acceptodds
Under review as a conference paper at ICLR 2027

FockTrust: Photonic Natural Gradients in Detector-Count Space

Abstract

Photonic learners observe detector counts rather than optical states, so their natural gradient must follow the distribution actually measured. Loss, missed photons, dark counts, and readout alter which parameter directions remain distinguishable. We introduce FockTrust, which defines the Fisher metric directly in detector-count space. Shared base and parameter-shift histograms estimate the task gradient and full metric, while covariance-whitened readout moments map count geometry into tangent space. We also bound the effect of finite counts on the resulting step. In matched 2-by-2 experiments at dimension 32, retaining off-diagonal count couplings reduces final loss by 17.6-24.3% relative to diagonal directions under either trust rule. With trust rules tuned on separate development tasks, full count geometry lowers loss by 25.7% on 12 unseen graphs. The advantage over state-QFIM persists in a two-photon study. FockTrust makes the detector model part of optimization and identifies detector-count space as the operative geometry of measured photonic learning.

Primary area:
optimization
Keywords:
photonic machine learning, natural gradient, Fisher information, photon counting, detector noise, quantum photonics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xA9JOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes FockTrust, a natural-gradient optimizer for photonic circuits that defines the Fisher metric using the observed detector count distribution.

The motivation is to align optimization with the measured outputs used by the objective, including measurement effects.

The method combines count-based Fisher estimation with a KL trust region and is evaluated in photonic simulations.

This paper is well outside my expertise, so my judgment should be taken with a grain of salt. I find the motivation sound, but the novelty seems primarily limited to specific applications in photonics, since, to my understanding, most of the components already exist, including Fisher estimation from count data and natural gradients based on observed distributions. I lack the domain expertise to confidently assess the significance of this application-specific contribution. I therefore lean toward a borderline rating, with low confidence.

Critical strengths

The motivation is sound, and the results support it. When the objective depends on detector outputs, it makes sense to define the Fisher metric using those outputs. The reported comparisons support this approach: the count-based optimizer achieves 25.5% and 34.7% lower final KL-divergence losses than the ideal-state QFIM optimizer in the single-photon and two-photon settings, respectively.

Critical weaknesses

The evaluation uses a controlled task whose broader relevance is unclear to me. The experiments train a circuit to match the detector-count distribution of another circuit with the same architecture. This provides a useful demonstration in a setting closely aligned with the count-based metric, but I remain unsure how well the gains extend to practical photonic applications. A clearer connection to those applications would strengthen the paper.

The practical significance of the experimental scale is also unclear to me. It would help to explain how the simulated settings relate to relevant real photonic systems, and whether estimating and solving with the full Fisher is practical / impratical. It will help defend the experiment methodology and practical usefulness. 

The novelty seems mainly specific to the photonics application. To my understanding, most of the ingredients already exist, including Fisher estimation from count data, natural gradients based on observed distributions, and KL trust regions. The main advance therefore seems to be adapting and combining these ideas for photonic detector counts. This is well motivated, but I lack the domain expertise to confidently judge the significance of this application-specific contribution.

My confidence is low because photonic optimization is outside my expertise.

Questions for authors
  1. Generality of the evaluation. How representative is the teacher–student task of practical photonic optimization? Could you provide evidence that the gains extend beyond matching targets generated by the same circuit architecture, ideally through a more application-driven task?

  2. Practical significance. Which photonic applications would benefit from FockTrust, and why would the reported improvements matter in those settings?

  3. Relevant scale. How do the simulated circuit sizes and measurement budgets compare with real photonic systems? At those scales, is estimating and solving with the full Fisher matrix practical?

2 comments

Thank you for the thoughtful review, and for finding the motivation sound and supported by the results. Your three questions are the ones a reader outside photonics would ask first, and we added experiments and text for each of them.

Could you provide evidence that the gains extend beyond matching targets generated by the same circuit architecture, ideally through a more application-driven task?

The in-family target was chosen to isolate the geometry. We added two d=32d=32 tasks without it, on the 12 held-out graphs. In the first, the target comes from a drifted detector, so no circuit parameter reproduces it. FF lowers excess loss by 24% relative to diagonal geometry and exact-QFIM NG, and by 50% relative to Adam. In the second, the target is known only through 2,048 measured samples, which is the setting of learning from experimental data. The reductions are 28%, 24% and 52% (Section 5.4, Appendix C.1).

Which photonic applications would benefit from FockTrust, and why would the reported improvements matter in those settings?

Any photonic learner whose loss is computed from detector counts, for example

  • photonic generative models (Born machines) trained on measured data, which is the second task above,
  • in-situ calibration of a programmable interferometer to a measured output distribution,
  • photonic classifiers that read out click patterns.

In these settings the detector is lossy and its efficiencies differ across modes. The new Proposition 1 shows that the count metric weights the state metric by these efficiencies, and FockTrust reads it from the gradient histograms at no extra measurement cost. The improvement holds across detection levels. With every method retuned, FF/QQ is 0.597, 0.591, 0.564 and 0.512 at mean efficiency 0.72, 0.60, 0.48 and 0.36. Section 4 now names Born machines and in-situ calibration as the target applications.

How do the simulated circuit sizes and measurement budgets compare with real photonic systems? At those scales, is estimating and solving with the full Fisher matrix practical?

A 20-mode universal integrated processor has 380 thermo-optic phase elements that switch at kHz rates (Taballione et al., Quantum 7, 1071, 2023). We added retuned runs at d=64d=64 and d=128d=128, both below this count. On the recursive-tree graphs FF/QFIM-NG is 0.665 [0.601, 0.719] and 0.442 [0.390, 0.488], so the advantage grows with the number of phases.

The full Fisher is practical at these sizes. Every gradient method needs the 1+2d1+2d parameter-shift settings per step, which at kHz reconfiguration is about 0.26 s of hardware time at d=128d=128. Forming and solving the dense count system takes 0.3 ms on one CPU thread, about a thousandth of that, and 0.17 s at d=1024d=1024. A diagonal-plus-rank-8 metric also keeps the dense gain (Appendix C.2).

The novelty seems mainly specific to the photonics application.

We agree that the components are standard. What the paper adds is what they give a photonic learner. The full metric comes from the gradient histograms with no extra circuits and no detector calibration. It is the state QFIM weighted by what the detector resolves (Proposition 1), and under the same controller its same-count estimate lowers final loss by 36% relative to the exact QFIM. Our reply to Reviewer pc1s compares the method with prior work in more detail.

Thank you again, and we would be glad to answer any further questions.

I appreicate the response. I do not think that I can accurately judge these responses still. I maintain the score yet I hope that other reviewers to give this paper the justice.

Reviewer_pc1sOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FockTrust which is an optimization framework based on natural gradients for photonic quantum circuits. Its motivation is that physical losses such as photon losses, dark counts or other measurement imperfections can alter the accuracy of parameter directions under the conventional state-based quantum Fisher information. The method constructs a Fisher information matrix from the observed distribution and estimates [1] the task gradient and [2] metric using consistent base and parameter-shift histograms. The authors perform experiments om simulated hardware single and two-photon quantum circuits. Authors report that full count geometry reduces loss value by 25.7 in percentage relative to diagonal count geometry on 12 graph circuits. The method outperforms an exact state-QFIM natural-gradient baseline.

Critical strengths

1. Problem formulation is clear I think the targeted problem is clear. During optimization of photonic quantum circuits, losses in photon detector can change which parameter changes are dintinguishable. Therefore, it has a physical motivation to define the optimization geometry in the detector-count distribution.

2. Good experimental results During the experimental section, the author tried [1] paired initializations [2] disjoint development [3] holdout configurations and [4] method-specific hyperparameter tuning. The experiments also account for emitted shots and circuit evaluations. It is observed that across the 12 held-out single-photon graph, there is a consistent improvement.

3. Robust theories behind the procedure In the paper, it connects detector-count Fisher information, compressed moment readouts, and finite-count natural-gradient estimation within the proposed framework. The mathematical formulation and related assumptions seem clean and generally well-specified.

Critical weaknesses

Within my knowledge, I found the following weakness:

Unclear novelty I can understand that the construction applies classical Fisher information to the observed detector-count distribution. This is well-motivated from its mathematical structure as an application. However, the metric used follows directly from the stadnard Fisher-information definition for a (parametrized) categorical distribution. Similarly, the work also uses KL divergence and the covariance-whitened projection of moment information, which are also close to established results of information geometry. Therefore, it is clear to me what the authors proposed as new that is different from a direct application of classical neural gradient to the measured probability distribution.

The experimental evaluation is restricted to synthetic distribution-fitting problems. The primary experiments consider simple photonic circuits with fitting targets generated from the same parametrized famility of models. Additionally, the main results uses single-photon circuits with up to 32 parameters, while the two-photon extension contains only eight trainable parameters. I also did not find actual photonic hardware experiments. If such experiments are inaccesible, the authors should be able to state it explicitly.

Additionally, the paper layout seems weird to me. Some space between figure caption and main text seem unecessarily stretched.

Questions for authors

No.

2 comments

Thank you for the review, and for finding the problem formulation clear, the experimental design careful and the theory well specified. Your question on novelty is the one we thought about most, and we hope the revision answers it more directly than the submission did.

the work also uses KL divergence and the covariance-whitened projection of moment information, which are also close to established results of information geometry

We agree that the categorical Fisher, the KL divergence and the whitened projection are standard, and we use them on purpose. What we hope the paper adds is what they imply for a photonic learner. Three points can be checked against prior work.

First, the full metric needs no extra circuits and no detector calibration. Estimating the QFIM on hardware needs overlap or Hadamard-test circuits, QN-SPSA needs extra evaluations per step, and classical-Fisher and random-basis methods use measured statistics as a surrogate for the state metric. Meyer (Quantum, 2021) notes that gradient data already determine the classical Fisher matrix. We turn this observation into an optimizer, bound the dependent errors of the reuse jointly (Proposition 2), and show that reuse lowers final loss relative to splitting the same shots between gradient and metric (0.308 [0.271, 0.351] against a 25:75 split).

Second, we now show what the count metric adds over the state QFIM. For single-photon circuits we prove ζmin⁡Q⪯Fcount⪯ζmax⁡Q\zeta_{\min}Q\preceq F_{\rm count}\preceq\zeta_{\max}Q (new Proposition 1). The exact count Fisher and the exact QFIM therefore differ by a bounded detector factor, and under one controller they perform alike (0.93 [0.83, 1.06]). The gain comes from the estimator, and the estimator is tied to count space. Only the detector outcome law is observed, so only the count metric can be read from the same counts at no extra cost. Under the same controller, FockTrust lowers final loss by 36% relative to the exact QFIM on the 12 held-out graphs (0.639 [0.581, 0.706]). The gain remains when QQ receives a tuned spectral floor (0.692 [0.661, 0.727]), and FF still lowers final loss by 9% when QQ receives the same shrinkage, same-count noise and EMA as an oracle (0.909 [0.876, 0.957]), and every held-out graph favors FF. Loss-aware geometries (Gill and Pal, 2026) rescale the step from the objective, whereas here the detector sets the metric.

Third, the moment result holds for any categorical detector channel and readout. Moment geometry equals X⊤PAX⪯FcountX^\top P_A X\preceq F_{\rm count}, with an exact condition for equality, whereas prior rank laws (Ait Haddou, 2026) assume isotropic tangents and computational-basis readout.

The revised introduction states these points in its contribution paragraph. The new Table 1 compares prior methods by the space on which the metric is defined, the role of the detector, and whether gradient and metric come from one set of counts.

The experimental evaluation is restricted to synthetic distribution-fitting problems.

Thank you for pushing on this. The revision adds two experiments on held-out graphs with every method retuned, and a structured-metric study.

  • Larger circuits. At d=64d=64 and d=128d=128 on the recursive-tree graphs, where most phases stay active, the advantage grows with dimension. At d=128d=128, FF lowers final loss by about 45 to 65% relative to diagonal geometry (0.535), exact-QFIM NG (0.442) and QQ (0.351), with every interval below one (Appendix C.3).
  • Targets outside the model family. In a drifted-detector task that no parameter reproduces, and in a task whose target is known only from 2,048 measured samples, FF improves on diagonal geometry, exact-QFIM NG and Adam, with every interval below one (Section 5.4).
  • A diagonal-plus-rank-8 count metric that keeps the dense gain (Appendix C.2).

I also did not find actual photonic hardware experiments.

All results are simulations with known detector laws. Section 4 now opens with this statement, and the conclusion names hardware runs with drifting efficiencies as the next test.

Additionally, the paper layout seems weird to me.

Thank you for noticing. The stretched spacing came from float placement, and it is fixed in the revision.

We hope this addresses your concerns, and we would be glad to answer any further questions.

We thank you again for your review. We would be grateful to hear whether our reply and the revised paper address your two main concerns. Here is a short summary of what changed.

On novelty, we agree that the categorical Fisher and the KL divergence are standard. What the paper adds is how to use them for a photonic learner. First, Proposition 1 is new. It shows that for single-photon circuits the count Fisher is the state QFIM weighted by the detector efficiencies, so the two metrics differ by a bounded detector factor. Second, reading the metric from the gradient histograms only works with care. Without our shrinkage the loss is 40 to 500 times higher in the low-dimensional controls, and estimating the metric from separate counts is clearly worse than reusing the same counts. Proposition 2 bounds the error of this reuse. Third, under the same controller, FockTrust has 36% lower final loss than the exact QFIM on the 12 unseen graphs, and the new experiments show that this gain comes from the same-count estimate, which only count space allows. Table 1 compares these points with prior methods.

On the experiments, the revision adds runs at d=64d=64 and d=128d=128, where the advantage grows with dimension, two tasks whose targets the model cannot reproduce, and a sweep over 16, 64 and 256 shots per arm. FockTrust has lower final loss than diagonal geometry and the exact-QFIM optimizers in all of them. Section 4 now states that all experiments are simulations, and the float spacing is fixed.

If any point is still unclear, we would be glad to explain it further. Thank you for your time.

Reviewer_kwCZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper uses the classical Fisher information of the observed detector-outcome distribution as the natural-gradient metric for photonic circuits, in place of the state QFIM. Gradient and metric are estimated from the same parameter-shift histograms, with frequency shrinkage, EMA smoothing, and a detector-KL trust rule. The paper also relates covariance-whitened moment geometry to a projection of the full count Fisher and derives finite-count perturbation bounds.

The main simulations fit in-family detector distributions by KL on single-photon graph circuits up to d=32d=32, followed by a six-mode two-photon experiment. After method-specific tuning, full vs. diagonal count geometry gives 25.7% lower final loss on 12 held-out graph circuits. Full count vs. the exact-QFIM optimizer gives a final-loss ratio of 0.745.

Critical strengths
  1. The 2×22\times2 FF/FD/DF/DD design is a strong experiment. It separates the matrix used for the search direction from that used for trust scaling while holding scalar hyperparameters, shots, and circuit evaluations fixed. The direction contrasts favor the full count matrix under both trust rules. Method-specific tuning is performed on development cells and the main comparisons are then evaluated on held-out graphs, targets, and design seeds.

  2. The count metric reuses the base and parameter-shift histograms already collected for the gradient, so estimating this geometry does not require an additional circuit/sample stream. The paper also explicitly analyzes the statistical dependence induced by this reuse.

Critical weaknesses
  1. The main benchmark favours the proposed geometry by construction, and the comparison against QFIM does not isolate geometry.

The high-dimensional objective is KL(q⋆∥qθ)KL(q^\star\|q_\theta) for an exact in-family detector target. At a realizable optimum qθ=q⋆q_\theta=q^\star, the Hessian of this objective is exactly the detector-count Fisher, making this loss intrinsically aligned with the proposed metric. The two-photon comparison uses the same type of realizable detector-KL objective. The paper includes other low-dimensional control objectives, but I do not see a non-KL high-dimensional or cross-optimizer experiment.

More importantly, the exact-QFIM baseline uses a different update controller. FF is tuned over learning rate, damping, EMA, detector trust radius, and late trust contraction, whereas the QFIM optimizer is tuned over learning rate and damping and does not use the detector-trust schedule. Therefore the reported full-count/QFIM ratio does not establish that detector-count geometry itself is better than state-QFIM geometry; it compares two complete optimization procedures. This is especially important because the paper's central interpretation concerns the choice of geometry.

  1. The evidence for a practical or scalable optimizer advantage is limited. In the two-photon experiment Adam has the lower point estimate: FF/Adam =1.278=1.278, with marginal 95% CI [1.057,1.521][1.057,1.521] (the familywise interval includes one). Dense full geometry also requires O(d2)O(d^2) memory and an O(d3)O(d^3) solve, while the experiments stop at d=32d=32 and contain no hardware execution or model mismatch.

The low-dimensional controls further suggest that the benefit of full geometry is task-dependent. In Table 9, the reported full and diagonal median losses are equal after rounding for Born fitting (0.003 vs. 0.003), and are similarly small for high-frequency fitting (5.45×10−55.45\times10^{-5} vs. 5.15×10−55.15\times10^{-5}). The finite-count concentration bounds are also extremely conservative: for current Fisher error, Table 6 reports median observed error 3.627 vs. a median concentration envelope of 1.05×1051.05\times10^5. Thus the theory establishes coverage but does not currently provide a practically informative error certificate.

A further interpretability issue is that, with fixed shrinkage fraction γ\gamma, the implemented metric is not consistent for the exact count Fisher as the shot count increases. Proposition 1 controls the step around a population-shrunken EMA reference, and the later KL bound explicitly retains a shrinkage-discrepancy term Bγ,tB_{\gamma,t}. Since removing shrinkage substantially worsens some controls, it would be useful to know whether the reported advantages persist as γ\gamma is reduced with increasing shot count.

The full-vs-diagonal experiments also do not establish that a dense O(d3)O(d^3) metric is necessary. No intermediate structured approximation (e.g. block, sparse, or low-rank count geometry) is evaluated, although such an approximation could potentially retain the useful off-diagonal couplings at substantially lower computational cost.

Questions for authors

Q1. What performance does an exact-QFIM direction achieve when combined with the same detector-count trust rule, Euclidean cap/projection, and comparable step-control schedule as FF? This is necessary to separate the proposed direction geometry from the controller.

Q2. Does full count geometry still outperform exact-QFIM NG and Adam in a high-dimensional experiment whose objective is not detector-distribution KL, or under model mismatch, where the proposed Fisher is not locally identical to the objective Hessian at a realizable optimum?

5 comments

Thank you very much for the detailed and technically precise review, and for singling out the direction/trust factorial and the analysis of the reuse dependence. Your Q1 identified the comparison the submission was missing, and it led to the main addition of the revision.

Q1. What performance does an exact-QFIM direction achieve when combined with the same detector-count trust rule, Euclidean cap/projection, and comparable step-control schedule as FF?

We agree that the submitted QFIM comparison was between two complete procedures, and we ran the arm you describe together with a stronger one. Both receive the exact QFIM QQ as an uncharged oracle and keep the EMA, terminal contraction, cap and box of FF. With the count trust rule (QF), FF/QF = 0.145 [0.109, 0.201]. When QQ also sets its own trust region (QQ, tuned on three times as many candidates as FF), FF/QQ = 0.639 [0.581, 0.706] on the 12 held-out graphs.

To locate the gain, we replaced the metric of FF one step at a time and tuned every arm separately.

StepRatio
Exact count Fisher / QQ0.93 [0.83, 1.06]
Shrunken / exact count Fisher0.89 [0.83, 0.96]
FF (same-count estimate with EMA) / shrunken0.81 [0.69, 0.94]

The first row is what the new Proposition 1, ζmin⁡Q⪯Fcount⪯ζmax⁡Q\zeta_{\min}Q\preceq F_{\rm count}\preceq\zeta_{\max}Q, leads one to expect, and the gain comes from the shrunken same-count estimate. A tuned spectral floor on QQ does not close the gap (FF/regularized QQ = 0.692 [0.661, 0.727]). Even when the oracle QQ receives the shrinkage, same-count noise and EMA of FF, FF lowers final loss by 9% (0.909 [0.876, 0.957]), and every held-out graph favors FF. These two comparisons are paired with the efficiency sweep at mean efficiency 0.72, where FF/QQ = 0.597. The gain also holds at uniform efficiency, where Fcount=ζQF_{\rm count}=\zeta Q (FF/QQ = 0.611 [0.547, 0.686]). Section 5.3 and Table 3 are new.

Q2. Does full count geometry still outperform exact-QFIM NG and Adam in a high-dimensional experiment whose objective is not detector-distribution KL, or under model mismatch, where the proposed Fisher is not locally identical to the objective Hessian at a realizable optimum?

We agree that a realizable target aligns the objective Hessian with the count Fisher, and we added two d=32d=32 tasks where no parameter reproduces the target, with every method retuned.

TaskFF / DDFF / QFIM-NGFF / QQFF / Adam
Drifted detector, excess loss0.757 [0.675, 0.827]0.756 [0.624, 0.874]0.559 [0.368, 0.727]0.500 [0.422, 0.571]
Target from 2,048 samples0.724 [0.666, 0.778]0.764 [0.674, 0.841]0.600 [0.456, 0.692]0.485 [0.435, 0.532]

Both tasks are in Section 5.4 and Appendix C.1. The low-dimensional controls (Tables 11 to 14) also cover moment fitting and Helmholtz residuals.

No intermediate structured approximation (e.g. block, sparse, or low-rank count geometry) is evaluated

Thank you for this suggestion. A diagonal-plus-rank-8 count metric keeps the dense gain (0.772 against 0.799 for dense FF, both relative to diagonal), and the dense gain grows with dimension, with FF/DD = 0.755, 0.722 and 0.535 at d=32d=32, 64 and 128 (Appendix C.2 and C.3).

the implemented metric is not consistent for the exact count Fisher as the shot count increases

You are right for fixed γ\gamma. Any schedule γn→0\gamma_n\to0 gives a consistent estimator, and Bγ,tB_{\gamma,t} in Eq. (14) then vanishes, as Section 3.3 now states. The exact count Fisher, which is this limit, performs like QQ under the same controller, and at 64 shots per arm the fixed shrinkage contributes the factor 0.89 in the table above.

On the other points, the two-photon FF/Adam familywise interval includes one, and in the new tasks above FF roughly halves the loss of Adam. The concentration bounds are loose in magnitude, but the quantity the trust rule uses is sharp, with median exact-KL/Fisher-quadratic ratio 0.999.

We hope this addresses your concerns, and we would be glad to answer any further questions.

Thank you for your response and the experiments.

Remaining stuff:

  1. Novelty clarification (in other replies) is okay but still not clear on what really goes beyond past work which already observes that gradient measurements already determine the classical Fisher. Novelty overall is weaker.
  2. Estimate with EMA improves over noise-free shrunken metric (0.808 [0.692, 0.940]). Possible to separate the contributions of EMA, same-count estimation, and their interaction?
  3. Can we check and verify the dependence on the Gradient noise over trace rescaling and PDF projection? That last 9% gap is how much sensitive to this construction ?

Also can you share the final-loss comparisons for 16/64/256 shot budgets? We can also check if the gain is there as shots increase.

Rating stays the same.

Thank you for reading our replies carefully and for the follow-up questions. We ran all the experiments you asked for. Each arm was retuned on development cells and evaluated on held-out cells. The main finding is that the exact count Fisher and the exact QFIM give similar results, as Proposition 1 predicts, and that the gain comes from estimating the metric from the same counts as the gradient. This estimator exists only in count space, because the counts are what a photonic learner measures. The revised paper states the result in this form.

still not clear on what really goes beyond past work which already observes that gradient measurements already determine the classical Fisher

We agree that Meyer (2021) makes this observation. That paper is a review, and there the classical Fisher is used as a bound on, or a stand-in for, the QFIM. Our paper builds an optimizer on this observation, and this needs more than the observation itself. If the gradient histograms are put directly into the Fisher formula, rare outcomes make the metric singular. Without our shrinkage, the median loss is 40 to 500 times higher in the low-dimensional controls. If the metric is estimated from a separate set of counts, which keeps gradient and metric independent, the loss is 14 to 98 times higher, and a 25:75 split of the same shots gives 0.308 [0.271, 0.351] in favor of reuse. So reuse works only with the shrinkage and with an error bound that allows gradient and metric to be dependent (Proposition 2).

Possible to separate the contributions of EMA, same-count estimation, and their interaction?

Yes. On the 512 mechanism cells we ran the noise-free shrunken metric and the same-count estimate, each with and without EMA. All four arms were retuned on the same grid.

ContrastRatio [95% CI]
Same-count / noise-free, no EMA0.744 [0.656, 0.861]
Same-count / noise-free, with EMA0.842 [0.671, 1.039]
EMA / no EMA, noise-free metric0.978 [0.904, 1.067]
EMA / no EMA, same-count estimate1.107 [1.015, 1.215]

The gain comes from same-count estimation. It lowers final loss by about 26%, and every graph is below one. EMA gives no extra gain once the damping is retuned, and the interaction is close to one (1.13 [0.98, 1.30]). Section 5.3 and Appendix C.3 now report this result.

That last 9% gap is how much sensitive to this construction ?

Thank you for this question. We retuned two other constructions on the same held-out cells.

Construction of the estimator-matched QQFF / arm
Trace rescaling, eigenvalue clipping0.909 [0.876, 0.957]
No trace rescaling1.064 [1.017, 1.114]
Eigenvalue shift instead of clipping1.020 [0.976, 1.065]

Across the three constructions, FF / estimator-matched QQ lies between 0.91 and 1.06. Section 5.3 and Table 3 now give this range, and the introduction is revised to match. The state metric matches FF only when an exact-QQ oracle is given the same treatment as the count estimator.

Also can you share the final-loss comparisons for 16/64/256 shot budgets?

Every method was retuned at each budget, with γ=0.64/m\gamma=0.64/\sqrt{m}.

ShotsFF / DDFF / QFIM-NGFF / QQFF / estimator-matched QQ
160.675 [0.588, 0.777]0.640 [0.574, 0.730]0.658 [0.579, 0.751]0.933 [0.886, 0.988]
640.743 [0.702, 0.782]0.745 [0.685, 0.805]0.597 [0.545, 0.655]0.909 [0.876, 0.957]
2560.873 [0.833, 0.914]0.734 [0.690, 0.782]0.826 [0.778, 0.883]0.980 [0.944, 1.015]

At every budget, FF has lower final loss than diagonal geometry, exact-QFIM NG and QQ, and every interval is below one. The ratio to QFIM-NG stays between 0.64 and 0.75 as the number of shots grows. Note that the QFIM arms get the exact QQ for free. On hardware, the QFIM needs extra circuits. Section 5.3, Table 3 and Appendix C.3 now report these results. The 64-shot DD and QFIM-NG values are from Table 2.

To summarize the contribution: FockTrust uses the histograms that a photonic learner already records to build the full natural-gradient metric, with no extra circuits and no detector calibration. Proposition 1 shows that this metric is the state QFIM weighted by the detector efficiencies. Proposition 2 bounds the error from using the same counts for gradient and metric. Under the same controller, FockTrust has lower final loss than diagonal geometry and the exact QFIM at every shot budget we tested.

Thank you again for your questions, which have improved the paper. We are happy to answer any further questions.

Comments on the paper

1 comment

We sincerely thank the three reviewers for their time and for the careful reviews. We were glad that the reviewers found the physical motivation sound (pc1s, xA9J), the direction/trust factorial a strong experiment (kwCZ), the reuse of the gradient histograms with an explicit analysis of the dependence useful (kwCZ), and the theory clean and well specified (pc1s). We have uploaded a revised PDF, with the changes in blue.

The central question, raised most directly by kwCZ and in different words by pc1s and xA9J, is a fair one. The exact-QFIM baseline of the submission used its own step controller, so the comparison did not separate the metric from the controller. And if the count Fisher is the classical Fisher of the outcome law, what does it add over the state QFIM? Most of our effort during the discussion period went into these two questions.

Our answer has a theoretical part and an experimental part. For single-photon circuits with click efficiencies ζj\zeta_j, we now prove (Proposition 1, Section 3.2) ζmin⁡ Q⪯Fcount⪯ζmax⁡ Q,\zeta_{\min}\,Q\preceq F_{\rm count}\preceq\zeta_{\max}\,Q, where QQ is the state QFIM. The exact count Fisher is therefore the state metric weighted by what the detector resolves. We then placed the exact QQ inside the FockTrust controller and replaced the metric of FF one step at a time, tuning every arm separately (new Section 5.3).

Table R1 (Table 3 of the revision). Final-loss ratios at d=32d=32 with 95% CIs. Values below one favor FF.

Comparison under the same controllerRatio
FF / QQ, exact QQ for direction and trust region0.639 [0.581, 0.706]
FF / QQ at uniform detection efficiency0.611 [0.547, 0.686]
FF / QQ with a tuned spectral floor0.692 [0.661, 0.727]
FF / QQ with the shrinkage, same-count noise and EMA of FF0.909 [0.876, 0.957]
Exact count Fisher / QQ (mechanism cells)0.932 [0.831, 1.057]

The spectral-floor and estimator-matched rows are paired with the FF runs of the efficiency sweep at mean efficiency 0.72, where FF/QQ = 0.597.

The picture agrees with the proposition. The exact count Fisher and the exact QFIM perform alike, and the gain comes from the shrunken same-count estimate that FockTrust forms from histograms it already records. This estimator belongs to count space, because only the detector outcome law is observed. Even when we give the exact QQ the same treatment as an oracle, which hardware cannot provide, FF still lowers final loss by 9%, and every held-out graph favors FF.

Other changes in the revision:

  • Targets the model cannot reproduce (kwCZ, xA9J, pc1s). In a drifted-detector task and a task whose target is known only from 2,048 measured samples, FF improves on diagonal geometry, exact-QFIM NG, QQ and Adam, with every interval below one (Section 5.4, Appendix C.1).
  • Larger circuits (all reviewers). With every method retuned on the recursive-tree graphs, where most phases stay active, FF/QFIM-NG goes from 0.828 at d=32d=32 to 0.665 at d=64d=64 and 0.442 at d=128d=128, and FF/QQ from 0.659 to 0.628 and 0.351 (Appendix C.3).
  • Detection efficiency (xA9J, pc1s). With every method retuned, FF/QQ is 0.597, 0.591, 0.564 and 0.512 at mean efficiency 0.72, 0.60, 0.48 and 0.36 (Section 5.3).
  • Cost and structure (kwCZ, xA9J). Forming and solving the dense count system takes 0.3 ms per update at d=128d=128, and a diagonal-plus-rank-8 count metric keeps the dense gain (Appendix C.2).
  • Positioning (pc1s, xA9J). Table 1 compares prior natural-gradient methods, Section 4 states that all experiments are simulations and relates our sizes to a 380-phase integrated processor, and the float spacing is fixed. The former Proposition 1 is now Proposition 2.

Thank you again for the feedback, which has clearly sharpened the paper. We look forward to the discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.