arXiv:2611.03422
Efficient thermalization of one-dimensional noncommuting Hamiltonians via system-bath interactions with $Θ(n)$ ancilla qubits
Zhiyan Ding, Bowen Li, Ruizhe Zhang
Abstract
We prove rapid convergence of a system-bath interaction algorithm for preparing Gibbs states of one-dimensional noncommuting Hamiltonians at every fixed positive temperature. For a chain of $n$ qudits of local dimension $2^q$ with bounded nearest-neighbor interactions, the algorithm prepares the Gibbs state from any initial state to trace-norm error $0<\varepsilon\le1/2$ in expected total Hamiltonian evolution time $\widetilde{O}(n/\varepsilon)$. This improves the previously proved $\widetilde{O}(n^6/\varepsilon)$ bound for single-ancilla system-bath Gibbs samplers, up to logarithmic factors. The proposed algorithm couples $4^q-1$ resettable bath qubits to each site in parallel and randomizes the interaction time. Our analysis bounds both the one-round approximation error and the stationarity defect of an approximate stationary state in Gibbs-weighted $L^2$ norms, with only linear dependence on the system size. It also uses the dephasing induced by randomized evolution to control the extensive Lamb shift without treating it as a small perturbation of the dissipative dynamics.
AI Summary
The paper asks whether repeated system–bath interactions can match the system-size scaling of engineered Lindbladian Gibbs samplers while avoiding controlled forward/backward Hamiltonian evolution. It considers one-dimensional noncommuting nearest-neighbor Hamiltonians with bounded interactions, fixed local dimension $2^q$, and any fixed inverse temperature $\beta>0$. The protocol couples $4^q-1$ resettable bath qubits to every site simultaneously, using Gaussian interaction pulses, independently sampled bath frequencies, and randomized round durations. From any initial state, it prepares the Gibbs state to trace-norm error $\varepsilon$ in expected total Hamiltonian evolution time $\widetilde O(n/\varepsilon)$, compared with the previous $\widetilde O(n^6/\varepsilon)$ guarantee for single-ancilla system–bath samplers. The explicit round bound is $O((n/\varepsilon)[\log^4(en)+\log(1/\varepsilon)])$, with expected duration $O(\sqrt{\log(en/\varepsilon)})$ per round. The analysis compares each physical interaction channel with an effective Lindbladian channel in Gibbs-weighted $L^2$ norm, obtaining $O(n\alpha^4)$ approximation and approximate-stationarity defects rather than naive quadratic-in-$n$ bounds. It uses one-dimensional complex-time locality, centered quasi-local field and moment estimates, and dephasing from randomized evolution to control the extensive Lamb shift. Combined with existing spectral-gap and enhanced-dissipation estimates, these yield stationary-state bias $O(n\alpha^2)$ and permit $\alpha^2=\Theta(\varepsilon/n)$. The guarantee fixes temperature and local dimension; their dependence is absorbed into constants. Complexity counts Hamiltonian evolution time, not a full gate-level implementation. Unlike exact Gibbs-preserving Lindbladian samplers, the protocol retains $1/\varepsilon$ scaling because of stationary-state bias. Extension to a continuously coupled reservoir remains open.
Later, with the authors' guided prompting using the notes, ChatGPT 5.6 Sol worked out a fixed-point proof for more general, noncommuting systems.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.