split-ensemble-qrc-training

name	split-ensemble-qrc-training
category	quantum-finance
description	Split-ensemble training methodology for quantum reservoir computing that reorganizes measurement shot records to create more training examples without additional quantum hardware cost.
trigger_words	["split-ensemble training","quantum measurement records","shot reorganization","QRC training efficiency","finite-shot noise","quantum time series"]

Split-Ensemble Training for Quantum Reservoir Computing

Paper Reference

arXiv:2604.28160 — "Reorganizing Quantum Measurement Records Improves Time-Series Prediction" Authors: Markus Baumann, Maximilian Zorn, Thomas Gabor, Claudia Linnhoff-Popien, Jonas Stein (2026-04-30)

Core Methodology

Problem

Standard QRC averages all shots from one labeled timestep into a single feature vector:

• Reduces finite-shot noise ✓
• But gives readout only ONE training example per timestep ✗
• With limited timesteps, readout is severely undertrained

Solution: Split-Ensemble Training

1. Execute circuit with N shots per timestep (e.g., 512 shots)
2. Split shots into K groups (e.g., 512 → 8 groups × 64 shots each)
3. Each group average becomes a separate feature vector for the SAME target
4. Readout sees K× more training examples without additional circuit executions

Key Insight

• Each group average is a partially denoised estimate of the true expectation value
• Groups are independent samples → readout learns from multiple noisy views
• Strongest gains observed on real hardware where noise is structured

Implementation

import numpy as np

def split_ensemble_features(shots_data, n_groups=8):
    """Split measurement shots into groups, return group averages"""
    # shots_data: (n_timesteps, n_shots, n_observables)
    n_timesteps, n_shots, n_obs = shots_data.shape
    group_size = n_shots // n_groups
    
    # Reshape: split shots into groups
    reshaped = shots_data[:, :n_groups * group_size, :].reshape(
        n_timesteps, n_groups, group_size, n_obs
    )
    
    # Average within each group
    group_averages = reshaped.mean(axis=2)  # (n_timesteps, n_groups, n_obs)
    
    # Flatten: each group becomes a separate training example
    X_split = group_averages.reshape(n_timesteps * n_groups, n_obs)
    y_split = np.repeat(y_targets, n_groups)  # Same target for all groups
    
    return X_split, y_split

# Usage
X_split, y_split = split_ensemble_features(measurement_data, n_groups=8)
readout.fit(X_split, y_split)

When to Use

Condition	Recommendation
Few training timesteps (< 100)	Use split-ensemble with K≥4 groups
Real hardware deployment	Strongest benefit — use K≥8 groups
Simulator with many timesteps (> 500)	Standard averaging sufficient
Very low shot count (< 64 per group)	May add too much noise — limit K

Pitfalls

1. Groups must be independent — don't reuse shots across groups
2. Too many groups with too few shots → excessive noise per group
3. Not a substitute for more data — only helps when the bottleneck is training sample count, not feature quality

Related

• Combine with post-training quantization (quantum-reservoir-forecasting-resource-efficient) for deployment
• Combine with distributed QRC (arXiv:2605.04991) for scaling

Activation

Keywords: split-ensemble training, quantum measurement records, shot reorganization, QRC training efficiency, finite-shot quantum ML, quantum time series training, measurement shot grouping, near-term quantum learning