arXiv:2610.03860
Driven-Dissipative Ground State Preparation: Mixing Time and Randomness
Paul Cazeaux, Marius Junge, Diyi Liu
Abstract
Dissipative protocols offer a route to ground state preparation that does not require a large initial overlap with the ground state. To move beyond time independent Lindbladians and randomized Lindbladians sampled from a fixed distribution, we study driven-dissipative protocols in which the generator varies over time. Our construction uses a family of randomized Lindbladians whose jump operators are obtained by averaging the unitary evolution of random matrices under a fixed target Hamiltonian. We show that empirical averages of these Lindbladians drive arbitrary initial density matrices to low energy states. The resulting mixing time bounds depend on the Hamiltonian only through its eigenvalue distribution and remain valid even for cold starts.
AI Summary
The paper studies whether time-dependent dissipative protocols can prepare low-energy states from arbitrary initial states using guarantees determined by the Hamiltonian’s eigenvalue distribution, without assumptions on its eigenbasis or initial ground-state overlap. Jump operators are constructed by frequency-filtering the Heisenberg evolution of Gaussian random matrices or random unitaries. Two constructions successively reduce nested low-energy spectral subspaces. For general spectra, stage-dependent filters direct jumps into a smaller subspace. Ideal dynamics contracts the complementary corner in diamond norm as $2e^{-\gamma_jt}+e^{-2\gamma_jt}$, where $\gamma_j$ measures available lower-energy spectral mass. Concentration bounds transfer this guarantee to empirical generators, with sufficient sample count $M_j=O(m^2t_j^2\epsilon^{-2}\chi_jp_j^4)$, where $\chi_j$ measures spectral enlargement and $p_j$ is logarithmic in dimension and inverse failure probability. For clustered spectra with distinct center differences separated by more than six cluster half-widths, frequency-resolved jumps preserve cumulative corners for every sample. Gaussian Gram-matrix bounds give $M=O(\gamma^{-1}+\log(J/\zeta))$ samples, independent of final accuracy, and stage times $t_j=2\gamma_j^{-1}\log(4J/\epsilon)$, with $\gamma_j=N_{j-1}/N_j$. Telescoping channel estimates yield uniform preparation guarantees, including for entangled inputs. Examples include semicircular quantiles and spin-number spectra. These are normalized evolution-time and sample-count bounds, not polynomial gate-complexity guarantees. Dense random operators, exponentially narrow filters, and stage normalization can entail large implementation costs. Spectral information is assumed known; random semicircular spectra require additional edge estimates. The clustered construction prepares only the lowest cluster unless it coincides with the ground space.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 53 pages, 7 figures