arXiv:2610.04176
Parallelizing Dissipative Quantum Algorithms
Sudatta Hor, Nathan Wiebe, Yongshan Ding
Abstract
Inspired by classical Markov Chain Monte Carlo (MCMC) methods, dissipative quantum algorithms are an emerging class of methods for preparing quantum states on a quantum computer. They operate by simulating a dissipative process governed by a mathematical operator known as the Lindbladian. The performance of these algorithms can be limited by the sequential implementation of the Lindbladian's constituent jump operators, which leads to high circuit depth. To address this issue, this work analyzes a parallelization scheme to implement jump operators simultaneously. We leverage the structure of geometric locality to show that individual jump operators remain spatially localized. This exponentially reduces the circuit depth per time step, resulting in an overall polynomial improvement. Numerical experiments demonstrate a 53x reduction in query depth for a 1D benchmark system of 100 qubits. The reduced circuit depth makes dissipative algorithms more practical for near-term quantum computing.
AI Summary
Dissipative state-preparation algorithms can gain faster mixing from bulk dissipation, but sequentially implementing many jump operators increases circuit depth. This paper studies parallel implementation for geometrically local Hamiltonians with filtered-Heisenberg jump operators $K_a=\int f(s)e^{iHs}A_ae^{-iHs}\,ds$, where $A_a$ is a single-site Pauli operator. Assuming exponential decay of both $f$ and its Fourier transform, with the time-domain decay rate exceeding a local interaction scale, Lieb–Robinson bounds give a localization radius $r=O(\log(nt/\epsilon))$ sufficient to approximate the full evolution to error $\epsilon$ in induced trace norm. Disjoint localized channels can then run in parallel using $O(r^D)$ layers per Trotter step. Combining first-order Trotterization, single-ancilla Stinespring dilations, quadrature, and explicit gate constructions yields query and circuit depth $\tilde O(nt^2/\epsilon)$ when filter and interaction parameters are fixed. The analyzed unlocalized sequential construction has circuit depth $\tilde O(n^5t^2/\epsilon)$. Stationary-state and mixing-time comparisons are also given, assuming unique stationary states. Classical simulations of a 100-site transverse-field Ising chain report a 53-fold reduction in estimated query depth at relative energy error $10^{-3}$. These are theoretical resource estimates, not hardware demonstrations. The useful numerical localization radii are smaller than the rigorous bound permits, and the numerical cooling filter only approximately suppresses upward transitions. Filter parameters can hide spectral-gap dependence; faster filter decay trades against mixing speed. Costs remain beyond current hardware, and higher-dimensional numerical tests are left open.
Central claim may not hold: The abstract's locality-based scope and 'exponential' depth-reduction language require qualification: the theorem imposes strong filter-decay assumptions and reduces O(n) jump-operator layers to polylogarithmically many, rather than giving an exponential speedup in system size. Table 1 labels circuit-depth bounds as query depth: the displayed derivations give sequential and single-jump query-depth scalings of approximately n⁴t² and n²t², not n⁵t² and n³t². The 53× numerical comparison establishes an energy-based resource estimate, not the theorem's full-state accuracy guarantee, and uses a localization radius for which stationary-state deviations are acknowledged. A nonzero exact cooling filter cannot satisfy the theorem's exponential time-decay assumption, because its Fourier transform would be analytic yet vanish on a real interval; approximate cooling requires additional stationary-state and mixing analysis not supplied here. The introduction's real-valued filter assumption makes the Pauli-derived jump operators Hermitian and the dynamics unital, contradicting unique ground-state cooling from arbitrary inputs.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.