac12\log_2\tfrac12=1$ bit. Why: rank $2>1$ independently confirms entanglement, and $S=1$ bit marks it as maximal for a qubit pair.

Go deeper · Derivation C · 1 minThe Werner state and its separability threshold

Derivation CThe Werner state and its separability threshold

The goal is to find the exact noise level at which the two-qubit Werner state — a singlet contaminated by white noise — stops being entangled and becomes separable, and to do it with the PPT criterion, which is both necessary and sufficient in this $2\times2$ setting. We define the family, compute the spectrum of its partial transpose in closed form, demand positivity, and read off the crossover, arriving at a clean threshold that later reappears as a noise budget for entanglement-based protocols.

Step 1. Define $\rho_W(p)=p\,|\Psi^-\rangle\langle\Psi^-|+\tfrac{1-p}{4}\mathbb{1}$ with the singlet $|\Psi^-\rangle=\tfrac1{\sqrt2}(|01\rangle-|10\rangle)$ and $0\le p\le1$ . It interpolates from the maximally mixed state at $p=0$ to the pure singlet at $p=1$ . Why: this one-parameter family models a maximally entangled resource degraded by depolarizing noise, the realistic situation in any lab.

Step 2. Compute the partial transpose. Transposing the $B$ block, the four eigenvalues of $\rho_W^{T_B}$ are $\tfrac{1+p}{4}$ (three-fold) and $\tfrac{1-3p}{4}$ (once). Direct diagonalization confirms these values exactly. Why: the noise term is proportional to the identity and transpose-invariant, so only the singlet's partial transpose, which has one negative-going eigenvalue, moves the spectrum.

Step 3. Impose positivity. Three eigenvalues stay positive for all $p$ , so the test rests entirely on the fourth: $\tfrac{1-3p}{4}\ge0\iff p\le\tfrac13$ . For $p>\tfrac13$ the partial transpose acquires a negative eigenvalue — the state is NPT. Why: a single sign change is what separates the PPT region from the entangled region, and it happens at $p=\tfrac13$ .

Step 4. Close the argument with sufficiency. Because this is a $2\times2$ system, the Horodecki theorem guarantees PPT is also sufficient, so $\rho_W(p)$ is separable for every $p\le\tfrac13$ and entangled for every $p>\tfrac13$ . The threshold is exactly $p=\tfrac13$ . Why: only in $2\times2$ and $2\times3$ does PPT close both ways, turning a one-sided witness into an exact separability certificate.

Limiting cases

Separability behaves differently depending on purity and on the dimensions of the two subsystems.

Case	Result	Meaning
Pure state	$\text{product}\iff\operatorname{rank}\rho_A=1\iff S(\rho_A)=0$	For pure $\|\psi\rangle_{AB}$ entanglement is equivalent to a mixed marginal; the marginal entropy measures it exactly.
$2\times2$ and $2\times3$	$\rho\ \text{separable}\iff\rho^{T_B}\ge0$	Peres–Horodecki: PPT is necessary AND sufficient (Horodecki theorem), so separability is fully decidable.
$d_A d_B>6$	$\rho^{T_B}\ge0\ \not\Rightarrow\ \text{separable}$	Bound-entangled (PPT) states exist; the partial transpose becomes only a necessary test and misses them.
Werner state	separable $\iff p\le\tfrac13$	The two-qubit singlet-plus-noise family crosses from separable to entangled and NPT exactly at $p=\tfrac13$ .
Maximally entangled	$\rho_A=\tfrac1d\mathbb{1},\ S=\log_2 d$	Flat Schmidt spectrum; the marginal carries no local information and all of it hides in the correlations.

Common trapA mixed state with correlated, mixed marginals must be entangled

It need not be. Mixed marginals and even strong correlations are perfectly compatible with separability, because classical correlation alone keeps a state inside the separable body. The state $\tfrac12(|00\rangle\langle00|+|11\rangle\langle11|)$ has both marginals maximally mixed and its outcomes are perfectly correlated, yet it is a plain classical mixture of two products — no entanglement anywhere. The reverse trap is just as common: do not read separability off a single witness. PPT is necessary everywhere but sufficient only in $2\times2$ and $2\times3$ ; once $d_A d_B>6$ there are bound-entangled states that pass the PPT test while being genuinely entangled, so a positive partial transpose in high dimension proves nothing on its own.

Worked examples

1 · Diagnosing a two-qubit pure state (fully worked)

Two states are on the table: $|\phi\rangle=\tfrac1{\sqrt2}(|00\rangle+|01\rangle)$ and $|\psi\rangle=\sqrt{0.9}\,|00\rangle+\sqrt{0.1}\,|11\rangle$ . For each, decide product versus entangled and quantify the entanglement.

Step 1. Factor $|\phi\rangle$ by inspection. Group the $A$ ket: $|\phi\rangle=\tfrac1{\sqrt2}|0\rangle_A\otimes(|0\rangle+|1\rangle)_B=|0\rangle_A\otimes|+\rangle_B$ . It is a clean tensor product, so its Schmidt rank is $1$ and it is a product state; its marginal $\rho_A=|0\rangle\langle0|$ is pure with $S=0$ . No entanglement.

Step 2. For $|\psi\rangle$ , read off the Schmidt form directly — it is already diagonal in $\{|0\rangle,|1\rangle\}\otimes\{|0\rangle,|1\rangle\}$ with coefficients $\sqrt{0.9}$ and $\sqrt{0.1}$ . Two nonzero terms means Schmidt rank $2$ , so $|\psi\rangle$ is entangled.

Step 3. Quantify it. The marginal is $\rho_A=\operatorname{diag}(0.9,0.1)$ , so the entanglement entropy is $S=-0.9\log_2 0.9-0.1\log_2 0.1$ . Numerically $\log_2 0.9=-0.152$ and $\log_2 0.1=-3.322$ , giving $S=0.9(0.152)+0.1(3.322)=0.137+0.332=\mathbf{0.469\ \text{bits}}$ .

Step 4. Sanity-check the extremes. If the weights were $\{1,0\}$ the state would be the product $|00\rangle$ with $S=0$ ; if they were $\{\tfrac12,\tfrac12\}$ it would be the Bell state with $S=1$ bit. Our $0.469$ bits sits between, marking $|\psi\rangle$ as partially entangled. Product $|\phi\rangle$ : S = 0; entangled $|\psi\rangle$ : Schmidt rank 2, S = 0.469 bits.

2 · Reading the Werner threshold at two noise levels

A source emits the two-qubit Werner state $\rho_W(p)$ . Test $p=0.2$ (heavy noise) and $p=0.8$ (light noise): is each separable or entangled?

Step 1. Use the closed-form partial-transpose spectrum from the derivation: eigenvalues $\tfrac{1+p}{4}$ (three of them) and $\tfrac{1-3p}{4}$ (one). Only the last can go negative, so it is the entire test.

Step 2. At $p=0.2$ : the critical eigenvalue is $\tfrac{1-0.6}{4}=\tfrac{0.4}{4}=\mathbf{+0.10}$ , and the triple is $\tfrac{1.2}{4}=0.30$ . All four eigenvalues are non-negative, so $\rho_W(0.2)$ is PPT; in $2\times2$ that means separable. Consistent with $0.2\le\tfrac13$ .

Step 3. At $p=0.8$ : the critical eigenvalue is $\tfrac{1-2.4}{4}=\tfrac{-1.4}{4}=\mathbf{-0.35}$ , negative, so the partial transpose is indefinite. $\rho_W(0.8)$ is NPT and therefore entangled. Consistent with $0.8>\tfrac13$ .

Step 4. Locate the boundary. Setting $\tfrac{1-3p}{4}=0$ gives $p=\tfrac13\approx0.333$ , the exact fidelity floor: any Werner source with singlet weight above one-third carries distillable entanglement, and anything at or below is a useless classical mixture. p = 0.2 separable, p = 0.8 entangled, crossover at p = 1/3.

Physics of the portrait

Separability as a precise mathematical predicate — as opposed to a vague sense that some states are "more correlated" than others — is a surprisingly modern invention. The word was pinned down only at the end of the 1980s, and the person who pinned it is the natural face for this article.

Origins · 1 minReinhard Werner and how this was discovered

Reinhard Werner

1954–, German mathematical physicist · defined separable versus entangled mixed states and the Werner states, 1989

For half a century after Einstein, Podolsky, Rosen and Schrödinger, "entanglement" was understood cleanly only for pure states, where a mixed marginal settles the question. Mixed states were a fog: a density matrix can be correlated for utterly classical reasons, and no one had a sharp line between correlation that a local-hidden-variable model can explain and correlation that it cannot. Reinhard Werner drew that line in a single 1989 paper. He defined a bipartite mixed state to be classically correlated — the term that became separable — exactly when it can be written as a convex mixture of product states $\sum_i p_i\,\rho_A^{(i)}\otimes\rho_B^{(i)}$ , and entangled otherwise, and he showed this is the operationally right notion by tying it to what local operations and shared randomness can prepare. To prove the definition had teeth he constructed the family now called Werner states: highly symmetric mixtures of a maximally entangled state with white noise that are provably entangled yet, for a range of parameters, admit an explicit local-hidden-variable model reproducing all projective-measurement statistics. That single example dissolved the naive hope that entanglement and Bell nonlocality are the same thing — some entangled states are local. Werner's convex-geometric framing set the agenda for everything that followed: the Peres–Horodecki criterion, entanglement witnesses, distillation, and the whole resource theory of entanglement are built on the separable-versus-entangled partition he made precise.

History

1935

— Einstein, Podolsky and Rosen challenge completeness; Schrödinger names the phenomenon Verschränkung (entanglement)

1935

— Schrödinger analyzes probability relations between separated systems, isolating the pure-state case

1964

— Bell derives an inequality obeyed by all local-hidden-variable models

1989

— Werner defines separable versus entangled mixed states and constructs the Werner states

1996

— Peres proposes the partial-transpose test; the Horodeckis prove it is sufficient in 2×2 and 2×3

1998

— Horodeckis exhibit bound-entangled PPT states, showing PPT is only necessary in higher dimensions

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §2.2.8 (Schmidt decomposition) and §12.5 (entanglement) · Preskill, Lecture Notes on Quantum Computation, Ch. 4 · Watrous, The Theory of Quantum Information, Ch. 6 (separability and PPT) · Wilde, Quantum Information Theory, Ch. 4 and 11 · Kitaev, Shen, Vyalyi, Classical and Quantum Computation, Ch. 11 · Horodecki, Horodecki, Horodecki, Horodecki, "Quantum entanglement", Rev. Mod. Phys. 81, 865 (2009). Key numerical anchors: the rank-one contradiction for $|\Phi^+\rangle$ yields the marginal $\rho_A=\tfrac12\mathbb{1}$ with $S=1$ bit; the entropy for weights $\{0.9,0.1\}$ is $S=0.469$ bits; the Werner partial-transpose eigenvalues are $\{\tfrac{1+p}{4}\times3,\ \tfrac{1-3p}{4}\}$ with threshold $p=\tfrac13$ (at $p=0.2$ the critical eigenvalue is $+0.10$ ; at $p=0.8$ it is $-0.35$ ).

Prerequisites. Dirac Bra-Ket Notation · The Qubit and Its State Space · Inner Product, Norm, and Normalization · Density Operators · Superposition Principle. Leads to. Schmidt Decomposition · Entanglement Measures · Bell States and Inequalities · Quantum Teleportation · Peres–Horodecki / PPT Criterion.

Ask the tutor: Ask me to walk through the partial transpose of any two-qubit state, or to show why a specific state sits inside or outside the separable body.

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

A product state is the special case; a separable state is the general one. A product state factorizes completely — $|\alpha\rangle\otimes|\beta\rangle$ for pure states, or $\rho_A\otimes\rho_B$ for mixed — so there is literally no correlation of any kind between the two subsystems. A separable state is any convex mixture of products, $\sum_i p_i\,\rho_A^{(i)}\otimes\rho_B^{(i)}$ . Such a state can be strongly correlated, but only classically: the correlation lives entirely in the shared label $i$ . Every product state is separable, but most separable states are not products. Both, crucially, are preparable without entanglement.

Why is separability so much harder to test for mixed states than for pure states?

For a pure state you compute one marginal and look: a pure marginal means product, a mixed marginal means entangled, and the Schmidt rank settles it in one step. A mixed state has no such shortcut because a single $\rho$ can be written as many different ensembles of products, and to certify separability you must show that at least one product decomposition exists among infinitely many candidates. Searching that space is a hard convex-geometry problem — proven NP-hard in the dimension. So instead of an exact algorithm one uses one-sided witnesses like PPT, which can prove entanglement quickly but cannot always prove separability.

Does a negative partial transpose always mean the state is entangled?

Yes — that direction is airtight in every dimension. If $\rho$ were separable, its partial transpose would be a convex sum of $\rho_A^{(i)}\otimes(\rho_B^{(i)})^{T}$ , each a valid positive operator, so $\rho^{T_B}\ge0$ necessarily. A single negative eigenvalue therefore rules out separability and certifies entanglement. What is not always true is the converse: a positive partial transpose does not guarantee separability once $d_A d_B>6$ , where bound-entangled PPT states live. So "NPT $\Rightarrow$ entangled" is universal, while "PPT $\Rightarrow$ separable" holds only in $2\times2$ and $2\times3$ .

Are entanglement and Bell nonlocality the same thing?

No, and Werner's own construction is the counterexample. Every state that violates a Bell inequality must be entangled, because separable states factorize their correlators and obey the CHSH bound of $2$ . But the reverse fails: Werner exhibited entangled mixed states that nonetheless admit an explicit local-hidden-variable model reproducing all projective-measurement statistics, so they are entangled yet Bell-local. Entanglement is the broader property — a statement about how the state is built — while Bell nonlocality is a statement about measurement statistics that only some entangled states display. Later work also separates both from quantum steering, a third notion sitting strictly between them.

Why can't local operations and classical communication create entanglement?

Because every LOCC step preserves the convex-mixture-of-products form. A local quantum channel acts as $\rho_A^{(i)}\mapsto\mathcal{E}_A(\rho_A^{(i)})$ on one factor and leaves the other alone, sending products to products. Classical communication only correlates or refines the shared label $i$ , which is exactly the ingredient that builds convex mixtures. So starting from a separable state and applying any sequence of local operations and classical messages, the output is still $\sum_i p_i\,\rho_A^{(i)}\otimes\rho_B^{(i)}$ — separable. This is the precise sense in which entanglement is a resource: it can be consumed, converted, or degraded locally, but never manufactured from scratch without a quantum interaction.

What is bound entanglement, and why does it matter?

Bound entanglement is entanglement you cannot get back out as pure singlets. In dimensions with $d_A d_B>6$ there exist entangled states whose partial transpose is positive — PPT-entangled states — and no LOCC protocol can distill even a single maximally entangled pair from any number of copies of them. They are genuinely entangled, yet the standard entanglement currency is frozen inside. This matters because it breaks the tidy picture where entanglement is one fungible resource: distillable and bound entanglement behave differently, PPT no longer detects entanglement, and stronger witnesses are needed. Remarkably, some bound-entangled states still support secure key distribution, so even undistillable entanglement can do useful cryptographic work.