quantum-paradigm-comparison-cv-dv

name	quantum-paradigm-comparison-cv-dv
category	quantum
description	Controlled comparison methodology for continuous-variable (CV) vs discrete-variable (DV) quantum computing paradigms. Uses shared classical backbone with interchangeable quantum heads to isolate quantum circuit as sole variable. CV outperforms DV with 18-point accuracy gap on wafer-map defect classification.
tags	["quantum","CV","DV","comparison","quantum-machine-learning","paradigm"]
arxiv_id	2607.00961v1
created	2026-07-07

CV vs DV Quantum Paradigm Comparison

Overview

When deploying quantum neural networks in industry, knowing which quantum computing paradigm suits which task is essential. This methodology provides a controlled comparison framework between continuous-variable (CV) and discrete-variable (DV) quantum computing paradigms, isolating the quantum circuit as the sole experimental variable.

Core Methodology

Controlled Experimental Design

To isolate the quantum circuit as the sole variable:

1. Shared Backbone: Use identical classical convolutional backbone (~4.3M parameters) for feature extraction
2. Interchangeable Heads: Swap only the classification head (classical dense, CV-QNN, or DV-QNN)
3. Scale Variation: Test each quantum head at multiple sizes (3, 4, 8 qumodes/qubits)
4. Fixed Conditions: Keep all other parameters constant (dataset, training procedure, evaluation)

Key Finding: CV Superiority

At 4 qumodes/qubits on WM-811K wafer-map defect classification:

• CV accuracy: 79.7 ± 1.8%
• DV accuracy: 61.6 ± 1.4%
• Gap: Non-overlapping 18-point advantage for CV

Fine-Grained Advantage

The CV advantage is sharpest on spatially localized defect types:

• Edge-Loc class: CV recall 0.66 ± 0.06 vs DV recall ≤ 0.05 at every size
• Edge-Loc is easily confused with Scratch — CV captures fine spatial distinctions that DV misses
• This shows the structured CV layer better captures spatial patterns

Root Cause Analysis

• DV limitation is a representational-capacity ceiling, not an optimization failure
• At Fock cutoff d=2, CV advantage reflects:
  1. A structured, neural-network-analogue layer
  2. Continuous phase-space encoding
• Not simply Hilbert-space dimensionality

Hardware Validation

On IBM hardware:

• DV accuracy holds at shallow depth
• Degrades only at the deepest circuit
• Both quantum heads remain below classical baseline (85.0%)

Implementation Pattern

Step 1: Choose Shared Backbone

• Select a classical feature extractor appropriate for the data modality
• Ensure sufficient capacity for the task
• Keep frozen during quantum head comparison

Step 2: Implement Interchangeable Heads

Classical Baseline: backbone → dense layer → output
CV Head:            backbone → CV-QNN (qumodes) → output  
DV Head:            backbone → DV-QNN (qubits) → output

Step 3: Scale Analysis

• Test at multiple sizes (small, medium, large)
• Plot accuracy vs size for each paradigm
• Identify where advantages emerge

Step 4: Class-Level Analysis

• Don't just look at overall accuracy
• Analyze per-class performance to find where paradigms differ
• Identify which data characteristics favor which paradigm

When to Use

• Choosing between CV and DV quantum computing for a specific task
• Benchmarking quantum vs classical approaches
• Understanding where quantum advantage might first appear
• Designing hybrid quantum-classical architectures
• Spatial pattern recognition tasks (image classification, defect detection)

Activation Keywords

CV-QNN, DV-QNN, continuous-variable, discrete-variable, quantum paradigm comparison, wafer-map classification, quantum head, qumodes, qubits, Fock cutoff, phase-space encoding, quantum benchmark