| name | qml-expressivity-trainability-paradox |
| description | Dynamical Lie Algebra framework for understanding and navigating the expressivity-trainability paradox in Quantum Machine Learning. Shows unstructured QML suffers quantum underfitting from barren plateaus. Symmetry-preserving structural regularization guarantees scalable gradient-rich landscapes. Trainability-by-Design approach. |
| version | 1.0.0 |
| tags | ["quantum","machine-learning","barren-plateau","dynamical-lie-algebra","trainability","expressivity"] |
| source | arXiv:2606.31536 |
| authors | ["Kung-Ming Lan","Edward Huang"] |
| published | 2026-06-30T00:00:00.000Z |
| trigger_words | ["QML expressivity trainability paradox","barren plateau DLA","dynamical lie algebra quantum","trainability by design","quantum underfitting","symmetry-preserving QML"] |
QML Expressivity-Trainability Paradox via DLA
Core Insight
QML suffers from a expressivity-trainability paradox: the vast Hilbert space capacity of PQCs is the direct mathematical cause of Barren Plateaus. Unstructured QML architectures suffer from quantum underfitting, not overfitting.
Key Findings
1. The Paradox
- Classical deep learning: increasing capacity risks overfitting
- Quantum ML: increasing capacity causes barren plateaus (quantum underfitting)
- The vast Hilbert space = the problem, not the solution
2. DLA Framework
- Links circuit generator algebraic dimension to optimization dynamics
- Exponential DLA growth = exponentially flat gradients
- Polynomial DLA growth = trainable landscapes
3. Trainability-by-Design
- Embed group-theoretic geometric priors as structural regularizers
- Restrict DLA growth to polynomial regime
- Sacrifice raw memorization capacity for scalable, gradient-rich landscapes
Implementation Pattern
- Analyze your PQC's DLA dimension
- If DLA grows exponentially with qubit count, expect barren plateaus
- Identify symmetries in your problem domain
- Embed symmetry-preserving constraints into circuit architecture
- Verify DLA growth is polynomial after constraint
- Train with guaranteed gradient-rich landscape
Practical Applications
Financial QML
- Stock price prediction QNNs: embed market symmetry constraints
- Portfolio optimization: embed permutation symmetries
- Avoid barren plateaus by design, not by luck
General QML
- Any QML task where circuit depth/qubit count scales
- Design circuits with built-in trainability guarantees
- Replace heuristic ansatz design with principled DLA analysis
Activation
Use when:
- Designing QML architectures and worried about barren plateaus
- Analyzing why a QNN fails to train
- Need principled ansatz design methodology
- Building scalable quantum neural networks