Warm Starts, Cold States: Exploiting Adiabaticity for Variational Ground-States

Reliable preparation of many-body ground states is an essential task in quantum computing, with applications spanning areas from chemistry and materials modeling to quantum optimization and benchmarking. A variety of approaches have been proposed to tackle this problem, including variational methods. However, variational training often struggle to navigate complex energy landscapes, frequently encountering suboptimal local minima or suffering from barren plateaus. In this work, we introduce an iterative strategy for ground-state preparation based on a stepwise (discretized) Hamiltonian deformation. By complementing the Variational Quantum Eigensolver (VQE) with adiabatic principles, we demonstrate that solving a sequence of intermediate problems facilitates tracking the ground-state manifold toward the target system, even as we scale the system size. We provide a rigorous theoretical foundation for this approach, proving a lower bound on the loss variance that suggests trainability throughout the deformation, provided the system remains away from gap closings. Numerical simulations, including the effects of shot noise, confirm that this path-dependent tracking consistently converges to the target ground state.

💡 Research Summary

The paper tackles the long‑standing challenge of preparing many‑body ground states on quantum computers, a task central to quantum chemistry, materials modeling, optimization, and sensing. While the Variational Quantum Eigensolver (VQE) is a popular hybrid quantum‑classical algorithm for this purpose, it often suffers from barren‑plateau phenomena and gets trapped in suboptimal local minima, especially as system size grows. To mitigate these issues, the authors propose an iterative “warm‑start” strategy that blends VQE with adiabatic‑inspired Hamiltonian deformation.

The core idea is to define a continuous interpolation between an easy‑to‑solve Hamiltonian (H_{\text{ini}}) (with a known ground state) and the target Hamiltonian (H_{\text{P}}). This interpolation is discretized into a sequence of intermediate Hamiltonians (H(x_k)=H_0 + x_k H_1) or equivalently (H(s)=(1-s)H_{\text{ini}}+sH_{\text{P}}) with parameters (x_k) (or (s_k)) increasing from the initial to the final value. At each step (k) a VQE optimization is performed for the current Hamiltonian, and the resulting optimal parameters (\theta^{}_{k}) are used to initialize the next VQE run. Concretely, the initialization is drawn from a small hyper‑cube around (\theta^{}_{k}), ensuring the optimizer starts in a region of non‑vanishing gradient.

The authors provide a rigorous theoretical analysis of gradient magnitudes along this path. They prove a lower bound on the variance of the gradient of the loss function that scales with the square of the minimal spectral gap (\Delta_{\min}) of the Hamiltonians involved, and only polynomially with the number of qubits. This bound guarantees that, as long as (i) the spectral gap remains open throughout the path, and (ii) successive Hamiltonians are sufficiently close so that (|\theta^{}_{k+1}-\theta^{}{k}|) stays within an (O(\Delta{\min})) radius, the gradient does not vanish exponentially. Consequently, the iterative procedure avoids barren plateaus and can reliably track the ground‑state manifold.

The framework is further extended to Meta‑VQE, where a single parameter vector (\theta) is trained to produce near‑optimal parameters for an entire family of Hamiltonians. By encoding the problem parameter (x) linearly into the rotation angles (i.e., (f_j(\theta_j,x)=g_j(x)\theta_j)), the same gradient‑variance bound applies to the averaged loss over the training set, provided the same gap conditions hold.

Numerical simulations validate the theory. The authors test the method on one‑dimensional Ising chains, Heisenberg spin chains, and small quantum‑chemical Hamiltonians (e.g., H₂). For each model they vary the number of interpolation steps (K) and include realistic shot noise (finite measurement samples). The results show that with modest (K) (10–50 steps) the final VQE converges to the true ground state with high fidelity (>0.99), even when the circuit depth is reduced compared to a single‑shot VQE. Importantly, the method remains robust under shot noise, confirming its practicality on near‑term devices. However, when the interpolation passes through a region where the spectral gap closes (e.g., near a quantum phase transition), the gradient bound fails, the warm‑start loses its advantage, and the optimizer can become trapped in a wrong basin—consistent with the theoretical prediction that ground‑state preparation becomes QMA‑complete in the presence of gap closings.

In summary, the paper introduces a discretized adiabatic‑inspired warm‑start VQE that systematically mitigates barren‑plateau issues and improves trainability for ground‑state preparation. The key requirements are (1) an easily preparable initial Hamiltonian, (2) a deformation path with a non‑vanishing spectral gap, and (3) sufficiently small steps to keep parameter updates within the gradient‑rich region. When these conditions are met, the approach enables shallow circuits to achieve high‑accuracy ground‑state preparation, offering a promising route for scaling variational algorithms on near‑term quantum hardware. Future work may explore optimal path design, gap‑preserving shortcuts, and experimental demonstrations on actual quantum processors.