Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Fixed-order photon-number losses need only polynomial samples to train photonic circuits.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 07:06 UTC pith:OC5WT6PD

load-bearing objection Worth engaging: the variance-ratio framework and negative results are solid, but the headline trainability claim for fixed-order photon-number monomials rests on an unproven polynomial-decay conjecture. the 3 major comments →

arxiv 2607.21544 v1 pith:OC5WT6PD submitted 2026-07-23 quant-ph

The trainability of photonic quantum circuits

Alexander Makarovskiy , Adam Taylor , Zhenghao Li , Michael Hanks , Aubrey Clark , M. S. Kim , Ian Walmsley , William Clements This is my paper
classification quant-ph PACS 03.67.Lx42.50.Ex
keywords variational quantum algorithmsphotonic circuitspassive linear opticsphoton-number observablesvariance ratiobarren plateaussample complexityquantum machine learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks when a variational photonic circuit—a passive linear-optical network acting on a fixed number of photons—can be trained without an exponential number of measurements. It argues that trainability is controlled by the ratio of sample variance to circuit variance: this ratio sets the number of shots needed to resolve loss differences and gradients. For fixed-order photon-number polynomials at constant photon density, the paper finds the ratio grows only polynomially with the number of modes, so these losses are efficiently trainable; high-order polynomials and output-probability observables give exponentially large ratios and are not trainable. A sympathetic reader would care because this identifies a practical class of photonic loss functions that avoid barren plateaus and still admit polynomial quantum speedups over term-by-term classical estimation.

Core claim

The central claim is that for passive linear-optical circuits with fixed photon number, trainability is governed by the ratio of sample variance to circuit variance rather than by circuit variance alone. Photonic observables are unbounded, so an exponentially large sample variance can mask a benign circuit variance. The paper proves closed-form first moments for photon-number monomials and computes second moments via representation-theoretic formulas, then uses the conjectured polynomial decay Var_circ(α) ~ µ1(α)^2/poly(M) to conclude that fixed-order monomials and typical fixed-order polynomials have polynomially scaling variance ratios, while high-order monomials and output-probability pro

What carries the argument

The central object is the variance ratio Varsamp( ˆO)/Varcirc( ˆO). For photon-number monomials, the first Haar moment reduces to products of ordered Bell polynomials built from Stirling numbers of the second kind, giving the asymptotic size of µ1; the second moment is evaluated through projections onto irreducible representations of U(M). The load-bearing identity is Eq. (14): the variance ratio equals poly(M) times a factor exponential in the order and support of the monomial, a factor that is constant when the order is fixed. A generalized parameter-shift rule converts loss estimates into gradient estimates with at most polynomial overhead, and a concentration-equivalence theorem shows th

Load-bearing premise

The results assume that the circuit variance of photon-number monomials decays only polynomially with system size (Eq. 13); this is a numerical conjecture for small systems, not an analytic proof, and if the true decay is exponential the fixed-order trainability claims collapse.

What would settle it

Compute the circuit variance for a fixed-order monomial (e.g., n^2 or n1 n2) at n=νM for M increasing to hundreds or thousands, using the exact second-moment formulas; if the gap between the second and first Haar moments decays exponentially in M, or the variance ratio increases exponentially, the central trainability claim fails. An analytic lower bound on the second-moment gap would settle the question definitively.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Fixed-order photon-number monomials and typical fixed-order polynomials can be trained with polynomially many samples at constant photon density, avoiding barren plateaus for these losses.
  • Observables built from single output probabilities or coarse-grained output bins require exponentially many samples, ruling them out for scalable variational training.
  • Neural network observables implement photon-number polynomials with exponentially many terms in O(LM^2) operations per sample, bypassing explicit polynomial expansion.
  • Quantum estimation of balanced photon-number polynomials can achieve a quadratic speedup over term-by-term classical estimation methods.
  • For the relevant circuit ensembles, exponential loss concentration is equivalent to barren plateaus across all parameters, so the variance ratio is a complete trainability diagnostic.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If Eq. (13) is later proved or replaced by a polynomial upper bound, the fixed-order trainability results become unconditional; if the true circuit-variance decay is exponential, the positive results collapse. This is an editorial caution, not a paper claim.
  • The framework likely extends to Gaussian and superposition photon-number inputs if the first-moment asymptotics can be generalized; the authors explicitly leave this open, and the extension is testable numerically.
  • The quadratic quantum speedup over classical estimation relies on cancellation between PNP coefficients; adversarial coefficient choices could restore exponential classical cost, so practical speedups depend on observable design.
  • Neural network observables give a concrete way to instantiate high-order PNPs without expanding terms, but their trainability still inherits the order dependence, so the useful regime is bounded by the network depth.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces a trainability criterion for passive linear-optical variational circuits based on the ratio of the sample variance (shot-noise spread at fixed parameters) to the circuit variance (fluctuation of the loss under Haar-random unitaries). It derives an exact first-moment formula for photon-number monomials (Theorem 1, Eqs. (9)–(10)), uses a second-moment decomposition from Mhiri et al. (Eq. (12)) to compute sample and circuit variances, and conjectures on numerical grounds that the circuit variance decays polynomially (Eq. (13)). Combining these ingredients, Eq. (14) predicts that fixed-order PNMs/PNPs at constant photon density require only polynomially many samples, while high-order PNMs and output-probability observables require exponentially many. The paper also proposes neural-network observables as efficient implementations of PNPs, claims a polynomial speed-up over the Lim-Oh classical estimator, and reports training simulations and a hardware demonstration on ORCA PT-2.

Significance. If Eq. (13) is eventually proven, the paper would provide a useful framework and a genuinely positive trainability result for photonic circuits, complementing the negative results. The first-moment theorem and the sample-variance bounds are carefully derived; the output-probability and coarse-grained-binning negative results (Eqs. (21)–(23), (26)) are exact and important. The paper is also commendably explicit about its key limitation. However, because the polynomial-sample conclusion for fixed-order observables depends on an unproven conjecture about the gap between the second and first moments, the central positive claim is currently conditional. The numerical evidence is too limited to distinguish polynomial from weak exponential decay. No fitted parameters are used in the central prediction, which is a strength; the load-bearing assumption is clearly identified rather than hidden.

major comments (3)
  1. [Sec. 3.1, Eq. (13); Sec. 5, Discussion] The central positive claim that fixed-order PNMs are trainable with polynomially many samples rests entirely on the conjectured polynomial decay Varcirc(α) ~ µ1(α)^2/poly(M). The paper itself concedes in the Discussion that this scaling is 'not yet established analytically.' The exact second-moment formula (12) only gives M-independent bounds µ1(α)^2 ≤ µ2(α) ≤ µ1(2α); the gap µ2(α) − µ1(α)^2 could decay polynomially or exponentially. Figures 3 and 4 cover modest M and cannot distinguish a power law from a weak exponential such as exp(−√M). If Eq. (13) fails, Eq. (14) and the main positive conclusion collapse, although the negative results for high-order PNMs and output probabilities would survive. Please either prove the conjecture (or a sufficient lower bound on Varcirc) or reclassify the polynomial-sample claim as conditional throughout the abstract, introduction, and conclusions.
  2. [Sec. 3.2.2; Appendix D.3–D.4] The claimed polynomial speed-up over classical estimation relies on the O(|A|) sample-variance bound for zero-mean random coefficients (Eq. (18) and Appendix D.3). For generic coefficients, Eq. (17) gives O(|A|^2) for the quantum estimator as well, and the term-by-term Lim-Oh approach also costs O(|A|^2), so no speed-up is shown in the worst case. Moreover, the comparison is to a specific classical algorithm (Lim-Oh) and to specific emulators; the paper does not rule out better classical estimators. The abstract's phrasing 'polynomial speed-up over multiple classical methods' therefore overstates what is established. Please restrict the speed-up claim to the random-coefficient ensemble and to the specific classical estimators analyzed.
  3. [Sec. 2.2; Appendix A.4] The variance-ratio criterion is sufficient for trainability only if a non-negligible fraction of randomly drawn parameter pairs have loss differences |D| of order sqrt(Varcirc). This is condition (i) or (ii) in Appendix A.4, which is assumed rather than proven for PNMs. Without such an anticoncentration-type condition, a polynomial variance ratio does not by itself guarantee that typical loss differences can be resolved; the landscape could be flat except on an exponentially rare set. The paper's main trainability conclusion therefore needs either a proof of this condition for fixed-order PNMs or an explicit statement that the conclusion is conditional on it.
minor comments (4)
  1. [Sec. 3.2.3, Eq. (20)] The degree bound in Eq. (20) is written as |α| ≤ 2L, but a network with L linear layers interleaved with quadratic activations should have degree 2^L. This is confirmed by Sec. 4.1, where a 3-layer network is said to contain terms up to order 8. Please correct Eq. (20) and adjust the term-count discussion accordingly.
  2. [Fig. 4 caption] The caption states the input state is |1,0⟩ with photon density ν = 0.5 while M is varied. For M > 2, |1,0⟩ has n = 1 and ν = 1/M, not 0.5. Please clarify the intended input state (e.g., an alternating 1,0 pattern with n = M/2) and align the notation.
  3. [Sec. 3.2.1, Eq. (16)] The lower bound Varsamp(α) ≥ (2^K − 1)ν^{2K} is stated without conditions. It is not valid when K > n, where the monomial vanishes identically, and the surrounding text explicitly considers the oversaturated regime ν > 1. Please state the required conditions (e.g., K ≤ n and/or ν ≥ 1) so the reader can see the scope of the high-order untrainability claim.
  4. [Throughout] There are a few typographical slips, e.g., 'Howeevr' in Sec. 3.2.4 and 'these variables' in Sec. 2.1. A careful proofread is recommended.

Circularity Check

0 steps flagged

No circularity: the fixed-order trainability claim is conditional on an explicitly labeled conjecture (Eq. 13), not on a fitted parameter renamed as a prediction or on a load-bearing self-citation.

full rationale

Walking the derivation chain, the first- and second-moment formulas (Theorem 1 and Eq. (12)) are analytic and are drawn from external, standard results (Mhiri et al., Weingarten calculus, Lim-Oh, parameter-shift rules). The paper's self-citations (Clements decomposition, time-bin hardness, QUBO examples) are technical or contextual and are not premises of the variance-ratio results. The positive fixed-order claim (Eq. 14) does use the conjectured polynomial decay Varcirc ~ µ1^2/poly(M) (Eq. 13) as its only bridge to poly(M) sample complexity. However, the paper explicitly labels Eq. (13) as a conjecture supported by numerics, and the Discussion concedes: 'Our analysis relies on a conjectured scaling of the gap between the first and second Haar-average moments... not yet established analytically.' This is a soundness/verification gap, not a circular reduction: Eq. (13) is not derived from Eq. (14), no fitted coefficient is reused as a prediction, and the negative results for high-order PNMs and output-probability observables are analytic and do not require Eq. (13). The variance-ratio trainability criterion is defined independently (Appendix A, Proposition 1), so the central framework is not self-definitional. Score 0.

Axiom & Free-Parameter Ledger

1 free parameters · 8 axioms · 0 invented entities

The paper introduces no fitted constants; the main unproven input is the polynomial-decay conjecture. It also imports second-moment formulas from [51], assumes constant photon density, Haar-random/dialled circuits, Fock inputs, and random balanced PNP coefficients. No new physical entities are postulated.

free parameters (1)
  • Photon density ν = n/M
    Choice of scaling regime n=νM with ν fixed (e.g. 0.5, 1); all fixed-order trainability conclusions are asymptotic in this limit.
axioms (8)
  • ad hoc to paper Circuit-variance polynomial decay conjecture, Eq. (13): Varcirc(α) ~ µ1(α)^2/poly(M)
    Unproved; used to convert first/second moments into the polynomial variance ratio and the trainability conclusion.
  • domain assumption Second-moment decomposition of Mhiri et al. [51], Eq. (12): µ2(O;ρ) = Σ_k ... ||ρ_k||^2 ||O_k||^2
    Central to all second-moment and variance results; cited as a preprint, not re-derived or machine-checked here.
  • domain assumption Constant photon-density limit n=νM with ν fixed
    All fixed-order trainability claims are asymptotic in this scaling.
  • domain assumption Haar-random or Haar-dialling circuit ensemble
    Trainability is assessed for random initialisation; does not cover structured initialisations or non-uniform parameter distributions.
  • domain assumption Fixed total photon-number input states; post-selection used for hardware
    Theory applies to definite-photon-number Fock inputs; Gaussian and superposition states are explicitly left open.
  • domain assumption For typical PNP coefficients, E[c_α]=0 and |c_α| independent of M (Appendix D.3)
    The O(|A|) sample-variance bound and the quantum-over-classical speedup rely on off-diagonal cancellation of random coefficients.
  • standard math Efron–Stein inequality, Poincaré inequalities, and log-concavity of the Haar-dialling measure (Appendix B)
    Proof of equivalence between loss concentration and barren plateaus rests on these; likely sound but not formalized.
  • domain assumption Anticoncentration or Gaussianity of loss differences (Appendix A.4)
    Needed to ensure the variance ratio controls typical, not just average, pair differences.

pith-pipeline@v1.3.0-alltime-deepseek · 34779 in / 14014 out tokens · 135611 ms · 2026-08-01T07:06:44.045105+00:00 · methodology

0 comments
read the original abstract

Variational quantum algorithms are a leading approach to near-term quantum computing, but their scalability can be limited by barren plateaus and the sampling cost of resolving small changes in the loss landscape. Here, we study the trainability of passive linear-optical quantum circuits and introduce a framework based on the ratio of sample variance to circuit variance. This ratio determines the number of circuit samples required to resolve local loss differences and gradients to proportional accuracy. We apply this framework to photon-number observables and identify both trainable and non-trainable regimes. Supported by analytic results and a numerically observed polynomial decay of the circuit variance, we find that fixed-order photon-number polynomials require only polynomially many samples as the system size grows, whereas high-order polynomials and observables based on output probabilities generally require exponentially many samples. Within the trainable regime, we further identify classes of observables in which quantum estimation achieves a polynomial speed-up over multiple classical methods. Within this family, neural network observables provide one practical construction that allow measurement outcomes to be efficiently processed into the desired polynomial. These results establish photonic variational quantum computing as a promising platform for near-term applications.

Figures

Figures reproduced from arXiv: 2607.21544 by Adam Taylor, Alexander Makarovskiy, Aubrey Clark, Ian Walmsley, Michael Hanks, M. S. Kim, William Clements, Zhenghao Li.

Figure 1
Figure 1. Figure 1: We illustrate the class of VQAs considered in this paper. The ratio between [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: We visualise the distribution of a finite-sample estimate of a fixed observable [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sample and circuit variances as functions of the number of modes [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Circuit variance as a function of the number of modes [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Numerical scaling of the circuit and sample variances for two classes of photon [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: We show a 3-layer neural network observable implementing a photon-number [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Training of simulated photonic circuits on a 3-layer neural network observable. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Illustration of the PT-2 photonic system. Squeezed states of light are generated [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Training of a PT-2 system to minimise loss. The PT-2 is run for 48 modes per [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Plot of a PNM first moment against the total weight [PITH_FULL_IMAGE:figures/full_fig_p043_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Classical simulation and model concentration in passive linear optics

    quant-ph 2026-07 conditional novelty 6.0

    In passive linear optics, expectation-value concentration is set by misalignment of state and observable irrep purities, and known non-concentrating regimes remain largely classically tractable via irrep truncation or...

Reference graph

Works this paper leans on

79 extracted references · 8 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Quantum computing in the NISQ era and beyond.Quantum2,79 (2018)

    Preskill, J. Quantum computing in the NISQ era and beyond.Quantum2,79 (2018)

    2018

  2. [2]

    Shor, P. W. Polynomial-time algorithms for prime factorization and discrete loga- rithms on a quantum computer.SIAM review41,303–332 (1999)

    1999

  3. [3]

    & Eker ˚ a, M

    Gidney, C. & Eker ˚ a, M. How to factor 2048 bit RSA integers in 8 hours using 20 million noisy qubits.Quantum5,433 (2021)

    2048

  4. [4]

    Cerezo, M.et al.Variational quantum algorithms.Nature Reviews Physics3,625– 644 (2021)

    2021

  5. [5]

    R., Romero, J., Babbush, R

    McClean, J. R., Romero, J., Babbush, R. & Aspuru-Guzik, A. The theory of vari- ational hybrid quantum-classical algorithms.New Journal of Physics18,023023 (2016)

    2016

  6. [6]

    & Gutmann, S

    Farhi, E., Goldstone, J. & Gutmann, S. A quantum approximate optimization algo- rithm.arXiv preprint arXiv:1411.4028(2014)

    Pith/arXiv arXiv 2014

  7. [7]

    & Lukin, M

    Zhou, L., Wang, S.-T., Choi, S., Pichler, H. & Lukin, M. D. Quantum approximate optimization algorithm: Performance, mechanism, and implementation on near-term devices.Physical Review X10,021067 (2020)

    2020

  8. [8]

    McArdle, S., Endo, S., Aspuru-Guzik, A., Benjamin, S. C. & Yuan, X. Quantum computational chemistry.Reviews of Modern Physics92,015003 (2020)

    2020

  9. [9]

    Nature communications5,4213 (2014)

    Peruzzo, A.et al.A variational eigenvalue solver on a photonic quantum processor. Nature communications5,4213 (2014)

    2014

  10. [10]

    Biamonte, J.et al.Quantum machine learning.Nature549,195–202 (2017)

    2017

  11. [11]

    Abbas, A.et al.The power of quantum neural networks.Nature computational science 1,403–409 (2021)

    2021

  12. [12]

    Xu, X.et al.Variational algorithms for linear algebra.Science Bulletin66,2181– 2188 (2021)

    2021

  13. [13]

    Consiglio, M.et al.Variational Gibbs state preparation on noisy intermediate-scale quantum devices.Phys. Rev. A110,012445 (1 2024)

    2024

  14. [14]

    C.et al.Variational quantum state preparation for quantum- enhanced metrology in noisy systems.Phys

    Zu˜ niga Castro, J. C.et al.Variational quantum state preparation for quantum- enhanced metrology in noisy systems.Phys. Rev. A110,052615 (5 2024)

    2024

  15. [15]

    & Benjamin, S

    Jones, T., Endo, S., McArdle, S., Yuan, X. & Benjamin, S. C. Variational quantum algorithms for discovering Hamiltonian spectra.Physical Review A99,062304 (2019)

    2019

  16. [16]

    Endo, S., Sun, J., Li, Y., Benjamin, S. C. & Yuan, X. Variational Quantum Simulation of General Processes.Phys. Rev. Lett.125,010501 (1 2020)

    2020

  17. [17]

    R., Boixo, S., Smelyanskiy, V

    McClean, J. R., Boixo, S., Smelyanskiy, V. N., Babbush, R. & Neven, H. Barren plateaus in quantum neural network training landscapes.Nature communications9, 4812 (2018). 22

    2018

  18. [18]

    & Coles, P

    Arrasmith, A., Holmes, Z., Cerezo, M. & Coles, P. J. Equivalence of quantum barren plateaus to cost concentration and narrow gorges.Quantum Sci. Technol.7(2022)

    2022

  19. [19]

    Larocca, M.et al.Barren plateaus in variational quantum computing.Nature Reviews Physics,1–16 (2025)

    2025

  20. [20]

    & Coles, P

    Cerezo, M., Sone, A., Volkoff, T., Cincio, L. & Coles, P. J. Cost function dependent barren plateaus in shallow parametrized quantum circuits.Nature communications 12,1791 (2021)

    2021

  21. [21]

    & Wiebe, N

    Ortiz Marrero, C., Kieferov´ a, M. & Wiebe, N. Entanglement-Induced Barren Plateaus. PRX Quantum2,040316 (4 2021)

    2021

  22. [22]

    L., Najafi, K., Gao, X

    Patti, T. L., Najafi, K., Gao, X. & Yelin, S. F. Entanglement devised barren plateau mitigation.Physical Review Research3,033090 (2021)

    2021

  23. [23]

    Nature communications12,6961 (2021)

    Wang, S.et al.Noise-induced barren plateaus in variational quantum algorithms. Nature communications12,6961 (2021)

    2021

  24. [24]

    Ragone, M.et al.A lie algebraic theory of barren plateaus for deep parameterized quantum circuits.Nature Communications15,7172 (2024)

    2024

  25. [25]

    Larocca, M.et al.Diagnosing Barren Plateaus with Tools from Quantum Optimal Control.Quantum6,824.issn: 2521-327X (2022)

    2022

  26. [26]

    & Benedetti, M

    Grant, E., Wossnig, L., Ostaszewski, M. & Benedetti, M. An initialization strategy for addressing barren plateaus in parametrized quantum circuits.Quantum3,214. issn: 2521-327X (2019)

    2019

  27. [27]

    Pesah, A.et al.Absence of Barren Plateaus in Quantum Convolutional Neural Net- works.Phys. Rev. X11,041011 (4 2021)

    2021

  28. [28]

    R., Mohseni, M., Van Der Smagt, P

    Skolik, A., McClean, J. R., Mohseni, M., Van Der Smagt, P. & Leib, M. Layerwise learning for quantum neural networks.Quantum Machine Intelligence3,5 (2021)

    2021

  29. [29]

    H., Medina, R

    Sack, S. H., Medina, R. A., Michailidis, A. A., Kueng, R. & Serbyn, M. Avoiding barren plateaus using classical shadows.PRX Quantum3,020365 (2022)

    2022

  30. [30]

    Cerezo, M.et al.Does provable absence of barren plateaus imply classical simulabil- ity?Nature Communications16,7907 (2025)

    2025

  31. [31]

    L., Larocca, M., Cincio, L., Cerezo, M

    Goh, M. L., Larocca, M., Cincio, L., Cerezo, M. & Sauvage, F. Lie-algebraic classical simulations for quantum computing.Phys. Rev. Res.7,033266 (3 2025)

    2025

  32. [32]

    Bermejo, P.et al.Quantum Convolutional Neural Networks are Effectively Classically Simulable.PRX Quantum7,020304 (2 2026)

    2026

  33. [33]

    arXiv preprint arXiv:2506.19707(2025)

    Gong, S.-Q.et al.Enhanced Image Recognition Using Gaussian Boson Sampling. arXiv preprint arXiv:2506.19707(2025)

    Pith/arXiv arXiv 2025

  34. [34]

    Cimini, V.et al.Large-scale quantum reservoir computing using a Gaussian Boson Sampler.npj Quantum Information(2026). 23

    2026

  35. [35]

    Aaronson, S. & Arkhipov, A.The computational complexity of linear opticsinPro- ceedings of the Forty-Third Annual ACM Symposium on Theory of Computing(As- sociation for Computing Machinery, San Jose, California, USA, 2011), 333–342.isbn: 9781450306911

    2011

  36. [36]

    Deshpande, A.et al.Quantum computational advantage via high-dimensional Gaus- sian boson sampling.Science advances8,eabi7894 (2022)

    2022

  37. [37]

    & Eisert, J

    Hangleiter, D. & Eisert, J. Computational advantage of quantum random sampling. Reviews of Modern Physics95,035001 (2023)

    2023

  38. [38]

    R., Bulmer, J

    Li, Z., Solomons, N. R., Bulmer, J. F., Patel, R. B. & Walmsley, I. A. A complexity transition in displaced Gaussian Boson sampling.npj Quantum Information11,119 (2025)

    2025

  39. [39]

    arXiv preprint arXiv:2602.23501(2026)

    Li, Z.et al.Machine learning of quantum data using optimal similarity measurements. arXiv preprint arXiv:2602.23501(2026)

    arXiv 2026

  40. [40]

    Hoch, F.et al.Quantum machine learning with adaptive boson sampling via post- selection.Nature Communications16,902 (2025)

    2025

  41. [41]

    & Zimbor´ as, Z

    Kolarovszki, Z., Bak´ o, B., Oszmaniec, M., Oh, C. & Zimbor´ as, Z. Generative mod- eling with Gaussian Boson Sampling: classically trainable Bosonic Born Machines. arXiv preprint arXiv:2603.11195(2026)

    arXiv 2026

  42. [42]

    Bacarreza, O.et al.Quantum latent distributions in deep generative models.arXiv preprint arXiv:2508.19857(2025)

    Pith/arXiv arXiv 2025

  43. [43]

    & Sohbi, A

    Chabaud, U., Markham, D. & Sohbi, A. Quantum machine learning with adaptive linear optics.Quantum5,496.issn: 2521-327X (2021)

    2021

  44. [44]

    Monbroussou, L.et al.Toward quantum advantage with photonic state injection. Phys. Rev. Res.7,033051 (3 2025)

    2025

  45. [45]

    Monbroussou, L.et al.Photonic quantum convolutional neural networks with adap- tive state injection.Advanced Photonics7,066012–066012 (2025)

    2025

  46. [46]

    Reck, M., Zeilinger, A., Bernstein, H. J. & Bertani, P. Experimental realization of any discrete unitary operator.Phys. Rev. Lett.73,58–61 (1 1994)

    1994

  47. [47]

    R., Humphreys, P

    Clements, W. R., Humphreys, P. C., Metcalf, B. J., Kolthammer, W. S. & Walmsley, I. A. Optimal design for universal multiport interferometers.Optica3,1460–1465 (2016)

    2016

  48. [48]

    Bell, B. A. & Walmsley, I. A. Further compactifying linear optical unitaries.Apl Photonics6(2021)

    2021

  49. [49]

    & Aaronson, S

    Bouland, A. & Aaronson, S. Generation of universal linear optics by any beam split- ter.Physical Review A89,062316 (2014)

    2014

  50. [50]

    & Quesada, N

    Cardin, Y. & Quesada, N. Photon-number moments and cumulants of Gaussian states.Quantum8,1521 (2024)

    2024

  51. [51]

    Boson sampling beyond the dilute regime: second moments and anti- concentration2026

    Mhiri, H.et al. Boson sampling beyond the dilute regime: second moments and anti- concentration2026. 24

  52. [52]

    & Oh, C.Classical Algorithms for Estimating Expectation Values in Linear- Optical Circuits2025

    Lim, Y. & Oh, C.Classical Algorithms for Estimating Expectation Values in Linear- Optical Circuits2025

  53. [53]

    Villalonga, B.et al.Efficient approximation of experimental Gaussian boson sam- pling.arXiv preprint arXiv:2109.11525(2021)

    Pith/arXiv arXiv 2021

  54. [54]

    T., Quesada, N

    Dodd, T., Mart ´ ınez-Cifuentes, J., Brown, O. T., Quesada, N. & Garc ´ ıa-Patr´ on, R. A fast and frugal Gaussian Boson Sampling emulator.arXiv preprint arXiv:2511.14923 (2025)

    arXiv 2025

  55. [55]

    & Bowles, J

    Recio-Armengol, E., Ahmed, S. & Bowles, J. Train on classical, deploy on quantum: scaling generative quantum machine learning to a thousand qubits.arXiv preprint arXiv:2503.02934(2025)

    arXiv 2025

  56. [56]

    Br´ adler, K., Friedland, S., Izaac, J., Killoran, N. & Su, D. Graph isomorphism and Gaussian boson sampling.arXiv preprint arXiv:1810.10644(2018)

    Pith/arXiv arXiv 2018

  57. [57]

    Mezher, R., Carvalho, A. F. & Mansfield, S. Solving graph problems with single photons and linear optics.Physical Review A108,032405 (2023)

    2023

  58. [58]

    Nikolopoulos, G. M. & Brougham, T. Decision and function problems based on boson sampling.Physical Review A94,012315 (2016)

    2016

  59. [59]

    & Duan, L.-M

    Wang, S.-T. & Duan, L.-M. Certification of boson sampling devices with coarse- grained measurements.arXiv preprint arXiv:1601.02627(2016)

    Pith/arXiv arXiv 2016

  60. [60]

    Singh, D.et al.Proof-of-work consensus by quantum sampling.Quantum Science and Technology10,025020 (2025)

    2025

  61. [61]

    Nikolopoulos, G. M. Cryptographic one-way function based on boson sampling: GM Nikolopoulos.Quantum Information Processing18,259 (2019)

    2019

  62. [62]

    D., Makarovskiy, A., Garc ´ ıa-Patr´ on, R

    Nov´ ak, S., Roberts, D. D., Makarovskiy, A., Garc ´ ıa-Patr´ on, R. & Clements, W. R. Boundaries for quantum advantage with single photons and loop-based time-bin in- terferometers.Quantum9,1915 (2025)

    1915

  63. [63]

    Makarovskiy, A.et al.A Binary Optimisation Algorithm for Near-Term Photonic Quantum Processors.arXiv preprint arXiv:2510.08274(2025)

    arXiv 2025

  64. [64]

    & Shah, T.Gaussian Boson Sampling for binary optimizationin2023 IEEE International Conference on Quantum Computing and Engineering (QCE)2(2023), 332–333

    Cazalis, J., Chai, Y., Jansen, K., K¨ uhn, S. & Shah, T.Gaussian Boson Sampling for binary optimizationin2023 IEEE International Conference on Quantum Computing and Engineering (QCE)2(2023), 332–333

    2023

  65. [65]

    Lee, D.et al.Error-mitigated photonic variational quantum eigensolver using a single- photon ququart.Optica9,88–95 (2022)

    2022

  66. [66]

    & Kim, M

    Taylor, A., Bressanini, G., Kwon, H. & Kim, M. S. Quantum error cancellation in photonic systems: Undoing photon losses.Phys. Rev. A110,022622 (2 2024)

    2024

  67. [67]

    & Kashefi, E.Classical simula- tion and model concentration in passive linear optics2026

    Monbroussou, L., Thomas, H., Mhiri, H., Holmes, Z. & Kashefi, E.Classical simula- tion and model concentration in passive linear optics2026

  68. [68]

    Facelli, G.et al.Exact gradients for linear optics with single photons.arXiv preprint arXiv:2409.16369(2024). 25

    Pith/arXiv arXiv 2024

  69. [69]

    Pappalardo, A.et al.Photonic parameter-shift rule: Enabling gradient computation for photonic quantum computers.Physical Review A111,032429 (2025)

    2025

  70. [70]

    J., Chakhmakhchyan, L., O’Brien, J

    Russell, N. J., Chakhmakhchyan, L., O’Brien, J. L. & Laing, A. Direct dialling of Haar random unitary matrices.New journal of physics19,033007 (2017)

    2017

  71. [71]

    & Stein, C

    Efron, B. & Stein, C. The Jackknife Estimate of Variance.The Annals of Statistics 9,586–596 (1981)

    1981

  72. [72]

    Isoperimetric and analytic inequalities for log-concave probability measures.The An- nals of Probability27,1903–1921 (1999)

    1903

  73. [73]

    & ´Sniady, P

    Collins, B. & ´Sniady, P. Integration with respect to the Haar measure on unitary, orthogonal and symplectic group.Communications in Mathematical Physics264, 773–795 (2006)

    2006

  74. [74]

    & Saad, N

    Collins, B., Matsumoto, S. & Saad, N. Integration of invariant matrices and moments of inverses of Ginibre and Wishart matrices.Journal of Multivariate Analysis126, 1–13 (2014)

    2014

  75. [75]

    & Kim, M

    Taylor, A., Hanks, M., Kwon, H. & Kim, M. S.Optimal Quantum Information Trans- mission Under a Continuous-Variable Erasure Channel2025

  76. [76]

    Belovas, I.Mathematica Slovaca73,335–342 (2023)

    2023

  77. [77]

    & Quesada, N

    Trudeau, M., Emeriau, P.-E. & Quesada, N. Parameter-Shift Rules for Gradients in Boson Sampling Experiments.arXiv preprint arXiv:2607.15160(2026)

    Pith/arXiv arXiv 2026

  78. [78]

    Hoch, F.et al.Variational approach to photonic quantum circuits via the parameter shift rule.Physical Review Research7,023227 (2025)

    2025

  79. [79]

    & Lin, C

    Wierichs, D., Izaac, J., Wang, C. & Lin, C. Y.-Y. General parameter-shift rules for quantum gradients.Quantum6,677 (2022). 26 A Sample complexity for resolving loss differences In this section we justify the claims of Sec. 2.2 that a number of samples scaling with the variance ratio Varsamp( ˆO)/Varcirc( ˆO) is sufficient to resolve loss differences at di...

    2022