Statistical Analysis of Quantum State Learning Process in Quantum Neural Networks

Summary

This paper describes a formalization and experimental verification of the thesis that a quantum state $\vert\phi\rangle$ is the local minimum in the process of the QNN training. The paper is well written and provides all necessary support documents for its understanding. This work is an extension and is complementary to the work on the observed plateaus in the learning landscape during the training of QNN. The most interesting finding is that the probability that a state $\vert\phi\rangle$ being the local minimum is inversely proportional to the number of qubits and proportionally growing with the depth of the QNN (number of layers in the network). This seems to be in opposition with the original work "Barren plateaus in quantum neural network training landscapes" where the learnability decreases with the number of qubits in the state. The formal description seems to uphold the hypothesis.

Strengths

- Problem description and formalization - The experimental verification of the observed phenomenon

Weaknesses

- The main weakness is the significance of the result. While the result is interesting and proven, it is a bit expected. In particular, the result shows us that with increasing embedding i.e. the number of parameters the representational power increases while it also vanishes with increasing number of qubits. So the first conclusion is not surprising and the second conclusion seems to follow the previous works. However, more importantly, because this papers concerns an unknown state, a conclusion should be drawn if this effect can be avoided at all. Because independently of this initial setting an unknown state can occur.

Questions

- What are the more formal conclusions beyond a simple better initialization or problem aware initialization? Because your paper concerns a quite serious issue a p@rediciable discussion should be provided. - Is there a ratio between the depth $d$ and $n$ so that the occurrence of the local minima is minimized? What is the order of such ratio?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

- The paper should be better discussed as for the consequences and unique conclusions. While there is a considerable amount of explanation I feel the authors failed to provide details on how to avoid the observed effect

Summary

The authors present a no-go theorem that reveals the limitations of learning unknown quantum states using QNNs, even with high-quality initial states. They prove that the probability of avoiding local minima decreases exponentially with the number of qubits but grows polynomially with circuit depth. The curvature of local minima is determined by the quantum Fisher information and a loss-dependent constant. These findings provide insights into the role of prior information and the scalability of QNNs, impacting their learnability and effectiveness.

Strengths

The work explores the trainability of quantum neural networks. The theoretical findings give the limits on the learnability of QNN in general cases which provide some insight into the development of QNN in future studies.

Questions

1. In this work, it focus on learning pure states through QNN, and whether the statement is also true for the mixed states. 2. Whether the no-go theorem is also true for other loss functions, such as using another metric as the distance in loss? 3. In the numerical experiments, how to calculate the probability $Pr_{\mathbb{T}}[LocalMin(\theta^*,\epsilon)]$?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Summary

The paper studies parameterized quantum circuits (aka QNNs). These architectures face trainability issues (e.g. barren plateaus) as the number of qubit grows, and several approaches have been explored to mitigate these. The paper analyzes these strategies using the task of training a circuit to transform |0> input state into a desired output state, and provides theoretical and numerical evidence of difficulty of training the circuit.

Strengths

The paper adds new, original result to an important, open problem of training QNNs. The result provide scaling law of the probability of avoiding local minima irrespective of techniques such as special initialization strategy. Importantly, the result explicitly involves the precision parameter, which is one of crucial differentiators between classical networks and QNNs that necessarily use quantum measurement. Theoretical results are followed with extensive numerical simulations confirming the results in practice.

Weaknesses

The paper is focused on a specific loss, specific task, which leads e.g. to a specific type of dependence of local minima on parameter count (e.g. discussion in lines 206-209 on pg. 6). It is not fully clear how insights from this task translate to other losses/tasks. The result showing that the number of local minima decreases with the expressibility of the circuit is in line with previous work, but those earlier results are not discussed in detail in the discussion. E.g., Larocca et al. Theory of overparametrization in quantum neural networks. 2021 [Ref 57] is cited in the introduction but not in discussion.

Questions

Discussion in 3.3 mentions complementarity of the abundance of local minima and barren plateaus on the expressibility axis. How do recent results on mitigating barren plateaus via architectural choices instead of initialization (e.g. Wang et al., ICLR'23) affect this understanding?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

There does not seem to be negative social impact of this theoretical research.

Summary

This paper investigates the learnability of the QNN in the task of quantum state learning from a statistical perspective. The paper develops a no-go theorem that proves that when the loss function value is lower than a critical threshold, the probability of avoiding local minima decreases exponentially with the number of qubits, while only increasing polynomially with the circuit depth. Moreover, the paper conducts some numerical experiments to validate the proposed theorem.

Strengths

1. The paper studies a novel research problem, namely the limitation of quantum neural networks in quantum state learning tasks, and analyzes the influence of loss function value information on training difficulty from a statistical perspective, which has certain value for understanding and improving the principles and methods of quantum machine learning. 2. It provides a rigorous and quantitative theoretical analysis showing that when the loss function value is below a critical threshold, the probability of avoiding local minimum decay exponentially with the number of qubits, but only grows polynomially with the circuit depth, revealing the tradeoff between learnability and scalability of QNNs.

Weaknesses

1. The theoretical analysis is only applicable to pure state learning tasks, and in fact, mixed states or noisy states may be encountered in quantum machine learning, so the conclusions and methods may need further generalization and verification. 2. Numerical experiments only use one kind of QNN but do not consider other possible circuit structures and parameterization methods, the results may have certain biases and limitations.

Questions

1. Why choose the ALT structure as the research study in the numerical experiment? ALT in fact provably does not suffer from the vanishing gradient problem, and does this affect the experimental results as well as the theoretical proof? 2. The abstract mentioned that "the results hold for any circuit structure", and is there any theoretical or experimental proof of this?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The paper discusses the limitations and addresses them.

Summary

This paper introduces a new statistical analysis for training variational quantum circuits (e.g., in terms of quantum neural networks), where the alternating-layered ansatz (Nakaji et al.; ALT), also referred to as the entanglement circuit for quantum ML in Chen et al. 2019 [4], has been characterized as a general quantum state learning task. In general, the presentation flow is quite good, including a proper introduction to the $l_p$ norm and Dirac notations. The authors have made considerable efforts to guide the readers from the existing learning definition in vector-to-vector mapping to standard encoding based parameterized quantum state learning. Although some recent works on error analysis in quantum circuit learning [1] and classical encoding circuits [2,3] are unfortunately omitted, the paper actually provides a careful and detailed review of related work in the appendix. In general, while the theorem is neat, moving from elaborating fidelity loss to fisher information based bound analysis, this paper also conducts a solid local minima analysis. Despite the fact that learnability is considered a no-go perspective and some related work (e.g., [1]) is missing, I believe the theoretical findings and its supporting numerical results conducted good takeaways to the community. In general, I tend to accept this paper. *** **References** 1. "Theoretical error performance analysis for variational quantum circuit based functional regression." J Qi et al. npj Quantum Information 9.1 (2023):. Nature 2. "Quantum Circuit Learning," K. Mitarai et al., Physical Review A 98.3 (2018): 032309. 3. "Quantum machine learning in feature hilbert spaces, M. Schuld, Physical Review Letters" 4. "Variational Quantum Circuits for Deep Reinforcement Learning" 2019

Strengths

- The paper provides a clear characterization of the curvature of local minima, which is important to understand the sensitivity of output state with respect to QNN (VQC learning) parameters. - It provides quantitative limits on good initial guesses related to no free lunch (NFL) theories and adaptive methods for improving the learnability and scalability of QNNs. - Good presentation quality. - The paper suggests that no single QNN is universally the best-performing model for learning all target quantum states. This introduces additional complexity for practical applications as it may necessitate more structured QNN architectures and innovative optimization tools.

Weaknesses

- the ensemble setting in Appendix A. 1 is not clear on the motivation of using model ensemble. - despite the results, there is a level of uncertainty remaining as the exact scaling of QNN depth needed to form a subspace 2-design is not very clear

Questions

1. the ensemble setting in Appendix A. 1 is not clear on the motivation of using model ensemble.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

- The no-go theorem, while important for understanding the limitations of QNNs, might be a potential barrier to the application of quantum neural networks in real-world scenarios. - The paper implies that significant future progress will be needed, potentially borrowing insights from the field of deep learning, to overcome the limitations of current QNNs.