INTERNATIONAL

JOURNAL

rOR

NUMERICAL

METHODS

IN ENGINEERING.

VOL. 3. 317-325

(1971)

ON THE CALCULATION OF CONSISTENT STRESS

DISTRIBUTIONS IN FINITE ELEMENT APPROXIMATIONS
J. T. ODEN

R".ll'arch /I/s'i""t'.

AND

II. J. nRAUCIIU

UI/il'ersily of Alahama ill HIiI/I.willt'. HIiI/lsl'illt'. Alahama

SUMMARY

The Iheory of conjugale approximations! is used 10 oblain consistent approximations of stress fields in
finite element approximations based on displacement assumptions. These consistent stresses are continuous across interclernent boundaries and involve less mean error than those computed by the
conventional approach.
INTRODUCTION
The problem of calculating stresses in finite element models based on displacement approximations has long been a source of difficulty. Typically. the procedure is as follows.:!
I. The displacement components u. r and II' over a typical finite clement are represented
approximately as a linear combination of their values (and possibly values of their derivatives)
at a prescribed number of nodal points.
2. Following standard procedures,:! element stiffness matrices klel are calculated using. say.
the principle of minimum potential energy.
3. A global stiffness matrix K is obtained by connecting all of the elements together and
applying appropriate boundary conditions. This leads to equilibrium equations governing the
connected assembly of elements.
4. The equilibrium equations are solved and stresses in a typical element are evaluated by
direct substitution of local displacement field into an appropriate constitutive equation. for
example
(I)
Here aIel is a vector of local stresss components. Dlt) is the elasticity matrix. HIe) is an operator
appearing in the strain displacement relations. Qle) is a matrix transforming the global generalized
displacements U to the local nodal displacements ulC) and P is a vector of generalized forces.
As pointed out by Melosh.3 stresses computed using the above procedure represent. at best,
only averages of the true stresses over a finite element. The stress field so computed will. in
general, exhibit finite discontinuities at the boundaries of each element. To evaluate stresses at a
nodal point. it is common practice to compute some kind of weighted average of stresses over
elements meeting at that node. Wilson4 noted that such element stresses generally do not represent
the true state of stress in anyone point of the clement: he developed a 'weighted average method'
which gave good results for both interior and boundary elements. The paper by Turner and coworkers:; includes a detailed discussion of an averaging technique. based on an equivalence of nodal
forces and element stresses. that overcomes large stress discontinuities in finite elements. Under
reasonable smoothness conditions on the local displacement fields. it can be shown that element
stresses computed using the procedure outlined previously do indeed converge in the mean to the
Reeeh'ed 30 Jail/wry /970
Revist'd 6 April /970
@

1971 by John Wiley &: Son5. Ltd.

3t7

318

J. T. ODEN ANO H. J. BRAUCHLI

exact values.II-1I In his 1964 monograph, GallagherlO proposed a 'direct' method of formulation of
stiffness matrices which involved independent developments related to both stress and displacement fields for an element. and which thercby overcame many of the usual dilliculties in
interpreting boundary stresses.
In the present paper. we give a consistent procedure for calculating stresscs in finite elements
which is based on the idea of conjugate approximations.l
While we make no claim that the method
is computationally more eff1cient than those generally used to compute stresses, it is nevertheless
consistent. sincc the distribution of such stresses in the finitc element modcl is directly dictated by
the displacemcnt approximations
used in formulating element properties. In other words. the
stresses so computed arc consistent with the displacement approximations.
Moreover if the
displacement approximations are continuous. the consistent stresses will also be continuous and
will be more accurate (in a certain mean square sense) than those computed following conventional
procedures.
The procedure for computing consistent stresses arises from the notions of biorthogonal bases
for finite-dimensional conjugate spaces. If the function u(x. y. z) (say. a component of displacement) is approximated by functions of the form
ii = p(x,y.

z) S

(2)

where p(x, y. z) is a matrix of suitably chosen interpolation functions and S a vector of generalized
displacements, thcn the identification of p(x. y. z) defines a basis for a finitc-dimensional
subspace of the spacc containing the actual displacement function II. Since the functions p(x, y. z)
are not orthogonal with respect to an appropriately defined inner product 0/1 the space, it is
possible to construct a different set of so-called conjugate approximation
functions A(X. y, z)
which form the basis of a space conjugate to that spanned by p(x. y, z). The significance of this
observation. which is to be used in this paper, is that element stresses should be referred to the
conjugate basis rather than mapped completely out of the subspace of thc functions p(x, y. z),
as is done conventionally by the procedure outlined above.
THE CONCEPT

OF CONJUGATE

APPROXIMATIONS

Although it is not essential that the details of the theory of conjugate approximations be digested
in order to apply it to practical computations. we outline briefly here certain of its underlying
concepts. Further details can be found in Reference 1.
The displacement field u described in the previous section can be regarded as an element in an
infinite-dimensional
linear space J(' in which an inner product between two elements u and w is
defined by
(u.w) = j~lITWd&

(3)

Here ,rJt is the bounded domain of the functions u(x.y,z).

w(x,)'.z).
In most methods of appro ximation. we deal with a finite-dimensional
subspace (I>, instead of
whercin functions arc
approximatcd as linear combinations of a finite number of linearly independent base functions.
For example, each displacement component II, I' and Ii' can be represcnted as the sum of G
linearly independent functions cp;(x. y. z):

.n:

II;;::CPl(X.y,Z)Ut+CP2(X.y.Z)U2+
1';;::

CPl(X,y.Z)

Vt+CP2(X,y.Z)

II';;::

CPl(X,y.Z)

Wt+CP2(X,y,Z)

.. +CPa(x.y,z)Ua
V2+ .. +CP(;(x,y.z)
W2+ .. +CPa(x.y.z)

}
Vu
WIl

(4)

CONSISTENT

STRESS DlSTRIIlUnONS

IN FINITE ELEMENT

AJ>PROXIMA nONS

319

wherc Vi' Vi and Wi arc thc values of u. to and \I' (and/or thcir derivatives) at various nodal points
in .rJ. It is often convenient to recast equation (4) in matrix form. as indieated in equation (2), by
introducing

(5)

Then
ct>8

II =

(6)

Now corresponding to any finite set of base functions cplx, y. z), we can always construct a set
of conjugate approximation functions which form a countable biorthogonal basis to <[) and its
conjugate space $*: that is, a unique set of functions 1\.(.\",)',z) = [Ai(x.y.z)]
which satisfy the
biorthogonality condition

r <t>T1\. d&l' = I

(7)

lit

where I is the G x G identity matrix. To construct

fundamental mat,.ix

the conjugatc

C=(<t>.<t=

functions. we first compute the

<t>T<t>d~

(8)

lit

Since the elements of <t>are linearly independent,

singular. positive-definite matrix. Define

it can be shown that C is a symmetric,

1\.'1' = C-l <t>T

Then

(9)

j' <t>T 1\. d3i = r <t>T<t>d.9jlC-l

lit

non-

= CC-l = I

(10)

,!Of

as required. Moreover.

C-l

j~1\.'1' 1\.

d,gf.

(II)

The conjugate approximation functions 1\.(x. y, z) provide a basis for the conjugate subspace <[)*.
Thus elements in $* (i.e. elements appearing as linear functionals of elements in (I are linear
combinations of the functions 1\.(x. y. z).
We recall that in finite element approximations.
the functions <t>(x. y. z) are sums of elemental
approximation functions ~(")(x. y, z) that are defined only loeally. Thus, for a collection of E
elements,
E

<t>(x,y. z) =

L ~(el(x,y.

z)Qle)

(12)

wherein Qle) is the Boolean transformation appearing in equation (I). However. the functions are
generally designed so that <t>(x, y. z) is continuous over [fl. The local character of the functions
~(c)(x.y. z) is described by saying that they have almost disjoint support. Formally, we would like
also to define local conjugate approximation functions )"Ie)(x, y, z) for the purpose of calculating
local approximations of the stress fields ovcr an element. Sincc, in finite element approximations,
1\.'1' = C-l <t>'1' =

C-I ~ QlelT ~(e)'l'

e

(13)

320

J.

T.

aDEN

H. J.

AND

IIRAUCHLI

we can set
( 14)

so that locally
A(el

C-!

(15)

Lnl/lT~I/ITnltIT

Two important observations immediately become available:

I. Each local conjugate function AI)(x.y,:) is a linear combination of all of the local approximations \jJICI(X.y,Z);
thus, the functions AIel do not have disjoint support and are defined over
the whole connected domain :3i.
2. Each local conjugate function is a linear combination of the global functions cp(x, y, z);
thus. they arc continuous across inter-clement boundaries.
These two properties are markedly different from those of the stresses calculated by means of
equation (I).
As a concluding comment in this section. we remark that because of equation (9) the subspaces
(Il and (ll* coincide. Thus, any function It belonging to (II can bc expressed as a lincar combination
of either thc cp(x,)'. z) or the 1\(x.y, z). For example,

I/(X,)'.z)

cp(x.y.z)U

1\(x,y.z)1)

where

0=

1\T I/d.~

( 16)

(17)

, .it

and

u=CU
Technically.

1\(x.y, z)

(18)

belongs to the conjugate space (Il*.

STRESS

CALCULATIONS

The strain energy in an elastic finite element is

IV lei

= ~ /'

ale)'l' el"1

2, ar

d3f

( 19)

where ale) is the stress vector and IC} is the vector of strain components. If A is the vector of
generalized displacements for the connected model. we have, in accordance with Reference 2,
(20)

Assuming that elements of le) belong to the subspace (II, equation (19) defines a linear functional
on <I:> and shows that the elements of aiel belong 10 <fl*. Consequently, the functions a should be
represented as linear approximations
of the conjugate approximation functions Alel(x,)', z). The
validity of the assumption that 1") belongs to (Il rests on the commutativity of the projection II
of Yl into III and partial differentiation which. for finite element models. is at least approximately
true.! It follows that a given stress component. say a~). should be represented by

a<.f) =Ale)(x.y.z)S(rl

(21)

where
SIC}

j'
,[Jf

\jJlelT

axd.:#

(\jJll'l'l',

ax)

(22)

CONSISTENT

STI(ESS DISTRIIlUTIONS

Here ij.r is the conventional

stress
material (i.e. a(~l = DI~J E(cl).

11'\ FINITE ElEMI.NT

component

computed

from

321

APPROXIMATIONS

the constitutivc

equation

for the

Procedure
We are now in a position toconstruct
a systematic procedure for computing
consistent. continuous
stress distributions
in finite elcmcnt approximations.
Continuing
with the 'conventional'
procedure described
prcviously.
suppose that the componcnts
of displacement
over the finite element
model are given by cquations
of the form in cquation (6) ano that the discontinuous
stress field
is given by the eonstitutive
lav.'

a = DE = DHu = DBcI>L1
We proceed
I. If there

(23)

as follows.
were unknown

displacements

at G nodes,

c=

construct

the G x G fundamcntal

matrix

cj>'l'cp d.jfl

(24)

.I>r

whcre cj> is the row Illatrix of intcrpolation

functions.
If the body is homogeneous
and of mass
density p. the matrix C may be already available since it is lip times a submatrix
on the diagonal
of the consistent
mass matrix for the finite elcment approximation
of the body.
2. Construct
the I x G matrices of conjugate
approximation
functions
A

= cj>C-l

(25)

3. The vector Ci of discontinuous

(conventional)
stress cOlllponents
is of order R x l. where
= 6 for a three-dimensional
solid, R = 3 for plane stress. R = I for a straight rod. etc. Construct
the R x G matrix H defined by

aTcpdJ?

The matrix R has G columns. one corresponding

to each node.
4. The consistent
strcss distribution
over the entire collection
the formula

(26)

of elements

is now computed

by
(27)

However. in praclice
obtained
automatically

it is sufficient to have only the consistent

by computing

S
The columns
points.

of the

x G matrix

S contain

stresses

RC-l

the values of consistent

at eaeh

node.

These

are

(28)
stress at each of the G nodal

EXAMPLES
To demonstrate
the basic ideas. we now examine applications
of the theory to selected problems.
Consider first the one-dimensional
example of a straight bar with a linearly modulus of elasticity.
E(x) = Eo(\ +x) subjected
to a prescribed
longitudinal
displacement
of u(x) = ,~(I-x2/36).
The exact stress distribution
is a(x) = E(x) duldx = - <l:Eo x( I + x)/l8. To analyse the bar using
finite elements. we adopt a rather crudc displacement
approximation
consisting
of six elements
over cach of which the displacement
is assumed to vary linearly: this problem then amounts to a
slight gencralization
of one described
in Reference
\. In this case, the approximate
displaccment
field is given by
(29)
u = cj>(x) U

322

J. T. ODEN AND H. J. BRAUCHLI

where
CPl(X)

= [41,(X), cp:!(X), .... CP7(X)

= I-x, O~X~ I

cplx)

CP7(X)

= 6+x,

cp(x)

5lS;xlS;6

h(x - i) being the unit step function.

CPI.-(x)

u=
Since the conventional.

= 0 elsewhere,

36{36.35,32.27,20.

'6 -

(31 )

II.O}

EGO +x) dcp(x) U

.

the matrix S of consistent

and

stresses a(x) are simply

discontinuous
a(x)

we can now generate

into equation (26):

(30)

I),

x+2-i+2(i-.\')h(x-i+

a(x) cp(x) dx

(32)

dx

nodal stresses by introducing

equations

(29)-(31)

Eo ex

= - 216 [5,26.74, 146,242,362.220]

(33)

.0

We next calculate

the fundamental

matrix

C of equation

1351

-362
724

C-l = r 1'6cpTcpdx
II

\-1

=- 1
390

97
-194
675

(24) and then its inverse

-26
52

7
-14

- 182
676

-2

45
-182
679

I
4

Then. the consistent

-14
52
-194
724

Sym.

-2
-26

(34)

97
-362
1351

stresses at each node arc

S

RC-t

= -

Eo:
;6 [0'8,3,11,1.23'8,39'2.60'9,80'0]

(35)

Finally, the complete stress distribution

a(x) is Scp T.
The results of these calculations are shown in Figure I. It is seen that the consistent stresses are
continuous and lead to a better approximation
of the maximum stress.
As a final example, we demonstrate
that once the conventional
elemcnt stresses are known.
improved continuous stress distributions
can be obtained by using almost any convenient set of
conjugate approximation
functions cp and A in the construction
of equations (26) and (27). The
matrix cp used in such calculations
need not correspond
to that used to obtain the conventional
stresses. even though the actual consistent stresses computed as described previously will lead to
the most accurate approximation
for a given cp(x. y. z). Computationally,
however, the usc of
lower-order approximations
in cp to compute stresses may be more practical and will generally
provide better stresses than those given by conventional methods. Mathematically,
such approximations amount to projecting the conventional
stresses into a subspace difTcrent than that spanned
by the original collection of approximating
functions. The results of such calculations
applied

,r--L

84
80~

Exocl
0--0---0

70'

---

Conventional

As;;

----i

n
0

Consistent

tn

~l

60

::;;;

-i
",

-i
tn

-i
;c
m

tn
tn

~o

'0

~ 50
b

XISL

-i

6
z
tn

b"<
N

-<::

oil

I
en

E.'
iii

'"
~
in

If)

30

:::
::l

~
.....

~ 40

-i

c:

l:l

~
,r,

in

.,.,

-1000

Z
::j
m

,."

r0-----0

20

Conventional element
stresses (Reference 8)

Conjugate stress
approximation

>
"0
"0

;c
0
X

>

1-0

08

-i

-2000
10

,."

~
m

-i

2
3
4
5
6x
Figure I. Comparison of convcntional finitc elemcnt strcsscs with consistcnt
stresses

6
z

tn

Figurc 2. Normal stresses at top fibres along centre line in a centrally

loaded square plate
w
IV
W

324

J. T. ODEN AND H. J. BRAUCHLI

to the problem of bending of a clamped. sq uare plate subjected to a centrally located concentrated
force P are shown in Figure 2. Hcre the conventional stresses arc those obtained by Bazeley and
co-workersll using non-conforming elements. The improved continuous stress distribution shown
was obtained by using the very crude but conforming set of piecewise linear interpolation functions
in equations (26) and (27). It is observed that even in this casc the conjugate stress approximations
tend to give better estimatcs of stresses near points of stress concentration.

ACKNOWLEDGEMENT

The support of this work by the U.S. Air Force Office of Scientific Research
F44620-69-C-0124
is gratefully acknowledged.

under Contract

APPENDIX
The proof that the best approximation
in a mean square scnse arises from the use of conjugate
approximations was presented in Reference I in connection with quadratic funetionals defined
on self-conjugate spaces. Here we generalize this result to the case in which a linear spacc and its
dual are considered. We first consider cases in which the character of the spaces involved is
arbitrary. and then we apply the result to linear elasticity.
Let . and <!J denote dual linear spaces the elements of which are functions denoted f and g
respectively. Let (f.g) dcnotc the inner product on Y; and <!J. Further. let (II denote a finitedimensional subspace of .~. and \f!' its dual in rg. The spaces q> and 'l' are spanned by biorthogonal
bases CPkE$,l/1kEIF
such that

=8%'.

(cpl.:.ifJlII)

off and

Thcn, if F and G are the projections

F=

lII.k=

(36)

/1

g into $ and \f'.

Pcp/;.

G=G

wherein the repeated indices are summed from I to

F/;

1.2 .....

= (f. ifJk}

and

/1

G",

(37)

,/.111
m'P

= (g, cp",)

(38)

Let J\k CPk

and M", if!''' denote arbitrary elements in (I> and IV and consider the bilinear functional
J[Ak, M".1 = (f-J\k'CPk,gwhich. upon simplification,

M",ifJ"')

(39)

can be wrillen
J [AI.:, MII.1

(f-

F.g- G)+(FI.:-Ak)

(Gk- Mk)

(40)

Obviously. either of the projections defined by equation (36) makes the second term in equation
(40) vanish. If we now introduce a linear. positive-definite.
regular mapping K from .~ to <!J.
such that K(ct c'l'. then K defines a matrix Kill/; = {cp"" K(cp/..)} such that
K",/;j\k

Then J [AI.:. Mill]

1\<1,"

= J reV. k"'kAk] becomes a positive-definite

J [.\1.:]

= (f- F. k(f- F)}+k"'k(AIII-

(41)
quadratic

form

FIII)(J\k' - Fk)

(42)

Consequently cquation (42) assumes a minimum value when Ak is chosen according to equation
(38).
In the case of lincar elasticity. letfcorrespond
to the stress tensor G;j and g correspond to the
strain tensor eij' Then (f g) represents. with a constant. the strain energy. Let eij be the strain

CONSISTENT

STRESS D1STRIIIUTIONS

IN FINITE

ELE\lENT

APPROXI~lATlONS

325

resulting from a finite clement model of the displacement field. Then. in view of the above results,
the best approximation of aij' in the sense of minimizing the square error equation (42), is that
function in III for which the coefficients are computed using equation (38).
REFERENCES
J. H. J. Brallchli and J. T. aden. 'Conjugate approximation functions in finite clement analysis'. Q. appl.
Malll. to appear.
2. O. C. Zienkiewicz and Y. K. Chung. The Finile Elemtll1 Mt'thod in Slrllcll/ral and Cominl/lIm Aleehanies.
McGraw-Hili. London. 1967.
3. R. J. Melosh. 'Basis of derivation of matrices for the direct stiffness method" A!AA Jill. 1. 1631-]637
(1963).
4. E. L. Wilson. Finile /l'/Ilell1 Analysis ofTlI'o-dime"sio"al SrrllClIlrl's. Ph.D. Thesis. University of California.
Berkeley. 1963.
5. M. J. Turner. H. C. Martin and 13. C. Weikel. 'Further developments and applications of the stiffness
method'. A'lalrix Melh. SlrI/CI. Analysis. AGARD 72.203-266 (1964).
6. M. W. Johnson. Jr. and R. W. McLay. 'Convergence of the finite element method in the theory of elasticity', J. uppl. Mech. 35. E. 2. 274-278 ( 1968).
7. S. W. Key, A COfll'erg"I/('t'!m't'sligatioll oflhe DireI'I SI((flless lvlelhod, Ph. D. Thesis, University of Washington. Seattle. 1966.
8. E. R. A. Oliviera, 'Theoretical foundations of the tinite element mClhod. 1111.J. Solids SlrIICI. 4. 929-952
(1968).
9. P. Tong and T. H. II. Pian. 'The convergence of Ihe finite element method in solving linear elaslicity
problems', 1111.J. Solids SlrIICI. 3, 865-879 (1967).
10. R. H. Gallagher, A Correlation Study of Melhods of ,"-latrix Slrllelllral Analysis, Pergamon Press. London.
1964.
II. G. P. llazeley, Y. K. Cheung. B. M. Irons and O. C. Zienkicwicz, 'Triangu]ar elements in plate bendingconforming and non-conforming solulions'. Proc. Matrix ,\-Il'lh. Slrllcl. Mech. TR-66-80. 547-576 (1966).