Zoë Holmes – Variational quantum computing after the hype: barren plateaus and classical simulability

Zoë Holmes – Variational quantum computing after the hype: barren plateaus and classical simulability

Zoë Holmes, EPFL

Leuchs-Russell-Auditorium, A.1.500, Staudtstr. 2
Location Details

Abstract
Variational quantum computing aims to use quantum computers as trainable models. A prominent approach uses parametrized quantum circuits whose parameters are optimized, much like the weights of a neural network, to minimize a loss function. These models initially generated considerable excitement as a flexible way of tackling a wide range of problems. Increasingly, however, they come with a rather formidable collection of challenges.
One of the best studied is the barren plateau phenomenon, in which the loss landscape becomes exponentially flat as the system grows, making useful gradients prohibitively difficult to resolve. Much effort has therefore gone into designing models that provably avoid barren plateaus. But this raises a potentially uncomfortable question: can the same structure that makes these models trainable also make them classically simulable?
I will present evidence that many commonly used models with provably absent barren plateaus can indeed be simulated classically, provided one can collect some data from a quantum device during an initial acquisition phase. The underlying picture is that barren plateaus arise from a curse of dimensionality, while many proposed solutions avoid this curse by restricting the relevant dynamics to small, classically manageable subspaces. I will discuss what this means for the prospects of quantum advantage, where the argument may break down, and whether smarter initialization strategies offer a way forward.