Decoy-State Loss Tomography for Free-Space Quantum Key Distribution Networks

Quantum key distribution (QKD) lets two parties grow a shared secret key whose security rests on physics rather than computational hardness. Practical systems send faint laser pulses, and a faint laser pulse sometimes carries two or more photons. An eavesdropper can quietly peel one photon off such a pulse. The standard defense is the decoy-state method: the sender randomly switches the pulse brightness among a few settings (a bright "signal", a weak "decoy", and an empty "vacuum"). Comparing how often each brightness arrives reveals how the channel treats true single photons. Every link runs this analysis on its own and summarizes the air between the telescopes as one number, the transmittance η. This proposal treats a network of many free-space links over one region differently. Each link is a ray through a shared three-dimensional volume of air, and its decoy statistics are a noisy measurement of how much light that ray loses. That is the same problem neural radiance fields solve when they rebuild a 3D scene from many camera rays.

We propose to fit a compact low-rank model of the region's loss field to decoy-state detection counts from all links jointly. The model is a vector-matrix (VM) decomposition, which writes a 3D grid as a sum of products of 1D vectors and 2D matrices. The fitted field is then used for two things that the per-link scalar view cannot do. The first is to predict loss on links that have little or no history, so the sender can choose better decoy brightness settings in advance. The second is to tell a weather front, which dims many neighbouring rays in a spatially consistent way, apart from a single link whose single-photon yield falls below what the surrounding physics predicts. The security guarantee is left entirely to standard decoy bounds; the field model only steers operations and diagnoses failures.

Core Claim

Consider a free-space QKD mesh in which rays overlap in a shared volume, such as ground-to-drone links, rooftop links in a city, or several ground stations tracking one satellite pass. In such a mesh, a VM-factorized loss field fitted to per-intensity decoy counts predicts held-out link transmittance better than per-link time-series models, and it raises finite-key secret-key rate through adaptive intensity selection. It also separates attacks from weather with higher accuracy than per-link consistency tests. The claim is falsifiable in a specific place. If rays share little volume, or if loss changes slowly enough that each link's own history suffices, the spatial model should lose to per-link forecasting. We will measure where that crossover sits.

One constraint shapes the whole design. Any physical prior that narrows the decoy estimate of single-photon yield would be insecure, because the eavesdropper controls the channel and can make it look however she likes. The field model therefore never enters the security proof. Intensity settings for block $k$ are chosen only from publicly announced data of blocks before $k$. Each block's key is then certified with an existing finite-key decoy analysis, following the composable asymmetric-channel treatment of Cao et al.. Because the choice of settings depends only on past public information, the per-block proof is unchanged.

Why these two directions meet here

The decoy literature motivates the gap. The 1,200 km satellite demonstration of Liao et al. handles a high-loss, fluctuating atmospheric path with multi-intensity decoys, but it reduces the atmosphere to an aggregate transmittance. Conference-key work by Cao et al. and Cao et al. shows that key rate is highly sensitive to channel asymmetry. The asymmetric-channel paper reports gains of at least an order of magnitude at a 100 km length difference once asymmetry is handled properly. Yet both papers treat each channel's parameters as independent scalars. Yang et al. study compromised nodes on abstract graphs and find that end-to-end paths vanish at about 60% compromised nodes. Their graphs carry no physical medium, so a lossy link and a hijacked link look alike.

Turbulence physics constrains the simulator we build. Chan shows that atmospheric distortion is tilt-then-blur and does not commute. This matters for QKD receivers because beam wander (tilt) and higher-order blur affect single-mode coupling differently. Our simulator applies them in that order. We also take a warning from the same paper: simpler approximations can be acceptable in smooth regimes, which is exactly the scalar-η baseline we must beat.

The vector-matrix literature supplies the method. TensoRF by Chen et al. represents a volume as sums of vector–matrix outer products along the three axis-plane pairs. This costs about $O(N^2)$ parameters per component instead of $O(N^3)$, and a scene trains in under ten minutes. The same paper shows that the purely 1D CP (canonical polyadic) variant loses fine structure. Guo et al. show that VM fields can absorb sparse, noisy auxiliary measurements (depth) and still suppress artifacts. That is the closest analogue to our setting, where each link gives a sparse, shot-noise-limited line integral. For drifting weather, the bandlimited time extension of Ramasinghe et al. shows one route from static VM fields to space-time fields.

Research Program

The observation model links the two fields. For link $\ell$ with ray $r_\ell$ and time block $k$, the physical transmittance is

$$\eta_{\ell,k} = c_{\ell,k}\,\exp\!\Big(-\int_{r_\ell} \sigma_k(\mathbf{x})\,d\mathbf{x}\Big).$$

Here $\sigma_k(\mathbf{x})$ is the extinction field at position $\mathbf{x}$, VM-factorized as in TensoRF. The factor $c_{\ell,k}$ is a receiver coupling efficiency that depends on beam wander and blur along the ray, computed with the tilt-then-blur model. For a phase-randomized pulse of mean photon number $\mu$, the detection probability is

$$Q_{\mu,\ell,k} = 1 - (1 - Y_0)\,e^{-\eta_{\ell,k}\,\epsilon\,\mu},$$

where $Y_0$ is the dark-count yield and $\epsilon$ the detector efficiency. The vacuum decoy pins down $Y_0$ separately from transmission. Without it, dark counts would pass for residual transmittance at high loss.

We fit $\sigma_k$ and $c$ by maximizing the binomial likelihood of the observed counts across all links and intensities. A smoothness penalty on the VM factors acts as regularization, and time is handled either by a slowly varying low-rank temporal factor or by per-block warm starts.

Two uses follow. For adaptive intensities, the predicted $\eta_{\ell,k+1}$ is plugged into the standard decoy key-rate optimizer to choose signal and decoy intensities and their probabilities for the next block. For attribution, we compare each link's decoy-certified single-photon yield lower bound against the value the field predicts from the other links. A photon-number-splitting attacker who blocks single photons pushes that bound below the physics-consistent value on one ray. A weather front lowers predictions on many neighbouring rays at once.

Baselines

The naive reference fixes the decoy intensities once from each link's long-run mean transmittance and runs the standard three-intensity analysis with no forecasting. It uses the same simulated counts, detector efficiency of 56%, dark count rate of $10^{-8}$, and error-correction efficiency of 1.1, which are the simulation parameters of Cao et al.. Any method that cannot beat this in finite-key rate is not worth deploying.

The strongest competitor is per-link temporal forecasting. Each link keeps its own autoregressive forecast of η from its past decoy counts, which is the natural extension of the scalar treatment in Liao et al.. This competitor needs no geometry, and wherever links rarely share air it should do as well as ours. It gets the same count history, the same hyperparameter search budget, and the same intensity optimizer. The spatial model must beat it on held-out and newly activated links and on attribution AUROC (area under the curve separating attack events from weather events). Matching it only on long-history links would not count as a win.

For attribution, the per-link baseline is a standard consistency test. It flags a block when the link's certified yield drops by more than a threshold relative to its own recent past, with the threshold tuned on the same validation weather.

Two ablations isolate the key ingredient. One replaces VM with CP factorization at matched parameter count, and the other replaces it with a coarse dense voxel grid. Together they test whether the vector–matrix structure matters or whether any shared spatial model suffices, given that Chen et al. report CP losing detail on radiance fields. A third ablation drops the vacuum and weak decoys from the fitting data and uses signal counts only, to measure what decoy statistics add to tomography. A fourth replaces tilt-then-blur coupling with commutative blur-then-tilt, following Chan.

No reported result was found on Lacuna for transmittance tomography or attack attribution in free-space QKD meshes, so there is no published number to beat. The quantities we can anchor to are the reported TensoRF-VM PSNR of 33.14 on Synthetic NeRF Chen et al., which we use only to validate our VM implementation, and the distance losses under device imperfection reported by Ma et al., which set the imperfection regime we simulate.

Why not just…?

Why not just keep a scalar η per link and forecast it over time? This is the real threat. If loss drifts slowly and every link has a long history, per-link forecasting captures most of the gain in intensity optimization, and the spatial model adds cost without benefit. The spatial model has an inherent advantage in only two situations. One is cold start, when a link is newly pointed, re-routed, or a satellite pass begins. The other is when many links change together, so that information about one tells you about its neighbours. Our experiments vary ray overlap and drift speed and report the regime where per-link forecasting wins. If that regime covers realistic mesh geometries, we will say so and narrow the claim to cold-start prediction.

Why not just use CP decomposition, which is far smaller? On radiance fields, Chen et al. find that CP misses fine structure that VM keeps. Aerosol and cloud fields, however, may be smooth enough that CP suffices. The matched-parameter ablation settles this. If CP matches VM on held-out transmittance, we adopt CP, because its linear memory makes on-site fitting cheaper.

Why not just rely on per-link decoy consistency tests for detecting eavesdroppers? The decoy bounds already keep the key secure against photon-number splitting, and our method does not improve security. What a per-link test cannot do is say why a link's yield fell. A drop in one link's history looks the same whether a storm cell arrived or someone is removing single photons. Cross-link consistency is the only information that separates the two. The attribution experiment measures how much it adds. If the per-link test reaches the same AUROC, the attribution use case fails.

Why not a coordinate MLP instead of a factorized grid? Blocks arrive every few seconds to minutes, and the field must be refitted within a block. Chen et al. report that VM fields train about 100 times faster than MLP radiance fields. We include an MLP field only as a timing reference and do not expect it to be usable in the loop.

Proposed Experiments

The testbed is a simulator we build. It combines a region of roughly one kilometre per side with an extinction field made of drifting aerosol and cloud structures, layered Kolmogorov phase screens applied tilt-then-blur for receiver coupling, and 20 to 100 transmitter–receiver links whose geometry ranges from nearly disjoint rays to heavily crossing meshes. Detection events are drawn per pulse under three-intensity decoy settings with finite block sizes. Device imperfections follow the state-preparation and after-pulse regime of Ma et al., and intensity fluctuations follow the concern raised by Cao et al..

Before any QKD experiment, we check our VM code on the Synthetic NeRF benchmark against the reported 33.14 PSNR, so that later failures cannot be blamed on the implementation.

The first experiment is novel-link prediction, the analogue of novel-view synthesis. We hold out 20% of links, fit the field on the rest, and report the relative error of predicted η on held-out links as a function of ray overlap, block size, and drift speed.

The second experiment measures finite-key secret-key rate summed over the network under adaptive intensities. We compare fixed settings, per-link forecasting, and the VM field, and certify every block with the same finite-key analysis.

The third experiment is attribution. We inject photon-number-splitting attacks with loss compensation on single links, correlated weather events, and node hijacks placed on the same topologies Yang et al. study. We report AUROC for separating attacks from weather and the delay until detection.

The decisive checks are simple to state. The proposal succeeds if the VM field beats per-link forecasting on held-out links and on attribution AUROC in overlapping meshes, and if that advantage shrinks toward zero as ray overlap shrinks, which would show the gain comes from shared volume rather than extra capacity. Removing the vacuum and weak decoys from the fitting data should measurably hurt high-loss links. If none of these hold, the result is still worth publishing: scalar per-link channel models would then be adequate for free-space decoy networks, and the question can be closed.