Speaker
Description
This work studies the trainability–expressibility trade-off in variational quantum machine learning. Highly expressive parameterized quantum circuits can suffer from barren plateaus, while overly restricted circuits may lack useful representational power. We introduce a finite-sample, dimension-free trainability certificate based on gradient fluctuations, allowing trainability to be used as a quantitative objective in ansatz search. Combining this criterion with expressibility and Bayesian optimization, we identify circuit architectures with favorable metric-level properties. We further study a hardware-constrained search inspired by the 5-qubit IQM Spark device and apply the framework to a quantum neural network benchmark, where the discovered ansatz achieves improved effective dimension with fewer parameters. Overall, the work provides a practical framework for trainability-aware design of quantum variational circuits.
| Topic | Quantum variational circuits and their applications |
|---|