acceptodds
Under review as a conference paper at ICLR 2027

How Much Fault Tolerance Does a QNN Need? Rethinking Precision Through Convergence Analysis

Abstract

The transition from NISQ to fault tolerance changes how a quantum neural network (QNN) is executed. Trainable rotations are no longer applied directly, but synthesized into sequences of logical gates at a chosen approximation error. Convergence behavior under such error is barely understood. As a result, the error QNN training can tolerate to converge, and the quantum resources that tolerance demands, are set by convention rather than by analysis. We derive a convergence-based theory of synthesis error that gives a principled way to set precision during fault-tolerant QNN training. From a biased-gradient convergence bound, we obtain the gradient distortion an optimizer can tolerate at a given convergence state. We translate this into closed-form tolerance laws for synthesis precision and magic-state fidelity. The laws cover both deterministic synthesis and mixed synthesis, which averages over compilations of the same rotation. We then prove a T-cost theorem for reaching a target stationarity under each strategy, with a matching lower bound for generic rotation angles. The two strategies affect the gradient differently. Deterministic synthesis introduces a first-order bias that persists during training, so the allowable error must shrink as convergence approaches. Mixed synthesis cancels this bias in expectation, leaving only a second-order residual: the error becomes gradient variance, and the same target is reached with fewer T-gates. Experiments with explicit Clifford+T synthesis and end-to-end training on three datasets confirm these predictions: mixed synthesis roughly halves the T cost per circuit relative to convergence-scheduled deterministic synthesis without loss of accuracy, and scheduling magic-state fidelity reduces distillation work by up to . These results show that the fault-tolerant precision of QNN training can be derived from learning rather than fixed by convention.

Primary area:
learning theory
Keywords:
fault-tolerant quantum computing, quantum neural network, convergence analysis

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_8U2bOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks how much rotation synthesis precision is actually needed to train a quantum neural network (QNN) on fault-tolerant hardware. In this setting each parameterized rotation is compiled to a Clifford+T sequence at some tolerance, and every T gate consumes a distilled magic state. Building on biased SGD theory (Ajalloeian & Stich, 2020), the authors require that the gradient bias at each update stay below a fixed fraction of the larger of the current gradient norm and a scheduled floor, and that the synthesis variance stay within a per-step budget. From this they derive tolerance laws for two regimes. Deterministic Ross-Selinger synthesis has gradient bias that is first order in the tolerance (Lemma 1), so the deterministic tolerance scales linearly with the target. "Three-axis" mixed synthesis solves an LP over a set of offset candidates and combines it with S-conjugation to center the full SU(2) error generator. This pushes the bias down to second order in the tolerance and turns the first-order term into variance (Lemma 2), so the mixed tolerance scales with the square root of the target until the variance cap binds (Proposition 1). Theorem 1 bounds the per-rotation T cost at the convergence floor and shows that mixed synthesis has an asymptotic factor-of-two advantage. Theorem 2 gives a stationarity guarantee for SGD under these budgets and extends it to momentum and Adam. The same bias budget is applied to magic-state infidelity, and the resulting time-varying fidelity requirement is replayed through a simple 15-to-1 distillation server model. Experiments use exact Clifford+T statevector simulation of 4 to 8 qubit QNNs on digits (0 vs. 1), UNSW-NB15 and Waveform. Mixed synthesis shows the predicted bias exponents (0.99 for deterministic, 1.98 for mixed), and at an equal number of updates it uses about 50% to 56% fewer T gates per circuit than scheduled deterministic synthesis. Fidelity scheduling lowers normalized factory work by 11% to 28%.

Critical strengths
  1. The question is timely and well posed, and it connects two research communities. Asking how much synthesis precision learning actually needs, rather than how cheaply a given precision can be reached, is a useful way to look at QML in the fault-tolerant era. The paper follows this question consistently, from gradient error (RQ1) to tolerance (RQ2), T cost (RQ3), and magic-state fidelity and provisioning (RQ4-5). The comparison with Sayginel et al. (2024), Kang et al. (2026) and the mixed synthesis literature (Appendix G, Table 23) is fair and accurate.
  2. The error model goes beyond angle-only noise, and the matching ablation is convincing. The paper models the full three-dimensional error generator of each gate (Section 3.1) and shows that centering only the tangential component still leaves first-order bias, because about 59% of the error lies on the transverse axes (Fig. 2a/b, Table 12, Fig. 7a). This is a concrete finding, and it is a useful caution for the angle-perturbation noise models common in QNN robustness work (Wang et al., 2022; Rabinovich et al., 2024). The LP plus S-conjugation construction (Appendix F.2) is simple, and its exactness can be checked: the residual mean generator is at floating-point round-off level relative to the tolerance.
  3. The predicted orders and variance laws are verified carefully. On all three datasets the fitted exponents agree with Lemmas 1-2: 0.99 (standard deviation 0.02) for deterministic, 1.98 (standard deviation 0.01) for mixed, and 0.997 for magic-state bias (Table 14). The trace of the measured covariance matches the predicted sum over gates of each generator covariance mapped through that gate's response matrix to within 0.1%, and the per-shot law (variance falling inversely with the number of shots) holds to within about 2% (Fig. 2c). The predictions were fixed before calibration (Table 16), which is good practice.
  4. The proofs I checked are clear, and the authors are candid about the limits of their results. I checked Lemma 3 (including the per-shot variance comparison, Eqs. 16-28), Lemmas 1-2 (Eqs. 33-46), Proposition 1, Theorem 1 (Eqs. 53-68) and the SGD part of Theorem 2 (Eqs. 72-89), and found no errors. The paper is also explicit about what it does not cover: the subtlety in the objective in Appendix B.1 (Eq. 29), the fact that the factor-of-two advantage is only asymptotic (the remark after Corollary 1), the path-length correction factor in Corollary 2, the oracle monitor, and the replay-only baselines.
  5. Synthesis is simulated exactly. Every circuit runs the sequences that pygridsynth actually compiles rather than a noise surrogate (Appendix F.5), and magic-state noise uses the exact injection channel. This makes the T counts and bias measurements at the scales studied credible.
Critical weaknesses
  1. The theory does not guarantee the reported savings, and the experiments run outside its assumptions. Theorem 2 and Corollary 2 are for SGD with an oracle-scheduled floor that is strictly positive and proportional to the target gap. The experiments instead use Adam with step decay, a gradient floor of zero, a hard tolerance floor of one in a million, and an empirical variance rule in place of the per-step variance budget (Section 3.1, Appendix F.4). More importantly, the Adam variance rule lets the synthesis perturbation of an update be as large as the intrinsic noise (FQ), or lets the total reach twice the noise (PS). This corresponds roughly to allowing synthesis variance equal to the intrinsic noise variance, i.e. a variance ratio of about one in Corollary 2, and at that value the certified cumulative ratio from Corollary 2 is close to one, which means there is no certified cumulative saving. So the headline "about half" saving is an empirical per-circuit saving at an equal number of updates, not a consequence of the theory. Please state which variance ratio the experimental variance rule corresponds to, run at least one SGD experiment that follows the schedule of Theorem 2, and report how far the realized cumulative ratio is from the certified one.
  2. A cheap baseline is missing: deterministic synthesis at a fixed loose tolerance. FH is fixed after the fact at the tightest tolerance used anywhere (0.000003), so it is a weak comparison. Fixed loose deterministic synthesis (tolerance 0.003) is listed, but only as a cost replay (Appendix C.1); its training quality is never reported. The paper's own Table 18 suggests this baseline may be quite competitive. Under a cap of 17 million T gates per B32 round, deterministic synthesis at a fixed tolerance of 0.006 trains stably and reaches a held-out MSE of 0.0976, close to scheduled deterministic synthesis (0.0959, Table 1) and FQ (0.0964). FQ does reach 0.0870 under the same cap, but that gap is within the seed-to-seed standard deviation of about 0.019. By my estimate, 17 million T gates per round is about 2.45 thousand T gates per circuit, roughly the same as the cost of FQ (2.44 thousand). On these easy tasks, a well-chosen fixed loose tolerance may match FT-QNN at the same cost. Please actually train (not replay) the fixed loose deterministic setting at several tolerances, and show where it fails to reach the accuracy or gradient targets while FT-QNN succeeds. Otherwise the claim could be restated as "matches a tuned fixed tolerance without tuning".
  3. The tasks are too small and too easy to show whether precision matters. Digits 0 vs. 1 reaches about 99.6% to 100% accuracy under every setting. UNSW-NB15 and Waveform use only 128 training samples (about 0.12% of UNSW-NB15). The standard deviation of held-out MSE often exceeds the mean (for example 8.07 with a standard deviation of 10.20 in Table 1), so with only three seeds "without loss of accuracy" cannot be decided statistically. The gradient targets are also rarely reached: no SD or FQ seed gets the gradient norm down to 0.005, and only 1 of 24 runs reaches 0.002 (Table 17). The cumulative cost claim in Contribution 2 therefore rests only on the 0.01 target. It would help to add harder tasks where precision clearly affects the outcome (multiclass, regression, or tasks on which the ideal model does not saturate), to give confidence intervals or paired tests, and to use more seeds for the cumulative time-to-target comparison.
  4. The method depends on several constants and an oracle monitor whose costs are not counted. At every update the controller needs the exact full-data gradient norm, and it also needs the first- and second-order bias constants (75 and 35) and the magic-state response constant (2571.3), all calibrated in advance from classical finite-difference response tensors (Appendix C.4, F.4). None of these are available on large hardware, and the T or shot cost of estimating the gradient norm accurately enough is not included in the T counts. In addition, the first-order bias constant grows roughly like a constant times the sum of the response-matrix norms over gates, so it grows with the number of rotations, while the variance constant grows like the sum of the squared norms, which is a different rate (Eqs. 34, 46). Table 2 hints at some degradation at 8q/9L, yet how the trade-off between deterministic and mixed synthesis changes with circuit size is never analyzed. Please analyze or measure how the first-order bias, second-order bias and variance constants scale with the number of rotations, evaluate a noisy shot-based monitor and count its cost, and show how robust the method is when the constants are off by a factor of 2 to 10.
Questions for authors
  1. In the reported runs, what effective variance ratio from Corollary 2 (and what value of the related variance-budget parameter) does the empirical Adam variance rule correspond to? Given Corollary 2, do you agree that when this ratio is about one the theory guarantees no cumulative advantage? If not, why not?
  2. Can you report the fixed loose deterministic setting actually trained (not replayed) at tolerances of 0.003, 0.006 and 0.001, including held-out MSE and accuracy, whether the gradient targets are reached, and T gates per circuit? Given Table 18, is there any setting in which a fixed loose deterministic tolerance fails while FQ succeeds?
  3. Why is per-shot (PS) mixed synthesis more expensive than fixed-query (3.01 thousand vs. 2.44 thousand T gates per circuit at B32/S500), and why is its predicted variance cap lower (Table 16: 0.00743 vs. 0.0105 at S500)? Lemma 3(c) says PS divides the synthesis variance by the number of shots, so I would expect PS to allow a looser tolerance. Is this due to the different acceptance rules ("synthesis no larger than the noise" vs. "total no larger than twice the noise")?
  4. The 36 calibration tests only check that the predicted tolerance is safe, which is a one-sided test. At update 368, with the second-order bias constant of 35 and a bias fraction of 0.1, the predicted bias bound is about 0.0014, while the calibrated interval is 0.0024 to 0.0027. How tight are the predictions in general? Would an overly conservative predictor also pass all 36 tests?
  5. If the gradient-norm monitor were estimated with shots on FT hardware, how would its cost compare with the T savings? How does the scheme behave with a noisy monitor (for example, simply using the same mini-batch gradient used for the update)?
  6. On the 4q-8q and 4L-9L models in Table 2, how do the first-order bias, second-order bias and variance constants (or the empirical calibrated constants) scale with qubits times layers? Is the higher held-out MSE of FQ at 8q/9L (0.1005 vs. 0.0845) an early sign that the variance cap is harder to satisfy at larger scale?
  7. Can you estimate the non-cancellation constant in Assumption 2 on your circuits? Is there a deterministic strategy, such as alternating between different compilations across parameter-shift queries, or angle-dependent choices, that could partly cancel the first-order bias without introducing randomness and so weaken the deterministic lower bound?
  8. The Waveform class pair was chosen "after results on another pair" (Appendix D.3). What were the results on the original pair?
6 comments
AuthorsOpenReview ↗

W1 / Q1, effective variance ratio and SGD run; W2 / Q2, trained fixed baselines

We thank the reviewer for the detailed review and highly encourage the reviewer to read the revised manuscript with all concerns addressed.

1) W1 / Q1: Effective variance ratio, SGD run, and certified versus realized cost.

The conditional is correct: at λ=1\lambda=1, Corollary 2 certifies no cumulative advantage. The premise, however, is not. The experimental rule does not correspond to λ≈1\lambda\approx1.

The variance rule is a ceiling, not the admitted variance. The Adam rule admits a mixed tolerance only if the synthesis perturbation is at most the intrinsic noise, so λ=1\lambda=1 is an upper limit. We measured the variance the rule actually admits. At updates 100, 400, and 800 of every FQ run, the admitted tolerance is ϵ=10−2\epsilon=10^{-2}, and the ratio of synthesis gradient variance to intrinsic shot and mini-batch variance is λeff=0.012\lambda_{\mathrm{eff}}=0.012–0.0190.019 (mean 0.015) with Adam and 0.0110.011–0.0200.020 (mean 0.015) with SGD. The admitted synthesis variance is about 1.5% of the intrinsic noise, so the certified cumulative ratio of Corollary 2 is 2/(1+λeff)≈1.972/(1+\lambda_{\mathrm{eff}})\approx1.97, not 1.

SGD run following Theorem 2. We trained SD and FQ with plain SGD under the schedule of Theorem 2: bias budget ρ=0.1\rho=0.1, floor gΔ=ϕΔg_\Delta=\phi\Delta, and one constant step size shared by both arms, on the main circuit (6q/9L, B32/S500, three paired seeds). The tuned step ηΔ∝Δ2\eta_\Delta\propto\Delta^2 in Theorem 2 is a worst-case certificate, so we use a practical constant step and compare at matched update 1,000 and at ∥∇L∥=0.04\Vert\nabla L\Vert=0.04, which every run reaches.

Certified versus realized cumulative cost ratio (SD/FQ), same setting. Mean ± SD.

Optimizerλeff\lambda_{\mathrm{eff}}Certified 2/(1+λeff)2/(1+\lambda_{\mathrm{eff}})Realized ratioFQ saving (%)
Adam, update 8000.0151.972.0952.2±1.3
Adam, stability0.0151.972.1152.5±1.1
SGD (Theorem 2), update 1,0000.0151.972.05±0.1051.2±2.3
SGD (Theorem 2), ∥∇L∥=0.04\Vert\nabla L\Vert=0.040.0151.972.24±0.5153.9±9.6

Under the schedule of Theorem 2, FQ saves 51.2±2.3% cumulative T at matched updates and 53.9±9.6% at a common gradient target, matching the 52% obtained with Adam. The realized ratio of 2.05–2.24 agrees with the certified ratio of 1.97. Corollary 2 is an asymptotic ratio of certified bounds at the floor, while the realized ratio also includes cheaper updates above it. The extra synthesis variance does not lengthen training: FQ reaches the target in 419±155 updates, against 465±173 for SD. The saving therefore does not depend on Adam, and its size agrees with the certified ratio once the admitted variance is measured. The Adam condition κ≤50\kappa\le50 of Theorem 2 also holds (maximum 25.6).

2) W2 / Q2: Trained fixed loose deterministic baselines.

We trained fixed deterministic synthesis at ϵ=10−3\epsilon=10^{-3}, 3×10−33\times10^{-3}, and 6×10−36\times10^{-3} on the main circuit (6q/9L, B32/S500, three paired seeds), under the same pipeline and stopping rule as SD and FQ. These are full training runs, not cost replays, as in our W2 / Q2 response to Reviewer bVyJ, which also adds a fixed mixed baseline and the tolerance-search cost.

Trained fixed deterministic baselines, same setting. Mean ± SD; held-out accuracy is 99.6±0.4%99.6\pm0.4\% for every arm at update 800.

ArmT/circuitCum. T at 800 (×1012\times10^{12})Seeds reaching 0.02/0.01/0.005Stable seedsHeld-out MSE (×10−2\times10^{-2})
FQ (ours)2,324±386.57±0.073/3/13/37.91±0.73
Fixed det., 10−310^{-3}3,260±109.06±0.053/3/13/39.58±3.12
Fixed det., 3×10−33\times10^{-3}2,661±347.35±0.073/1/03/38.29±1.24
Fixed det., 6×10−36\times10^{-3}2,310±586.35±0.011/1/01/38.61±1.02

Yes, fixed loose tolerances fail where FQ succeeds. At ϵ=6×10−3\epsilon=6\times10^{-3}, which matches FQ's cost per circuit and corresponds to the 17M-per-round ceiling the reviewer estimates from Table 18, two of three seeds fail to stabilize within 1,500 updates, and only one reaches ∥∇L∥≤0.01\Vert\nabla L\Vert\le0.01. At 3×10−33\times10^{-3}, the runs stabilize, but two of three seeds never reach ∥∇L∥=0.01\Vert\nabla L\Vert=0.01, which FQ reaches in every seed. A fixed loose tolerance stalls at a higher gradient floor, exactly the first-order bias floor of Lemma 1. Held-out MSE and a plateau-based stopping rule cannot reveal this, which is why Table 18 appears favorable to loose tolerances. Relative to FQ, cumulative T at update 800 is +37.9±1.6% at 10−310^{-3}, +11.9±1.2% at 3×10−33\times10^{-3}, and −3.4±0.9% at 6×10−36\times10^{-3}. The only fixed tolerance that matches FQ on every gradient target is 10−310^{-3}, which costs 37.9% more. Finding a working fixed tolerance also requires trial runs (Prop 1), costing 22.122.1–34.6×101234.6\times10^{12} T gates against 10.4×101210.4\times10^{12} for one FT-QNN run. FT-QNN therefore reaches targets that the cheaper fixed tolerances miss, and is cheaper than the fixed tolerance that reaches them, with no tuning.

AuthorsOpenReview ↗

W3, task difficulty and statistics; W4 / Q5 / Q6, monitor, constants, and robustness

3) W3: Task difficulty, stringent targets, and statistics.

Harder tasks and more data. We trained SD and FQ on UNSW-NB15 with 8×8\times more training data at 6, 10, and 12 qubits, and on a three-class UNSW-NB15 task with 96×96\times more data (10q/5L), where held-out accuracy is 63% and does not saturate. FQ saves 56.1–56.8% T per circuit at every width and 52.9% on the three-class task, with comparable held-out MSE and accuracy.

Stringent targets. We trained ideal training with exact rotations, and FH end to end, on the main circuit. No arm reaches ∥∇L∥≤0.003\Vert\nabla L\Vert\le0.003, and ideal training has the highest gradient floor (0.0078±0.00120.0078\pm0.0012). The rarely reached targets are set by shot and mini-batch noise, not by synthesis error, since removing synthesis error entirely does not lower the floor.

Seeds, paired tests, and confidence intervals. We extended SD and FQ to five paired seeds on the main circuit and on UNSW-NB15 and Waveform. Paired comparisons remove the seed-to-seed variation that dominates the unpaired standard deviations.

Paired comparisons over five seeds. Saving: cumulative T of FQ relative to SD. Differences are FQ minus SD, with 95% confidence intervals. At ∥∇L∥=0.01\Vert\nabla L\Vert=0.01, MSE and accuracy are evaluated at stability.

TaskPointSaving (95% CI)MSE difference (95% CI)Accuracy difference, pts (95% CI)
Digits 6q/9LUpdate 80052.2% [51.0, 53.3], p<10−6p<10^{-6}−0.015 [−0.044, +0.013], p=0.21p=0.21+0.25 [−0.44, +0.94], p=0.37p=0.37
Digits 6q/9L∥∇L∥=0.01\Vert\nabla L\Vert=0.0157.8% [47.3, 68.2], p=0.002p=0.002−0.015 [−0.045, +0.014], p=0.22p=0.22+0.25 [−0.44, +0.94], p=0.37p=0.37
UNSW-NB15 6q/4LUpdate 50048.6% [45.6, 51.5], p<10−5p<10^{-5}+0.008 [−0.015, +0.030], p=0.39p=0.39−1.3 [−5.7, +3.2], p=0.48p=0.48
Waveform 6q/4LUpdate 50051.6% [49.3, 54.0], p<10−5p<10^{-5}+0.012 [−0.017, +0.042], p=0.30p=0.30−1.3 [−4.2, +1.7], p=0.31p=0.31

The cumulative saving is statistically significant with tight intervals on every task, and no task shows a significant difference in held-out MSE or accuracy. The cumulative claim now rests on matched updates, a common gradient target, three datasets, and the Theorem 2 schedule with SGD (W1), not on the 0.01 target alone.

4) W4 / Q5 / Q6: Monitor cost, constants, and scaling.

Noisy monitor. We evaluated exactly the monitor the reviewer suggests: the mini-batch gradient of the preceding update, together with a split-batch estimate. Both cost zero additional circuits, shots, or T gates, since they reuse queries already executed for training. FQ retains a 47–50% cumulative saving against SD, and updates to stability are unchanged. Theorem 2 now covers estimated monitors with bounded relative error, and a new proposition gives a direct SGD bound for the lagged mini-batch monitor.

Scaling of the constants. We measured the constants at update 500 of the SD runs at ϵ=10−3\epsilon=10^{-3}.

Measured response constants versus number of synthesized trainable rotations NN (digits, B32/S500). 6q/9L: mean ± SD over three seeds; other models: one saved seed.

ModelNN∥bdet⁡∥/ϵ\Vert b_{\det}\Vert/\epsilon∥bmix∥/ϵ2\Vert b_{\mathrm{mix}}\Vert/\epsilon^2tr ΣFT/ϵ2\mathrm{tr}\,\Sigma_{\mathrm{FT}}/\epsilon^2γ\gamma
6q/4L486.43.0173.3
4q/9L725.03.4232.5
6q/6L726.87.2543.4
6q/9L1086.3±0.85.8±1.039±183.2±0.4
8q/9L14410.39.5685.1

The bias constants grow at most linearly in NN: over a 3×3\times increase in rotations, the deterministic bias constant grows by 1.6×1.6\times and the mixed bias constant by 3×3\times, so the deterministic bias does not accumulate as a sum of response norms. The variance constant grows by 4×4\times, which tightens the variance cap by only 4=2×\sqrt4=2\times in tolerance, since the cap scales as Λv−1/2\Lambda_v^{-1/2}. The quantity that sets the deterministic-to-mixed tolerance ratio, Λ1/Λ2\Lambda_1/\sqrt{\Lambda_2}, stays between 2.5 and 3.7 for every model without a downward trend, consistent with the unchanged saving at 12 qubits. On Q6: at 8q/9L, the variance constant is 1.8×1.8\times that of 6q/9L, which tightens the variance cap by only 1.8≈1.3×\sqrt{1.8}\approx1.3\times. The higher FQ MSE at 8q/9L (0.1005 vs. 0.0845) lies within one seed SD, and the 12-qubit runs show no degradation.

Robustness to misspecified constants. We trained SD and FQ with the bias constants scaled by 0.1 and by 10 (main circuit, three paired seeds).

Controller constants scaled by 0.1 and 10, same setting. Mean over seeds.

ArmScaleCum. T at 800 (×1012\times10^{12})Seeds reaching 0.01Outcome
FQ0.16.423/3Stable, 3/3
FQ108.253/3Stable, 3/3
SD0.110.383/3Stable, 3/3
SD1016.852/3Infeasible, 3/3

FQ is robust to a 10×10\times error in either direction: every seed stabilizes and reaches ∥∇L∥=0.01\Vert\nabla L\Vert=0.01, with unchanged held-out MSE. Underestimating its constant changes the cost by −2%, and overestimating it by +26%.

AuthorsOpenReview ↗

W4 (cont.); Q3, FQ vs. PS; Q4, tightness; Q7, non-cancellation; Q8, Waveform pair

Deterministic synthesis is not: with AA overestimated 10×10\times, the required tolerance falls below the 10−610^{-6} grid in all three seeds. This is the ϵ∝Δ\epsilon\propto\Delta requirement of Theorem 1 in practice, while mixed synthesis keeps a Δ\sqrt\Delta margin. At matched scaling, FQ saves 38.1±2.9%38.1\pm2.9\% (×0.1) and 51.1±0.4%51.1\pm0.4\% (×10). Even with FQ's constant overestimated 10×10\times and SD's underestimated 10×10\times, FQ uses 20% less cumulative T. The calibrated constants A=75A=75 and B2=35B_2=35 are 6–12× the measured values above, so they are conservative by design, and this conservatism costs little. RMSR_{\mathrm{MS}} is likewise taken as the maximum over checkpoints (Table 22).

5) Q3: Fixed-query versus per-shot synthesis.

The reviewer correctly notes that Lemma 3(c) gives per-shot synthesis a 1/S1/S reduction in the leading synthesis variance at equal tolerance. The observed T-cost ordering does not contradict this result. In the reported B32/S500 experiments, FQ uses a tolerance ceiling of 10−210^{-2}, whereas PS is restricted to 3×10−33\times10^{-3}. Their predicted variance caps are 0.0105 and 0.00743, respectively, under different empirical Adam acceptance rules. Consequently, PS operates with a substantially tighter tolerance, which explains its higher T cost. The FQ–PS comparison is therefore policy-specific and does not establish that fixed-query synthesis is intrinsically cheaper than per-shot synthesis. The difference between the predicted variance caps reflects the distinct empirical acceptance tests rather than the common variance budget of Lemma 3(c). The caps do not enter the SD–FQ comparison: with A=75A=75 and ρ=0.1\rho=0.1, the deterministic tolerance reaches 3×10−33\times10^{-3} only for ∥∇L∥>2.25\Vert\nabla L\Vert>2.25, while the largest gradient norm in training is 0.68, so SD never reaches its cap, and the principal T-saving result is unaffected. We clarify these policy-specific ceilings and the scope of the comparison in the revised manuscript.

6) Q4: Tightness of the calibrated predictions.

The 36 tests check safety, which is the property the controller needs: a predicted tolerance must never violate the budget. We additionally quantify tightness. At update 368, the predicted bias-limited tolerance is 0.0014 and the calibrated bracket is [0.00237, 0.00274], so the prediction is within a factor of 1.7–2.0 of the calibrated limit. This matches the measured constants: B2=35B_2=35 is about 6×6\times the measured mixed bias constant, which gives 6≈2.4×\sqrt6\approx2.4\times in tolerance. At 10.5 T gates per decade, a factor of 1.85 costs about 2.8 T gates per rotation. At updates 0 and 200, the calibrated caps exceed 10−210^{-2} and are upper-censored, so the predictions there are safe by a margin that cannot be resolved further. The robustness study gives the practical answer: a 10×10\times looser constant changes FQ's cost by only −2%, since the tolerance is mostly at its cap, and a 10×10\times more conservative one by +26%, with every seed still reaching the 0.01 target.

7) Q7: Non-cancellation constant and deterministic cancellation.

We estimated γ\gamma directly: γ=∥bdet⁡∥/max⁡j∥aj∥\gamma=\Vert b_{\det}\Vert/\max_j\Vert a_j\Vert ranges from 2.5 to 5.1 across all five models (constants table in W4), bounded well away from zero. Assumption 2 holds on our circuits, and the deterministic errors do not cancel across gates. A fixed pairing of compilations within one update cannot remove the first-order term for generic angles, since the base and shifted queries enter the estimator through different response matrices, and the errors do not align with them. Alternating compilations across updates is a derandomized form of the mixture: it centers the error on average over time but not at each update, so it does not reduce the per-update bias that sets the tolerance. The lower bound of Theorem 1 applies to every schedule with one fixed compilation per rotation instance, and three-axis mixing is the construction that removes the first-order term at each update.

8) Q8: Original Waveform class pair.

The original pair, classes 0 vs. 1, was first run as an exploratory single-seed study, in which FQ had lower held-out accuracy (82.0% vs. 87.5%). We have now trained two further seeds under the identical protocol. Over three paired seeds, FQ saves 55.5±0.3%55.5\pm0.3\% T per circuit, and the held-out MSE is the same for both arms (0.458 for FQ and 0.457 for SD). The accuracy difference is −1.8 points and not significant (p=0.43p=0.43): it varies in sign across seeds (−5.5, −0.8, +0.8 points). The accuracy gap in the original run was therefore specific to that seed. The saving on this pair matches that on classes 1 vs. 2 (55.8% per circuit over five seeds), so the choice of class pair does not affect the result.

All new results are added to Appendices D.2 and D.3.

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

The paper studies how gate synthesis errors affect the training of fault-tolerant quantum neural networks. The authors analyze how deterministic and mixed synthesis introduce different levels of gradient bias and variance, then use these results to derive precision requirements and bounds on T-gate costs during optimization. They also propose scheduling magic-state fidelity based on the convergence state. Experiments using explicit Clifford+T synthesis on small QNNs test the predicted error scaling and compare resource costs across synthesis methods, circuit sizes, and datasets. The results show interesting reductions in T-gate consumption with similar predictive performance.

Critical strengths

The authors put a lot of effort into the mathematical analysis, clearly stating their assumptions and examining how compilation errors affect training.

S1. The T-gate savings relative to scheduled deterministic synthesis hold across several datasets, circuit sizes, and shot budgets. The authors also compare cumulative T consumption at matched gradient targets, where mixed synthesis retains a substantial saving for the targets reached by both methods.

S2. The treatment of compilation error includes all three rotation directions, rather than assuming that the error only changes the intended rotation angle. The experiments show that cancelling the angle error alone still leaves a first-order error. Accounting for the other components is therefore important for obtaining the quadratic bias reduction that the method relies on.

Critical weaknesses

W1. The controller requires the exact full-data gradient norm at every update, which is available in these simulations but would need to be estimated in practice. The reported T-gate savings do not account for obtaining this information. Since the experiments only cover 4–8 logical qubits, it remains unclear how much of the saving would survive the cost of running the controller on larger problems.

W2. The fixed-precision mixed baseline is evaluated through cost replays along an existing training trajectory, rather than separate training runs. This leaves an important comparison unresolved. Mixed synthesis already reduces the leading-order compilation bias, so I would like to know how much the proposed scheduling adds over a well-chosen fixed precision. The replay cannot establish whether that baseline would reach comparable accuracy at a similar total cost.

In addition to these, I’m also curious whether the same picture holds for the deterministic baseline. Fixed-loose deterministic synthesis (ε = 3×10⁻³) is also evaluated only through cost replay, and Table 18 suggests that loose deterministic tolerances can be sufficient at the budgets examined. Training this baseline to completion would help show how strict the deterministic schedule really needs to be.

Questions for authors

Q1 (Suggestion). The controller uses the exact full-data gradient norm at every update. An evaluation using noisy gradient estimates already available during training, possibly with periodic updates to the precision schedule, would help establish whether the reported savings survive the cost of monitoring convergence. It would be useful to compare both convergence and cumulative T consumption.

Q2 (Suggestion). The fixed-precision mixed baseline (ε = 3×10⁻⁴, a tolerance tighter than what FQ uses for much of finite-shot training) and the fixed-loose deterministic baseline (ε = 3×10⁻³) are evaluated only through cost replays. Table 18 also suggests that relatively loose deterministic tolerances can be sufficient at the budgets examined. Separately training these fixed-precision baselines and comparing cumulative T consumption at matched accuracy or gradient targets would help establish how much of the reported savings comes from scheduling rather than mixing, and how much the deterministic schedule actually needs to tighten.

Q3 (Clarification). The paper notes that the empirical Adam variance rule is not covered by Theorem 2. Does the convergence guarantee still apply to the preconditioner used in the experiments? Could the authors report the observed preconditioner range and explain how informative the resulting bound is at β₂ = 0.999?

5 comments
AuthorsOpenReview ↗

W1 / Q1, gradient monitors and scale; W2 / Q2, trained fixed-precision baselines

We thank the reviewer for the careful reading and for the constructive suggestions . We substantially address each point in detail and we encourage the reviewer to read the revised manuscript, where every change made in response to the reviews is highlighted in blue.

1) W1 / Q1: Oracle gradient monitor and scale.

The oracle monitor was a deliberate scoping choice to isolate the synthesis mechanism. We now evaluate the practical setting the reviewer suggests: noisy gradient estimates already available during training. For more details, including the full theoretical statements and monitor definitions, please refer to the revised main paper (Section 4 and RQ3 in Section 5.2, highlighted in blue) and W1 / Q1 response to Reviewer pQU6.

Monitors at zero additional quantum cost. We replace the oracle with two estimates computed from the gradient queries of the preceding update: the mini-batch gradient norm, and a split-batch estimate from two half-batch gradients. Both require no additional quantum circuits, shots, or T gates. The schedule is updated at every step using these estimates, with only classical monitoring overhead.

Oracle vs. estimated monitors (Digits, 6q/9L, B32/S500). Mean ± SD over three paired seeds.

PointMonitorCumulative T, SD (×1012\times10^{12})Cumulative T, FQ (×1012\times10^{12})FQ saving (%)
Update 800Oracle13.74 ± 0.246.57 ± 0.0752.2 ± 1.3
Update 800Split-batch12.78 ± 0.316.45 ± 0.0649.5 ± 1.7
Update 800Mini-batch12.32 ± 0.046.41 ± 0.0648.0 ± 0.6
StabilityOracle21.82 ± 0.3910.36 ± 0.4152.5 ± 1.1
StabilitySplit-batch19.45 ± 0.639.85 ± 0.3149.4 ± 0.9
StabilityMini-batch18.49 ± 0.979.77 ± 0.3747.1 ± 2.1

FQ retains a 47 to 50% cumulative saving with estimated monitors, within a few percentage points of the oracle's 52%. Training duration remains comparable: FQ reaches stability after 1223 to 1225 updates across monitors, with comparable held-out MSE. On the theory side, Theorem 2 now covers estimated monitors with bounded relative error, and a new proposition establishes a direct SGD bound for the lagged mini-batch monitor without that assumption.

Scale. We trained SD and FQ on UNSW-NB15 with 8×8\times more training data at 6, 10, and 12 qubits. FQ saves 56.1 to 56.8% T per circuit and 49.9 to 51.5% cumulative T across these widths. Held-out MSE remains comparable, with accuracy differences of at most two percentage points. On a three-class task with 96×96\times more data (10q/5L), FQ saves 52.9% per circuit and 50.3% cumulative T, with comparable held-out MSE and accuracy. The cumulative saving remains approximately 50% across the tested circuit widths, dataset sizes, and binary and multiclass tasks.

Width, data, and task scaling on UNSW-NB15 (B32/S500). Mean ± SD over three paired seeds.

Model (task, update)Cum. T SD (×1012\times10^{12})Cum. T FQ (×1012\times10^{12})Saving, cumulative (%)Saving, per circuit (%)Held-out MSE SD / FQHeld-out acc. (%) SD / FQ
6q/4L (binary, 800)2.76 ± 0.081.34 ± 0.0151.5 ± 1.256.8 ± 1.10.507 / 0.50589.8 / 90.4
10q/4L (binary, 800)7.57 ± 0.313.72 ± 0.0250.8 ± 1.956.1 ± 1.30.514 / 0.51489.8 / 87.8
12q/4L (binary, 800)10.64 ± 0.235.33 ± 0.0149.9 ± 1.056.2 ± 0.40.508 / 0.51489.6 / 89.8
10q/5L (three-class, 500)21.25 ± 0.4210.56 ± 0.2050.3 ± 0.152.9 ± 0.40.6735 / 0.672763.1 / 62.9

Binary runs use 1,024 training examples (8×8\times the original); the three-class run uses 12,288 (96×96\times). These results are in RQ3 (Section 5.2, highlighted in blue) and Appendix D.3, and are discussed further in our W2 / Q2 response to Reviewer pQU6.

W2 / Q2: Trained fixed-precision baselines.

In the submission, fixed-precision baselines were evaluated through cost replays. A fixed tolerance is a valid alternative, but selecting a suitable value requires problem-specific calibration: the tolerance that works depends on the target and noise level (Proposition 1), and trial runs themselves consume T gates. Moreover, a fixed loose tolerance need not satisfy the synthesis bias budget as the gradient decreases. At the reviewer's request, we now train these baselines independently to completion, isolating the contributions of scheduling and mixing. All arms use the main circuit (6q/9L, B32/S500, three paired seeds), with the same training pipeline and stability criterion as SD and FQ.

AuthorsOpenReview ↗

W2 / Q2 (cont.), fixed-precision results and tolerance selection; Q3, Adam preconditioner

Independently trained fixed-precision baselines. Mean ± SD. *For fixed deterministic synthesis at ϵ=6×10−3\epsilon=6\times10^{-3}, two seeds did not stabilize within 1,500 updates (cumulative T approximately 12.0×101212.0\times10^{12} per unsuccessful run at the cap); stability cost and MSE are reported for the single stabilized seed.

ArmT/circuit (update 800)Cum. T at 800 (×1012\times10^{12})Stable seedsCum. T at stability (×1012\times10^{12})Held-out MSE (×10−2\times10^{-2})
FQ (scheduled mixed, ours)2,324 ± 386.57 ± 0.073/310.36 ± 0.417.88 ± 0.62
Fixed mixed, ϵ=3×10−4\epsilon=3\times10^{-4}4,222 ± 2611.74 ± 0.013/317.33 ± 1.038.31 ± 1.05
Fixed det., ϵ=10−3\epsilon=10^{-3}3,260 ± 109.06 ± 0.053/312.51 ± 1.919.59 ± 3.16
Fixed det., ϵ=3×10−3\epsilon=3\times10^{-3}2,661 ± 347.35 ± 0.073/311.01 ± 0.678.30 ± 1.22
Fixed det., ϵ=6×10−3\epsilon=6\times10^{-3}2,310 ± 586.35 ± 0.011/3*9.17*9.79*
SD (scheduled det.)5,279 ± 32313.74 ± 0.243/321.82 ± 0.399.42 ± 3.27

Scheduling versus mixing. Fixed mixed synthesis at ϵ=3×10−4\epsilon=3\times10^{-4} uses the same three-axis mixture as FQ, without scheduling. Relative to this independently trained fixed-mixed baseline, scheduling reduces cumulative T consumption by 44.0% at update 800 and 40.2% at stability, with comparable held-out MSE. Against SD, which schedules without mixing, FQ saves approximately 52%. Both scheduling and mixing therefore make substantial contributions in the evaluated comparisons.

Fixed loose deterministic tolerances. These can be cheaper than SD because they never tighten, but their reliability depends strongly on the selected tolerance. The reliable range is narrow in our experiments: ϵ=3×10−3\epsilon=3\times10^{-3} stabilizes in all three seeds, while ϵ=6×10−3\epsilon=6\times10^{-3} fails to stabilize in two of three seeds. Under the evaluated sequential tolerance-search procedures, identifying a reliable setting requires complete trial runs, including unsuccessful ones. Each trial costs its cumulative T at stability, or at the 1,500-update cap if it never stabilizes: 11.06×101211.06\times10^{12} for 6×10−36\times10^{-3} (mean of 12.02 and 12.00 at the cap and 9.17 at stability), 11.01×101211.01\times10^{12} for 3×10−33\times10^{-3}, and 12.51×101212.51\times10^{12} for 10−310^{-3}. Searching from loose to tight (6×10−36\times10^{-3}, which fails, then 3×10−33\times10^{-3}) therefore consumes 11.06+11.01=22.1×101211.06+11.01=22.1\times10^{12} T gates. Searching from tight to loose (10−310^{-3} and 3×10−33\times10^{-3}, then 6×10−36\times10^{-3}, which fails) consumes 12.51+11.01+11.06=34.6×101212.51+11.01+11.06=34.6\times10^{12}. A single FQ run costs 10.36×101210.36\times10^{12} T gates, 53 to 70% less than these fixed-tolerance selection procedures.

Even with a suitable fixed tolerance known in advance, FQ retains an advantage. Among the tested fixed deterministic tolerances that stabilize in all three seeds, ϵ=3×10−3\epsilon=3\times10^{-3} is the cheapest. Nevertheless, it consumes 11.9±1.2%11.9\pm1.2\% more cumulative T than FQ at matched update. Tightening to ϵ=10−3\epsilon=10^{-3} increases cost further, while loosening to ϵ=6×10−3\epsilon=6\times10^{-3} causes two of three seeds to fail. At ϵ=6×10−3\epsilon=6\times10^{-3}, which approximately matches FQ's per-circuit cost, reliable convergence is lost. FQ also achieves the lowest mean held-out MSE and smallest MSE standard deviation among arms stabilizing in all three seeds. It achieves this without a separate fixed-tolerance grid search.

These full-training experiments establish that the observed advantage is not merely an artifact of cost replay. Scheduling provides a substantial additional saving over the evaluated fixed-mixed baseline, while FQ achieves lower cumulative cost than the reliable fixed deterministic settings and avoids their sensitivity to tolerance selection. All new results and tolerance-search procedures are included in Appendix D.3.

3) Q3: Adam preconditioner.

Yes. Theorem 2 covers Adam when the preconditioner entries lie in [pmin⁡,pmax⁡][p_{\min},p_{\max}] with cκ=1−κρ2≥1/2c_\kappa=1-\kappa\rho^2\ge1/2, which at ρ=0.1\rho=0.1 requires κ≤50\kappa\le50. For the preconditioner Adam actually applies, 1/(v^t+ε)1/(\sqrt{\hat v_t}+\varepsilon), the largest condition number from update 100 to stability is κ=19.0±6.0\kappa=19.0\pm6.0 for SD and 20.0±2.520.0\pm2.5 for FQ, with a maximum of 25.6 over all seeds. The condition therefore holds throughout training, with cκ≥0.74c_\kappa\ge0.74, so the guarantee applies to the preconditioner used in the experiments. The remaining drift term scales as 1−β2≈0.03\sqrt{1-\beta_2}\approx0.03 at β2=0.999\beta_2=0.999 and vanishes as β2→1\beta_2\to1. Its constant remains worst-case, so the bound certifies the structure of convergence rather than a tight rate. The quantities that the savings rest on (the bias orders, the variance law, and the tolerance laws) are verified directly on the Adam runs in RQ1 and RQ2, including all 36 pre-declared tests.

AuthorsOpenReview ↗

Additional Experimental Results and Request for Further Discussion

We thank the reviewer again for the constructive feedback.

We have added further fixed-mixed synthesis experiments at ϵ∈{10−2,6×10−3,3×10−3}\epsilon \in \{10^{-2}, 6\times10^{-3}, 3\times10^{-3}\}. The results show that scheduled FQ achieves more reliable convergence than loose fixed tolerances and saves 19–23% cumulative T cost even against the strongest tested fixed-mixed baseline ablation, with comparable held-out performance. These results are included in the revised manuscript, with changes highlighted in blue.

We believe these additional results have addressed all your concerns. If any issues remain, we would be happy to clarify them before the discussion window closes. We therefore kindly request that you reassess the revised manuscript and consider whether the additional evidence addresses the concerns underlying your initial evaluation.

Thank you again for your valuable time and feedback.

Comments on the paper

5 comments
AuthorsOpenReview ↗

W1 / Q1, practical gradient monitors

Note: We encourage the reviewer to read the revised manuscript, where every change made in response to the reviews is highlighted in blue.

1) W1 / Q1: Dependence on an ideal gradient monitor.

We thank the reviewer for raising the practical monitoring question. The ideal-gradient monitor in the submitted experiments serves as a controlled reference isolating synthesis precision from convergence. It is not intrinsic to the synthesis mechanism or its resource advantage. However, we extend the convergence analysis and evaluate two practical controllers using gradients already available during training.

Theoretical extension:

We establish two complementary results. First, let st=max⁡{∥∇L(θt)∥,gΔ}s_t=\max\lbrace\Vert\nabla L(\theta_t)\Vert,g_\Delta\rbrace and s^t=max⁡{G^t,gΔ}\widehat s_t=\max\lbrace\widehat G_t,g_\Delta\rbrace. If s^t≤(1+δ)st\widehat s_t\le(1+\delta)s_t almost surely at every update, the controller's bias and variance budgets imply

∥bt∥≤ρ′st,E[∥ξt∥2∣Ft]≤(1+λ)σ2+μ′st2,\Vert b_t\Vert\le\rho' s_t,\qquad \mathbb E\left[\Vert\xi_t\Vert^2\mid\mathcal F_t\right]\le(1+\lambda)\sigma^2+\mu' s_t^2,

where ρ′=(1+δ)ρ\rho'=(1+\delta)\rho and μ′=(1+δ)2μ\mu'=(1+\delta)^2\mu. Theorem 2 therefore retains its SGD and Adam convergence bounds with (ρ′,μ′)(\rho',\mu') replacing (ρ,μ)(\rho,\mu), provided ρ′≤1/2\rho'\le1/\sqrt2 for SGD, κρ′2≤1/2\kappa\rho'^2\le1/2 for Adam, and the remaining assumptions hold. The SGD step-size ratio is unchanged since ρ′2/μ′=ρ2/μ\rho'^2/\mu'=\rho^2/\mu for μ>0\mu>0. The oracle is recovered with δ=0\delta=0.

Second, we introduce a new proposition directly analyzing the practical mini-batch monitor s^t+1=max⁡{∥g^t∥,gΔ}\widehat s_{t+1}=\max\lbrace\Vert\hat g_t\Vert,g_\Delta\rbrace, without requiring bounded relative monitoring error. Let B=(1+λ)σ2B=(1+\lambda)\sigma^2, a=2ρ2+μa=2\rho^2+\mu, K=ρ2+ℓημK=\rho^2+\ell\eta\mu, and d=1−a−2K>0d=1-a-2K>0. Under the enforced bias and variance budgets, SGD with η≤1/ℓ\eta\le1/\ell satisfies

1T∑t<TE∥∇L(θt)∥2≤2D(1−a)ηTd+Kd(E[s^02]T+gΔ2+B)+ℓη(1−a)Bd.\frac1T\sum_{t<T}\mathbb E\Vert\nabla L(\theta_t)\Vert^2\le\frac{2D(1-a)}{\eta Td}+\frac{K}{d}\left(\frac{\mathbb E\left[\widehat s_0^2\right]}{T}+g_\Delta^2+B\right)+\frac{\ell\eta(1-a)B}{d}.

Thus, even with gΔ=0g_\Delta=0, the lagged mini-batch monitor admits an SGD convergence bound to a noise-dependent stationarity neighborhood. The complete proof is provided in the revised Appendix B. Practical Adam and split-batch monitoring are also evaluated empirically below.

Practical monitoring experiments:

We evaluated two estimated-gradient monitors on the primary Digits configuration (6 logical qubits, 9 layers, batch size 32, 500 shots), using three paired seeds. Both reuse measurements already collected for parameter-shift optimization:

  • Mini-batch monitor: The preceding optimizer gradient norm, G^t=∥g^t−1∥\widehat G_t=\Vert\hat g_{t-1}\Vert.
  • Split-batch monitor: Two half-batch gradients give G^t=⟨g^t−1(1),g^t−1(2)⟩\widehat G_t=\sqrt{\langle\hat g_{t-1}^{(1)},\hat g_{t-1}^{(2)}\rangle} when the inner product is nonnegative; otherwise, the preceding monitor value is retained.

The model, optimizer, shot budget, stopping criteria, and seeds are matched across monitoring policies. Scheduled deterministic synthesis (SD) and fixed-query mixed synthesis (FQ) are independently trained under each monitor.

Cumulative executed logical T cost (×1012\times10^{12}) through update 800 and at operational stability (6q/9L, B32/S500). Mean ± SD over three paired seeds. Savings compare FQ against SD under the same monitor.

MonitorSD (update 800)FQ (update 800)Saving (%)SD (stability)FQ (stability)Saving (%)Updates to stability, SD / FQ
Oracle13.74 ± 0.246.57 ± 0.0752.2 ± 1.321.82 ± 0.3910.36 ± 0.4152.5 ± 1.11221 ± 39 / 1225 ± 40
Split-batch12.78 ± 0.316.45 ± 0.0649.5 ± 1.719.45 ± 0.639.85 ± 0.3149.4 ± 0.91207 ± 49 / 1225 ± 39
Mini-batch12.32 ± 0.046.41 ± 0.0648.0 ± 0.618.49 ± 0.979.77 ± 0.3747.1 ± 2.11192 ± 63 / 1223 ± 35

Results and resource accounting.

Both estimated monitors preserve the principal resource advantage. Through update 800, FQ reduces cumulative executed T consumption by 49.5% under split-batch monitoring and 48.0% under mini-batch monitoring, compared with 52.2% under the oracle. At operational stability, savings remain 49.4% and 47.1%, respectively, versus 52.5% for the oracle.

Training trajectories also remain comparable. FQ reaches stability after 1225 ± 39 updates under split-batch monitoring and 1223 ± 35 under mini-batch monitoring, compared with 1225 ± 40 under the oracle. Held-out accuracy remains at least 99.4% across evaluated runs.

AuthorsOpenReview ↗

W1 / Q1 (cont.) and W2 / Q2, stationarity targets and scale

Neither practical monitor introduces additional quantum circuit executions, shots, or logical T gates. Both reuse parameter-shift measurements, adding only classical processing. Exact full-data gradients are used separately for convergence diagnostics; they do not influence practical controller decisions and are excluded from training costs. The oracle row is an ideal reference, with its monitoring overhead likewise excluded.

Unlike the instantaneous per-circuit T cost at update 800 reported in the submission, the monitor table measures cumulative execution cost over training.

Summary and changes in manuscript:

The extended Theorem 2 establishes convergence under bounded relative monitoring error, while the new proposition directly analyzes the lagged mini-batch monitor for SGD without that assumption. Independently, the paired-seed experiments demonstrate that both practical monitors retain 47.1–49.4% cumulative T savings at stability, with comparable training duration and high held-out accuracy.

The revised manuscript includes these theoretical results, practical monitoring implementations, and experimental evidence (Section 4, RQ3 in Section 5.2, and Appendices B and F.4). Together, they demonstrate that the observed T-cost advantage persists without an ideal-gradient controller.

2) W2 / Q2: Common stationarity targets, stringent targets, and scale.

We thank the reviewer for this question and address each part with new evidence.

Cost at a common achieved target. At ∥∇L∥=0.01\Vert\nabla L\Vert=0.01, the tightest target reached by all paired SD and FQ seeds on the main circuit, FQ uses 51.5% less cumulative executed T than SD (8.46±0.888.46\pm0.88 versus 17.46±3.35×101217.46\pm3.35\times10^{12}, Table 17). The saving holds at a matched stationarity target, not only at matched updates.

Stringent targets. We trained two additional arms on the main circuit (6q/9L, B32/S500, three paired seeds): ideal training with exact rotations, which has no synthesis error, and fixed-high synthesis (ϵ=3×10−6\epsilon=3\times10^{-6}). No arm reaches ∥∇L∥≤0.003\Vert\nabla L\Vert\le0.003, and ideal training has the highest floor of all (table below). Removing synthesis error entirely does not lower the floor, so synthesis error is not what limits stationarity at this budget. The floor is consistent with the shot and mini-batch variance term of Theorem 2. No evaluated arm, including ideal training, reaches the stringent targets, so they do not reflect a limitation of FT-QNN.

Noise floor on the main circuit (6q/9L, B32/S500). Mean ± SD over three paired seeds.

ArmLowest rolling median ∥∇L∥\Vert\nabla L\VertHeld-out MSE (×10−2\times10^{-2})
Ideal (no synthesis error)0.0078 ± 0.00128.46 ± 1.19
FH, ϵ=3×10−6\epsilon=3\times10^{-6}0.0056 ± 0.00088.14 ± 1.13
SD0.0062 ± 0.00129.42 ± 3.27
FQ0.0052 ± 0.00057.88 ± 0.62

Larger QNNs and more data. We trained SD and FQ on UNSW-NB15 with 1,024 training examples (8×8\times the original) at 6, 10, and 12 qubits (4 layers, B32/S500, three paired seeds). At matched update 800, FQ saves 51.5%, 50.8%, and 49.9% in cumulative T and 56.1–56.8% per circuit, with comparable held-out MSE and accuracy (table below). The saving is unchanged from 4 to 12 qubits and with 8×8\times more training data.

Width and data scaling on UNSW-NB15 (1,024 training examples, 4 layers, B32/S500, update 800). Mean ± SD over three paired seeds.

WidthCum. T SD (×1012\times10^{12})Cum. T FQ (×1012\times10^{12})Saving, cumulative (%)Saving, per circuit (%)Held-out MSE SD / FQHeld-out acc. (%) SD / FQ
6q2.76 ± 0.081.34 ± 0.0151.5 ± 1.256.8 ± 1.10.507 / 0.50589.8 / 90.4
10q7.57 ± 0.313.72 ± 0.0250.8 ± 1.956.1 ± 1.30.514 / 0.51489.8 / 87.8
12q10.64 ± 0.235.33 ± 0.0149.9 ± 1.056.2 ± 0.40.508 / 0.51489.6 / 89.8

Multiclass task. The experiments above show that the saving holds across small and large QNNs. We therefore also tested a larger and harder setting: three-class classification on UNSW-NB15 with 12,288 training examples (96×96\times the original) and a 10q/5L model (B32/S500, three paired seeds). At update 500, FQ saves 52.9±0.4%52.9\pm0.4\% T per circuit and 50.3±0.1%50.3\pm0.1\% cumulative T. Held-out MSE is 0.6727 for FQ versus 0.6735 for SD, and held-out accuracy is matched (table below). The saving carries over to a multiclass task with 96×96\times more data, with no loss in held-out performance.

Three-class UNSW-NB15 (12,288 training examples, 10q/5L, B32/S500, update 500). Mean ± SD over three paired seeds.

ArmT per circuitCumulative T (×1012\times10^{12})Held-out MSEHeld-out acc. (%)
SD4,917 ± 3621.25 ± 0.420.67349 ± 0.0000363.1 ± 1.0
FQ2,318 ± 1410.56 ± 0.200.67273 ± 0.0001062.9 ± 1.2
Saving (%)52.9 ± 0.450.3 ± 0.1––
AuthorsOpenReview ↗

W2 / Q2 (cont.) and W3 / Q3, fault-tolerant resource assumptions

How the conclusions change. They do not. The saving is stable in width, data size, and number of classes. It holds at the tightest common target, and the stringent targets are unreached by every evaluated arm, including ideal training. For tighter targets, Theorem 1 shows that the per-rotation advantage of mixed synthesis persists and approaches a factor of 2 in the bias-limited regime, since the mixed tolerance scales as Δ\sqrt{\Delta} and the deterministic tolerance as Δ\Delta. These results are added to RQ3 in Section 5.2 and Appendix D.3.

3) W3 / Q3: Sensitivity to realistic fault-tolerant resource assumptions.

We thank the reviewer for this question. We first separate two levels of the resource claim. The headline result, a logical T saving of about 50% relative to scheduled deterministic synthesis, is counted directly from the compiled Clifford+T sequences executed in training. It therefore does not depend on any distillation or hardware model. Only the factory-level savings depend on the provisioning model, and for those we tested each of the four assumptions the reviewer raises. The table in Part 4/4 summarizes the results.

Residual logical errors. The analysis already accounts for residual logical noise: it enters the shared bias budget through νt=RMSpMS,t+Rdeppdep\nu_t=R_{\mathrm{MS}}p_{\mathrm{MS},t}+R_{\mathrm{dep}}p_{\mathrm{dep}} in Section 4B, so the tolerance laws reserve part of the budget for it. To quantify RdepR_{\mathrm{dep}}, we applied logical depolarizing noise after every logical gate of the compiled SD and FQ circuits at late-training checkpoints (updates 800 and 1100, three seeds), where the gradient norm is smallest and the budget is tightest. We used the rates implied by the same surface-code model at d=15d=15, 13, and 11, namely pdep=10−9p_{\mathrm{dep}}=10^{-9}, 10−810^{-8}, and 10−710^{-7}. The induced gradient bias is linear in pdepp_{\mathrm{dep}} (fitted exponent 0.9998). Even at the largest rate, which corresponds to a smaller distance than any selected in the paper, it uses 2.9% of the bias budget ρ∥∇L∥\rho\Vert\nabla L\Vert for FQ (update 800) and 5.5–6.2% for SD. Since the bias is linear in pdepp_{\mathrm{dep}}, it is two orders of magnitude smaller at the selected distances (d≥15d\ge15). FQ is also less sensitive than SD (Rdep=807±110R_{\mathrm{dep}}=807\pm110 versus 1403±4011403\pm401), because its circuits contain about half as many T gates and therefore accumulate less residual noise. Residual logical errors thus leave the savings intact, and they favor mixed synthesis.

Physical qubits and runtime. The normalized factory volume 15d3/(1−15pin)15d^3/(1-15p_{\mathrm{in}}) used in the submission is proportional to physical qubits multiplied by code cycles, so it converts directly to physical units. We anchored it to a one-level 15-to-1 factory with 4,620 physical qubits and 42.6 code cycles per output state at d=17d=17 (Litinski, 2019), scaled the footprint as d2d^2 and the time as dd for each selected distance, and assumed a 1 µs code cycle. On the RQ4 traces, scheduled provisioning then saves 25.7±2.1%25.7\pm2.1\% in physical qubit-seconds, which agrees with the normalized saving of 25.9%. For a fixed number of factories (10 to 1,000), it also shortens wall-clock factory time by 9.5±0.6%9.5\pm0.6\%. The normalized model therefore does not overstate the physical benefit.

Factory reconfiguration overhead. We charged each physical configuration switch one complete distillation round plus dd code cycles, during which all factories are idle. This conservative charge adds less than 10−610^{-6}% to total factory time. The overhead is negligible because a single training update consumes about 101010^{10} magic states, whereas a reconfiguration lasts on the order of 100 µs, and the trace contains only about 15 physical switches.

Replay versus retraining. To verify that replaying recorded demand does not bias the provisioning results, we retrained UNSW-NB15 (6q/4L, updates 250 to 500, three seeds) with joint fidelity selection and the exact injection channel at every T gate. Retraining reproduces the replayed saving over the same window, 36.5±6.1%36.5\pm6.1\% against 36.5±5.7%36.5\pm5.7\%, with unchanged held-out MSE (0.489 versus 0.488) and accuracy. This window exceeds the full-trace value in Table 5 because fidelity requests vary most in this phase of training.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.