PPT Criterion and the Peres–Horodecki Test

Deciding whether a bipartite mixed state $\rho_{AB}$ is separable — writable as a probabilistic mixture of product states $\sum_k p_k\,\rho_A^{(k)}\otimes\rho_B^{(k)}$ — is one of the genuinely hard problems of quantum information, NP-hard in the dimension of the state space. Yet there is a test so cheap it fits on a napkin: transpose one half of the density matrix and look at the eigenvalues. If any is negative, the state is entangled. Asher Peres proposed exactly this in 1996, and within months the three Horodeckis — Michał, Paweł, and Ryszard — pinned down precisely when it captures all entanglement, tying the whole question to the abstract theory of positive maps. The conceptual heart is a single asymmetry. Transposition $T$ is a positive map: it sends any legitimate density matrix to another one, with the same eigenvalues. But it is not completely positive: extended to act on only half of a larger system, $\mathbb{1}\otimes T$ can produce an operator with negative eigenvalues. A separable state is a mixture of products, and on each product $\mathbb{1}\otimes T$ merely transposes the $B$ factor into another valid state — so separability forces the partial transpose $\rho^{T_B}\ge 0$ . Entangled states can break this, and the maximally entangled ones break it maximally. The Horodecki theorem then reframes separability through the whole family of positive maps at once, and shows that in the two smallest cases — two qubits ( $2\times2$ ) and qubit–qutrit ( $2\times3$ ) — transposition alone decides everything. From $3\times3$ and $2\times4$ upward, entangled states that survive the partial transpose appear: the strange, undistillable bound entangled states. Beyond the yes/no verdict, the partial transpose also furnishes a computable entanglement measure, the negativity, placing the criterion at the working centre of entanglement detection.

Key takeaways

The partial transpose $\rho^{T_B}$ transposes only subsystem $B$ of a bipartite density matrix; if $\rho$ is separable then $\rho^{T_B}\ge 0$ , so any negative eigenvalue is a certificate of entanglement.
This positive-partial-transpose (PPT) condition is the Peres necessary criterion — one of the most-used entanglement tests in the field.
The Horodeckis proved PPT is also sufficient for separability exactly in dimensions $2\times2$ and $2\times3$ , and only there, by connecting separability to the full set of positive maps.
Above $2\times3$ there exist PPT entangled (bound entangled) states, whose entanglement cannot be distilled by any LOCC protocol — PPT misses them.
Quantitatively the negativity $\mathcal{N}(\rho)=\tfrac12(\lVert\rho^{T_B}\rVert_1-1)$ turns the sign test into an efficiently computable entanglement monotone.
Key idea: transposition is positive but not completely positive, so a separable state stays positive under $\mathbb{1}\otimes T$ while an entangled one can go negative — making a negative eigenvalue of $\rho^{T_B}$ an unambiguous signature of entanglement: necessary in every dimension, sufficient only in $2\times2$ and $2\times3$ .

The core insight

Transposition is positive but not completely positive: a separable state stays positive under the partial transpose $\mathbb{1}\otimes T$ , so a negative eigenvalue of $\rho^{T_B}$ is an unambiguous signature of entanglement — necessary in every dimension, sufficient only in $2\times2$ and $2\times3$ .

Where this lives today

The PPT criterion and its negativity are the default first move in experimental entanglement verification. Given a density matrix reconstructed by quantum-state tomography — from photon pairs, trapped ions, superconducting transmons, or nitrogen-vacancy centres — a laboratory computes $\rho^{T_B}$ and its eigenvalues in microseconds; a negative one is a model-independent certificate, and the negativity reports how much noise the entanglement can survive. In continuous-variable optics the same idea, specialised to Gaussian states, becomes the Simon–Duan criterion: partial transpose is a sign flip of one mode's momentum, and logarithmic negativity from symplectic eigenvalues is the entanglement measure for squeezed-light and optomechanical experiments. Bound entanglement — PPT states that are nonetheless entangled — has moved from curiosity to resource, with photonic demonstrations of its preparation and its role in secure key distribution. And the criterion still marks one of the field's most stubborn open problems: whether every entangled state that does violate PPT can be distilled, or whether there exist NPT bound entangled states — a question sharpened around 2000 that remains unresolved today.

The physical picture

Transposition is the operator version of time reversal: it flips the sign of every momentum and reverses the order of a product, yet it leaves any single density matrix a legitimate state, because it does not touch the eigenvalues. The trouble starts when you reverse time on only half of an entangled pair. A separable state is a recipe — a bag of independently prepared products $\rho_A^{(k)}\otimes\rho_B^{(k)}$ — and transposing the $B$ factor of each just swaps one valid local state for another, so the whole mixture stays a valid state. Entanglement is exactly the correlation that has no such product recipe, and it is precisely what the one-sided time reversal can expose: an eigenvalue — a would-be probability — is driven negative. The first figure shows this happening for the archetype, the Bell state, whose partial transpose carries a single $-\tfrac12$ eigenvalue.

Figure

Partial transpose of the Bell state Φ⁺: three eigenvalues sit at +1/2 and one is driven to −1/2. The lone negative eigenvalue certifies entanglement, and its magnitude is the negativity N = 1/2, the maximal value for two qubits

Not every positive map is up to the job in every dimension, and this is where the criterion acquires its subtlety. Think of a positive map as one specific "light" you can shine on a state: separable states stay bright (positive) under all of them, but an entangled state throws a shadow (goes negative) under at least one. Transposition is a single, especially convenient light. The Horodeckis showed that in $2\times2$ and $2\times3$ that one light is enough — every entangled state is caught by it — whereas from $3\times3$ upward some entangled states are camouflaged and need other, indecomposable lights to be seen. The second figure turns the amount of negativity into a meter, tracing the exact boundary at which the two-qubit Werner state crosses from separable to entangled as its smallest partial-transpose eigenvalue $(1-3p)/4$ passes through zero at $p=\tfrac13$ .

Figure

The two-qubit Werner state ρ_W(p) = p·Φ⁺⟨Φ⁺ + (1−p)/4·𝟙. Its smallest partial-transpose eigenvalue is λ_min = (1−3p)/4, which crosses zero at p = 1/3: separable for p ≤ 1/3, entangled for p > 1/3, and because it is 2×2 this boundary is exact

The core equations

The starting object is the partial transpose itself: transpose the matrix indices of subsystem $B$ only, leaving $A$ untouched. In a fixed product basis this is nothing more than swapping the two $B$ -labels of every matrix element, which is what makes it so cheap to compute.

\rho_{AB}^{\,T_B}=(\mathbb{1}_A\otimes T)\,\rho_{AB},\qquad \langle i\alpha|\rho^{T_B}|j\beta\rangle=\langle i\beta|\rho|j\alpha\rangle \qquad (1)

where: $\rho_{AB}$ — bipartite density matrix on $\mathcal{H}_A\otimes\mathcal{H}_B$ ; $T$ — transposition on system $B$ in a fixed basis; $i,j$ — basis indices of subsystem $A$ ; $\alpha,\beta$ — basis indices of subsystem $B$ . Reading it: only the $B$ -indices $\alpha,\beta$ exchange places, so $\rho^{T_B}$ is still Hermitian and still has unit trace — but no longer need be positive. The Peres criterion is the one-line consequence for separable states. Because the partial transpose is linear and transposition preserves the spectrum of each local factor, a mixture of products maps to another mixture of positive operators.

\rho_{AB}=\sum_k p_k\,\rho_A^{(k)}\otimes\rho_B^{(k)} \;\Longrightarrow\; \rho_{AB}^{T_B}=\sum_k p_k\,\rho_A^{(k)}\otimes\big(\rho_B^{(k)}\big)^{T}\ge 0 \qquad (2)

where: $p_k$ — probabilities, $p_k\ge0,\ \sum_k p_k=1$ ; $\rho_A^{(k)},\rho_B^{(k)}$ — local density matrices of the product decomposition; $(\cdot)^{T}$ — matrix transpose, which preserves both positivity and trace. Reading it: separability implies PPT, so the contrapositive — a negative eigenvalue implies entanglement — holds in every dimension. The Horodecki theorem lifts this from one map to all of them, and that is what makes it a full characterisation rather than a one-sided test. Separability is equivalent to surviving every positive map applied to one side.

\rho \text{ separable} \iff (\mathbb{1}\otimes\Lambda)\,\rho\ge 0 \ \ \text{for all positive maps } \Lambda \qquad (3)

where: $\Lambda$ — any positive linear map on operators of system $B$ ; $\mathbb{1}\otimes\Lambda$ — the map acting trivially on $A$ and by $\Lambda$ on $B$ . Reading it: transposition is just one such $\Lambda$ , so PPT is always necessary; whether it is also sufficient depends on whether transposition can stand in for the whole family. In the two smallest bipartite dimensions it can, because there every positive map decomposes into completely positive pieces plus transposition. This is the Størmer–Woronowicz result, and it collapses the infinite family of conditions in (3) to the single check $\rho^{T_B}\ge0$ .

\rho \text{ separable in } 2\times2 \text{ or } 2\times3 \iff \rho^{T_B}\ge 0 \qquad (4)

where: $2\times2$ — two qubits, $\dim=4$ ; $2\times3$ — qubit and qutrit, $\dim=6$ . Reading it: these are the only dimensions where a single transpose settles separability outright; $2\times3$ is the last one. To go beyond a verdict, the partial transpose supplies a magnitude — the negativity, the absolute sum of the negative eigenvalues, equivalently half the amount by which the trace norm exceeds one.

\mathcal{N}(\rho)=\frac{\lVert\rho^{T_B}\rVert_1-1}{2}=\sum_{\lambda_i<0}|\lambda_i|,\qquad E_{\mathcal{N}}(\rho)=\log_2\lVert\rho^{T_B}\rVert_1 \qquad (5)

where: $\lVert X\rVert_1=\mathrm{tr}\sqrt{X^\dagger X}=\sum_i|\lambda_i|$ — trace norm; $\lambda_i$ — the (real) eigenvalues of $\rho^{T_B}$ ; $\mathcal{N}$ — negativity, with $\mathcal{N}=0\Leftrightarrow$ PPT; $E_{\mathcal{N}}$ — logarithmic negativity, an upper bound on distillable entanglement. Reading it: $\mathcal{N}$ is an entanglement monotone computable in $O(d^3)$ time, so PPT doubles as a practical measure. Two benchmark states anchor the whole machinery. The maximally entangled Bell state realises the largest possible two-qubit negativity, and the Werner family sets the exact separability boundary for two qubits.

\operatorname{spec}\big[(|\Phi^+\rangle\langle\Phi^+|)^{T_B}\big]=\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\},\qquad \rho_W(p)=p\,|\Phi^+\rangle\langle\Phi^+|+\tfrac{1-p}{4}\,\mathbb{1}_4 \ \text{ entangled}\iff p>\tfrac13 \qquad (6)

where: $|\Phi^+\rangle=\tfrac1{\sqrt2}(|00\rangle+|11\rangle)$ — Bell state; $p$ — mixing weight of the Bell component, $0\le p\le1$ ; $\mathbb{1}_4$ — the maximally mixed two-qubit state. Reading it: the Bell state gives $\mathcal{N}=\tfrac12$ , $E_{\mathcal{N}}=1$ (one ebit), and mixing in white noise erodes that entanglement until it vanishes exactly at $p=\tfrac13$ .

Derivation ASeparable states have a positive partial transpose

The goal is to establish the Peres criterion — that separability implies $\rho^{T_B}\ge0$ — in full generality, valid in every dimension and for every number of terms in the mixture, using nothing but the convex definition of separability and the elementary fact that the matrix transpose preserves the spectrum of a density matrix. This is the direction of the implication that makes a negative eigenvalue a genuine, dimension-independent certificate of entanglement, and it is worth seeing exactly which property of transposition is doing the work.

Step 1. Write the separable state in its defining form, $\rho=\sum_k p_k\,\rho_A^{(k)}\otimes\rho_B^{(k)}$ with $p_k\ge0$ and $\sum_k p_k=1$ , where each $\rho_A^{(k)},\rho_B^{(k)}$ is a legitimate local density matrix. Why: this convex-combination-of-products form is the precise meaning of "not entangled", and the whole proof is just tracking what the partial transpose does to it.

Step 2. Apply $\mathbb{1}\otimes T$ term by term, using linearity, to get $\rho^{T_B}=\sum_k p_k\,\rho_A^{(k)}\otimes\big(\rho_B^{(k)}\big)^{T}$ . Why: the partial transpose touches only the $B$ factor of each product, so it cannot mix the terms of the sum — it acts inside each product independently.

Step 3. Observe that $\big(\rho_B^{(k)}\big)^{T}$ has exactly the same eigenvalues as $\rho_B^{(k)}$ (transposition never changes a matrix's spectrum), so it is again a positive, unit-trace density matrix. Why: this is the one place complete positivity is not needed — for a stand-alone matrix, ordinary transposition is harmless, which is exactly why a single local state survives it.

Step 4. Conclude that each term $p_k\,\rho_A^{(k)}\otimes\big(\rho_B^{(k)}\big)^{T}$ is a nonnegative multiple of a tensor product of positive operators, hence positive, and a sum of positive operators is positive: $\rho^{T_B}\ge0$ . Why: positivity is preserved under tensor products and nonnegative sums, so the entangled-detecting power comes entirely from states that have no such product decomposition in the first place. The logic verifies directly: each transposed factor is PSD with the same spectrum, tensor products of PSD operators are PSD, and convex combinations of PSD operators are PSD — so any $\rho^{T_B}$ with a negative eigenvalue cannot have arisen from a separable $\rho$ .

Go deeper · Derivation B · 1 minThe Bell state and its single negative eigenvalue

Derivation BThe Bell state and its single negative eigenvalue

The goal is to compute the partial transpose of the maximally entangled two-qubit state $|\Phi^+\rangle=\tfrac1{\sqrt2}(|00\rangle+|11\rangle)$ explicitly in the computational basis, extract its full spectrum $\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ , and read off the maximal two-qubit negativity. This is the concrete calculation behind the first figure, and it shows both how the index-swap rule of equation (1) operates in practice and why the maximally entangled state violates PPT as strongly as any two-qubit state can.

Step 2. Apply $T_B$ using the swap rule $\langle i\alpha|\rho^{T_B}|j\beta\rangle=\langle i\beta|\rho|j\alpha\rangle$ : the term $|00\rangle\langle11|=|0_A0_B\rangle\langle1_A1_B|$ becomes $|0_A1_B\rangle\langle1_A0_B|=|01\rangle\langle10|$ , and likewise $|11\rangle\langle00|\to|10\rangle\langle01|$ . Why: swapping only the $B$ -labels moves the off-diagonal weight from the $\{00,11\}$ pair into the $\{01,10\}$ pair — a completely different block of the matrix.

Step 3. Assemble $\rho^{T_B}=\tfrac12\big(|00\rangle\langle00|+|11\rangle\langle11|+|01\rangle\langle10|+|10\rangle\langle01|\big)$ : two diagonal entries of $\tfrac12$ on $|00\rangle,|11\rangle$ , and an off-diagonal $2\times2$ block coupling $|01\rangle$ and $|10\rangle$ . Why: the diagonal pieces are trivially positive; all the action is in the little exchange block, which is where a negative eigenvalue can hide.

Step 4. Diagonalise that block, $\tfrac12\left(\begin{smallmatrix}0&1\\1&0\end{smallmatrix}\right)$ , whose eigenvalues are $\pm\tfrac12$ with eigenvectors $\tfrac1{\sqrt2}(|01\rangle\pm|10\rangle)$ ; combined with the two $\tfrac12$ 's from the diagonal this gives the full spectrum $\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ . Why: the antisymmetric combination $\tfrac1{\sqrt2}(|01\rangle-|10\rangle)$ — the singlet — carries the negative eigenvalue, the operator ghost of the entanglement.

Step 5. Read off the negativity $\mathcal{N}=|-\tfrac12|=\tfrac12$ and logarithmic negativity $E_{\mathcal{N}}=\log_2(2\cdot\tfrac12+1)=\log_2 2=1$ . Why: these are the maximal values attainable by any two-qubit state — one ebit — confirming that the maximally entangled state is also maximally PPT-violating. The spectrum checks against trace preservation: $\tfrac12+\tfrac12+\tfrac12-\tfrac12=1=\mathrm{tr}\,\rho^{T_B}$ , and a direct numerical diagonalisation returns eigenvalues $\{-0.5,0.5,0.5,0.5\}$ with $\mathcal{N}=0.5$ , $E_{\mathcal{N}}=1$ .

Go deeper · Derivation C · 1 minThe Werner-state separability threshold

Derivation CThe Werner-state separability threshold

The goal is to find the exact mixing parameter at which the two-qubit Werner state $\rho_W(p)=p\,|\Phi^+\rangle\langle\Phi^+|+\tfrac{1-p}{4}\mathbb{1}_4$ crosses from separable to entangled, by tracking the smallest eigenvalue of its partial transpose down to zero. Because the system is $2\times2$ , sufficiency (equation 4) guarantees the PPT boundary is the true separability boundary — not merely a necessary condition — so this single calculation yields the complete phase diagram drawn in the second figure.

Step 1. Partial-transpose the mixture, using that the identity is invariant under $T_B$ : $\rho_W^{T_B}=p\,(|\Phi^+\rangle\langle\Phi^+|)^{T_B}+\tfrac{1-p}{4}\mathbb{1}_4$ . Why: the white-noise part contributes only a uniform shift, so all the structure comes from the Bell part whose spectrum we already know.

Step 2. Shift the known Bell spectrum by the identity term: the eigenvalues become $\lambda_i=p\,\mu_i+\tfrac{1-p}{4}$ with $\mu_i\in\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ . Why: adding a multiple of the identity moves every eigenvalue up by the same amount without rotating any eigenvector, so we can reuse the previous derivation wholesale.

Step 3. Pick out the smallest one, coming from $\mu=-\tfrac12$ : $\lambda_{\min}=p\cdot(-\tfrac12)+\tfrac{1-p}{4}=\tfrac{1-3p}{4}$ . Why: only this eigenvalue can ever go negative, so the entire separability question reduces to the sign of a single linear function of $p$ .

Step 4. Impose positivity: $\lambda_{\min}\ge0\iff 1-3p\ge0\iff p\le\tfrac13$ ; for $p>\tfrac13$ the state is PPT-violating and therefore entangled, and for $p\le\tfrac13$ it is PPT and — since $2\times2$ — genuinely separable. Why: the exact threshold $p=\tfrac13$ is a hard boundary, not an estimate, precisely because sufficiency holds in this dimension. Numerically, $\lambda_{\min}$ evaluates to $+0.025$ at $p=0.30$ , exactly $0$ at $p=\tfrac13$ , and $-0.005$ at $p=0.34$ , each matching $(1-3p)/4$ to the digit — confirming the threshold and its sign change.

Limiting cases

Case	Result	Meaning
$2\times2$ (two qubits)	$\rho\text{ separable}\iff\rho^{T_B}\ge0$	PPT is necessary AND sufficient; equivalent to Wootters' concurrence being zero
$2\times3$ (qubit–qutrit)	$\rho\text{ separable}\iff\rho^{T_B}\ge0$	Still exactly sufficient — the last dimension where transposition alone decides separability
$3\times3,\ 2\times4$ and higher	$\rho^{T_B}\ge0\not\Rightarrow\text{separable}$	PPT entangled (bound entangled) states exist; PPT is necessary but no longer sufficient
Pure states, any dimension	$\|\psi\rangle\langle\psi\|\ \text{PPT}\iff\text{product}$	Any entangled pure state (Schmidt rank $>1$ ) violates PPT — no bound entanglement among pure states
Choice of subsystem	$\operatorname{spec}(\rho^{T_A})=\operatorname{spec}(\rho^{T_B})$	It never matters which half you transpose; the verdict and the negativity are identical

Common trapPPT means separable

This is the single most common error, and it is only true in the two smallest dimensions. PPT (positive partial transpose) is necessary for separability everywhere, but sufficient only for $2\times2$ and $2\times3$ . From $3\times3$ and $2\times4$ upward there exist PPT entangled states — the bound entangled states — so concluding " $\rho^{T_B}\ge0$ , therefore separable" is wrong in those dimensions and will silently misclassify real states. To settle separability above $2\times3$ you need stronger tools: indecomposable entanglement witnesses (built from indecomposable positive maps such as the Choi map in $3\times3$ ), the realignment / CCNR criterion, or the range criterion. A second, harmless subtlety worth stating so it never trips you: $\rho^{T_A}$ and $\rho^{T_B}$ always share the same spectrum, so it is irrelevant which subsystem you transpose — but this is a convenience, not a loophole, and it does nothing to rescue sufficiency above $2\times3$ .

Worked examples

1 · Werner state: certify and quantify the entanglement (fully worked)

Take the two-qubit Werner state $\rho_W(p)=p\,|\Phi^+\rangle\langle\Phi^+|+\tfrac{1-p}{4}\mathbb{1}_4$ at $p=0.6$ . Decide whether it is entangled, and if so, compute its negativity and logarithmic negativity.

Step 1. Get the partial-transpose spectrum by shifting the Bell spectrum $\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ : each eigenvalue is $\lambda_i=p\,\mu_i+\tfrac{1-p}{4}$ . With $p=0.6$ the uniform offset is $\tfrac{1-0.6}{4}=0.1$ , so the three positive eigenvalues are $0.6\cdot\tfrac12+0.1=\mathbf{0.4}$ each, and the lone candidate is $\lambda_{\min}=0.6\cdot(-\tfrac12)+0.1=-0.3+0.1=\mathbf{-0.2}$ .

Step 2. Apply the criterion. Since $p=0.6>\tfrac13$ we expect entanglement, and indeed $\lambda_{\min}=-0.2<0$ , so $\rho^{T_B}\not\ge0$ : the state is entangled. This matches $(1-3p)/4=(1-1.8)/4=-0.2$ .

Step 3. Compute the negativity as the absolute sum of negative eigenvalues: there is exactly one, so $\mathcal{N}=|-0.2|=\mathbf{0.2}$ . Equivalently via the trace norm, $\lVert\rho^{T_B}\rVert_1=3(0.4)+0.2=1.4$ , giving $\mathcal{N}=\tfrac{1.4-1}{2}=0.2$ — the two forms agree.

Step 4. Logarithmic negativity: $E_{\mathcal{N}}=\log_2\lVert\rho^{T_B}\rVert_1=\log_2 1.4=\mathbf{0.485}$ ebits. This is well below the Bell value of $1$ , as expected for a noisy state that is only $60\%$ Bell. Verdict: entangled, with $\mathcal{N}=0.2$ and $E_{\mathcal{N}}\approx0.485$ ebits.

2 · A product state passes cleanly

Check the criterion on the manifestly separable state $\rho=|0\rangle\langle0|\otimes|+\rangle\langle+|$ , where $|+\rangle=\tfrac1{\sqrt2}(|0\rangle+|1\rangle)$ .

Step 1. The $B$ factor is $|+\rangle\langle+|=\tfrac12\left(\begin{smallmatrix}1&1\\1&1\end{smallmatrix}\right)$ , a real symmetric matrix, so it equals its own transpose: $\big(|+\rangle\langle+|\big)^{T}=|+\rangle\langle+|$ .

Step 2. Since the state is a single product, the partial transpose is $\rho^{T_B}=|0\rangle\langle0|\otimes\big(|+\rangle\langle+|\big)^{T}=|0\rangle\langle0|\otimes|+\rangle\langle+|=\rho$ itself — completely unchanged.

Step 3. Its eigenvalues are those of the original: a single $1$ (for $|0\rangle|+\rangle$ ) and three $0$ 's, all $\ge0$ . So $\rho^{T_B}\ge0$ , $\mathcal{N}=0$ , and the state is (correctly) certified not entangled — consistent with the Peres criterion, since every product state is trivially separable.

Physics of the portrait

The PPT criterion is the rare result whose elegance is inseparable from the person who saw it: a one-line test born from a physicist who spent his career insisting that quantum theory be stated in terms of what can actually be measured.

Origins · 1 minAsher Peres and how this was discovered

Asher Peres

1934–2005 · entanglement and quantum information

Asher Peres was an Israeli theoretical physicist at the Technion in Haifa who moved fluidly between general relativity and the foundations of quantum mechanics, and who helped invent the field of quantum information before it had a name. He co-authored the 1993 quantum teleportation protocol and produced a stream of sharp, foundational results, but the separability criterion of 1996 is his most-cited single idea. Peres asked a disarmingly simple question — what happens to a density matrix if you transpose only one of its two subsystems? — and realised that separable states must survive it while entangled ones can fail, turning an abstract question about convex mixtures into a two-minute eigenvalue calculation. He conjectured that PPT might detect all entanglement; the Horodeckis proved him exactly right in $2\times2$ and $2\times3$ and exactly wrong beyond, discovering bound entanglement in the process. Famous for his textbook Quantum Theory: Concepts and Methods and for aphorisms that doubled as physics — "unperformed experiments have no results" — he championed an operational reading of quantum mechanics throughout. His criterion remains, three decades on, the first thing anyone computes when handed an unknown bipartite state.

History

1963

— Størmer characterises positive maps on low-dimensional matrix algebras

1976

— Woronowicz completes the decomposability result for 2×2 and 2×3

1989

— Werner defines the states that separate classical correlation from entanglement

1996

— Peres proposes the positive-partial-transpose criterion for separability

1996

— the Horodeckis prove PPT is necessary and sufficient exactly in 2×2 and 2×3

1998

— Horodeckis discover bound entangled PPT states; negativity is introduced as a measure

2002

— Vidal and Werner establish the logarithmic negativity as a computable monotone

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §12.5 (separability and the partial transpose) · Preskill, Lecture Notes on Quantum Information, Ch. 4 · Watrous, The Theory of Quantum Information, Ch. 6 (separability and positive maps) · Wilde, Quantum Information Theory, Ch. 4 · Bengtsson & Życzkowski, Geometry of Quantum States, Ch. 16.

Papers. Peres, Separability Criterion for Density Matrices, Phys. Rev. Lett. 77, 1413 (1996) · Horodecki, Horodecki & Horodecki, Separability of Mixed States: Necessary and Sufficient Conditions, Phys. Lett. A 223, 1 (1996) · Horodecki, Horodecki & Horodecki, Mixed-State Entanglement and Distillation: Is There a Bound Entanglement in Nature?, Phys. Rev. Lett. 80, 5239 (1998) · Życzkowski, Horodecki, Sanpera & Lewenstein, Phys. Rev. A 58, 883 (1998) · Vidal & Werner, Computable Measure of Entanglement, Phys. Rev. A 65, 032314 (2002). The key numerical benchmarks are: the Bell partial-transpose spectrum $\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ , negativity $\tfrac12$ , log-negativity $1$ ; the Werner threshold from $\lambda_{\min}=(1-3p)/4$ , giving $p=\tfrac13$ ; the trace-norm identity $\mathcal{N}=\tfrac12(\lVert\rho^{T_B}\rVert_1-1)$ ; and the worked $p=0.6$ Werner values $\mathcal{N}=0.2$ , $E_{\mathcal{N}}=\log_2 1.4\approx0.485$ . All are confirmed by direct numerical diagonalisation.

Prerequisites. Separable versus Entangled States · Schmidt Decomposition · Bell States and EPR Pairs. Leads to. Entanglement Witnesses · Concurrence and Wootters' Formula · Entanglement of Formation.

Ask the tutor: Hand me any bipartite density matrix and I will walk you through its partial transpose, its negativity, and whether PPT settles the question in your dimension.

What exactly does "partial" mean in partial transpose?

It means you transpose the matrix indices of only one subsystem while leaving the other alone. A bipartite operator has four indices — two for $A$ and two for $B$ — and the full transpose swaps both pairs at once. The partial transpose $T_B$ swaps only the $B$ pair, $\langle i\alpha|\rho^{T_B}|j\beta\rangle=\langle i\beta|\rho|j\alpha\rangle$ , which in block form means transposing each $B$ -block in place. The result is still Hermitian and still has unit trace, so it looks like a density matrix in every respect except one: it may have a negative eigenvalue, and that is the whole point.

Why does transposing one subsystem detect entanglement at all?

Because transposition is positive but not completely positive. On a single system it is harmless — it preserves eigenvalues, so any density matrix stays a density matrix. But "completely positive" is the stronger demand that it stay positive even when applied to half of a larger entangled system, and transposition fails that test. Separable states are built from products, on which the one-sided transpose only rewrites a valid local state into another valid one, so they always stay positive. Entangled states have correlations with no product decomposition, and those correlations can be flipped into a negative eigenvalue. The gap between "positive" and "completely positive" is exactly the gap between separable and entangled.

What is a bound entangled state?

It is an entangled state whose entanglement cannot be distilled — no amount of local operations and classical communication can concentrate it into pure Bell pairs. In $3\times3$ and larger dimensions there exist entangled states that nonetheless satisfy PPT, and a PPT state is provably undistillable. So these states carry genuine entanglement that is "locked up": real, but useless for the standard tasks that consume Bell pairs, such as teleportation. Their existence is why PPT stops being sufficient above $2\times3$ .

Does it matter whether I transpose system A or system B?

No — the verdict and the negativity are identical either way. The two partial transposes are related by a full transpose, $\rho^{T_A}=(\rho^{T_B})^{T}$ , and a full transpose never changes eigenvalues. So $\rho^{T_A}$ and $\rho^{T_B}$ always have exactly the same spectrum; one is positive if and only if the other is, and both give the same negativity. Choose whichever subsystem has the smaller dimension to lighten the arithmetic — the physics is indifferent to the choice.

How is the negativity different from just saying "entangled or not"?

The sign of the smallest eigenvalue gives a yes/no answer; the negativity gives a magnitude. Formally $\mathcal{N}(\rho)=\sum_{\lambda_i<0}|\lambda_i|=\tfrac12(\lVert\rho^{T_B}\rVert_1-1)$ , the total negative weight in the partial-transpose spectrum. It is an entanglement monotone — it cannot increase under local operations and classical communication — so it is a legitimate measure of how much entanglement is present, not merely whether some is. It is also efficiently computable for any state, unlike the entanglement of formation, which is why it is the workhorse quantitative measure in experiments; its logarithm $E_{\mathcal{N}}=\log_2\lVert\rho^{T_B}\rVert_1$ upper-bounds the distillable entanglement.

If PPT misses bound entanglement, why is it still the standard test?

Because it is cheap, universal, and exact where it counts most. For the two-qubit and qubit-qutrit systems that dominate real experiments, PPT is not just a hint but a complete answer — necessary and sufficient. For larger systems it remains necessary, so a PPT violation is always a rigorous certificate of entanglement, and it costs only one matrix diagonalisation. What it cannot do is rule out entanglement above $2\times3$ : passing the test proves nothing there. So PPT is the correct first test everywhere and the final word in low dimensions; when it comes back positive in a large system, you reach for witnesses or the CCNR criterion to keep looking.

How does the PPT criterion relate to entanglement witnesses?

They are two faces of the same mathematics, linked by the Choi–Jamiołkowski isomorphism, which turns any positive-but-not-completely-positive map into a Hermitian operator — a witness — nonnegative on all separable states but negative on some entangled one. Transposition is the canonical such map, and the witness it generates is the most symmetric member of the family. Detecting the bound entanglement that PPT misses requires indecomposable witnesses built from indecomposable positive maps, such as the Choi map in $3\times3$ . So PPT is the simplest witness, and the general theory of entanglement witnesses is where this criterion naturally leads.