qml-expressivity-separation

name

qml-expressivity-separation

description

Quantum Machine Learning expressivity separation methodology. Based on Anschuetz & Gao (Quantum 10, 1976, 2026). Provides framework for constructing efficiently trainable QNNs with provable polynomial memory separations over classical neural networks. Use when: (1) designing QNN architectures with provable quantum advantage, (2) analyzing expressivity vs trainability trade-offs, (3) implementing quantum contextuality as computational resource, (4) comparing quantum vs classical sequence modeling capabilities. Keywords: quantum machine learning, QNN, expressivity, contextuality, polynomial separation, trainable.

QML Expressivity Separation

Based on Anschuetz & Gao, "Arbitrary Polynomial Separations in Trainable Quantum Machine Learning" (Quantum 10, 1976, 2026).

Core Result

Constructed hierarchy of efficiently trainable QNNs that exhibit unconditionally provable polynomial memory separations of arbitrary constant degree over classical neural networks (including Transformers) for classical sequence modeling tasks.

Key Concepts

Expressivity-Trainability Trade-off

Previous work showed a fundamental tension:

• High expressivity QNNs → exponential training time
• Trainable QNNs → limited expressivity advantage

This work circumvents the trade-off by:

1. Constructing QNNs with constant gate complexity per unit cell
2. Achieving polynomial (not exponential) but arbitrary degree separations
3. Using quantum contextuality as the resource

Quantum Contextuality

Definition: Quantitative notion of semantic ambiguity — the inability to assign consistent classical values to all measurement outcomes simultaneously.

Role in QML: Contextuality in the quantum model's latent space is the source of expressivity separation. Tasks with semantic ambiguity are natural candidates for quantum advantage.