Augmenting Imaginary-Time Evolution with Local Geometric Information

Problem Statement
Imaginary-time evolution (ITE) serves as a foundational mechanism for ground-state preparation across quantum simulation, condensed matter physics, and quantum chemistry. By applying the non-unitary propagator $e^{-\tau \hat{H}}$, ITE exponentially suppresses excited-state components, driving a system toward its ground state. However, the convergence of standard ITE is governed strictly by the energy variance of the instantaneous state. This results in an asymptotic exponential decay of the energy error, meaning the system approaches the ground state only as $\tau \to \infty$. The rate of convergence is sensitive to the system's energy gap and correlation length, often leading to prohibitively slow convergence in practical applications. While significant effort has been devoted to implementing ITE efficiently across various algorithmic branches (direct, stochastic, variational, double-bracket, and spectral-filter methods), the underlying dynamics have remained largely unchanged, leaving a gap between the conceptual utility of ITE and its computational efficiency.

Methodology
The authors introduce Augmented Imaginary-Time Evolution (AITE), a framework that replaces the standard gradient flow on the energy landscape with a geometrically informed descent along locally optimal directions.

1. Double-Bracket Formalism: The approach builds upon the double-bracket formulation of ITE, where the evolution is expressed as $\partial_\tau |\psi(\tau)\rangle = -[\hat{H}, \hat{\rho}(\tau)]|\psi(\tau)\rangle$. The authors generalize this by replacing the projection operator $\hat{\rho}(\tau) = |\psi(\tau)\rangle\langle\psi(\tau)|$ with a general projection operator $\hat{O}(\tau)$, yielding $\partial_\tau |\psi(\tau)\rangle = -[\hat{H}, \hat{O}(\tau)]|\psi(\tau)\rangle$.
2. Optimization of Descent Direction: To maximize the rate of energy decrease, the authors minimize the linear functional $\text{Tr}[\hat{\chi}(\tau)\hat{O}(\tau)]$, where $\hat{\chi}(\tau) = -[\hat{H}, [\hat{H}, \hat{\rho}(\tau)]]$ is the double-bracket operator encoding the local curvature of the energy landscape.
3. Krylov Subspace Restriction: To make the optimization tractable, the search for the optimal $\hat{O}(\tau)$ is restricted to the second-order Krylov subspace spanned by the instantaneous state $|\psi(\tau)\rangle$ and the centered Hamiltonian action $\bar{H}|\psi(\tau)\rangle$.
4. Role of Skewness: Within this subspace, the projected double-bracket operator depends on the energy variance $\mu_2(\tau)$ and the instantaneous Fisher–Pearson skewness coefficient $\kappa(\tau) = \mu_3(\tau)/\mu_2^{3/2}(\tau)$. The optimal descent operator $\hat{\rho}_A(\tau)$ is identified as the projector onto the lowest eigenvector of this projected operator. Crucially, this direction incorporates the asymmetry (skewness) of the instantaneous energy distribution, a feature ignored by standard ITE which implicitly assumes a symmetric (Gaussian-like) distribution.
5. Augmented Hamiltonian: The authors demonstrate that this optimal flow can be reproduced by an augmented non-unitary ITE using an effective Hamiltonian $\hat{H}_A(\tau) = \cos(2\phi)\bar{H} + \sin(2\phi)\frac{1}{2\sqrt{\mu_2}}\bar{H}^2$, where $\phi$ is a mixing angle determined by the skewness.
6. Heuristics: To reduce computational overhead associated with estimating the third moment $\mu_3$ (required for exact skewness), the authors propose a mean-field approximation $\langle \hat{H}^3 \rangle \approx \langle \hat{H}^2 \rangle \langle \hat{H} \rangle$, which expresses skewness in terms of first and second moments.

Key Contributions and Results

• Superlinear Convergence: Theoretical analysis reveals that AITE strictly outperforms standard ITE. The convergence rate includes a superlinear speedup factor $\sqrt{1 + \Delta/4\varepsilon}$, where $\Delta$ is the spectral gap and $\varepsilon$ is the energy error. This factor diverges as the error approaches zero, indicating increasingly aggressive acceleration near the ground state.
• Finite-Time Extinction: Unlike standard ITE, which exhibits asymptotic exponential decay, AITE achieves finite-time extinction of the energy error. The energy error vanishes exactly at a finite imaginary time $\tau^*$, following a power-law decay $\varepsilon(\tau) \propto (\tau^* - \tau)^2$ near extinction. This places AITE in a qualitatively distinct convergence class compared to ITE.
• Universal Upgrade: The framework is designed as a universal upgrade for the ITE algorithmic family. By substituting $\hat{H} \to \hat{H}_A(\tau)$, the acceleration can be applied to direct, stochastic, variational, double-bracket, and spectral-filter implementations without altering their fundamental ansatz or circuit structures.
• Numerical Validation: Benchmarks on the hydrogen chain ($H_8$) and the 1D Fermi-Hubbard model across weakly and strongly correlated regimes confirm the existence of three dynamical regimes: a linear regime (tracking standard ITE), a superlinear regime (driven by skewness), and a power-law extinction regime. The mean-field skewness heuristic is shown to capture the convergence characteristics of the exact method effectively.

Significance
The paper claims to close the gap between the conceptual power of ITE and its practical computational efficiency by addressing a structural limitation previously unaddressed in the literature: the implicit assumption of symmetric energy distributions in standard ITE. By exploiting higher-order statistical structure (specifically skewness) to identify locally optimal descent directions, AITE transforms the convergence behavior from asymptotic exponential decay to finite-time power-law extinction. The authors emphasize that this improvement is not tied to a specific implementation but offers a systematic pathway to accelerate the entire family of ITE-based algorithms, providing a provably faster route to ground-state preparation in quantum simulation.