WorksPhysics and Astronomy
How many events does entanglement-based positronium imaging need? Information limits of Compton polarimetry with CZT detectors
Abstract
The degree of polarisation correlation of the two annihilation photons, measured by Compton polarimetry, has been proposed as a PET biomarker of tissue microstructure, but the statistics it requires have not been quantified. We parameterise the measurement by the linear-polarisation correlation coefficient ( maximally entangled, separable), derive the classical Fisher information of double-Compton events and compare it with the quantum Fisher information, relate to the annihilation channels of positronium in tissue, and simulate pixelated CdZnTe (CZT) rings. The model is checked against three published measurements and an independent GATE/Geant4 simulation. A double-Compton event carries at most 0.12-0.14 units of Fisher information about and 0.015 on average, 0.1-0.6% of the quantum limit. Event-wise maximum likelihood reaches the Cramer-Rao bound, whereas the conventional ratio R needs 1.2-2.3 times more events. The observed coefficient is , with a pick-off share of 0.13-0.29 in tissue; it is sensitive to lipid fraction but not to oxygen, and carries no tissue information if pick-off photons are maximally entangled. Published porous-polymer data imply = 0.34-0.58. In a 15-mm CZT module, ordering ambiguity, position errors and electron transport roughly halve the information per event, and sequence reconstruction of multi-site events recovers this loss. With F and an eight-module ring, can be measured to in a 10-mm tumour, but separating tissues by alone requires 6-74 times more decays than are available.
Introduction
Positronium imaging extends positron emission tomography (PET) from where a tracer accumulates to the environment in which positrons annihilate (Bass et al 2023, Moskal et al 2024, 2021). Three observables are under study: the mean ortho-positronium (o-Ps) lifetime, the ratio of three- to two-photon annihilations (Fujimoto et al 2026) and the polarisation correlation of the two annihilation photons (Moskal 2026, Moskal et al 2025, Watts et al 2021). The last, measured by the azimuthal correlation of two Compton scatterings [25] (Snyder et al 1948), has been interpreted as a degree of quantum entanglement. In a porous polymer the measured correlation was smaller than expected for the maximally entangled state but larger than for separable photons, which was explained by assuming that photons from o-Ps pick-off annihilation are not entangled (Moskal et al 2025). Because the pick-off fraction depends on free-volume size and composition, the correlation has been proposed as a biomarker of tissue type, and the first in vivo images of this quantity in a human subject have been reported [23]; a roadmap towards quantum-entanglement PET has been outlined [24].
Three questions remain open. The first is what the measurement determines. A Compton polarimeter is insensitive to circular polarisation, and separable states can reproduce the double-Compton cross section of the entangled state (Clarke et al 2026, Hiesmayr and Moskal 2019). Single-scattering observables built on the Pryce–Ward cross section therefore admit local hidden-variable descriptions and cannot certify entanglement on their own (Clarke et al 2026); lower bounds on Bell-state fidelities that can be extracted from coarse-grained coincidences were derived only recently (Sutter and Hiesmayr 2026), and it has been argued that the full double-Compton distribution of the maximally entangled state is reproduced only by that state (Joos and Kling 2026). The second concerns what happens to the correlation when a photon scatters before detection: experiments disagree (Bordes et al 2024, Ivashkin et al 2023, Tkachev et al 2025, Watts et al 2021), and theoretical treatments of the evolution of the two-photon state are still being developed (AskariPour Ravari and Riazi 2026, Caradonna and Shimazoe 2025, Caradonna et al 2024). The third concerns the pick-off channel: the interpretation of the J-PET result assumes that pick-off photons are separable [23] (Moskal et al 2025), but this has not been derived from the quantum electrodynamics of annihilation with bound molecular electrons and remains to be verified.
Underlying all three is a more practical question that has received little attention: how many events are needed. In X-ray astronomy the statistics of single-photon polarimetry are well developed, through the minimum detectable polarisation and per-event Stokes-parameter estimators (Kislat et al 2015, Weisskopf et al 2010). For two-photon correlations the corresponding analysis has not been done. The most recent proposal for a Bell test with annihilation photons gives an order-of-magnitude rate estimate and defers the required number of events to a dedicated feasibility study (Clarke et al 2026). Counts required to discriminate between two prescribed modulation levels, with finite acceptance, resolution and uniform background, have been derived within an idealised framework for coincidence-conditioned single-arm modulation (Caradonna et al 2026). An estimation-theoretic treatment of the correlation coefficient itself, referred to the quantum limit and to realistic detectors, has not been given. For imaging, the question is sharper still, because tissue contrasts in the correlation coefficient are expected to be of order .
In a related study that is under review (Zhou et al 2026), we combined positronium lifetime, the three-to-two-photon ratio and the polarisation correlation in a source model and found that lifetime plus correlation is the combination that separates microstructure from oxygenation. That work assumed a constant analysing power and a usable fraction of double-Compton events. Here we derive these quantities from first principles and detector simulation and ask: (i) how much information a double-Compton event carries about the correlation coefficient, compared with the quantum limit; (ii) which estimator attains it; (iii) how the observed coefficient depends on annihilation channels, microstructure and oxygen, and on the unresolved pick-off state; (iv) how many usable events a pixelated CZT system delivers per decay; (v) whether in-body scattering biases the measurement; and (vi) whether tissue-level contrasts are reachable in a small-animal setting with . We test the model against published measurements before using it for these predictions.
Methods
Measurement model
We consider two back-to-back 511-keV photons, each undergoing a Compton scattering through polar angle and azimuth , with the angle between the two scattering planes. A single Compton scattering is described by a positive-operator-valued measure parameterised by the analysing power (Clarke et al 2026)
which reaches its maximum at for . Let be the unpolarised Klein–Nishina cross section (constants omitted). For a source without net polarisation, invariant under rotation about the line of flight and under parity, the joint distribution is
which defines the linear-polarisation correlation coefficient . For the maximally entangled annihilation state ; for a separable mixture of orthogonally polarised photons with uniformly distributed orientation ; for uncorrelated photons . The widely used ratio with follows directly. A uniform background fraction in multiplies by . Because the marginal distribution of in (2) does not depend on , all information on resides in .
Equation (2) is the general form under the stated symmetries: the measurement determines one number, the common linear-polarisation correlation. Appendix § 12 shows that, for the Bell-diagonal states compatible with these symmetries, constrains the concurrence only to the interval , because the circular correlation is not measured. Throughout this paper we therefore treat as the measured quantity and refer to entanglement only where this interval is invoked.
Fisher information and quantum limit
For fixed , the conditional density of is with . Its Fisher information about is (Appendix § 13)
The information per double-Compton event is the average of (3) over the scattering-angle distribution , optionally restricted to an angular window. With independent events, the Cramér–Rao bound is .
To assess how much of the available information Compton polarimetry extracts, we computed the quantum Fisher information (QFI) for two one-parameter families of Bell-diagonal states. In family A, the maximally entangled state is mixed with the separable state described above, so that ; in family B, it is mixed with white noise, so that . In both families the eigenbasis does not depend on the parameter, and the QFI equals the classical Fisher information of the Bell-basis weights:
As an intermediate reference we also evaluated (3) with , which corresponds to an ideal linear-polarisation measurement with a continuous azimuth.
Estimators and toy Monte Carlo
We compared three estimators: event-wise maximum likelihood (ML) based on (2); a moment estimator weighted by the analysing power, , which is the two-photon analogue of the per-event Stokes estimators used in X-ray polarimetry ((Kislat et al 2015)); and the ratio formed from counts within of and and converted to with the mean of the accepted events. Events were drawn from (2) for , $0.75 and 0.95, with events per experiment and 300–400 repetitions, using scattering-angle windows of 60–110° and 20–160°.
Annihilation channels and tissue model
Photons from direct annihilation, para-positronium (p-Ps) self-annihilation and O-induced conversion of o-Ps to p-Ps were taken to be maximally entangled (). Pick-off photons were assigned an unknown coefficient . The observed coefficient of the two-photon sample is then
where is the pick-off share of all two-photon annihilations. This is identical to the entanglement observable of Zhou et al (2026), in which the partial-entanglement parameter corresponds to . We evaluated three hypotheses: H, (the working assumption of Moskal et al (2025)); H, ; and H, .
The channel shares were computed with the source model of Zhou et al (2026), an absorbing Markov chain in which o-Ps and p-Ps interconvert through O until annihilation, with parameters derived from published tissue positron-lifetime data (Avachat et al 2024, Moskal 2026, Moskal et al 2021, 2023, Stepanov et al 2020). Microstructure was varied through a shift of the Tao–Eldrup free-volume radius ([30] (Eldrup et al 1981)) and through the lipid fraction, at pO = 30 mmHg. To test how the pick-off hypothesis propagates to oxygen estimation, we recomputed the lower bounds (floors) on (pO) when four microstructure nuisance parameters are free within priors set by the spread between measured tissues (Zhou et al (2026)).
CZT detector Monte Carlo
We modelled a commercial pixelated CZT Compton-camera module (DT-PA05): CdZnTe (CdZnTe, 5.78 g cm), mm, pixels of 2 mm, sub-pixel lateral resolution 0.5 mm and depth resolution 1 mm (both treated as FWHM), and 1% energy resolution at 662 keV. Compton attenuation was computed from the Klein–Nishina cross section with ; photoabsorption was parameterised as cm g, which gives a total cm g at 511 keV. Coherent scattering, fluorescence escape and charge sharing were neglected. Transport within the crystal was unpolarised; the effect of the polarisation correlation on escape probabilities is of second order.
A photon was counted as a usable polarimetric event if (i) its full energy was deposited, (ii) its first interaction was Compton scattering, (iii) its interactions formed exactly two pixel clusters, the first containing only the first interaction, and (iv) both clusters exceeded an energy threshold (10, 15 or 30 keV). The scattering angle was reconstructed from the energy of the first cluster, and the azimuthal resolution was propagated as a modulation factor with , where is the transverse cluster separation. Events with three or more clusters, which would require sequence reconstruction, were counted separately as an upper bound. This idealised model, which assumes correct cluster ordering, was used for the ring geometry and the pre-scattering study.
We simulated a single module, then two rings with a point source at the centre: G8, eight modules in an octagon with crystal front faces at cm, which reflects the 83.5-mm housing; and G16, sixteen crystals of the same size forming a gapless polygon ( cm), which represents tighter packing. Photons could cross between modules. The per-decay Fisher information was the probability that both photons gave usable events multiplied by the mean of (3) over these pairs, with and a time-averaged live fraction of 0.596 for double coincidences at 10 MBq, as in Zhou et al (2026). Random coincidences between usable single events in opposite modules were estimated for a 24-ns window at 10 MBq.
Detector realism
A second single-module simulation removed the main idealisations. (i) Bound-electron Doppler broadening was included in the impulse approximation: the scattered energy was shifted by , with the momentum transfer and the electron momentum projection drawn from a Gaussian of standard deviation 1.5 atomic units (1 and 2 a.u. as variants). (ii) Positions were the energy-weighted centroids of the interactions within each 2-mm pixel, smeared by the sub-pixel lateral (0.5 mm FWHM) and depth (1 mm FWHM) resolutions; energies were smeared by the energy resolution and a full-energy window of was applied. (iii) The order of two clusters was not assumed. Two rules were compared: the larger deposit is the first scattering ((Watts et al 2021)), and, using the incident direction known from the line of response, the order whose geometric scattering angle agrees best with the angle computed from the energies. (iv) Events with three or four clusters were reconstructed by testing all orderings and choosing the one that minimised the summed squared difference between geometric and kinematic scattering cosines along the sequence.
For each accepted event we recorded the true analysing power of the first scattering, the reconstructed angle and the error of the reconstructed scattering-plane axis, which includes ordering errors. The modulation retained by reconstructed events is . Because a real analysis can use only , the effective analysing power was calibrated in 18 bins of , , as would be done with a calibration source; the information per pair was then the average of (3) with over random pairs of accepted events (, ). The ratio of the information per pair of interacting photons to that of the idealised model, taken from the independent simulation of § 2.7, was used to rescale the ring results. One parameter at a time was varied to obtain the sensitivity to the photoabsorption cross section (), pixel size (1 and 3 mm), crystal thickness (10 and 20 mm), threshold (10 and 30 keV), energy resolution (2%) and Doppler width.
The total attenuation coefficient of the model was compared with the tabulated values for obtained from the elemental tables of Hubbell and Seltzer (2004) by mass-weighted addition.
Independent Monte Carlo
The detector model and the correlation model were checked with GATE 10 (Sarrut et al 2022), which is based on Geant4 (Allison et al 2016), using the Livermore polarised electromagnetic physics, which includes Doppler broadening, fluorescence and electron transport. Two simulations were run. In the first, 511-keV photons from a point 10.5 cm in front of a DT-PA05 module were tracked; the deposits were processed with the same clustering, ordering and calibration as above. In the second, positrons annihilated at rest in a 1-mm water sphere between two CZT slabs at cm, with Geant4’s entangled annihilation-photon option enabled or disabled; was fitted by maximum likelihood with (2) to the true first-Compton angles and scattering planes of both photons.
Comparison with published measurements
Three published measurements were reproduced with the model. (i) Watts et al (2021) measured the modulation ratio with two CZT polarimeters ( detectors with sub-millimetre pixels, 4.35 cm from the source), selecting events with two pixels and assigning the first scattering to the larger deposit; they reported for scattering angles of – and for –. We simulated their geometry with the transport model of § 2.5, reconstructed angles and azimuths from pixel centres with their ordering rule, and computed for and . (ii) Ivashkin et al (2023) used scintillator scatterers with 16 azimuthal counters of and scattering angles of –, reporting ; we folded (2) with their angular acceptance. (iii) Moskal et al (2025) measured for photons from the porous polymer XAD-4 at and , with circular angular regions of radius, and quoted ideal and detector-simulated ratios for the maximally entangled and separable cases. We used their simulated ratios as a scale, checked that our model reproduces their separable reference, and inverted their measured ratios to and, through (5) with their pick-off share, to .
In-body pre-scattering
A photon that scatters by a small angle in the body may remain inside the photopeak window, and its correlation after scattering is disputed [23] (Ivashkin et al 2023; Tkachev et al 2025; Watts et al 2021). We therefore treated the post-scattering coefficient as unknown in , so that the bias is bounded by , where is the fraction of accepted pairs with at least one pre-scattered photon. The mouse was modelled as an infinite water cylinder of radius 1.0 or 1.5 cm with the source on the axis or 0.5–1.0 cm off axis. Only Compton scattering in water was tracked, and photons were accepted if they reached a module of G8. The photopeak window was of the CZT energy resolution at 511 keV (FWHM 5.8 keV), which admits scattering angles below ; a case without an energy window was simulated for comparison with detectors of low energy resolution. The quantity of interest is the difference in between source positions, since a depth-dependent bias could mimic a tissue contrast.
Imaging scenarios
We considered with 10 MBq injected in a mouse and acquisition until complete decay. Two regions were used: a 10-mm tumour (0.5 mL at 400 kBq mL, decays) and a liver of about 1.5 mL ( decays). The decays required per region were for a target precision , and for a separation of two regions whose coefficients differ by .
Implementation
All calculations used Python with NumPy; the source model of Zhou et al (2026) additionally used SciPy, and the independent simulations used GATE 10.1. Analytical Fisher information was checked against direct numerical integration, and the QFI expressions against numerical differentiation of the Bell weights. Code and numerical results will be released with the preprint.
Results
Benchmarks
The analysing power reached at , in agreement with Clarke et al (2026). With both photons at , (2) gave for and for , reproducing the values usually quoted (2.85 and 1.63) (Ivashkin et al 2023, Moskal et al 2025). Equation (3) agreed with numerical integration to six digits for between 0 and 1, and (4) with numerical differentiation to five digits.
Comparison with published measurements
Table 1 summarises the comparison. For the CZT polarimeter of Watts et al (2021), the model predicted (70–110°) and 1.98 (93–103°) for , against measured values of and ; for it predicted 1.37 and 1.40. The larger-deposit rule ordered 69% of events correctly, close to the 60% reported, and pixelisation and misordering lowered from 2.06 and 2.15. For the scintillator polarimeter of Ivashkin et al (2023), folding with the counter width gave , against . Both measurements are thus consistent with for direct annihilation in these sources and with our treatment of angular acceptance.
For the J-PET data ((Moskal et al 2025)), our model reproduced the separable reference from their simulated ratios ( and at and ). Their measured ratios correspond to and ; with their pick-off share this gives and . Both values exclude , but they differ from each other by about three standard deviations, and the XAD-4 mixture implied by their own simulation at (0.895) differs from that implied by their stated pick-off share (0.845). The inference therefore depends on how the published ratios were corrected and should be taken as indicative.
How much information a double-Compton event carries
Figure 3 shows over the plane. The information peaks where both photons scatter near , at 0.120 for and 0.139 for , so that would need – such events. The Klein–Nishina distribution, however, places most scatterings at forward angles where is small. Averaged over all angles, a double-Compton event carries 0.0146–0.0157, and requires – events, about nine times more than the optimal-angle estimate.
Figure 3. Fisher information about per double-Compton event as a function of the two scattering angles, for (a) and (b) . White contours: relative density of double-Compton events (25, 50, 75% of maximum). Cross: optimal angle 81.7°.
Figure 4 places these numbers on the scale of the quantum limit. For family A at the QFI is 4.0; an ideal linear-polarisation measurement extracts 0.62 (15.5%); Compton polarimetry at the optimal angles extracts 0.12 (3.0%); and Klein–Nishina-averaged Compton polarimetry extracts 0.015 (0.37%). For family B the corresponding fractions at are 25.8%, 5.0% and 0.61%. The first step reflects the blindness to circular polarisation and the restriction to a single azimuthal outcome; the second reflects ; the third reflects the scattering-angle distribution. As the QFI of family A diverges, whereas the Compton information stays finite, so the efficiency falls further (0.14% at ).
Figure 4. Information about per photon pair: quantum Fisher information, ideal linear-polarisation measurement, Compton polarimetry at the optimal angle and Compton polarimetry averaged over the Klein–Nishina distribution, for (a) family A and (b) family B.
Angular selection and estimators
Selecting a window of scattering angles raises the information per accepted event but discards events, and the product always decreased (Figure 5(a)). Relative to using all angles, the windows –, – and – required 1.9, 5.4 and 14 times as many double-Compton events. A window of – lost only 1%.
Figure 5. (a) Double-Compton events needed for a given precision when a scattering-angle window is imposed, relative to using all angles (). (b) Efficiency of three estimators (Cramér–Rao bound divided by Monte Carlo variance) with a 20–160° window and events.
Event-wise ML and the -weighted moment estimator were unbiased and reached the Cramér–Rao bound within Monte Carlo error at all three values of (efficiency 0.96–1.03; Figure 5(b)). The ratio had an efficiency of 0.43–0.45 with the – window, that is it needed 2.2–2.3 times as many events. With the narrower – window the penalty was smaller (variance ratio 1.24) because varies less within the window.
Annihilation channels, microstructure and the pick-off hypothesis
Across the tissue parameter set, the pick-off share of two-photon annihilations ranged from 0.13 (porcine liver) to 0.29 (human adipose, (Moskal et al. 2023)), giving between 0.854 and 0.934 under (Table 2). Under the coefficients moved halfway towards 1, and under they were all exactly 1.
| Tissue (source of lifetime data) | H0 () | H () | H1 () | |
| Human adipose (Moskal et al (2021)) | 0.217 | 0.891 | 0.946 | 1 |
| Human cardiac myxoma (Moskal et al (2021)) | 0.228 | 0.886 | 0.943 | 1 |
| Human adipose (Moskal et al (2023)) | 0.292 | 0.854 | 0.927 | 1 |
| Human cardiac myxoma (Moskal et al (2023)) | 0.189 | 0.905 | 0.953 | 1 |
| Porcine adipose (Avachat et al (2024)) | 0.187 | 0.907 | 0.953 | 1 |
| Porcine liver (Avachat et al (2024)) | 0.132 | 0.934 | 0.967 | 1 |
| Porcine muscle (Avachat et al (2024)) | 0.148 | 0.926 | 0.963 | 1 |
Table 2. Pick-off share of two-photon annihilations and observed correlation coefficient at mmHg under three hypotheses for the pick-off coefficient.
The coefficient was sensitive to composition but almost insensitive to oxygen (Figure 2). In muscle under , raising from 20 to 40 mmHg changed by , a free-volume radius shift of nm by , and a lipid fraction of 0.1 by . Contrasts between tissue pairs were (muscle vs porcine adipose), 0.0080 (muscle vs liver) and 0.0055 (cardiac myxoma vs adipose) under , and half of these under .
Figure 2. Observed correlation coefficient as a function of (a) pO2, (b) free-volume radius shift and (c) lipid fraction, for three pick-off hypotheses. Muscle base tissue in (b) and (c), pO2 = 30 mmHg.
Table 3 shows how the hypothesis propagates to oxygen estimation. In the analysis of Zhou et al. (2026), adding the correlation to lifetime events lowers the floor on from 2214 to 100 mmHg in muscle and from 339 to 14 mmHg in adipose. These floors were identical under and , because the floor depends on how the observables vary with the parameters and not on the size of that variation; the number of decays needed to approach the floor scales as . Under the correlation coefficient carries no information and the floor reverts to that of lifetime alone.
| Hypothesis | Muscle | Adipose | ||||
| L | L+C | L+R+C | L | L+C | L+R+C | |
| H0 | 2214 | 100 | 8.8 | 339 | 14 | 2.7 |
| H | 2214 | 100 | 8.8 | 339 | 14 | 2.7 |
| H1 | 2214 | 2214 | 1661 | 339 | 339 | 275 |
Table 3. Floors on (mmHg; decays, microstructure nuisances with inter-tissue priors, long-axial-field-of-view scenario of Zhou et al (2026)) under the three pick-off hypotheses. L: lifetime events; C: double-Compton events; R: three-photon events.
Usable events in CZT
In a single 15-mm module (Table 4), 47.0% of incident 511-keV photons interacted; of these, 50.4% deposited their full energy and 84.1% interacted first by Compton scattering. Only 17.3% of interacting photons produced a clean two-cluster, full-energy event, and 6.6% did so within the window commonly used for . The main loss is lateral escape: a photon scattered near carries about 250 keV sideways and must be absorbed within the same 44-mm crystal. Thicknesses of 10 and 20 mm gave windowed fractions of 6.1% and 6.8%, and thresholds between 10 and 30 keV changed the fractions by less than 0.002. Clean events had an approximately symmetric distribution of reconstructed angles centred near (Figure 6(a)), unlike the forward-peaked distribution of first scatterings. Within the window, the effective analysing power was 0.391, and the 2-mm pixels reduced the modulation by less than 1% (median factor 0.994). A further 8.5% of interacting photons produced three or more clusters with full energy.
Table 4: Usable polarimetric events in DT-PA05 CZT ( keV). Single-module fractions are relative to photons that interact in the crystal; ring quantities are per decay. Information per decay includes a live fraction of 0.596 and .
| Quantity | Single module | G8 ring | G16 gapless ring |
| Interaction probability (photon / both photons) | 0.470 | ||
| Clean two-cluster fraction, all angles | 0.173 | 0.180 | – |
| Clean two-cluster fraction, | 0.066 | 0.067 | – |
| Usable pairs per decay, all angles | – | ||
| Cross-module share of usable events | – | 1.7% | 5.1% |
| Information per usable pair, all angles | 0.023 | 0.022 | 0.024 |
| Fisher information per decay, all angles (ML) | – | ||
| Fisher information per decay, | – | – | |
| Random coincidences at 10 MBq | – | 0.6% | 0.5% |
Table 4.
Figure 6. (a) Reconstructed scattering angle of clean two-cluster, full-energy events in a DT-PA05 module, with the Klein–Nishina distribution of first scatterings (peak-scaled) and the window. (b) Fisher information about per decay (live time included) assumed previously and obtained here with the idealised model and with the GATE-corrected response using two clusters only or including sequence reconstruction of events with three or four clusters.
In the eight-module ring, the probability that both photons interacted was per decay; usable pairs occurred at per decay over all angles and within the window. Only 1.7% of usable events involved two modules, because the housings leave wide gaps between crystals. The Fisher information per decay including live time was with all angles and ML, and with the window (Figure 6(b)). In the gapless 16-module ring, both photons interacted with probability , 5.1% of usable events crossed modules, and the information per decay rose to . Random coincidences contributed 0.5–0.6% of usable pairs at 10 MBq in both geometries.
Zhou et al (2026) assumed a usable single-photon fraction of 0.22 (range 0.15–0.30) within the window. The simulation gives 0.066, which reduces the information per decay by a factor of 11 with windowed analysis and of 5.4 with all angles and ML. Its per-event information (effective analysing power 0.41) was consistent with the Klein–Nishina average within the window (0.415).
The realistic model changed these numbers in two opposite directions (table 5). With positions from sub-pixel centroids and the kinematic ordering rule, 19% of interacting photons gave usable two-cluster events, and the order was correct for 70% of them. Misordering is intrinsic near 90°, where both deposits are close to 256 keV, and it reverses the scattering-plane axis only partly; together with position errors it reduced the mean to 0.84, and 14% of events had axis errors above 20°. After calibration in , the information per usable pair was 0.013, against 0.023 in the idealised model, so that two-cluster events alone gave 0.68 times the idealised information per pair of interacting photons. Doppler broadening of 1–2 a.u. had no measurable effect (relative change below 3%), because the angular resolution was dominated by the lever arm between clusters. Ordering by the larger deposit lowered the information by a further 19%. Including events with three or four clusters raised the usable fraction to 27%; their first pair was identified correctly in 87% of cases, and the information per pair of interacting photons rose to 1.26 times the idealised two-cluster value.
Table 6 lists the sensitivity. Crystal thickness had the largest effect (0.75 and 1.21 for 10 and 20 mm), followed by pixel size (1.12 and 0.87 for 1 and 3 mm) and threshold (1.03 and 0.90 for 10 and 30 keV); a ±10% change in photoabsorption changed the information by −4% and +6%, and 2% energy resolution by +2%. The model’s total attenuation coefficient was 6% below the tabulated value at 511 keV and 5–13% below it between 100 and 600 keV, mainly because coherent scattering was omitted; raising photoabsorption by 25%, more than this deficit, increased the information per pair of interacting photons by 10%, so the estimates are conservative by about this amount.
The independent GATE simulation supported both models. For annihilation pairs in which both photons first scattered by Compton scattering, the fitted coefficient was with entanglement enabled and with it disabled, close to the values 1 and of the maximally entangled and the separable orthogonally polarised states; the common excess of about 4% indicates a slightly larger azimuthal modulation in the Livermore polarised Compton model than in the free-electron Klein–Nishina analysing power used here. For photons incident on a module, GATE gave the same usable fraction as our model (0.266 versus 0.265 with sequence reconstruction; 0.181 versus 0.188 for two clusters), but lower correct-ordering rates (0.65 versus 0.70 for two clusters; 0.74 versus 0.87 for the first pair of multi-cluster events) and a lower information per usable pair (0.0100 versus 0.0126). We attribute the difference to electron ranges, fluorescence and bremsstrahlung, which displace cluster centroids and add spurious clusters. Relative to the idealised model, the information per pair of interacting photons in GATE was 1.03 with sequence reconstruction and 0.58 with two clusters only; we used these more conservative factors for the ring results.
In-body pre-scattering
Without an energy window, 3–6% of accepted pairs contained a photon scattered in the water cylinder (Table 7). The CZT photopeak window reduced this to 0.23–0.40%. The worst-case bias was then at most 0.0037, and its variation between source positions at most 0.0004, below 10% of the smallest tissue contrast considered. Without an energy window, the worst-case bias reached 0.052 and varied by up to 0.013 between source positions in the 1.5-cm cylinder, which exceeds the muscle–liver contrast of 0.008.
| Radius / offset (cm) | CZT photopeak window () | No energy window | ||||
| difference | difference | |||||
| 1.0 / 0 | 0.0023 | 0.0022 | – | 0.027 | 0.025 | – |
| 1.0 / 0.5 | 0.0025 | 0.0023 | 0.0001 | 0.030 | 0.028 | 0.003 |
| 1.5 / 0 | 0.0036 | 0.0034 | – | 0.042 | 0.039 | – |
| 1.5 / 0.5 | 0.0040 | 0.0037 | 0.0004 | 0.049 | 0.046 | 0.007 |
| 1.5 / 1.0 | 0.0038 | 0.0035 | 0.0002 | 0.056 | 0.052 | 0.013 |
Table 7. In-body pre-scattering in a water cylinder. : fraction of accepted pairs with at least one pre-scattered photon. Differences are relative to an on-axis source in the same cylinder.
Feasibility in a small-animal setting
Figure 7 and Table 8 combine these results, using the GATE-corrected detector response with sequence reconstruction. With G8 the Fisher information per decay was . Then required decays per region, about half of the decays available in a 10-mm tumour; required , 1.8 times the available decays. Separating muscle from adipose at under needed decays per region (6.0 times the tumour’s decays, 2.0 times the liver’s), muscle from liver (35 and 12 times), and myxoma from adipose (74 and 25 times). With two-cluster events only, all requirements increased by a factor of 1.8. The gapless ring reduced all requirements by a factor of 2.2; under they increased by a factor of 4.
| Target | Geometry | Decays per region | Tumour ratio | Liver ratio |
| G8 | () | 0.44 | 0.15 | |
| G16 | () | 0.20 | 0.07 | |
| G8 | () | 1.8 | 0.59 | |
| G16 | () | 0.80 | 0.27 | |
| Muscle vs adipose () | G8 | () | 6.0 | 2.0 |
| G16 | () | 2.7 | 0.90 | |
| Muscle vs liver () | G8 | () | 35 | 12 |
| G16 | () | 16 | 5.2 | |
| Myxoma vs adipose () | G8 | () | 74 | 25 |
| G16 | () | 33 | 11 | |
| Muscle vs liver, H () | G8 | () | 139 | 46 |
Table 8. Decays required per region with (GATE-corrected detector response with sequence reconstruction; in parentheses, two-cluster events only), and their ratio to all decays in a 10-mm tumour () and a mouse liver (). Tissue separations are under H0 unless stated.

Figure 7. Standard deviation of versus decays per region for the eight-module and gapless 16-module rings (GATE-corrected detector response with sequence reconstruction; dashed, two clusters only). Dotted lines: precision per region required for a separation of tissue pairs (H0). Dashed lines: all decays of in a 10-mm tumour and a mouse liver at 400 kBq mL.
Discussion
What is measured. Double-Compton polarimetry of annihilation photons measures one number, the linear-polarisation correlation . Its relation to entanglement requires assumptions about the unmeasured circular correlation: for Bell-diagonal states, implies entanglement and bounds the concurrence below by , but the same is compatible with concurrences up to (Appendix § 12). This is consistent with recent analyses showing that separable states can reproduce the double-Compton cross section and that single-scattering observables admit local hidden-variable models ((Clarke et al 2026, Hiesmayr and Moskal 2019, Sutter and Hiesmayr 2026)). For imaging, the distinction matters less than it might seem: what carries tissue information is the mixture of channels in (5), and is a well-defined, model-independent observable of that mixture. We therefore suggest reporting (or together with the analysing power) and reserving the term entanglement degree for statements supported by the bounds above.
Statistical requirements. The information per double-Compton event is small for reasons that cannot be engineered away: Compton scattering does not resolve circular polarisation, the analysing power at 511 keV does not exceed 0.69, and most scatterings are forward. In the small-modulation limit, (3) gives , the two-photon analogue of the per-event Stokes uncertainty of X-ray polarimetry (Kislat et al 2015) with replaced by the mean squared product of analysing powers. Two practical consequences follow. First, scattering-angle windows should not be imposed for estimation; every event should enter the likelihood with its own analysing power. Second, the ratio, although convenient for display, should not be the estimator: it needed up to 2.3 times as many events as ML in our simulations, because a single mean analysing power is applied to events with widely different . The ML and -weighted estimators are simple to implement and reached the Cramér–Rao bound.
Detector design. The usable fraction in CZT was limited not by energy or position resolution, which were more than adequate, but by containment of the laterally scattered photon. Tighter packing helped: a gapless ring more than doubled the information per decay, mostly through geometric coverage. Once ordering and position errors are included, two further properties matter: sequence reconstruction of events with three or more clusters, which nearly doubled the information per pair of interacting photons, and the use of the known incident direction for ordering, which gained 19% over ordering by deposit size. Doppler broadening and energy resolution between 1% and 2% were unimportant. Thicker crystals (20 mm: +21%), finer pixels (1 mm: +12%) and layered geometries in which the scattered photon is absorbed in a second detector, as in Compton cameras, would add further gains. Table 6 can serve as a list of design requirements. Rates of random coincidences were negligible at 10 MBq, and polarisation-correlated Compton events have been measured to have a higher signal-to-random ratio than photoelectric events at clinical activities (Kožuljević et al 2026).
Table 6: Sensitivity of the information per pair of interacting photons to detector parameters (realistic model with sequence reconstruction; reference: 15 mm, 2-mm pixels, 15 keV, 1% at 662 keV, Doppler 1.5 a.u.).
| Parameter | Value | Relative information |
| Crystal thickness | 10 / 20 mm | 0.75 / 1.21 |
| Pixel size | 1 / 3 mm | 1.12 / 0.87 |
| Energy threshold | 10 / 30 keV | 1.03 / 0.90 |
| Photoabsorption cross section | −10% / +10% | 0.96 / 1.06 |
| Energy resolution at 662 keV | 2% | 1.02 |
| Doppler width | 1 / 2 a.u. | 1.00 / 0.97 |
| Cluster ordering | larger deposit instead of kinematic | 0.81 |
| Sequence reconstruction | two clusters only | 0.55 |
Table 6.
Energy resolution and the decoherence question. Whether a photon keeps its polarisation correlation after scattering in the body is disputed. One view holds that scattering below about 40° does not reduce the correlation [23] (Ivashkin et al 2023); another measurement reported a reduction (Watts et al 2021). With the energy resolution of CZT, pre-scattered photons were reduced to below 0.4% of accepted pairs, and the worst-case bias varied by less than across positions, so tissue contrasts can be measured whichever view is correct. Without energy discrimination, the worst case would vary by more than the muscle–liver contrast across a mouse-sized object, and more in humans. Energy resolution therefore makes imaging robust to an open question in the physics of the measurement.
The pick-off question. (5) makes explicit that carries tissue information only through the pick-off channel, and only to the extent that . The interpretation of the JPET result relies on separable pick-off photons (Moskal et al 2025), an assumption that remains to be verified. Our inversion of their data gave between 0.34 and 0.58, which disfavours maximally entangled pick-off photons; if confirmed, tissue contrasts would be 0.8–1.3 times those computed under . The inference, however, rests on the published correction from measured to ideal ratios, and the two angles disagree. A first-principles calculation of for annihilation with bound molecular electrons is therefore the key theoretical input for the whole programme. Our results show what is at stake: the floors on oxygen estimation are unchanged for any , but the required statistics scale as , and the biomarker vanishes at . Measurements in materials with known pick-off fractions and different electronic structure, analysed with the estimators recommended here, could discriminate between these hypotheses.
Feasibility. With and an eight-module CZT ring, can be measured to about in a 10-mm tumour, which is sufficient to distinguish separable from maximally entangled pick-off photons in phantoms with large pick-off fractions. It is not sufficient for tissue classification by alone, which needs 6–74 times the available decays. Classification in such systems will rely on lifetime and three-photon observables, as in Zhou et al (2026), with contributing as a complementary observable in combined estimation. For that study the present results leave the identifiability floors unchanged but increase the number of decays required in the CZT scenario: by factors of 1.1–1.3 for myxoma versus adipose, where lifetime and three-photon events dominate, by 3–7 for muscle versus liver, and by 5–11 when is used alone.
Limitations. Transport in CZT was unpolarised and neglected coherent scattering, charge sharing and fluorescence escape; Doppler broadening was approximated by a Gaussian momentum profile. The effects of cluster ordering, sequence reconstruction and electron transport were computed for a single module and transferred to the rings as a common factor, and the effective analysing power was calibrated with the true scattering angles, which in practice requires a calibration measurement or a validated simulation. The comparison with published data used reported ratios rather than raw data. The mouse was a uniform water cylinder, and the source was a point at the centre of the ring for the detector simulation. The channel shares rely on ex vivo positron-lifetime data and on the source model of Zhou et al (2026), including its treatment of oxygen-converted p-Ps. The post-scattering coefficient was bounded rather than modelled. Detector parameters not provided by the manufacturer (timing, count-rate behaviour) were taken from Zhou et al (2026) and must be measured.
Conclusion
Compton polarimetry of annihilation photons extracts less than 1% of the quantum Fisher information about their polarisation correlation, and a double-Compton event carries about 0.015 units of information on . Using all scattering angles with event-wise maximum likelihood is optimal; angular windows and the ratio waste up to half the data. The model reproduces published modulation ratios of CZT and scintillator polarimeters and agrees with an independent Geant4 simulation. In a pixelated CZT ring, lateral containment limits the usable events, ordering errors, position errors and electron transport halve the information per event, and sequence reconstruction of multi-site events recovers this loss; energy resolution makes the measurement robust to in-body scattering. With in a mouse, can be measured to in centimetre-sized regions, but tissue classification by alone is out of reach. The biomarker value of depends entirely on the polarisation state of pick-off photons; published porous-polymer data suggest , and this should be the next target of theory and experiment.
Acknowledgments
This work was supported by the National Natural Science Foundation of China (Nos. 62005150, U22A6008 and 82027804), the Shanxi Province Higher Education “Billion Project” Science and Technology Guidance Project, the Fundamental Research Program of Shanxi Province (No. 202503021211260), the Natural Science Foundation for Young Scientists of Shanxi Province (No. 202303021222344) and the First Hospital of Shanxi Medical University Hospital Fund Project (No. SYYYRC-2023004).
Data availability statement
The code, simulation scripts and numerical results that support the findings of this study are openly available at Zenodo (DOI to be assigned on acceptance; an anonymised copy is provided to the reviewers).
Author contributions
Haitao Zhou: conceptualization, methodology, software, formal analysis, validation, visualization, funding acquisition, writing – original draft, writing – review and editing. Ruonan Wang: data curation (tissue parameter set), funding acquisition, writing – review and editing. Jianyong Hu: methodology (quantum optics and polarimetry), validation, writing – review and editing. Zhifang Wu: resources, writing – review and editing. Sijin Li: conceptualization, supervision, project administration, funding acquisition, writing – review and editing.
Declaration of generative AI use
During the preparation of this work the authors used Claude (Anthropic) to assist in writing and debugging the Python code for the Fisher-information calculations, the toy Monte Carlo and the detector simulations; to help organise and summarise the literature; and to draft and edit the English text. All assumptions, parameter choices, results and conclusions were designed, checked and verified by the authors, and all cited references were read and verified by the authors. The authors take full responsibility for the content of this article.
Ethics statement
This study is entirely theoretical and computational. No human participants, human samples, animals or patient data were involved; tissue parameters were taken from published measurements.
Conflict of interest
The authors declare no competing interests.
Identifiability: from to concurrence
Write the two-photon polarisation state in the Pauli basis as , with the linear Stokes components and the circular one. We assume no single-photon polarisation () and a correlation matrix that is diagonal in a frame in which the annihilation state has , i.e. a Bell-diagonal state. Rotational invariance about the line of flight requires , and the double-Compton distribution (2) depends on only through ; the circular correlation is not measured. The Bell weights are for the annihilation state, and . Positivity requires . The concurrence of a Bell-diagonal state is , which over the allowed range of gives
where is the fidelity with the annihilation Bell state (Figure 8). Thus certifies entanglement within this family, and fixes the maximally entangled state, in agreement with Joos and Kling (2026); for a separable state with the same exists. The family A and B states of § 2.2 correspond to and , respectively.
Figure 8. Concurrence compatible with a measured for Bell-diagonal states (Appendix A).
Fisher information
For on with and , and
using . For the quantum Fisher information, the states of families A and B are diagonal in the Bell basis with parameter-independent eigenvectors, so . Family A has weights , giving ; family B has , giving (4).
Tables
| Measurement | Selection | Model, | Model, | Measured |
| Watts et al (2021) , CZT | – | 1.91 | 1.37 | |
| – | 1.98 | 1.40 | ||
| Ivashkin et al (2023) , scintillators | –, counters | 2.46 | – | |
| Separable reference (expected 0.5) | Inferred | |||
| Moskal et al (2025), XAD-4 | 0.508 | |||
| 0.512 |
Table 1. Comparison with published measurements of the modulation ratio and inferred coefficients.
Table 5: Single DT-PA05 module: idealised and realistic detector models ( keV). Fractions are relative to photons that interact in the crystal. Information per pair of interacting photons is the product of the squared usable fraction and the information per usable pair.
| Model | Usable fraction | Correct order | Info per usable pair | Relative infob |
| Idealised (correct order, two clusters) | 0.173 | 1 | 0.023 | 1 |
| Realistic, no Doppler, kinematic order, two clusters | 0.189 | 0.70 | 0.013 | 0.68 |
| Realistic, Doppler 1.5 a.u., larger-deposit order, two clusters | 0.146 | 0.66 | 0.018 | 0.55 |
| Realistic, Doppler 1.5 a.u., kinematic order, two clusters | 0.190 | 0.70 | 0.013 | 0.69 |
| Realistic, Doppler 1.5 a.u., kinematic order, clusters | 0.266 | 0.012 | 1.26 | |
| GATE, two clusters | 0.181 | 0.65 | 0.012 | 0.58 |
| GATE, clusters | 0.266 | 0.010 | 1.03 |
Table 5. Two-cluster events / first pair of events with three or four clusters. Per pair of interacting photons, relative to the idealised model.
Figures
Figure 1. (a) Annihilation channels and their polarisation-correlation coefficients; the pick-off coefficient is unknown. (b) Double-Compton measurement in an eight-module CZT ring with a source in a water cylinder.
References
- [23]Moskal P et al 2026a First-in-human quantum entanglement imaging arXiv:2606.29421arxiv.org/abs/2606.29421
- [24]Moskal P et al 2026b Roadmap towards quantum entanglement positron emission tomography (QE-PET) arXiv:2607.12182arxiv.org/abs/2607.12182
- [25]Pryce M H L and Ward J C 1947 Angular correlation effects with annihilation radiation Nature 160 435 (https://doi.org/10.1038/160435a0)
- [30]Tao S J 1972 Positronium annihilation in molecular substances J. Chem. Phys. 56 5499–5510 (https://doi.org/10.1063/1.1677067)





