Quantum
Restricting Trainable Lie-Algebra Growth in Equivariant Quantum Networks via Hierarchical Ancilla-Controlled Subspace Projections
Publication: quant-ph updates on arXiv.org
Author: Ting Li, Zhiming Xiao, Qibiao Tang
Source published: 28/09/2026, 04:00:00 UTC
Briefing created: 28/09/2026, 04:34:36 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 28/09/2026, 04:13:24 UTC.
Briefings
This research introduces a hierarchical ancilla-controlled architecture to limit Lie-algebra growth in equivariant quantum networks. By using commuting invariant-sector projectors on the data register, trainable dynamics is confined to a shared ancilla register, decomposing the circuit into compatible joint sectors. The authors prove a group-independent bound on the dimension of the Lie algebra, showing it scales polynomially with data qubits when ancilla dimension and control modes remain constant along a logarithmic hierarchy. Simulations indicate this approach yields slower gradient-variance decay and larger initialization gradients compared to conventional rotationally equivariant circuits, yet it supports sparse classification and Heisenberg ground-state energies.
Key points
- The authors propose a hierarchical ancilla-controlled architecture that confines trainable dynamics to a shared ancilla register using commuting invariant-sector projectors.
- A group-independent bound on the Lie algebra dimension is proven, scaling polynomially with data qubits when ancilla dimension and control modes remain constant along a logarithmic hierarchy.
Context and limitations
Finite-size state-vector simulations show slower gradient-variance decay and larger initialization gradients compared to generic and conventional rotationally equivariant circuits over the studied system sizes.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 28/09/2026, 04:34:36 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeicqg2yi57gew3brmlmvgzjy6t65dvwslz3yg63wutr7k4lmv6h3ka