Pitfalls when tackling the exponential concentration of parameterized quantum models
read the original abstract
Identifying scalable circuit architectures remains a central challenge in variational quantum computing and quantum machine learning. Many approaches have been proposed to mitigate or avoid the barren plateau phenomenon or, more broadly, exponential concentration. However, due to the intricate interplay between quantum measurements and classical post-processing, we argue these techniques often fail to circumvent concentration effects in practice. Here, by analyzing concentration at the level of measurement outcome probabilities and leveraging tools from hypothesis testing, we develop a practical framework for diagnosing whether a parameterized quantum model is inhibited by exponential concentration. Applying this framework, we argue that several widely used methods (including quantum natural gradient, sample-based optimization, and certain neural-network-inspired initializations) do not overcome exponential concentration with finite measurement budgets, though they may still aid training in other ways.
This paper has not been read by Pith yet.
Forward citations
Cited by 1 Pith paper
-
Trainability Beyond Linearity in Variational Quantum Objectives
The trainability boundary for variational quantum objectives is the affine regime; non-affine amplification-capable losses can mitigate barren plateaus when using coarse-grained statistics at polynomial widths.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.