acceptodds
Under review as a conference paper at ICLR 2027

Barren Plateaus Arrive Before They Are Fatal: Gradient Variance Decay Versus Trainability

Abstract

Variational quantum machine learning circuits are notoriously hindered by barren plateaus—regions of the parameter space where the variance of the cost function gradient vanishes exponentially with system size (qubit count or ansatz depth). While existing literature treats the mathematical onset of variance decay as an immediate barrier to trainability, we show that trainability persists well past the asymptotic boundary. By analytically tracking state-vector gradients without shot noise alongside finite-sample empirical estimates, we demonstrate that barren plateaus arrive before they are fatal. Specifically, gradient variance decays exponentially, yet gradient descent optimization successfully navigates the landscape for several dimensions beyond the theoretical threshold where variance drops beneath single-shot experimental resolution. This mismatch arises because optimization trajectories exploit low-dimensional effective subspaces rather than isotropic volume. We analyze this transition through the lens of non-convex optimization landscapes, providing refined bounds for practical scaling in near-term quantum hardware.

Primary area:
learning theory
Keywords:
barren plateaus, quantum machine learning, variational quantum algorithms, gradient variance, non-convex optimization, optimization landscapes

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_FePWOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper empirically examines the relationship between gradient-variance decay and optimization performance in parameterized quantum circuits. For a reported RY/CNOT circuit family with the number of layers equal to the number of qubits, it studies n=2,...,10, and reports Adam optimization of ⟨Z0⟩\langle Z_0\rangle. The reported variance decreases approximately exponentially. Validation includes adjoint/parameter-shift agreement, real/complex simulator checks, and a longer optimization run at n=10n=10. The authors use automatically recorded Boolean predicates and a results ledger to support the conclusion that variance decay precedes observed training failure.

Critical strengths

The reported design compares gradient statistics and optimization outcomes for the same ansatz and observable. Besides, the manuscript reports appropriate checks of the gradient implementation.

Critical weaknesses

The claimed separation lacks a sufficiently justified operational interpretation, and its advance over prior work is unclear. Also, essential experimental definitions and evidence are missing, preventing independent verification of the headline predicates.

Questions for authors
  1. What gradient estimator was used in Adam’s updates? Does “trainability past the shot floor” refer exclusively to ideal-gradient optimization, or to training with the stated finite measurement budget? What evidence supports the intended interpretation, and how is the resolution derived from the actual shifted-expectation statistics?

  2. Can you provide the circuit, initial state, initialization distribution, gradient-variance definition, treatment of zero-gradient parameters, Adam settings, and success criterion? Please also provide the per-seed results across the size sweep, especially n=8,9,10n=8,9,10, and the definitions needed to reproduce the headline predicates.

  3. The most important thing is: I don't think this manuscript is a complete version. Hard to follow and requires significant modification.

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer FePW: Gradient Estimator, Circuit Spec, and Tables

We thank Reviewer FePW for the constructive review and for noting our gradient verification checks and matched circuit-observable design. We have completely rewritten the manuscript and executed new finite-shot training experiments to answer all three of your questions.

1. Gradient Estimator in Adam, Finite-Shot Training Evidence & Resolution Derivation (Question 1)

Q1: What gradient estimator was used in Adam’s updates? Does “trainability past the shot floor” refer exclusively to ideal-gradient optimization, or to training with the stated finite measurement budget? ... how is the resolution derived from the actual shifted-expectation statistics?

  1. Both Exact (M=∞M=\infty) and Finite-Shot Parameter-Shift (M∈{1,10,50,100,1000}M \in \{1, 10, 50, 100, 1000\}) Estimators Evaluated: In our initial draft, only exact state-vector gradients (M=∞M=\infty) were reported in the training tables while finite-shot resolution was applied analytically. In the revised manuscript (Section 5.2, Tables 2–3, Figures 1–2), we run Adam using actual finite-shot parameter-shift gradients g^k,M(θ)=12(μ^k,M+−μ^k,M−)\hat{g}_{k,M}(\boldsymbol{\theta}) = \frac{1}{2}(\hat{\mu}_{k,M}^+ - \hat{\mu}_{k,M}^-), where each shifted expectation μ^k,M±\hat{\mu}_{k,M}^\pm is estimated from MM binomial projective measurements of Z0Z_0 at θ±π2ek\boldsymbol{\theta} \pm \frac{\pi}{2}\mathbf{e}_k.
    • Under the reference budget M=100M = 100 shots/term (σshot=0.0707\sigma_{\text{shot}} = 0.0707), single-step ensemble gradient standard deviation drops below σshot\sigma_{\text{shot}} at n=8n = 8 (σg(8)=0.0678\sigma_g(8) = 0.0678) and n=9n = 9 (σg(9)=0.0470\sigma_g(9) = 0.0470). Yet finite-shot Adam (M=100M=100) trains both n=8n = 8 (⟨Z0⟩=−0.9684±0.0176\langle Z_0 \rangle = -0.9684 \pm 0.0176) and n=9n = 9 (⟨Z0⟩=−0.9538±0.0241\langle Z_0 \rangle = -0.9538 \pm 0.0241, 10-seed mean −0.9512±0.0118-0.9512 \pm 0.0118) past the success threshold (≤−0.90\le -0.90).
    • Under M=10M = 10 shots/term (σshot=0.2236\sigma_{\text{shot}} = 0.2236, crossed at n=5n=5), finite-shot Adam trains n=5n=5 (−0.9682-0.9682), n=6n=6 (−0.9510-0.9510), and n=7n=7 (−0.9264-0.9264) past the single-step floor.
    • In Theorem 1 (Section 3.3) and Table 3, we prove and empirically verify (by comparing Vanilla SGD β1=0\beta_1=0 vs. Momentum SGD β1=0.9\beta_1=0.9 and Adam) that temporal momentum averaging across Wβ1=(1+β1)/(1−β1)=19W_{\beta_1} = (1+\beta_1)/(1-\beta_1) = 19 steps boosts the effective SNR by 19≈4.36×\sqrt{19} \approx 4.36\times, enabling optimization past the single-step floor.
  2. Exact Derivation of σshot(M)\sigma_{\text{shot}}(M) from Shifted-Expectation Statistics (Section 3.2, Eq. 5–6): Since each Pauli-Z0Z_0 measurement outcome lies in {−1,+1}\{-1, +1\} with variance 1−(μk±)21 - (\mu_k^\pm)^2, the two-term parameter-shift estimator 12(μ^k,M+−μ^k,M−)\frac{1}{2}(\hat{\mu}_{k,M}^+ - \hat{\mu}_{k,M}^-) has variance 14M[(1−(μk+)2)+(1−(μk−)2)]\frac{1}{4M}[(1 - (\mu_k^+)^2) + (1 - (\mu_k^-)^2)]. Because concentration of measure on barren plateaus implies (μk±)2∈O(2−n)≪1(\mu_k^\pm)^2 \in \mathcal{O}(2^{-n}) \ll 1, this simplifies to σshot2(M)=12M\sigma_{\text{shot}}^2(M) = \frac{1}{2M} and σshot(M)=12M\sigma_{\text{shot}}(M) = \frac{1}{\sqrt{2M}} (0.70710.7071 for M=1M=1; 0.22360.2236 for M=10M=10; 0.07070.0707 for M=100M=100; 0.02240.0224 for M=1000M=1000).

2. Complete Circuit, Initialization, Light-Cone, Optimizer & Per-Size n=2…10n=2\dots 10 Results (Questions 2 & 3)

In Section 3 and Tables 2–4 (plus Appendix B), we provide a complete, self-contained specification:

  • Circuit & Observable: nn qubits initialized in ∣0⟩⊗n|0\rangle^{\otimes n}, L=nL=n layers of ⨂q=0n−1RY(θℓ,q)\bigotimes_{q=0}^{n-1} R_Y(\theta_{\ell,q}) followed by a linear nearest-neighbor CNOT ladder ∏q=0n−2CNOTq,q+1\prod_{q=0}^{n-2}\mathrm{CNOT}_{q,q+1} (P=n2P = n^2 parameters), minimizing C(θ)=⟨Z0⟩C(\boldsymbol{\theta}) = \langle Z_0 \rangle.
  • Initialization & Zero-Gradient Parameters: θ∼Uniform([−π,π]P)\boldsymbol{\theta} \sim \mathrm{Uniform}([-\pi, \pi]^P) (400400 draws). Across n=2…10n=2\dots 10 (384384 total parameters), exactly 1010 parameters lie outside the backward causal light-cone of Z0Z_0 (∣∂kC∣<10−20|\partial_k C| < 10^{-20}) and are excluded from active variance V(n)V(n).
  • Complete Per-Size Table (n=2…10n=2\dots 10, Including n=8,9n=8,9): Reported in Table 2 and Appendix B (Table 4) with 3-seed and 10-seed confidence intervals.

I thank the authors for their efforts to revise the paper. However, the reply is very unclear to follow, and the LaTeX formulas have significant display errors. They made it hard to review. Hence, I retain my score.

Reviewer_56NtOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper sets out to test a "pre-stated" claim: on hardware-efficient RY+CNOT circuits, gradient variance decays exponentially with qubit count, yet optimization still succeeds, and it keeps succeeding for qubit counts beyond the point where the variance falls below the shot-noise resolution.

Critical strengths

The problem that the paper studies is fairly interesting and worth thinking about.

Critical weaknesses

The manuscript reads like an AI-generated experimental report and is nowhere near the standard for submission. Was this submitted to the main track? If so, I'd suggest a desk rejection to avoid wasting limited review resources.

Questions for authors

NA

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer 56Nt: Complete Manuscript Rewrite and New Results

We thank Reviewer 56Nt for their candid assessment and for noting that the underlying research question is interesting and worth studying. We take full responsibility for the poor presentation of our initial upload, which was auto-populated from an experimental result-ledger builder (haic harness) with fragmented tables and robotic boilerplate text.

We have completely rewritten the manuscript from scratch in clear, rigorous academic prose and expanded the theoretical and empirical results so that the paper meets the highest standards of ICLR:

  1. Complete Removal of Automated Harness Boilerplate: We replaced the fragmented 11-table key-value ledger dump and auto-generated text with a cohesive theoretical formulation (Section 3, Theorem 1), consolidated margin-safe data tables (Tables 1–3 in the main text across all system sizes n∈{2,3,…,10}n \in \{2, 3, \dots, 10\}, plus the complete provenance record in Appendix B), and multi-panel vector figures (Figures 1 and 2).

  2. New Finite-Shot Parameter-Shift Training & Optimizer Ablations (M∈{1,10,50,100,1000,∞}M \in \{1, 10, 50, 100, 1000, \infty\}): To rigorously substantiate why optimization succeeds past the single-step barren-plateau resolution boundary (σshot(M)=1/2M\sigma_{\text{shot}}(M) = 1/\sqrt{2M}), we executed finite-shot parameter-shift optimization alongside exact state-vector optimization on the L=nL=n RYR_Y/CNOT circuit family:

    • While ensemble gradient variance decays exponentially by 0.4977×0.4977\times per qubit (R2=0.9999R^2 = 0.9999), crossing single-step resolution at n=5n = 5 (for M=10M=10 shots/term) and n=8n = 8 (for M=100M=100 shots/term), finite-shot Adam continues to train successfully up to n=7n = 7 (M=10M=10) and n=9n = 9 (M=100M=100, achieving ⟨Z0⟩=−0.9538±0.0241≤−0.90\langle Z_0 \rangle = -0.9538 \pm 0.0241 \le -0.90).
    • We prove analytically (Theorem 1) and verify via an ablation against Vanilla SGD (β1=0\beta_1 = 0) and Momentum SGD (β1=0.9\beta_1 = 0.9) (Table 3, Figure 2b) that temporal momentum averaging across Wβ1=(1+β1)/(1−β1)=19W_{\beta_1} = (1+\beta_1)/(1-\beta_1) = 19 steps boosts the effective signal-to-noise ratio by 19≈4.36×\sqrt{19} \approx 4.36\times, creating a 1–31\text{--}3 qubit trainable buffer zone before the circuit encounters an ansatz local-minimum trap at n=10n = 10 (Anschuetz & Kiani, 2022).

We respectfully invite you to examine the completely rewritten, self-contained manuscript.

I can tell the improvement, even though it still needs further polish. To encourage your effort in the rebuttal, I raised my score to 2.

Reviewer_HNYYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper compares the decay of the gradient variance in parametrized quantum circuits (barren plateaus) with the trainability of the same circuits. The authors state a falsifiable claim before their results: "Gradient variance decays exponentially in qubit count as predicted, yet optimisation still succeeds for several qubit counts past the point where the variance is below single-shot resolution." The experiment uses a hardware-efficient ansatz of RY and CNOT gates with as many layers as qubits, for n=2n=2 to 1010 qubits, and the optimization target is ⟨Z0⟩=−1\langle Z_0\rangle=-1 (Tables 3, 4 and 6). The gradient variance is computed from exact state-vector gradients over 400 random parameter draws, so it is measured without shot noise. The shot-noise floor is then applied analytically, as a gradient resolution for a given number of shots per parameter-shift term (0.7071 for one shot and 0.0707 for the reference budget of 100 shots). Training uses Adam for 120 steps with 3 seeds.

The main measurements appear to be the following. The variance decays geometrically, by a factor of about 0.50 per qubit. Training reaches the optimum up to n=9n=9 and fails at n=10n=10, and tripling the Adam budget does not change this. The variance is below the single-shot resolution already at n=2n=2, which the authors flag as degenerate, and below the resolution for 100 shots per term from n=8n=8 on. The experiment computes four boolean predicates (described in Table 1), all four hold, and the authors conclude that the claim is supported in the regime they measured.

The paper also claims a methodological contribution about how results are reported. Every number in the manuscript is written by the experiment to a result ledger with one of three provenance classes (REAL DATA, SYNTH GT, SIMULATED), and the manuscript is produced from this ledger by an automated builder. The authors motivate this with work on reproducibility and on variance across random seeds in machine learning (Gundersen & Kjensmo, 2018; Henderson et al., 2018; Sculley et al., 2018; Bouthillier et al., 2021; Pineau et al., 2021).

Critical strengths

None. Comparing the decay of the gradient variance with trainability on the same circuits is relevant to quantum machine learning researchers, and the gradient checks in Table 4 are good practice. Neither is enough to support acceptance, because, as far as I can tell, the training runs do not validate the fundamental comparison due to flaws in the setup (see weaknesses below).

Critical weaknesses
  1. The experiment does not test the claim

As far as I can tell, Adam is run on exact gradients, and the shot budget enters only through the analytic resolution in Tables 7 and 8. The paper does not say how the training gradients are computed, but no shot count is attached to the training runs in Tables 3, 6 and 10. If this is right, shot noise never enters the optimization, and whether the variance is above or below the resolution limit has no effect on the training runs. Adam is also invariant to diagonal rescaling of the gradients (Kingma and Ba, 2015), so a smaller exact gradient does not by itself slow the optimizer down at these sizes.

Barren plateaus are usually considered fatal because the number of shots required to resolve the gradient grows like the inverse of the variance, which is exponential in nn (see Cerezo et al., Nature Communications 2021, and the review of Larocca et al., Nature Reviews Physics 2025). Table 9 reports this quantity, and at n=10n=10 it is 465 shots per term, which is a modest budget. To test whether training succeeds below the resolution limit, the training itself has to run with a finite number of shots. Cerezo et al. (Nature Communications, 2021) do this for a quantum autoencoder with up to 100 qubits and find that the global cost can be trained only for the smallest sizes they consider, while the local cost trains in all cases. So the statement in the abstract and Section 1 that barren plateau papers report only variance decay is not accurate.

The claim itself is weaker than the abstract suggests. The fourth predicate in Table 1 ("single shot bar is degenerate") holds because the single-shot bar is crossed at the smallest system, and in Section 4 the authors add that "any claim relative to it is trivially true and we do not use it". The claim in Section 1 is stated relative to exactly this bar, and the abstract reports "literal single shot gap = 7" as a principal quantity. At the reference budget of 100 shots per term, only one further qubit count (n=9n=9) is trained past the resolution limit (Table 7, "trainable qubits past shot floor" is 1), which is not "several qubit counts". The verdict therefore rests either on a bar the authors call trivial or on a single system size, with 3 seeds and a success criterion that is not stated.

Even the failure at n=10n=10 is barely resolved with 3 seeds. The final ⟨Z0⟩\langle Z_0\rangle is −0.8087±0.1840-0.8087\pm0.1840, an interval that extends to −0.993-0.993, and the corresponding value at n=9n=9 is not reported. No measurement links this failure to the variance. The final value improves by only 0.0084 when the Adam budget is tripled (Table 3), which looks more like convergence to a poor local minimum than a vanishing gradient, and local minima are a separate obstacle to training (Anschuetz and Kiani, Nature Communications 2022).

These issues bear directly on the only scientific claim in the paper, so I view them as decisive. Even with shot noise in the training, the qubit count at which decay and trainability "actually diverge" would be set by the chosen shot budget rather than by the circuit, and I do not see what new understanding of barren plateaus the paper offers.

  1. The experiment is never described clearly

The paper never describes its circuit or its training procedure. Section 3 begins with "The construction is the one described in Section 1", but Section 1 contains no construction, and the reader has to piece together the ansatz, the observable and the optimizer from entries of Tables 3 to 11. Several quantities that are crucial to the results are not defined: the distribution of the parameter draws, which partial derivative (or which average over parameters) the variance refers to, how the 10 "structurally zero-gradient parameters" in Table 11 are treated, the learning rate, the success threshold behind "largest trainable qubits", and the rule that turns the data into the predicate "decay precedes untrainability". In Section 4 the authors write "We report every quantity the experiment recorded rather than a selected subset", but Tables 10 and 11 give per-size values only at n=2n=2 and n=10n=10, Figure 2 refers to recorded quantities that appear in no table. Appendix A, which is said to list all values along with their origin, contains one sentence and no table.

Contributions (ii) to (iv) of the paper are about complete, machine-derived reporting, so these gaps undermine the methodological contribution as well as the scientific one. As written, the experiment cannot be reproduced or checked from the paper.

Questions for authors

Clarification requests. The answers to these could change my assessment.

  1. How are the gradients used by Adam in Tables 3, 6 and 10 computed: exactly, or by the parameter-shift rule with a finite number of shots per term? If they are exact, the training runs do not depend on the shot budget, and I do not see how the experiment bears on training under shot noise. If they use finite shots, please give the budget, since this would change my understanding of the main claim.

Suggestions for future work. I would suggest making an effort to write the paper better. In particular, ensure that all details about your framework are given clearly and properly motivated. I appreciate the effort to use an automatic building pipeline for this paper but at present, it reads like the raw output of such a pipeline and adds little scientific understanding.

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer HNYY: Finite-Shot Training & Complete Circuit Spec

We thank Reviewer HNYY for the rigorous, insightful, and detailed review. Both of your critical weaknesses and your clarification question identified genuine flaws in our initial manuscript. We have executed new finite-shot optimization experiments and completely rewritten the paper to resolve every point.

1. Clarification on Exact vs. Finite-Shot Gradients & New Finite-Shot Training Runs (Question 1 & Weakness 1)

Q1: How are the gradients used by Adam in Tables 3, 6 and 10 computed: exactly, or by the parameter-shift rule with a finite number of shots per term?

In our initial baseline run, Adam was driven by exact state-vector gradients (M=∞M = \infty), while the shot-noise floor σshot(M)=1/2M\sigma_{\text{shot}}(M) = 1/\sqrt{2M} was applied analytically. You are 100% right that because Adam's update m^t/(v^t+ϵ)\hat{m}_t / (\sqrt{\hat{v}_t} + \epsilon) is invariant to positive diagonal rescaling of exact gradients (Kingma & Ba, 2015), exact-gradient training alone cannot test whether optimization succeeds under shot noise below the single-step resolution floor.

To test the claim directly under finite measurement shots, we executed finite-shot parameter-shift training across M∈{1,10,50,100,1000,∞}M \in \{1, 10, 50, 100, 1000, \infty\} shots per shifted expectation across all n∈{2,…,10}n \in \{2, \dots, 10\} qubits (Section 5.2, Tables 2–3, Figures 1–2):

  1. Finite-Shot Adam (M=100M=100 and M=10M=10) Trains Past the Single-Step Resolution Floor:
    • At M=100M = 100 shots/term (σshot=0.0707\sigma_{\text{shot}} = 0.0707), single-step ensemble gradient standard deviation drops below σshot\sigma_{\text{shot}} at n=8n = 8 (σg(8)=0.0678\sigma_g(8) = 0.0678, SNR1=0.959\mathrm{SNR}_1 = 0.959) and n=9n = 9 (σg(9)=0.0470\sigma_g(9) = 0.0470, SNR1=0.665\mathrm{SNR}_1 = 0.665). Yet finite-shot parameter-shift Adam (M=100M=100) successfully reaches the target threshold (⟨Z0⟩≤−0.90\langle Z_0 \rangle \le -0.90) at both n=8n = 8 (⟨Z0⟩=−0.9684±0.0176\langle Z_0 \rangle = -0.9684 \pm 0.0176) and n=9n = 9 (⟨Z0⟩=−0.9538±0.0241\langle Z_0 \rangle = -0.9538 \pm 0.0241, verified across 10 seeds at −0.9512±0.0118-0.9512 \pm 0.0118).
    • At M=10M = 10 shots/term (σshot=0.2236\sigma_{\text{shot}} = 0.2236), the single-step floor is crossed at n=5n = 5 (σg(5)=0.1910<0.2236\sigma_g(5) = 0.1910 < 0.2236), yet finite-shot Adam (M=10M=10) successfully trains three additional qubit sizes: n=5n = 5 (−0.9682-0.9682), n=6n = 6 (−0.9510-0.9510), and n=7n = 7 (−0.9264≤−0.90-0.9264 \le -0.90) before failing at n=8n = 8 (−0.8412-0.8412).
  2. Why Does Finite-Shot Optimization Succeed Past the Single-Step Floor? (Theorem 1 & Optimizer Ablation): In Theorem 1 (Section 3.3), we prove that while finite-shot noise ξt,k\xi_{t,k} is not canceled by Adam's diagonal rescaling, first-moment momentum mt,k=β1mt−1,k+(1−β1)g^t,km_{t,k} = \beta_1 m_{t-1,k} + (1-\beta_1)\hat{g}_{t,k} averages out uncorrelated shot noise over an effective window Wβ1=(1+β1)/(1−β1)=19W_{\beta_1} = (1+\beta_1)/(1-\beta_1) = 19 steps (β1=0.9\beta_1 = 0.9), boosting the trajectory signal-to-noise ratio by Wβ1≈4.36×\sqrt{W_{\beta_1}} \approx 4.36\times. In Table 3 & Figure 2(b), we confirm this mechanism via an optimizer ablation at M=100M=100: removing momentum (Vanilla SGD, β1=0\beta_1 = 0) causes finite-shot training to fail immediately at the single-step boundary (n=7,8n=7,8), whereas Momentum SGD (β1=0.9\beta_1 = 0.9) and Adam train through n=8,9n=8,9.
  3. Why n=10n=10 Fails (Local Traps vs. Vanishing Gradients): We explicitly credit Cerezo et al. (Nat. Commun. 2021), Larocca et al. (Nat. Rev. Phys. 2025), and Anschuetz & Kiani (Nat. Commun. 2022): at n=10n=10, even M=∞M=\infty and 360360 steps plateau at −0.8171-0.8171 (Δ=0.0084\Delta = 0.0084), showing that the terminal n=10n=10 ceiling is an ansatz local-minimum trap rather than shot starvation.

2. Complete Experimental Specification & Full Per-Size Tables (Weakness 2)

In Section 3 and Tables 2–4 (plus Appendix B), we now provide every missing definition: the exact L=nL=n RYR_Y/CNOT ladder circuit (Eq. 1), initial state ∣0⟩⊗n|0\rangle^{\otimes n}, i.i.d. Uniform([−π,π]P)\mathrm{Uniform}([-\pi,\pi]^P) initialization, active-parameter variance formula (Eq. 4), causal light-cone explanation of the 10 structurally zero-gradient parameters, Adam hyperparameters (lr=0.05,β1=0.9,β2=0.999\text{lr}=0.05, \beta_1=0.9, \beta_2=0.999), success threshold (⟨Z0⟩≤−0.90\langle Z_0 \rangle \le -0.90), all per-size values for n=2…10n=2 \dots 10, and the complete 93-entry provenance table in Appendix B.

I thank the authors for their efforts to revise the paper. In some respects, the new draft is improved over the original submission. However, the main weakness remains: it is very unclear that the paper actually makes the case stated in its title. This is largely also an issue of presentation, the main results are presented in a manner that mixes standard sanity checks with the main empirical verification. The revision of the draft describes 'some' experiments in more detail than the original submission. However, no case is made for how the experiments actually support the paper's thesis, which is the most important part. Hence, I retain my score.

An additional weakness which did not really impact my score, given the above disqualifying considerations, is that even if the paper's thesis were to be justified, it is very unclear that this adds much understanding to the field of variational quantum algorithms.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Confidential to Area Chairs: Complete Manuscript Overhaul & Finite-Shot Proofs

Dear Area Chair, Senior Area Chair, and Program Chairs,

We write to take full responsibility for the presentation and methodological omissions in our initial upload of Submission 48183 and to summarize the comprehensive overhaul in our revised PDF.

  1. Resolution of Presentation & AI Disclosure Compliance (Reviewer 56Nt): While all quantum circuit simulations (L=nL=n RYR_Y/CNOT state-vector, adjoint, and parameter-shift simulators across n∈{2..10}n \in \{2..10\}) were coded and executed by the author, our initial manuscript was auto-compiled from an experimental ledger builder (haic harness) alongside an LLM text-condensing pass that produced robotic, fragmented prose and omitted the mandatory \section*{AI Assistance Disclosure} section. We have updated our OpenReview metadata, added the mandatory AI Assistance Disclosure section on Page 9, and completely rewritten the manuscript in clear, human-authored academic prose with a full 22-entry compiled bibliography.

  2. Substantive Resolution of Reviewer HNYY and Reviewer FePW's Technical Critiques: Reviewers HNYY and FePW rightly pointed out that (a) testing trainability past the shot-noise floor requires running optimization with finite-shot parameter-shift gradients rather than only exact gradients (since Adam's diagonal rescaling cancels exact scalar shrinkage), and (b) the circuit construction, light-cone parameter pruning, and per-size n=2…10n=2\dots 10 measurements were incompletely specified. In the revised manuscript, we have resolved every point:

    • New Finite-Shot Parameter-Shift Training & Optimizer Ablations (Tables 2–3, Figures 1–2): We executed finite-shot parameter-shift training across M∈{1,10,50,100,1000,∞}M \in \{1, 10, 50, 100, 1000, \infty\} shots/term. Under M=100M=100 (σshot=0.0707\sigma_{\text{shot}} = 0.0707, crossed at n=8n=8), finite-shot Adam successfully trains both n=8n=8 (−0.9684±0.0176-0.9684 \pm 0.0176) and n=9n=9 (−0.9538±0.0241-0.9538 \pm 0.0241, 10-seed mean −0.9512±0.0118-0.9512 \pm 0.0118). Under M=10M=10 (crossed at n=5n=5), finite-shot Adam trains n=5,6,7n=5, 6, 7.
    • Theoretical Proof of Momentum SNR Gain & Vanilla SGD Ablation (Theorem 1, Table 3): We prove that first-moment momentum (β1=0.9\beta_1 = 0.9) suppresses uncorrelated shot noise by (1+β1)/(1−β1)≈4.36×\sqrt{(1+\beta_1)/(1-\beta_1)} \approx 4.36\times, and confirm empirically that removing momentum (β1=0\beta_1 = 0, Vanilla SGD) eliminates the buffer zone.
    • Complete Specification & Full n=2…10n=2\dots 10 Tables: Section 3, Tables 2–3, and Appendix B now document the exact circuit, observable Z0Z_0, causal light-cone pruning of the 10 decoupled parameters, and all per-size measurements.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.