Quantum Tomography Neural Network Library

A Python library for neural network-based quantum state tomography, implementing the methods from the paper "Neural-network quantum state tomography" by Koutný et al.

The Problem: Quantum State Tomography

Mathematical Foundation

In quantum mechanics, a quantum state is described by a density matrix $\rho$, which is a positive semidefinite matrix with unit trace ($\text{Tr}(\rho) = 1$). For a d-dimensional quantum system, $\rho$ is a $d \times d$ complex matrix.

The goal of quantum state tomography is to reconstruct this unknown density matrix $\rho$ from experimental measurements. According to Born's rule, the probability of obtaining measurement outcome $i$ is:

$$ p_i = \text{Tr}(\rho \Pi_i) $$

where ${\Pi_i}$ forms a Positive Operator-Valued Measure (POVM) satisfying:

• $\Pi_i \geq 0$ (positive semidefinite)
• $\sum_i \Pi_i = I$ (completeness)

The Challenge

Traditional tomography methods face several challenges:

1. Positivity constraint: Reconstructed states must be physically valid (positive semidefinite)
2. Noise sensitivity: Real measurements have finite statistics and noise
3. Computational cost: Maximum likelihood estimation can be slow for large systems

Our Solution: Neural Network Approach

We implement a deep neural network that learns the inverse mapping from measurement probabilities to density matrices:

$$ f: [p_1, p_2, \ldots, p_{d^2}] \rightarrow \rho $$

Mathematical Details

Cholesky Decomposition

Any density matrix can be written as:

$$ \rho = \frac{LL^\dagger}{\text{Tr}(LL^\dagger)} $$

where $L$ is lower triangular with:

• Real diagonal elements
• Complex off-diagonal elements

This representation has exactly $d^2$ real parameters and automatically ensures $\rho$ is positive semidefinite.

Hilbert-Schmidt Distance

The reconstruction quality is measured by:

$$ D_{HS}(\rho, \sigma) = \sqrt{\text{Tr}((\rho-\sigma)^2)} $$

which quantifies the distance between true state $\rho$ and reconstructed state $\sigma$.

Square-Root Measurements

The POVM elements are constructed as:

$$ \Pi_i = S^{-1/2} |\psi_i\rangle\langle\psi_i| S^{-1/2} $$

where $S = \sum_i |\psi_i\rangle\langle\psi_i|$ and ${|\psi_i\rangle}$ are Haar-random states.

Installation

git clone https://github.com/yourusername/quantum-tomography-nn
cd quantum-tomography-nn
pip install -e .

Quick Start

#!/usr/bin/env python3
"""
Proper minimal example using library functions correctly.
"""

import numpy as np
import os
from quantum_tomography_nn.cli.generate_data import generate_povms, generate_training_data
from quantum_tomography_nn.models.training import train_model
from quantum_tomography_nn.core.random_states import random_HS_state
from quantum_tomography_nn.evaluation.metrics import hilbert_schmidt_distance

# Generate POVM and data
generate_povms([2], num_states=1000)
generate_training_data([2], num_states=1000)

# Train the model
model, history = train_model(2, 'data/quantum_training_data_d2.npz', 'models/demo_model.h5')

# Test
from quantum_tomography_nn.core.data_generation import get_probabilities, sample_measurements
from quantum_tomography_nn.core.state_representation import cholesky_to_density
import keras

#model = keras.models.load_model('models/demo_model.h5', compile=False)
povm = np.load('data/SRMpom2.npz')['povm']

test_rho = random_HS_state(2)
probs = sample_measurements(get_probabilities(test_rho, povm), 1000) / 1000
pred = cholesky_to_density(model.predict(probs.reshape(1,-1), verbose=0)[0], 2)

print(f"Reconstruction error: {hilbert_schmidt_distance(test_rho, pred):.4f}")
print("True state:\n", np.round(test_rho, 3))
print("Predicted state:\n", np.round(pred, 3))

Example Explanation

This minimal example demonstrates the complete workflow:

• POVM Generation: Creates a fixed set of measurement operators using generate_povms([2], num_states=1000)

• Generation: Generates random quantum states and their measurement statistics using generate_training_data([2], num_states=1000)

• Network Training: Learns the mapping from probabilities to states using train_model()

• Reconstruction: Uses the trained network to predict unknown states from new measurements

• Evaluation: Compares true vs predicted states using Hilbert-Schmidt distance

The neural network automatically learns to produce physically valid density matrices while being robust to measurement noise and finite statistics.

Neural Network Architecture

The network follows the architecture from the original paper:

• Input layer: $d^2$ nodes (measurement probabilities)

• Hidden layers: 200 → 180 → 180 → 160 → 160 → 160 → 160 → 100 neurons

• Activation: ReLU for hidden layers, tanh for output layer

• Output layer: $d^2$ nodes (Cholesky decomposition elements)

• Optimizer: Nadam with early stopping

Citation

If you use this library in your research, please cite the original paper:

@article{koutny2022neural,
  title={Neural-network quantum state tomography},
  author={Koutn{\`y}, Dominik and Motka, Libor and Hradil, Zdenek and Reh{\'a}{\v{c}}ek, Jaroslav and S{\'a}nchez-Soto, Luis L},
  journal={Physical Review A},
  volume={106},
  number={1},
  pages={012409},
  year={2022},
  publisher={APS}
}

Authors

This project was created by:

• Manuel A. Garcia
• Johan Garzón

License

This project is licensed under the GNU General Public License v3.0 or later (GPL-3.0-or-later).
See the LICENSE file for details.