Reproduce “Barren plateaus in quantum neural network training landscapes” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jarrod R. McClean, Sergio Boixo, Vadim N. Smelyanskiy, Ryan Babbush, Hartmut Neven
Randomly initialized, sufficiently expressive quantum circuits can develop “barren plateaus”: wide regions where expectation-value gradients shrink exponentially with qubit count. Analytical results based on unitary-design statistics and numerical tests show that this suppression occurs for both local Pauli objectives and state-overlap objectives, making gradient estimation—and therefore variational training—impractical at scale.
Key insights
- A barren plateau is not simply a conventional local minimum: it is a broad, statistically typical region where both objective changes and directional gradients are nearly absent.
- Greater circuit expressibility and depth can worsen trainability, because circuits approaching 2-design behavior exhibit exponentially concentrated gradients.
- Quantum measurement costs amplify the problem: resolving an exponentially small gradient requires a correspondingly prohibitive number of circuit evaluations.
- Structured initialization, problem-informed ansätze, or layerwise training may be needed instead of fully random initialization, although these strategies are not established by the study.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.