Entanglement Embezzlement in Quantum Systems

• The paper introduces entanglement embezzlement as the process where a shared catalyst state is nearly unchanged while approximately producing additional entangled states via LOCC.
• It relies on spectral properties of entangled states, notably the van Dam–Hayden family with characteristic 1/x Schmidt coefficient decay, to achieve universality.
• The phenomenon distinguishes tensor-product from commuting-operator models, impacting nonlocal games, self-testing, and operator-algebraic structures in quantum information.

Entanglement embezzlement is the phenomenon whereby two separated parties use a shared entangled resource state as a catalyst to approximately produce additional entanglement by local operations, while returning the catalyst essentially unchanged. In finite-dimensional quantum information, the phenomenon is classically associated with van Dam–Hayden families whose Schmidt coefficients decay like $1/x$; in operator-algebraic settings it extends to exact and universal forms tied to commuting-operator models, Cuntz-algebraic structure, and type III von Neumann factors (Zanoni et al., 2023, Cleve et al., 2016, Luijk et al., 2024).

1. Operational notion and canonical constructions

In the finite-dimensional formulation, Alice and Bob share a bipartite catalyst $\chi$ and seek an LOCC protocol of the form

$$\chi_{AB}\xrightarrow{\mathrm{LOCC}} \tilde\chi_{AB}\otimes \tilde\sigma_{A'B'},$$

with $\tilde\chi\approx \chi$ and $\tilde\sigma\approx \sigma$, where $\sigma$ is an arbitrary finite-dimensional target state. The natural error metric in the modern LOCC-based theory is the trace distance

$T(\rho,\sigma)=\frac12\|\rho-\sigma\|_1,$

and the corresponding conversion distance is

$$T(\rho\to\sigma)=\inf_{\mathcal N\in \mathrm{LOCC}} T(\mathcal N(\rho),\sigma).$$

A family $\{\chi_n\}$ is a universal embezzling family exactly when

$$\lim_{n\to\infty} T(\chi_n\to \chi_n\otimes \sigma)=0$$

for every finite-dimensional bipartite state $\chi$0 (Zanoni et al., 2023).

The canonical van Dam–Hayden family is

$\chi$1

Its Schmidt coefficients are $\chi$2, and van Dam–Hayden showed that for any finite-dimensional target state $\chi$3, sufficiently large $\chi$4 allows arbitrarily small-error embezzlement using local operations only; the resource is only slightly perturbed (Zanoni et al., 2023).

A complementary formulation, used in two-player nonlocal games, emphasizes coherent state exchange. There one considers an embezzling family $\chi$5 together with local unitaries $\chi$6 and $\chi$7 such that

$\chi$8

An explicit family is obtained from left-shift unitaries acting on

$\chi$9

which is a concrete realization of van Dam–Hayden embezzlers with error scaling $$\chi_{AB}\xrightarrow{\mathrm{LOCC}} \tilde\chi_{AB}\otimes \tilde\sigma_{A'B'},$$0 (1904.02350).

The basic operational point is therefore not entanglement creation ex nihilo. Rather, a highly spread Schmidt spectrum permits a local reshuffling that is operationally close to leaving the catalyst unchanged while extracting a prescribed finite entangled state.

2. Universal families and Schmidt-coefficient characterization

The modern one-shot theory reduces universal embezzlement to a purely spectral condition. First, it suffices to embezzle maximally entangled states $$\chi_{AB}\xrightarrow{\mathrm{LOCC}} \tilde\chi_{AB}\otimes \tilde\sigma_{A'B'},$$1, and in fact it is enough to check $$\chi_{AB}\xrightarrow{\mathrm{LOCC}} \tilde\chi_{AB}\otimes \tilde\sigma_{A'B'},$$2: universality is equivalent to

$$\chi_{AB}\xrightarrow{\mathrm{LOCC}} \tilde\chi_{AB}\otimes \tilde\sigma_{A'B'},$$3

where $$\chi_{AB}\xrightarrow{\mathrm{LOCC}} \tilde\chi_{AB}\otimes \tilde\sigma_{A'B'},$$4 is the star conversion distance on pure states (Zanoni et al., 2023).

For pure states $$\chi_{AB}\xrightarrow{\mathrm{LOCC}} \tilde\chi_{AB}\otimes \tilde\sigma_{A'B'},$$5 with Schmidt vectors $$\chi_{AB}\xrightarrow{\mathrm{LOCC}} \tilde\chi_{AB}\otimes \tilde\sigma_{A'B'},$$6,

$$\chi_{AB}\xrightarrow{\mathrm{LOCC}} \tilde\chi_{AB}\otimes \tilde\sigma_{A'B'},$$7

and the closed-form expression is

$$\chi_{AB}\xrightarrow{\mathrm{LOCC}} \tilde\chi_{AB}\otimes \tilde\sigma_{A'B'},$$8

with $$\chi_{AB}\xrightarrow{\mathrm{LOCC}} \tilde\chi_{AB}\otimes \tilde\sigma_{A'B'},$$9. The true LOCC conversion distance and $\tilde\chi\approx \chi$0 are topologically equivalent: $\tilde\chi\approx \chi$1 This replaces an optimization over all LOCC channels by a finite maximization over Ky Fan sums (Zanoni et al., 2023).

For a family of pure states $\tilde\chi\approx \chi$2 with Schmidt vectors $\tilde\chi\approx \chi$3, the complete characterization of universality is

$\tilde\chi\approx \chi$4

The condition states that every dyadic block $\tilde\chi\approx \chi$5 must become uniformly light. A direct corollary is $\tilde\chi\approx \chi$6, hence $\tilde\chi\approx \chi$7 for every universal family (Zanoni et al., 2023).

Under asymptotic regularity assumptions, this singles out the van Dam–Hayden profile. For regular families defined by

$\tilde\chi\approx \chi$8

universality forces $\tilde\chi\approx \chi$9 to be asymptotically $\tilde\sigma\approx \sigma$0 in the sense that

$\tilde\sigma\approx \sigma$1

together with $\tilde\sigma\approx \sigma$2. In particular, for the power-law family $\tilde\sigma\approx \sigma$3, universality occurs iff $\tilde\sigma\approx \sigma$4 (Zanoni et al., 2023).

A common misconception is that any slowly decaying Schmidt tail should suffice. The complete characterization shows that universal embezzlement is much more rigid: under the stated regularity hypotheses, the asymptotic $\tilde\sigma\approx \sigma$5 law is essentially unique.

3. Exact embezzlement: impossibility and possibility across models

Approximate embezzlement admits finite-dimensional catalytic families, but exact embezzlement sharply separates tensor-product and commuting-operator frameworks. In the tensor-product model, perfect embezzlement is impossible even with infinite-dimensional Hilbert spaces and infinite entanglement entropy; in the commuting-operator model it becomes possible (Cleve et al., 2016).

The tensor-product no-go is structural. Exact embezzlement would require a local-unitary transformation that changes Schmidt data while leaving the catalyst exactly invariant, which is incompatible with Schmidt-coefficient preservation under local unitaries. Cleve, Liu, and Paulsen establish this impossibility even when Alice and Bob are given arbitrary ancillas (Cleve et al., 2016).

By contrast, in the commuting-operator framework one replaces literal tensor-factor locality by commuting local algebras on a common Hilbert space. In that setting exact Bell-pair embezzlement is achievable. The construction can be phrased through a universal $\tilde\sigma\approx \sigma$6-algebra $\tilde\sigma\approx \sigma$7 and a state on $\tilde\sigma\approx \sigma$8 satisfying

$\tilde\sigma\approx \sigma$9

and it can also be realized explicitly on a separable infinite-dimensional Hilbert space via shift-and-swap operations of “Hilbert hotel” type (Cleve et al., 2016).

Later work recasts exact embezzlement in operator-algebraic form. For a target state

$\sigma$0

an exact protocol has the form

$\sigma$1

or more generally the same equation with contractions in place of unitaries; such contraction protocols can be dilated to unitary ones on the same algebras (Harris, 21 May 2026).

The impossibility/possibility dichotomy is therefore not about “more entanglement” alone. Exact embezzlement is obstructed in the spatial tensor-product model but enabled by commuting-operator and type III operator-algebraic structures.

4. Self-testing, Cuntz algebras, and exact universal constructions

Exact embezzlement has a strong rigidity theory. From any exact protocol one can extract Cuntz isometries $\sigma$2 and $\sigma$3 on the effective support, satisfying

$\sigma$4

and similarly for the $\sigma$5. This identifies each side with a representation of the Cuntz algebra $\sigma$6, and the protocol induces a unique quasi-free state characterized by

$\sigma$7

on $\sigma$8, and a unique state on $\sigma$9 satisfying

$T(\rho,\sigma)=\frac12\|\rho-\sigma\|_1,$0

After compression by support projections, any two exact protocols for the same target state are unitarily equivalent; exact embezzlement is therefore a self-test for the relevant Cuntz isometries, the quasi-free state, and the generated von Neumann-algebraic structure (Harris, 21 May 2026).

Modular theory then fixes the type of the minimal embezzlement factor. If $T(\rho,\sigma)=\frac12\|\rho-\sigma\|_1,$1 is the closed multiplicative subgroup generated by $T(\rho,\sigma)=\frac12\|\rho-\sigma\|_1,$2, then the minimal factor is the unique separable AFD type $T(\rho,\sigma)=\frac12\|\rho-\sigma\|_1,$3 factor when $T(\rho,\sigma)=\frac12\|\rho-\sigma\|_1,$4 is countable, and the unique separable AFD type $T(\rho,\sigma)=\frac12\|\rho-\sigma\|_1,$5 factor when $T(\rho,\sigma)=\frac12\|\rho-\sigma\|_1,$6. Equivalently, it is type $T(\rho,\sigma)=\frac12\|\rho-\sigma\|_1,$7 iff $T(\rho,\sigma)=\frac12\|\rho-\sigma\|_1,$8 for some pair $T(\rho,\sigma)=\frac12\|\rho-\sigma\|_1,$9 (Harris, 21 May 2026).

A related recent perspective treats exact embezzlement itself as a structural certification principle. Exact embezzlement of a pure target state $$T(\rho\to\sigma)=\inf_{\mathcal N\in \mathrm{LOCC}} T(\mathcal N(\rho),\sigma).$$0 from a catalyst $$T(\rho\to\sigma)=\inf_{\mathcal N\in \mathrm{LOCC}} T(\mathcal N(\rho),\sigma).$$1 is equivalent to $$T(\rho\to\sigma)=\inf_{\mathcal N\in \mathrm{LOCC}} T(\mathcal N(\rho),\sigma).$$2 locally containing infinitely many mutually commuting copies of $$T(\rho\to\sigma)=\inf_{\mathcal N\in \mathrm{LOCC}} T(\mathcal N(\rho),\sigma).$$3; in this sense embezzlement behaves like a self-test for an infinite tensor-product structure inside the catalyst. This viewpoint yields an elementary route from universal exact embezzlement to type $$T(\rho\to\sigma)=\inf_{\mathcal N\in \mathrm{LOCC}} T(\mathcal N(\rho),\sigma).$$4 factors (Liu, 5 Sep 2025).

Exact universal constructions are now explicit in $$T(\rho\to\sigma)=\inf_{\mathcal N\in \mathrm{LOCC}} T(\mathcal N(\rho),\sigma).$$5-algebraic language. A single fixed catalyst on infinite tensor products of CAR algebras can exactly embezzle arbitrary bipartite pure states via local $$T(\rho\to\sigma)=\inf_{\mathcal N\in \mathrm{LOCC}} T(\mathcal N(\rho),\sigma).$$6-automorphisms of Hilbert-hotel type. In the dense-state case the associated GNS von Neumann algebra is type $$T(\rho\to\sigma)=\inf_{\mathcal N\in \mathrm{LOCC}} T(\mathcal N(\rho),\sigma).$$7; exact embezzlement of all states can be achieved as well, but only by moving to a non-separable $$T(\rho\to\sigma)=\inf_{\mathcal N\in \mathrm{LOCC}} T(\mathcal N(\rho),\sigma).$$8-algebra (Liu, 12 Jun 2025).

5. Nonlocal games, dimension witnesses, and universality for strategies

Entanglement embezzlement has become a mechanism for certification in nonlocal games. Coladangelo constructs a two-player game $$T(\rho\to\sigma)=\inf_{\mathcal N\in \mathrm{LOCC}} T(\mathcal N(\rho),\sigma).$$9 with Alice question set of size $\{\chi_n\}$0, Bob question set of size $\{\chi_n\}$1, and answer alphabets of size $\{\chi_n\}$2, combining a $\{\chi_n\}$3-dimensional CHSH-type self-test, a tilted-CHSH self-test, and a consistency test. The honest strategy starts from

$\{\chi_n\}$4

uses the qutrit Bell pair for the $\{\chi_n\}$5-CHSH branch, and invokes an embezzlement unitary on the tilted-CHSH branch to approximate the ideal state

$\{\chi_n\}$6

The resulting value satisfies

$\{\chi_n\}$7

so $\{\chi_n\}$8-optimal play is achievable with local dimension $\{\chi_n\}$9. Conversely, any strategy with value at least $$\lim_{n\to\infty} T(\chi_n\to \chi_n\otimes \sigma)=0$$0 must use dimension

$$\lim_{n\to\infty} T(\chi_n\to \chi_n\otimes \sigma)=0$$1

Thus $$\lim_{n\to\infty} T(\chi_n\to \chi_n\otimes \sigma)=0$$2 is a two-player dimension witness with exponential tradeoff, and the same analysis yields

$$\lim_{n\to\infty} T(\chi_n\to \chi_n\otimes \sigma)=0$$3

giving an elementary embezzlement-based proof of the non-closure of the set of quantum correlations (1904.02350).

At the level of interactive strategies, embezzlement states are universal for two-prover nonlocal games. Any strategy for a $$\lim_{n\to\infty} T(\chi_n\to \chi_n\otimes \sigma)=0$$4-round two-prover game using a shared state on $$\lim_{n\to\infty} T(\chi_n\to \chi_n\otimes \sigma)=0$$5 qubits and winning with probability $$\lim_{n\to\infty} T(\chi_n\to \chi_n\otimes \sigma)=0$$6 can be simulated, for any $$\lim_{n\to\infty} T(\chi_n\to \chi_n\otimes \sigma)=0$$7, by a strategy using a van Dam–Hayden embezzlement state on $$\lim_{n\to\infty} T(\chi_n\to \chi_n\otimes \sigma)=0$$8 qubits and winning with probability at least $$\lim_{n\to\infty} T(\chi_n\to \chi_n\otimes \sigma)=0$$9. Consequently,

$\chi$00

so restricting the provers to embezzlement states does not change these entangled proof-system classes (Oliveira, 2010).

These results illustrate two distinct certification roles. In dimension-witness constructions, embezzlement is the only way to reconcile incompatible self-tested states coherently. In complexity-theoretic universality, it provides a canonical shared resource from which arbitrary entangled prover strategies can be approximated.

6. Type III factors, quantum fields, multipartite and Gaussian regimes, and complexity

The operator-algebraic classification of embezzlement is controlled by invariants of the underlying factor. For a normal state $\chi$01 on a von Neumann algebra $\chi$02, the performance functional

$\chi$03

measures worst-case catalytic power. For factors, $\chi$04; semifinite factors admit no embezzling states and satisfy $\chi$05 for all normal $\chi$06, whereas type $\chi$07 factors with $\chi$08 admit embezzling states. The worst-case value is

$\chi$09

for type $\chi$10, and $\chi$11 for type $\chi$12; thus type $\chi$13 factors are universal embezzlers, meaning every normal state is embezzling (Luijk et al., 2024).

This dovetails with the LOCC classification of pure-state entanglement for commuting factors. In bipartite systems modeled by commuting factors in Haag duality, all pure states have infinite single-shot entanglement iff the local factors are not type I; type III factors are characterized by arbitrary-precision LOCC transitions between any two pure states; and for type $\chi$14 factors this already holds without classical communication (Luijk et al., 2024). A plausible implication is that embezzlement in type $\chi$15 is the catalytic counterpart of a much broader collapse of pure-state entanglement ordering.

For relativistic quantum field theory, the conclusion is operational. Wedge algebras, and under additional conditions local algebras of bounded regions, are type $\chi$16 factors; hence they are universal embezzlers. In particular, the vacuum and other states can be used to embezzle arbitrary finite-dimensional entangled states by local operations in spacelike separated regions. The same framework yields

$\chi$17

so for type $\chi$18 factors, where $\chi$19, one obtains maximal CHSH violation $\chi$20 (Luijk et al., 2024).

Multipartite embezzlement extends these ideas beyond two parties. Finite-dimensional approximations of multipartite embezzling states form multipartite embezzling families, but not every embezzling family converges to an embezzling state. An additional consistency condition distinguishes the van Dam–Hayden family from the Leung–Toner–Watrous family; the latter admits a multipartite generalization whose limit is a multipartite system of commuting type $\chi$21 factors on which every state is embezzling (Luijk et al., 2024).

In fermionic Gaussian systems, embezzlement becomes a spectral property of covariance matrices. If a covariance matrix $\chi$22 has $\chi$23-dense spectrum, then for any $\chi$24 Gaussian target covariances $\chi$25,

$\chi$26

This shows that ground states of non-interacting, critical fermions in one spatial dimension form natural Gaussian embezzling resources, and it suggests that the embezzling property is generic among fermionic Gaussian states with sufficiently dense covariance spectra (Kera et al., 19 Sep 2025).

Finally, embezzlement is kinematically powerful but dynamically costly. Under general locality assumptions on the circuit model, the complexity of embezzlement grows with the target entanglement and the precision. A generic lower bound derived from entropy-growth constraints is

$\chi$27

where $\chi$28 is the embezzled entanglement and $\chi$29 is the local catalyst dimension; the abstract further states that, for a $\chi$30 critical resource system, an exponentially growing lower bound can be derived. In this sense, circuit complexity acts as a physical obstruction to perfect embezzlement even when universal embezzlers exist kinematically (Schwartzman, 2024).