5 Hirsch-Fye quantum Monte Carlo method for ... - komet 337

Hirsch-Fye QMC 5.23
Note the convergence factoreiωnη which is essential in order to get the correct result. Obviously,
such a term is difficult to handle numerically; in practice, η may be replaced, e.g., by the time
discretization parameter ∆τ, which also determines the cutoff frequency in the infinite sum.
This approximation can be avoided by evaluating the noninteracting part separately. For the
kinetic energy, this implies
Ekin = lim
η→0 +2T
∞�
e iωnη
� ∞
1
dǫǫρ(ǫ)
(43)
iωn −ǫ+µ−Σ(iωn)
� ∞
= 2
−∞
� ∞
≈ 2
−∞
n=−∞
−∞
ǫρ(ǫ)
dǫ
eβ(ǫ−µ) +1 +2T
ǫρ(ǫ)
dǫ
eβ(ǫ−µ) +1 +2T
∞�
n=−∞
L/2 �
� ∞
n=−L/2+1
dǫ ǫρ(ǫ)
−∞
� Gǫ(iωn)−G 0 ǫ
(iωn) �
(44)
� ∞
dǫ ǫρ(ǫ) � Gǫ(iωn)−G 0 ǫ(iωn) � , (45)
where we have assumed the paramagnetic case. Here, the interacting and noninteracting “momentumdependent”
Green functions read
Gǫ(iωn) =
−∞
1
iωn −ǫ+µ−Σ(iωn) ; G0 ǫ (iωn) =
1
. (46)
iωn −ǫ+µ
In (45), the terms in the Matsubara sum fall off at least as1/ω 2 , which makes it well-defined also
without convergence factor. At the same time, the truncation error is reduced significantly. The
complementary ingredient to the energy is the double occupancyD withE = Ekin+UD. In the
context of QMC calculations, this observable is best calculated directly from Wick’s theorem
(i.e. as 〈ni↑ni↓〉 or the corresponding expression in Grassmann variables) when sampling over
the auxiliary field. The overall behavior of D and E in the Hubbard model can be read off (for
T = 0.1) from the middle and lower parts of Fig. 13, respectively. For small U, the kinetic
energy increases quadratically while D and, consequently, Epot and E increase linearly. The
potential energy reaches a maximum below U = 3. A region of strong curvature of D, Epot,
and Ekin near U = 4.6 gives a rough indication of the metal-insulator crossover. The total
energyE, however, hardly shows any anomalies at this scale. Note also that the solutions forD
and E are quite close to the results of plain zero-temperature second-order perturbation theory.
The agreement actually becomes better for low-temperature QMC data, extrapolated to T → 0
(not shown). The offset of the curves for E gives (for not too large U) an indication of the
specific heatcV = dE/dT which is linear within the Fermi-liquid phase and is, in general, best
evaluated by fitting the temperature dependence ofE. 18
Susceptibilities
The direct evaluation of the compressibility (as of most other susceptibilities) is numerically
costly since it requires the QMC computation of 2-particle vertex functions. The formalism is
omitted here; it can be found, e.g., in [1, 46].
18 In Fermi liquid phases, the linear coefficient γ of the specific heat may also be obtained via the quasiparticle
weightZ (extrapolated toT = 0).