Numerical Solution

of the Navier-Stokes
Equations*
By Alexandre Joel Chorin
Abstract. A finite-difference method for solving the time-dependent NavierStokes equations for an incompressible fluid is introduced. This method uses the
primitive variables, i.e. the velocities and the pressure, and is equally applicable to
problems in two and three space dimensions. Test problems are solved, and an application to a three-dimensional
convection problem is presented.
Introduction.

The equations of motion of an incompressible fluid are

dtUi 4- UjdjUi = dip + vV2Ui + Ei}

( V2 = Yl d2 ) ,

PO

'

djUj = 0 ,

where Ui are the velocity components, p is the pressure, p0 is the density, Ei are
the components of the external forces per unit mass, v is the coefficient of kinematic
viscosity, t is the time, and the indices i, j refer to the space coordinates Xi, x, i, j =
1, 2, 3. d, denotes differentiation with respect to Xi, and dt differentiation with
respect to the time t. The summation convention is used in writing the equations.

We write
,

Uj

Ui -

u '

e/ = ($)e,

Xj

, _ (

Xi " d '

d \

p - \povur

f-(),

where 77 is a reference velocity, and d a reference length. We then drop the primes.
The equations become

(1)

dtu, 4- RujdjUi = dip 4- V2u, + E, ,

(2)

dfij = 0 ,

where R = Ud/v is the Reynolds number. It is our purpose to present a finitedifference method for solving these equations in a bounded region 3), in either twoor three-dimensional space. The distinguishing feature of this method lies in the
use of Eqs. (1) and (2), rather than higher-order derived equations. This makes it
possible to solve the equations and to satisfy the imposed boundary conditions
while achieving adequate computational
efficiency, even in problems involving
three space variables and time. The author is not aware of any other method for
which such claims can be made.
Received February 5, 1968.
* The work presented in this report is supported
matics Center, Courant Institute of Mathematical

by the AEC Computing and Applied MatheSciences, New York University, under Con-

tract AT(30-1 )-1480 with the U. S. Atomic Energy Commission.

745

746

ALEXANDRE

JOEL CHORIN

Principle of the Method. Equation (1) can be written in the form

(1)'

dtUi + dip = <5nt,

where CF.-wdepends on u, and Ei but not on p; Eq. (2) can be differentiated

to

yield
(2)'

diidtUi) = 0.

The proposed method can be summarized as follows: the time t is discretized; at

every time step i is evaluated; it is then decomposed into the sum of a vector
with zero divergence and a vector with zero curl. The component with zero divergence is dtUi, which can be used to obtain u,- at the next time level; the component
with zero curl is 3,-p. This decomposition exists and is uniquely determined whenever the initial value problem for the Navier-Stokes equations is well posed; it has
also been extensively used in existence and uniqueness proofs for the solution of
these equations (see e.g. [1]).
Let Ui, p denote not only the solution of (1) and (2) but also its discrete approximation, and let Du be a difference approximation to djU. It is assumed that
at time t = nAt a velocity field m," is given, satisfying Du" = 0. The task at hand
is to evaluate <n+1 from Eq. (1), so that Dun+l 0.
Let Tu i = buin+1 Bu i approximate dw, where b is a constant and Bui a
suitable linear combination of Uin~>, j 0. An auxiliary field w;aux is first evaluated

through
(3)

&waux Bu = F{u

where F{u approximates

F.-w. waux differs from utn+l because the pressure term
and Eq. (2) have not been taken into account. ,aux may be evaluated by an implicit scheme, i.e. F At, may depend on u<n, w,aux and intermediate
fields, say u*,
Ui**. bu?** Bui now approximates m to within an error which may depend

on Ai.
Let Gip approximate
decomposition

dip. To obtain un+l, p"+1 it is necessary to perform the

Fu = bus - But = Tut + Gipn+l,

DQTu) = 0 .

It is, however, assumed that Dun~> = 0, j ^ 0. It is necessary therefore only

to perform the decomposition
(4)

waux = Uin+1 +

b~lGipn+l,

where Duin+l = 0, and m,-"+1satisfies the prescribed boundary conditions. Since

p" is usually available and is a good first guess for the values of pn+l, the decomposition (4) is probably best done by iteration. For that purpose we introduce the
following iteration scheme :
(5a)

(5b)

u*+i.i*h

Miaux _ i)-'Gimp

p+l,m+l _ pn+X,m _ \QUn+l,m+l

where X is a parameter,

m ^

1 ,

WI

M,n+Im+1 and p"+1.+1are successive

Uin+l, pn+1, and Gimp is a function

of pn+x-m+1 and p"+lm which

as |pn+1.">+1 p+i.m| tends to zero. We set

1 ,

approximations
converges

to

to Gipn+l

NUMERICAL SOLUTION OF THE NAVIER-STOKES EQUATIONS

747

The iterations (5a) are to be performed in the interior of 2D, and the iterations
(5b) in 20 and on its boundary.
It is evident that (5a) tends to (4) if the iterations converge. We are using
dmp instead of dp in (5a) so as to be able to improve the rate of convergence of
the iterations. This will be discussed in detail in a later section. The form of Eq.
(5b) was suggested by experience with the artificial compressibility method [2]
where, for the purpose of finding steady solutions of Eqs. (1) and (2), p was related to Ui by the equation
dtp = constant (j-My).

When for some I and a small predetermined

,

max \p

n+X,l+X

constant
n+l,|

we set

The iterations (5) ensure that Eq. (1), including the pressure term, is satisfied inside 3D,and Eq. (2) is satisfied in 3Dand on its boundary.
The question of stability and convergence for methods of this type has not
been fully investigated. I conjecture that the over-all scheme which yields w,-n+1in
terms of w," is stable if the scheme

Tu i = F ai
is stable. The numerical evidence lends support to this conjecture.
We shall now introduce specific schemes for evaluating ttaux and specific representations for Du, dp, Gjmp. Many other schemes and representations
can be
found. The ones we shall be using are efficient, but suitable mainly for problems
in which the boundary data are smooth and the domain 3Dhas a relatively simple

shape.
Evaluation

of waux. We shall first present

schemes

for evaluating

waux, defined

by (3).
Equation

(3) represents one step in time for the solution of the Burgers equation
dtUi = 3,-u ,

which can be approximated in numerous ways. We have looked for schemes which
are convenient to use, implicit, and accurate to 0(A7) + 0(Aa;2), where Ax is one
of the space increments Ax, i = 1, 2, 3. Implicit schemes were sought because
explicit ones typically require, in three space dimensions, that
At < iAx2

which is an unduly restrictive condition. On the other hand, implicit schemes of

accuracy higher than 0(Ai) would require the solution of nonlinear equations at
every step, and make it necessary to evaluate w,aux and ui"+1 simultaneously

748

ALEXANDRE JOEL CHORIN

rather than in succession. Since we assume throughout that Ai = 0(Aa;2), the gain
in accuracy would not justify the effort.
Two schemes have been retained after some experimentation. For both of them
Tu i = iuin+l - Uin)/At,

Qj-1 = At,

Bu = "/Ai)

They are both variants of the alternating direction implicit method.

(A) In two-dimensional problems we use a Peaceman-Rachford
proposed by Wilkes in [3] in a different context. This takes the form
UHq,t)

Ui(q.r)

I ^

(6a)

+-~2

aux

Ui(z,r)

U(5,r)

U2(1,r){Ui(q,r+X)

Ui{q,r-X))

+ M*(3-i,r) -

2 iUi(,q,r+X)

/ aux

2Aa;2

r>

4Ar

^-t

(6b)

Ui(q-X.r))

iu%+i,r)

2Aii

Wl(g,r)(M(3+i,r)

Wi(,,r_l)

Ml^'r>^Mi9+1'rt

scheme, as

2u%,r))

2-U(3lI.))

aux

~~ Ui(q-l,r))

AAX M2('r'l-M''('r+1> M.r-1);

+ T"

h iu*(q+1,r) + M*(g-l,r) M*(,,r))

2Aa;i

&t

2A^2

aux

2 \M"(.r+l)

aux

I UHq,r-X)

r,

aux

Unq,r))

+ -;,

where w* are auxiary fields, and w,-(4,r) = UiiqAxx, rAx2). As usual, the onedimensional systems of algebraic equations can be solved by Gaussian elimination.
(B) In two-dimensional and three-dimensional problems we use a variant of
the alternating direction method analyzed by Samarskii in [4]. This takes the form
1t>i(qlT,s)

^(q.r.s)

O A'v

~T"

Mi(t,r,i)

ttto,r,j)

2 V^f+l.r.s)

^T~T

_1_
"T

^1(3 r,s) V^'tif+l",)

Ax2

~T ^i(q-~l,r,s)

W2(,r,)lW,'(g,r+l,s)

(,,**
2 V.*(8,r+l,s)

J_ 1,**
"T (?,r-l,s)

^i'(g l,r,a))

^Hq,r,8))

M (3,r_l, s) J

9,/**
i*i(,r,s);

"\

NUMERICAL SOLUTION OP THE NAVIER-SROKES

aux

Mi(3,r,s)

jjcjs

Ui(gtT,s)

&t

t3(,r,)

^t

aux

\MHq,T.t+l)

aux

2 lM(9,r,s+l)

749

Mj(5l,-,g_l)^

aux

"t" (s,r,s-l)

Ax3

+ AtEi(qra)

aux

EQUATIONS

aux

W(9,,.,))

Ui(q,r,s) = Mi(gAxi, rAx2, sAz3),

Ei(Q,r,S) = EiiqAxx,

rAx2, sAx3).

Ui*, U** are auxiliary fields. These equations can be written in the symbolic form

(7 - AiQiK* = uf,
(7 - AQ,)***=

(7)

(7 -

where 7 is the identity

Ui* ,

AiQ3)Wiaux = Ui** 4- AtEi,

operator,

and Q involves differentiations

with respect to

the variable xi only.

It can be verified that when R = 0 scheme (6) is accurate to 0(Ai2) 4- 0(Ax2).
When ^ 0 however, they are both accurate to the same order. Scheme (7) is
stable in three-dimensional problems; the author does not know of a simple extension of scheme (6) to the three-dimensional case. Scheme (7) has two useful
properties: It requires fewer arithmetic operations per time step than scheme (6),
and because of the simple structure of the right-hand sides, the intermediate fields
Ui*, ui** do not have to be stored separately.
If either scheme is to be used in a problem in which the velocities m"+1are
prescribed at the boundary, values of u*, w**, w,aux at the boundary have to be
provided in advance so that the several implicit operators can be inverted. Consider the case of the scheme (7). We have

Uin+1= (7 + AtQx + AtQ2 4- AiQ3)u," 4- AtEi - AG.-p4- 0(A2) ,

Ui* = (7 + AtQx)uin + 0(Ai2) ,

Ui** = (7 + AiQi + AtQ2)uin + 0(Ai2) ,

M.aux = (/+

AiQi 4- AQ2 4- AtQ3)Uin+ AtEt 4- 0(A<2) .

From these relations it can be deduced that if we set at the boundary

(8)

Ui* = Uin+X-

AtQ2Uin+i -

AtQsUin+1-

ui** = Ui+1 -

AtQ3Uin+1-

AtEi + AtGip ,

taux

= Uin+1 4- AtGip

the scheme will remain accurate

AtEi 4- AtGip ,

to 0(A). Here Gi does not have to be identical

with d; all we need is

Gip" = dp" + O(Ai) .

The reason for introducing the new operator (?< is that at the boundary the normal
component of (? has to be approximated
by one-sided differences, while this is
not necessary in the interior of the domain 3Dwhere Eq. (4) is assumed to hold.
More accurate expressions for the auxiliary fields at the boundaries can be

750

ALEXANDRE

JOEL CHORIN

used, provided one is willing to invest the additional programming effort required
to implement them on the computer. Appropriate expressions for u*, u?ux at the
boundary can be derived for use with the scheme (6).
It should be noted that for problems in which the viscosity is negligible, it is
possible to devise explicit schemes accurate to 0(A2) 4- 0(A.r2) and stable when
At = OiAx). Such schemes will be discussed elsewhere.

The Dufort-Frankel

Scheme and Successive Point Over-Relaxation.

In order

to explain our construction of D, G,m and our choice of X for use in (5a) and (5b),
we need a few facts concerning the Dufort-Frankel
scheme for the heat equation
and its relation to the relaxation method for solving the Laplace equation.
Consider the equation
(9)

-V2w = /,

(V2 = dx2 + dx2)

in some nice domain 3D, say a rectangle,

3D.We approximate this equation by

(10)

u is assumed

known on the boundary

of

-Lu=f,

where L is the usual five-point approximation to the Laplacian, and u and / are
now m-component vectors, m is the number of internal nodes of the resulting difference scheme. For the sake of simplicity we assume that the mesh spacings in the
Xx and x2 directions are equal, Axx = Ax2 Ax; this implies no essential restriction. The operatorL is represented by an m X m matrix A.
We write

A = A' - E - E'
where E, E' are respectively strictly upper and lower triangular matrices, and Ah
is diagonal. The convergent relaxation iteration scheme for solving (10) is defined

by
(11)

(A' - co7>"+' = !(1 - u)A' + wE'}u 4- uf

(see e.g. Varga [5]). cois the relaxation factor, 0 < w < 2, and the u" arc the suc^
cessive iterates. The evaluation of the optimal relaxation factor a!opt depends on
the fact that A satisfies "Young's condition (A)," i.e. that there exists a permutation matrix P such that

(12)
where A is diagonal,

P-lAP

= A- A ,

and A has the normal form

(
\G'

)
0/

the zero submatrices being square. Under this condition, cooptcan be readily determined.
The matrix A depends on the order in which the components of w"+1 are computed from un. Changing that order is equivalent to transforming
A into I'~lAP,
where P is a permutation matrix.
We now consider the solution of (14) to be the asymptotic steady solution of

NUMERICAL SOLUTION OF THE NAVIER-STOKES

(13)

n+l

751

dTU=V2U+f

and approximate
Uq,r

EQUATIONS

nX

Uq.r

the latter equation by the Dufort-Frankel

il^T

-2

Ax

, n

(."s+l.r

M-l,r

W,i>r+i -f- M,,r_l

scheme
0

+1

on_1\

Uqr

Zq,r ) +

o a

At}

uq,T = u{qAxx, rAx2, nAr)

which approximates

(13) when Ai = o(Ax). Grouping terms, we obtain

(14)
= 2-

Ar
Ax

(Mg+l.r +

Ma-I.r

Uq.r+l +

Uq.r-x)

2At/

Since m",, does not appear in (14), the calculation separates into two independent calculations on intertwined meshes, one of which can be omitted. When
this is done, we can write
7Jn+1 = ( 2n+i )

(7J"+1 has m components)

If we then write
(15)

r--^/^
1 4- 4Ar/ Ax

we see that the iteration (14) reduces to an iteration of the form (11) where the
new components of Un+l are calculated in an order such that A has the normal
form (12). The Dufort-Frankel
scheme appears therefore to be a particular ordering of the over-relaxation method whose existence is equivalent to Young's con-

dition (A).
The best value of Ar, Aropt, can be determined
from cooptand relation (15).
We find that Aropt = O(Ax), therefore for At = Aropt the Dufort-Frankel
scheme
approximates, not Eq. (13), but rather the equation
dTu = V2u - 2 \~)

d2u + f.

This is the equation which Garabedian in [6] used to estimate coopt.It can be used
here to estimate ATopt- These remarks obviously generalize to problems where
Azi 9e Ax2 or where there are more than two space variables.
The following remark will be of use: We could have approximated
Eq. (13)
by the usual explicit formula
(16)

uYr Uq,r =

-;

Ax

(W+I,r 4- K-l.r

+ K.r+X + w",r-l

4:Uq,r) +

Ar/

and used this formula as an iteration procedure for solving (10). The resulting
iteration converges only when At/ Ax2 < 1/4, and the convergence is very slow.
The rapidly converging iteration procedure (14) can be obtained from (16) by
splitting the term w",r on the right-hand
side into hiuYr + U"Y^-

752

ALEXANDRE JOEL CHORIN

Representation of D, d and 67, and the Iteration Procedure for Determining

w"+1, p"+1. For the sake of clarity we shall assume in this section that the domain
3Dis two-dimensional and rectangular, and that the velocities are prescribed at the
boundary. Extension of the procedure to three-dimensional
problems is immediate, and extension to problems with other types of boundary conditions often possible. Stress-free boundaries and periodicity conditions in particular offer no difficulty. Domains of more complicated shape can be treated with the help of appropriate interpolation procedures.
Our first task is to define D. Let (B denote the boundary of 3Dand Q the set of
mesh nodes with a neighbor in (B. In 3D (B we approximate the equation of continuity by centered differences, i.e. we set

(17)

Du =

1
iUl(q+l,r)

Ux(q-X.r))

2Azi ^i"1'"

~-"

i 2Az2

(2(8,r+1)

M2(a,r-1))

At the points of (B we use second-order one-sided differences, so that Du is

accurate to 0(Az2) everywhere. Consider the boundary line x2 = 0, represented

by j = 1 (Fig. 1). We have on that line

- -T~

(18)

Ax2

[2(5,2)

2(5,1)

(2(8,3)

2(5,1))]

2Azi (mi(+i,i) ks-i.d)

with similar expressions at the other boundaries. Equation (17) states that the
total flow of fluid into a rectangle of sides 2Axi, 2Az2 is zero. Equation (18) does
not have this elementary interpretation.

Figure

1. Mesh Near a Boundary.

We now define G ip at every point of 3D (B by

GxV = 2^T (Pq+l.r - Pl-l.r)

G2p = 1^7

iVq,r+X -

Vq.r = piqAxx,rAx2)

Pq.r-x)

i.e. dip is approximated by centered differences. It should be emphasized

these forms of dp and Du are not the only possible ones.

that

NUMERICAL SOLUTION OF THE NAVIER-STOKES

It is our purpose now to perform the decomposition

EQUATIONS

753

(4). "+1 is given on the

boundary (B, aux is given in 3D (B (the values of aux on 03, used in (6) or (7),
are of no further use). pn+1 is to be found in 3D (including the boundary) and un+1

in 3D 63, so that in 3D 03
aux = Uin+i 4- Atdp

and in 3D(including the boundary)

Dun+1 = 0 .

This is to be done using the iterations

(5), where the form of dmp has not yet been

specified.
At a point (g, r) in 3D 03 C, i.e. far from the boundary,

one can substitute

Eq. (5a) into Eq. (5b), and obtain

(19)

pn+l,m+l

This is an iterative

pn+l.m

-\)uau*

_|_ M\DGmp

procedure for solving the equation

(20)

an \

Lp = j^E

where Lp = DGp approximates the Laplacian of p. With our choice of D and d,

Lp is a five-point formula using a stencil whose nodes are separated by 2 Axx, 2 Ax2.
Equation (20) is of course a finite-difference analogue of the equation

(21)

V2p = didjUiUj 4- djEj,

which can be obtained from Eq. (1) by taking its divergence. At points of 03 or e
if it is not possible to substitute (5a) into (5b) because at the boundary n+1 is
prescribed, w"+1"I+1 = wn+1for all m, (5a) does not hold and therefore (19) is not
true. Near the boundary the iterations (5) provide boundary data for (20) and
ensure that the constraint of incompressibility is satisfied. We proceed as follows :
dmp and X are chosen so that (19) is a rapidly converging iteration for solving
(20); Gimp at the boundary are then chosen so that the iterations (5) converge
everywhere.
Let

iq, r) again

be a node in 3D 03 C. utn+l-m and

known. We shall evaluate

simultaneously

pn+l'm are assumed

pY1,m+i and the velocity

components

involved in the equation 7>u"+1= 0 at iq, r), i.e. hT,o, Ktl'ZtV) (FiS- 2)These velocity components depend on the value of p at (q, r) and on the values
of p at other points. Following the spirit of the remark at the end of the last section, the value of p at iq, r) is taken to be

UpZUm+1 + pZUm)
while at other points we use p"+Im.
This leads to the following formulae

(22a)

pZUm+1 = pZUm - X7>w"+1'm+1

(7) given by (17))

/nrvu\

(22b)

n+l,m+l

aux

^t

1(8+1.r) = 1(8+1,r) - 2^T

nA~l,m

iPq+l.r -

\ /

n-f-l.m-f-1

\iPq.r

n+l,m\\

4" Pq.r

)) ,

754

ALEXANDRE

7i+l,m+]

(22c)

1(9-1,

(22d)

2(5,7-+!)

(22e)

2(8,

r)

r)

aux
2(5,r+1)

n+1,771+1

7i+l,77i+l

r-1)

At
+
n-r-l,n\
n+l,m\
2A.)j](iGC1'"*1 Pq.r ) - Pq-i.r)

aux

1(8-1,

JOEL CHORIN

At

aux

2(8, r-

n+l,m

i/

iPq.r+2 -

2Ax2

n+l,m+l

hiPq.r

i\iptr'm+l

n+l,jn\\

4- pq,r

)) ,

n-\-l,m\

+ pZUm)

Pq.r-Y .

2Ax2

These equations define Gimp. Clearly, dmp >G,p.

Equations (22) can be solved for p'Y/'m+l, yielding

(23a)

71+1,771+1
p4,V"

(1 + ax + a2)

'[(1

ai a2)pn+1,m

n+l,m

ai(P8+2,r

\Duu*

n+l,i\

n+l,m

4- Pg_2,r) 4- a2(pg,r+l

n+l,m\i

4- Pg,r_2)J ,

where a, = XA/4Aa,\, i = 1, 2. This can be seen to be a Dufort-Frankel

relaxation scheme for the solution of (20), as was to be expected. The At of the preceding section is replaced here by \At/2. Corresponding
to Aropt (or coopt) wTe find
Xopt-If p were known on 03 and C, convergence of the iterations (23a) would follow
from the discussion in the preceding section, and X = Xoptwould lead to the highest
rate of convergence.

9U2(q,r+l)

Pq.r

>-uKq+i,r)

ui(q-i,r)

U2(q,r-i)
Figure

2. Iteration Scheme

In 03 and C formulae (22) are modified by the use of the known values of ,+1
at the boundary whenever necessary. This leads to the formulae, for iq, r) in e :

(23b)

pn+X.m+X= {l + ai+

la2)-'[(l

_ ai _ ia2)pYYm
+

n+l,m

XDa
|

n+l,m\

\iPq.i

<*l(P5+2.2 4" Pq-i.i)

n+l,mi

+ a2pqA

] ,

Pq.1

))J

and for iq, r) on 03:

(23c)

71+1,771+1

V"-1

/-,

\ If/,

(! + ai)

71+1,771

Kl - ai)P8.i

-. 7-,

BX

- >^7>m

2a2(P8.3

etc. In (23b) Du is given by (17), and in (23c) by (18). w,aux at the boundary is
interpreted
as Uin+1. Although no proof is offered, a heuristic argument and the

numerical evidence lead us to state that the whole iteration systemEqs. (23a),
(23b), (23c)converges for all X > 0 and converges fastest when X ~ Xopt. None
of the boundary instabilities which arise in two-dimensional vorticity-stream
function calculation has been observed.
It can be seen that because our representation of Du = 0 expresses the balance
of mass in a rectangle of sides 2Ax,, i = 1, 2, the pressure iterations split into

NUMERICAL SOLUTION OF THE NAVIER-STOKES

EQUATIONS

755

two calculations on intertwined meshes, coupled at the boundary. The most efficient orderings for performing the iterations are such that the resulting over-all
scheme is a Dufort-Frankel
scheme for each one of the intertwined meshes. This
involves no particular difficulty; a possible ordering for a rectangular grid is shown
in Fig. 3. The iterations are to be performed until for some I
I

71+1, +1

71+1, I

max \pYr + - pYr I

5.1-

for a predetermined
e.
The new velocities W;n+I,i = 1, 2, are to be evaluated using (22b), (22c), (22d),
(22e). This has to be done only after the pn+l<mhave converged. There is no need
to evaluate and store the intermediate fields n+1'm+1.A saving in computing time
can be made by evaluating 7)aux at the beginning of each iteration. We notice
two advantages of our iteration procedure: Dun+l can be made as small as one
wishes independently
of the error in Dun; and when p"+li+1 and pn+1'1 differ by
less than e, Dun+1 = 0(f/X); it can be seen that Xopt = OiAxr1), hence Dun+1 =
0(Ax). A gain in accuracy appears, which can be used to relax the convergence
criterion for the iterations. This gain in accuracy is due to the fact that the "+1
are evaluated using an appropriate combination of pn+1-1 and p+i.'+lt rather than
only the latest iterate p"+i,H-i<

The domain is swept in the order AB,CD,EF, GH,IJ,K

Figure

3. An Ordering for the Iteration Scheme

Solution of a Simple Test Problem. The proposed method was first applied to
a simple two-dimensional
test problem, used as a test problem by Pearson in [7]
for a vorticity-stream
function method. 3Dis the square 0 = x ^ tt, i = 1, 2;
Ex E2 = 0; the boundary data are
i = cos Xx sin x2e~~2',

The initial data are

2 = sin Xx cos x2e~2'.

756

ALEXANDRE

JOEL CHORIN

ux = cos xx sin x2,

u2 = sin xx cos x2.

The exact solution of the problem is

i = cos xx sin x2e~2',

u2 = sin xx cos x2e~2t,

p = R Kcos 2xi 4- cos 2x2)e~u ,

where R is the Reynolds number. This solution has the property

dip = RujdjUi ;
hence curl (w) satisfies a linear equation.
of our method

because

We first evaluate

Nevertheless,

this problem is a fair test

7)aux ^ 0.

Xopt-For the equation

Lu = f
in 3D,with a grid of mesh widths 2 Axx, 2 Ax2, and known on the boundary,

we

have
"opt-1

(i-ay/2'

where a = |(cos 2Axi 4- cos 2Ax2) is the largest eigenvalue

Jacobi matrix (see [5]).

of the associated

We put

q~

Equation

2 \Axx2

Ax,2)'

(15) can be written as

^1
wopt ~
.14COopt

4g '

therefore
(X

(1 - a 2\l/2
)

and

iAt/Axx2 + Ai/Ace^)

(1 - a2)1'2 '

We now assume Axx = A.r2 = Ax, obtaining

ont

2Ax2
At sin (2Ax) '

In Tables I, II, and III we display results of some sample calculations, n is

the number of time steps; e(), i = 1, 2, are the maxima over 3Dof the differences
between the exact and the computed solutions w,-. It is not clear how the error in
the pressure is to be represented;
pn is defined at a time intermediate
between
(n l)At and nAt; it is proportional to R in our nondimensionalization.
There

NUMERICAL SOLUTION OF THE NAVIER-STOKES

EQUATIONS

757

are errors in p due to the fact, discussed at the end of the preceding section, that
the iterations can be stopped before the pn-m have truly converged. e(p) in the
tables represents the maximum over the grid of the differences between the exact
pressure at time nAt and the computed pn, divided by R; it is given mainly for the
sake of completeness. The accuracy of the scheme is to be judged by the smallness
of e{ui). I is the number of iterations; it is to be noted that the first iteration always has to be performed in order that Eq. (1) be satisfied. "Scheme A" means
that Miaux was evaluated

using Eq. (6), and "Scheme B" means Eq. (7) were used.

Table

Scheme A; Ax = tt/39; Ai = 2 Ax2 = 0.01397; e = Ax2; It = 1

n

e(i)

1
2
3
4
5
6
7
9
10

2.8 X
2.7 X
1.5 X
1.8 X
1.3 X
1.3 X
1.6 X
1.4 X
1.3 X

10-4
10-4
lO"4
10-4
10"4
10-4
10-4
10-4
10"4

20

1.8 X lO"4

e(w2)

e(p)

10"4
10-4
10-4
lO"4
10-'
lO4
10-4
10-"
10 l

0.0243
0.0136
0.0069
0.0145
0.0089
0.0116
0.0144
0.0147
0.0156

1
7
4
4
5
4
4
4
4

2.3 X 10"4

0.0241

2.6 X
2.0 X
1.3 X
1.9 X
1.7 X
1.8 X
1.9 X
1.7 X
1.6 X

Table

II

Scheme A; Ax = tt/39; Ai = 2A.r.2= 0.01397; e = Ax3;R = 1

n

eiux)

1
2
3
4
5
6
7
8
9
10

8.5
1.0
1.0
1.0
1.0
9.7
9.4
9.0
8.7
8.3

X
X
X
X
X
X
X
X
X
X

10-5
10"4
10-4
10-4
10-4
10-5
lO"5
10-5
10"5
10-5

20

1.0 X10-4

e(2)

3.8
5.7
7.0
7.8
8.3
8.6
8.7
8.7
8.7
8.5

X
X
X
X
X
X
X
X
X
X

e(p)

10~5
10-5
10-6
10-5
10"5
10~5
10"5
10-5
lO"6
10"5

0.0059
0.0067
0.0068
0.0068
0.0069
0.0070
0.0071
0.0073
0.0077
0.0082

10
10
10
10
10
10
10
10
10
10

1.0 X 10"4

0.0216

Table III
Scheme B; Ax = tt/39; At = A.r2 = 0.00324; e = A.u2;R = 20
n

1
3
5
7
9

20

e(i)

1.1
1.9
2.5
3.3
4.0

X
X
X
X
X

e(w2)

10"3
10"3
lO"3
lO"3
10-3

5.8 X 10"3

e(p)

lO-3
10"3
10"3
10"3
lO"3

0.0217
0.0234
0.0242
0.0249
0.0253

15
9
9
9
8

3.9 X 10~3

0.0258

1.2
2.1
2.8
3.2
3.5

X
X
X
X
X

758

ALEXANDRE

JOEL CHORIN

Tables I and II describe computations which differ only in the value of e.

They show that = Ax2 is an adequate convergence criterion. Table III indicates
that fair results can be obtained even when RAt is fairly large; when R = 20,
Ax = 7r/39, Ai = 2 A.r2, we have

1.5 Aar1.

The errors are of the order of 1 %. Additional computational

results were presented

in [8].
Application to Thermal Convection. Suppose a plane layer of fluid, in the field
of gravity, of thickness d and infinite lateral extent, is heated from below. The
lower boundary x3 = 0 is maintained at a temperature
T0, the upper boundary
Xi = d at a temperature
Tx < T0. The warmer fluid at the bottom expands and
tends to move upward ; this motion is inhibited by the viscous stresses.
In the Boussinesq approximation
(see e.g. [9]) the equations describing the
possible motions are

dtUi 4- UjdjUi = dip 4- eVV - y(l - (7 7'0))<5,-,

Po

tT 4- UjdjT = kV2T ,

d,Uj = 0 ,

where T is the temperature,

k the coefficient of thermal conductivity, a the coefficient of thermal expansion, and 5, the components of the unit vector pointing
upwards.

We write
ut

\v^>

*~ d '

n>

To_ Ti ,

L \n

. (Tx - Tp)dgxi
_C3

P - p, \'y J P +

and drop the primes. The equations

2
V

now are

R*

dtUi 4- UjdjUi = dip 4- VV 4- (7 1)5, ,

a

d,T-\-UjdjT

= Y2T ,

djUi = 0,

where R* = agd3irT0 Tx)kv is the Rayleigh number, and <r = v/k the Prandtl
number. The rigid boundaries are now situated at x3 0 and xz = 1, where it is

assumed that w = 0, i = 1, 2, 3.
It is known that for R* < R*, the state of rest is stable and no steady convection can arise, where R* = 1707.762.
When R* = R*, steady infinitesimal convection can first appear, and the field
quantities are given by

us = CWixs)<t>,

Ui = -2 id3Wix3))di<t>,
a

T = CTix3)<p

i = 1,2 ,

NUMERICAL SOLUTION OF THE NAVIER-STOKES

where <p = <f>ixx,x2) determines

the horizontal

EQUATIONS

759

planform of the motion and satisfies

Oi2 + dt* 4- a2)tf= 0 ,

Wix3), Tix3) are fully determined functions of x3, a = 3.117, and C is a small but
undetermined amplitude.
In two-dimensional motion ux = 0 and the motion does not depend on xx. We
then have
<p = cos ax2.

The motion is periodic in x2 with period 2-ir/a.

The Nusselt number Nu is defined as the ratio of the total heat transfer to the
heat transfer which would have occurred if no convection were present. For R*
^ R*, Nu = 1. In our dimensionless variables
A = #- / * " (u3T - d3T)dx2.
IT J 0

A similar expression holds in the three-dimensional case. When the convection is

steady A does not depend on x3.
When R* > R* steady cellular convection sets in. It is of interest to determine
its magnitude and its spatial configuration. The problem of its magnitude, and in
particular the dependence of A on R* and a, when the motion is steady, has been
studied by the author in previous work [2], [10]. As to the shape of the convection
cells, it is known that flows may exist in which the cells, when viewed from above,
look like hexagons, or like rectangles with various ratios of length to width, or like
rolls, i.e. two-dimensional convection cells (see [11]). However, only cellular structures which are stable with respect to small perturbations
can persist in nature or
be exhibited by our method. It has been shown, numerically by the author [10],
experimentally by Koschmieder [12] and Rossby [13], theoretically, in the case of
infinite o- and small perturbations,
by Busse [14], that for R*/Rc* < 10 the preferred cellular mode is a roll. Busse showed that the rolls are stable for wave numbers in a certain range. We shall now demonstrate numerically the impermanence
of hexagonal convection and the emergence of a roll.
Consider the case R*/R* = 2, o- = 1. We assume the motion to be periodic
in the xx and x% directions, with periods respectively 4tt/cz V 3 and 4tt/ (the first
period is apparently in the range of stable periods for rolls as predicted by Busse).
These are the periods of the hexagonal cells which could arise when R* = R*.
The state of rest is perturbed by adding to the temperature in the plane x3 = Aj3
a multiple of the function <f>ixx,x2) which corresponds to a hexagonal cell, and
adding a small constant to the temperature
on the line xx (3/4)(4x/a V3),
x2 = (3/4)(4,r/a).
We then follow the evolution of the convection in time, using a

net of 24 X 24 X 25, i.e.

Axi = (4Tr/a V 3)/24 ,

A.r2 = (4Tr/a)/24 ,

Ax3 = 1/24 .

We choose e = Ax22, Ai = 3Ax32. The convection pattern is visualized as follows: the velocities in the plane x3 17A^3 are examined. If 3(8,r,is) > 0 an * is

printed, if u3iq,,,xs) ^ 0, a 0 is printed.

ALEXANDRE

Figure 4. Evi

JOEL CHORIN

of a Convection Cell

*********0000000000000
**

*******000000000000000
**
*******00000000000000000
*******0000000000000****
******00000000000000*0**
****00000000000000000
***
000000000000000000000000
000000000000000000000000
000000000000000000000000
0000000000*0*00000000000
000000000000000000000000
000000*****0*00000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
0000000000000000***00000
0000000000000000***00000
0000000000000000
***00000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000

******000000000000
******
***000000000000000000
***
000000000000000000000000
000000000*******00000000
00000************0000000
0000***********000000000
000**********00000000000
000********0000000000000
oooo*****000000000000000
000000000000000000000000
000000000000000000000000
00000000000000*
******000
0000000000000*********00
0000000000000*********oo
0000000000000*********00
00000000000000
*******ooo
000000000000000000000000
00000*000000000000000000
00*****00000000000000000
********0000000000000000

4a. After 1 step (Nu = 1)

4b. After 10 steps (Nu = 1)

**********00000000000000
***********000000000000*
***********000000000000*

**********00000000000000
***********000000000000*

**********
*ooooooooo
****
***0000000**************

*************
*ooooooo
***
************000000000000

*********
*ooooooooo
*****
*********
*ooooooo
*******
*******
*ooooooooo
*******

***********
*nooooooooo
**
************
*qoooooooo
**
*************
*oooooooo
**
**************oooooooo*
*

*************
*oooooooo
**

************
*oooooooooo
*

************ *QQQQQf)00000

***********0000000000000
0**********0000000000000
00 ********00000000000000
00000***0000000000000000
0000000000000*******oooo
00000000000***********00
00000000000************0

00**********000000000000
0000********000000000000
0000000*****000000000000
00000000*******000000000
00000000**********000000
00000000*************ooo
00000000**************Q()

0000000000 **************

00000000 ***************q
00000000 ************** *n
00000000 ************** *n
000000000 ************* *q

0000000000**************
0000000000 **************
00000000000*************
000000000000**********00
000000000000000000000000
00******0000000000000000
0*********00000000000000
4c. After 125 steps (Nu = 1.25)

00000000000***********00
000000000000000000000000
000000000000000000000000
00*****00000000000000000
4d. After 225 steps (Nu = 1.72)

NUMERICAL SOLUTION OF THE NAVIER-STOKES EQUATIONS

761

0000********000000000000

0000***********000000000

o**************
*oooooooo

0000************00000000

00************oooooooooo0000***********000000000
0000***********000000000
o*************
*ooooooooo
o**************
*oooooooo
o**************
*oooooooo
00**************00000000
oo*************000000000
000************000000000
0000***********000000000
00000**********000000000
00000**********000000000
00000***********00000000
00000************0000000
00000*************000000
00000*************000000

0000**************
*ooooo
00000**************00000

00000*************
*ooooo

000000************000000
0000000**********0000000
000000000******000000000
000000000000000000000000
000000000000000000000000
4e. After 325 steps (Nu = 1.76)

000 *************00000000

000 *************00000000
000*************00000000
000*************00000000
0000***********000000000
0000***********000000000

0ooo***********ooooooooo
0ooo***********ooooooooo
0000***********000000000
0000***********000000000
000************000000000

000************
*oooooooo
000************
*00000000
000*************00000000
0000***********000000000
0000***********000000000

00oo***********ooooooooo
0000***********000000000
00000*********0000000000
00000*********0000000000
4f. After 430 steps (Nu = 1.77)

The evolution of the convection is shown in Figs. 4a, 4b, 4c, 4d, 4e, and 4f.
The hexagonal pattern introduced into the cell is not preserved. The system evolves
through various stages, and finally settles as a roll with period 4x/a V 3. The value
of Nu evaluated at the lower boundary is printed at the bottom of each figure.
The steady state value for a roll is 1.76. The final configuration of the system is
independent of the initial perturbation.
The calculation was not pursued until a
completely steady state had been achieved because that would have been excessively time consuming on the computer. It is known from previous work that
steady rolls can be achieved, and that the mesh used here provides an adequate
representation.

Conclusion and Applications. The Benard convection problem is not considered

to be an easy problem to solve numerically even in the two-dimensional case. The
fact that with our method reliable time-dependent results can be obtained even
in three space dimensions indicates that the Navier-Stokes equations do indeed
lend themselves to numerical solution. A number of applications to convection
problems, with or without rotation, can be contemplated; in particular, it appears
to be of interest to study systematically the stability of Benard convection cells
when o- t, and when the perturbations
have a finite amplitude.
Other applications should include the study of the finite amplitude
of Poiseuille flow, the stability of Couette flow, and similar problems.

instability

Acknowledgements.
The author would like to thank Professors Peter D. Lax
and Herbert B. Keller for their interest and for helpful discussions and comments.

762

ALEXANDRE

JOEL CHORIN

New York University

Courant

Institute

of Mathematical

Sciences

New York, New York 10012

1. H. Fujita

& T. Kato,

"On the Navier-Stokes

initial value problem.

I," Arch. Rational

Mech. Anal, v. 16, 1964, pp. 269-315. MR 29 #3774.

2. A. J. Chorin,

"A numerical

method

for solving incompressible

viscous flow problems,"

J. Computational Physics, v. 2, 1967, p. 12.

3. J. O. Wilkes,

"The finite difference

computation

of natural

convection

in an enclosed

cavity," Ph.D. Thesis, Univ. of Michigan, Ann Arbor, Mich., 1963.

4. A. A. Samarskii,

"An efficient

bolic equation in an arbitrary

difference

method

for solving

a multi-dimensional

para-

domain," Z. Vycisl. Mat. i Mat. Fiz., v. 2, 1962, pp. 787-811 =

U.S.S.R. Comput. Math, and Math. Phys., v. 1963, 1964, no. 5, pp. 894-926. MR 32 #609.
5. R. Varga, Matrix Iterative Analysis, Prentice-Hall,
6. P. R. Garabedian,

"Estimation

Englewood Cliffs, N. J., 1962.

of the relaxation

factor

for small

mesh size,"

Math.

Comp., v. 10, 1956, pp. 183-185. MR 19, 583.

7. C. E. Pearson,
"A computational
pressible viscous flow problems," Report

method for time dependent two dimensional incomNo. SBRC-RR-64-17,

Sperry Rand Research Center,

Sudbury, Mass., 1964.

8. A. J. Chorin,

"The numerical

solution of the Navier-Stokes

equations

for incompressible

fluid," AEC Research and Development Report No. NYO-1480-82, New York Univ., Nov. 1967.
9. S. Chandrasekhar,

Hydrodynamic

and Hydromagnetic

Stability,

Internat.

Series of Mono-

graphs on Physics, Clarendon Press, Oxford, 1961. MR 23 #1270.

10. A. J. Chorin,

"Numerical

study of thermal convection

in a fluid layer heated from below,"

AEC Research and Development Report No. NYO-1480-61, New York Univ., Aug. 1966.
11. P. H. Rabinowitz,
"Nonuniqueness
Arch. Rational Mech. Anal. (To appear.)
12. E. L. Koschmieder,
"On convection

of rectangular

solutions

of the Benard

on a uniformly

heated plane,"

problem,"

Beitr. Physik. Ahn.,

v. 39, 1966, p. 1.
13. H. T. Rossby, "Experimental
study of Benard convection with and without rotation,"
Ph.D. Thesis, Massachusetts Institute of Technology, Cambridge, Mass., 1966.
14. F. Busse, "On the stability of two dimensional convection in a layer heated from below,"

J. Math. Phys., v. 46, 1967, p. 140.