SOLVING GENERALIZED
NAVIERSTOKES EQUATIONS
Stelian Ion
Abstract
In this paper we set up a numerical algorithm for computing the
ow of a class of pseudoplastic uids. The method uses the nite vol
ume technique for space discretization and a semiimplicit two steps
backward dierentiation formula for time integration. As primitive
variables the algorithm uses the velocity eld and the pressure eld.
In this scheme quadrilateral structured primaldual meshes are used.
The velocity and the pressure elds are discretized on the primal mesh
and the dual mesh respectively. A certain advantage of the method is
that the velocity and pressure can be computed without any articial
boundary conditions and initial data for the pressure. Based on the
numerical algorithm we have written a numerical code. We have also
performed a series of numerical simulations.
MSC: 35Q30, 65M08, 76A99, 76D05, 76M12.
keywords: generalized NavierStokes equations, pseudoplastic uids, nite
volume methods, admissible primaldual mesh, discrete derivatives operators,
discrete Hodge formula.
ab
(u) = 2([
u[)
u
ab
(2)
where
u is the strain rate tensor given by
u
ab
=
1
2
(
a
u
b
+
b
u
a
) ,
a
standing for the partial derivative with respect to the space coordinate
x
a
, and for any square matrix e, [e[ being dened as
[e[ =
_
_
i,j
e
2
ij
_
_
1/2
.
Concerning the viscosity function (s), we assume that it is a continuous
dierentiable, decreasing function, with bounded range
_
0 <
(s)
0
< , s > 0,
((s
1
) (s
2
)) (s
1
s
2
) < 0, s
1
, s
2
> 0,
(3)
and it satises the constraint
(s) +s (s) > c > 0. (4)
The model of the Newtonian uid corresponds to = constant.
We consider the case when the ow takes place inside a xed and bounded
domain R
2
and we assume that the uid adheres to its boundary ,
hence we impose a Dirichlet type boundary condition for the velocity eld,
Solving Generalized NavierStokes Equations 147
u = u
D
(x), x , t > 0. (5)
To the equations (1) we append the initial condition for the velocity
u(x, 0) = u
0
(x), x . (6)
The initial boundary value problem (IBV), which we intend to solve nu
merically, consists in nding the velocity eld u(x, t) and the pressure eld
p(x, t) that satisfy the partial dierential equations (1), boundary condition
(5) and the initial condition (6).
A constitutive function as (2) is used , for example, to describe the be
havior of polymeric uids, [5], [8], [11], and the ow of the blood through the
vessels, [19], [9], [6], [18].
In writing down a numerical algorithm for the nonstationary incom
pressible generalized NavierStokes equations three main diculties occur,
namely: (i) the velocity eld and the pressure eld are coupled by the in
compressibility constraint [12], (ii) the presence of the nonlinear convection
term and (iii) the nonlinear dependence of the viscosity on the share rate.
The rst two problems are common to the NavierStokes equations and
in the last fty years several methods were developed to overcome them: the
projection method, [12], [13], [7],[14], [3], and gauge method, [20] to mention
the most signicant methods for our case.
When one deals with a nonNewtonian uid, the nonlinearity of the vis
cosity rises a new problem in obtaining a discrete form for the generalized
NavierStokes equations. The new issue is the development of an appropriate
discrete form of the action of the stress tensor on the boundary of the volume
control. A similar diculty is raised by the discretization of the plaplacean,
see [2] for that.
The outline of the paper is as follows. In Section 2 we dene the weak
solution of IBV (1), (5) and (6) and we present an existence theorem of the
weak solution for a class of pseudoplastic uids that satisfy (4). In Section
3 we establish the semidiscrete, space discrete coordinates and continuum
time variable form of the equation (1) and we present some general concepts
concerning the space discretization and related notions like admissible mesh,
primal and dual mesh, the discretization of the derivative operators etc. In
Section 4 we present an algorithm for solving a 2D model. In the last section
we present the results of some numerical simulations of the lid driven cavity
ow.
148 Stelian Ion, Anca Veronica Ion
2 The Existence of the Weak Solution
To dene the weak solution we need the following functional frame, [16], [17].
By L
p
() and W
m,p
(), m = 0, 1, , we denote the usual Lebesgue and
Sobolev spaces, respectively. The scalar product in L
2
is indicated by (, ).
For u, v vector functions dened on we put
(u, v) =
_
u
a
v
a
dx,
(u, v) =
a,b=1
_
a
u
b
a
v
b
dx.
We denote by [[ [[ the norm in L
2
associate to (, ). The norm in W
m,p
is denoted by [[ [[
m,p
. Consider the space
1 = C
0
(), div = 0 .
We dene H() the completion of 1 in the space L
2
(). We denote by
H
1
() the completion of 1 in the space W
1,2
.
For T (0, ] we set Q
T
= [0, T) and dene
1
T
= C
0
(Q
T
); div(x, t) = 0 in Q
T
.
The weak solution of IBV is dened as follow.
Denition 1. Let f L
2
(). Let u
0
(x) L
2
() and u
D
be such that
_
_
_
div u
0
= 0,
u
D
n = 0, x ,
u
0
= u
D
, x ,
(7)
and there exists v W
1,2
() L
4
() a vector function that satises
_
div v = 0,
v = u
D
, x .
(8)
Then u is a weak solution of IBV (1,5,6) if
u v L
2
_
(0, T); H
1
()
_
L
_
0
_
u,
t
_
dt
_
0
(u u, ) dt +
_
0
_
(u),
_
dt =
=
_
0
(f , ) dt + (u
0
, )
(10)
for any test function 1
T
.
Concerning the existence of the weak solution of the IBV we proved the
following result, [15]:
Theorem 1. If the constitutive function () satises the relations (3) and
(4) then there exists a weak solution of the IBV (1), (5) and (6).
3 Semidiscrete Finite Volume Method
The nite volume method (FVM) is a method for approximating the solution
of a partial dierential equation (PDE). It basically consists in partitioning
the domain , on which the PDE is formulated, into small polygonal domains
i
(control volumes) on which the unknown is approximated by constant
values, [10].
We consider a class of nitevolume schemes that includes two types of
meshes: the primal mesh, T =
I
, r
I
and the dual mesh,
T =
J
, r
J
.
The space discrete form of the GNS equations are obtained from the integral
form of the balance of momentum equation and mass balance equation on
the primal mesh and the dual mesh respectively.
For any
i
of the primal mesh T the integral form of the balance of
momentum equation reads as,
t
_
i
u(x, t)dx +
_
i
uu nds +
_
i
pdx =
_
i
nds, (11)
and for any
u nds = 0. (12)
150 Stelian Ion, Anca Veronica Ion
The velocity eld u(x, t) and the pressure eld p(x, t) are approximated
by the piecewise constant functions on the primal mesh and the dual mesh
respectively,
u(x, t) u
i
(t), x
i
, p(x, t) p
(t), x
.
By using certain approximation schemes of the integrals as functions of the
discrete variables u
i
(t)
iI
, p
(t)
J
one can dene:
T
i
(u)
_
i
uu nds, o
i
(u)
_
i
nds,
Grad
i
(p)
_
i
pdx, , Div
(u)
_
e
u nds.
(13)
The semidiscrete form of GNS equations, continuous with respect to time
variable and discrete with respect to space variable, can be written as:
m
i
du
i
dt
+T
i
(u) +Grad
i
(p) o
i
(u) = 0, i 1
Div
(u) = 0,
(14)
where m
i
stands for the volume of the
i
.
Now the problem is to nd the functions u
i
(t)
iI
, p
(t)
J
that sat
isfy the dierential algebraic system of equations (DAE) (14) and the initial
condition
u
i
(t)[
t=t
0
= u
0
i
, i 1. (15)
In solving the Cauchy problem (14) and (15), an essential step is to
dene a primaldual mesh
_
T ,
T
_
that allows one to calculate the velocity
eld independent of the pressure eld.
In the next subsections we dene a pair of quadrilateral admissible primal
dual (QAPD) meshes
_
T ,
T
_
, and we dene the discrete gradient of the
scalar functions and the discrete divergence of the vector functions such that
the discrete space of the vector elds admits an orthogonal decomposition
into two subspaces: one of discrete divergences free vectors elds, and other
consisting of vectors that are the discrete gradient of some scalar elds.
3.1 Quadrilateral primaldual meshes
Let be a polygonal domain in R
2
. Let T =
I
, r
I
be a quadrilateral
mesh dened as follows:
Solving Generalized NavierStokes Equations 151
(1)
i
is a quadrilateral,
iI
i
= ,
(2) i ,= j I and
i
j
,= , either H
1
(
i
j
) = 0, or
ij
:=
i
j
is a common (n 1) face of
i
and
j
,
(3) r
i
i
, if
i
= [ABCD], then r
i
= [M
AB
M
DC
] [M
AD
M
BC
],
(4) for any vertex P there exists only four quadrilateral
with the common vertex P,
where H
1
is the onedimensional Hausdor measure, and M
AB
denotes the
midpoint of the line segment [AB].
Let
T =
J
, r
J
be another mesh dened as follows:
(1) , r
is a vertex of T ,
(2) r
, ,
(3) r
, the poligon
,
where by "center" of the quadrilateral we understand the intersection of the
two segments determined by the midpoints of two opposed sides.
Figure 1: Quadrilateral mesh.
We call (T ,
T )  a pair of QAPD meshes.
We denote by H
e
T
() the space of piecewise constant scalar functions
that are constant on each volume
J
, by H
T
() the space of piecewise
constant vectorial functions that are constant on each volume
i
I
and
by H H
e
T
() the space of piecewise constant tensorial functions of order
two that are constant on each volume
a
J
.
For any quantity that is piecewise constant on
J
we denote by
the
constant value of on
a
, analogously
i
stands for the constant value of a
piecewise constant quantity on
I
.
152 Stelian Ion, Anca Veronica Ion
We dene the discrete derivative operators:
Div
(T ,
e
T )
: H
T
() H
e
T
(), by
Div
(u) :=
_
e
u nds =
i
u
a
i
_
e
i
n
a
ds, (16)
(T ,
e
T )
: H
T
() HH
e
T
() by
b
u
a
[
=:
1
m(
)
_
e
u
b
n
a
ds =
1
m(
a
)
i
u
b
i
_
e
i
n
a
ds, (17)
Grad
(T ,
e
T )
: H
e
T
() H
T
() by
Grad
i
()
_
i
nds =
i
e
nds, (18)
rot
(T ,
e
T )
: H
e
T
() H
T
() by
rot
i
() :=
1
m(
i
)
_
i
dr =
1
m(
i
)
i
e
dr. (19)
On the space H
T
() we dene the scalar product , )) by
u, v)) =
iI
u
i
v
i
, (20)
and on the space H
e
T
() we dene the scalar product , ) by
, ) =
. (21)
In the next lemma we prove certain properties of the discrete derivative
operators.
Lemma 1. Let
_
T ,
T
_
be a pair of QAPD meshes and the discrete diver
gence, the discrete gradient and the discrete rotation be dened, respectively,
by (16), (18), and (19). Then:
Solving Generalized NavierStokes Equations 153
(a1) Discrete Stokes formula. For any u H
T
() and any H
e
T
(), a
discrete integration by parts formula holds, that is
_
Div
(T ,
e
T )
(u),
_
+
__
u, Grad
(T ,
e
T )
()
__
= 0. (22)
(a2) For any H
e
T
(), [
= 0 one has
Div
(T ,
e
T )
rot
(T ,
e
T )
= 0. (23)
Proof. To prove (a1) we use the fact that for any domain
_
nds = 0
and the denitions of the two operators.
To prove (a2), we note rstly that
Div
(rot
(T ,
e
T )
) =
i
rot
i
()
_
e
i
nds =
=
i
1
m(
i
)
_
e
i
dr
_
e
i
nds.
Then, let
i
a
, a = 1, 4 be the primal volumes with the common vertex P
a
and
i
a+1
have a common side. For each i
a
let
P
a
b
, b = 1, 4 be the vertexes of the quadrilateral
i
a
anticlockwise numbered
and P
a
1
= P
. We have
1
m(
i
a
)
a
b
_
e
a
b
a
dr
_
e
i
a
nds =
a
2
a
4
.
Finally, by summing up for a = 1, 4, we have
Div
(rot
(T ,
e
T )
) =
a
(
a
2
a
4
) = 0,
for any such that P
. If for some , P
.
154 Stelian Ion, Anca Veronica Ion
Now we prove an orthogonal decomposition of the space H
T
() that
resembles the one for the nondiscrete case. Let
J
,
H
e
T
() be
a basis of the space H
e
T
() given by
(x) =
_
1, if x w
,
0, if , x / w
.
(24)
Dene the discrete vector eld 
H
T
() by

= rot(
). (25)
Let W
T
() be the linear closure of the set 
)) = 0, Int(). (27)
We construct a function H
e
T
() such that
Grad
i
() = u
i
, i 1.
For a given
i
we denote by P
i
b
, b = 1, 4 its vertexes counterclockwise
numbered. The gradient of a scalar eld can be written as
Grad
i
() =
1,3
(
i
3
i
1
) +
2,4
(
i
4
i
2
),
where
1,3
is a vector orthogonal to
P
i
2
P
i
4
oriented from P
i
1
to P
i
3
and
[
1,3
[ =
i
2
P
i
4
/2 and
2,4
is a vector orthogonal to
P
i
1
P
i
3
oriented
from P
i
2
to P
i
4
and [
2,4
[ =
i
1
P
i
3
i
2
P
i
4
m(
i
)
=
i
4
i
2
,
u
i
i
1
P
i
3
m(
i
)
=
i
3
i
1
.
(28)
Solving Generalized NavierStokes Equations 155
The point is that if u satises the ortogonalty conditions (27) then one can
solve the equations (28) inductively, i.e. starting from two adjacent values
and following some path of continuation. For a general discrete eld u the
dierent paths lead to dierent values!
Corollary 1 (Discrete Hodge formula). Let (T ,
T ) be a pair of QAPD
meshes. Then for any w H
T
() there exists an element u H
T
()
and a scalar function H
e
T
() such that
w = u +Grad() with Div
(T ,
e
T )
(u) = 0. (29)
Proof. We search for a divergence free vector u of the form
u =
aJ
a

a
.
By inserting this form into (29), one obtains a linear algebraic system of
equation for the determination of the unknowns
a
aJ
,
__
w, 
b
__
=
aJ
a
__

a
, 
b
__
.
The matrix of the system is the Gram matrix of a linear independent family,
hence there exists an unique solution u.
Since wu, 
a
)) = 0 for any function in the basis, it follows that wu
is orthogonal to G
kK
of the boundary mesh induced by the primal mesh i.e
k
=
i
k
, =
kK
k
.
On each
k
the boundary data u
D
are approximated by constant values
u
Dk
.
Several formulas to calculate the numerical convective ux (NCF) are
available, most of them derived from the theory of hyperbolic equations.
156 Stelian Ion, Anca Veronica Ion
In the case of a hyperbolic equation, the numerical convective ux must
satisfy, besides the accuracy of the approximation requirements, a number of
conditions in order that the implied solution be physically relevant. In the
case of NavierStokes equation at high Reynolds number, the way in which
NCF is evaluated is also very important. We propose to dene the NCF as
follow. Consider the tensorial product uu constant on the dual mesh and,
for any control volume
i
that does not lie on the boundary T, set for the
NCF:
T
a
i
=
(u
a
u
b
)
_
e
i
n
b
ds. (30)
The tensorial product u u is approximated by
(u
a
u
b
)
=
1
m(
)
_
e
u
a
dx
1
m(
)
_
e
u
b
dx. (31)
The numerical stress ux is set up by considering that the gradient of the
velocity is piecewise constant on the dual mesh. This fact implies that the
stress tensor is also piecewise constant on the dual mesh. So we can write
for the numerical stress ux
o
i
(u) =
(u)
_
i
e
nds. (32)
The values of
(u) = 2([D
(u)[)D
(u), (33)
where the discrete strain rate tensor D
is given by
D
ab
(u)[
=
1
2
(
a
u
b
+
b
u
a
)
. (34)
Dirichlet Boundary conditions. The boundary conditions for the veloc
ity are taken into account through the numerical convective ux and numer
ical stress ux. If for some the dual volume
a
u
b
=
1
m(
)
_
e
u
b
n
a
ds =
Solving Generalized NavierStokes Equations 157
=
1
m(
)
_
_
_
_
ext
e
u
b
D
n
a
ds +
i
u
b
i
_
int
e
i
n
a
ds
_
_
_
. (35)
For a primal volume
i
adjacent to the boundary the NCF (30) is
given by
T
a
i
= u
a
D
u
b
D
_
i
n
b
ds +
(u
a
u
b
)
i
un
b
ds. (36)
4 FullyDiscrete Finite Volume Method
We set up a time integration scheme of the Cauchy problem (14) and (15)
that determines the velocity eld independently on the pressure eld. The
pressure eld results from the discrete balance momentum equation (141).
The scheme resembles the Galerkin method and it makes use of the orthogo
nal decomposition (26) of the space H
T
() and of the set of divergence free
vectorial elds 
J
0.
We write the unknown velocity eld u(t) as linear combination of 
J
0
u =
(t)
(37)
where the coecients
dt
__
m
, 
__
+
__
T(), 
__
__
o(), 
__
= 0,
0
, (38)
with the initial conditions
(0)
__

, 
__
=
__
u
0
, 
__
,
0
. (39)
If the functions
(t
n
), u
n
=
(t
n+1
)
dt
=
_
3
2
n+1
2
n
+
1
2
n1
_
/t
that leads to the following nonlinear equations for
n+1
I
3
2
n+1
__
m
, 
__
+ t
T(
n+1
), 
__
t
o(
n+1
), 
__
=
=
2u
n
0.5u
n1
, m
__
.
(41)
To overcome the diculties implied by the nonlinearity, we consider a
linear algorithm:
3
2
n+1
__
m
, 
__
t
__
o(u;
n+1
), 
__
=
0.5
u
n
u
n1
, m
__
t
__
3
2
T(u
n
)
1
2
T(u
n1
), 
__
+t
o(u
n
), 
__
,
(42)
where
n+1
:=
n+1
n
.
For the rst step one can use a Euler step
n+1
__
m
, 
__
t
__
o(u
n
;
n+1
), 
__
=
t
T(u
n
), 
__
+t
o(u
n
), 
__
.
(43)
Solving Generalized NavierStokes Equations 159
In both (42) and (43) schemes we use the notation
o(u
n
;
n+1
) = 2([D(u
n
)[)
n+1
D(
).
5 Numerical Simulations
We present the results of some numerical experiments designed to test the
numerical method presented in the previous sections.
We consider the pseudoplastic uid modeled by the CarreauYasuda law,
( ) =
+ (
0
) (1 + ( )
a
)
(n1)/a
.
In the current study the problem was solved for a series of rectangular regular
or nonregular meshes. The code incorporates the time integration scheme
(42) and (43); the numerical convective ux T dened by the formulae (30),
(31), (36) and the numerical stress ux o dened by formulae (32), (33),
(34), (17), (35). In all the numerical simulations we consider that at the
initial time the uid is at rest.
Lid Driven Cavity Flow
Figure 2: Lid Driven 2D Cavity Flow.
The uid is moving in a rectangular box, the side and bottom walls are
static while the top wall is moving across the cavity with a constant velocity
160 Stelian Ion, Anca Veronica Ion
u = U, v = 0 as in Fig. 2. We assume the nonslip boundary conditions on
the walls.
In the rst set of computations we test the capabilty of the method to
catch the behavior of the pseudoplastic uid. To be more precise, we chose a
pseudoplastic uid and two Navier Stokes uids having the viscosities equal
to
0
, and
, respectively.
Figure 3 shows the contours plots of the steady solutions for the three
type of uids. Each ow consists of a core of uid undergoing solid body
rotation and small regions in the bottom corners of counterrotating vortex.
The intensity of the counterrotating vortex is decreasing with respect to
viscosity. The velocity prole along the vertical centerline is shown in Figure
4. We observe that, in the lower part of the cavity, the uid is moving in
contrary sense to the sense of the motion of the top wall. The maximum of
the negative velocity depends decreasingly on the viscosity of the uid.
NS Eqs. GNS Eqs. NS Eqs.
Re
0
= 10
2
/1.57 Re
0
Re Re
Re
= 10
3
/1.57
Figure 3: U = 0.01ms
1
, a = 0.144. Contour plot of stream functions,
steady solutions. Regular grid, 51 51 grid points.
The second set of computations analyzes the response of the numerical
method to the variation of the parameters of the uid. The results are shown
in Figure 5.
Final Remarks
A certain advantage of our method is that there is no need to introduce arti
cial boundary conditions for the pressure eld or supplementary boundary
conditions for additional velocity eld as in the projection methods or gauge
methods. The preliminary numerical results prove a good agreement with
Solving Generalized NavierStokes Equations 161
Figure 4: U = 0.01ms
1
, a = 0.144. Distribution of uvelocity along at
vertical centre line of the cavity. Regular grid, 51 51 grid points.
a = 0.644 a = 0.144
steady solution t = 200
Figure 5: GNS Eqs. U = 0.1ms
1
Re
= 10
4
/1.57, Re
0
= 10
3
/1.57 .
Stretched grid, 51 51 grid points.
the results obtained by other methods. At the present moment we do not
know if it is possible to extended the method to the 3D case and this is a
drawback of the method. The study of this extension might be a task for our
future work.
162 Stelian Ion, Anca Veronica Ion
References
[1] P.D. Anderson, O.S. Galaktionov, G.W.M. Peters, F.N. van de Vose,
H.E.H. Meijer. Mixing of nonNewtonian uids in timeperiodic cavity
ow. J. NonNewtonian Fluid Mech. 93:265286, 2000.
[2] B. Andreianov, F. Boyer and F. Hubert. Finite volume scheme for the
plaplacian on cartezian meshes. M
2
AN. 38(6):931960, 2004.
[3] J. Bell, P. Colella, and H. Glaz. A secondorder projection method for in
compressible NavierStokes equations. J. Comp. Phys. 85:257283, 1989.
[4] K.L. Brenan, S.L. Campbell, L.R. Petzold. Numerical Solution of Initial
Values Problems in DierentialAlgebraic Equations. Classics in Applied
Mathematics, SIAM, 1996.
[5] G.F. Carey, W. Barth, J. A. Woods, B. S. Kork, M. L. Anderson, S.
Chow, and W. Bangerth. Modelling error and constitutive relations in
simulattion of ow and transport. Int. J. Num. Meth. Fluids. 46:1211
1236, 2004.
[6] P. Carreau. Rheological equations from molecular network theories.
Transactions of Society of Rheology. 16:99127, 1972.
[7] A.J. Chorin. Numerical solution of the NavierStokes equations. Math.
Comput. 22:745762, 1968.
[8] S.D. Cramer, J.M. Marchello. Numerical evolution of models describing
nonNewtonian behavior. AIChE Journal. 4(6):980983, 1968.
[9] M.M. Cross. Rheology of non newtonian uids: A new ow equation for
pseudoplastic systems. Journal of Colloid and Interface Science. 20:417
437, 1965.
[10] R. Eymard, Th. Gallouet and R. Herbin. Finit Volume Method. In P.G.
Ciarlet and J.L. Lions, (eds.) Handbook of Numerical Analysis. North
Holland, 2000.
[11] H. Eyring. Viscosity, plasticity, and diusion as examples of absolute
reaction rates. Journal of Chemical Physics. 4:283287, 1936.
Solving Generalized NavierStokes Equations 163
[12] J.L. Guermond, P. Minev, and Jie Shen. An overview of projection
methods for incompressible ows. Comput. Methods Appl. Mech. En
grg. 195:60116045, 2006.
[13] F.H. Harlow and J.E. Welch. Numerical calculation of timedependent
viscous incompressible ow of uid with free surface. Phys. Fluids.
8:21822189, 1965.
[14] J. Kim and P. Moin. Application of a fractionalstep method to incom
pressible NavierStokes equations. J. Comp. Phys. 59:308323, 1985.
[15] S. Ion. Solving Generalized 2D NavierStokes Equations. Tech
nical report, Institute of Mathematical Statistics and Ap
plied Mathematics, Department of Applied Mathematics, 2009,
http://www.ima.ro/publications/reports/rth_030605.pdf.
[16] O. A. Ladyzhenskaya. The Mathematical Theory of Viscous Incompress
ible Flow. Gordon and Breach, 1969.
[17] R. Temam. NavierStokes Equations. Elsevier Science, 1984.
[18] K. Yasuda, R.C. Armstrong, and R.E. Cohen. Shear ow properties of
concentrated solutions of linear and star branched polystyrenes. Rheo
logica Acta. 20:163178, 1981.
[19] K.K. Yeleswarapu. Evaluation of continuum models for characterizing
the constitutive behavior of blood. Ph. D. thesis, University of Pittsburgh,
1996.
[20] Weinan E and J.G. Liu. Gauge method for viscous incompressible ows.
Comm. Math. Sci. 1:317332, 2003.