Abstract:Quantum neural networks (QNNs) have been a promising framework in pursuing near-term quantum advantage in various fields, where many applications can be viewed as learning a quantum state that encodes useful data. As a quantum analog of probability distribution learning, quantum state learning is theoretically and practically essential in quantum machine learning. In this paper, we develop a no-go theorem for learning an unknown quantum state with QNNs even starting from a high-fidelity initial state. We prove that when the loss value is lower than a critical threshold, the probability of avoiding local minima vanishes exponentially with the qubit count, while only grows polynomially with the circuit depth. The curvature of local minima is concentrated to the quantum Fisher information times a loss-dependent constant, which characterizes the sensibility of the output state with respect to parameters in QNNs. These results hold for any circuit structures, initialization strategies, and work for both fixed ansatzes and adaptive methods. Extensive numerical simulations are performed to validate our theoretical results. Our findings place generic limits on good initial guesses and adaptive methods for improving the learnability and scalability of QNNs, and deepen the understanding of prior information's role in QNNs.

What problem does this paper attempt to address?

This paper attempts to solve the local minimum problem encountered by quantum neural networks (QNNs) in the process of learning unknown quantum states. Specifically, the author studies, through statistical analysis methods, the phenomenon that when the loss function value is below a certain critical threshold, the probability of QNNs avoiding local minima in learning unknown quantum states decreases exponentially with the number of qubits. This problem is of great significance for understanding the training difficulty and learnability in quantum machine learning. ### Main contributions of the paper 1. **No - solution theorem**: - Proved a no - solution theorem, pointing out that in the process of learning unknown quantum states, if the loss function value is below a certain critical threshold \( L_c = 1-\frac{1}{d} \), the probability of avoiding local minima decreases exponentially with the number of qubits \( N \) and increases polynomially with the circuit depth \( D \). Specifically, the probability is \( O\left(\frac{N^2 2^{-ND}}{\epsilon^2}\right) \). - The curvature of local minima is concentrated on the quantum Fisher information (QFI) multiplied by a constant related to the loss value. 2. **Numerical experiments**: - Conducted extensive numerical experiments to verify the theoretical results. The experiments include: - Comparing the theoretical boundaries with the actual loss curves to show the predictive ability of the theory on the statistical behavior of the actual training process. - Sampling the loss landscape contour map to visualize the existence of asymptotic local minima. - Calculating the gradients and diagonalizing the Hessian matrix to directly verify the correctness of the theoretical boundaries. 3. **Limitations on the learnability of QNNs**: - The results pose general limitations on the learnability of QNNs, especially for those training strategies that are not within the scope of randomly initialized deep QNNs (such as high - fidelity initial guesses, initialization heuristics close to the identity matrix, adaptive and layer - by - layer training methods). - Provide a theoretical basis for introducing prior information into QNN design, providing guidance for the future development of QNNs. ### Related work - **Wilderness phenomenon**: First discovered by [36], pointing out that the gradient variance of randomly initialized QNNs decreases exponentially as the system scale increases. - **Other related research**: [37] studied the wilderness phenomenon under the local observable loss function; [51] proved the general NP - hard problem of training QNNs; [52] introduced the wilderness phenomenon from the perspective of uncertainty; [54] and [55] pointed out that QNNs are also affected by local minima in some cases. ### Mathematical symbols and basic knowledge - Use \( \| \cdot \|_p \) to represent the \( l_p \) norm of vectors and the Schatten \( p \) norm of matrices. - The basic unit of a qubit is a quantum bit or qubit, and a single - qubit pure state is described by a unit vector in the Hilbert space \( \mathbb{C}^2 \). - Common single - qubit rotation gates include \( R_x(\theta)=e^{-i\theta X/2} \), \( R_y(\theta)=e^{-i\theta Y/2} \), \( R_z(\theta)=e^{-i\theta Z/2} \), where \( X, Y, Z \) are Pauli matrices. - A quantum neural network (QNNs) usually refers to a parameterized quantum circuit \( U(\theta) \), and the parameter \( \theta \) can be trained by a classical optimizer based on the results of quantum measurements. ### Paper structure 1. **Introduction**: Introduce the research background and motivation. 2. **Fundamentals and symbols of quantum computing**: Define basic quantum computing concepts and symbols. 3. **Statistical characteristics of quantum state learning**: Analyze in detail the statistical characteristics of QNNs in quantum state learning tasks. 4. **Numerical experiments**: Verify the theoretical results through experiments. 5. **Conclusion**: Summarize the research contributions and future work directions. ### Conclusion This paper reveals the training difficulty of QNNs in learning unknown quantum states through statistical analysis methods and provides a theory for designing more effective QNNs.

Statistical Analysis of Quantum State Learning Process in Quantum Neural Networks

Predictive Modelling of Quantum Process with Neural Networks

Quantum neural networks with deep residual learning

Quantum Algorithms for Feedforward Neural Networks

Toward Trainability of Deep Quantum Neural Networks

Infinite Neural Network Quantum States: Entanglement and Training Dynamics

Quantum-data-driven dynamical transition in quantum learning

Learning To Optimize Quantum Neural Network Without Gradients

QuST: Optimizing Quantum Neural Network Against Spatial and Temporal Noise Biases

A Unified Theory of Quantum Neural Network Loss Landscapes

The Dilemma of Quantum Neural Networks

A Study of Quantum Neural Networks

Deep quantum neural networks form Gaussian processes

Analyzing Convergence in Quantum Neural Networks: Deviations from Neural Tangent Kernels

Quantum Generative Training Using Rényi Divergences

Absence of Barren Plateaus in Quantum Convolutional Neural Networks

Coherent Feed Forward Quantum Neural Network

Geometry of learning neural quantum states

Toward Trainability of Quantum Neural Networks

Quantum Neural Network and Its Applications

State Classification via a Random-Walk-Based Quantum Neural Network