Lie Geometry Trainability Theory: A New Path to Solving the Barren Plateau Problem in Quantum Neural Networks

This study proposes a theoretical framework based on Lie geometry. By restricting the parameter space of quantum neural networks (QNNs) to low-dimensional Lie subalgebras, it fundamentally eliminates the barren plateau problem that hinders the practical application of QNNs, providing a solid theoretical foundation and experimental validation for the scalable training of quantum machine learning.

Quantum neural networks (QNNs) are an important application direction of quantum computing, but the "barren plateau" phenomenon restricts their development: as the number of qubits increases, the gradient of the cost function decays exponentially, making gradient optimization algorithms difficult to converge. Traditional mitigation methods (such as special parameter initialization and local cost function design) lack theoretical guarantees and are difficult to scale to large-scale systems.

The core contribution is the establishment of a Lie geometry framework that links the trainability of QNNs to the algebraic and representation-theoretic structure of the generator space. Key insight: Barren plateaus are geometric concentration phenomena caused by the action of high-dimensional Lie groups; restricting the parameter space to Lie subalgebras of polynomial dimension can avoid geometric concentration effects, thereby eliminating barren plateaus and providing a theoretical basis for scalable QNN design.

The project implements a Lie-QNN benchmark model based on collective spins, using symmetry to reduce the effective dimension: the variational ansatz is generated by collective spin operators, and the dynamics remain in the permutation-symmetric Dicke subspace (dimension n+1, where n is the number of qubits). Numerical experiments show that as n increases from 3 to 12, the gradient variance ratio between the Lie-constrained model and the full Hilbert baseline increases from 11.68 to 827.51. Three core experimental scripts (collective spin variance, initial residual scaling, and appendix variance experiments) form a complete evidence chain to verify the theoretical predictions.

The theoretical significance includes: 1) Revealing the trade-off between representation dimension and expressive power (full parameterization has strong expressive power but poor trainability, while Lie-constrained parameterization trades this for scalability); 2) Establishing a triangular relationship between geometry, algebra, and optimization (the optimization landscape depends on the algebraic structure of the parameter space and the geometric properties of group actions); 3) Transforming the barren plateau from an empirical obstacle into a phenomenon that can be accurately predicted and eliminated, marking an increase in the maturity of quantum machine learning theory.

Future directions: Systematically analyze the expressive power boundaries of different Lie subalgebras; explore constraints such as spatial symmetry and particle number symmetry; extend the framework to general homogeneous spaces; use geometric concentration theory to accurately characterize gradient variance scaling behavior. The goal is to establish a "Lie algebra-aware" quantum machine learning theory to guide architecture design, optimization algorithms, and theoretical analysis.

This project demonstrates the value of pure mathematics such as Lie theory and representation theory in solving practical problems in quantum computing. By understanding the algebraic structure of the parameter space, it not only explains the causes of barren plateaus but also provides a constructive method to avoid them, marking the transition of quantum machine learning from an empirical trial-and-error phase to a system design phase based on rigorous mathematical theory.