PHYSICAL REVIEW E VOLUME 62, NUMBER 6 DECEMBER 2000

Stabilization of the lattice Boltzmann method by the H theorem: A numerical test

Santosh Ansumali and Iliya V. Karlin*
ETH-Zürich, Department of Materials, Institute of Polymers, ETH-Zentrum, Sonneggstrasse 3, ML J 27, CH-8092 Zürich, Switzerland
共Received 7 June 2000兲
For a one-dimensional benchmark shock tube problem, we implement the lattice Boltzmann method based
on the H theorem 关I. Karlin, A. Ferrante, and H. C. Öttinger, Europhys. Lett. 47, 182 共1999兲兴. Results of
simulation demonstrate significant improvement of stability, as compared to realizations without explicit en-
tropic estimations.

PACS number共s兲: 47.11.⫹j, 05.20.Dd

I. INTRODUCTION lision integral ⌬ based solely on the knowledge of the en-

Since the invention of the lattice-gas model 关1兴, lattice-
based methods for simulations of complex hydrodynamic
phenomena received much attention over the past decade. In
these methods, hydrodynamic equations are not addressed by
a direct discretization procedure, rather, a simple pseudopar-
ticle kinetics is introduced in such a way that the hydrody-
namic equations are obtained on the large space and time
scale. Particularly promising is the well-known lattice Bolt-
zmann method 共LBM兲 关2兴. It is based on the fully discrete
velocity-space-time kinetic equation of the form,
N共 x⫹c,t⫹1 兲 ⫺N共 x,t 兲 ⫽⌬关 N共 x,t 兲兴 . 共1兲
Here N(x,t) is the b-component vector of populations N i of
the pseudoparticles with velocities c i , at the sites x of a
lattice at discrete time t. The system of discrete velocities at
any site is formed by the outgoing links of the lattice, and it
also may include the zero vector.
One of the most important problems related to the LBM,
recognized by many authors, is the problem of numerical
stability. For the LBM related to incompressible flow simu-
lations, numerical instabilities preclude so far a study of high
Reynolds number flow situations. Instabilities become even
more annoying for compressible flows 关3兴.
It has been discussed for some time in the literature that
stability of the LBM could be improved if the method could
be equipped with an analog of the Boltzmann H theorem.
Recently, theoretical progress in this direction has been
achieved 关4–8兴. In particular, for the isothermal LBM, the
hydrodynamic fields are the density, ␳ ⫽(1,N), and the av-
erage momentum, ␳ u ␣ ⫽(c␣ ,N), where (•,•) denotes the
standard scalar product in the b-dimensional space of popu-
lation vectors. In this case, one can construct entropy func-
tions in such a way that its local equilibrium implies the
crucial relation for the stress tensor,
共 c␣ c␤ ,Neq兲 ⫽c s2 ␳ ␦ ␣␤ ⫹ ␳ u ␣ u ␤ ,
up to the admissible degree of accuracy of the LBM 关6兴.
Furthermore, it has been suggested how to construct the col-
*Corresponding author. FAX: ⫹41 01 632 10 76. Email address:
ikarlin@ifp.mat.ethz.ch
tropy function, and how to stabilize the updates on the basis
of the discrete-time H theorem 关6,8兴. In the sequel, we term
the LBM based on the H theorem the entropic lattice Boltz-
mann method 共ELBM兲.
It is the goal of this paper to test the aforementioned the-
oretical developments for a shock tube problem. This one-
dimensional benchmark problem has been suggested some
time ago 关9兴 for testing various ideas in the LBM. Though
this model is based on a very simple three-velocity lattice, it
provides a stringent test of stability. Implementation of the
ELBM demonstrated a large improvement of stability in this
benchmark problem. In fact, we were able to reach the values
of the kinematic viscosity as low as 10⫺12 without any sign
of numerical instabilities. The most important part of the
realization is a robust root-finding procedure which imple-
ments the H theorem.
II. CONSTRUCTION OF THE ELBM
An advantage of the LBM in comparison to the lattice-gas
method is that the Galilean invariance of the Navier-Stokes
equation is easier to control in the former than in the latter.
In order that this advantage should not get lost in the ELBM,
entropy functions should be found for each lattice separately.
We here consider the one-dimensional lattice with spacing c,
and the population vector at each site x has three compo-
nents, N⫽(N ⫹ ,N 0 ,N ⫺ ) † , corresponding to velocities c ⫹
⫽c, c 0 ⫽0, and c ⫺ ⫽⫺c, respectively. For this model, the
entropy function has been found in 关6兴
H⫽N 0 ln共 N 0 /4兲 ⫹N ⫺ ln共 N ⫺ 兲 ⫹N ⫹ ln共 N ⫹ 兲 . 共2兲
Realizations of the ELBM based on the entropy function 共2兲
result in the one-dimensional Navier-Stokes equation with
the sound speed c s ⫽ 冑1/3c, within the accuracy of the order
(u/c s ) 4 . Construction of the ELBM involves the following
two steps 共these steps are independent of the choice of the
entropy兲.
First, we specify bare collision integral ⌬ in such a way
as to satisfy the admissibility condition, (⌬,1)⫽(⌬,c␣ )⫽0,
(⌬,ⵜH)⭐0, and ⌬(Neq)⫽0. Here ⵜH is the gradient of H
in the space of population vectors. The choice of the bare
collision integral is not unique. We here consider three cases,
⌬⫽ 共 g⫹ ⫺g⫺ 兲 兵 exp关共 “H,g⫺ 兲兴 ⫺exp关共 “H,g⫹ 兲兴 其 , 共3a兲

1063-651X/2000/62共6兲/7999共5兲/$15.00 PRE 62 7999 ©2000 The American Physical Society

8000 SANTOSH ANSUMALI AND ILIYA V. KARLIN PRE 62

⌬⫽ 共 g⫹ ⫺g⫺ 兲共 “H,g⫺ ⫺g⫹ 兲 , 共3b兲 where M 2 ⫽u 2 /c s2 is the Mach number squared. However,
result 共7兲 is the exclusive case which does not happen in
⌬⫽Neq共 N兲 ⫺N, 共3c兲 higher dimensions.

where g⫹ ⫽(1,0,1) † , and g⫺ ⫽(0,2,0) † are positive and
negative parts of the vector g⫽g⫹ ⫺g⫺ , the latter is orthogo-
nal to the vectors of conserved fields, (g,1)⫽0, and (g,c)
⫽0. Collision integral of the form 共3a兲 has been suggested in
Ref. 关6兴, and is motivated by well-known models of chemical
kinetics in the framework of so-called Marcelin-De Donder
kinetic function 关10兴. Collision integral 共3b兲 is a linearized
form thereof, and it is motivated by recently introduced
GENERIC models of nonequilibrium thermodynamics 关11兴.
Equation 共3c兲 is the familiar Bhatnagar-Gross-Krook 共BGK兲
form.
Second, the population vector is updated according to the
kinetic equation
N共 x⫹c,t⫹1 兲 ⫺N共 x,t 兲 ⫽⌬* 关 N共 x,t 兲兴 , 共4兲
where ⌬* is a dressed 共or stabilized兲 collision integral,
⌬* 关 N共 x,t 兲兴 ⫽ ␤␣ 关 N共 x,t 兲兴 ⌬ 关 N共 x,t 兲兴 . 共5兲
III. IDENTIFICATION OF VISCOSITY
Identification of the viscosity coefficient in the ELBM is
done on the basis of the Chapman-Enskog analysis 关12兴 in
the vicinity of the local equilibrium, in the same way as in
the standard lattice Boltzmann realizations. Let us do this
derivation in some detail for the example given here.
Linearization of the dressed collision integral 共5兲 may be
written as
␦ ⌬* ⫽ ␤␣ 共 Neq兲共 “⌬兩 Neq, ␦ N兲 ⫹ ␤ 共 “ ␣ 兩 Neq, ␦ N兲 ⌬共 Neq兲 .
共8兲
The last term on the right-hand side of the latter expression is
equal to zero by the construction of the bare collision inte-
gral. In the sequel, we denote as L the matrix of the deriva-
tives of the bare collision integral at the local equilibrium,
and write ␣ eq⫽ ␣ (Neq),
␦ ⌬* ⫽ ␤␣ 共 Neq兲 L␦ N. 共9兲

Here ␤ 苸 关 0,1兴 is a parameter related to viscosity 关see Eq. For the bare collision integrals 共3a兲 and 共3b兲, the components
共20兲 below兴, and ␣ is the scalar function of the population of the matrix L have the form
vector. Function ␣ ensures the discrete-time H theorem, and
is the nontrivial root of the scalar nonlinear equation,
L i j ⫽⫺K eqg i
⳵ 2H
兺k ⳵ N j ⳵ N k 冏 gk ,
Neq
共10兲
H 共 N兲 ⫽H 共 N⫹ ␣ ⌬ 关 N兴 兲 . 共6兲
where positive scalar functions K eq are K eq
Put differently, bare collision integrals are stripped of any ⫽exp关(“H兩Neq,g⫾ ) 兴 and K eq⫽1 for the collision integrals
relaxation time parameters, and are merely directions in the 共3a兲 and 共3b兲, respectively. Function ␣ eq is found upon ex-
space of populations, pointing towards the change of the panding Eq. 共6兲 at equilibrium up to the quadratic in ␦ N
state in the collision event. Parameter ␣ defines the maximal terms. 关Note that a substitution of Neq into Eq. 共6兲 does not
admissible collision step along this direction so that the en- give an equation for ␣ eq兴. This results in the following qua-
tropy will not decrease. The combination ␤␣ is thus the dratic equation:
effective relaxation time in the fully discrete kinetic picture.
An advantage of the bare BGK collision integral in the
ELBM scheme is not obvious: In most cases, the local equi-
librium is not known as an explicit function of the hydrody-
␣ eq 冉 1
2
␣ eq 兺
i jk
␦Ni
⳵ 2H
⳵ N i⳵ N j
冏 Neq
L jk ␦ N k

namic fields, and has to be evaluated numerically on each

iteration of the method for each lattice cite. However, after
the local equilibrium is found, the resulting bare BGK colli-
⫹ 兺
i jkl
Li j␦N j
⳵ 2H
⳵ N i⳵ N k
冏 Neq
冊
L kl ␦ N l ⫽0. 共11兲

sion integral must be dressed by numerically solving Eq. 共6兲.

Thus, in comparison to bare collision integrals 共3a兲 and 共3b兲
which only require the knowledge of the entropy but not of
the local equilibrium, numerical efforts roughly double. In
the example considered here we were able to find analyti-
cally the local equilibrium of the H function 共2兲,
2␳
N 0⫽ 关 2⫺ 冑1⫹M 2 兴 ,
3
For the bare BGK collision integral 共3c兲, it has been already
demonstrated elsewhere 关4兴 that the nontrivial solution to
this equation results in ␣ eq⫽2, so we shall discuss only the
cases 共3a兲 and 共3b兲. Using the explicit form of the linearized
bare collision integral 共10兲, and the explicit form of the sec-
ond derivative of the entropy function, the nontrivial root of
Eq. 共11兲 is found to be
2
␣ eq⫽ 共12兲

冋 册
.

N ⫹⫽
␳ uc⫺c s2
⫹ 冑1⫹M 2 , 共7兲
K eq 兺i j g i 共 N eq
i 兲
⫺1
gi
3 2c s2
Thus, Eq. 共12兲 together with Eq. 共10兲 defines the linearized
␳
N ⫺⫽ ⫺
3 冋
uc⫹c s2
2
2c s
⫹ 冑1⫹M 2 , 册 dressed collision integral,

␦ ⌬* ⫽⫺2 ␤ L* ␦ N, 共13兲
PRE 62 STABILIZATION OF THE LATTICE BOLTZMANN . . . 8001

where c s2 共 1⫺ ␤ 兲
␯⫽ . 共20兲
⫺1 2␤
g i 共 N eq
j 兲 gj
ij⫽
L* . 共14兲
Thus, the ELBM is able to retain full control over the vis-
兺m eq ⫺1
g m共 N m 兲 gm cosity, while variation of the parameter ␤ in the interval
关 0,1兴 covers the full linear stability interval, and the limit
Now it is important to notice that operator L* has the same ␤ →1 corresponds to the zero velocity limit 关4,6兴.
projector property as the linearized BGK operator, Several remarks to the derivation just given are in order:
While the local equilibrium Neq formally appears at the in-
L* L* ⫽L* . 共15兲 termediate state of computation, and, in particular, in the
formula for ␣ eq 共12兲, the result for the viscosity 共20兲 is inde-
The image of the operator L* is the linear subspace spanned pendent of it, and, in fact, the derivation has circumvented
by the vector g. the explicit use of Neq. This is a direct consequence of the
With this description of the linearized dressed collision projector property 共15兲. In our example, the projector prop-
integral, we now follow the standard Chapman-Enskog erty, in turn, follows from the fact that the kinetic subspace
analysis, and seek the solution to the kinetic equation 共4兲 in of the model is one dimensional. Thus, all the admissible
the form, N⫽Neq⫹ ␦ Nne, where the nonequilibrium part collision integrals like Eq. 共3兲, and any others, become
␦ Nne is orthogonal to the hydrodynamic subspace, (1, ␦ Nne) equivalent near the local equilibrium, and they all result in
⫽(c, ␦ Nne)⫽0, and is found in terms of the expansion, the same viscosity coefficient. This is not the case when the
␦ Nne⫽ ⑀ ␦ N(1) ⫹ ⑀ 2 ␦ N(2) ⫹O( ⑀ 3 ), subject to the multiscale dimension of the kinetic subspace is larger than one. In that
expansion of the time and space derivatives, ⳵ t ⫽ ⑀ ⳵ (1) t
case, various admissible bare collision integrals may lead to
⫹ ⑀ 2 ⳵ (2) different expressions for the viscosity. Nevertheless, it is al-
t ⫹O( ⑀ ), ⳵ x ⫽ ⑀ ⳵ x ⫹O( ⑀ ). Then,
3 (1) 2
ways possible to construct bare collision integrals which, un-
like the BGK, do not use the local equilibrium explicitly, and
⫺2 ␤ 兺j L *i j ␦ N (1)
j ⫽ 关 ⳵ t ⫹c i ⳵ x 兴 N i ,
(1) eq
共16兲 at the same time their linearization satisfies the projector
property 共15兲. This construction will be reported in a sepa-
rate publication 关14兴. Finally, it should be stressed that the
⫺2 ␤ 兺j L *i j ␦ N (2)
j ⫽ ⳵ t N i ⫹ 关 ⳵ t ⫹c i ⳵ x 兴
(2) eq (1) theoretical derivation of the viscosity coefficient is always
strictly applicable only within the domain of validity of the

冋 册
Chapman-Enskog analysis in the vicinity of the local equi-
⫻ ⫺␤ 兺j L *i j ␦ N (1)
j ⫹␦Ni
(1)
. librium.

共17兲 IV. IMPLEMENTATION

The ELBM algorithm differs from the standard LBM in

By the Fredholm alternative, the solution to Eq. 共16兲 is writ-
that the nonlinear equation 共6兲 has to be solved at each time
ten
step on each lattice site. Although the time required for solv-
ing this equation does not contribute too much to the total
␦ N(1) ⫽ ␦ Nspec
(1)
⫹ ␦ Nhom
(1)
,
run time, a robust algorithm is required for solving this
highly nonlinear equation. Near the local equilibrium, the
where ␦ Nhom
(1)
is the general solution to the homogenous
Newton-Raphson method fails because the first derivative
equation, L* ␦ Nhom
(1)
⫽0, and ␦ Nspec
(1)
is a special solution to the tends to zero very rapidly. For this reason, we have con-
inhomogeneous equation 共16兲. The homogeneous solution is structed an algorithm which uses successive substitutions
equal to zero by the orthogonality condition mentioned near the local equilibrium 共if possible兲, and it uses a combi-
above. The special solution has the form nation of the Newton-Raphson and of the bisection method
关15兴 far away from the local equilibrium. The initial approxi-
␦ Nspec
(1)
⫽Ag. mation, in most cases, was taken as the solution obtained
from the quadratic expansion, but very far away from the
Thus, by the projector property 共15兲, Eqs. 共16兲 and 共17兲 are local equilibrium, the solution to the equation, N⫹ ␣ ⌬⫽0,
equivalent to the following two equations for the special so- ␣ ⬎0, is a better guess 关6兴. Details of the code are available
lution 共we omit the subscript spec兲: from the authors.
⫺2 ␤ ␦ N (1)
i ⫽ 关 ⳵ t ⫹c i ⳵ x 兴 N i ,
(1) eq
共18兲
V. SHOCK TUBE TESTS

⫺2 ␤ ␦ N (2)
i ⫽ ⳵ t N i ⫹ 共 1⫺ ␤ 兲关 ⳵ t ⫹c i ⳵ x 兴 ␦ N i . 共19兲
(2) eq (1) (1)
We have studied the time evolution of a one-dimensional
front in a shock tube, a very classical problem in which it
The latter set of equations coincides with the well known appears a compressive shock front, moving in the low den-
case of the LBGK, in which the BGK relaxation parameter sity, and a rarefaction front moving in the high-density re-
␶ ⫺1 is replaced by 2 ␤ , and we are immediately led to the gion 关9兴. These two fronts leave an intermediate region in the
following viscosity coefficient for each of the bare collision central portion of the tube with uniform density ␳ c , and uni-
integrals 共3兲: form velocity u c . The tube is filled at time t⫽0 with a gas at
8002 SANTOSH ANSUMALI AND ILIYA V. KARLIN PRE 62

FIG. 1. Density profile 共dimensionless lattice units兲 at t⫽500 FIG. 2. Velocity profile. Simulation setup and notation same as
for viscosity ␯ ⫽3.3333⫻10⫺2 . Thin line: Exact solution. Symbol: in Fig. 1.
Simulation.
significant decrease of ␣ in the simulation, its deviations
rest with uniform density ␳ ⫺ (u ⫺ ⫽0) for x⬍0, and ␳ ⫹ (u ⫹ from Eq. 共6兲 may explain smoothening of the density and of
⫽0) for x⬎0. For the inviscid case, ␯ ⫽0, the density and the velocity profiles at the shocks. On the other hand, we
velocity profiles present a discontinuity across the shock. have not observed a development of postshock oscillations
The shock speed 关9兴 is given by the Rankine-Hugoniot rela- by the ELBM algorithm.
tions, In the ELBM based on the entropy 共2兲, the sound velocity,
c s⫽ 冑1/3c, is not kept constant by imposing a constraint.
r c ⫺1 ␳c Deviations of the effective sound velocity,
v s2 ⫽r c c s2 , u c⫽ cs , r c⫽
冑r c ␳⫹

and r c and be obtained by using the Navier-Stokes relation, c s⬘ ⫽ 关 ␳ ⫺1 c 2 共 N ⫺

eq
⫹N ⫹
eq
兲 ⫺u 2 兴 ,

r c ⫺1 ␳⫺ from c s measure the influence of anomalous terms 关we note

logr c ⫹ ⫽log .
冑r c ␳⫹ that these terms are of the order u 4 by the choice of the
entropy 共2兲兴. The error E⫽(c s⬘ ⫺c s)/c s was evaluated using
Simulations were performed in order to compare the stability the exact local equilibrium 共7兲. In Fig. 3 the error is given for
of the three LBM algorithms: the nonlinear LBE 关9兴 共LBE the simulation with the viscosity ␯ ⫽1.6⫻10⫺13. It was
hereafter兲, the LBGK method with the polynomial equilib- found that the error was of the order of 0.1% even though the
rium ansatz 关13兴, and the present ELBM algorithm with the Mach number was as high as 0.3464 in the present simula-
bare collision integral 共3a兲. tion.
Runs were performed on the lattice with 800 nodes. At t
⫽0 the lattice was populated as to give the density ␳ ⫺
⫽1.5 for 0⭐x⭐400, and ␳ ⫹ ⫽0.75 for 400⬍x⭐800. Stan-
dard bounce back boundary conditions were applied at both
ends of the tube. Results for the three algorithms are demon-
strated in Figs. 1 and 2 for a relatively high value of viscosity
␯ ⫽3.3333⫻10⫺2 . This value was taken in order to compare
all the three algorithms because it is close to the instability of
the LBE 关9兴. It has been found that the LBGK and the ELBM
never showed divergence when the viscosity was smaller
than ␯ ⬍10⫺3 . However, in contrast to the ELBM, the results
of the LBGK demonstrate large fluctuations already at ␯
⬍1. For this reason, results of the ELBM were always better
in comparison to the LBGK at low viscosity. It should be
also stressed that because the theoretical derivation of the
viscosity coefficient 共20兲 is valid only in the vicinity of the
local equilibrium, it may become invalid at the shocks where
the populations may be far away from the local equilibrium.
Increase in the effective viscosity near shocks would mean
that the evaluated parameter ␣ 共6兲 is smaller than the near- FIG. 3. Deviation of effective sound velocity from the exact
equilibrium bound 共12兲. While we have not observed a very sound velocity, E⫽(c s⬘ ⫺c s)/c s , at ␯ ⫽1.6⫻10⫺13.
PRE 62 STABILIZATION OF THE LATTICE BOLTZMANN . . .
VI. DISCUSSION AND CONCLUSIONS
The numerical test of the entropic lattice Boltzmann
method demonstrated much better stability of this method in
comparison to the LBM without the H theorem. The natural
working window of this approach is low velocity (u→0)
and low kinematic viscosity ( ␤ →1). Low values of velocity
are required in order to keep small the anomalous terms of
the order u 4 , whereas the entropic estimation 共6兲 allowed
implementations very close to the zero viscosity limit.
While the entropy-based lattice Boltzmann methods are,
per construction, unconditionally stable, a word of caution
about their realizations is in order. On the one hand, the
theoretical limit of zero kinematic viscosity is at the same
time the limit of no convergence to the local equilibrium, the
population vectors becomes trapped on the fixed level of the
entropy function. Furthermore, implementation of the colli-
关1兴 U. Frisch, B. Hasslacher, and Y. Pomeau, Phys. Rev. Lett. 56,
1505 共1986兲.
关2兴 Excellent reviews of the lattice Boltzmann method can be
found in R. Benzi, S. Succi, and M. Vergassola, Phys. Rep.
222, 147 共1992兲; A. J. C. Ladd, J. Fluid Mech. 271, 285
共1994兲; Special issue on lattice gas, edited by Y. H. Qian, J.
Lebowitz, and S. Orszag, J. Stat. Phys. 81, 1-2, 共1995兲; Y. H.
Qian, S. Succi, and S. Orszag, Annu. Rev. Comput. Phys. 3,
195 共1995兲; S. Chen and G. D. Doolen, Annu. Rev. Fluid
Mech. 30, 329 共1998兲.
关3兴 G. Mc Namara, G. Garcia, and B. Alder, J. Stat. Phys. 81, 359
共1995兲.
关4兴 I. V. Karlin, A. N. Gorban, S. Succi, and V. Boffi, Phys. Rev.
Lett. 81, 6 共1998兲.
关5兴 I. V. Karlin and S. Succi, Phys. Rev. E 58, 4053 共1998兲.
关6兴 I. V. Karlin, A. Ferrante, and H. C. Öttinger, Europhys. Lett.
47, 182 共1999兲.
关7兴 A. Wagner, Europhys. Lett. 44, 144 共1998兲.
关8兴 B. M. Boghosian, J. Yepez, P. V. Coveney, and A. Wagner,
8003
sion integrals like Eqs. 共3a兲 or 共3b兲 requires working with
logarithmic functions which makes them more vulnerable to
the round-off errors. Finally, since the whole construction is
largely based on convexity of the entropy functions, and can-
not tolerate any nonpositivity of the populations 共unlike the
standard realizations which may tolerate negative popula-
tions to some extent兲, care must be taken when implementing
various entropy-based estimations of the collision step like
Eq. 共6兲 in order not to violate these properties. All these
questions require a further detailed study, which is the sub-
ject of our current work.
ACKNOWLEDGMENT
We thank Professor H. C. Öttinger for the encouragement
and valuable discussions of results.
e-print cond-mat/0005260.
关9兴 Y. H. Qian, D. d’Humieres, and P. Lallemand, in Advances in
Kinetic Theory and Continuum Mechanics, edited by R. Gatig-
nol and Soubbaramayer 共Springer, Berlin, 1991兲, p. 127.
关10兴 See, e.g., A. N. Gorban, I. V. Karlin, V. B. Zmievskii, and S.
V. Dymova, Physica A 275, 361 共2000兲, and references
therein.
关11兴 M. Grmela and H. C. Öttinger, Phys. Rev. E 56, 6620 共1997兲;
H. C. Öttinger and M. Grmela, ibid. 56, 6633 共1997兲.
关12兴 S. Chapman and T. Cowling, The Mathematical Theory of
Non-Uniform Gases 共Cambridge University Press, Cambridge,
England, 1970兲.
关13兴 Y. H. Qian, D. d’Humiéres, and P. Lallemand, Europhys. Lett.
17, 479 共1992兲.
关14兴 S. Ansumali and I. V. Karlin 共unpublished兲.
关15兴 W. H. Press, S. A. Teukolsky, W. T. Vellerling, and B. P.
Flannery, Numerical Recipes in C 共Cambridge University
Press, Cambridge, England, 1998兲, p. 365.