Quantum Algorithms for Non-smooth Non-convex Optimization

This paper considers the problem for finding the $(\delta,\epsilon)$-Goldstein stationary point of Lipschitz continuous objective, which is a rich function class to cover a great number of important applications. We construct a zeroth-order quantum estimator for the gradient of the smoothed surrogate. Based on such estimator, we propose a novel quantum algorithm that achieves a query complexity of $\tilde{\mathcal{O}}(d^{3/2}\delta^{-1}\epsilon^{-3})$ on the stochastic function value oracle, where $d$ is the dimension of the problem. We also enhance the query complexity to $\tilde{\mathcal{O}}(d^{3/2}\delta^{-1}\epsilon^{-7/3})$ by introducing a variance reduction variant. Our findings demonstrate the clear advantages of utilizing quantum techniques for non-convex non-smooth optimization, as they outperform the optimal classical methods on the dependency of $\epsilon$ by a factor of $\epsilon^{-2/3}$.

Paper

Similar papers

Peer review

Reviewer RNco6/10 · confidence 1/52024-07-01

Summary

This paper considers using quantum methods for stochastic optimization, using zeroth order queries. It looks like the main idea is that using quantum methods, one can summarize over finite difference calculations quickly and efficiently, to arrive at approximate subgradients efficiently; this would usually be very inefficient for classical methods. Overall, they are able to show speedup, from $O(\epsilon^{-4})$ to $O(\epsilon^{-3})$.

Strengths

The problem is well contained and the premise is believable. The classical optimization bits looks reasonable, and the results make sense. I skimmed through the appendix, and the classical optimization parts are reasonable.

Weaknesses

Section 3 is a bit hard to follow. The specific speedup offered by the quantum method is not entirely clear, though it is likely coming from Theorem B1. Perhaps a deeper discussion of this, and why this quantum speedup exists (e.g. is it a consequence of Deusch Josza? Can you provide a more complete argument for where the speedup appears? )

Questions

(minor) Why do you say that finding a Clarke subdifferential is harder than finding a smooth differential? Generally speaking the complexities are comparable.

Rating

6

Confidence

1

Soundness

3

Presentation

3

Contribution

3

Limitations

no societal limitations

Reviewer mFHz6/10 · confidence 4/52024-07-11

Summary

This paper investigates quantum algorithms for finding the $(\delta,\epsilon)$-Goldstein stationary point of a potentially nonconvex and nonsmooth objective function $f$. Utilizing quantum variance reduction techniques as outlined in [42], the authors have developed a zeroth-order quantum estimator for the gradient of the smoothed surrogate of $f$. The stationary point of this smoothed surrogate is also the Goldstein stationary point of $f$ when using an appropriate smoothing parameter $\delta$. Leveraging this zeroth-order quantum estimator, the authors propose two algorithms, QGFM and QGFM+, to find the Goldstein stationary point, achieving a quantum speedup on the order of $\epsilon^{-2/3}$. Additionally, the QGFM+ framework adjusts the variance level during each variance reduction step, providing further acceleration to the Q-SPIDER algorithm described in [42] for smooth nonconvex optimization.

Strengths

This paper initiates the study of quantum algorithms for finding Goldstein stationary points, a significant problem in continuous optimization. Additionally, the authors present an explicit construction of the quantum sampling oracle using the quantum zeroth-order oracle, including a detailed discussion on the number of qubits required.

Weaknesses

Despite the detailed implementation and calculations, the overall technical approach remains relatively straightforward. The zeroth-order quantum estimator combines the classical stochastic gradient estimator for the smoothed surrogate with the quantum variance reduction algorithm in [42]. The quantum algorithms for finding the Goldstein stationary point are obtained by replacing the classical estimators with quantum estimators. Moreover, the narrative is somewhat incomplete due to the absence of lower bound results.

Questions

Is it possible to improve the $\delta$ dependence using quantum algorithms? Minor issues: 1. Consistency of big-O notation. For example, $O$ is used in line 139 and $\mathcal{O}$ in line 183. Similarly, there are consistency issues with the quantum oracle notation, where $\mathcal{O}$ is used in line 168 and $\mathbf{O}$ in line 184. 2. Typo on the RHS of the inequality in line 125. 3. The use of dashes '-' is a bit odd. For example, the dashes in line 139, line 210, and line 251 can be removed. 4. The name initials in the citation format are not precise. For example, in entry [1], it should be 'G. Arfken' instead of 'G Arfken'. 5. Line 310: "adjust" -> "adjusts". Line 311: "fixed" -> "fixes".

Rating

6

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

N/A

Reviewer hDMr7/10 · confidence 3/52024-07-12

Summary

This paper studies quantum algorithm for non-smooth non-convex stochastic optimization with zeroth-order oracle. It introduces an effective quantum estimator that reduces the variance compared to classical zeroth-order estimators. Upon substituting this estimator into known zeroth-order non-smooth optimizers, namely GFM and GDM+, the resulting quantum optimizer achieves improved rate $\tilde O(d^{3/2}\delta^{-1}\epsilon^{-3})$ and $\tilde O(d^{3/2}\delta^{-1}\epsilon^{-7/3})$ respectively for finding a $(\delta,\epsilon)$-Goldstein stationary point. Notably, quantum speedup improves upon the classical lower bound $\delta^{-1}\epsilon^{-3}$ by a factor of $\epsilon^{2/3}$. Moreover, a modified algorithm achieves $O(\sqrt{d}\epsilon^{-7/3})$ for smooth optimization, improving upon the best known rate.

Strengths

This paper proposes a new zeroth-order quantum estimator. This leads to new quantum algorithms that solves zeroth-order non-smooth non-convex optimization problem, which is not well studied in the literature. Moreover, the proposed algorithms show quantum speedup compared to their classical (non-quantum) counterparts. Notably, it improves over the classical lower bound of $\Omega(\delta^{-1}\epsilon^{-3})$ by a factor of $\epsilon^{2/3}$. Overall, these results represent a significant contribution to the understanding of optimization with quantum oracles. Given my expertise lies primarily in optimization and not in quantum computation, I am only able to assess the optimization-related aspects of this work.

Weaknesses

Although the dependence on $\delta,\epsilon$ is improved, the dimension dependence is suboptimal. In particular, since GFM and GFM+ are known to have suboptimal dimension dependence $d^{3/2}$, so do QGFM and QGFM+. On the other hand, as observed by Kornowsky and Shamir [1], optimizing the random smoothing $f_\delta$ with a non-smooth optimizer, such as online-to-non-convex (o2nc) [2], eliminates this $\sqrt{d}$ factor and achieves $O(d)$ in dimension. Hence, my intuition suggests that upon substituting the quantum estimator into o2nc and following a similar approach to Kornowsky and Shamir, the authors might be able to recover $O(d)$ (or even better) dimension dependence. [1] Kornowski, G. and Shamir, O., “An Algorithm with Optimal Dimension-Dependence for Zero-Order Nonsmooth Nonconvex Stochastic Optimization”, 2023. doi:10.48550/arXiv.2307.04504. [2] Cutkosky, A., Mehta, H., and Orabona, F., “Optimal Stochastic Non-smooth Non-convex Optimization through Online-to-Non-convex Conversion”, 2023. doi:10.48550/arXiv.2302.03775.

Questions

- As someone unfamiliar with quantum computation, I have a general question: Is the proposed quantum oracle practically feasible to implement, or is it purely theoretical? - line 87: does state $|i\rangle$ denote the $i$-th orthonormal basis of $\mathcal{H}^m$? - line 100: what does it mean by $|\mathbf{x}\rangle |q\rangle$? Is it a shorthand for tensor product? - Thm 3.4 part 1: should it be $Var(\hat g) \le \hat\sigma_1^2$ instead of $\hat \sigma_1$? part 2: number of queries should be $\frac{d^{3/2}L\\|y-x\\|}{\delta \hat\sigma_2}$ (i.e., currently it's missing $1/\delta$)? Since this theorem is the main result of the quantum oracle, I encourage the authors to carefully check its correctness. also in the proof (line 471): $\sigma_1^2$ => $\hat\sigma_1^2$? Minor comments: - line 98: $C_{f(x)} = f(x)$ => $C_f(x) = f(x)$? - Proposition 2.1: the properties of smooth surrogate $f_\delta$ are known in [1] and [2], and Lin et. al. and Chen et. al. are restating these results in their papers. Hence, these should be more appropriate references. [1] Yousefian, F., Nedić, A., and Shanbhag, U. V., “On Stochastic Gradient and Subgradient Methods with Adaptive Steplength Sequences”, 2011. doi:10.48550/arXiv.1105.4549. [2] Duchi, J. C., Bartlett, P. L., and Wainwright, M. J., “Randomized Smoothing for Stochastic Optimization”, 2011. doi:10.48550/arXiv.1103.4296.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer KzCk6/10 · confidence 4/52024-07-13

Summary

This paper introduces new quantum algorithms for non-smooth non-convex optimization problems. The authors propose a quantum gradient estimator for smoothed objectives and develop the Quantum Gradient-Free Method (QGFM) and its enhanced version, QGFM+, which achieve better query complexities than their classical counterparts. These complexities demonstrate a marked quantum speedup over classical counterparts, indicating the potential of quantum computing in optimizing complex functions more efficiently. The paper also discusses the construction of quantum oracles and the application of variance reduction techniques, paving the way for future research in quantum optimization.

Strengths

- The paper proposed new zeroth order quantum optimization algorithms achieving better computational complexities compared to classical methods for non-smooth and non-convex optimization. - Technically, they construct efficient quantum gradient estimators and quantum superpositions over required distributions as a key subroutine. - They also proposed a quantum algorithm for non-convex smooth problems with an adaptive variance level, accelerating prior quantum algorithms to get more speedups.

Weaknesses

- The assumptions of having a quantum stochastic function value oracle may be strong. Could the authors explain more about why it is reasonable and important to have such a function oracle? - The technical core for quantum speedups seems to be the quantum mean value estimation procedure, which is already used in many other optimization problems and scenarios. Could the authors explain more about the technical novelty of their work?

Questions

Besides the questions raised in the weakness part, I have some minor issues with the submission as follows: - In line 89, the definition of the tensor product may be a little confusing. - In the explicit construction of quantum sampling oracles, it seems that the time complexity of the quantum algorithm may be much larger than the query complexity, due to the sampling on the unit sphere. However, for such optimization algorithms, time complexity may be more crucial in real-world applications. Could the authors state the actual time complexity of their algorithm in terms of gate counts?

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

The quantum complexity lower bound on this problem is not proved in this paper, which is mentioned in the conclusion part. Also, as noted in remark 3.7, implementing quantum sample oracle may require the uses of QRAM, which is currently limited by the physical realizations. This is a theoretical work, so there is no potential negative societal impact of their work.

Reviewer KzCk2024-08-10

Thank the authors for the detailed response! Since the quantum upper bound provided in the paper is $\tilde{\mathcal{O}}(d^{3/2}\delta^{-1}\epsilon^{-7/3})$, while a possible classical upper bound is $\mathcal{O}(d\delta^{-1}\epsilon^{-3})$ the classical lower bound is $\delta^{-1}\epsilon^{-3}$, does it mean that quantum speedup can only occur when the requirement of the precision is relatively high compared with the dimension, given the dimension fixed? If so, regarding current physical realization constraints like noises in quantum devices, it seems that this result is more of a theoretical one with few potential applications. Another problem is the time/gate complexity part. Thank the authors for their effort to analyze and calculate the gate cost in great detail of constructing one quantum estimated gradient oracle. However, it seems that for one call of the gradient oracle, the cost is at least proportional to the dimension, which is relatively expensive for some problems and may be a major concern in solving them. Due to the above reasons, I'll keep my rating currently and keep in mind your detailed comments for future discussions.

Authorsrebuttal2024-08-10

Thanks for your response and the follow-up questions. We present detailed answers as follows. > Since the quantum upper bound provided in the paper is $d^{3/2}\delta^{-1}\epsilon^{-7/3}$, while a possible classical upper bound is $d\delta^{-1}\delta^{-3}$ the classical lower bound is $\delta^{-1}\epsilon^{-3}$, does it mean that quantum speedup can only occur when the requirement of the precision is relatively high compared with the dimension, given the dimension fixed? Yes, the quantum speedup occurs when $d$ and $\epsilon$ satisfies that $d\leq \epsilon^{-4/3}$. We want to point out it is very common to show the quantum speed-up under the case that the dimension is not too large in non-convex optimization. For the smooth case, there is NO quantum speedup for finding $\epsilon$-stationary point with stochastic gradient inputs when the dimension is large such that $d\geq \epsilon^{-3}$ [Proposition D.12, A]. Besides, [41] presents a quantum upper bound of $\epsilon^{-5/2}\sqrt{d}$ for smooth non-convex optimization, where the quantum speedup only occurs when $d\leq \epsilon^{-1}$. The function class considered in this paper can be non-smooth, which is more general than the previous works. Thus we think the speed-up region $d\leq \epsilon^{-4/3}$ obtained in this paper is reasonable. > If so, regarding current physical realization constraints like noises in quantum devices, it seems that this result is more of a theoretical one with few potential applications. We would like to point out that the device's noise mentioned by you can be taken care of by quantum error correction code [B] such as surface code. On the other hand, the problem's dimension $d$ is limited by the scale of existing quantum computers and thus cannot be very large. Hence we think the quantum speedup region $d\leq \epsilon^{-4/3}$ is not that strict. Though the number of precise qubits is still limited by the current technique, the quantum community believes that getting more enough precise qubits is just a matter of time [C]. > Another problem is the time/gate complexity part. Thank the authors for their effort to analyze and calculate the gate cost in great detail of constructing one quantum estimated gradient oracle. However, it seems that for one call of the gradient oracle, the cost is at least proportional to the dimension $d$, which is relatively expensive for some problems and may be a major concern in solving them. '' Thanks for acknowledging our effort! We think the $d$-dependency computation cost for one call of the quantum estimated gradient oracle is unavoidable and should not be regarded as a major concern. Evaluating a function value at a $d$-dimension input will at least take $\mathcal{O}(d)$ time to even just read the input. Hence, our construction of the quantum estimated gradient oracle doesn't introduce much more cost than necessary, and it is natural to focus on the query complexities on stochastic function value oracle, which usually is the dominant part in the computation. For instance, even for a simple function with the form $f({\bf x}) = {\bf x}^{\top}{\bf A}({\bf x}){\bf x}$, where ${\bf A}({\bf x})\in R^{d\times d}$, its computation cost of a given $\bf x$ will be at least proportional to $d^2$, which is much larger than $\mathcal{O}(d)$ for constructing one quantum estimated gradient oracle. **References:** [A]. Chenyi Zhang, and Tongyang Li. Quantum Lower Bounds for Finding Stationary Points of Nonconvex Functions. ICML, 2023. [B]. Roffe J. Quantum error correction: an introductory guide. Contemporary Physics, 2019. [C] Gambetta, Jay M., Jerry M. Chow, and Matthias Steffen. Building logical qubits in a superconducting quantum computing system. npj quantum information, 2017. We hope our response can address your concern and are happy to answer any further questions.

Reviewer KzCk2024-08-11

Thank you for the detailed response! It has addressed all of my questions, and I adjusted the score accordingly.

Reviewer hDMr2024-08-09

Thank the authors for their detailed response. Regarding weakness, thanks for providing the additional background. Now I understand improving dimension dependence is a non-trivial challenge. Given this, I adjusted my score accordingly. As a quick follow-up question, the authors mentioned that quantum estimators can only estimate one-point stochastic function value $f(x,z)$, but not 2-point function values nor stochastic gradients. Is this limitation proven impossible in the quantum computing literature, or is it still an open problem?

Authorsrebuttal2024-08-09

Thanks for your positive comments and raising your score. We present the answer for your follow-up question as follows: We do not mean that ‘’quantum estimators can only estimate one-point stochastic function value $f(x,z)$, but not 2-point function values nor stochastic gradients''. In fact, we can use quantum oracles to construct gradient estimators by $2$-queries of stochastic function value oracle (see our result in Theorem 3.2) or a query of stochastic gradient oracle [42]. However, the speed-up by the quantum mean estimators requires to use mini-batch queries of stochastic function value oracles or mini-batch queries of stochastic gradient oracles to reduce oracle calls with a given variance level (see our Theorem 3.4 and Remark 3.5). Hence, whether one can construct the stochastic gradient estimators by $2$-queries (or constant queries) of stochastic function value oracle or a query of stochastic gradient oracle (as in the algorithms of [30, 41]) while still maintain the quantum speed-up remains an open problem. We hope this answers your following-up question and are happy to address any further concern.

Reviewer hDMr2024-08-10

Thank you to the authors for their thorough follow-up response. It has addressed all of my questions, and I have no further questions.

Reviewer mFHz2024-08-11

Thank you to the authors for their detailed response. I remain my rating

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC