Distillable Entanglement and Bound Entanglement

The theory of entanglement is, at heart, an economics of conversion. A quantum state is a resource not because it is "spooky" but because it can be traded — under local operations and classical communication (LOCC) — into a single hard currency: the maximally entangled Bell pair, worth one ebit. This replaces the flat question "is $\rho$ entangled?" with two accounting questions. How many Bell pairs can you refine out of many copies of $\rho$ ? That optimal asymptotic yield is the distillable entanglement $E_D$ . And how many must you spend to manufacture $\rho$ ? That is the entanglement cost $E_C$ . For pure states the books balance perfectly: $E_D = E_C = S(\rho_A)$ , the entropy of entanglement, so pure-state entanglement is one reversibly tradeable number set by the Schmidt spectrum. Mixed states break the balance. In general $E_D \le E_C$ , and the inequality can be strict — a thermodynamic-style irreversibility in which entanglement, once poured into a state, can never be fully poured back out. The extreme case is the most startling object in the subject: bound entanglement — states that are provably entangled (a witness fires, they fail every separability test) yet have $E_D = 0$ . Not one clean Bell pair can ever be extracted from them by LOCC, no matter how many copies you hold. The bridge to a computable criterion is a theorem of the Horodecki family: any state with a positive partial transpose (PPT) is undistillable, because PPT survives LOCC while the Bell pair violates it. Every PPT entangled state — and they exist in $3\times3$ and $2\times4$ systems — is therefore bound. This article builds the operational definitions, proves the pure-state reversibility theorem, proves PPT-implies-undistillable, and writes down the canonical Horodecki state explicitly.

Key takeaways

Distillable entanglement $E_D(\rho)$ is the optimal asymptotic rate at which near-perfect Bell pairs can be extracted from many copies of $\rho$ using only LOCC; entanglement cost $E_C(\rho)$ is the reverse rate.
For pure states both collapse to the entropy of entanglement, $E_D = E_C = S(\rho_A) = -\sum_i \lambda_i \log_2 \lambda_i$ , so pure-state entanglement is reversible and uniquely quantified.
For mixed states $E_D \le E_C$ , and the gap can be strict — the irreversibility of entanglement manipulation.
Bound entangled states are entangled yet have $E_D = 0$ : no ebit is extractable, though creating them still costs entanglement, $E_C > 0$ .
Every PPT entangled state is bound, because PPT is preserved by LOCC while the Bell pair is NPT — its partial transpose carries the eigenvalue $-\tfrac12$ .
Bound entanglement is not inert: it activates distillation of other states and can yield secret key, so the operational landscape is richer than a single number.
Key idea: LOCC is the free operation and the Bell pair the unit of currency; distillable entanglement is the asymptotic out-rate of Bell pairs, and a bound entangled state is one with a positive out-rate of zero — entangled but locked — which every PPT entangled state must be, since positivity of the partial transpose survives LOCC while the Bell pair has a negative one.

The core insight

LOCC defines a free operation and the Bell pair a unit of currency. Distillable entanglement is the asymptotic out-rate of Bell pairs; bound entangled states are entangled states with zero out-rate, and every PPT entangled state is bound because positivity of the partial transpose survives LOCC while the Bell pair has a negative one.

Where this lives today

Bound entanglement began as a paradox and became a working part of the quantum-information toolkit. In quantum key distribution it produced its most striking surprise: certain bound entangled states, with $E_D = 0$ , still yield a nonzero distillable secret key — the private states of Horodecki, Horodecki, Horodecki and Oppenheim (2005), so key distillation and entanglement distillation are genuinely different tasks. Bound entanglement also refuses to be inert through activation: individually undistillable states can, combined, become distillable, and Smolin-type four-party bound states can be "unlocked" by classical communication. Experimentally, PPT entangled states have been prepared and certified with tailored non-decomposable witnesses in photonic platforms (Amselem–Bourennane 2009 and later realizations), precisely because the Peres–Horodecki test is blind to them. And the central open question stands: does NPT bound entanglement exist? If some NPT states are undistillable, the clean equivalence NPT $\Leftrightarrow$ distillable fails in high dimensions, and the hierarchy must be redrawn.

The physical picture

Think of entanglement as a fuel. Separable states are empty tanks — LOCC can never pour fuel in. Distillable states are tanks you can siphon into clean, standardized cans: run the right LOCC protocol on enough copies and out come Bell pairs, ready to teleport a qubit or key a cipher. Bound entangled states are the strange third kind: sealed tanks that demonstrably hold fuel — a witness lights up — yet the fuel is locked in a chemistry from which LOCC can pour nothing out. You paid entanglement to fill the tank ( $E_C > 0$ ), and you will never get an ebit back ( $E_D = 0$ ). The lock has a precise mechanism: a conserved quantity LOCC cannot violate. Positivity of the partial transpose is preserved by every LOCC map, and the Bell pair breaks it — its partial transpose owns the eigenvalue $-\tfrac12$ . You cannot manufacture negativity out of positivity with free operations, so a PPT state can never be nudged toward a Bell target. Distillability and PPT sit on opposite sides of an LOCC-invariant wall.

Figure

The three-shell hierarchy of bipartite mixed states: separable states in the green core, the amber shell of PPT-entangled bound states with E_D = 0 but E_C greater than 0, and the blue region of NPT distillable states; the Bell pair sits outside every wall with log-negativity 1 ebit and a partial-transpose eigenvalue of minus one half

Pure states escape the trap because they carry no wasted classical ignorance. Their entire content is the Schmidt spectrum $\{\lambda_i\}$ , which governs both extraction and creation, so the conversion runs reversibly at rate $S(\rho_A)$ . Mixing introduces classical uncertainty that can be irreversibly burned, opening the gap $E_D < E_C$ and, at the extreme, $E_D = 0 < E_C$ . The second figure makes the currency concrete: the LOCC engine that turns copies of $\rho$ into ebits, and the negative eigenvalue that powers the distinction.

Figure

The LOCC distillation engine and the negative eigenvalue that powers it: many copies of ρ are refined into Bell pairs at output rate E_D, or synthesized from Bell pairs at cost E_C, with E_D never exceeding E_C; the Bell state's partial transpose has spectrum one half, one half, one half, minus one half, giving negativity one half and log-negativity 1 ebit

The core equations

Distillation is asymptotic and error-tolerant: with $n$ copies you aim to output roughly $Rn$ Bell pairs, demanding only that the trace-norm error vanish as $n\to\infty$ . $E_D$ is the best such rate.

E_D(\rho) = \sup\Big\{ R : \lim_{n\to\infty}\Big( \inf_{\Lambda}\, \big\|\Lambda(\rho^{\otimes n}) - \Phi_2^{\otimes \lfloor Rn\rfloor}\big\|_1 \Big) = 0 \Big\} \qquad (1)

where: $\rho$ — bipartite state on $\mathcal{H}_A\otimes\mathcal{H}_B$ ; $\Lambda$ — an LOCC channel on all $n$ copies; $\Phi_2$ — the two-qubit Bell state, one ebit; $R$ — Bell-pair output rate per input copy; $\|\cdot\|_1$ — trace norm, the approximation error. Reading: $E_D$ is the largest yield at which the error still vanishes. Running the machine backwards defines the cost: how many Bell pairs must be fed in to build $\rho^{\otimes n}$ to vanishing error.

E_C(\rho) = \inf\Big\{ R : \lim_{n\to\infty}\Big( \inf_{\Lambda}\, \big\|\Lambda(\Phi_2^{\otimes \lceil Rn\rceil}) - \rho^{\otimes n}\big\|_1 \Big) = 0 \Big\} \qquad (2)

where: $R$ — Bell pairs consumed per copy produced; $\Lambda$ — LOCC channel synthesizing $\rho^{\otimes n}$ from Bell pairs. Reading: $E_C$ is the cheapest input rate that still succeeds. The cost is not independent; it equals the regularized entanglement of formation — the minimal average pure-state entanglement needed to mix $\rho$ , optimized over asymptotically many copies.

E_F(\rho) = \min_{\{p_k,|\psi_k\rangle\}} \sum_k p_k\, S\big(\mathrm{Tr}_B|\psi_k\rangle\langle\psi_k|\big),\qquad E_C(\rho) = \lim_{n\to\infty} \frac{1}{n}\, E_F(\rho^{\otimes n}) \qquad (3)

where: $\{p_k,|\psi_k\rangle\}$ — a pure-state decomposition of $\rho$ ; $S(\cdot)$ — von Neumann entropy of the reduced state; regularization is needed because $E_F$ is not additive (Hayden–Horodecki–Terhal). The two rates are never in the wrong order — you cannot get more out than you put in, or LOCC would create entanglement free.

E_D(\rho) \;\le\; E_C(\rho) \qquad (4)

where: $E_D$ — out-rate; $E_C$ — in-rate. A strict gap $E_D < E_C$ signals irreversibility; the maximal gap $E_D = 0 < E_C$ is bound entanglement. The central theorem ties distillability to a computable criterion — the partial transpose's sign.

\rho^{T_B}\ge 0 \;\;(\text{PPT}) \;\Longrightarrow\; E_D(\rho) = 0 \qquad (5)

where: $\rho^{T_B}$ — partial transpose on $B$ , elements $\langle i j'|\rho^{T_B}|i' j\rangle = \langle i j|\rho|i' j'\rangle$ ; PPT — the property $\rho^{T_B}\ge 0$ . Reading: positive partial transpose certifies undistillability. A computable upper bound on $E_D$ comes from the same object — the logarithmic negativity, the log trace-norm of $\rho^{T_B}$ .

E_N(\rho) = \log_2 \big\| \rho^{T_B} \big\|_1 = \log_2\Big( 1 + 2\!\!\sum_{\lambda_i<0}\!\! |\lambda_i| \Big) \qquad (6)

where: $\lambda_i$ — eigenvalues of $\rho^{T_B}$ ; $\mathcal{N}=\sum_{\lambda_i<0}|\lambda_i|$ is the negativity. Reading: $E_N \ge E_D$ always, and $E_N = 0$ on every PPT state — re-deriving $(5)$ . The Bell pair violates $(5)$ : it is NPT, its negativity the origin of one full ebit.

\Phi_2 = \tfrac12\big(|00\rangle+|11\rangle\big)\big(\langle00|+\langle11|\big),\qquad \mathrm{eig}\big(\Phi_2^{T_B}\big) = \{\,\tfrac12,\tfrac12,\tfrac12,-\tfrac12\,\} \qquad (7)

where: $T_B$ — partial transpose on the second qubit; the eigenvalue $-\tfrac12$ certifies NPT, negativity $\tfrac12$ , and $E_N = \log_2 2 = 1$ ebit. Finally, the object that proves $E_D$ and "entangled" differ — the canonical Horodecki $3\times3$ PPT entangled state, in the basis $|00\rangle,\ldots,|22\rangle$ .

\rho_a = \frac{1}{8a+1}\begin{pmatrix} a&0&0&0&a&0&0&0&a\\ 0&a&0&0&0&0&0&0&0\\ 0&0&a&0&0&0&0&0&0\\ 0&0&0&a&0&0&0&0&0\\ a&0&0&0&a&0&0&0&a\\ 0&0&0&0&0&a&0&0&0\\ 0&0&0&0&0&0&\tfrac{1+a}{2}&0&\tfrac{\sqrt{1-a^2}}{2}\\ 0&0&0&0&0&0&0&a&0\\ a&0&0&0&a&0&\tfrac{\sqrt{1-a^2}}{2}&0&\tfrac{1+a}{2}\end{pmatrix},\quad 0<a<1 \qquad (8)

where: $a$ — family parameter; for $0<a<1$ the state $\rho_a$ is a valid density matrix ( $\mathrm{Tr}=1$ , positive), PPT, and entangled — hence bound. At $a=0,1$ it degenerates to a separable state. The three theorems behind the picture — pure-state reversibility, the PPT obstruction, and the Bell negativity — are derived below.

Derivation APure-state entanglement equals the entropy of entanglement

The goal is to show that for any pure bipartite state the two operational rates coincide with a single number, $E_D = E_C = S(\rho_A)$ , so pure-state entanglement is perfectly reversible and measured by the marginal entropy alone. The engine is the typical subspace: $n$ independent copies concentrate their Schmidt weight on a set of size $2^{nS(\rho_A)}$ of nearly equal amplitudes — exactly a stack of $nS(\rho_A)$ Bell pairs in disguise. We extract that stack (distillation), rebuild the state from it (dilution), and squeeze the two bounds together.

Step 1. Write the Schmidt decomposition $|\psi\rangle_{AB} = \sum_i \sqrt{\lambda_i}\,|i\rangle_A|i\rangle_B$ , so $\rho_A = \sum_i \lambda_i|i\rangle\langle i|$ with entropy $S(\rho_A) = -\sum_i \lambda_i\log_2\lambda_i$ . Why: the Schmidt coefficients are the only local invariants of a pure bipartite state, so any LOCC rate can depend on nothing else.

Step 2. Tensor $n$ copies: $|\psi\rangle^{\otimes n} = \sum_{\mathbf i}\sqrt{\lambda_{\mathbf i}}\,|\mathbf i\rangle_A|\mathbf i\rangle_B$ with $\lambda_{\mathbf i} = \prod_k \lambda_{i_k}$ . By the asymptotic equipartition property these weights concentrate on a typical set of $\approx 2^{nS(\rho_A)}$ strings, each of weight $\approx 2^{-nS(\rho_A)}$ . Why: many i.i.d. copies flatten the Schmidt spectrum onto an almost-uniform plateau — the fingerprint of maximal entanglement.

Step 3. Distill. Alice and Bob each locally project onto the typical subspace, with success approaching one. The shared state is then approximately maximally entangled across $\approx 2^{nS(\rho_A)}$ equal terms — that is, $nS(\rho_A)$ Bell pairs. Hence $E_D \ge S(\rho_A)$ . Why: a flat Schmidt vector of dimension $2^m$ is precisely $m$ ebits.

Step 4. Dilute. Conversely, from $nS(\rho_A)$ Bell pairs, Schumacher compression plus classical communication synthesizes $|\psi\rangle^{\otimes n}$ to fidelity approaching one. Hence $E_C \le S(\rho_A)$ . Why: the same typical subspace tells you exactly how many ebits are needed to pack the state back in.

Step 5. Squeeze. With the universal ordering $E_D \le E_C$ , the two bounds give $S(\rho_A) \le E_D \le E_C \le S(\rho_A)$ , forcing all three equal and the round trip reversible. Why: two inequalities pointing the opposite way through a general one pin the quantity to a formula. *Sanity checks: a Bell pair has $\lambda=\{\tfrac12,\tfrac12\}$ , giving $S=1$ ebit; a product state has $\lambda=\{1\}$ , giving $S=0$ .*

Go deeper · Derivation B · 2 minPositive partial transpose forbids distillation

Derivation BPositive partial transpose forbids distillation

The goal is to prove the load-bearing theorem of the subject: if $\rho^{T_B}\ge 0$ then $E_D(\rho)=0$ , so no LOCC protocol on any number of copies can squeeze out a single clean Bell pair. The strategy is to find an invariant — positivity of the partial transpose — that every LOCC map preserves, then observe that the distillation target lies strictly outside that invariant set. A protocol that never leaves a closed region can never reach a point outside it, so the achievable rate is zero.

Step 1. Every LOCC operation is a special separable map, $\rho\mapsto \sum_k (A_k\otimes B_k)\,\rho\,(A_k\otimes B_k)^\dagger$ , conditioned on classical messages. Why: local Kraus operators tensor across the cut, and classical communication only selects among local branches — it introduces no non-local operator.

Step 2. The partial transpose intertwines with local maps: $\big[(A\otimes B)\,\sigma\,(A\otimes B)^\dagger\big]^{T_B} = (A\otimes \bar B)\,\sigma^{T_B}\,(A\otimes \bar B)^\dagger$ , with $\bar B$ the complex conjugate. Why: transposing $B$ turns the right action of $B^\dagger$ into a left action of $\bar B$ , so the expression stays a congruence — and congruences preserve positivity.

Step 3. So if $\rho^{T_B}\ge 0$ , each output term $(A_k\otimes\bar B_k)\,\rho^{T_B}\,(A_k\otimes\bar B_k)^\dagger\ge 0$ and their sum stays $\ge 0$ : PPT is a closed, LOCC-invariant set, stable under tensoring, so $\rho^{\otimes n}$ remains PPT. Why: a sum of positive operators is positive — no branch, no number of copies, can manufacture a negative eigenvalue.

Step 4. The target lies outside this set. The Bell pair is NPT ( $\Phi_2^{T_B}$ has eigenvalue $-\tfrac12$ ), and any state within small trace-norm distance of $\Phi_2^{\otimes m}$ is NPT too, since eigenvalues move continuously. Why: NPT is an open condition, so a whole neighbourhood of the target is forbidden to a PPT trajectory.

Step 5. Conclude. From PPT $\rho^{\otimes n}$ , no LOCC channel can enter the NPT neighbourhood of any Bell target, so the best rate with vanishing error is $R=0$ : $E_D(\rho)=0$ . With $E_C>0$ for entangled $\rho$ , every PPT entangled state is bound. Why: an invariant separating start from target is a conserved obstruction, like a topological charge. *Consistency check: in $2\times2$ and $2\times3$ , PPT is equivalent to separable, so no PPT entangled state exists and the theorem is non-vacuous only from $3\times3$ and $2\times4$ upward.*

Go deeper · Derivation C · 1 minThe Bell state has a negative partial transpose

Derivation CThe Bell state has a negative partial transpose

The goal is a direct, by-hand computation that anchors everything else: the maximally entangled two-qubit state violates PPT, with partial-transpose spectrum $\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ , negativity $\tfrac12$ , and logarithmic negativity exactly $1$ ebit. This single negative number makes the wall of the previous derivation a real barrier, and it fixes the currency unit — one Bell pair is worth one ebit of negativity.

Step 2. Apply $T_B$ , transposing the $B$ index of each bra–ket pair: $|00\rangle\langle11|\to|01\rangle\langle10|$ and $|11\rangle\langle00|\to|10\rangle\langle01|$ , diagonal terms untouched. Why: transposition swaps the second-slot labels, moving coherence from the $\{00,11\}$ sector into the $\{01,10\}$ sector.

Step 3. The result is $\Phi_2^{T_B} = \tfrac12\big(|00\rangle\langle00| + |11\rangle\langle11| + |01\rangle\langle10| + |10\rangle\langle01|\big)$ . The $\{|01\rangle,|10\rangle\}$ block is $\tfrac12\left(\begin{smallmatrix}0&1\\1&0\end{smallmatrix}\right)$ , eigenvalues $\pm\tfrac12$ ; the diagonal contributes $\tfrac12,\tfrac12$ . Why: an antidiagonal $2\times2$ block splits into symmetric and antisymmetric eigenvectors — one positive, one negative eigenvalue of equal magnitude.

Step 4. Collect the spectrum $\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ : negativity $\mathcal{N}=\tfrac12$ , trace norm $\|\Phi_2^{T_B}\|_1 = 2$ , and $E_N = \log_2 2 = 1$ ebit. Why: one full ebit of logarithmic negativity defines the currency unit against which every other $E_N$ is quoted. *The partial transpose returns eigenvalues $\{-0.5,0.5,0.5,0.5\}$ , trace norm $2.0$ , and $E_N = 1.0$ , matching the standard negativity $\tfrac12$ of a maximally entangled qubit pair.*

Limiting cases

How the two rates specialize across the hierarchy:

Case	Result	Meaning
Pure state $\vert\psi\rangle_{AB}$	$E_D = E_C = S(\rho_A) = -\sum_i \lambda_i\log_2\lambda_i$	Reversible regime; a Bell pair gives $E_D=E_C=1$ ebit, a product state $0$ .
Separable $\rho=\sum_k p_k\,\rho_A^k\otimes\rho_B^k$	$E_D = E_C = 0$	No entanglement; freely creatable, trivially undistillable.
PPT entangled (bound)	$E_D = 0,\ E_C > 0$	Entangled but undistillable — maximally irreversible. Witness fires, partial transpose stays $\ge 0$ .
NPT distillable	$0 < E_D \le E_C$	Generic entangled states; some pure entanglement is recoverable, negativity $>0$ .
Two-qubit, qubit-qutrit ( $2\times2$ , $2\times3$ )	$\text{NPT} \Leftrightarrow \text{distillable} \Leftrightarrow \text{entangled}$	PPT equals separable here; no bound entanglement exists in these dimensions.

Common trapEntangled does not mean useful, and undistillable does not mean separable

It is tempting to collapse three distinct properties — non-separable, distillable, useful — into one, and every collapse is wrong. Bound entangled states fail every separability test and are flagged by witnesses, yet yield $E_D = 0$ : being entangled does not make a state a source of ebits. The reverse implication fails too: $E_D = 0$ does not imply separable, since PPT entangled states have $E_D = 0$ while remaining firmly non-separable. And the PPT route proves undistillability only for PPT states — whether NPT bound entanglement exists is a famous open problem. Keep the three ideas apart: separability is about product decompositions, distillability about LOCC extraction, and usefulness about the task.

Worked examples

1 · When does white noise kill distillability? The two-qubit Werner state (fully worked)

Consider the Werner state $\rho_p = p\,|\Psi^-\rangle\langle\Psi^-| + (1-p)\tfrac{I}{4}$ , a singlet blended with maximally mixed noise. For which mixing $p\in[0,1]$ is it entangled, and hence (in $2\times2$ ) distillable? Find the negativity.

Step 1. Take the partial transpose. Since $I^{T_B}=I$ and the singlet, like any Bell state, has partial-transpose spectrum $\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ , linearity gives $\rho_p^{T_B} = p\,|\Psi^-\rangle\langle\Psi^-|^{T_B} + (1-p)\tfrac{I}{4}$ , whose eigenvalues are $p\cdot(\text{Bell PT eig}) + \tfrac{1-p}{4}$ .

Step 2. Evaluate the four eigenvalues. Three come from the $+\tfrac12$ Bell eigenvalues: $p\cdot\tfrac12 + \tfrac{1-p}{4} = \tfrac{1+p}{4}$ (always positive). One comes from the $-\tfrac12$ eigenvalue: $p\cdot(-\tfrac12) + \tfrac{1-p}{4} = \tfrac{1-3p}{4}$ .

Step 3. The lone eigenvalue $\tfrac{1-3p}{4}$ turns negative exactly when $p > \tfrac13$ . So $\rho_p$ is NPT — hence entangled and, in $2\times2$ , distillable — precisely for $p > 1/3$ ; for $p\le\tfrac13$ it is separable, $E_D = E_C = 0$ .

Step 4. Read off the negativity: $\mathcal{N} = \max\!\big(0,\ \tfrac{3p-1}{4}\big)$ . At $p=1$ (pure singlet) $\mathcal{N} = \tfrac12$ , recovering the Bell result; at $p=\tfrac12$ , $\mathcal{N} = \mathbf{0.125}$ ; at the boundary $p=\tfrac13$ , $\mathcal{N}=0$ . The critical noise level is $p=1/3$ : below it the singlet's entanglement is fully washed out. Numerically confirmed: at $p=\tfrac12$ the partial-transpose eigenvalues are $\{0.375,0.375,0.375,-0.125\}$ , negativity $0.125$ ; at $p=\tfrac13$ they are $\{\tfrac13,\tfrac13,\tfrac13,0\}$ , exactly the separability edge.

2 · The Bell pair versus a PPT state: one ebit against zero (fully worked)

Compare the logarithmic negativity, and hence the distillability, of the Bell state $\Phi_2$ with that of any PPT state $\sigma$ .

Step 1. For $\Phi_2$ , Derivation 3 gives spectrum $\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ , so $\|\Phi_2^{T_B}\|_1 = 2$ and $E_N(\Phi_2) = \log_2 2 = \mathbf{1\ ebit}$ . Since $E_N \ge E_D$ and $\Phi_2$ is pure with $S(\rho_A)=1$ , in fact $E_D = E_C = 1$ ebit — the currency unit.

Step 2. For any PPT state $\sigma$ , positivity gives $\sigma^{T_B}\ge 0$ , so $\|\sigma^{T_B}\|_1 = \mathrm{Tr}\,\sigma^{T_B} = \mathrm{Tr}\,\sigma = 1$ and $E_N(\sigma) = \log_2 1 = \mathbf{0}$ .

Step 3. Since $E_N \ge E_D$ , a value of $0$ forces $E_D(\sigma) = 0$ : no PPT state can be distilled, re-deriving equation $(5)$ from negativity alone. The Bell pair carries a full ebit, while any PPT state — even the entangled Horodecki $\rho_a$ — carries exactly zero. But $E_N(\rho_a)=0$ does not make $\rho_a$ separable: it certifies undistillability, not separability. $E_N$ separates the distillable Bell pair (1 ebit) from every bound state (0 ebits) with a single scalar.

Physics of the portrait

Bound entanglement is the work not of one theorist but of a family that turned entanglement from a yes/no property into a structured, computable hierarchy in a few years in the late 1990s.

Origins · 1 minHorodecki family and how this was discovered

Horodecki family

collaborating from 1996 · bound entanglement and the PPT separability criterion

The Horodecki family — Ryszard Horodecki and his sons Paweł and Michał, working in Gdańsk — reshaped entanglement theory in a burst between 1996 and 1998. In 1996, building on Asher Peres's proposal that a separable state must have a positive partial transpose, the three proved the converse in the smallest cases: in $2\times2$ and $2\times3$ systems PPT is not merely necessary but sufficient for separability, so entanglement there is exactly negativity of the partial transpose. The same paper introduced entanglement witnesses via the Hahn–Banach separation theorem. The deeper surprise came next: in 1997 Paweł Horodecki constructed an explicit $3\times3$ state — the matrix $\rho_a$ of this article — that is PPT yet demonstrably entangled, using a range criterion to rule out separability. Then in 1998 all three proved that any PPT state is undistillable, coining bound entanglement: entanglement that costs ebits to create but returns none. This split the once-monolithic notion of "entanglement" into a layered resource theory of free, bound, and distillable strata, mapped in their standard 2009 Reviews of Modern Physics survey. The through-line is operational: an entanglement measure earns its name only by controlling a physical conversion rate, and bound entanglement is the proof that detection and distillation are different questions.

History

1989

— Werner constructs mixed states that are entangled yet admit a local hidden-variable model

1996

— Bennett, Bernstein, Popescu and Schumacher quantify concentration and dilution, establishing pure-state reversibility

1996

— Peres proposes the positive-partial-transpose test as a necessary condition for separability

1997

— Paweł Horodecki exhibits the first PPT entangled state in a 3×3 system

1998

— the Horodecki family proves that every PPT entangled state is undistillable, defining bound entanglement

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §12.5 · Preskill, Lecture Notes on Quantum Computation, Ch. 4 · Watrous, The Theory of Quantum Information, Ch. 6 · Wilde, Quantum Information Theory, Ch. 19–21 · Horodecki⁴, Quantum entanglement, Rev. Mod. Phys. 81, 865 (2009) · Horodecki³, Phys. Rev. Lett. 80, 5239 (1998). The key numerical anchors are: (1) the Bell partial-transpose spectrum $\{\tfrac12,\tfrac12,\tfrac12,-\tfrac12\}$ , negativity $\tfrac12$ , $E_N=1$ ebit; (2) the Werner PT eigenvalues $\tfrac{1+p}{4}$ (triple) and $\tfrac{1-3p}{4}$ , entangled iff $p>\tfrac13$ , negativity $\tfrac{3p-1}{4} = 0.125$ at $p=\tfrac12$ ; (3) $\rho_a$ of equation $(8)$ has unit trace, is positive semidefinite, and is PPT for $0<a<1$ . The pure-state squeeze $S(\rho_A)\le E_D\le E_C\le S(\rho_A)$ establishes reversibility; PPT preservation under LOCC gives $E_D=0$ for all PPT states; and PPT $\Leftrightarrow$ separable in $2\times2$ and $2\times3$ .

Prerequisites. Separable versus Entangled States · PPT Criterion and the Peres–Horodecki Test · Schmidt Decomposition · Entanglement Entropy · Entanglement Witnesses. Leads to. Bell States and EPR Pairs · Concurrence and Wootters' Formula · Nonlocality and the No-Signalling Principle.

Ask the tutor: Ask me to walk through why the partial transpose is the exact obstruction to distillation, or to compute the negativity of any state you name.

What is the difference between distillable entanglement and entanglement cost?

They are the two directions of the same conversion. Distillable entanglement $E_D$ is the best rate at which you can refine many noisy copies into near-perfect Bell pairs using only LOCC — the output rate of the currency. Entanglement cost $E_C$ is the rate at which you must feed Bell pairs in to build the state — the input rate. For any state $E_D \le E_C$ , because otherwise you could create a state cheaply and distill more than you spent, generating entanglement for free under LOCC. When $E_D < E_C$ the manipulation is irreversible; when $E_D = 0 < E_C$ the state is bound.

How can a state be entangled but have zero distillable entanglement?

Entanglement means the state cannot be written as a mixture of product states — a structural fact, certified by a witness or by $E_C > 0$ . Distillability is a stronger, operational demand: that LOCC can actually extract clean Bell pairs from many copies. A PPT entangled state satisfies the first but fails the second, because its partial transpose stays positive under every LOCC map, so the state can never be driven toward a Bell pair, which is NPT. The entanglement is really there — you paid for it — but it is locked in a form no free operation can convert into usable ebits. That locked form is exactly what "bound" names.

Why is the partial transpose the right tool here?

Two facts conspire. First, transposition is positive but not completely positive, so applying it to one half of an entangled state can produce a negative eigenvalue — a flag no separable state can raise. Second, and crucially for distillation, the positivity of the partial transpose is preserved by all LOCC operations, because local Kraus operators intertwine with $T_B$ as a congruence. That makes PPT a conserved obstruction: an LOCC trajectory starting inside the PPT region stays there forever, while the Bell target lives outside. The partial transpose is thus both a separability probe and a distillability barrier.

Is logarithmic negativity a perfect measure of distillable entanglement?

No — it is a convenient bound, not an equality. Logarithmic negativity $E_N = \log_2\|\rho^{T_B}\|_1$ is easy to compute (just diagonalize the partial transpose) and it upper-bounds $E_D$ , so $E_N = 0$ guarantees $E_D = 0$ . But the gap can be large: many NPT states have $E_N$ strictly above their true $E_D$ , and computing $E_D$ exactly is generally intractable. Use $E_N$ as a fast certificate of undistillability and a ceiling on the yield, not as the yield itself.

Does bound entanglement have any use at all, or is it just a curiosity?

It is genuinely useful, which is what makes it interesting rather than pathological. Bound entangled states can be activated: two states individually undistillable can, used together, allow distillation, and bound states can boost the teleportation fidelity achievable from another resource. More dramatically, some support a nonzero distillable secret key, so they are useful for cryptography even though they yield no Bell pairs. These phenomena show that a single scalar cannot capture a state's worth — the value depends on the task and on what other resources are at hand, which is the whole point of a resource theory.

Why does bound entanglement need at least a 3×3 or 2×4 system?

Because in the smallest systems — $2\times2$ (two qubits) and $2\times3$ (qubit and qutrit) — the Horodecki family proved that a positive partial transpose is not merely necessary but sufficient for separability. So there every entangled state is NPT, every NPT state is distillable, and there is no room for a state that is entangled yet PPT. The equivalence breaks at $3\times3$ and $2\times4$ : the set of PPT states becomes strictly larger than the separable set, opening a shell of PPT entangled states. Those are the smallest bound entangled states, and Paweł Horodecki's $3\times3$ matrix is the canonical inhabitant of that shell.