学术文献库

We present NFNet, a PyTorch-based framework for polynomial-time simulation of large-scale, continuously controlled quantum systems, supporting parallel matrix computation and auto-differentiation of network parameters. It is based on the non-interacting Fermionic formalism that relates the Matchgates by Valiant to a physical analogy of non-interacting Fermions in one dimension as introduced by Terhal and DiVincenzo. Given an input bit string $\boldsymbol{x}$, NFNet computes the probability $p(\boldsymbol{y}|\boldsymbol{x})=\langle x|U_θ^\dagger Π_y U_θ|x\rangle$ of observing the bit string $\boldsymbol{y}$, which can be a sub or full-system measurement on the evolved quantum state $U_{\mathbfθ}|x\rangle$, where $\mathbfθ$ is the set of continuous rotation parameters, and the unitary $U_{\mathbfθ}$'s underlying Hamiltonians are not restricted to nearest-neighbor interactions. We first review the mathematical formulation of the Matchgate to Fermionic mapping with additional matrix decomposition derivations, and then show that on top of the pair-wise circuit gates documented in Terhal and DiVincenzo, the Fermionic formalism can also simulate evolutions whose Hamiltonians are sums of arbitrary two-Fermion-mode interactions. We then document the design philosophy of NFNet, its software structure, and demonstrate its usage in various quantum system simulation, benchmarking, and quantum learning tasks involving 512+ qubits. As NFNet is both an efficient large-scale quantum simulator, and a quantum-inspired classical computing network structure, many more exciting topics are worth exploring, such as its connection to recurrent neural networks, discrete generative learning and discrete normalizing flow. NFNet source code can be found at https://github.com/BILLYZZ/NFNet.