Barren Plateaus Beyond Observable Concentration

TL;DR

This paper introduces a unified framework to understand barren plateaus in quantum circuits, distinguishing observable concentration from other causes, and identifies new mid-circuit phenomena affecting trainability.

Contribution

It develops a statistical framework separating observable concentration from other effects, and uncovers two new mid-circuit causes of gradient suppression in quantum circuits.

Findings

Observable concentration explains many barren plateau phenomena.

Avoiding observable concentration is necessary but not sufficient for trainability.

Mid-circuit information loss and scrambling are new causes of gradient suppression.

Abstract

Parameterized quantum circuits (PQCs) are central to quantum machine learning and near-term quantum simulation, but their scalability is often hindered by barren plateaus (BPs), where gradients decay exponentially with system size. Prior explanations, including expressivity, entanglement, locality, and noise, are often presented in ways that conflate two distinct issues: concentration of the measured observable and loss of parameter sensitivity caused by circuit dynamics. We develop a unified statistical framework that separates these mechanisms. We show that several standard BP explanations, including locality- and entanglement-related effects, can be understood through a single phenomenon that we term observable concentration (OC). Importantly, we prove that avoiding OC is necessary but not sufficient for trainability. Beyond OC, we identify two distinct mid-circuit sources of gradient suppression. First, mid-circuit information loss occurs when parameter perturbations propagate into degrees of freedom that are inaccessible to the final measurement, yielding little or no response. Second, mid-circuit information scrambling occurs when local perturbations rapidly spread across the system and become effectively undetectable on the measured subsystem. We support our theory with explicit constructions and numerical evidence, including quantum convolutional neural network architectures that exhibit information-loss-induced barren plateaus despite the absence of observable concentration.