Loop-Multiplexed Heralded Photons for Dose-Limited Reflectance Imaging

When a sample can only tolerate a small number of photons, as with living cells, light-sensitive materials, or dim scenes that should not be disturbed, the goal is to learn as much as possible about each photon that hits it. The quantity most worth measuring is the sample's reflectance or transmittance: the fraction of light it returns, independent of how it happens to be lit. Recovering this from an ordinary image is underdetermined. The image is reflectance multiplied by illumination, and the illumination is unknown. This proposal makes the problem well posed by probing the sample with heralded single photons. In a heralded source, each probe photon is created as one half of a pair, and detecting its twin tells you exactly when a probe photon was sent. Because the experimenter then knows the probe count, every pixel's reflectance becomes a counting problem with a known denominator. The unknown room light shows up separately, as counts that are not tied to any herald.

The difficulty is speed. Heralded sources fire at random, and turning up the pump to fire more often creates unwanted extra pairs. A temporal loop multiplexer addresses this. It is a fiber delay loop plus a fast switch that stores an early photon and releases it into a target time slot, so more slots contain a photon without pushing the pump harder. But each trip around the loop loses light, and that loss turns the known photon number back into a guess. The central question here is whether there is a regime, and an optimal amount of loop storage, in which a loop-multiplexed heralded probe recovers reflectance with lower error per photon on the sample than a time-gated attenuated laser. The research also asks whether a Retinex-style decomposition can turn that per-pixel advantage into clean reflectance and illumination maps.

Core Claim

For a pixel with reflectance $R$, suppose a heralded probe photon that spent $j$ round trips in the loop is detected after the sample with probability $p_j = \eta_h T^j \eta_d R$. Here $\eta_h$ is the source's heralding efficiency (the chance a heralded photon actually leaves the source), $T$ is the loop transmission per round trip, and $\eta_d$ is the collection and detection efficiency after the sample. Each herald is then a single yes-or-no trial. The Fisher information about $R$, which sets the best achievable estimation variance, per photon actually incident on the sample is

$$F_{\text{herald}}(j) = \frac{\eta_d}{R\,(1 - \eta_h T^j \eta_d R)}, \qquad F_{\text{laser}} = \frac{\eta_d}{R}.$$

The laser expression follows from Poisson photon statistics, the herald expression from binomial statistics. Their ratio, $1/(1-p_j)$, is the largest per-dose gain the quantum probe can deliver before background light is counted. Loop storage multiplies the rate of useful probe slots, by up to the 1.54(1) per-bin enhancement measured over four pulses in Resource-efficient fibre-integrated temporal multiplexing of heralded single photons. At the same time it divides $p_j$ by $T^j$, which erodes the per-photon gain. The claim is that information per unit time at fixed dose has an interior optimum in the storage depth. A herald-aware Retinex estimator that weights each event by its storage history $j$ should reach this optimum. Treating the multiplexed output as an ordinary light source should not.

This advantage is modest under realistic numbers. With the 30% heralding efficiency reported for a birefringent-fiber source in Spectrally pure single photons, a 1 dB loop, and good collection, $p$ sits near 0.1 to 0.2. The per-dose gain is then roughly 10 to 25%, not orders of magnitude. The proposal is therefore framed as a measurement of where the crossover lies, with an explicit criterion for abandoning the quantum probe.

How the Pieces Fit

Illumination-reflectance separation, also called Retinex decomposition, supplies the estimator and its priors. Retinex models an image as reflectance times a smooth illumination field. Deep Retinex Decomposition for Low-Light Enhancement learns this split from paired images using a smoothness loss that respects edges. R2RNet adds a denoising stage guided by the illumination map. Both must resolve the decomposition's inherent ambiguity using learned priors alone, and that ambiguity becomes total below about 0.1 photons per pixel, as Photon-Starved Scene Inference using Single Photon Cameras shows for classification. Our measurement removes most of the ambiguity physically. Herald-correlated counts measure $R$ times a known probe field. Uncorrelated counts outside the herald window measure the ambient illumination. The learned prior is needed only for spatial regularization at very low counts.

The single-photon camera literature shows that the detector model must be built in from the start. Photon-Flooded Single-Photon 3D Cameras and CASPI show that background light causes pile-up. Pile-up happens because, after each detection, the detector is blind for a "dead time," so early background photons hide later signal photons and distort the timing histogram. Our likelihood therefore uses the saturating count model rather than a linear one.

Temporal loop multiplexing is the delivery mechanism, and its loss per round trip is the constraint we optimize against (Mosley; Temporal Loop Multiplexing).

Reversible quantum state transfer, based on cavity adiabatic passage, enters in a narrower role than its name suggests. In this technique, called vacuum-stimulated Raman adiabatic passage (vSTIRAP), an atom in an optical cavity is steered by a slowly ramped laser so that it emits one photon into the cavity mode, with a pulse shape set by the laser ramp (Kuhn et al.). We do not use the time-reversed absorption that makes this technique attractive for quantum networks. We use it as the deterministic-source comparison arm, where $\eta_h$ approaches the source's intrinsic efficiency. That arm exposes a trade-off specific to imaging. The photons from Quantum Logic with Cavity Photons From Single Atoms last 200 ns and have $g^{(2)}(0)=0.15$, a measure of how often two photons come out at once. A 200 ns photon forces a detection gate at least that wide, and a wider gate admits proportionally more ambient light. Shortening the photon requires faster pulses, which breaks the adiabatic condition. This is the latency-fidelity tension that Accurate and Scalable Simulation of Cavity-Based Networks identified for network links, reappearing here as a trade between efficiency and background rejection. The insight from time reversal that we keep is that the detection gate should be a matched filter for the emitted wavepacket shape.

Research Program

The first component is an analytic and simulation model of the full measurement chain. It covers the source (heralding efficiency, multi-pair rate as a function of pump power), the loop (transmission per round trip, switching schedule, distribution of storage depth $j$), the sample (reflectance map), the ambient field, and a detector model with dead time. From this we derive the Fisher information rate at fixed sample dose as a function of storage depth, background flux, and gate width. The same calculation is run for three sources: a time-gated attenuated laser, an SFWM heralded source with and without the loop, and a vSTIRAP-like deterministic source using the parameters from Holleczek et al.. The output is a phase diagram showing where each source wins.

The second component is the estimator. Each pixel records two counts per acquisition: herald-correlated counts labeled by $j$, and uncorrelated counts outside the gate. Writing $N_j$ for heralds at depth $j$, $k_j$ for correlated detections, $B(x)$ for the ambient rate at pixel $x$, and $w$ for the gate width, the per-pixel negative log-likelihood is

$$\mathcal{L}(R,B) = -\sum_j \Big[k_j \log q_j + (N_j-k_j)\log(1-q_j)\Big] + \mathcal{L}_{\text{amb}}(B), \qquad q_j = 1-e^{-(\eta_h T^j \eta_d R + B w)}.$$

The saturating form of $q_j$ accounts for dead time, and $\mathcal{L}_{\text{amb}}$ is the Poisson likelihood of the out-of-gate counts. We minimize this jointly with a learned Retinex prior. Concretely, a small decomposition network in the style of Retinex-Net provides the reflectance and illumination priors as a regularizer, inside an unrolled optimization, so the physics term and the prior each have a defined role. The illumination output is $B(x)$, the actual ambient field, which makes the decomposition physically meaningful rather than a matter of convention.

The third component is a bench experiment. A fiber SFWM heralded source with a 200 ns switched loop, in the style of Hoggarth and Mosley, illuminates a USAF resolution target in transmission, with controllable stray light added. This is the geometry of Quantum microscopy of cells at the Heisenberg limit, which needed 17 minutes to reach a contrast-to-noise ratio of 3. The detector is a gated single-photon detector, raster-scanned or arrayed. The same optical path is also fed by an attenuated pulsed laser at matched dose.

Proposed Experiments

In simulation, we render photon-count measurements from the multi-illumination scenes underlying MIW+ (Color Preserving CMOS-SPAD Fusion). These scenes show the same object under many lightings, so a correct reflectance map should not change across them. That gives a test of illumination invariance that LOL-style paired data cannot provide. We also use SID ground-truth images as reflectance-times-illumination targets at sub-photon flux (Retinex-RAWMamba). The metrics are reflectance PSNR, variance of the reflectance estimate across lightings of the same scene, ambient-map error, and reflectance root-mean-square error as a function of sample dose. All are reported against the Fisher bound from the first component. On the bench, the metrics are contrast-to-noise ratio and resolvable USAF element per unit dose and per unit time, plus the measured per-pixel variance ratio between herald and laser probing.

No reported result was found on Lacuna for reflectance estimation from photon counts on MIW+, or for any heralded-versus-laser comparison at matched dose. These comparisons will therefore be established inside the proposal.

Baselines

The do-nothing reference is passive imaging of the same sample under ambient light only, followed by a classical smoothing Retinex split. It shows how much of the reflectance is recoverable without any probe and calibrates the inherent ambiguity at each flux level.

The strongest competitor is a time-gated attenuated pulsed laser, processed both by our likelihood with the Poisson probe term and by CASPI-style spatio-temporal filtering (CASPI). This is the strongest competitor because gating already separates the probe from ambient light, which is half of our claimed benefit. The laser shares the source's repetition rate, detector, gate, optical path, and sample dose. Both arms get the same prior network, trained with identical data and step budget. Beating it means a statistically significant reduction in reflectance error at equal dose that matches the predicted $1/(1-p_j)$ factor within error bars. A gain present only at equal time, and absent at equal dose, does not count. CASPI reports no number in the file. The closest reported laser-side figure is the reduction in depth error from 18.4 cm to 2.2 cm under 15,000 lux achieved by flux optimization in Photon-Flooded Single-Photon 3D Cameras. That figure is for depth, not reflectance, so we reproduce flux-optimized laser operation as part of this baseline rather than quoting it.

The quantum-imaging baseline is coincidence imaging with an unmultiplexed SPDC source. The reference is the result in He et al.: 1.4 µm resolution, and a contrast-to-noise ratio of 3 after 17 minutes. Beating it means reaching that contrast in clearly less time at equal dose.

The learned baseline is R2RNet and Retinex-Net retrained on our simulated count images without herald information. On LOL, R2RNet reports 20.207 dB PSNR, which shows what a learned prior alone achieves on a different input.

Two ablations isolate the key ingredient. In the first, the loop is on but the estimator ignores $j$ and treats all events as Poisson. In the second, the estimator is herald-aware but the loop is off. The claim stands only if the herald-aware estimator with the loop beats both ablations.

Why not just…?

Why not just use a pulsed laser with time gating? This is the main threat. Gating already isolates ambient light, and lasers are bright, cheap, and lossless to store. The analysis predicts that the laser wins whenever $p_j$ is small: at low collection efficiency, low reflectance, or deep loop storage. The quantum probe should win only in the high-efficiency, dose-limited corner. The equal-dose bench comparison settles the question directly. If the measured variance ratio stays below about 1.05 across the realistic parameter range, the honest conclusion is that this application does not need heralded photons. We would then publish the crossover map as a negative result rather than redesign around it.

Why not just build a brighter source or several sources in parallel instead of a loop? Pumping harder raises the multi-pair rate, which lowers single-photon purity (Barz et al.). That directly corrupts the known-photon-number assumption the advantage rests on. Parallel sources cost one detector and one pump path per source (Mosley). Our simulation includes the brighter-pump arm at equal herald rate so the purity penalty is measured rather than assumed.

Why not just train an end-to-end network on raw photon counts? Learned Retinex models do outperform direct mappings on LOL and SID (R2RNet; Retinex-RAWMamba). However, a network that never sees herald labels cannot exploit binomial statistics, and at sub-photon flux it will fill in reflectance from its prior. The learned baseline and the multi-illumination consistency test measure exactly this. Hallucinated reflectance shows up as estimates that change with lighting.

Why not just use a deterministic atom-cavity source and skip multiplexing? If such a source reaches high fiber-coupled efficiency, it dominates on $\eta_h$, and the loop becomes unnecessary. Its 200 ns wavepackets, however, widen the detection gate and admit more ambient light. The deterministic arm of the phase diagram quantifies at what background level this penalty outweighs the efficiency gain. We expect the answer to depend on the scene, and we report it either way.

Decisive Checks

The first check is whether the measured equal-dose variance ratio between the herald and laser probes tracks $1/(1-p_j)$ as storage depth $j$ is varied. That requires showing the predicted rise with efficiency and the predicted fall with loop loss. The second is whether the information rate at fixed dose peaks at an interior storage depth, rather than at zero or at maximum storage. The third is whether reflectance maps recovered on multi-illumination scenes stay consistent across lightings more tightly than all baselines, at matched dose. Failure of the first check ends the quantum-probe claim. Success there combined with failure of the third means the physics holds but the Retinex estimator is not extracting the advantage.