Adaptive optics corrects blur by measuring how a sample or medium distorts light and then cancelling that distortion with a programmable phase plate, called a spatial light modulator (SLM). Normally this needs a "guidestar", a bright point of light inside or behind the sample. Pairs of entangled photons offer a label-free substitute. Because the two photons of a pair leave the source at the same point, the spread of their joint arrival positions on a camera traces the blur of the optics. Feeding that spread back to the SLM corrects the image without placing anything in the sample. The method works, but it is very slow: each correlation frame takes about a minute, so a correction takes hours.
This note argues that part of the slowness comes from a symmetry problem rather than from photon starvation. The two photons of a pair travel with opposite transverse momenta. If both pass through the same aberrated pupil, the pair accumulates the phase at $k$ plus the phase at $-k$. That simple argument predicts two failures of the usual correlation signal. It cannot see odd aberrations such as coma or tilt. It cannot tell the sign of even aberrations such as defocus, astigmatism or spherical aberration. Sign ambiguity of this kind is a polarity-discrimination problem, the same one radio detectors solve by comparing a pulse's sign against an external reference. The proposal is to verify the predicted blindness, then remove it by making the pupil asymmetric, and to measure how many correlation frames are saved once the optimizer stops wandering between equivalent wrong answers.
Core Claim
Write $\phi(k)$ for the aberration phase across the pupil, where $k$ is the transverse momentum. In the configuration where both photons of a momentum-anticorrelated pair cross the same pupil, the two-photon amplitude carries the effective phase
$$\Phi(k) = \phi(k) + \phi(-k) = 2\,\phi_{\mathrm{even}}(k),$$
where $\phi_{\mathrm{even}}$ is the centrosymmetric part of $\phi$. For a centrosymmetric pupil, the sum-coordinate correlation $C^+$ is the photons' joint position projected onto their summed coordinate, the quantity that Adaptive Optical Imaging with Entangled Photons showed is proportional to the system's blur pattern (the point spread function). Under this effective phase, $C^+$ satisfies
$$C^+[\phi] = C^+[-\phi(-k)].$$
This is the conjugate-flip ambiguity of classical phase retrieval. Here it is combined with complete invisibility of the odd part of $\phi$.
We claim three things. First, this degeneracy exists in standard entangled-photon feedback geometries; this is a hypothesis to confirm in simulation and on the bench before building on it. Second, it explains a puzzling published result. Non-classical optimization of entangled photons through complex media found that masks optimized on coincidence counts leave classical laser light as uncorrected speckle, while giving a 5-fold boost to the two-photon peak. If the feedback signal constrains only $\phi_{\mathrm{even}}$, then the odd part of the optimized mask is unconstrained. Classical light, which does see the odd part, would then fail to focus. Third, putting an asymmetric aperture in a plane both photons cross breaks the degeneracy. Each aberration then produces a distinct correlation pattern, and the aberration can be inferred from a handful of correlation frames instead of found by long trial and error.
Research Program
The work proceeds in three stages.
Stage 1: confirm the symmetry analysis numerically. We will simulate the two-photon field (spontaneous parametric down-conversion with realistic pump width and crystal length) through Zernike-parameterized aberrations, meaning combinations of standard optical aberration shapes, in two geometries. In the shared-pupil geometry, both photons cross the aberration in the far field. In the split geometry, photons are separated or the aberration sits in an image plane. The simulation records which aberration modes leave $C^+$, the full joint distribution $\Gamma$, and the difference-coordinate projection unchanged under a sign flip. This stage also decides whether the issue lies only in $C^+$ or in the whole correlation function. If the full $\Gamma$ already breaks the ambiguity, the cheaper fix is to use more of $\Gamma$, and the aperture becomes an ablation rather than the method.
Stage 2: restore identifiability. The pupil-engineering idea comes from Guidestar-Free Adaptive Optics with Asymmetric Apertures. That work places a triangular aperture stop so that every phase aberration produces a unique blur pattern for incoherent scenes. It then trains a pair of networks on MIT Places2 images to invert blur to phase, converging in 2–4 iterations from 4 measurements. We will not reuse its networks. They rely on natural-image priors and intensity imaging, and our measurement is a correlation pattern that depends only on the source and the optics. Instead we will train a small regressor, or use a direct least-squares inversion, from asymmetric-pupil $C^+$ frames to Zernike coefficients, on simulated pairs.
Because the target is specimen-independent, the training distribution is set by physics, not by an image dataset. That is the main structural difference from the classical method. The asymmetric aperture costs light, and coincidence rates scale with the square of transmission, so the aperture shape and fill factor become explicit design variables. We will report the information gained per frame against the photon loss.
Stage 3: handle drift. Minute-long frames accumulate laser-power drift, mechanical drift and SLM phase drift; the open problem is noted in the entangled-imaging work above. Following the orthogonal-channel logic of the Daya Bay hydrogen-capture analysis, Independent measurement of the neutrino mixing angle θ13 via neutron capture on hydrogen at Daya Bay, which checked one detection channel against an independent one with different systematics, we treat classical light as a second channel. That paper reports nH and nGd agreement and inter-detector consistency within 0.6%.
The procedure has two parts. After the quantum correction is chosen, a co-aligned laser is sent through the final mask. The odd part of the aberration predicted from classical focusing must match the estimate from the asymmetric quantum channel. Separately, flat-mask reference frames are interleaved throughout the run so that frame-to-frame drift is fitted as a nuisance term rather than absorbed into the objective.
The connection to exponential state-space scaling is kept modest and is a measurement in its own right. We will measure how the number of frames to convergence scales with the number of corrected modes $M$. A modal inverse should scale roughly linearly in $M$, from counting unknowns. Pixel-wise random-partition search scales with SLM pixel count. We will also test whether correcting the odd modes recovers some of the entanglement dimensionality lost through scattering in Manipulation and certification of high-dimensional entanglement through a scattering medium, which reports 28 certified dimensions in a clear path and 17 through parafilm with classical-beacon correction.
Baselines
Uncorrected reference. The naive point of comparison is the uncorrected system through the same aberrator. For thick PDMS (a soft silicone layer used as an aberrator) and the honeybee-head specimen in Adaptive Optical Imaging with Entangled Photons, the reported uncorrected values are SNR 3 and SSIM 77.89%.
Modal hill-climbing on $C^+$ sharpness. The strongest directly comparable method is the one in that same paper: Zernike coefficients are adjusted one by one to maximize the sharpness of $C^+$. It reports SNR 3→15 and SSIM 77.89%→98.41%. It is the right competitor because it uses the same label-free signal and the same modal basis. Our method changes only the pupil and the inversion. We will run it on identical aberrators, the same camera, the same pair flux, and the same budget of correlation frames. Beating it means two things. We must reach its final residual wavefront error, or its SSIM, in substantially fewer frames, targeting several-fold fewer. We must also correct odd modes that it leaves untouched, verified by classical focusing through the final mask. The file contains no reported frame count to convergence for this method, so we will measure it ourselves.
Random-partition coincidence feedback. The pixel-level competitor is the random-partition algorithm of Non-classical optimization of entangled photons through complex media, which reports a 5-fold coincidence enhancement. It is the strongest option for scattering media that modal bases cannot describe. We will run it on the parafilm and diffuser samples with matched acquisition time. The comparison is coincidence enhancement against the number of frames.
Classical-beacon correction. The transmission-matrix method of the high-dimensional certification paper serves as a reference for what classical correction achieves: 17 certified dimensions after parafilm.
Symmetric-aperture ablation. This ablation isolates the key ingredient. It uses our identical inverse model and probing schedule but a circular aperture of equal light throughput. If the asymmetric pupil is what matters, the circular version should show sign errors on even modes and no recovery of odd modes.
Why not just…?
Why not just add known bias aberrations of both signs, the standard sensorless adaptive-optics trick? Applying $+b$ and $-b$ probes per mode also resolves sign, and it is the most serious competitor. It cannot recover odd modes in the shared-pupil geometry, because the pair still sees only $\phi_{\mathrm{even}}$ however the probes are chosen. It also costs at least two extra frames per mode, which at a minute per frame is the dominant expense. We will run bias probing as an additional baseline at matched frame budget. If it matches our residual error on even modes with comparable frames, we will report that honestly. The remaining claimed advantage would then rest on odd-mode recovery alone.
Why not just measure the classical transmission matrix with a laser and display its inverse? That is fast, but the orthogonality result in Non-classical optimization of entangled photons through complex media shows that classical and quantum optima differ. Classical-beacon correction in the certification paper still lost 11 of 28 dimensions. A co-propagating beacon also requires access to the beam path. Our classical-consistency check doubles as a test of where the two kinds of correction agree and where they diverge.
Why not just use the full correlation function $\Gamma$ instead of changing hardware? This may well suffice if the aberration plane breaks the momentum symmetry. Stage 1 exists to answer exactly this question. If simulation shows that $\Gamma$ or the difference coordinate identifies both sign and odd modes in realistic geometries, the project pivots to software-only inversion and drops the aperture.
Why not just run the classical asymmetric-aperture networks on ordinary camera images? They depend on natural-scene priors and specimen texture. Adaptive Optical Imaging with Entangled Photons shows that sample-dependent sharpness metrics fail on multi-depth specimens, which is the regime where a specimen-independent guidestar matters.
Proposed Experiments
Simulation benchmark. Using an ensemble of random Zernike aberrations up to fourth order, we will report the per-mode sign-error rate and the odd-mode residual for each geometry and baseline. The decisive check is whether the circular-pupil $C^+$ inverse fails at chance level on even-mode signs and leaves odd modes uncorrected.
Bench validation. The aberrators will be SLM-imposed known aberrations, then PDMS layers, then parafilm. Metrics are residual wavefront RMS against ground truth on SLM-imposed cases, frames to threshold, coincidence-peak enhancement, and SSIM on a biological specimen.
Orthogonality test. This tests the explanatory claim directly. After random-partition optimization, we fit the mask to Zernike modes. The prediction is that its odd part is uncorrelated with the true aberration while its even part is correlated. Classical focusing should then be restored when the true odd part is added.
Drift test. We will inject controlled pump-power and SLM phase drifts and report correction bias with and without interleaved reference frames.
Dimensionality. Entanglement dimensionality will be certified through parafilm after our correction and compared against the 17 dimensions reported with classical-beacon correction.
The project fails if Stage 1 shows that $C^+$ in realistic geometries already resolves sign and odd modes. It also fails if aperture photon losses erase the frame savings. Both outcomes would be measured, not assumed.
References
- Adaptive optical imaging with entangled photons
- Non-classical optimization of entangled photons through complex media
- Guidestar-Free Adaptive Optics with Asymmetric Apertures
- Independent measurement of the neutrino mixing angle<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>θ</mml:mi></mml:mrow><mml:mrow><mml:mn>13</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>via neutron capture on hydrogen at Daya Bay
- Manipulation and certification of high-dimensional entanglement through a scattering medium