Escolar Documentos
Profissional Documentos
Cultura Documentos
Neil Drummond
TCM Group, Cavendish Laboratory, University of Cambridge
XX
R i>j
qiqj
,
|ri rj R|
where qi is the charge of particle i and {R} are the lattice vectors.
Unfortunately this sum is conditionally convergent.
Riemann series theorem: can rearrange terms of a conditionally convergent sum to
get any answer you like. . .
Physically, O(r2) distant, neutral cells at distance r make dipole contributions
[O(r2)] to the electrostatic potential at any given point.
Practical solution: use Ewald method to calculate interaction energy.1
1
XX
R
where Q =
i qi
Q
qi(r ri R) ,
G6=0
G6=0
3/2
i
R
"
#
XX
3/2
2
qi (r ri R)
+
exp |r ri R|
.
qi
exp |r ri R|
3/2
exp(|r|2).
3/2 Z
All
0 2
|r|
exp(|r | ) 0 erf
dr
=
|r0 r|
|r|
space
X X 1 erf |r ri R|
Q
b(r) =
qi
,
|r
R|
i
i
R
b(r) dr = 0.
r
X X erfc |rj ri R|
Q
0
+
qi
2
qj
,
|rj ri R|
i
R
where
P0
i
1X 2
1X
qiqj vE (ri rj ) +
q vM
2
2 j j
i6=j
1X
qiqj [vE (ri rj ) vM ]
2
if Q = 0,
i6=j
where
vE (r) =
vM
2
X
X
erfc |r R|
exp[|G| /(4)] exp(iG r)
4
+
|G|2
|r R|
R
G6=0
r
2
X
X
erfc
|R|
4
exp[|G| /(4)]
.
2
|G|
|R|
G6=0
R6=0
+
+
+
+
+
+
+
+
Minimum Images
Minimum image distance between particles A and B: distance from A to closest
periodic image of B.
Translational Symmetry
Translational symmetries of the Hamiltonian:
1.
2.
1, . . . , ri + Rs, . . . , rN ) = H(r
1, . . . , ri, . . . , rN )
H(r
i {1, . . . , N },
1 + Rp, . . . , ri + Rp, . . . , rN + Rp) = H(r
1, . . . , ri, . . . , rN )
H(r
G. Rajagopal et al., Phys. Rev. Lett. 73, 1959 (1994); G. Rajagopal et al., Phys. Rev. B 51, 10591 (1995).
1.15
PBC
TABC
Inf. sys.
1.1
1.05
1
1
10
100
1000
Number of electrons
10000
-0.50
-0.55
-0.60
-0.65
-0.70
1
PBC
TABC
Inf. sys.
10
100
1000
Number of electrons
Twist averaging greatly dampens the energy fluctuations caused by shell filling.
10000
1
[k(r) + k(r)] = 2Re [k(r)]
2
1
[k(r) k(r)] = 2Im [k(r)]
2i
1
0
0
1
GS
000000
111111
0
GP
Finite-Size Extrapolation
Fit QMC data {EN } to
EN = E + a
TNHF
HF
T
b
,
N
where a, b and E are fitting parameters and TNHF is the HartreeFock (or DFT)
kinetic energy of an N -particle system.
HF
For electron gases etc. T
can be calculated exactly; otherwise it must be taken
from a HF/DFT calculation with a fine k-point mesh.
For real systems, obtaining EN at several different cell sizes can be time-consuming.
There are several variants of this fitting formula in the literature.
Better to correct finite-size errors than to rely on fitting.
1
h[
(r) (r)][
(r0) (r0)]i
N
Z
S (r0 + r, r0) dr0.
S (G) =
(G)
(G) (G) (G) ,
N
where (G) =
i exp(iG
Vee =
=
|(R)|2 12
N vM
i6=j vE (ri rj ) dR
R
+
2
|(R)|2 dR
Z Z
1
[(r, r0) (r)(r0)] [vE (r r0) vM ] dr dr0
2
Z Z
1
+
vE (r r0)(r)(r0) dr dr0
2
X 2
N X 4
+
[S(G)
1]
+
(G)
v
(G),
M
2
2
2
|G|
|G|
G6=0
DP
G6=0
E
0
where (r, r0) =
(r
r
)(r
rj ) is the pair density.
i
i6=j
First term: exchange-correlation energy (interaction of electrons with their XC holes).
Second term: Hartree energy (interaction of charge densities).
vE (r) vM
1 2 T
r
= +
,
r Wr + O
r 3
5/3
L. M. Fraser et al., Phys. Rev. B 53, 1814 (1996); A. J. Williamson et al., Phys. Rev. B 55, R4851 (1997).
XZ
1X
f (ri rj ) +
(r) [vE(ri r) f (ri r)] dr
=
2
i
i6=j
Z
1
1X
2
i6=j
X
X
1
4
iGri
f (ri rj ) +
e
G
A,
G
i
|G|2
G6=0
X
i
1
f0A,0
2
X
G6=0
1
4
A,
G
G
0
A,
0
A,
G
A,
0
|G|2
2
Fourier coefficients fG are evaluated numerically. Requires care because f (r) diverges
at r = 0 and is non-differentiable at the boundary of the Wigner-Seitz cell.
N
4 2
N
4 2
Z
Z
S(k) 1
2N X S(G) 1 N vM
dk
|k|2
|G|2
2
G6=0
S(k)
dk,
2
D |k|
To see this, insert factors of exp(|k|2 ) and exp(|G|2 ) in the integrand and summand and choose = 1/(2 )
in the Ewald expression for vM ; finally take the limit 0.
6
W. M. C. Foulkes, unpublished.
p/(4N ) (a.u.)
0.00154
0.000817
0.000369
EEwald is total SJ-DMC energy per particle obtained using Ewald interaction.
EMPC is total SJ-DMC energy per particle obtained using MPC interaction.8
EEwald was used in the branching factor in DMC, so the kinetic energy is the same in the two cases.
XX
XX
= s exp
u (ri rj ) +
u(ri rj) ,
> i,j
i>j
XX
1
u (G)
(G)
(G) + K .
= s exp
2
, G6=0
where Ts = 2 log(s(R))/4.
Use 2 [
(G)
(G)] = 2|G|2 [
(G)
(G) N ] to show that
X
X
1 X
2
hT i = hTsi +
|G|
u (G) h
(G)
(G)i
Nu(G) .
4
G6=0
(k) is only nonzero for G vectors of the primitive lattice. Assuming the sum runs
only over small G,
X
X
N X
1 X
2
2
hT i = hTsi +
|G|
|G|
u (G)S (G)
Nu(G).
4
4
G6=0
G6=0
-20
54 electrons
102 electrons
226 electrons
RPA
k u(k) (a.u.)
-10
-30
-40
0
k (a.u.)
k
u
(k) dk
3
4(2)
0
in 3D, where u
(k) is the spherical average of u(k) and D is the radius of a sphere of
volume (2)3/.
1
1
u(k) = 4A
.
2
2
k
k + 1/(2A)
where A is a parameter to be determined by fitting. Hence
1
tan
2AD
N
T =
D
4
2A
=
N A
+ O(N 2/3),
-0.065
-0.066
-0.067
-0.068
-0.069
0
Ewald
MPC
Ewald (corr.)
MPC (corr.)
-1
226
-1
102
-1
N
54
Correction: p/(2N ) for Ewald data and p/(4N ) for MPC data.
-1
Conclusions
QMC simulations of periodic systems suffer from finite-size errors.
Finite-size errors must be accounted for in any QMC study of condensed matter.
Methods for dealing with finite-size errors (use appropriate combinations of these):
1.
2.
3.
4.
5.
Can usually expect reasonable cancellation of finite-size biases when energy differences
are taken.