Counting Photons to Certainty in Sub-Wavelength Coherent Absorbers

Nanophotonic devices can squeeze light into volumes far smaller than its wavelength. They do this with metal films, sharp tips, and standing waves that pin an absorbing film at a point of maximum field. Experiments have shown that a film only 50 nm thick, placed at the bright spot of a standing wave, absorbs single photons almost deterministically. A free-running traveling wave, by contrast, cannot be absorbed by such a thin symmetric film more than half the time. What nobody has asked is how much information per photon these devices deliver about the thing they are meant to sense. The question matters because the imperfections of real devices act like a constant background of stray counts. These imperfections include metal losses of roughly half the light, interference that swings between 10% and 90% instead of 0% and 100%, and detector dead time. In photon-limited statistics, a constant background is exactly what kills sensitivity.

This note proposes a statistical figure of merit for sub-wavelength absorbers: the number of incident photons needed to detect a nanoscale perturbation at a fixed significance, for example 5σ. We would compute it from the Fisher information of realistic count models, which includes the interference floor, metal loss, dark counts and dead time. We would validate the models on released single-photon coherent-absorption data. Then we would use the metric to find where deep sub-wavelength plasmonic confinement actually beats a lossless dielectric interferometer, and where it does not. The metric replaces the speed, energy and footprint scaling rules used to rank nanophotonic devices today with a question about how many photons a measurement costs.

Core Claim

Two coherent-absorption experiments showed that deterministic single-photon absorption in deeply sub-wavelength films is real. Coherent perfect absorption in deeply subwavelength films in the single-photon regime and Metamaterial Coherent Plasmonic Absorption With a Single Photon use heralded down-conversion photons, a gold split-ring metamaterial or graphene film, and an interferometer. Heralded counts swing between 10% and 90%. Neither paper turns that swing into a sensitivity statement. These papers motivate the proposal: they supply the effect and expose the missing analysis.

The swing has a direct cost. Model each heralded photon as either reaching the output detector or not, with detection probability $p(\phi)=\tfrac12(1-v\cos\phi)$. Here $\phi$ is the standing-wave phase and $v$ is the fringe visibility, meaning the normalized depth of the swing. The Fisher information per heralded photon, which is the inverse of the best achievable phase variance per photon, is $F(\phi)=p'(\phi)^2/[p(1-p)]$. Its maximum over $\phi$ is

$$F_{\max}=1-\sqrt{1-v^2}.$$

This follows by optimizing the binary-outcome information over the operating phase. At $v=1$ the value is 1. At the reported 10% to 90% swing ($v=0.8$) it is 0.4. So the imperfect node alone costs 2.5 times more photons for the same precision, before any metal loss or dark counts are counted.

Our claim is that a small set of physically meaningful quantities fixes how many photons a sub-wavelength absorber needs to detect a perturbation localized in its mode. Those quantities are this floor-limited information, the collection efficiency $\eta$, the dark count rate, and how strongly the perturbation shifts the phase. The last of these grows as the mode volume shrinks. Confinement raises the signal per photon, while loss and residual floor lower the information per photon. The product has a computable optimum. The optimum should land at intermediate confinement, echoing the non-monotonic efficiency curve reported in Fundamental Scaling Laws in Nanophotonics. That work ranks devices by speed over energy times footprint; we rank them by photons-to-significance.

Research Program

Stage 1: count model. We would build a full likelihood for heralded coincidence counts in a coherent absorber. It would combine the interference term with visibility $v$, overall efficiency $\eta$, dark and accidental coincidences, and SPAD dead time. Dead time is modeled the way Photon-Flooded Single-Photon 3D Cameras models pile-up. That paper also derives a closed-form optimal attenuation, and we reuse its dead-time machinery as a method. Unlike it, we apply the model to a phase-modulated standing wave rather than a depth histogram.

We would fit the model to the released phase-stepping count traces in the supporting data for single-photon CPA. The fit asks two things. First, what fraction of the 10% floor comes from background counts, which could in principle be subtracted, and what fraction comes from true incomplete destructive interference, which cannot? The papers attribute the floor to quadrupole transitions at nodes and to phase jitter. Second, could detector nonlinearity mimic any part of the modulation, the concern raised in the file's gap 4.4? Given heralded rates this is unlikely, but the fit will test it.

Stage 2: significance budgets. From the fitted model we would compute two budgets. The first is the number of heralds needed to reject the classical 50% absorption cap at 5σ. The second is the Cramér-Rao phase bound per herald. We would extend both to the two-photon case using the reported 61% single-pass graphene absorption from Coherent Absorption of N00N States. That paper switches between one- and two-photon absorption but never computes phase information. We would ask whether the two-photon state keeps any per-photon advantage once the non-ideal film is included. We expect it may not.

Stage 3: perturbation-sensing figure of merit. For a small particle or refractive-index change inside the mode, the phase shift per perturbation scales roughly as $Q/V$, the quality factor over the mode volume. This is standard cavity perturbation theory. Combining it with Stage 1 gives photons-to-5σ as a function of $V$, $Q$, $\eta$, $v$ and dark rate.

We would evaluate this function for the platforms in the file using their reported parameters. These are:

We would then sweep new designs in 3D finite-difference time-domain simulations to map the crossover between confinement and loss.

The nanobeam's thermo-optic bistability at about 17 nW matters here. It caps the usable photon flux, which turns the photon count from a free resource into a hard budget.

Baselines

The naive reference is a traveling-wave thin absorber read by a SPAD, with no standing wave. Its absorption is capped at 50% and there is no interference. We would run it on the same simulated perturbations with identical detector models. The proposal's Stage 3 metric must show fewer photons-to-5σ than this reference for any localized perturbation.

The strongest existing sub-wavelength result is the coherent absorber exactly as reported. Metamaterial Coherent Plasmonic Absorption With a Single Photon reports heralded count modulation from 10% to 90% with near-100% plasmon coupling at the antinode. No reported Fisher information or phase-sensitivity number was found on Lacuna for this or any CPA experiment. The baseline number we compare against is therefore our own derived $F_{\max}=0.4$ per herald at the reported visibility. "Beating it" means one of two things. Either a design reaches a higher visibility-limited information per incident photon, or an operating-point choice reaches the same information with fewer photons after accounting for dark counts. We would test operating points with our model on the released traces.

The strongest overall competitor is a lossless dielectric interferometer or cavity read by a high-efficiency SPAD. Its information per photon is about $\eta$ at full visibility. Its limit is that it cannot confine light below roughly 100 nm, as Fundamental Scaling Laws in Nanophotonics notes. For phase sensing of a perturbation that fills its mode, this competitor should win, and we would say so. The comparison is fair only if both designs see the same perturbation, photon budget and detector. The plasmonic design counts as winning only in the size range where its $Q/V$ gain outweighs its $\eta\,F_{\max}$ penalty.

A dead-time baseline applies the optimal attenuation of Photon-Flooded Single-Photon 3D Cameras to a coherent absorber under added ambient background. That paper reports depth RMSE falling from 18.4 cm to 2.2 cm at 15,000 lux. We would reuse the attenuation rule to separate how much of the lost information is recoverable by flux control and how much is fixed by the interference floor.

The key ablation runs the full model with the floor set to zero, then with metal loss set to zero. This isolates which imperfection dominates the photon budget on each platform.

Why not just…?

Why not just use a lossless dielectric cavity and a good SPAD? This is the real threat, and for any perturbation larger than about 100 nm it probably wins. A 50% metal loss halves the counts, as the hybrid nanocavity paper notes, and a 0.8 visibility costs another factor of 2.5. The proposal does not argue around this. Stage 3 computes the crossover size explicitly. If no realistic plasmonic design crosses over at any perturbation size, that negative result is itself the finding. It would redirect sub-wavelength absorbers away from sensing claims.

Why not just integrate longer? In Poisson counting, the 2.5× penalty is just 2.5× more time, which is acceptable when photons are free. They often are not. The nanobeam turns bistable near 17 nW. Fragile nanoscale samples limit dose. Heralded sources are slow. In the rare-event physics surveyed in the file, exposure is capped by nature or cost. The photons-to-5σ metric is designed for exactly the regime where the photon budget is fixed, and in that regime a floor-limited information loss cannot be bought back.

Why not just fix it in software, with a better fit or a robust learned estimator? A better likelihood removes bias. It cannot restore information the measurement never captured, because post-processing cannot increase Fisher information. Single-Photon Image Classification makes the same point for classification: accuracy on a single detected photon is capped at 22.96% on MNIST without optical interference before detection, and reaches 41.27% with it. Robust training such as SharpDRO helps models tolerate shot noise but cannot lift the Cramér-Rao bound. We would still fit our estimators to that bound, so software gets full credit.

Proposed Experiments

The first decisive check is on released data. Fitting the Stage 1 model to the CPA phase-stepping traces must reproduce the 10% to 90% swing. It must also either confirm or rule out the claim that the floor is mostly true interference leakage rather than subtractable background. We would verify the fit by showing that a maximum-likelihood phase estimator on held-out traces reaches the predicted Cramér-Rao variance within its confidence interval.

The second check is a significance budget. We would report, from the same data, the number of heralds needed to reject the 50% classical cap at 5σ. We would then compare it with the number actually collected.

The third check is the crossover map. For simulated nanoparticle perturbations from 5 nm to 200 nm, we would compute photons-to-5σ for the hybrid nanocavity, the bowtie nanobeam, and a lossless dielectric reference, all at matched $\eta$ and dark rate. The proposal succeeds if it predicts a specific perturbation size below which the plasmonic device needs fewer photons. It succeeds equally, in a negative sense, if it shows that no such size exists at the reported loss and visibility.

The final check is the two-photon case. Using the 61% graphene film, we would test whether the two-photon state retains per-photon phase information above the single-photon $F_{\max}$. If it does not, that gives a quantitative reason why scaling beyond two photons in lossy thin films is premature.