acceptodds
Under review as a conference paper at ICLR 2027

Accelerating Feedback-based Algorithms for Quantum Optimization Using Gradient Descent

Abstract

Feedback-based methods, built on Quantum Lyapunov Control (QLC), have gained significant attention as an alternative training paradigm for Variational Quantum Algorithms (VQAs) in solving combinatorial optimization problems such as MAX-CUT. These methods aim to reduce the training overhead associated with VQAs, and can mitigate barren plateaus. QLC methods often require long control sequences, and may show sensitivity to the evolution timestep, which can result in slow convergence compared to gradient-based VQA approaches. In this work, we address these challenges by proposing a hybrid method that incorporates per-layer gradient estimation to accelerate the convergence of QLC while preserving its low training overhead and stability guarantees. By leveraging layer-wise gradient information, our approach selects improved control parameters and achieves faster convergence rates compared to common QLC methods such as FALQON. We validate our algorithm through various empirical studies across a range of problem instances and optimization settings. We empirically show our algorithm's robustness against noise and variations in the timestep parameter. We also demonstrate a parameter initialization scheme for QAOA using our method that can lead to improved solution quality.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Quantum Optimization, Quantum Lyapunov Control, Gradient Descent, QAOA

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KU22OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Gradient Descent Quantum Lyapunov Control (GD-QLC) which is a feedback based QLC method. Standard QLC methods like FALQON require deep circuits for smaller time steps Δt\Delta t, but when the circuit is shortened with larger time steps Δt\Delta t it can become unstable. To address this the author proposes a greedy layerwise training variant of QLC. Each layer’s control parameter is initialized at zero and updated with L gradient descent steps, before it is frozen and the next layer is added. They show GD-QLC makes the energy non-increasing with a sufficiently small step size (Theorem 1) and that the inner loop converges to a stationary point under noisy gradients (Theorem 2). Empirically GD-QLC converges with fewer layers than FALQON like methods including under finite shots and simulated hardware noise.

Critical strengths

The benchmark results are strong. GD-QLC reaches better approximation ratios than FALQON often with much fewer layers and it also matches or exceeds SO-FALQON / FOCQS. It is more robust to the time step Δt\Delta t where the other methods begin to become unstable at larger values such as Δt\Delta t = 0.3. The shot experiment looks particularly strong where GD-QLC remains much more stable. This suggests the method is more suitable in NISQ.

I also thought the study of GD-QLC as a warm start for QAOA was promising. Here GD-QLC was used to build a fixed size circuit (10 layers) whose parameters then initialize QAOA training. I think practically this may be the most relevant use case of GD-QLC when one is constrained by the number of layers they can use and want to initialize the QAOA circuit in a good training regime.

Critical weaknesses

My main concern is novelty. GD-QLC is essentially a greedy layerwise gradient descent method. Each layer's parameter is initialized at zero, trained with some finite GD steps, before they freeze the layer and add a new one. Layerwise training like this is well established for VQA’s, (See [1, 2, 3, 4]). By Lemma 1 the feedback of observable A is the derivative of the energy w.r.t the new layers control parameters, so GD-QLC is just gradient descent on that parameter. Growing a QAOA circuit using commutator gradients is the core idea of ADAPT-QAOA [3] and similar to this body of work. The main difference is that GD-QLC never re-trains earlier layers, but this work doesn't cite layerwise training methods or discusses how GD-QLC relates to it and I think this is needed. Adding one such method to the benchmarks would also be valuable.

Another concern is how the benchmarks are reported. GD-QLC performs L GD steps per layer, so plotting performance only against the number of layers understates its cost relative to QLC like methods (at L = 5, 200 layers of GD-QLC require five times as many evaluations as 200 layers of FALQON). The noiseless comparisons should therefore also be plotted against total gradient steps or measurement cost. Fig. 7 partially addresses this by fixing the per-iteration shot budget but the other experiments should also include a budget style plot.

Questions for authors

Experiments use η=0.1/(l+1)log⁡k\eta = 0.1/\sqrt{(l+1) \log k} which is much larger than the bound in Theorem 1 & 2 (η<=1\eta <= 1/Smoothness constant). Do the theorems cover the reported runs? How sensitive is GD-QLC to learning rate choice?

The abstract states that feedback-based methods can mitigate barren plateaus. Could the authors support this with a citation or experiment and discuss it in relation to GD-QLC, which was proposed partly for the same purpose [1]?

[1] Skolik, Andrea, et al. "Layerwise learning for quantum neural networks." Quantum Machine Intelligence 3.1 (2021): 5.

[2] Campos, Ernesto, Aly Nasrallah, and Jacob Biamonte. "Abrupt transitions in variational quantum circuit training." Physical Review A 103.3 (2021): 032607.

[3] Zhu, Linghua, et al. "Adaptive quantum approximate optimization algorithm for solving combinatorial problems on a quantum computer." Physical Review Research 4.3 (2022): 033029.

[4] Kverne, C., Akewar, M., Huo, Y., Patel, T. & Bhimani, J.. (2026). WSBD: Freezing-Based Optimizer for Quantum Neural Networks . Proceedings of The 29th International Conference on Artificial Intelligence and Statistics, in Proceedings of Machine Learning Research 300:4150-4158 Available from https://proceedings.mlr.press/v300/kverne26a.html.

2 comments

We thank the reviewer for this insightful feedback and for highlighting the connection to established layerwise VQA training methods. Below we will address their careful comments point by point.

Weakness 1

ADAPT-QAOA uses commutator gradients to select operators from a candidate pool at each layer and subsequently re-optimizes all variational parameters globally. This leads to compounding optimization overhead relative to GD-QLC as new layers are added. In contrast, GD-QLC keeps the driver Hamiltonian HdH_d fixed and uses the commutator expectation value A(l)=⟨i[Hd,Hp]⟩A^{(l)} = \langle i[H_d, H_p] \rangle directly to update a single control variable.

While GD-QLC shares the high-level structure of freezing earlier layers, it differs from existing layerwise VQAs. Heuristic layerwise VQAs rely on off-the-shelf black-box classical optimizers (and multi-evaluation parameter-shift or finite-difference rules) without global Lipschitz smoothness bounds, Lyapunov stability conditions, or finite-shot convergence guarantees. By contrast, GD-QLC obtains exact gradients directly from the single commutator observable Ak(l)A_k^{(l)} and ensures satisfaction of the Lyapunov condition with strict monotone energy descent (Theorem 1), alongside a convergence rate in expectation to a per-layer stationary point under unbiased shot noise (Theorem 2). Relative to established layerwise VQA training, GD-QLC serves as a hybrid bridge between variational optimization and feedback-based quantum Lyapunov control.

We agree that explicitly situating GD-QLC relative to this literature will strengthen the paper, and we will add a dedicated discussion and citations covering layerwise circuit learning (Skolik et al.) and adaptive commutator-based methods such as ADAPT-QAOA (Zhu et al.).

Following the reviewer’s suggestion, we will add a comparison against a frozen layerwise QAOA baseline (parameter-fixing QAOA) in the revised manuscript to empirically highlight the trade-offs in circuit depth, measurement cost, and convergence stability.

Weakness 2

We agree that plotting against layer count alone understates GD-QLC’s cost when L>1L > 1. We will add noiseless comparisons against cumulative gradient/feedback evaluations and estimated measurement shots, counting each of the L updates and the observable groups used to estimate them. Layer-count plots will be presented as depth comparisons, while resource-matched plots will support claims about efficiency. Figure 7 addresses finite-shot performance under its stated budget convention.

Question 1

We thank the reviewer for raising the distinction between the theorem and the empirical schedule. In our empirical studies we have found that we can choose much larger learning rates for the system compared to the learning rates used in Theorem 1 & 2. We have observed in our research that GD-QLC performs well over a wide range of learning rates. We recognize that a more thorough discussion on learning rates is needed to clarify the difference between the theorems and the empirical results. In the revised paper we will include a discussion of this point and we will include a plot illustrating the effect of learning rate on performance gains.

Question 2

We are not aware of a concrete research paper demonstrating barren plateau mitigation for quantum Lyapunov control methods. But historically, QLC methods have been motivated based on this. For example, (Magann et al., 2022a) discusses feedback structure as a possible way to avoid some variational-training difficulties. These results motivate the question but do not establish barren-plateau mitigation for GD-QLC. We will add the relevant citations, and describe this as an open empirical question for GD-QLC.

We look forward to further insightful discussions regarding these points.

Thank you for addressing my review. I appreciate the updated framing and novelty discussion and your plan to add more experiments.

However as these additions are phrased as "future work" I cannot evaluate your response further until a new version of your work is uploaded. If I am not mistaken the current draft is still the old one. I'd be happy to go further over your response once your updated draft is in.

Reviewer_vq1POpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes GD-QLC to solve combinatorial optimization problems. It improves FALQON by updating the control parameter at each layer through multiple gradient descent steps. The authors compare the performance of the proposed method with common QLC-based methods and FALQON, to demonstrate that GD-QLC can achieve higher approximation ratios and success probabilities on Max-Cut problems.

Critical strengths
  • The manuscript provides theoretical analysis of the condition under which the energy is non-increasing during per-layer updates of the control parameters.

  • The experiment section discusses the multiple hyper-parameters that may affect the performance of the proposed method.

Critical weaknesses
  • The theoretical analysis lacks convergence guarantees as the number of circuit layers increases. Although the proposed method guarantees a monotonic decrease in energy under certain conditions, this property alone does not characterize the overall convergence behavior.

  • Since the update of control parameters introduces additional measurement cost, the benefit of fewer circuit layers may be weakened. The trade-off between solution quality and required measurement cost should be considered.

  • It lacks the comparison with vanilla QAOA and QAOA with layer-wise training.

Questions for authors

What are the primary differences between GD-QLC and QAOA with layer-wise training, particularly when the parameters of cost Hamiltonian are fixed?

2 comments

We thank the reviewer for their careful and insightful review. Below we will address the points raised by the reviewer on a point by point basis.

Weakness 1

Stating overall convergence guarantees is difficult in quantum control theory as the theory is not well developed and oftentimes only asymptotic convergence results can be stated. Methods in quantum Lyapunov control theory commonly state Lyapunov guarantees which only guarantee a monotonic non-increasing performance in the objective function (when the objective is formulated as a minimization).

Weakness 2

We aim to address this with the equal budget comparison in Fig 7: Given the same per-layer budget, a method is free to allocate shots either to estimating one observable really well (FALQON) or to update a parameter multiple times with less measurements (GD-QLC). In this regime, we still see GD-QLC exhibit stability and converge in fewer layers. Moreover, GD-QLC can be reduced to FALQON when taking L=1L=1, so we never do worse than the baseline.

Weakness 3

We have compared our method with vanilla QAOA in Fig 8(b). The comparison is over a 10 layer circuit in the noisy setting. The results show that in this setting, GD-QLC can achieve higher benchmarks than vanilla QAOA with random initializations. The figure further shows that using GD-QLC as a warm starting strategy for QAOA leads to even better results. We thank the reviewer for pointing out the missing comparison. Where computationally feasible, we will add layerwise-trained QAOA on the same instances, with comparable initialization, optimizer, depth, and total resource budget. We will report both solution quality and total evaluations. If these benchmarks cannot be completed in the revision period, we will acknowledge the limitation and discuss the relevant layerwise-QAOA literature.

Question 1

GD-QLC can be viewed as a continuous layer-wise training scheme that bridges quantum control with the notion of layer-wise training. Under the discretization scheme discussed in the paper, GD-QLC performs layerwise training on a QAOA-like circuit. Because of this connection, we can utilize tools such as the Lyapunov guarantee (Theorem 1) for our method. Layerwise training is mainly defined for VQA's where each layer consists of a parameterized gate for which we can use parameter shift rule. To the best of our knowledge, this feedback-based layerwise scheme with Lyapunov guarantees has not been studied in the literature before.

We will include a discussion on layerwise training and its connection to GD-QLC including relevant citations for prior layerwise-circuit and iterative layerwise-QAOA work in the revised paper.

Please let us know if you would like to discuss any of these points further.

We thank the authors for their responses and clarifications.

W1. I understand that establishing overall convergence guarantees is challenging in quantum control theory. But the current theoretical analysis still does not characterize the convergence behavior as the number of circuit layers increases, which I think weakens the theoretical contribution of this work.

W2. The response addresses my concern.

W3 and Q1. The authors clarify the connection between GD-QLC and layer-wise QAOA, but the primary methodological differences still remain unclear. A direct comparison with layer-wise QAOA is necessary.

Besides, the manuscript has not been updated yet.

Overall, I do not think the response completely addresses my concerns regarding the methodological novelty of the proposed approach.

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

This paper asks the following questions: Can we improve feedback-based quantum optimization by choosing each new circuit parameter more carefully so that fewer circuit layers are needed?

More precisely, the authors consider min_{||psi||=1} <psi | H_p | psi >, where H_p encodes a combinatorial problem and its minimum eigenvalue E_0 is optimal.

The authors' goal is to improve better solution quality with fewer layers, compared with existing feedback-based algorithms such as FALQON, FOCQS, etc.

This question is important because circuit depth in quantum computing constitutes a practical limitation of feedback-based quantum optimization (and other algorithms) as more layers require longer coherent execution. A convincing improvement would potentially facilitate the implementation of these algorithms.

Critical strengths
  1. The authors provide experimental supports for potentially useful improvements relative to the benchmarks mentioned above.
  2. GD-QLC achieves comparable solution quality with fewer layers than the feedback-based baselines across several optimization tasks.
  3. Fig. 7 reports gains under equal per-iteration shot budgets, supporting its benefits under limited measurement resources.
Critical weaknesses
  1. The baseline comparisons need validation. Reported FALQON controls exceed 4000 whereas the stated setup implies |beta_k|<= 6n <=108. So, the stability advantage is not clear.
  2. FOCQS already proposes repeated gradient refinement. On the other hand, the theoretical guarantees do not show acceleration over other methods. Thus, the contribution depends heavily on empirical comparisons, but as indicated in 1), the advantages aren't clear.
Questions for authors
  1. How the reported FALQON controls above 4000 reconcile with the bound |beta_k|<=6n discussed earlier?
  2. What does GDQLC add beyond FOCQS repeated gradient refinement?
  3. Please reconcile Fig. 8 shot counts and circuit depths and provide enough information to reproduce the experiments.
  4. Do GDQLC gains persist when each baseline is reasonably tuned and all measurements are accounted for? Is this how the experiment is done?
  5. Minor: Algorithm 1: the stepsize divides by log k, although k starts at 1. The experimental stepsizes also exceed the theorem’s bound.
  6. Appendix C, around eq (30) the gradient identity is false for general noncommuting control gates, please check.
1 comment

We thank the reviewer for their insightful review. Below we will present our responses to the reviewer's questions broken down question by question. We aim to also address weakness 1 with question 1 and weakness 2 with question 2.

Question 1 and weakness 1

The statement at line 325 is our mistake: we meant to state the magnitude claim only about SO-FALQON, as FALQON clearly does not exhibit such large magnitude swings. FALQON typically exhibits instability through rapid oscillations as seen in Figure 2 (middle panel). We will fix this typo in the revised paper.

Question 2 and weakness 2

FOCQS's update rule updates every parameter every iteration, while our approach is layerwise and freezes previous layers. FOCQS's strong point is that it does not need to remeasure the previous layers and only uses the stored classical output history to update the parameters at every step, and so it builds a better approximation of β\beta as it adds more layers. On the other hand, FOCQS requires updating the entire circuit per added layer as all parameters are updated at once. In contrast, GD-QLC freezes previous layers and only updates one parameter at a time, avoiding the need to retune the entire circuit at each step. This can be beneficial from a practical point of view as re-tuning quantum gates can be costly especially in the near-term NISQ era. When we compare FOCQS to GD-QLC, in the noiseless setting they both perform roughly the same (Fig 2), but in the noisy setting our approach seems to perform better as FOCQS's performance degrades over time while our approach stays stable (Fig 7).

Question 3

We thank the reviewer for pointing out this discrepancy. The stated budgets in Figure 8(b) (100 and 1000) are the correct values and we will fix the typo in the figure caption. Figure 8(a) uses 100 shots as noted in the paper. The discrepancy in layer count between Figure 8(a) (30 layers) and Figure 8(b) (10 layers) is due to computational constraints related to computing the gradient vector for QAOA. We will provide all code in the supplamentary material and will revise the paper with more comprehensive experimental details.

Question 4

Wherever applicable we have referred to the original paper for any choice of hyperparameters. The main tunable parameter shared between all methods is the timestep Δt\Delta t for which we have provided experiments under multiple choices, and these timesteps are within the common range of choices in the literature. For FOCQS, we have used β0=1\beta_0=1 as the default value consistent with the original paper. We have tried to implement every method faithfully and have taken care to ensure our implemented code is not affecting the baseline methods negatively.

Question 5

We thank the reviewer for raising the distinction between the theorem and the empirical schedule. We add a small linear shift to the log function in order to avoid zeros in the logarithm. In the implementation we use log⁡10(k+10)\log_{10}(k+10) to normalize the log to 1 at k=0k=0. We will clarify this in the algorithm description.

The larger rates used in some experiments exceed the sufficient bound for smoothness and are supported empirically instead of Theorem 1. We have observed in our research that GD-QLC performs well over a wide range of learning rates. We recognize that a more thorough discussion on learning rates is needed to clarify the difference between the theorems and the empirical results. In the revised paper we will include a discussion of this point and we will include a plot illustrating the effect of learning rate on performance gains.

Question 6

Thank you for your correction. The definition of A(t)A(t) has to be updated to reflect the correct gradient identity when multiple non-commuting gates are involved and equation (30) must be replaced with the derivative obtained by differentiating that ordered product. We will correct this calculation in the revised paper. Short outline of the correction to equation (30): Assuming the rightmost gate acts first, for a single layer we write the overall unitary as Uk=Ud,J⋯Ud,0UpU_k=U_{d,J}\cdots U_{d,0}U_p. For parameter βk,j\beta_{k,j}, define the state just before its gate by ∣ϕk,j⟩=Ud,j−1⋯Ud,0Up∣ψk−1⟩|\phi_{k,j}\rangle=U_{d,j-1}\cdots U_{d,0}U_p|\psi_{k-1}\rangle, the state after it by ∣χk,j⟩=Ud,j∣ϕk,j⟩|\chi_{k,j}\rangle=U_{d,j}|\phi_{k,j}\rangle, and the later-gate product by Vd,j=Ud,J⋯Ud,j+1V_{d,j}=U_{d,J}\cdots U_{d,j+1}. Replace the commutator in equation (30) by the feedback Ak,j=⟨χk,j∣i[Hd,j,Vk,j†HpVk,j]∣χk,j⟩.A_{k,j}=\langle\chi_{k,j}|i[H_{d,j},V_{k,j}^{\dagger}H_pV_{k,j}]|\chi_{k,j}\rangle. Then ∂Ek/∂βk,j=ΔtAk,j\partial E_k/\partial\beta_{k,j}=\Delta t A_{k,j}, so for the scaled objective fk=Ekf_k=E_k, ∇fk=/ΔtA⃗k\nabla f_k=/\Delta t\vec A_k and equation (31) becomes β⃗k(ℓ+1)=β⃗k(ℓ)−ηk,ℓA⃗k(ℓ)\vec \beta_k^{(\ell+1)}=\vec\beta_k^{(\ell)}-\eta_{k,\ell}\vec A_k^{(\ell)}. We will then state a multiparameter smoothness bound and state a per-layer descent claim.

We hope we have addressed your concerns and we look forward to discussing this further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.