| name | compositional-quantum-heuristics |
| description | Compositional quantum heuristics for mitigating barren plateaus in quantum machine learning. Assembles larger quantum models from smaller subcomponents with group-invariant loss functions introducing symmetry-induced inductive bias for improved gradient behavior. Use when: barren plateau mitigation, quantum graph neural networks, permutation-equivariant quantum models, recursive quantum-classical hybrid optimization, QIRO-inspired quantum heuristics, max-clique quantum detection, group-invariant quantum loss functions, symmetry-induced quantum inductive bias. Triggered by: compositional quantum circuits, barren plateau quantum ML, quantum graph neural network, permutation-equivariant QGNN, group-invariant loss quantum, recursive quantum optimization, QIRO quantum informed recursive optimization, max-clique quantum detection. |
Compositional Quantum Heuristics
Mitigating barren plateaus by assembling larger quantum models from smaller
subcomponents with symmetry-induced inductive bias.
Paper
arXiv: 2605.07611v1 — Compositional Quantum Heuristics for Max-Clique Detection
by Tiffany Duneau, Colin Krawchuk, Anna Pearson (May 2026).
Core Approach
- Compositional Assembly: Build large quantum models from smaller, trainable subcomponents.
- Group-Invariant Loss Functions: Construct loss functions invariant under group actions,
introducing symmetry-induced inductive bias for improved gradient behavior and generalization.
- Permutation-Equivariant QGNNs: Design quantum graph neural networks that respect
graph permutation symmetry for max-clique detection.
- Recursive Hybrid Heuristic: Use trained quantum models to guide classical search,
inspired by QIRO (Quantum-Informed Recursive Optimization).
Key Results
- Superior training gradients through symmetry-induced bias
- Generalization to larger, more complex problem instances
- Improved inference accuracy and scalability via recursive hybrid quantum-classical procedure
- Viable pathway to scalable quantum learning models that remain hard to simulate classically
When to Use
- Quantum ML model design suffering from barren plateaus
- Graph optimization problems (max-clique, max-cut, etc.)
- Building trainable quantum circuits with expressivity
- Hybrid quantum-classical recursive optimization workflows
