Sculpting Quantum Landscapes: Fubini-Study Metric Conditioning for Geometry Aware Learning in Parameterized Quantum Circuits

TL;DR

This paper introduces Sculpture, a meta learning framework that conditions the Fubini-Study metric of quantum circuits to improve trainability and generalization by mitigating barren plateaus in variational quantum algorithms.

Contribution

The paper proposes a novel meta learning approach that explicitly conditions the quantum circuit's geometry, specifically the Fubini-Study metric, to enhance optimization and generalization.

Findings

Meta training reduces the logarithmic condition number of the Fubini-Study metric.

Improved conditioning leads to better optimization and generalization in quantum tasks.

Enhanced initializations accelerate convergence and improve accuracy in a quantum classification task.

Abstract

We present a novel meta learning framework called Sculpture that explicitly conditions the Fubini Study metric tensor of parameterized quantum circuits to mitigate barren plateaus in variational quantum algorithms. Our theoretical analysis identifies the logarithmic condition number of the Fubini Study metric as a critical geometric quantity governing trainability, optimization dynamics, and generalization. Sculpture uses a classical meta model trained to generate data dependent quantum circuit initializations that minimize the logarithmic condition number, thereby promoting an isotropic and well conditioned parameter space. Empirical results show that meta training reduces the logarithmic condition number from approximately 1.47 to 0.64 by significantly increasing the minimum eigenvalue and slightly decreasing the maximum eigenvalue of the metric, effectively alleviating barren plateaus. This improved conditioning generalizes well to unseen data, consistently producing well conditioned quantum circuit initializations. In a downstream hybrid quantum classical classification task on the Kaggle diabetes dataset, increasing the meta scaling coefficient accelerates convergence, reduces training loss and gradient norms, and crucially improves generalization, with test accuracy increasing from about 0.68 to over 0.78. These findings demonstrate that sculpting the quantum landscape via meta learning serves as a principled geometric regularizer, substantially enhancing trainability, optimization, and generalization of parameterized quantum circuits and enabling more robust and efficient variational quantum algorithms.