pith. sign in

arxiv: 2312.09135 · v3 · pith:XNBLU36Hnew · submitted 2023-12-14 · 🪐 quant-ph · cond-mat.stat-mech· cond-mat.str-el

Measurement-Induced Landscape Transitions and Coding Barren Plateaus in Hybrid Variational Quantum Circuits

classification 🪐 quant-ph cond-mat.stat-mechcond-mat.str-el
keywords barrenplateauinformationlandscapemeasurement-inducedmiltquantumtext
0
0 comments X
read the original abstract

The entanglement-induced barren plateau is an exponential vanishing of the parameter gradients with system size that limits the practical application of variational quantum algorithms (VQA). A landscape transition from barren plateau to no-barren plateau was recently observed in monitored quantum circuits, hypothesized to coincide with the measurement-induced phase transition (MIPT) that separates the area-law states from the volume-law states. We argue from an information theory perspective that these are different transitions. This hypothesis is supported by a numerical study that includes cost-gradient variances, visualizations of the optimization runs and cost-landscape, and a quantum-classical channel mutual information measure. The results are evidence for a universal measurement-induced landscape transition (MILT) at $p_c^{\text{MILT}} \approx 0.2 < p_c^{\text{MIPT}}$ and that throughout $0<p<p_c^{\text{MILT}}$, there is a finite quantum-classical channel mutual information in the limit of a large number of qubits. Unlike the barren plateau without measurements, a non-zero rate of measurements induces a coding barren plateau where, typically, information about the parameters is available to a local cost function despite a vanishing gradient.

This paper has not been read by Pith yet.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Large Language Models Can Help Mitigate Barren Plateaus in Quantum Neural Networks

    quant-ph 2025-02 unverdicted novelty 5.0

    AdaInit uses LLMs with submartingale properties to iteratively synthesize QNN initial parameters that maintain non-negligible gradient variance and mitigate barren plateaus, with claimed theoretical convergence guaran...