- 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
- 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.
- 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.
- 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.
- 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.
- 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
- 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.
- 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".
- 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.
- 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
- 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?
- 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?
- 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")?
- 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?
- 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)?
- 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?
- 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?
- The Waveform class pair was chosen "after results on another pair" (Appendix D.3). What were the results on the original pair?
6 comments
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 more training data at 6, 10, and 12 qubits, and on a three-class UNSW-NB15 task with 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 , and ideal training has the highest gradient floor (). 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 , MSE and accuracy are evaluated at stability.
| Task | Point | Saving (95% CI) | MSE difference (95% CI) | Accuracy difference, pts (95% CI) |
|---|---|---|---|---|
| Digits 6q/9L | Update 800 | 52.2% [51.0, 53.3], | −0.015 [−0.044, +0.013], | +0.25 [−0.44, +0.94], |
| Digits 6q/9L | 57.8% [47.3, 68.2], | −0.015 [−0.045, +0.014], | +0.25 [−0.44, +0.94], | |
| UNSW-NB15 6q/4L | Update 500 | 48.6% [45.6, 51.5], | +0.008 [−0.015, +0.030], | −1.3 [−5.7, +3.2], |
| Waveform 6q/4L | Update 500 | 51.6% [49.3, 54.0], | +0.012 [−0.017, +0.042], | −1.3 [−4.2, +1.7], |
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 .
Measured response constants versus number of synthesized trainable rotations (digits, B32/S500). 6q/9L: mean ± SD over three seeds; other models: one saved seed.
| Model | |||||
|---|---|---|---|---|---|
| 6q/4L | 48 | 6.4 | 3.0 | 17 | 3.3 |
| 4q/9L | 72 | 5.0 | 3.4 | 23 | 2.5 |
| 6q/6L | 72 | 6.8 | 7.2 | 54 | 3.4 |
| 6q/9L | 108 | 6.3±0.8 | 5.8±1.0 | 39±18 | 3.2±0.4 |
| 8q/9L | 144 | 10.3 | 9.5 | 68 | 5.1 |
The bias constants grow at most linearly in : over a increase in rotations, the deterministic bias constant grows by and the mixed bias constant by , so the deterministic bias does not accumulate as a sum of response norms. The variance constant grows by , which tightens the variance cap by only in tolerance, since the cap scales as . The quantity that sets the deterministic-to-mixed tolerance ratio, , 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 that of 6q/9L, which tightens the variance cap by only . 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.
| Arm | Scale | Cum. T at 800 () | Seeds reaching 0.01 | Outcome |
|---|---|---|---|---|
| FQ | 0.1 | 6.42 | 3/3 | Stable, 3/3 |
| FQ | 10 | 8.25 | 3/3 | Stable, 3/3 |
| SD | 0.1 | 10.38 | 3/3 | Stable, 3/3 |
| SD | 10 | 16.85 | 2/3 | Infeasible, 3/3 |
FQ is robust to a error in either direction: every seed stabilizes and reaches , with unchanged held-out MSE. Underestimating its constant changes the cost by −2%, and overestimating it by +26%.
W4 (cont.); Q3, FQ vs. PS; Q4, tightness; Q7, non-cancellation; Q8, Waveform pair
Deterministic synthesis is not: with overestimated , the required tolerance falls below the grid in all three seeds. This is the requirement of Theorem 1 in practice, while mixed synthesis keeps a margin. At matched scaling, FQ saves (×0.1) and (×10). Even with FQ's constant overestimated and SD's underestimated , FQ uses 20% less cumulative T. The calibrated constants and are 6–12× the measured values above, so they are conservative by design, and this conservatism costs little. 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 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 , whereas PS is restricted to . 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 and , the deterministic tolerance reaches only for , 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: is about the measured mixed bias constant, which gives 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 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 looser constant changes FQ's cost by only −2%, since the tolerance is mostly at its cap, and a more conservative one by +26%, with every seed still reaching the 0.01 target.
7) Q7: Non-cancellation constant and deterministic cancellation.
We estimated directly: 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 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 (): 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.
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, Corollary 2 certifies no cumulative advantage. The premise, however, is not. The experimental rule does not correspond to λ≈1.
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 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, and the ratio of synthesis gradient variance to intrinsic shot and mini-batch variance is λeff=0.012–0.019 (mean 0.015) with Adam and 0.011–0.020 (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)≈1.97, 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, floor gΔ=ϕΔ, 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 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, which every run reaches.
Certified versus realized cumulative cost ratio (SD/FQ), same setting. Mean ± SD.
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 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, 3×10−3, and 6×10−3 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% for every arm at update 800.
Yes, fixed loose tolerances fail where FQ succeeds. At ϵ=6×10−3, 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. At 3×10−3, the runs stabilize, but two of three seeds never reach ∥∇L∥=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 10−3, +11.9±1.2% at 3×10−3, and −3.4±0.9% at 6×10−3. The only fixed tolerance that matches FQ on every gradient target is 10−3, which costs 37.9% more. Finding a working fixed tolerance also requires trial runs (Prop 1), costing 22.1–34.6×1012 T gates against 10.4×1012 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.