Entanglement Quasiprobabilities in Quantum Systems

• The entanglement quasiprobability framework decomposes a quantum state into product projectors, where nonnegative weights indicate separability and negative weights certify entanglement.
• It employs a separability‐eigenvalue problem and Gram-matrix inversion to optimize the product-state decomposition, enabling precise detection of quantum entanglement.
• Experimental and deep-learning approaches have validated the method, demonstrating its effectiveness in reconstructing entangled states even from incomplete measurements.

Searching arXiv for relevant papers on entanglement quasiprobabilities and related formulations. Entanglement quasiprobabilities are quasiprobability distributions over pure product projectors whose sign structure provides a necessary and sufficient characterization of bipartite, and in suitable extensions multipartite, entanglement. In this framework, a density operator $\rho$ is expanded in separable rank-one operators, and separability is equivalent to the existence of a nonnegative distribution, whereas unavoidable negative weights certify inseparability. The formalism is constructed through the separability-eigenvalue problem, which selects an optimized product-state decomposition and thereby converts entanglement verification into the analysis of a quasiprobability representation closely analogous, but not identical, to the Glauber–Sudarshan $P$-representation of quantum optics [(Sperling et al., 2012); (Bohmann et al., 2017); (Thomas et al., 2017)]. Subsequent work established explicit reconstructions for noisy continuous-variable and finite-dimensional states, an experimental realization for polarization Bell states, and a deep-learning approach that infers entanglement quasiprobabilities from noisy and incomplete local measurements (Sperling et al., 2018, Li et al., 19 Mar 2026).

1. Conceptual definition and relation to separability

For a bipartite system on $\mathcal H_A\otimes\mathcal H_B$, the central object is a decomposition of the density operator into product projectors,

$$\rho=\sum_n q_n\,|\phi_n^a\rangle\langle\phi_n^a|\otimes|\psi_n^b\rangle\langle\psi_n^b| ,$$

or, in continuous form,

$$\rho=\int d\mu(a,b)\,P_{\rm Ent}(a,b)\,|a,b\rangle\langle a,b| .$$

The coefficients $\{q_n\}$, or equivalently $P_{\rm Ent}(a,b)$, define the entanglement quasiprobability distribution (Li et al., 19 Mar 2026, Thomas et al., 2017).

The entanglement criterion is exact. If $\rho$ is separable, then there exists a decomposition with all coefficients nonnegative. If $\rho$ is entangled, then every decomposition of this form necessarily contains at least one negative coefficient. Hence negativity of the optimized entanglement quasiprobability is both necessary and sufficient for entanglement [(Sperling et al., 2012); (Thomas et al., 2017); (Sperling et al., 2018)].

This formulation is conceptually parallel to optical phase-space methods. In the Glauber–Sudarshan representation, negativities or strong singularities of $P(\alpha)$ signal nonclassicality of a field. Entanglement quasiprobabilities transpose this logic from coherent-state expansions to product-state expansions. The crucial distinction is that $P$0 is tailored to separability rather than to optical classicality: a state may possess a singular optical $P$1-function while remaining separable, and in such cases the entanglement quasiprobability can remain nonnegative (Bohmann et al., 2017, Thomas et al., 2017).

A related but distinct line of work introduced operational quasiprobabilities for qudits and a marginal quasiprobability that acts as an entanglement witness for two qudits. That construction is witness-like rather than a full necessary-and-sufficient separability characterization, and therefore should not be conflated with the entanglement-quasiprobability formalism based on separability eigenvectors (Ryu et al., 2012).

2. Separability-eigenvalue construction

The optimized construction of entanglement quasiprobabilities proceeds from the separability-eigenvalue problem. Given a Hermitian operator $P$2, one defines the local reductions

$P$3

The separability-eigenvalue equations are

$P$4

When $P$5, their solutions $P$6 provide the product vectors relevant to the decomposition of the state itself (Bohmann et al., 2017, Sperling et al., 2018).

From these solutions one forms a Gram matrix and solves a linear system for the quasiprobability coefficients. The standard workflow is: $P$7 in the formulation summarized for noisy N00N states (Bohmann et al., 2017), or

$P$8

in the formulations used for dephased Werner states, squeezed light, and experimental reconstruction [(Thomas et al., 2017); (Sperling et al., 2012); (Sperling et al., 2018)]. One then solves

$P$9

for the coefficient vector $\mathcal H_A\otimes\mathcal H_B$0, with $\mathcal H_A\otimes\mathcal H_B$1. If $\mathcal H_A\otimes\mathcal H_B$2 is singular, one uses an inverse or pseudoinverse and projects out the kernel to obtain an optimized solution (Bohmann et al., 2017, Thomas et al., 2017).

The resulting decomposition,

$\mathcal H_A\otimes\mathcal H_B$3

is optimized in the sense that negativities appearing in the coefficients are directly attributable to entanglement. In this representation, negative entries are necessary and sufficient for inseparability [(Bohmann et al., 2017); (Sperling et al., 2012); (Thomas et al., 2017)].

The relation to entanglement witnesses is explicit. Any witness can be written as $\mathcal H_A\otimes\mathcal H_B$4, where $\mathcal H_A\otimes\mathcal H_B$5. Witnesses therefore probe a single extremal direction in the separability-eigenvalue geometry, whereas the quasiprobability reconstruction uses the full set of separability eigenvectors. This suggests why entanglement quasiprobabilities can detect states that evade simpler witness constructions (Bohmann et al., 2017).

3. Canonical examples: squeezed light, N00N states, and Werner states

The earliest detailed applications treated non-Gaussian continuous-variable radiation fields. For a two-mode squeezed vacuum undergoing one-mode Gaussian dephasing, the state

$\mathcal H_A\otimes\mathcal H_B$6

with $\mathcal H_A\otimes\mathcal H_B$7 was analyzed by truncation to finite Fock subspaces. Solving the separability-eigenvalue equations and reconstructing $\mathcal H_A\otimes\mathcal H_B$8 yielded explicit negative quasiprobability entries whenever the state remained entangled (Sperling et al., 2012). In a two-qubit truncation at $\mathcal H_A\otimes\mathcal H_B$9 and $$\rho=\sum_n q_n\,|\phi_n^a\rangle\langle\phi_n^a|\otimes|\psi_n^b\rangle\langle\psi_n^b| ,$$0, the reported values include $$\rho=\sum_n q_n\,|\phi_n^a\rangle\langle\phi_n^a|\otimes|\psi_n^b\rangle\langle\psi_n^b| ,$$1, while increasing dephasing suppresses the negativities until they vanish within numerical precision (Sperling et al., 2012).

For ideal N00N states,

$$\rho=\sum_n q_n\,|\phi_n^a\rangle\langle\phi_n^a|\otimes|\psi_n^b\rangle\langle\psi_n^b| ,$$2

the formalism was extended to noisy variants including mixing, constant or fluctuating losses, and dephasing (Bohmann et al., 2017). Under Gaussian dephasing in mode $$\rho=\sum_n q_n\,|\phi_n^a\rangle\langle\phi_n^a|\otimes|\psi_n^b\rangle\langle\psi_n^b| ,$$3, the state becomes

$$\rho=\sum_n q_n\,|\phi_n^a\rangle\langle\phi_n^a|\otimes|\psi_n^b\rangle\langle\psi_n^b| ,$$4

with

$$\rho=\sum_n q_n\,|\phi_n^a\rangle\langle\phi_n^a|\otimes|\psi_n^b\rangle\langle\psi_n^b| ,$$5

The reconstructed quasiprobabilities are

$$\rho=\sum_n q_n\,|\phi_n^a\rangle\langle\phi_n^a|\otimes|\psi_n^b\rangle\langle\psi_n^b| ,$$6

The persistence of the four negative entries $$\rho=\sum_n q_n\,|\phi_n^a\rangle\langle\phi_n^a|\otimes|\psi_n^b\rangle\langle\psi_n^b| ,$$7 for any finite $$\rho=\sum_n q_n\,|\phi_n^a\rangle\langle\phi_n^a|\otimes|\psi_n^b\rangle\langle\psi_n^b| ,$$8 shows that entanglement survives arbitrary partial dephasing and disappears only in the limit $$\rho=\sum_n q_n\,|\phi_n^a\rangle\langle\phi_n^a|\otimes|\psi_n^b\rangle\langle\psi_n^b| ,$$9 (Bohmann et al., 2017).

A second N00N example treated correlated fluctuating atmospheric losses for $$\rho=\int d\mu(a,b)\,P_{\rm Ent}(a,b)\,|a,b\rangle\langle a,b| .$$0, with mixed output

$$\rho=\int d\mu(a,b)\,P_{\rm Ent}(a,b)\,|a,b\rangle\langle a,b| .$$1

The corresponding quasiprobabilities contain four entries $$\rho=\int d\mu(a,b)\,P_{\rm Ent}(a,b)\,|a,b\rangle\langle a,b| .$$2, so entanglement survives whenever $$\rho=\int d\mu(a,b)\,P_{\rm Ent}(a,b)\,|a,b\rangle\langle a,b| .$$3, that is, for any nonzero transmissivity (Bohmann et al., 2017).

Dephased Werner states provide an especially important finite-dimensional benchmark because they expose the distinction between complete and incomplete entanglement criteria. For

$$\rho=\int d\mu(a,b)\,P_{\rm Ent}(a,b)\,|a,b\rangle\langle a,b| .$$4

local dephasing introduces coefficients

$$\rho=\int d\mu(a,b)\,P_{\rm Ent}(a,b)\,|a,b\rangle\langle a,b| .$$5

In the qubit case, the smallest quasiprobability component is

$$\rho=\int d\mu(a,b)\,P_{\rm Ent}(a,b)\,|a,b\rangle\langle a,b| .$$6

whose negativity threshold

$$\rho=\int d\mu(a,b)\,P_{\rm Ent}(a,b)\,|a,b\rangle\langle a,b| .$$7

coincides with the PPT bound (Thomas et al., 2017). In the qutrit case, the dominant negative component yields the threshold

$$\rho=\int d\mu(a,b)\,P_{\rm Ent}(a,b)\,|a,b\rangle\langle a,b| .$$8

while the PPT criterion gives

$$\rho=\int d\mu(a,b)\,P_{\rm Ent}(a,b)\,|a,b\rangle\langle a,b| .$$9

For Gaussian dephasing with $\{q_n\}$0 and $\{q_n\}$1, there exists a nonempty interval

$\{q_n\}$2

of PPT-entangled, i.e. bound-entangled, states that are detected by the entanglement quasiprobability but not by partial transposition (Thomas et al., 2017).

4. Relation to witnesses, PPT, and optical quasiprobabilities

Entanglement quasiprobabilities occupy a distinct position among entanglement criteria. Entanglement witnesses can be practical, but they test specific operator directions and generally provide only sufficient conditions. The PPT criterion is necessary and sufficient only for $\{q_n\}$3 and $\{q_n\}$4, and it fails to certify bound entanglement in higher dimensions. Wigner-function negativity is not a reliable entanglement diagnostic; for example, the two-mode squeezed vacuum has a positive Gaussian Wigner function (Thomas et al., 2017, Sperling et al., 2018).

By contrast, the defining feature of the entanglement-quasiprobability formalism is completeness: $\{q_n\}$5 if and only if the state is separable, and negative values occur if and only if the state is entangled [(Sperling et al., 2012); (Thomas et al., 2017); (Sperling et al., 2018)]. The qutrit Werner example makes this distinction concrete by detecting PPT-entangled states through negative quasiprobability weights (Thomas et al., 2017).

The relation to optical nonclassicality is more subtle. In multimode quantum optics, regularized phase-space quasiprobabilities $\{q_n\}$6 were introduced to detect general quantum correlations encoded in the multimode Glauber–Sudarshan function (Agudelo et al., 2012). Negativities of $\{q_n\}$7 are necessary and sufficient for nonclassical multimode correlations, and any entangled state must have a nonclassical multimode $\{q_n\}$8-function (Agudelo et al., 2012). However, $\{q_n\}$9 does not by itself identify the type of correlation; local nonclassicality, discord-like correlations, and entanglement can all contribute. Entanglement quasiprobabilities remove this ambiguity by using arbitrary product vectors as the reference basis and directly targeting separability (Bohmann et al., 2017).

An instructive example is the fully dephased N00N mixture

$P_{\rm Ent}(a,b)$0

Its optical $P_{\rm Ent}(a,b)$1-function is highly singular, yet its entanglement quasiprobability is strictly nonnegative, correctly identifying the state as separable (Bohmann et al., 2017). This contrast shows that optical nonclassicality and entanglement, while related, are not coextensive notions.

A complementary representation was later developed using quasistates. In that approach, negativity can be shifted from the distribution to nonpositive semidefinite local operators, yielding classical probabilities multiplied by tensor-product quasistates. The corresponding theorem states, informally, that entanglement is present if and only if any decomposition with nonnegative weights necessarily uses unphysical local quasistates (Sperling et al., 2018). This suggests a dual viewpoint: one may encode nonclassicality either in the coefficients or in the operator-valued basis.

5. Experimental reconstruction

The first experimental reconstruction of an entanglement quasiprobability was reported for two-qubit polarization states (Sperling et al., 2018). The experiment prepared the Bell singlet

$P_{\rm Ent}(a,b)$2

using Type-II PDC in a periodically poled Ti:PPLN waveguide in a Sagnac-loop geometry, with orthogonally polarized photon pairs at $P_{\rm Ent}(a,b)$3 produced by bidirectional pumping at $P_{\rm Ent}(a,b)$4 (Sperling et al., 2018).

The reconstruction protocol combined tomographic sampling with a standard-form reduction for two qubits. Coincidence counts $P_{\rm Ent}(a,b)$5 were measured for $P_{\rm Ent}(a,b)$6, arranged into a $P_{\rm Ent}(a,b)$7 matrix, and converted into the sampled correlation matrix $P_{\rm Ent}(a,b)$8 by

$P_{\rm Ent}(a,b)$9

with a specified $\rho$0 sampling matrix $\rho$1 whose rows correspond to $\rho$2 (Sperling et al., 2018). Local reversible operations $\rho$3 brought the state to the standard form

$\rho$4

for which the exact entanglement-quasiprobability decomposition in the local Pauli eigenbases is known (Sperling et al., 2018).

The reconstructed distribution exhibited several negative entries with high significance. The reported minimum was

$\rho$5

corresponding to a negativity of up to $\rho$6, thereby certifying entanglement directly from the quasiprobability representation without correcting for imperfections (Sperling et al., 2018). As a benchmark, maximum-likelihood tomography yielded fidelity $\rho$7 and purity $\rho$8, while the PPT test gave $\rho$9 (Sperling et al., 2018). A separable reference state reconstructed with the same method showed no negative entries within errors (Sperling et al., 2018).

This experiment is significant because it moved the formalism from a purely constructive criterion based on reconstructed density operators to a laboratory protocol in which entanglement is visualized directly as negative quasiprobability components. A plausible implication is that the method is especially attractive when interpretability of the reconstructed correlation structure is as important as binary certification.

6. Computational scaling and machine-learning reconstruction

A recurring limitation of the traditional approach is that it typically requires full knowledge of $\rho$0, either from tomography or from analytic modeling, followed by solution of the separability-eigenvalue problem and Gram-matrix inversion [(Sperling et al., 2012); (Thomas et al., 2017); (Sperling et al., 2018)]. The details summarized across these works repeatedly identify SEP solution, construction of a sufficiently rich set of product vectors, and inversion or pseudoinversion of $\rho$1 as numerically demanding steps, especially in larger Hilbert spaces (Thomas et al., 2017).

A 2026 development addressed this bottleneck with a deep-learning framework, the Entanglement Quasiprobability Prediction Network with Adaptive Incomplete Measurement Encoding (EQPs-AIME-Net), which reconstructs entanglement quasiprobabilities directly from incomplete local projective measurements (Li et al., 19 Mar 2026). The input consists of measured probabilities $\rho$2 together with a numerical encoding of the chosen projectors $\rho$3, processed by a residual neural network with shortcut connections and a final linear output layer producing the EQP weights $\rho$4 for a fixed ansatz set of separable projectors (Li et al., 19 Mar 2026).

The training target is obtained offline from full-tomography-based ground-truth EQPs on random training states, and the loss function is the mean squared error

$\rho$5

No positivity constraint is imposed on the outputs, so the network learns negative weights when appropriate (Li et al., 19 Mar 2026).

The reported numerical benchmarks for two and three qubits show substantial gains over MaxLik and MLME under incomplete data. For two qubits, with up to 36 projectors drawn from mutually unbiased bases, the network achieved $\rho$6 versus $\rho$7 and $\rho$8, with coefficient of variation $\rho$9 versus $P(\alpha)$0 and $P(\alpha)$1 (Li et al., 19 Mar 2026). For out-of-distribution Werner states in the sparse regime of 2–10 projectors, the RMSE fell from approximately $P(\alpha)$2 to approximately $P(\alpha)$3, while the baseline methods fluctuated at $P(\alpha)$4 (Li et al., 19 Mar 2026). For three qubits, the network achieved RMSE values in $P(\alpha)$5, with mean $P(\alpha)$6, while MaxLik and MLME degraded to $P(\alpha)$7–$P(\alpha)$8 in the sparse limit (Li et al., 19 Mar 2026).

Experimental validation on polarization-entangled photon pairs at $P(\alpha)$9 showed that the network reproduced the correct sign pattern of the negative EQP components already at 10 projectors, whereas MaxLik required $P$00 projectors to avoid sign-flips. Reported entanglement-detection success exceeded $P$01 for the network at 10 projectors, versus $P$02 for MaxLik (Li et al., 19 Mar 2026). This suggests that machine-learning-assisted EQP reconstruction may mitigate the tomography bottleneck without changing the underlying separability criterion.

7. Extensions, generalizations, and open technical issues

The formalism extends beyond bipartite finite-dimensional settings. For $P$03 parties, one generalizes the separability-eigenvalue equations to operators on $P$04, obtaining a multipartite expansion

$P$05

whose negativity again is necessary and sufficient for entanglement across every partition (Bohmann et al., 2017). Witnesses of $P$06-separability can then be built from maximal separability-eigenvalues over product vectors respecting the corresponding partition structure (Bohmann et al., 2017).

In continuous-variable systems, the dephased squeezed-light example demonstrated that the method remains applicable to non-Gaussian states and can be implemented after finite-dimensional truncation in Fock space (Sperling et al., 2012). For discrete systems, exact closed forms are known for families such as Bell-diagonal two-qubit states and dephased Werner states (Thomas et al., 2017, Sperling et al., 2018, Sperling et al., 2018).

The main technical limitations recur across the literature. They include the difficulty of solving the separability-eigenvalue problem in high dimension, the numerical cost of generating a sufficiently large product-vector set, the poor scaling of Gram-matrix inversion or SDP-based optimization, and, in standard implementations, the need for full state reconstruction (Thomas et al., 2017, Sperling et al., 2018). Machine-learning methods address only part of this difficulty: they reduce measurement and reconstruction overhead, but their output remains tied to a predetermined separable ansatz set and to the training distribution (Li et al., 19 Mar 2026).

A common misconception is that entanglement quasiprobabilities are merely another family of entanglement witnesses. The published constructions show otherwise: witnesses correspond to inequalities derived from extremal separability eigenvalues, whereas entanglement quasiprobabilities reconstruct a full decomposition whose negativity is equivalent to entanglement itself (Bohmann et al., 2017, Sperling et al., 2018). Another possible misconception is that any quasiprobability negativity in phase space should be interpreted as entanglement; the contrast between $P$07, the optical $P$08-function, operational qudit quasiprobabilities, and $P$09 demonstrates that the meaning of negativity depends on the underlying operator basis and on the operational or geometric constraints built into the representation [(Ryu et al., 2012); (Agudelo et al., 2012); (Bohmann et al., 2017)].

Taken together, these developments place entanglement quasiprobabilities at the intersection of convex geometry, quantum-state representation theory, and experimental entanglement verification. Their distinctive contribution is not merely to witness entanglement, but to encode separability itself as positivity of an optimized quasiprobability distribution.