Post-Selected Entanglement Witnesses

• Post-selected entanglement witnesses are schemes that refine detection by conditioning measurement data through optimized separability bounds and auxiliary observables.
• They employ methods such as mirrored constructions, ultrafine techniques, and subspace restrictions to extract more information from the same experimental statistics.
• These approaches improve efficiency by reducing the need for additional measurement settings and allowing post-experimental optimization of witness parameters.

Searching arXiv for the cited and closely related papers on post-processed / post-selected entanglement witnesses. Post-selected entanglement witnesses are entanglement-detection schemes in which the witness value is not fixed solely by a single pre-chosen linear observable on the full ensemble, but is refined by conditioning, constrained optimization, subspace restriction, or classical post-processing of fixed measurement data. In the literature, the label covers several closely related mechanisms: extracting two-sided separability bounds from one observable, optimizing over a family of witnesses after data acquisition, conditioning witness bounds on auxiliary observables, restricting fidelity tests to a relevant subspace, and genuinely heralding entangled states by measurement outcomes such as photon detections. The unifying theme is that more entanglement information is obtained from the same experimental statistics, or from a conditional subset of runs, than in a conventional single-threshold witness test (Bae et al., 2018).

1. Definitions and conceptual scope

A standard entanglement witness is a Hermitian operator $W$ such that

$$\operatorname{Tr}(W\sigma)\ge 0 \quad \forall \text{ separable }\sigma,$$

and

$\operatorname{Tr}(W\rho)<0$

for at least one entangled state $\rho$. Geometrically, witnesses are supporting hyperplanes of the convex set of separable states. Several of the works associated with post-selected or post-processed witnessing retain this definition but enlarge the decision rule applied to the measured data (Bae et al., 2018).

The phrase “post-selected entanglement witness” is not used uniformly. In one line of work, the relevant operation is classical post-processing rather than probabilistic discarding of runs. “Entanglement Witness 2.0” shows that a single observable can yield two independent linear witness tests through a lower and an upper separability bound, even though the paper does not use the term post-selection (Bae et al., 2018). In the gravitational-entanglement setting, “with post-processing” means minimizing the expectation value over a continuous witness parameter after measuring a fixed set of five observables; the analysis explicitly distinguishes this from conditioning on subsets of runs (Guff et al., 2021). In measurement-device-independent witnessing, the same fixed dataset $p(1,1|\omega_x,\tau_y)$ is reprocessed with optimized coefficients $\beta^{x,y}$, which is naturally interpreted as selecting the witness after the experiment rather than the events within it (Yuan et al., 2015). By contrast, photon-detection-induced spin squeezing is postselection in the literal sense: the atomic state is conditioned on $\nu$ detected photons, and entanglement is then certified on the conditional state (Rosario et al., 18 Nov 2025).

A second distinction concerns what is being conditioned. Ultrafine entanglement witnessing conditions the separable bound on an auxiliary observable $C$, so that separability is tested inside constrained subsets $\{\sigma:\operatorname{Tr}(C\sigma)\le c\}$ or $\{\sigma:\operatorname{Tr}(C\sigma)\ge c\}$ rather than on the full separable set (Shahandeh et al., 2016). Subspace witnesses condition on a target subspace and maximize fidelity over phase-twisted target states inside that subspace, thereby absorbing local phase errors while keeping the entanglement test fidelity-based (Sun et al., 2019). In prepare-and-measure QKD, effective entanglement witnesses act on a post-selected and renormalized effective state reconstructed from detection events that correspond to valid logical outcomes (Rezazadeh et al., 1 Dec 2025).

2. Single-observable post-processing and constrained separability bounds

The most direct post-processing paradigm is the “mirrored” or “Witness 2.0” construction. For a positive observable $$\operatorname{Tr}(W\sigma)\ge 0 \quad \forall \text{ separable }\sigma,$$0 with $$\operatorname{Tr}(W\sigma)\ge 0 \quad \forall \text{ separable }\sigma,$$1, one defines the separability window

$$\operatorname{Tr}(W\sigma)\ge 0 \quad \forall \text{ separable }\sigma,$$2

Every separable state obeys

$$\operatorname{Tr}(W\sigma)\ge 0 \quad \forall \text{ separable }\sigma,$$3

If the window is a proper subset of the spectrum of $$\operatorname{Tr}(W\sigma)\ge 0 \quad \forall \text{ separable }\sigma,$$4, then either

$$\operatorname{Tr}(W\sigma)\ge 0 \quad \forall \text{ separable }\sigma,$$5

or

$$\operatorname{Tr}(W\sigma)\ge 0 \quad \forall \text{ separable }\sigma,$$6

certifies entanglement. Thus one measured expectation value $$\operatorname{Tr}(W\sigma)\ge 0 \quad \forall \text{ separable }\sigma,$$7 supports two linear witness tests, one below the window and one above it (Bae et al., 2018).

The corresponding mirrored witnesses are

$$\operatorname{Tr}(W\sigma)\ge 0 \quad \forall \text{ separable }\sigma,$$8

with

$$\operatorname{Tr}(W\sigma)\ge 0 \quad \forall \text{ separable }\sigma,$$9

They satisfy

$\operatorname{Tr}(W\rho)<0$0

Experimentally, the same local POVM decomposition implementing $\operatorname{Tr}(W\rho)<0$1 implements both witnesses; the difference is entirely in the classical weights and thresholds (Bae et al., 2018).

The structural physical approximation (SPA) gives a constructive route from an ordinary witness $\operatorname{Tr}(W\rho)<0$2 to such a two-bound picture. The positive SPA is

$\operatorname{Tr}(W\rho)<0$3

with minimal $\operatorname{Tr}(W\rho)<0$4 such that $\operatorname{Tr}(W\rho)<0$5; the negative SPA is

$\operatorname{Tr}(W\rho)<0$6

with $\operatorname{Tr}(W\rho)<0$7 maximal. The paper proves that for a mirrored pair $\operatorname{Tr}(W\rho)<0$8, the p-SPA of $\operatorname{Tr}(W\rho)<0$9 coincides with the n-SPA of $\rho$0, so that one positive observable encodes the separability window and both mirrored witnesses (Bae et al., 2018).

Ultrafine entanglement witnessing generalizes the same logic by constraining the separable set with an auxiliary observable $\rho$1. For a test operator $\rho$2 and threshold $\rho$3, one defines

$\rho$4

where $\rho$5 and $\rho$6 is defined analogously for $\rho$7. Entanglement is certified if

$\rho$8

or

$\rho$9

This conditional bound is stronger than the global separable bound because the optimization is restricted to a subset of separable states compatible with the same auxiliary statistic (Shahandeh et al., 2016).

3. Witness families derived from fixed datasets

A second major paradigm is to reconstruct a whole family of witnesses from one measurement record and choose the most informative member afterward. In the gravitational-entanglement proposal of Bose et al., the target state

$p(1,1|\omega_x,\tau_y)$0

is separable iff $p(1,1|\omega_x,\tau_y)$1 and maximally entangled iff $p(1,1|\omega_x,\tau_y)$2. The optimized fidelity-witness family is

$p(1,1|\omega_x,\tau_y)$3

where

$p(1,1|\omega_x,\tau_y)$4

Expanded in Pauli products,

$p(1,1|\omega_x,\tau_y)$5

Only five non-trivial spin measurements are needed: $p(1,1|\omega_x,\tau_y)$6 After collecting those correlators once, one minimizes $p(1,1|\omega_x,\tau_y)$7 over $p(1,1|\omega_x,\tau_y)$8. This post-processing detects entanglement for any choice of phases in the setup, up to a set of measure zero (Guff et al., 2021).

Measurement-device-independent entanglement witnesses sharpen the same idea under untrusted measurements. From fixed input states $p(1,1|\omega_x,\tau_y)$9 and experimental probabilities $\beta^{x,y}$0, one defines

$\beta^{x,y}$1

The original construction of Buscemi and Branciard provides reliability: for separable shared states and arbitrary measurement devices, $\beta^{x,y}$2 (Branciard et al., 2012). The robust variant of Wu et al. then optimizes the coefficients $\beta^{x,y}$3 after data collection, under entanglement-witness constraints and normalization $\beta^{x,y}$4, so that the same measured probabilities can yield the most negative admissible $\beta^{x,y}$5 compatible with reliability (Yuan et al., 2015). In this sense, the witness itself is post-selected from the data.

Subspace witnesses introduce another family-based construction. For a target subspace $\beta^{x,y}$6 and phase-parametrized state

$\beta^{x,y}$7

the subspace witness value is

$\beta^{x,y}$8

For two-qubit Bell-subspace states, this becomes

$\beta^{x,y}$9

so the witness depends on the coherence magnitude rather than one fixed phase projection. This yields a strictly larger violation than the conventional fidelity witness for states with unknown local phases (Sun et al., 2019).

4. Nonlinear and conditionally improved witnesses

Nonlinear post-processing extends linear witnesses by adding quadratic or higher-order terms built from the same measured statistics. In the mirrored-witness framework, if $\nu$0 arise from a single observable $\nu$1, one can define nonlinear functionals

$\nu$2

and corresponding nonlinear bounds

$\nu$3

For separable states,

$\nu$4

$\nu$5

Thus the lower bound is improved by subtracting a nonlinear term, and the upper bound is improved by adding one, while still using the statistics of the same observable $\nu$6 (Bae et al., 2018).

The orbital-angular-momentum experiment on nonlinear witnesses implements a related strategy starting from a linear PPT witness $\nu$7. For the correlated Bell state $\nu$8, the asymptotic nonlinear witness takes the form

$\nu$9

The experimental result is that one nonlinear witness detects entanglement for essentially the whole phase family of the target mixed state, whereas the linear witnesses $C$0 or $C$1 each fail on roughly half of the phase range (Agnew et al., 2012).

Conditioning can also be explicit and dynamical. In photon-detection-induced entanglement of emitter ensembles, the conditional state after $C$2 detected photons in direction $C$3 is

$C$4

Entanglement is certified by a generalized field-based spin-squeezing parameter

$C$5

with the rigorous criterion

$C$6

Here the entanglement is genuinely postselected: only runs with specified photon-detection records are retained, and successive detections act as a purification process (Rosario et al., 18 Nov 2025).

5. Experimental realizations and operational domains

The operational advantage of post-selected or post-processed witnesses is that they often avoid extra measurement settings. “Entanglement Witness 2.0” emphasizes that mirrored witnesses use the same local measurement decomposition and the same set of POVM elements as the original observable, differing only by classical thresholds $C$7 and $C$8 (Bae et al., 2018). In the gravitational-entanglement proposal, five non-trivial spin measurements replace full two-qubit tomography with 15 local settings, while still allowing an a posteriori optimization over the continuous parameter $C$9 (Guff et al., 2021).

Multipartite experiments use the same principle at larger scale. The intactness witnesses

$\{\sigma:\operatorname{Tr}(C\sigma)\le c\}$0

and the depth witnesses

$\{\sigma:\operatorname{Tr}(C\sigma)\le c\}$1

require only two local measurement settings per party, yet the same global dataset can be classically postprocessed to bound entanglement intactness, entanglement depth, and even infer entanglement partitioning by evaluating witnesses on subsets of parties (1711.01784). The paper is explicit that the free parameters $\{\sigma:\operatorname{Tr}(C\sigma)\le c\}$2 and $\{\sigma:\operatorname{Tr}(C\sigma)\le c\}$3 may be optimized a posteriori.

Prepare-and-measure QKD supplies a further variant. In the COW protocol, effective entanglement is witnessed not on the full optical state but on a post-selected effective two-time-bin state reconstructed from valid logical detection events. The witness family

$\{\sigma:\operatorname{Tr}(C\sigma)\le c\}$4

acts on this effective qubit subspace, and the probabilities used to evaluate it are explicitly renormalized after discarding no-click and other irrelevant events. The paper identifies parameter regions in which $\{\sigma:\operatorname{Tr}(C\sigma)\le c\}$5 is a valid witness and reports clear signatures of effective entanglement in previously obtained COW data (Rezazadeh et al., 1 Dec 2025).

Measurement-device-independent witnessing provides a complementary operational guarantee: no post-selection on detected events is needed, because losses are absorbed into the measurement model and the separable bound remains $\{\sigma:\operatorname{Tr}(C\sigma)\le c\}$6 on raw probabilities. In that setting, the decisive post-processing step is the choice of witness coefficients, not the selection of detection events (Branciard et al., 2012).

6. Limitations, ambiguities, and open problems

A recurrent limitation is that not every witness admits a nontrivial post-selected enhancement of the same type. In the mirrored construction, some would-be mirrors are positive operators rather than entanglement witnesses, so the upper-bound test becomes trivial (Bae et al., 2018). In ultrafine witnessing, the gain depends on whether the constraint operator $\{\sigma:\operatorname{Tr}(C\sigma)\le c\}$7 is incompatible with the test operator $\{\sigma:\operatorname{Tr}(C\sigma)\le c\}$8; if separable $\{\sigma:\operatorname{Tr}(C\sigma)\le c\}$9 and $\{\sigma:\operatorname{Tr}(C\sigma)\ge c\}$0 commute, the ultrafine improvement does not detect entanglement (Shahandeh et al., 2016).

Another limitation is computational. Optimizing witness bounds often requires hard constrained optimizations over separable or $\{\sigma:\operatorname{Tr}(C\sigma)\ge c\}$1-producible states. Mirrored witnesses require computing $\{\sigma:\operatorname{Tr}(C\sigma)\ge c\}$2 and $\{\sigma:\operatorname{Tr}(C\sigma)\ge c\}$3, which is a separability optimization problem (Bae et al., 2018). Optimized MDI witnesses replace exact entanglement-witness constraints by $\{\sigma:\operatorname{Tr}(C\sigma)\ge c\}$4-level witnesses and sampled product-state constraints so that the search becomes a semidefinite program, trading exact reliability for a controlled approximation (Yuan et al., 2015). Entanglement-depth bounds in multipartite experiments are similarly tightened by SDP relaxations (1711.01784).

The term itself also invites misconception. Some schemes called “post-selection” are purely classical post-processing of fixed data and do not discard experimental runs. This is explicit for the gravitational witness family $\{\sigma:\operatorname{Tr}(C\sigma)\ge c\}$5, where all runs and all outcomes are used and only the witness parameter is chosen after the fact (Guff et al., 2021). Subspace witnesses likewise optimize over a family of target phases inside a chosen subspace rather than conditioning on outcome subsets (Sun et al., 2019). By contrast, photon-detection-generated squeezing and effective-entanglement analysis in COW do involve genuine conditioning on selected events or selected Hilbert-space sectors (Rosario et al., 18 Nov 2025).

Open directions stated across the literature include constructing more than two useful bounds from a single observable, understanding the relation of mirrored witnesses to extremal or decomposable/non-decomposable witnesses, extending subspace and ultrafine methods to more general multipartite and higher-dimensional settings, and using witness expectation values not only for detection but also for lower-bound entanglement quantification and QKD security estimates (Bae et al., 2018). Recent work on machine-learning-derived witnesses adds a different post-processing layer: a linear SVM defines a hyperplane in the measured feature space, and the optimized coefficients of local observables directly specify an entanglement witness, permitting feature ranking and systematic reduction of measurement terms for target states such as W states (Greenwood et al., 2021).

Post-selected entanglement witnesses therefore designate not one formalism but a family of closely related strategies. Some turn a single observable into multiple linear or nonlinear tests, some optimize a witness over the same data after the experiment, some impose conditional separability constraints, some restrict the witness to a task-relevant subspace, and some herald entangled states by measurement outcomes. Across these variants, the central objective is unchanged: certify entanglement more efficiently, more robustly, or under more realistic experimental assumptions than a single fixed linear witness permits.