Title: Algorithmic Phases in Variational Quantum Ground-State Preparation

Abstract: The fidelity of a variational quantum circuit state prepared within stochastic gradient descent depends on, in addition to the circuit architecture, the number $N_s$ of measurements performed to estimate the gradient components. Simulating the variational quantum eigensolver (VQE) approach applied to two-dimensional frustrated quantum magnets, we observe that this dependence has systematic features. First, the algorithm manifests pronouncedly separated regimes on the $N_s$ axis with state fidelity $\mathcal{F}$ vanishing at $N_s < N_s^c$ and rapidly growing at $N_s > N_s^c$. {The point of transition $N_s^c$ is marked by a peak of energy variance, resembling the behaviour of specific heat in second-order phase transitions.} The extrapolation of the system-dependent threshold value $N_s^c$ to the thermodynamic limit suggests the possibility of obtaining sizable state fidelities with an affordable shots budget, even for large-scale spin clusters. Second, above $N_s^c$, the state infidelity $\mathcal{I} = 1 - \mathcal{F}$ satisfies $\mathcal{I} - \mathcal{I}_0 \propto 1/(\Delta^2 N_s)$, with $\mathcal{I}_0$ representing the circuit's inability to express the exact state, $\mathcal{F}$ is the achieved state fidelity, and $\Delta$ represents the system energy gap. This $1 / \Delta^2$ empirical law implies optimization resources increase inversely proportional to the squared gap of the system. We provide a symmetry-enhanced simulation protocol, which, in case of a closing gap, can significantly reduce the frustrated magnets simulation costs in quantum computers.