Trainability ยท Algorithms

Barren-plateau gradient variance

Theorem statement

For a parametrized quantum circuit whose dynamical Lie algebra is the orthogonal direct sum $\mathfrak{g} = \bigoplus_j \mathfrak{g}_j$ of its ideals, the loss variance is the reductive Ragone sum $$\operatorname{Var}_\theta[\ell] = \sum_j \frac{P_{\mathfrak{g}_j}(\rho) P_{\mathfrak{g}_j}(O)}{\dim\mathfrak{g}_j},$$ where $P_{\mathfrak{g}_j}$ is the $\mathfrak{g}_j$-purity (derived from cross-ideal Casimir orthogonality); the simple-Lie-algebra case is its single-ideal corollary. When $\dim\mathfrak{g}$ grows exponentially in the qubit count, an exponential barren plateau follows (given the Haar second-moment / Schur projection the formalization carries as a named hypothesis).

Sources

  1. A Lie Algebraic Theory of Barren Plateaus for Deep Parameterized Quantum Circuits

    Michael Ragone, Bojko N. Bakalov, Frederic Sauvage, Alexander F. Kemper, Carlos Ortiz Marrero, Martin Larocca, M. Cerezo, 2023

  2. Barren plateaus in variational quantum computing

    Martin Larocca, Supanut Thanasilp, Samson Wang, Kunal Sharma, Jacob Biamonte, Patrick J. Coles, et al., 2024

  3. Barren plateaus in quantum neural network training landscapes

    Jarrod R. McClean, Sergio Boixo, Vadim N. Smelyanskiy, Ryan Babbush, Hartmut Neven, 2018

Lean context

Copy a short prompt with the import, theorem name, citations, and public source link.

Open Lean source