Photon loss is the dominant noise source in linear optical quantum circuits. Each lost photon turns an ideal n-photon outcome into a pattern with fewer photons, so the detector counts no longer follow the ideal output distribution, and the lossless events that do follow it become exponentially rare as the photon number grows. Discarding the lossy events (postselection) therefore wastes most of the data at realistic loss rates, and the ideal distribution has to be inferred from the lossy counts themselves. Recycling-based mitigation (Mills & Mezher, Quantum 2026; Quandela's photon_recycling) reweights lossy samples to recover lossless probabilities, but its estimators are biased and restricted to the no-collision sector, where each output mode holds at most one photon. We instead treat the raw counts as Multinomial observations under per-photon binomial loss and infer the ideal output distribution and the loss rate jointly from a Bayesian posterior, which extends naturally to outputs with collisions. On data from the independent Perceval simulator at the scale of prior work (n=4 photons in m=20 and m=10 modes), the posterior has lower KL divergence than both recycling and postselection at every tested loss rate in the no-collision benchmark. In the collision regime it recovers the loss rate and the collision-pattern probabilities, which recycling cannot estimate, and stays below recycling at every loss rate and below postselection except at low loss, where the two are comparable. Its advantage over recycling grows with the sample count, and it supplies credible intervals that the baselines lack.
AI Summary
Photon-loss mitigation seeks to reconstruct lossless linear-optical output probabilities without discarding all shots with missing photons. This paper replaces recycling-based reconstruction with Bayesian inference over an unstructured output distribution $p$ and uniform per-photon loss rate $\eta$. A multinomial likelihood uses independent binomial thinning in each mode, including factors $\prod_i\binom{s_i}{y_i}$ for collision patterns. Dirichlet and Beta priors are combined with sparse-kernel NUTS sampling to obtain posterior means and credible intervals.
Tests use Perceval simulations of four photons, 30 Haar-random interferometers per condition, and typically $N=10^5$ shots. For $m=20$, scored on the no-collision subspace, posterior KL divergence is lower than postselection and the shipped exponential-extrapolation recycling implementation in all 15 tested loss–indistinguishability cells. At $\eta=0.3$ and full indistinguishability, KL values are $0.105$, $0.195$, and $0.315$, respectively. For collision-aware inference at $m=10$, collisions carry about 71% of ideal probability mass; full-distribution KL is $0.012$, $0.125$, and $0.285$ at losses $0.3$, $0.7$, and $0.9$. Sample-budget sweeps show continued posterior improvement while recycling errors plateau.
The evidence is simulation-only and limited to four photons. Exact sampling becomes difficult around $10^4$ probabilities and fails at sufficiently large sample budgets. The no-collision benchmark omits substantial collision mass and biases inferred loss. Partial distinguishability shifts the inferred target rather than being corrected; credible-interval coverage against the ideal distribution can fall to 67%. High-loss postselection KL depends strongly on probability flooring. Uniform loss, photon-number-resolving detection, and unmodelled multiphoton emission further limit applicability; asymptotic unbiasedness is not established.
Likely largely AI-generatedAI use disclosedIncrementalSmall numerics onlyMissing comparisonStrong assumptions
for literature and citation checking, drafting and revising text, reformatting, writing and running the analysis code under the authors' supervision, and simulated pre-submission review.
Error found: Table 1's postselection KL of 0.680 at 90% loss is inconsistent with the specified empirical-frequency estimator and floor-and-renormalize scoring rule, which predict a much larger KL for the approximately five retained outcomes.
The collision-aware m=20 appendix reports mean loss estimates whose deviations from the true rates are incompatible with the closed-form Beta posterior and sampling fluctuations under its stated matched-model experiment.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
Photon loss is the dominant noise source in linear optical quantum circuits. Each lost photon turns an ideal n-photon outcome into a pattern with fewer photons, so the detector counts no longer follow the ideal output distribution, and the lossless events that do follow it become exponentially rare as the photon number grows. Discarding the lossy events (postselection) therefore wastes most of the data at realistic loss rates, and the ideal distribution has to be inferred from the lossy counts themselves. Recycling-based mitigation (Mills & Mezher, Quantum 2026; Quandela's photon_recycling) reweights lossy samples to recover lossless probabilities, but its estimators are biased and restricted to the no-collision sector, where each output mode holds at most one photon. We instead treat the raw counts as Multinomial observations under per-photon binomial loss and infer the ideal output distribution and the loss rate jointly from a Bayesian posterior, which extends naturally to outputs with collisions. On data from the independent Perceval simulator at the scale of prior work (n=4 photons in m=20 and m=10 modes), the posterior has lower KL divergence than both recycling and postselection at every tested loss rate in the no-collision benchmark. In the collision regime it recovers the loss rate and the collision-pattern probabilities, which recycling cannot estimate, and stays below recycling at every loss rate and below postselection except at low loss, where the two are comparable. Its advantage over recycling grows with the sample count, and it supplies credible intervals that the baselines lack.
Photon-loss mitigation seeks to reconstruct lossless linear-optical output probabilities without discarding all shots with missing photons. This paper replaces recycling-based reconstruction with Bayesian inference over an unstructured output distribution p and uniform per-photon loss rate η. A multinomial likelihood uses independent binomial thinning in each mode, including factors ∏i(yisi) for collision patterns. Dirichlet and Beta priors are combined with sparse-kernel NUTS sampling to obtain posterior means and credible intervals.
Tests use Perceval simulations of four photons, 30 Haar-random interferometers per condition, and typically N=105 shots. For m=20, scored on the no-collision subspace, posterior KL divergence is lower than postselection and the shipped exponential-extrapolation recycling implementation in all 15 tested loss–indistinguishability cells. At η=0.3 and full indistinguishability, KL values are 0.105, 0.195, and 0.315, respectively. For collision-aware inference at m=10, collisions carry about 71% of ideal probability mass; full-distribution KL is 0.012, 0.125, and 0.285 at losses 0.3, 0.7, and 0.9. Sample-budget sweeps show continued posterior improvement while recycling errors plateau.
The evidence is simulation-only and limited to four photons. Exact sampling becomes difficult around 104 probabilities and fails at sufficiently large sample budgets. The no-collision benchmark omits substantial collision mass and biases inferred loss. Partial distinguishability shifts the inferred target rather than being corrected; credible-interval coverage against the ideal distribution can fall to 67%. High-loss postselection KL depends strongly on probability flooring. Uniform loss, photon-number-resolving detection, and unmodelled multiphoton emission further limit applicability; asymptotic unbiasedness is not established.
Photon loss rates set an effective upper limit on the size of computations that can be run on current linear optical quantum devices. We present a family of techniques designed to mitigate the effects of photon loss on both output probabilities and expectation values derived from noisy linear optical circuits composed of an input of n photons, an m --mode interferometer, and m single photon detectors. Central to these techniques is the construction recycled probabilities . Recycled probabilities are constructed from output statistics affected by loss, and are designed to amplify the signal of the ideal (lossless) probabilities. Classical postprocessing techniques then take recycled probabilities as input and output a set of loss-mitigated probabilities, or expectation values. Our postprocessing methods result in biased estimators of the lossless probabilities. Nevertheless, we provide both analytical and numerical evidence that these methods can be applied, up to large sample sizes, to produce output probabilities with lower combined bias and statistical errors than the statistical errors of the output probabilities obtained from postselection. Therefore, these methods can outperform postselection - currently the standard method of coping with photon loss when sampling from discrete variable linear optical quantum circuits. In contrast, we provide evidence that the popular zero-noise extrapolation technique cannot improve on the performance of postselection for any photon loss rate.
Understanding how photon loss changes the classical complexity of BosonSampling is central to assessing the robustness of near-term photonic quantum advantage. In this work, we present a polynomial-time classical simulation algorithm for lossy BosonSampling that applies whenever a constant fraction of the input photons survives. We consider the standard fixed-loss model, where N input photons propagate through an M-mode interferometer and exactly N−n photons are lost uniformly at random, leaving n surviving photons at the output. For any constant surviving fraction n/N∈(0,1), the algorithm samples from a distribution within total variation distance ε of the fixed-loss BosonSampling distribution with probability at least 1−δ over Haar-random interferometers, in time poly(M,N,ε−1,δ−1). We further extend the fixed-loss analysis to the uniform-transmission-loss model and obtain analogous polynomial-time simulability for any fixed transmission rate η∈(0,1). These results substantially extend the known provably simulable regime of n=O(εN) surviving photons or η=O(ε/N) transmission rate, and demonstrate that even the experimentally relevant linear-survival regime can admit efficient classical simulation under the conditions considered here.
BosonSampling is an intermediate model of quantum computation where linear-optical networks are used to solve sampling problems expected to be hard for classical computers. Since these devices are not expected to be universal for quantum computation, it remains an open question of whether any error-correction techniques can be applied to them, and thus it is important to investigate how robust the model is under natural experimental imperfections, such as losses and imperfect control of parameters. Here, we investigate the complexity of BosonSampling under photon losses, more specifically, the case where an unknown subset of the photons is randomly lost at the sources. We show that if k out of n photons are lost, then we cannot sample classically from a distribution that is $1/{n}^{\mathrm{\ensuremath{\Theta}}(k)}$ close (in total variation distance) to the ideal distribution, unless a BPPNP machine can estimate the permanents of Gaussian matrices in nO(k) time. In particular, if k is constant, this implies that simulating lossy BosonSampling is hard for a classical computer, under exactly the same complexity assumption used for the original lossless case.
For near-term quantum devices, an important challenge is to develop efficient methods to certify that noise levels are low enough to allow potentially useful applications to be carried out. We present such a method tailored to photonic quantum devices consisting of single photon sources coupled to linear optical circuits coupled to photon detectors. It uses the output statistics of BosonSampling experiments with input size n (n input photons in the ideal case). We propose a series of benchmark tests targetting two main sources of noise, namely photon loss and distinguishability. Our method results in a single-number metric, the Photonic Quality Factor, defined as the largest number of input photons for which the output statistics pass all tests. We provide strong evidence that passing all tests implies that our experiments are not efficiently classically simulable, by showing how several existing classical algorithms for efficiently simulating noisy BosonSampling fail the tests. Finally we show that BosonSampling experiments with average photon loss rate per mode scaling as o(1) and average fidelity of $ (1-o(\frac{1}{n^6}))^2$ between any two single photon states is sufficient to keep passing our tests. Unsurprisingly, our results highlight that scaling in a manner that avoids efficient classical simulability will at some point necessarily require error correction and mitigation.
We analyze the impact of photon loss on the photon-number statistics of Gaussian states. Specifically, we propose and carefully evaluate several methods to mitigate deviations in the photon-number distributions of lossy (displaced) squeezed vacuum states from those of their lossless counterparts. These methods rely on appropriately redefining the parameters of Gaussian states when the loss budget is known in order to recover, as closely as possible, the desired photon-number distribution associated with each target state. While it is intrinsically hard to directly optimize the photon-number distribution of high-dimensional, correlated multimode Gaussian states, the proposed methods are instead based on optimizing specific key properties such as fidelity, phase-space functions, low-order moments of the underlying photon-number statistics, or overlap with the vacuum state. In particular, our results show that optimizing the fidelity between a pure Gaussian target state and a modified Gaussian state that has passed through a loss channel does typically not result in closeness of the corresponding photon-number distributions. Furthermore, we show that correcting for the vacuum overlap minimizes the deviation in the photon-number distribution for large parameter ranges which we explicitly prove for single-mode squeezed vacuum and provide numerical evidence for general (displaced) squeezed vacuum states. As photon loss is a key limitation for Gaussian boson sampling, our results provide insights into the feasibility and limitations of such photonic quantum simulations in lossy environments and offer guidelines for mitigating these imperfections.
Photon loss makes Gaussian boson sampling easier to simulate classically. A standard approach replaces the lossy input by classical light, and it stays accurate while the mean number of surviving photons grows no faster than the square root of the number N of squeezed inputs at fixed squeezing and accuracy. We prove that this scale is a limit of replacement by classical light. For explicit interferometers, no state with a nonnegative Glauber--Sudarshan P representation, Gaussian or not and however correlated, keeps the error in the photon count distribution small beyond this scale. We then surpass this limit by approximating the detection instead of the input. The sampler evolves the state exactly between detections and, after each detection, replaces the conditional state by a Gaussian with the same covariance. For every interferometer its total variation error is O(Nη3) at polynomial expected cost, where η is the photon survival probability. The mean surviving photon number can therefore grow as N2/3 at fixed accuracy. A separate pair construction reaches N3/4 when every input mode is equally squeezed.