arXiv:2610.11537
An Efficient Quantum Circuit for Flow Model Execution Using Quantum Neural Networks
Rui Che, Ludvig af Klinteberg
Abstract
Flow models generate trajectories from an initial distribution to a target distribution by solving an ordinary differential equation defined by a velocity field. Flow matching learns this velocity field by modeling the transport dynamics between the two distributions. Wavefunction flow establishes a formal connection between flow models and quantum dynamics by introducing a continuity Hamiltonian, which drives the Schrödinger evolution of quantum states. In this paper, we investigate accurate and efficient quantum simulation of the wavefunction flow, thereby realizing the efficient implementation of flow models on quantum computers. We first leverage a quantum read-only memory (QROM)-based phase kickback framework for the wavefunction flow simulation, generating probability densities that closely match those produced by the corresponding conventional flow model. To address the high circuit-resource cost, we further incorporate a trained quantum neural network (QNN) into the phase kickback framework, replacing QROM for data encoding. Numerical experiments demonstrate that our proposed method implements flow models on quantum computers more efficiently, since it maintains the accuracy of wavefunction flow simulation compared with the QROM-based framework, and significantly reduces the circuit resources.
AI Summary
The paper studies circuit implementations of “wavefunction flow,” which represents classical distribution transport under a conservative velocity field $v_t=\nabla V_t$ as quantum evolution generated by the continuity Hamiltonian $H_t=i[K,V_t]$. It asks whether learned circuits can reduce the data-loading cost of simulating this evolution. The implementation uses Fourier spatial discretization, commutator product formulas, and phase kickback for potential and kinetic evolution. A baseline loads binary-encoded diagonal entries using QROM. The proposed replacement trains parameterized circuits, built from single-qubit and controlled rotations, to reproduce those binary entries on ancillas. Binary cross-entropy training is compared with value- and fidelity-based losses. For fixed ancilla count $m$ and layer count $l$, the proposed encoder uses $6mnl$ CNOTs, versus $\Theta(4^n)$ for the particular QROM decomposition analyzed; no general accuracy-versus-depth bound is established. GPU state-vector simulations test one-dimensional grids with 16 points and two-dimensional grids with $8\times8$ points. Across prescribed periodic potentials, QNN- and QROM-based final distributions have similar errors against numerical continuity-equation solutions: relative maximum errors are at most 0.022 and KL divergences below $2\times10^{-4}$. One five-address-qubit comparison reports 1,680 versus 39,424 encoder CNOTs. A separate flow-matching example learns a potential transporting a uniform distribution toward four Gaussian peaks; quantum simulation gives KL divergence 0.2935 against the target. Stated limitations include substantial training time because all discretized times enter training, and increased circuit resources because multidimensional operators require more than eight QNN layers for high precision.
Central claim may not hold: The claimed linear QNN execution cost assumes fixed depth, but the paper does not establish that fixed depth preserves accuracy as resolution or dimension increases, and its principal accuracy tests remain small. The resource comparison excludes training over all spatial addresses and time steps, so it establishes execution-gate savings on examples rather than end-to-end computational efficiency. The advertised Theta(4^n) QROM baseline comes from a particularly expensive lookup construction and gate decomposition, not an intrinsic QROM requirement, so it does not establish a general exponential-to-linear improvement.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.