Researchers map qubit states to probability distributions in phase space

Mapping Qubit States onto Classical Probability Distributions

Researchers from ICFO and the University of Jagiellonian have introduced a novel way to simulate quantum circuits by representing an N‑qubit state as a probability distribution on a 2N‑dimensional manifold. The approach leverages the Q‑function—a positive‑definite phase‑space representation of quantum states—and models each unitary gate as a normalizing flow, a type of generative model that transforms one probability distribution into another.

Phase‑Space Representation of Multi‑Qubit Systems

In conventional quantum‑simulation techniques, the exponential growth of the Hilbert space quickly becomes intractable. By projecting the state onto the compact manifold \((S^2)^N\) and using the Q‑function, the authors obtain a fully positive distribution that can be handled with classical probability tools. This sidesteps the need for complex amplitudes and enables the use of well‑established machine‑learning methods.

Normalizing Flows as Quantum Gates

Normalizing flows, originally designed for density‑estimation tasks in machine learning, are repurposed here to encode the action of a quantum gate. Each gate receives its own flow, which maps the input distribution (the Q‑function before the gate) to the output distribution (the Q‑function after the gate). Because flows are composable, a whole circuit can be simulated by chaining the corresponding flows without retraining a global architecture after each gate.

Benchmarking Against Neural‑Network Baselines

The team evaluated the framework on three families of circuits:

1. 1‑D cluster‑state preparation
2. GHZ‑state preparation
3. Ground‑state preparation of a 1‑D transverse‑field Ising model (VQE style)

Across increasing qubit counts (up to 10 qubits) and circuit depths, the classical fidelity—the overlap between the exact Q‑function and the flow‑generated distribution—remained higher than that achieved by two neural‑network baselines (a neural quantum state and a transformer‑based model). The advantage grew as the system size expanded, suggesting better scalability.

Implications and Future Directions

By shifting the bottleneck from storing exponentially many amplitudes to learning expressive flow transformations, the method opens a new pathway for classical simulation of larger quantum circuits. Potential extensions include:

• Transfer learning between flows to reduce training cost for similar circuits.
• Hybrid schemes that combine flow‑based models with tensor‑network techniques.
• Application to noisy, open‑system dynamics by adapting flow architectures to Lindblad‑type evolutions.

While the current work focuses on closed, multi‑qubit systems, the underlying idea—using continuous, positive phase‑space representations—could inspire broader algorithmic innovations in quantum‑classical hybrid computing.

---

The research is available as a pre‑print on arXiv (see link below). As the work has not yet undergone peer review, readers should treat the results as promising but preliminary.