Barren plateaus are swamped with traps

TL;DR

This paper reveals that barren plateaus in quantum machine learning are filled with many trivial local minima, which challenges the effectiveness of large-gradient initialization strategies for training quantum models.

Contribution

It identifies a mechanism creating numerous poor local minima in barren plateau regimes, highlighting limitations of current initialization approaches.

Findings

Existence of approximate local minima optimizing single loss terms

Conjecture of exact local minima optimizing a logarithmic fraction of loss terms

Large gradients alone do not solve the barren plateau problem

Abstract

Two main challenges preventing efficient training of variational quantum algorithms and quantum machine learning models are local minima and barren plateaus. Typically, barren plateaus are associated with deep circuits, while shallow circuits have been shown to suffer from suboptimal local minima. We point out a simple mechanism that creates exponentially many poor local minima specifically in the barren plateau regime. These local minima are trivial solutions, optimizing only a few terms in the loss function, leaving the rest on their barren plateaus. More precisely, we show the existence of approximate local minima, optimizing a single loss term, and conjecture the existence of exact local minima, optimizing only a logarithmic fraction of all loss function terms. One implication of our findings is that simply yielding large gradients is not sufficient to render an initialization strategy a meaningful solution to the barren plateau problem.