COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING 65 (1987) 85-96

NORTH-HOLLAND

A NEW FINITE ELEMENT FO~ULATION FOR COMPUTATIONAL

FLUID DYNAMICS: VH. THE STOKES PROBLEM WITH VARIOUS
WELL-POSED BOUNDARY CONDITIONS: SYMMETRIC FORMULATIONS
THAT CONVERGE FOR ALL VELOCITY /PRESSURE SPACES*

Thomas J.R. HUGHES and Leopold0 P. FRANCA*”

institute for Computer methods in Applied technics and engineering, ~iv~ion of Applied ~ec~un~cs,
Stanford University, Stanford, CA 94305, U.S.A.

Received 14 April 1987

Symmetric finite element formulations are proposed for the primitive-variables form of the Stokes
equations and shown to be convergent for any combination of pressure and velocity interpolations.
Various boundary conditions, such as pressure, are accommodated.

1. Introduction

In [ll], (hereafter referred to as Part V) we developed a finite element method for the
Stokes problem in primitive variables which was proved convergent for any combination of
continuous inte~olations. This is an advantage compared with the classical Galerkin formula-
tion which requires satisfaction of the BabuSka-Brezzi condition (see ]2,3]). The Babuska-
Brezzi condition precludes the use of many seemingly natural combinations of velocity and
pressure interpolations.
In this paper we continue our investigation of the Stokes problem and generalize and
improve upon the formulation presented in Part V. The following shortcomings of Part V were
noted: For all but the simplest elements, the matrix problem was nonsymmetric; and the
formulation apparently did not improve upon the stability of the classical Galerkin formula-
tion when the pressure was assumed discontinuous. In the approach proposed herein
symmetry is attained for all elements, and convergence is proved for ail velocity and pressure
inte~olations. Specifically, a pressure inte~olation of arbitrary order, continuous or discon-
tinuous, may be combined with a velocity interpolation of any other order, and the resulting
method is stable and convergent.
We apply our approach to two different boundary-value problems for the Stokes equations.
The first problem considers boundary conditions recently proposed and studied by Pironneau
*This research was sponsored by the NASA Langley Research Center under Grant NASA-NAG-1-36~ and the
IBM Almaden Research Center under Grant No. 604912.
**Graduate Research Assistant. Partially supported by Brazilian Government fellowship CAPES Proc.
4093182-5.

004.5-7825/87/$3.50 0 1987, Eisevier Science Publishers B.V. (North-Holland)

86 T. J. R. Hughes, I... l? Franca, The Stokes problem: convergence for all velocitieslpresswes

[16]. The variational formulation of this problem accommodates pressure as a weakly (or
strongly) enforced boundary condition, and seems particularly well suited to fluid mechanics.
Velocity and vorticity boundary conditions are also permitted. The second problem is the
traditional one involving velocity and traction boundary conditions. It is identical to the
standard displacement-pressure formulation of classical isotropic incompressible elasticity
theory (due to Herrmann [9] and derived from the Hellinger-Reissner principle; see [8, 17]),
although it has been used extensively for fluid mechanics as well.
An outline of the paper follows. In Section 2 we present the Stokes boundary value
problem for Pironneau’s boundary conditions. In Section 3 we present the finite element
formulation and in Section 4 we derive error estimates. In Section 5 we present the traditional
Stokes boundary value problem, its finite element formulation and error estimates. In Section
6 we draw conclusions.
The symbol q is used to denote the end of a proof.

2. Stokes equations with velocity and pressure boundary conditions

Let 0 be an open, bounded region of R”, where d = 2 or 3. with piecewise smooth

boundary F. Let {r, . J2, r3, F,} be a partition of F. The unit outward normal vector to r is
denoted by n.
Let w be a vector field on R. The vector Luplucian is defined by (see [12, pp. 88, 116, 1181):

Aw = grad div w - curl curl w . (1)

The equations of Stokes flow are,

PAu-gradp+f=O, (2)

div u = 0 , (3)

where p+ is the viscosity (assumed constant), u is the velocity, p is the pressure, and f is the
body force.
Vector fields w defined on F may be written as,

w = w, + w, , (4)
where

w, = w,n = (n * w)la (normal component) , (5)

W,=nqWXn) (tangential component) . (6)

We consider the following boundary conditions:

u=g (7)
 T. .I. R. Hughes, L. P. Franca, The Stokes problem: convergence for all velocitieslpressures 87

n*u=g,
n X ((curl u) X n) = co, I on r, , (8)

p=dJ
n X (u X n) = g, 1 on r, , (9)

P=+
n X ((curl U) X n) = co, I on r, . (10)

By virtue of the identity,

n X w, = n X (n X (IV X n))

= n X (w - w,n)
=nXw, (11)

the tangential velocity and vorticity boundary conditions may be expressed in the following
alternative forms, respectively:

nXu=nXg,, (12)

n X curl u = n X w, . (13)

Physical and mathematical justification for the boundary conditions on r, U r, is presented in

WI.
The object of the Stokes problem is to find u and p which satisfy (2), (3), and (7)-(10).
In subsequent sections it is convenient to employ the notations,

r, = r, u r,, )
(14)
r, = r, u r, . (1%
Remark
The well-posedness of the boundary value problem has only been rigorously established in
specific instances. For example, Pironneau [16] presents the following result: If fi is bounded
and simply-connected, with piecewise C* boundary, if r - r, is not empty, if r, is simply-
connected and C’, and if r’ is empty, then for sufficiently regular given data, there exists a
unique solution {u, p} E H*(R)3 x L*(O). (If r3 is empty, the pressure is unique up to a
constant.)

3. Finite element formulation

Consider an element partition of a. Let W be the interior of the eth element, let r” be its
boundary, and
88 T. J. R. Hugiaes. L. P. Francn, The Stokes problem: convergence for all velocities/pressures

ii=!Jfr (element interiors) , (16)

I’= /J r” - r (element interfaces) . (17)

Let ?@ be the set of velocity weighting functions, let Yh be the set of velocity trial solutions,
and let 9’ be the set of pressure weighting functions and trial solutions. %“‘Iand 9” consist of
typical C” finite element interpolations. In addition,

12= g on r, , n- uh = g, on &, n X uh = n X g, on 1; vu” E 3’” .

(19)

Functions in 9’ may be continuous or discontinuous across element interfaces.

The finite element problem is to find Uh = {u”, p”} E 9’ X 9’, such that for all W” =
(Wh, qh) E Yh x Ph,

B(Wh,U”) = L(Wh), (20)

where
B(Wh, U”) = (curl wh, p curl u”){] + (div wh, p div u’*)~]- (div WI’,P”)~)

- ( qh, div u”)~ - wh - grad qh), p Au” - grad p” _

i i f1

- - ($ qh-Phjii
(5 us”ll,IPhIl), ’ (21)

L(Wh) = (w” -t- $ (,u hwk -grad qk)+ f), - (w”_ ,un x u,~)~,w

(22)

in which ( 0 , *ID denotes the L’(D) inner product, h is the mesh parameter, [ ] is the jump l

operator, and a and /3 are nondimensional stability constants. We assume h is defined locally.
The formal consistency of the method is apparent from the Euler-Lagrange form of (20):

O=-(wh+$(pAwh- grad qh), p Aah - grad ph +f), - (q”, div u”)~~

+ (n * wh, p div u~)~+ - (wh, &n x curl ah - n x w.~))~~,

+ (n - wk. p.Bdiv u”j)i:

- (wh, pn X ([curl uh])f

(23)
 T. J. R. Hughes, L. P. Franca, The Stokes problem: convergence for all velocities lpressures 89

In obtaining (23) we have employed the following integration-by-parts formulas:

(curl wh, curl u”)~ + (div wh, div u”>~ =

- (wh, Au’)~ + (n l wh, div ‘c~)~. - (wh, n x curl ZJ~)~.+ (n l wh, I[div uhl)f

- (wh, n x [curl uh])p , (24)

- (div wh, ph)n = (wh, grad ph)h - (n l wh, I[P”]>~= - (n l wh7 ph)r; . (25)

Remarks
(1) Note that the pressure and tangential vorticity boundary conditions are satisfied
weakly.
(2) If (Y= p = 0, we reduce to a Galerkin formulation. In general we assume cyB 0 and
p 30.

4. Error analysis

Let u = {u, p} denote the solution of the Stokes problem (2), (3), and (7)-(10). It follows
from (23) that,

B(Wh,U) = L(Wh), (26)

and thus, by (20)

B(Wh,E) = 0, (27)

where E = {e,, eP} = Uh - U is the error in the finite element solution. We refer to (27) as the
consistency condition.
For quasi-uniform mesh refinements, it considerably simplifies the error analysis if we
assume h in (21) and (22) is constant for each mesh. This represents no loss in generality.
We will need to assume the following inverse estimate:

h2[1Awh11; d C( Jlcurl ~“11: + lldiv whlli) Vwh E “Irh , w-9

where F is a constant.
Let

= g (Ilcurl
lllWh11(2 W*11: + lldiv w”IIk) + $ IIt@ qhII:

+ ph
2P ms”nllf + Ilqhl12
r*1- (2%
90 T. 1. R. Hughes. L. P. Franca. The Stokes problem: convergence for all vel~cit~es~~ressur~~

LEMMA 4.1. Let ti” = (w”, -q”> E Y” X 9’. Assame the inverse e&mare holds and 0 <
(Ysc -- ‘, where C is the constant in (28). Then.
B(Wh,w”p fllWhl((’ VWhE YPh
x Ph . (30)

We refer to (23) us the stability condition (a&u known as the coercivity condition).

PROOF.

2 II(WhlI12
. cl (31)

Let fi” E 9’” x ph denote an interpolant of U. The interpolation error is denoted by

H=I?h-U={qu,~n}. (32)

Thus,

E=Eh+H, (33)

where

Eh = (e:, eij E Yfh X P’” . (34)

We need to assume the following interpolation estimates:

where c, and c, are functions of u and p, respectively. These notations are used subsequently,
it being understood that the values may change in each instance.
The following integration-by-parts formula is also used subsequently:

(37)
 T. J. R. Hughes, L. P. Franca, The Stokes problem: convergence for all velocitieslpressures 91

THEOREM 4.2. Assume the inverse estimate and the interpolation estimates hold; assume the
consistency and stability conditions also hold. Then,

)ljE111*d c,h2k + cph2(‘+l) . (38)

PROOF. Let Eh = {eh,, -$} E Vh x Ph, and let a, b, and c be constants whose values will be
assigned later. We estimate IlI~~lll as follows:

IllEhl((2 S B(Eh, Eh) (stability)

= B(Eh, E - H)
= -B(Eh, H) (consistency)
d IB(Eh, H)I
= I p(cur1 et, curl v,)~ + p(div e:, div qU)* - (div e:, T,)~

- (grad eh,, S)JL + (Ikh,D,n * s>f + (eh,, n * s,>, 4

ah*
- {(grad eh,, grad rl,h - p*(Ae:, AT,>A - Agrad et, Arl,)~
+ 2E.L

(by definition of B and the integration- by-parts formula (37))

+ 2Fb([curl q,lli + $ b-‘Ildiv e:lli

+ 4pbljdiv q,IIi + F b-‘)ldiv e:[[i + $ bllt),IIf,.

+ $ a-‘(Igrad e”,lla + ($)-‘allqull~ + 5 c-‘lle”,ll’,

+ ph -l
( 2p )

ah*
~ a-‘llgrad e”,IIi + $ allgradq,11: + T h2b-111Ae~I)~
+ 2p

+ 5 a2h2bllAqu11i + $ a-‘llgrad e”,lli + 5 ah*a(lAq,((~

+ ; h*b-‘IlAe:I)f+ + F blkradq,lk + $f c-‘llUe”,llllf

+ @ cII[[~~nil’ + L2.f
2j.L p l- 2j.k
92 T. J. R. Hughes. L. P. Franca, The Stokes problem: convergence for all velocitiesipressures

+ E 2cem’(llUefJll~
+ llefllI;Y)}+ p{bllcurlrl,Il~J+2blldiv%IIi

By selecting a = 3, b = 1 + 2C, and c = 2, the first term on the right-hand side of (39) becomes
i (ll~“ljl’. The second and third terms may be estimated by (35) and (36), respectively.
Therefore.

II(EhJ((‘s cuhzX:+ cph7(‘+‘) . (40)

Likewise, by (29), (3.5), and (36),

I((Hlll* d c,hzk + cph”‘+” . (41)

This completes the proof of the theorem. Cl

Remarks
(1) The theorem is valid for 0 < (Y< C- ‘, and 0 < p. Note that there is no upper bound
imposed upon p. To balance errors, we may as well take p = 0( 1). The upper bound on (Y
emanates from the necessity of controlling Ae: on element interiors. If the velocity is
piecewise linear, this term vanishes identically, and the upper bound on (Y is no longer
necessary. In this case we may also take (Y= O(1).
(2) By a generalized Poincari-Friedrichs inequality, proved by Arnold [l], which in the
present circumstances can be written as,

m, IG4 c,(h’lkrade,ll~~
+ hIllIe,llil~
+ hlle,lI~j)3 (42)

we may obtain an L’(0) convergence result for pressure, namely,

lle,II~ s c,h2’k-” + c,h” . (43)

This is suboptimal and not useful when either k = 1 or I= 0. In many circumstances (43) is
pessimistic because, for appropriate hypotheses, duality arguments can be used to attain the
improved estimate:
 T.J. R. Hughes, L. P. Franca, The Stokes problem: convergence for all ve~~ci~ie~~~ressures 93

(44)
(3) Within the proposed formulation we may distinguish between four different types of
methods:
Type I: a! = 0, /3 = 0 (Galerkin). Classical convergence theory requires satisfaction of the
BabuBka-Brezzi condition. Various texts summarize known examples (e.g., see [4,7,10,18]).
Among the most popular are the continuous pressure elements P2Pl and QZQl, and the
discontinuous pressure element Q2Pl.
Type II: CY> 0, p = 0. All continuous-pressure elements are convergent. Equal-order
elements (i.e., PkPk and QkQk, k a 1) are particularly attractive due to implementational
considerations.
Type III: cx = 0, p > 0. The /3 terms are sufficient to render convergent certain low order,
discontinuous pressure elements such as PlPO and QlPO.
Type IV: CY> 0, p > 0. The (Y and p terms are both required for higher-order discon-
tinuous-pressure elements. Equal-order elements again appear to be the most attractive. This
is of course the general case and is applicable to any element.
(4) The /3 terms necessitate a nonstandard finite element assembly algorithm. For this
reason, continuous pressure elements, which oniy require the presence of the CYterms, may be
more convenient in practice.
(5) The traction vector along ‘no-slip’ walls may be calculated from

t=-pn+t,, (45)
tLt = -pn X curl u . (46)

Accurate calculation of this quantity is facilitated by the following post-processing procedure

(see [Ml, p, 1071 for an elementary description of the basic ideas): Let W; denote weighting
functions expanded in terms of velocity basis functions which do not vanish on r,. Let t:
denote a vector field expanded in terms of the same basis functions. Define ti by:

(M$., tt)r, = (curl wh,, p curl u”)~ + (div w:, p div uh - ph)n

-( $(uh2Aw~, p Auh -grad ph 4f)d - (w:, f)e . (47)

In forming the right-hand-side, only those elements with a boundary segment contained in r,
need to be included. The approximate traction is then given by:

th = -phn + t; . (48)
It is interesting that, despite traction not being a boundary condition for the problem
considered, a natural post-processing procedure for wall traction nevertheless ensues.
94 T. J. R. Hughes, L. P. Francu, The Stokes problem: convergence for all velocities/pressures

5. Stokes equations with velocity and traction boundary conditions

Let { rg, r,} be a partition of r. This time we work with the following form of the Stokes
equations:

div u +f= 0 on R , (49)

div u = 0 on a , (50)

u = -pZ + 2jx(u) on 0 , (W

u=g on rx, (52)

n*a=h on r,,, (53)

where u is the Cauchy stress tensor, Z is the identity tensor, F(U) is the symmetric part of the
velocity gradient, and h is the boundary traction. This form of the Stokes problem is more
useful in the case of incompressible elasticity.
A finite element formulation of (49)-(53) analogous to the one presented in Section 3 is
given as follows: We require all uh E Yh to satisfy uh = g on rp, and all wh E ‘Vh to satisfy
wh = 0 on rp. Find Uh E .!Y”X Cl” such that for all Wh E ‘Vh x ph,

B(Wh, U”) = L(Wh), (54)

where

B(w~, u”) = (&(wh), 2p4uh)), - (div wh, ph)n - (qh, div u”)~>

-( $ div u(W"),div u(U”))~ - (e 14”k l[P”1l), 7 (55)

L(Wh) = (Wh + g div u(W"), f), +(wh,h),,. (56)

The Euler-Lagrange form of (54) is

o=-
(wh+ $ div u(W”), div u(U”) .+f), + (wh, n* u(U”) - h),,

+(wh, n. [u(U”)])f - ( qh, div u”)~ - (57)

For the Dirichlet problem we are able to prove the following error estimate:

~IIhA~~ + $ Ilgrad e,I/i + e Il[eJll~ B c,hZk + Cph’(‘+‘)

(58)
 T. J.R. Hughes, L.P. Franca, The Stokes problem: convergence for all velocitieslpressures 95

To obtain (58) we need to assume (Y< l/(2;), where c”is the constant in the inverse estimate,

h*]]div e(rv”)]]i < C”](E(W~)](~ Vwh E 7fh .

The analysis is closely related to that presented in Section 4 and so it is omitted. (A proof
for the case p = 0 may be found in [5].) We are unable to prove convergence for the case in
which r, is not empty, but presume that this case is also well behaved. _

6. Conclusions

The symmetric methods presented herein were proved to be convergent for arbitrary
combinations of velocity and pressure interpolations. Two modifications of the classical
Galerkin formulation are required to achieve this result. Both involve the addition of
‘least-squares’ forms of residuals: one is the momentum equation residual on element
interiors, the other is the pressure continuity residual on element boundaries. The bilinear
form of the modified formulation is rendered coercive, in contrast to the classical Galerkin
formulation, and this enables us to circumvent the Babuska-Brezzi test. (We have also
employed similar techniques in the development of mixed methods which satisfy the BabuSka-
Brezzi condition for a wider range of functions than the classical Galerkin formulation; for
applications to elasticity and various structural model problems, see [S, 6, 13-151). The
methodology appears general and offers a new line of attack on the problem of developing
convergent mixed finite element methods.

References

PI D. Arnold, An interior penalty finite element method with discontinuous elements, SIAM J. Numer. Anal. 19
(1982) 742-760.
I. BabuSka, Error bounds for finite element method, Numer. Math. 16 (1971) 322-333.
;i; F. Brezzi, On the existence, uniqueness and approximation of saddle-point problems arising from Lagrange
multipliers, RAIRO Ser. Rouge Anal. Numer. R-2 (1979) 129-151.
141 G.F. Carey and J.T. Oden, Finite Elements: A Second Course II (Prentice-Hall, Englewood Cliffs, NJ, 1983).
PI L.P. Franca, New mixed finite element. methods, Ph.D. Thesis, Applied Mechanics Division, Stanford
University, Stanford, CA, 1987.
161 L.P. Franca, T.J.R. Hughes, A.F.D. Loula and I. Miranda, A new family of stable elements for nearly
incompressible elasticity based on a mixed Petrov-Galerkin finite element method, presented at the
conference on The Impact of Mathematical Analysis on the Solution of Engineering Problems, University of
Maryland, College Park, MD, 1986; also Numer. Math. (to appear).
171 V. Girault and P.A. Raviart, Finite Element Methods for Navier-Stokes Equations, Theory and Algorithms
(Springer, Berlin, 1986).
PI E. Hellinger, Der Allgemeine Ansatz der Mechanik der Kontinua, in: Encyclopidie der Mathematischen
Wissenschaften 4 (4) (1914) 602-694.
[91 L.R. Herrmann, Elasticity equations for nearly incompressible materials by a variational theorem, AIAA J. 3
(1965) 1896-1900.
WY T.J.R. Hughes, The Finite Element Method: Linear Static and Dynamic Finite Element Analysis (Prentice-
Hall, Englewood Cliffs, NJ, 1987).
96 T. J. R. Hughes, L. P. Frarlca, The Stokes problem: con~~ergerzcefor a11 ~~elocitieslpressttrt~s

[ll] T.J.R. Hughes, L.P. Franca and M. Balestra, A new finite element formulation for computational fluid
dynamics: V Circumventing the Babuska-Brezzi condition: A stable Petrov-Galerkin formulation of the
Stokes problem accommodating equal-order interpolations, Comput. Meths. Appl. Mech. Engrg. 59 ( 1YXh)
85-99.
[12] W. Jaunzemis, Continuum Mechanics (MacMillan, New York. 1967).
[13] A.F.D. Loula, L.P. Franca, T.J.R. Hughes and 1. Miranda, Stability, convergence and accuracy of a new
finite element method for the circular arch problem. Comput. Meths. Appl. Mech. Engrg. 63 (1987) 281-303.
[14] A.F.D. Loula, T.J.R. Hughes, L.P. Franca and I. Miranda, Mixed Petrov-Galerkin method for the
Timoshenko beam, Comput. Meths. Appl. Mech. Engrg. 63 (1987) 133-1.54.
[15] A.F.D. Loula, I. Miranda. T.J.R. Hughes and L.P. Franca, A successful mixed formulation for axisymmetric
shell analysis employing discontinuous stress fields of the same order as the displacement field, Fourth
Brazilian Symposium on Piping and Pressure Vessels. Salvador, Brazil, 1986.
(161 0. Pironneau. Conditions aux limites sur la pression pour les equations de Stokes et de Navier-Stokes, C.K.
Acad. SC. Paris, t. 303, S&e I (9) (1986) 403-406.
[17] E. Reissner, On a variational theorem in elasticity. J. Math. Phys. 2Y (2) (lYS0) 90-95.
[18] F. Thomasset. Implementation of Finite Element Methods for Navier-Stokes Equations (Springer, New
York, 1981).