Reproduce “Theory of overparametrization in quantum neural networks” on OpenResearch

About this paper

Martin Larocca, Nathan Ju, Diego García-Martín, Patrick J. Coles, M. Cerezo

Researchers from Los Alamos National Laboratory developed the first rigorous theoretical framework for overparametrization in quantum neural networks (QNNs), linking it to the QNN's Dynamical Lie Algebra (DLA). The work demonstrates that exceeding a critical number of parameters, determined by the DLA dimension, leads to a computational phase transition that improves trainability by removing spurious local minima in the loss landscape.

Key insights

• Overparametrization in QNNs is characterized by the number of parameters exceeding a critical value, M_c, where the QFIM for training states reaches its maximum rank, bounded by the dimension of the QNN's Dynamical Lie Algebra (DLA).
• The onset of overparametrization corresponds to a computational phase transition where spurious local minima in the loss landscape tend to disappear, significantly improving the QNN's trainability and convergence.
• The DLA dimension serves as an upper bound for both the maximum achievable rank of the QFIM (at optimal and random points), the QNN's capacity, and the rank of the Hessian matrix at optimal solutions, indicating the effective dimensionality of the optimization problem.

Summary by alphaXiv.