Normalizing Flows for Classical Simulation of Quantum Circuits in Phase Space

Technical Summary: Normalizing Flows for Classical Simulation of Quantum Circuits in Phase Space

Problem Statement
The classical simulation of quantum many-body systems remains a critical challenge for understanding quantum advantage and controlling quantum hardware. Traditional methods face significant limitations: exact state-vector and path-integral approaches suffer from exponential algorithmic complexity relative to system size, while structure-exploiting techniques (e.g., Clifford circuits, tensor networks) are restricted to specific circuit classes or low-entanglement regimes. Furthermore, existing machine learning approaches, such as Neural Quantum States (NQS) and generative models for quantum dynamics, typically exhibit degraded accuracy as system size and evolution time increase. A key difficulty lies in representing quantum states as probability distributions without resorting to discrete relaxations or hand-crafted embeddings that may lose fidelity.

Methodology
The authors propose a framework that represents the output states of quantum circuits using the phase space formulation of many-body qubit systems, specifically utilizing the Husimi Q function.

1. Phase Space Representation:

   • $N$-qubit states are mapped to probability distributions over the smooth manifold $(S^2)^N$. The Q function for a state $|\psi\rangle$ is defined as $Q_{|\psi\rangle}(\Omega) = \frac{1}{(2\pi)^N} |\langle \Omega | \psi \rangle|^2$, where $\Omega$ represents coordinates on the product of Bloch spheres.
   • Unitary gates $U$ correspond to convolution operations with integral kernels $K_U$ in this phase space. While these kernels can contain negative values (preventing direct sampling), they allow for the exact evolution of the distribution.

2. Normalizing Flow Formulation:

   • The authors model the evolution of the quantum circuit as a sequence of normalizing flows. Instead of directly sampling from the potentially negative kernel convolution, they learn a flow transformation $f_k$ for each gate $U_k$ that maps the distribution of the previous state $Q_{|\psi_{k-1}\rangle}$ to the next state $Q_{|\psi_k\rangle}$.
   • The transformation is defined by the change-of-variables formula:
     $$\tilde{Q}_{|\psi_k\rangle}(\Omega) = Q_{|\psi_{k-1}\rangle}(f_k^{-1}(\Omega)) \left| \det \frac{\partial f_k^{-1}}{\partial \Omega} \right|$$
   • This architecture is naturally composable, mirroring the sequential structure of the quantum circuit.

3. Training Strategy:

   • The model is trained to minimize the Kullback-Leibler (KL) divergence between the distribution obtained via the exact transition kernel convolution and the distribution generated by the learned flow.
   • To ensure stability, flows are pre-trained sequentially ($k=1$ to $K$) and then jointly fine-tuned.
   • The initial state is typically the fiducial state $|00\dots0\rangle$, which has a known, efficient product-form Q function.

Key Contributions

• Continuous Phase Space Modeling: Unlike prior generative approaches that operate on discrete domains or require continuous relaxations, this method operates directly on the continuous manifold $(S^2)^N$, avoiding the need for hand-crafted embeddings.
• Composable Architecture: The framework assigns a specific normalizing flow to each gate, allowing the model to scale with circuit depth without retraining a fixed architecture after every step (a limitation of previous variational methods).
• Efficient Sampling: By leveraging the composability of flows, the method enables efficient sampling and likelihood estimation for large system sizes, transforming the problem of Hilbert space exponential growth into a problem of model expressivity.

Experimental Results
The framework was evaluated on three families of quantum circuits: 1D cluster state preparation, GHZ state preparation, and ground state preparation for a 1D transverse-field Ising model (TFIM) via Variational Quantum Eigensolver (VQE).

• Performance: The method was benchmarked against state-of-the-art neural network baselines, specifically IC-POVM (using transformers) and Neural Quantum States (NQS) with TDVP.
• Scalability:
  • For 1D cluster states at $N=64$ qubits, the model achieved a classical fidelity of $\approx 0.96$, outperforming IC-POVM ($0.91$ at $N=60$) and NQS ($0.84$ at $N=60$).
  • For GHZ states, the model reached $\approx 0.965$ fidelity, surpassing both baselines.
  • For the VQE circuit (10 qubits), the fidelity decreased linearly with gate count but remained higher ($0.985$) than the transformer baseline ($0.965$) over the tested depth.
• Observation: The accuracy of the model exceeds current neural network methods as system size and circuit depth increase.

Significance and Limitations
The paper claims that this approach opens a new avenue for simulating quantum circuits using generative models, particularly for large system sizes where other methods degrade.

• Computational Cost: The authors note a trade-off: while accuracy scales well, the training cost grows as $O(K^2)$ (where $K$ is the number of gates) because inverting previous flows is required for each training step. This contrasts with baseline methods that retrain fixed-size models ($O(K)$).
• Negativity and Approximation: The necessity of the flow arises from the "sign problem" (negative values in transition kernels). The approximation error compounds exponentially with circuit depth, manifesting as the observed fidelity decay in deeper circuits.
• State Dependence: The learned flows are approximate transport maps specific to the input distribution. While single-qubit gates admit exact, input-independent flows (rotations), entangling gates require retraining for different input states.

The authors conclude that this framework demonstrates the ability to scale to large system sizes and suggests future benchmarks in Trotterized dynamics, noisy evolution, and using the composed flow as a variational ansatz for optimizing circuit parameters.