Velocity of excitations in ordered, disordered, and critical antiferromagnets

We test three different approaches, based on quantum Monte Carlo simulations, for computing the velocity c of triplet excitations in antiferromagnets. We consider the standard S = 1/2 one- and two-dimensional Heisenberg models, as well as a bilayer Heisenberg model at its critical point. Computing correlation functions in imaginary time and using their long-time behavior, we extract the lowest excitation energy versus momentum using improved fitting procedures and a generalized moment method. The velocity is then obtained from the dispersion relation. We also exploit winding numbers to define a cubic space-time geometry, where the velocity is obtained as the ratio of the spatial and temporal lengths of the system when all winding number fluctuations are equal. The two methods give consistent results for both ordered and critical systems, but the winding number estimator is more precise. For the Heisenberg chain, we accurately reproduce the exactly known velocity. For the two-dimensional Heisenberg model, our results are consistent with other recent calculations, but with an improved statistical precision, c = 1.65847(4). We also use the hydrodynamic relation c(2) = rho(s)/chi(perpendicular to)(q -> 0) between c, the spin stiffness rho(s), and the transversal susceptibility chi(perpendicular to), using the smallest non-zero momentum q = 2 pi/L. This method also is well controlled in two dimensions, but the cubic criterion for winding numbers delivers better numerical precision. In one dimension, the hydrodynamic relation is affected by logarithmic corrections whichmake accurate extrapolations difficult. As an application of the winding number method, for the quantum-critical bilayer model our high-precision determination of the velocity enables us to quantitatively test, at an unprecedented level, the field-theoretic low-temperature scaling forms chi = (a/c(2))T and C = (b/c(2))T-2. We find agreement to within 3% with the leading 1/N results for the factors a and b in the O(N) model, but the agreement becomes significantly worse when the first corrections in 1/N are considered.