| name | non-unitary-qml-fisher-efficiency |
| description | Non-unitary quantum machine learning via Linear Combination of Unitaries (LCU) framework, with Fisher efficiency transitions and threshold-dependent parameter scaling in medical imaging tasks. Use when: implementing non-unitary quantum layers, benchmarking quantum vs classical performance across domains, analyzing Fisher information efficiency in QML, or designing quantum circuits for medical image classification. |
| metadata | {"arxiv_id":"2603.27377","published":"2026-03-28","tags":["quantum","machine-learning","non-unitary","LCU","fisher-information","medical-imaging","histopathology"]} |
Non-Unitary Quantum ML with Fisher Efficiency Transitions
Systematic evaluation of non-unitary QML via Linear Combination of Unitaries (LCU) framework in hybrid quantum-classical neural networks.
Core Finding
Non-unitary quantum layers consistently outperform unitary baselines across all tested domains (+0.2% to +5.8% improvement), with a critical Fisher efficiency transition in medical imaging: parameter efficiency shifts from negative to positive as qubit count crosses a threshold (10→12 qubits on PathMNIST).
LCU Framework
Implementation
Non-unitary operations implemented as: U_LCU = Σᵢ αᵢ Uᵢ where each Uᵢ is a unitary gate and αᵢ are learnable coefficients.
Circuit Variants
- IQP (Instantaneous Quantum Polynomial): Commuting gate circuits with proven complexity-theoretic hardness
- Hardware-efficient: Native gate set matching device topology
- Data re-uploading: Multiple encoding layers for expressivity
Fisher Efficiency Analysis
Metric Definition
Fisher Efficiency = (Δ performance) / (Δ parameters) — measures how much accuracy improvement per additional parameter.
Threshold Phenomenon
Medical imaging tasks (PathMNIST) show a sharp transition:
- Below threshold (≤10 qubits): Negative efficiency — adding parameters hurts
- Above threshold (≥12 qubits): Positive efficiency — quantum advantage emerges
- Transition point: ~11 qubits for this architecture
Interpretation
The threshold corresponds to the point where the quantum Hilbert space dimension becomes sufficient to represent the feature manifold of the medical imaging task. Below this, the quantum model is underparameterized relative to the task complexity.
Experimental Protocol
Domain Coverage
570+ experiments across:
- MNIST: Digit classification (baseline)
- PlantVillage: Agricultural disease detection
- QM9: Molecular property regression
- PathMNIST: Medical histopathology classification
Comparison Setup
- Structurally identical unitary baselines
- Matched parameter budgets
- Same classical preprocessing pipeline
- Multiple qubit counts (2-16)
Workflow for Non-Unitary QML
Step 1: Choose Circuit Architecture
Select IQP for theoretical guarantees, hardware-efficient for near-term deployment.
Step 2: Determine Qubit Budget
For medical imaging: start at ≥12 qubits to exceed the Fisher efficiency threshold. For simpler tasks: 6-8 qubits may suffice.
Step 3: LCU Implementation
def lcu_layer(input_state, coefficients, unitaries):
output = sum(α_i * U_i @ input_state for i in range(n_terms))
return normalize(output)
Step 4: Hybrid Integration
Insert non-unitary quantum layer between classical feature extraction and classical classifier.
Step 5: Fisher Efficiency Validation
Track efficiency vs qubit count to identify threshold behavior.
Pitfalls
- LCU overhead: Probabilistic implementation requires amplitude amplification; success probability decreases exponentially with circuit depth
- Threshold sensitivity: Fisher efficiency threshold depends on architecture — validate per-task
- Noise sensitivity: Non-unitary operations are more sensitive to hardware noise than unitary equivalents
- Classical comparison fairness: Ensure classical baseline has matched parameter budget; unfair comparisons are a common critique of QML
Related Work
- Unitary QML baselines: Standard VQC approaches
- Dequantization results: Classical algorithms matching some QML claims
- Quantum expressivity: Measures of representational power in quantum models