arXiv:2609.22425
Gibbs resampling transitions and structured fast-measurement protocols
Daan Timmers, Benedikt Placke, Siddharth A. Parameswaran
Abstract
Quantum Gibbs sampling algorithms generalize classical Markov chain Monte Carlo (cMCMC) methods, and prepare thermal states on quantum computers by implementing local dissipative evolution for a mixing time $t_{\rm mix}$. Since quantum measurements disrupt the prepared state, the rate of extracting unbiased information from the system depends on the recovery of the underlying thermal correlations. The corresponding ''Gibbs resampling'' time $t_{\rm res}$ is distinct both from $t_{\rm mix}$ and from the autocorrelation time relevant to cMCMC sampling. We study this problem in effective classical models that emulate disruptive quantum measurements of local observables on extensive subsystems. When such subsystems are chosen randomly, in models with non-local order parameters we uncover a ''resampling transition'' as a function of the fraction of sites $p$ that are measured, from $t_{\rm res}\propto \log L$ for small $p$ to $t_{\rm res}\propto {\rm poly} L$ for large $p$, in systems of linear size $L$. We argue that this transition is generically absent in systems with local order parameters, and relate this dichotomy to one between the coarsening of initial states that break 1- and 0-form symmetries under Glauber dynamics. Finally, we devise a structured fast-measurement protocol that evades slow resampling even with non-local order parameters, while leaving only a vanishing fraction of all sites unmeasured.
AI Summary
The paper asks how quickly a prepared thermal state can recover after extensive disruptive measurements, rather than how long preparation from an arbitrary state takes. It studies a classical proxy: independently reset spins with probability $p$ to infinite-temperature configurations, then evolve with local Glauber dynamics. The models are the 2D Ising ferromagnet, 3D XY ferromagnet, and 3D $\mathbb{Z}_2$ lattice gauge theory. Numerics and heuristic coarsening arguments indicate rapid recovery, $t_{\rm res}=O(\log L)$, for every fixed $p<1$ in the ferromagnets, with fidelity to the initial orientation approaching unity. The proposed mechanism is that residual local magnetization confines defects to small regions. In the gauge theory, simulations instead locate a transition at $p_c\approx0.226$, above which recovery scales approximately as $L^2$ and logical fidelity falls. Mapping recovery to toric-code decoding gives an upper bound $p_c\lesssim0.466$ for faithful recovery under any decoder. A structured protocol measures boxes separated by unmeasured collars; defect-confinement arguments predict rapid recovery while allowing the measured fraction to approach one as box size increases. These conclusions concern classical models, not quantum Gibbs samplers: entanglement-mediated measurement effects are omitted. The logarithmic scaling is heuristic, and numerics do not exclude other polylogarithmic behavior. Recovery times exclude runs trapped in metastable states, although fidelity counts these as failures. Continuous-spin recovery uses a magnetization threshold rather than exact restoration. Generalization to all local-order-parameter phases is conjectural; point defects lack a general confinement argument, and finite-temperature collar estimates are heuristic.
Central claim may not hold: The reported quadratic slow-resampling times are conditioned on avoiding frozen trajectories; the unconditional mean is infinite at strictly zero temperature when freezing has positive probability, and can have a larger polynomial scaling at finite temperature. The structured protocol establishes logarithmic resampling for fixed box size and an arbitrarily small fixed unmeasured fraction, but does not establish the joint limit of logarithmic resampling and an unmeasured fraction vanishing with system size. The finite-temperature structured protocol's claim of exactly unit logical fidelity is unsupported because its nonzero collar-crossing probabilities are not converted into a system-wide logical-failure bound.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.