Entanglement of Formation

A pure bipartite state carries an unambiguous entanglement measure — the von Neumann entropy of either reduced density operator, the entropy of entanglement. Mixed states break that clarity, because a single density operator ρ can be assembled from pure states in infinitely many ways, and the correlations inside ρ may be quantum, classical, or a blend of the two. The entanglement of formation resolves the ambiguity operationally: of all the recipes that synthesize ρ from pure entangled ingredients, which one spends the least entanglement? That minimum-over-recipes is the convex-roof construction, the canonical way to lift a pure-state measure to the mixed-state world, and it sits at the centre of the resource theory of entanglement alongside its siblings the distillable entanglement and the entanglement cost. Introduced by Bennett, DiVincenzo, Smolin and Wootters in 1996, the entanglement of formation $E_F$ was conceived as a formation cost — the entangled resource, in ebits, needed to build one copy of ρ by local operations and classical communication (LOCC) given a supply of singlets. Computing it is in general a hard convex optimization with no efficient algorithm known in arbitrary dimension. The triumph of the theory is Wootters' 1998 exact solution for two qubits, where $E_F$ collapses to an algebraic function of the concurrence — one of the very few entanglement measures with a closed form for an entire class of states. This article assembles the definition, proves the properties that make $E_F$ faithful, derives the two-qubit formula, and works the numbers, so the reader can compute $E_F$ for a real density matrix and know what it means.

Key takeaways

The entanglement of formation $E_F(\rho)$ is the minimal average pure-state entanglement — measured by the entropy of entanglement — needed to build a bipartite mixed state ρ from pure states.
It is defined by the convex roof: minimize $\sum_i p_i\,E(|\psi_i\rangle)$ over every pure-state ensemble $\{p_i,|\psi_i\rangle\}$ that reconstructs $\rho=\sum_i p_i|\psi_i\rangle\langle\psi_i|$ .
$E_F$ is a faithful entanglement monotone: it vanishes exactly on separable states and never increases under LOCC.
For two qubits, Wootters' formula gives the closed form $E_F=h\!\big(\tfrac{1+\sqrt{1-C^2}}{2}\big)$ in terms of the concurrence $C$ — the only class where the convex roof is solved exactly.
$E_F$ generally differs from the distillable entanglement $E_D$ and the entanglement cost $E_C$ ; the regularization $E_C=\lim_n E_F(\rho^{\otimes n})/n$ recovers the true asymptotic cost.
Additivity of $E_F$ was a landmark open problem; Hastings' 2009 counterexample to additivity of minimum output entropy showed $E_F$ is not additive in general.
Key idea: Lift the pure-state entropy of entanglement to mixed states by the convex roof — $E_F(\rho)$ is the minimum, over all pure-state decompositions of ρ, of the average entanglement entropy. The minimization strips away classical correlation and leaves only the genuinely quantum part.

The core insight

Lift the pure-state entropy of entanglement to mixed states by the convex roof — $E_F(\rho)$ is the minimum, over all pure-state decompositions of ρ, of the average entanglement entropy, a minimization that separates the genuinely quantum correlations from the merely classical ones.

Where this lives today

The entanglement of formation is the number a laboratory actually reports when it claims to have made an entangled pair. In photonic, trapped-ion and superconducting-qubit experiments the standard pipeline is tomography to reconstruct the two-qubit density matrix, then Wootters' formula: compute the spin-flip eigenvalues, extract the concurrence, and read off $E_F$ in ebits as the certified figure of merit. Beyond two qubits the problem is genuinely hard — computing $E_F$ for arbitrary bipartite states is believed to be intractable, and the practical frontier is semidefinite-programming lower bounds and symmetry-reduced special cases (isotropic and highly symmetric states) where the convex roof can be pinned down. Two deep open questions still hang on $E_F$ . The regularization gap — how far single-copy $E_F$ overshoots the asymptotic cost $E_C$ — is only partly understood, and Hastings' 2009 disproof of additivity showed the gap is real rather than an artifact. And the irreversibility it quantifies — that PPT bound entangled states cost entanglement to make ( $E_F>0$ ) yet yield none back ( $E_D=0$ ) — continues to drive the search for a "second law" of entanglement thermodynamics.

The physical picture

Think of a mixed state ρ as a probabilistic recipe for preparing quantum systems. Many recipes — many pure-state ensembles — produce exactly the same ρ, yet they differ enormously in how much genuine entanglement they pour in. A clumsy recipe uses highly entangled ingredients whose entanglement partly cancels into mere classical correlation once you average; a smart recipe uses the least entanglement that still rebuilds ρ. The entanglement of formation is the price tag on the smartest recipe. The first figure makes this concrete with a state whose correlations are entirely classical, yet which can be dressed up in a wasteful entangled disguise.

Figure

Two ensembles build the same separable ρ: the naive Bell-state recipe averages 1 ebit while the smart product-state recipe averages 0, and E_F is the minimum over recipes, so E_F = 0 ebit

Mixedness has two origins a naive measure fatally conflates: classical ignorance of which pure state was prepared, and genuine quantum entanglement inside each pure state. The reduced-state entropy $S(\rho_A)$ sees both at once and overcounts — the classically correlated state above has $S(\rho_A)=1$ bit yet zero entanglement. The convex roof cures this by handing you the freedom to choose the decomposition, and the optimal choice pours away every drop of avoidable entanglement, leaving only what is truly irreducible. Geometrically, $E_F$ is the largest convex function on density operators that agrees with the entropy of entanglement on the pure states — the "floor" obtained by draping a convex sheet beneath the pure-state values. For two qubits that floor depends on the state only through a single algebraic number, the concurrence, and the second figure plots the resulting universal curve.

Figure

The Wootters function ℰ(C) = h((1+√(1−C²))/2) maps concurrence to entanglement of formation, climbing convexly from 0 to 1 ebit and lying below the diagonal E = C; the Werner state at p = 1/2 has C = 0.25 giving E_F = 0.118 ebit, while a Bell state has C = 1 giving 1 ebit

The core equations

The definition is a minimization. A mixed state ρ has infinitely many pure-state decompositions; the entanglement of formation is the smallest average entanglement any of them can achieve. This single optimization is the object everything else in the article computes, bounds, or interprets.

E_F(\rho)=\min_{\{p_i,|\psi_i\rangle\}}\ \sum_i p_i\,E\!\left(|\psi_i\rangle\right),\qquad \rho=\sum_i p_i\,|\psi_i\rangle\langle\psi_i| \qquad (1)

where: $\rho$ — bipartite density operator on $\mathcal{H}_A\otimes\mathcal{H}_B$ ; $\{p_i,|\psi_i\rangle\}$ — a pure-state ensemble realizing ρ, with $p_i\ge 0$ and $\sum_i p_i=1$ ; $E(|\psi_i\rangle)$ — the entropy of entanglement of the pure state $|\psi_i\rangle$ . Read it as: the cheapest recipe wins, and its cost defines $E_F$ . The ingredient being averaged is the pure-state entanglement, which is unambiguous: the von Neumann entropy of the reduced state. This is the quantity the convex roof lifts, and it is why $E_F$ reduces to the entropy of entanglement whenever ρ is itself pure.

E\!\left(|\psi\rangle\right)=S(\rho_A)=-\mathrm{Tr}\!\left[\rho_A\log_2\rho_A\right],\quad \rho_A=\mathrm{Tr}_B|\psi\rangle\langle\psi| \qquad (2)

where: $\rho_A$ — reduced state on subsystem $A$ ; $S$ — von Neumann entropy in bits (base-2 logarithm), so that one Bell pair carries exactly one ebit. For a Schmidt decomposition $|\psi\rangle=\sum_k\sqrt{\lambda_k}|k\rangle_A|k\rangle_B$ this is just $-\sum_k\lambda_k\log_2\lambda_k$ . The binary entropy function packages the two-qubit case. It is the Shannon entropy of a biased coin, and it is the exact shape into which Wootters' formula pours the concurrence.

h(x)=-x\log_2 x-(1-x)\log_2(1-x) \qquad (3)

where: $x$ — probability argument, $0\le x\le 1$ ; $h$ — binary Shannon entropy, with $h(0)=h(1)=0$ and maximum $h(\tfrac12)=1$ . It is symmetric about $x=\tfrac12$ and concave. For two qubits the whole optimization collapses onto a single scalar, the concurrence, built from the antiunitary spin flip. This is the algebraic miracle that makes $E_F$ computable in closed form for a qubit pair.

C(\rho)=\max\{0,\ \lambda_1-\lambda_2-\lambda_3-\lambda_4\} \qquad (4)

where: $\lambda_1\ge\lambda_2\ge\lambda_3\ge\lambda_4$ — the decreasingly ordered square roots of the eigenvalues of $\rho\,\tilde\rho$ (equivalently the eigenvalues of $\sqrt{\sqrt{\rho}\,\tilde\rho\,\sqrt{\rho}}$ ); $\tilde\rho=(\sigma_y\otimes\sigma_y)\rho^*(\sigma_y\otimes\sigma_y)$ — the spin-flipped state, with $\rho^*$ the complex conjugate in the computational basis. The concurrence runs from $0$ (separable) to $1$ (maximally entangled). Wootters' formula is the payoff: the entanglement of formation of any two-qubit state is a fixed, monotonically increasing function of its concurrence. Compute $C$ from the eigenvalues above, feed it in, and read off $E_F$ in ebits.

E_F(\rho)=\mathcal{E}\big(C(\rho)\big)=h\!\left(\frac{1+\sqrt{1-C(\rho)^2}}{2}\right) \qquad (5)

where: $C(\rho)$ — concurrence of the mixed two-qubit state; $\mathcal{E}$ — the increasing, convex map from concurrence to formation plotted in the second figure. At $C=0$ , $E_F=0$ ; at $C=1$ , $E_F=h(\tfrac12)=1$ ebit. The single-copy $E_F$ is not the last word on cost. The true asymptotic price of preparing ρ is the regularized formation, obtained by amortizing over many copies — and it can be strictly smaller because $E_F$ is generally superadditive.

E_C(\rho)=\lim_{n\to\infty}\frac{E_F(\rho^{\otimes n})}{n}\ \le\ E_F(\rho) \qquad (6)

where: $E_C$ — asymptotic entanglement cost, the ebits per copy needed to prepare ρ by LOCC; $\rho^{\otimes n}$ — $n$ independent copies. Equality would hold if $E_F$ were additive, which in general it is not. Finally, $E_F$ (an upper bound on cost) sits above the distillable entanglement (what LOCC can extract), and the strict gap for many mixed states is the signature of irreversibility.

E_D(\rho)\ \le\ E_C(\rho)\ \le\ E_F(\rho) \qquad (7)

where: $E_D$ — distillable entanglement, ebits recoverable per copy by LOCC. For pure states all three coincide; for mixed states the chain can be strict, and PPT bound entangled states realize the extreme case $E_D=0<E_F$ .

Derivation AThe convex roof is a faithful entanglement monotone

The goal is to establish the two properties that qualify $E_F$ as a bona fide entanglement measure rather than an arbitrary functional: it is faithful, vanishing precisely on the separable states and strictly positive otherwise, and it is an LOCC monotone, never increasing under local operations and classical communication. Both follow directly from the convex-roof definition together with the corresponding facts about the pure-state entropy of entanglement. We take non-negativity, the two directions of faithfulness, and monotonicity in turn.

Step 1. Non-negativity is immediate: every term in the average of equation (1) is a von Neumann entropy $E(|\psi_i\rangle)=S(\rho_A^{(i)})\ge 0$ , so every candidate average is non-negative, and therefore so is their minimum, giving $E_F(\rho)\ge 0$ . Why: the roof inherits the sign of its ingredient — an average of non-negative numbers cannot dip below zero, and minimizing over such averages cannot either.

Step 2. A separable state has, by definition, a product-state decomposition $\rho=\sum_i p_i\,|a_i\rangle\langle a_i|\otimes|b_i\rangle\langle b_i|$ . Each ingredient $|a_i\rangle\otimes|b_i\rangle$ is a product pure state with $E=0$ , so this particular ensemble has average entanglement $0$ ; since $E_F$ is the minimum over ensembles and cannot be negative, $E_F(\rho)=0$ . Why: separability literally hands you a zero-cost recipe, and zero is already the floor from Step 1.

Step 3. Conversely suppose $E_F(\rho)=0$ , achieved by some optimal ensemble $\{p_i,|\psi_i\rangle\}$ . Then $\sum_i p_i E(|\psi_i\rangle)=0$ with every term non-negative, forcing $E(|\psi_i\rangle)=0$ for each $i$ with $p_i>0$ . A pure state with zero entropy of entanglement is a product state, so ρ is a mixture of product states — separable. Why: a sum of non-negative terms vanishes only if each vanishes, and vanishing pure-state entanglement means the ingredient factorizes, which is exactly separability.

Step 4. For monotonicity, any LOCC protocol can be written as a collection of local operations that, conditioned on classical outcomes $k$ , map ρ to $\rho_k$ with probability $q_k$ . Because the pure-state entropy of entanglement is non-increasing under local operations, and because the convex roof is itself convex, averaging over outcomes gives $\sum_k q_k\,E_F(\rho_k)\le E_F(\rho)$ . In particular no LOCC protocol raises the expected entanglement of formation. Why: local maps can only degrade entanglement of the pure ingredients, and convexity guarantees that classical mixing of the outcomes cannot conjure it back. *As a concrete check: the classically correlated state ρ = ½|00⟩⟨00| + ½|11⟩⟨11| gives $E_F = 0$ despite $S(\rho_A) = 1$ — the product-state recipe costs nothing, confirming that the roof correctly returns zero for a separable state.*

Go deeper · Derivation B · 1 minWhy the convex roof is convex

Derivation BWhy the convex roof is convex

The goal is to prove that $E_F$ is a convex function of the density operator — mixing two states can never cost more than the average of their formation costs — the property that gives the convex roof its name and its meaning as the largest convex function agreeing with the entropy of entanglement on pure states. The proof is short and general: take the optimal ensembles of the two states, glue them into one ensemble of the mixture, and observe that the minimum defining $E_F$ cannot exceed the cost of that glued recipe.

Step 1. Let $\{p_i,|\psi_i\rangle\}$ be an optimal ensemble for $\rho_1$ , achieving $E_F(\rho_1)=\sum_i p_i E(|\psi_i\rangle)$ , and $\{q_j,|\phi_j\rangle\}$ an optimal ensemble for $\rho_2$ , achieving $E_F(\rho_2)$ . These exist because the minimization in equation (1) is over a compact set and the objective is continuous. Why: we need concrete cheapest recipes for each state before we can combine them.

Step 2. Form the mixture $\rho=\lambda\rho_1+(1-\lambda)\rho_2$ with $0\le\lambda\le1$ . The combined ensemble $\{\lambda p_i,|\psi_i\rangle\}\cup\{(1-\lambda)q_j,|\phi_j\rangle\}$ has non-negative weights summing to one and reconstructs ρ exactly, so it is a valid — though generally non-optimal — decomposition of ρ. Why: probabilistically mixing the two recipes with weights λ and 1−λ produces exactly the mixed state, providing one admissible way to build it.

Step 3. Because $E_F(\rho)$ is the minimum over all decompositions, it is bounded above by the average entanglement of this specific one: $E_F(\rho)\le \sum_i\lambda p_i E(|\psi_i\rangle)+\sum_j(1-\lambda)q_j E(|\phi_j\rangle)$ . The two sums are precisely $\lambda E_F(\rho_1)$ and $(1-\lambda)E_F(\rho_2)$ , so $E_F(\rho)\le\lambda E_F(\rho_1)+(1-\lambda)E_F(\rho_2)$ . Why: a minimum is never larger than the value of any particular candidate, and the candidate's cost splits cleanly into the two optimal costs — that inequality is convexity.

Go deeper · Derivation C · 1 minWootters' formula applied to the Werner state

Derivation CWootters' formula applied to the Werner state

The goal is to carry Wootters' closed form from a physically natural mixed state to an explicit number, exhibiting the full pipeline — identify the eigenstructure, compute the concurrence, feed it through $\mathcal{E}$ — on the one-parameter Werner family $\rho_W = p\,|\Phi^+\rangle\langle\Phi^+| + (1-p)\,\tfrac{I}{4}$ , which runs from a Bell state at $p=1$ to the maximally mixed, separable state at $p=0$ . Along the way the separability threshold $p=\tfrac13$ emerges from the concurrence and matches the independent PPT criterion, a strong internal check.

Step 1. The state is Bell-diagonal. Since $\tfrac{I}{4}=\tfrac14\sum_{\text{Bell}}|\beta\rangle\langle\beta|$ , the weight on $|\Phi^+\rangle$ is $p+\tfrac{1-p}{4}=\tfrac{1+3p}{4}$ , and each of the other three Bell states carries weight $\tfrac{1-p}{4}$ . These are the eigenvalues of $\rho_W$ . Why: writing the identity in the Bell basis makes ρ_W diagonal there, so its spectrum is read off with no diagonalization.

Step 2. For a Bell-diagonal state the concurrence is $C=\max\{0,\ 2\lambda_{\max}-1\}$ , where $\lambda_{\max}$ is the largest Bell weight. Here $\lambda_{\max}=\tfrac{1+3p}{4}$ , so $C=\max\{0,\ \tfrac{1+3p}{2}-1\}=\max\!\left\{0,\ \tfrac{3p-1}{2}\right\}$ . The concurrence is positive precisely when $p>\tfrac13$ . Why: the spin-flip eigenvalues of a Bell-diagonal state are just its Bell weights, so the max-formula of equation (4) reduces to twice the dominant weight minus one.

Step 3. At $p=1$ this gives $C=1$ and hence $E_F=h(\tfrac12)=1$ ebit, the pure Bell state; at $p\le\tfrac13$ it gives $C=0$ and $E_F=0$ , a separable state. The crossover at $p=\tfrac13$ coincides exactly with the Peres–Horodecki PPT threshold for Werner states, confirming that the concurrence sees separability correctly. Why: the two extremes and the threshold are known independently, so their agreement validates the whole calculation.

Step 4. For $p>\tfrac13$ the explicit curve is $E_F(\rho_W)=h\!\left(\frac{1+\sqrt{1-\left(\frac{3p-1}{2}\right)^2}}{2}\right)$ , a smooth, monotonically increasing function of $p$ . At $p=\tfrac12$ , $C=0.25$ and $E_F=0.118$ ebit — the amber point in the second figure. Why: plugging the concurrence into $\mathcal{E}$ turns the abstract measure into a number a tomographer can compare against data. *As a direct check: $h\!\left(\tfrac{1+\sqrt{1-0.0625}}{2}\right) = 0.118$ ebit, confirming the value at $p = \tfrac{1}{2}$ .*

Limiting cases

The behaviour of $E_F$ across the canonical regimes fixes the meaning of the number and exposes the pitfall it avoids.

Case	Result	Meaning
Separable state	$E_F(\rho)=0 \iff \rho=\sum_i p_i\,\rho_i^A\otimes\rho_i^B$	Vanishes exactly on the separable set; a product-state recipe costs no entanglement.
Pure state	$E_F(\|\psi\rangle\langle\psi\|)=S(\rho_A)=-\sum_k\lambda_k\log_2\lambda_k$	Reduces to the entropy of entanglement; the only decomposition is the state itself, so the roof is trivial.
Maximally entangled qubits	$E_F(\|\Phi^+\rangle)=1$ ebit	Concurrence $C=1$ gives $E_F=h(\tfrac12)=1$ , the maximum for a qubit pair.
Maximally mixed qubits	$E_F(I/4)=0$	Separable despite $S(\rho_A)=1$ ; the reduced entropy is large but the entanglement is zero — the pitfall made vivid.
Werner state, $p>\tfrac13$	$E_F=h\!\big(\tfrac{1+\sqrt{1-((3p-1)/2)^2}}{2}\big)$	Rises smoothly from $0$ at threshold to $1$ ebit at $p=1$ .

Common trapThe reduced entropy is not the entanglement of a mixed state

The single most common error is to take $S(\rho_A)$ , the entropy of the reduced state, as the entanglement of a mixed ρ. For a pure state that is correct, but for a mixed state $S(\rho_A)$ conflates quantum entanglement with classical mixedness. The classically correlated state $\tfrac12(|00\rangle\langle00|+|11\rangle\langle11|)$ has $S(\rho_A)=1$ bit yet $E_F=0$ : no entanglement whatsoever, because it is a mixture of product states. One must minimize over decompositions, not read off a reduced entropy. A second trap: $E_F$ is a single-copy quantity and is not additive in general, so $E_F(\rho^{\otimes n})\ne n\,E_F(\rho)$ — only the regularization $E_C=\lim_n E_F(\rho^{\otimes n})/n$ gives the true asymptotic entanglement cost.

Worked examples

1 · Entanglement of formation of a pure state (confidence builder)

Compute $E_F$ for the pure state $|\psi\rangle=\sqrt{0.8}\,|00\rangle+\sqrt{0.2}\,|11\rangle$ , and confirm it agrees with Wootters' formula via the concurrence.

Step 1. For a pure state the convex roof is trivial — the only decomposition is the state itself — so $E_F=S(\rho_A)$ . Trace out $B$ : the state is already in Schmidt form with Schmidt coefficients $\lambda_1=0.8$ , $\lambda_2=0.2$ , giving $\rho_A=\mathrm{diag}(0.8,0.2)$ .

Step 2. Evaluate the entropy: $S(\rho_A)=-0.8\log_2 0.8-0.2\log_2 0.2=0.8(0.32193)+0.2(2.32193)=0.2575+0.4644=\mathbf{0.722}$ ebit.

Step 3. Cross-check through the concurrence. For a two-qubit pure state $C=2\sqrt{\lambda_1\lambda_2}=2\sqrt{0.8\cdot0.2}=2\sqrt{0.16}=0.8$ . Feed into Wootters: $E_F=h\!\left(\tfrac{1+\sqrt{1-0.64}}{2}\right)=h\!\left(\tfrac{1+0.6}{2}\right)=h(0.8)=0.722$ ebit — identical, as it must be. E_F = 0.722 ebit, whether computed as S(ρ_A) or as ℰ(C) with C = 0.8.

2 · A Werner state (fully worked, mixed)

Find $E_F$ for the Werner state $\rho_W=\tfrac12|\Phi^+\rangle\langle\Phi^+|+\tfrac12\,\tfrac{I}{4}$ , i.e. $p=\tfrac12$ .

Step 1. Bell weights: $\lambda_{\max}=\tfrac{1+3p}{4}=\tfrac{1+1.5}{4}=\tfrac{2.5}{4}=0.625$ for $|\Phi^+\rangle$ , and $\tfrac{1-p}{4}=0.125$ for each of the other three. The spectrum is $(0.625,0.125,0.125,0.125)$ , which sums to $1$ as required.

Step 2. Concurrence: $C=\max\{0,\ 2\lambda_{\max}-1\}=\max\{0,\ 1.25-1\}=\mathbf{0.25}$ . Positive, so the state is entangled — consistent with $p=\tfrac12>\tfrac13$ .

Step 3. Wootters' formula: $\sqrt{1-C^2}=\sqrt{1-0.0625}=\sqrt{0.9375}=0.96825$ , so the argument is $x=\tfrac{1+0.96825}{2}=0.98412$ . Then $E_F=h(0.98412)=-0.98412\log_2 0.98412-0.01588\log_2 0.01588=0.0227+0.0949=\mathbf{0.118}$ ebit.

Step 4. Sanity: $0.118$ ebit is far below the $1$ ebit of a pure Bell pair — the classical noise in the $I/4$ term has diluted most of the entanglement. E_F(ρ_W, p = ½) = 0.118 ebit at concurrence C = 0.25.

3 · The pitfall in numbers

Contrast $E_F$ with the reduced entropy for the classically correlated state $\rho=\tfrac12(|00\rangle\langle00|+|11\rangle\langle11|)$ .

Step 1. Reduced state: tracing out $B$ gives $\rho_A=\tfrac12(|0\rangle\langle0|+|1\rangle\langle1|)=\tfrac{I}{2}$ , so $S(\rho_A)=1$ bit — as large as a Bell state's.

Step 2. Concurrence: in the Bell basis $\rho=\tfrac12|\Phi^+\rangle\langle\Phi^+|+\tfrac12|\Phi^-\rangle\langle\Phi^-|$ , Bell-diagonal with $\lambda_{\max}=\tfrac12$ , so $C=\max\{0,\ 2\cdot\tfrac12-1\}=0$ and $E_F=0$ .

Step 3. The gap is the whole lesson: a coin-flip between $|00\rangle$ and $|11\rangle$ is perfectly correlated, entirely classical, zero entanglement. S(ρ_A) = 1 bit, yet E_F = 0 — reduced entropy is not entanglement.

Physics of the portrait

The entanglement of formation is inseparable from one physicist who twice bent the field — first proving you cannot copy a quantum state, then proving you can, for two qubits, put an exact number on how entangled a density matrix is.

Origins · 1 minWilliam Wootters and how this was discovered

William Wootters

b. 1951, American physicist · entanglement measures and quantum information theory

William Kent Wootters took his doctorate under John Archibald Wheeler at the University of Texas at Austin, where the questions that would define quantum information were being asked before the field had a name. In 1982, with Wojciech Zurek, he proved the no-cloning theorem — that an unknown quantum state cannot be copied — a one-line argument from linearity that turned out to underpin the security of quantum cryptography and the very possibility of quantum error correction. A decade later, entanglement had become a resource to be measured, and in 1996 Wootters joined Charles Bennett, David DiVincenzo and John Smolin to define the entanglement of formation through the convex roof, giving the mixed-state world its formation-cost measure. The masterstroke came in 1997–1998: with Scott Hill he introduced the concurrence, and then, in a single celebrated paper, Wootters proved that for any two-qubit state the convex-roof optimization — a minimization over a continuum of ensembles that looks utterly intractable — collapses to the closed algebraic form $E_F=\mathcal{E}(C)$ . The proof constructs an optimal ensemble in which every pure member shares the same concurrence, using the antiunitary spin-flip symmetry. It remains one of the few places in entanglement theory where an infinite optimization has a finite answer, and it is why concurrence and entanglement of formation carry, for a qubit pair, the same information.

History

1935

— Schrödinger coins "entanglement" (Verschränkung) and frames the EPR puzzle

1989

— Werner separates classical from quantum correlations and defines Werner states

1996

— Bennett, DiVincenzo, Smolin and Wootters define the entanglement of formation via the convex roof

1998

— Wootters proves the closed-form E_F = ℰ(C) for all two-qubit states

2009

— Hastings disproves additivity of minimum output entropy, so E_F is not additive in general

Going further

Texts. Nielsen and Chuang, Quantum Computation and Quantum Information, §12.5 (entanglement measures) · Watrous, The Theory of Quantum Information, Chapter 6 · Wilde, Quantum Information Theory, 2nd ed., Chapter 22 · Bengtsson and Życzkowski, Geometry of Quantum States, 2nd ed., Chapter 16 · Horodecki, Horodecki, Horodecki and Horodecki, Quantum entanglement, Rev. Mod. Phys. 81, 865 (2009).

Primary sources. Bennett, DiVincenzo, Smolin and Wootters, Phys. Rev. A 54, 3824 (1996) — definition · Hill and Wootters, Phys. Rev. Lett. 78, 5022 (1997) — concurrence · Wootters, Phys. Rev. Lett. 80, 2245 (1998) — closed form · Hayden, Horodecki and Terhal, J. Phys. A 34, 6891 (2001) — cost equals regularized $E_F$ · Hastings, Nature Physics 5, 255 (2009) — non-additivity. Key numerical anchors: $h(\tfrac{1}{2})=1$ ; the Werner Bell-basis eigenvalues are $\tfrac{1+3p}{4}$ and $\tfrac{1-p}{4}$ with concurrence threshold $p=\tfrac{1}{3}$ ; the pure-state value is $S(0.8)=0.722$ ebit with $C=2\sqrt{0.8\cdot0.2}=0.8$ ; and $E_F(p=\tfrac{1}{2})=0.118$ ebit.

Prerequisites. Entanglement Entropy of Pure States · Schmidt Decomposition · Separable versus Entangled States · Concurrence and Wootters' Formula. Leads to. Concurrence and Wootters' Formula · PPT Criterion and the Peres–Horodecki Test · Entanglement Witnesses.

Ask the tutor: Hand me any two-qubit density matrix and I will walk you through the spin flip, the concurrence, and Wootters' formula to its entanglement of formation in ebits.

Why not just use the reduced-state entropy S(ρ_A) as the entanglement measure?

For a pure state that is exactly right, and $E_F$ reduces to it. For a mixed state it fails, because $S(\rho_A)$ counts two things at once: genuine quantum entanglement and mere classical ignorance about which pure state was prepared. The classically correlated coin-flip $\tfrac12(|00\rangle\langle00|+|11\rangle\langle11|)$ has $S(\rho_A)=1$ bit yet contains no entanglement — it is a mixture of product states. The convex roof fixes this by letting you choose the decomposition that minimizes the residual entanglement, which for that state finds the product-state recipe and correctly returns zero.

What does "convex roof" actually mean geometrically?

Picture the entropy of entanglement defined on the pure states only, as a set of heights above the space of density operators. The convex roof is the largest convex function that stays below all those pure-state heights — the surface you get by draping a taut convex sheet underneath them. Equivalently, at any mixed state ρ you look at every way of writing ρ as a convex combination of pure states and take the lowest average height. Convexity (proved in the second derivation) is exactly the statement that this construction produces a convex function, which is what makes it well behaved as a resource measure.

Why is E_F easy for two qubits but hard in general?

The convex roof is a minimization over a continuum of ensembles, and in general there is no way to avoid searching that continuum — the problem is believed to be computationally intractable for arbitrary dimension. Two qubits are special because of a symmetry: the antiunitary spin flip $\sigma_y\otimes\sigma_y$ lets Wootters construct an optimal ensemble in which every pure member has the same concurrence. That reduces the whole optimization to computing four eigenvalues of a fixed matrix. No analogous symmetry is known in higher dimensions, which is why $E_F$ has a closed form for qubit pairs and essentially nowhere else.

How do E_F, entanglement cost E_C, and distillable entanglement E_D differ?

They answer three different operational questions. $E_C$ is the entanglement you must spend, in ebits per copy, to create ρ by LOCC in the many-copy limit. $E_D$ is the entanglement you can extract from ρ back into pure Bell pairs by LOCC. $E_F$ is the single-copy formation cost, an upper bound on $E_C$ . For pure states all three coincide, so entanglement manipulation is reversible. For mixed states the chain $E_D\le E_C\le E_F$ can be strict, and the gap $E_D<E_C$ is the irreversibility of mixed-state entanglement — most dramatically for bound entangled states with $E_D=0<E_F$ .

What is the concurrence really measuring?

The concurrence repackages "how entangled" a two-qubit state is into a single number between 0 and 1, built from the spin-flip operation. For a pure state $C=2\sqrt{\lambda_1\lambda_2}$ with $\lambda_1,\lambda_2$ the two Schmidt weights, so $C=0$ for a product state and $C=1$ when the weights are equal (a Bell state). For a mixed state it uses the ordered eigenvalues of $\rho\tilde\rho$ with a max-of-zero to keep it non-negative. Its reason for existence is that $E_F$ is a fixed monotonic function of it, so concurrence and entanglement of formation carry the same information for a qubit pair.

Why did the additivity of E_F matter so much?

If $E_F$ were additive — $E_F(\rho\otimes\sigma)=E_F(\rho)+E_F(\sigma)$ — then single-copy $E_F$ would equal the asymptotic entanglement cost $E_C$ , and a whole web of quantities (minimum output entropy, Holevo capacity) would be additive too, giving clean single-letter formulas for channel capacities. That web was conjectured to hold for years. Hastings' 2009 counterexample to additivity of minimum output entropy tore it down, proving $E_F$ is superadditive in general. The consequence is that you must regularize — average over many copies — to obtain the true entanglement cost, and single-copy $E_F$ only bounds it from above.