Technical Summary: Driven-Dissipative Ground State Preparation: Mixing Time and Randomness

1. Problem Statement

The preparation of ground states for large Hamiltonians is a central challenge in quantum information science, serving as a prerequisite for solving problems in quantum chemistry, materials science, and molecular modeling. Existing coherent quantum algorithms (e.g., phase estimation, adiabatic preparation, spectral filtering) typically suffer from costs that scale inversely with the spectral gap and the initial overlap between the starting state and the ground state. This overlap often decays exponentially with system size, rendering these methods inefficient for "cold starts" (arbitrary initial states).

While dissipative dynamics (reservoir engineering) offers a route to ground state preparation independent of initial overlap by making the ground state an attracting fixed point, previous constructions have largely relied on time-independent Lindbladians or randomized Lindbladians sampled from a fixed distribution. These approaches often struggle to provide rigorous mixing time bounds for general Hamiltonians without assuming specific locality or high-temperature regimes.

This paper addresses the gap by proposing driven-dissipative protocols where the generator (Lindbladian) varies over time. The goal is to construct iterative dissipative channels that drive arbitrary initial density matrices to low-energy subspaces, with mixing time bounds that depend solely on the Hamiltonian's eigenvalue distribution rather than its eigenbasis or locality.

2. Methodology

The authors propose two distinct iterative methods, both relying on the construction of randomized jump operators derived from the unitary evolution of random matrices under the target Hamiltonian $H$. The core strategy involves a sequence of stages $j=1, \dots, m$ (or $J$), where each stage reduces the support of the state from a larger spectral subspace $P_j$ to a smaller one $P_{j-1}$ (where $P_0$ is the ground state subspace).

General Framework

The protocol uses a sequence of Lindblad operators $L_j$. For a single jump operator $a$, the generator is $L_a(\rho) = 2a\rho a^* - a^*a\rho - \rho a^*a$. The protocol constructs $L_j$ as an empirical average of $M_j$ independent random jump operators:
$$L_j = \frac{1}{M_j} \sum_{k=1}^{M_j} L_{a_j(\omega_k)}$$
The jump operators are constructed via Fourier filtering of the unitary evolution of random matrices $g(\omega)$:
$$a(\omega) = \frac{1}{\sqrt{N}} \int_{-\infty}^{\infty} \phi(t) e^{itH} g(\omega) e^{-itH} dt$$
where $\phi(t)$ is a filter function chosen to select specific energy differences.

Method 1: General Spectrum (Section 3)

This method applies to Hamiltonians with a general spectrum where the eigenvalue distribution follows a "regular profile."

• Construction: At each stage $j$, a smooth Fourier filter $\hat{\phi}_j$ is designed to be non-zero only for energy differences in a specific interval $[\alpha'_j, \beta'_j]$. This ensures the jump operator maps states from the active subspace $P_j$ to a lower subspace $Q_j \subseteq P_{j-1}$.
• Randomness: The jump operators use random matrices $g$ with independent Gaussian entries (or unitary $t$-designs).
• Key Mechanism: The filter suppresses transitions that do not lower energy significantly. The "active corner" $P_j$ is compressed to $Q_j$, and the "enlarged" subspace $P^{\beta'_j}_j$ accounts for couplings created by the dissipator.
• Convergence: The mixing time is bounded by the spectral gap properties of the distribution. The authors prove that the empirical average converges to the ideal Lindbladian with high probability, provided the number of samples $M_j$ scales polynomially with system parameters.

Method 2: Clustered Spectrum (Section 4)

This method is designed for Hamiltonians where eigenvalues form distinct, well-separated clusters (e.g., spin chains, free fermions).

• Construction: The spectrum is partitioned into disjoint intervals (clusters) $I_j = [m_j - \delta, m_j + \delta]$. The jump operators are frequency components of the random matrix evolution, specifically targeting transitions between clusters separated by energy differences $w \in W = \{m_j - m_i\}$.
• Separation Assumption: Distinct cluster center differences must be separated by more than $6\delta$ to allow for precise filtering.
• Advantage: A single family of random matrices and filters serves all stages, differing only by normalization factors $1/N_j$. This simplifies the construction compared to Method 1, where filters change at every stage.
• Convergence: The method relies on the fact that the operator $K_M = \sum a^* a$ commutes with the cluster projections, ensuring that the dissipator preserves the cumulative corner structure while driving population from higher clusters to lower ones.

3. Key Contributions and Results

Theoretical Guarantees

The paper establishes rigorous bounds on the mixing time and resource requirements for both methods.

• Theorem 1.1 (Informal): For a Hamiltonian $H$ on $n$ qubits, there exist $m = O(\text{poly}(n))$ Lindblad operators and times $t_j$ such that the composition of channels drives any initial state $\rho$ to a state $\sigma$ supported on the ground state subspace $P_0$ with diamond norm error $\epsilon$.
• Mixing Time: The total evolution time depends on the spectral distribution (specifically the ratio of dimensions of subspaces and the spectral gaps) but is independent of the initial state's overlap with the ground state.
• Sample Complexity: The number of random matrices $M$ required to approximate the ideal Lindbladian is polynomial in the system size $n$, the inverse error $1/\epsilon$, and the inverse spectral gap parameter $\gamma^{-1}$.
  • For Gaussian samples, $M \sim O(\gamma^{-1} + \log(J/\zeta))$.
  • For unitary samples, $M \sim O(\gamma^{-2} p^4)$.

Specific Examples (Section 5)

The authors validate their methods on several Hamiltonian classes:

1. Number Operators with Geometric Coefficients ($H = \sum b^s n_s$): Method 1 successfully prepares the ground state for $b > 1$. The number of stages scales linearly with $n$ for $b < (3+\sqrt{5})/2$.
2. Geometric Eigenvalues: For Hamiltonians with eigenvalues $\Lambda_\ell = C x^{-\ell}$, the methods yield admissible stages with constant mixing parameters independent of system size.
3. Semicircular Quantiles (GUE-like): For Hamiltonians with eigenvalues distributed according to the semicircular law, the method achieves ground state preparation in $O(n)$ stages with polynomial sample complexity.
4. Spin Hamiltonians and Free Fermions: Method 2 is applied to the transverse field Ising model (mapped to free fermions) and spin chains. The method handles clustered spectra efficiently, with the number of stages equal to the number of clusters ($n$). The mixing time is bounded by $O(n \log n)$.

Numerical Validation

Numerical simulations (Figures 1, 4-7) demonstrate the convergence of the empirical dynamics. Trajectories show rapid decay of energy and population transfer from high-energy eigenstates to the ground state. The results confirm that the theoretical bounds on mixing time and sample counts are achievable in practice for moderate system sizes (e.g., $n=18$).

4. Significance and Claims

The paper claims to provide a systematic, driven-dissipative framework for ground state preparation that:

1. Eliminates the need for initial overlap: Unlike coherent algorithms, the cost does not depend on the initial state's fidelity with the ground state.
2. Relies only on spectral distribution: The construction requires a priori knowledge of the eigenvalue distribution (which can be estimated from a histogram) but does not require knowledge of the eigenbasis or the locality of the Hamiltonian.
3. Handles "Cold Starts": The protocols are valid for arbitrary initial density matrices.
4. Provides Polynomial Bounds: The resource costs (time and number of random matrices) are polynomial in the system size for a broad class of Hamiltonians, including those with geometric spectra and clustered structures.

The authors emphasize that while the theoretical construction uses idealized random matrices, the framework offers a concrete path to designing dissipative protocols where the generator varies in time to overcome the limitations of static reservoir engineering. The work bridges the gap between abstract dissipative state preparation and practical, time-dependent control strategies, offering a route to ground state preparation that is robust against the "overlap problem" inherent in many quantum algorithms.

The paper concludes by noting that the implementation cost of the random operators and filters remains a separate engineering challenge, but the theoretical bounds establish the feasibility of the approach for systems where the spectral distribution is known or estimable.