| name | quantum-encoding-selection |
| category | quantum-computing |
| description | Quantum machine learning data encoding selection methodology based on arXiv:2606.05387. Provides a three-axis taxonomy (cost-expressivity-robustness), depth-fidelity bounds under NISQ decoherence, and a five-regime decision framework for choosing optimal encoding strategies. |
| source | arXiv:2606.05387 |
| source_title | Feature Encoding in Quantum Machine Learning: A Survey and Practical Guidelines |
| source_author | Vincenzo Sammartino |
| keywords | ["quantum machine learning","data encoding","NISQ","amplitude encoding","angle encoding","barren plateaus","quantum kernels"] |
Quantum ML Encoding Selection
Overview
Systematic methodology for selecting optimal quantum data encoding strategies on NISQ devices. Based on a survey of 66 primary works (2017-2026) with a PRISMA-adapted protocol.
Trigger: When designing quantum machine learning pipelines, choosing data encoding strategies, analyzing QML trainability, or optimizing encoding circuits for NISQ hardware.
arXiv: 2606.05387 | Author: Vincenzo Sammartino
Core Framework: Three-Axis Taxonomy
Classify all encoding families along three independently measurable axes:
1. Cost Axis
- Gate depth: Total circuit depth required for encoding
- Qubit count: Number of qubits needed (amplitude encoding: log₂(D), basis encoding: D)
- Classical preprocessing: Computational overhead before quantum circuit
2. Expressivity Axis
- Fourier expressivity: Range of frequencies the encoding can represent
- Feature map rank: Dimension of the span of encoded states
- Kernel richness: Ability to separate data classes in Hilbert space
3. Robustness Axis
- Noise resilience: Sensitivity to NISQ decoherence channels
- Barren plateau resistance: Gradient scaling behavior with qubit count
- Kernel concentration: Variance of kernel values under noise
Encoding Families Reference
| Encoding | Qubits | Depth | Expressivity | NISQ Viable |
|---|
| Basis | D | O(1) | Low | Yes |
| Angle | n | O(n) | Medium | Yes |
| Dense-Angle | n | O(n) | Medium-High | Yes |
| Amplitude | log₂(D) | O(D) | High | Only if p < 10⁻³ |
| Data Re-uploading | n | O(n×L) | Very High | Limited |
| IQP | n | O(n²) | High | Limited |
Critical Threshold
Gate-error rate p ~ 10⁻³* is the critical threshold below which amplitude encoding is viable. For p ≥ 10⁻³ (current NISQ reality), shallow angle-based encodings consistently outperform amplitude encoding despite the latter's exponential qubit advantage.
Five-Regime Decision Framework
Map (D, n, p, τ) to encoding recommendation:
- Low-D, High-p: Basis encoding (minimal depth, safe)
- Medium-D, Medium-p: Angle/dense-angle encoding (balanced)
- High-D, Low-p (< 10⁻³): Amplitude encoding (exponential compression)
- Complex features, Any-p: Data re-uploading (iterative expressivity)
- Hardware-aware: IQP encoding when hardware connectivity permits
Trainability Analysis
Unified treatment of three trainability concerns as functions of encoding circuit:
- Barren plateau onset: Exponential gradient vanishing with qubit count
- Mitigated by: Local encodings, shallow circuits, structured data
- Quantum kernel concentration: Kernel values collapsing under noise
- Mitigated by: Encoding depth optimization, noise-aware kernels
- Fourier spectrum gaps: Missing frequency components in feature map
- Mitigated by: Data re-uploading, hybrid encoding strategies
Pitfalls
- Amplitude encoding's exponential qubit advantage is nullified by decoherence at current error rates
- Fixed embedding ansatz selection without data geometry analysis leads to suboptimal performance
- Ignoring the cost-expressivity-robustness tradeoff triad results in untrainable circuits
- Wasserstein distance in input space provides an a priori diagnostic for encoding optimization saturation
Verification Steps
- Compute gate-error rate p of target hardware
- If p ≥ 10⁻³, default to angle-based encoding
- Check barren plateau scaling: verify gradient norm doesn't vanish exponentially
- Validate kernel concentration: ensure kernel variance remains above noise floor
- Cross-reference with hardware connectivity for depth optimization