Fock-Space Quantum Neural Network (Continuous-Variable QML)

This repository implements a continuous-variable (CV) quantum neural network realizing a Fock-space classifier, inspired by

M. Schuld et al., Quantum machine learning in feature Hilbert space.

The model embeds classical data into an infinite-dimensional (truncated) Fock space and performs linear classification in feature space using a non-Gaussian CV quantum circuit.

---

Requirements (IMPORTANT)

This project only works with specific pinned versions. Do not upgrade packages.

Tested on:

• Ubuntu 22.04
• Python 3.10

---

Repository structure

schuld_qml/
├── .venv/                 # Python virtual environment (NOT committed)
├── qm_neural_net.py       # Full Fock-space classifier (training + plots)
├── test_00.py             # Backend sanity check (must return 0.0)
├── requirements.txt       # Fully pinned dependency versions
└── README.md

---

Setup instructions (MANDATORY)

1. Create a virtual environment inside the repository

cd schuld_qml
python3 -m venv .venv
source .venv/bin/activate

You must see:

(.venv)

---

2. Install dependencies

pip install --upgrade pip
pip install -r requirements.txt

---

3. Verify the CV Fock backend (CRITICAL STEP)

Run:

python test_00.py

Expected output:

0.0

If and only if this prints 0.0, the environment is correct.

---

Running the classifier

python qm_neural_net.py

This will:

• generate a two-moons dataset,

• train a 2-mode Fock-space classifier using stochastic gradient descent,

• plot:

  • training loss vs optimization steps,
  • decision regions showing ( p(y=1) ).

---

Model overview

• Feature map Classical input ( x \in \mathbb{R}^2 ) is encoded via displacement operators into a truncated 2-mode Fock space.

• Feature space Infinite-dimensional bosonic Hilbert space, truncated numerically.

• Variational circuit Repeated gate blocks consisting of:

  • Beamsplitter (BS)
  • Displacement (D)
  • Quadratic phase (P)
  • Cubic phase (C)

• Measurement Photon number expectation value.

• Loss Square loss with gentle ( \ell_2 ) regularization.

The classifier is:

• linear in Fock space,
• nonlinear in input space,
• non-Gaussian, due to cubic phase gates.

---

Dependency versions (DO NOT CHANGE)

The program is guaranteed to work only with:

PennyLane==0.29.1
PennyLane-SF==0.20.1
strawberryfields==0.23.0
thewalrus==0.19.0
autoray==0.6.12
numpy==1.23.5
scipy==1.10.1

These versions are pinned in requirements.txt.

---

Common issues

• strawberryfields.fock device not found → Ensure PennyLane-SF==0.20.1 is installed.

• SciPy import errors → Ensure scipy==1.10.1.

• Autoray errors → Ensure autoray==0.6.12.

If anything breaks, delete .venv/ and repeat the setup exactly.

---

Reproducibility rule

If it runs once, freeze the environment and never upgrade dependencies.