EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Controllability and trainability in VQC: A DLA framework for optimal expressivity

Pablo Vilaboy-Sachocos, Roberto Casado, Jesús-Ángel Román-Gallego, María-Luisa Pérez-Delgado, Moisés-Alberto García-Núñez, Juan M Corchado

Journal of Logic and Computation · 02.11.2026 · 36 (8)

DOI: 10.1093/logcom/exag055 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Logic and Computation, online erschienen am 13.11.2026. Verfasst von Pablo Vilaboy-Sachocos und 5 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Variational quantum circuits are a central framework in quantum machine learning, yet their trainability remains poorly understood, particularly in relation to circuit expressivity. In this work, we develop a unified framework that connects controllability and trainability through the structure of the dynamical Lie algebra (DLA) associated with a given circuit architecture. We show that while increasing controllability enhances expressivity, it may simultaneously degrade optimization due to concentration effects leading to vanishing gradients. Conversely, strongly constrained architectures may preserve large gradients, but their performance depends critically on the alignment between the induced algebraic structure and the learning task. We formalize this trade-off by introducing the notion of trainability-efficient controllability, which characterizes intermediate regimes where the DLA is sufficiently expressive while avoiding excessive dynamical complexity. Theoretical results establish that full controllability is not necessary for achieving optimal task performance, and that proper Lie subalgebras can be sufficient under appropriate conditions. These findings are supported by numerical experiments on representative architectures exhibiting distinct controllability regimes. Our results provide empirical evidence for a non-monotonic relationship between controllability and trainability, and highlight that optimal performance is achieved when the structure of the DLA is aligned with the requirements of the learning task.

Zitieren

APA

Vilaboy-Sachocos, P., Casado, R., Román-Gallego, J., Pérez-Delgado, M., García-Núñez, M., & Corchado, J. M. (2026). Controllability and trainability in VQC: A DLA framework for optimal expressivity. Journal of Logic and Computation, 36(8). https://doi.org/10.1093/logcom/exag055

Vancouver

Vilaboy-Sachocos P, Casado R, Román-Gallego J, Pérez-Delgado M, García-Núñez M, Corchado JM. Controllability and trainability in VQC: A DLA framework for optimal expressivity. Journal of Logic and Computation. 2026;36(8). doi:10.1093/logcom/exag055

BibTeX
@article{VilaboySachocos2026,
  author = {Vilaboy-Sachocos, Pablo and Casado, Roberto and Román-Gallego, Jesús-Ángel and Pérez-Delgado, María-Luisa and García-Núñez, Moisés-Alberto and Corchado, Juan M},
  title = {Controllability and trainability in VQC: A DLA framework for optimal expressivity},
  journal = {Journal of Logic and Computation},
  year = {2026},
  volume = {36},
  number = {8},
  doi = {10.1093/logcom/exag055}
}
Kostenlos starten →