-5
1975ARCH1EF
Nomenclature
= value of approximate
solu-tionatz =
g, G = exact, approximate
fwsc-fions (g = ce) nìiiiínitni value of g1, ja = functionals of variational
principles M1, M2, Jf = positive numbers 58/ MARCH 1971
liitrou cl io nIN the aipIualioit of tise flnite-elenwnt method [1,
2] Ito cuniplex problems il i.s often Ilsefill to make zspproxintatiolss that are outside of the I Iit'ory of varialional calculus, For
ex-ample, iii many dynamic tiialyses il. is useful to lump themass of
the system by some rule of experience. The 1(51 ut is a diagonal mass matrix that is easier tu handle thaii a colisisteuil mass matrix 13] which is coupled. iniilarly, tise stillness matrix for als
ele-ment is generally derived with certain simplifying assumptions.
The cross-sectional area, moment of inertia, material, etc., usually are assumed to be constants over ais element. These
approxima-Numbers in brackets de,.ig,iatc Ileferences at end ofpaper.
Contributed by the Applied i\Ic,lsnnks Division and presented at
the Winter Annual Meeting, New York, N. Y., November 29-December 3, 1970, of THE AMERICAN SOCIETY OF MECHANICAL
ENGINEERS.
Discussion of this paper should be addressed to the Editorial De-partment, ASME, Tnjte Engineering ('enter, 345 East 47th Street, New York, N. Y. 101)17, and will he accepted until April 20, 1971.
Disusiun received after the closirlg date will be returned. Manu-script received by ASME Applied Mechanics Division, October 20,
lt-liti; final revision, January 30, 1970 Paper No. 70-WA/APM-34.
On Certain Approximations in the
Finite-Elemeiit Method
Approximations made outsideof the variational principle used in the finite-element
iiiet hod are examined for a restricted problem in elasticity. They are s/town to be riorous
front the standpointofa generalized variational principle in which Lagrange multipliers are utilized. The convergence oflite Ritz solution to 1h e exact solution is demonstrated. The bound is shown to he a functionofthe qualityofboth the displacement functions and oilier approximate f unclions in tite a nalysis.
Lab.
y.
Scheepsbouwkunde
Technische Hogeschool
Deift
= extremal function of I
(exact solution)
ii0 = extremal function of I i5f
= first variations of
fune-tionals
Xg À = Lagrange multiplier
func-tions
Transactions o the ASME
at ('
ssu
di
X
X r O. XrL
Fig. i(a) Boundary-value problem
110115 are desirable 5juni I mvmeducc tite complexity of derivations luid simplify coding and (lcbugging of the digital programs. From
experience we know that they IISU:llly du not affect the quality of
the approximate solution. However, there still remain several questions. Why do we obtaiui a good approximate solutiotI to it
coulIplex problem when we usiake approximations outside nf the theory of variational calculus? how do we obtain the sollutioml? What are tise error relatioisships betweefl the approximate solis-lions and the exact solution to the problem? This work attempts
to partially answer these questions by examining a restricted
problem in the light of the approximation theory. An Investigation
We examine here the problem of tise rod loaded by a body force
q shown in Fig.
i(a), l'bis probleun is the simplest mixedboundary-value problem in elasticity. We
assume that g =
ae > 0, the product of the area and elastic modulus, is apiece-n = piece-number of elemepiece-nts ipiece-n
finite-element mesh
q, Q = exact, approximate
body-force fwsctions
U = approxi mate displacement
function
= Ritz function
R. W. MCLAY
Associate Professor, Department of Mechanical Engineering, University of Vermont,
X:O X Fig. 1(6) Example function g
X:O. )t=L
Fig. 1(c) Example function q
e diitreri i al >Je fi r ti ni ro id ib at q is a piecewise continuous
Irr. F;XilIii)i finiti iiiris for q, q are shown in Figs. i (b, e).
fl
t\ati soliti loi i to titeproblem posed in Fig. I will sal isfy i Li erer ii irtiequationd
f
ii\
-
(g
J+
q = Odx\
(LV,at p ii Is at which (dg/iI.r),q are coin ri nous arai t he t ranisit ion
11115
=
itt
iris ¿ where either or both (dg/dx), q
arediscorithuions.Tri zìdlitioii
lii irder to obtain roiripprcxi ni atesol uit ji ut tot his prinbleni lu'
using the flruite-elenieiit rnethuiil we make lise of the following f unit ronrui
11
'il
= 2j>
g Uix) dx
-
qUdx.J(i
whiri. U isan apptoximritefunctionsatisfying theessential
eon-(ijtioiiï
I
UC()
wilinini ihs elements,-, U
= at element boundaries, (5) 3UrO
= O. If we choose U asU =
a+
(asi - a) (
J (6)-Witiuiuia fInite element where a0 a+ are vali
lesof the approximate
solitI ioni at, the left arid right elenienitboundaries x0
respec-tively, it cari be shown[4-Ji] t hat the
flitz
solution obtained with (J arid the funìctionalof equation (4) coiiverge to the exactsoin-lion , on refinementof the finite-element mesh.
Let us
desig-nate the flitz soltition as U11. The sense of convergencet>f itSfollows:
fL(du
j)2dx <
!',
JUR - uJ
(L2)h/2
(7)Where n is the nuniiruirer ofcl>rireiiis in lite analysis and21f1,M2are
Journal of Applied Mechanics
(4)
positive ritniniliers. 1(luiit tue ?l>vel?l?lirenut if eciurrtjoni (7)is (nIy luruilrin' to tine work >f i efenrnii loi i t will riot, he prescrit ed here.
We i uext ileveli ilIi nie i iiI eint, rif this work, a t,u1(1g of
ap-poixi nrii inrisheytuiiil I lun ii illoweuli,v the essential corel t ionsof
t Ii> originalfinitiii tuai lu tii irtt.ioit (4). Suppose that itis
meurt-veu uienit to evntl ini le tine ini (grals ofequation (4) di IC lo i he
corn-l>leity of generati rig t lii fiiii ilions g, q. We crini sim Ill fy the
problem byapprox juunaiirg these functions. Titis eroi he ji isti fled
by the tise of Lagrriruge niuiil i pIers. l'ue funnuctiottal of equation (4) then becomes
i
1"
IdU\
¡'L= -
2J0
IG ( -
\dx/
I(Ix -
I QUdxJ0
+
j
iXg(G - g) + À5(Q - q)Jdx, (S) where G,Qare fuiictionuii ripproxirurit inng g, q, respectively, arid XQ À5 aie Lagrange muli iplier functions imposi rig approximateequality of the fiiiretiouu l)airs. Note that this approach brings
the approximation of the functions g, q within the variational principle. We can further extend nur discussion of equation (S)
h iimiting (lie variations of (lie approximate functions G, Qsuch that the third integral is zero. On doiuug this I becomes
1
fL
G
(?!)2
i/xL
QUdx. (9)
-
dxItenauise l is not t lu?' annre J the extreunril funictiout u0 of I
will sat isfy a differeni t l in lcr(II? ial ii uni nu rial ural troriditionis
than
tliuir.i' 1 '1'}us linI,i >Iluutit >tiof l(; 15
nl /
diu1\-- ( (i
- - J+
(J =Lr
nix,
'a-i t bui t lie elenueints Iul k, t he natural conudiiioiis are il,u G----"
=G---'
dx x= dx (1111G--'1
dx=0.
aI'he theory of Lrigr:uuige nuuuil t pliers shows us that the
ripproxi-inritiouis in eqnialioru (S) (arid (9) ris well) are correct. Yet, if u9 and 0G satisfy ditfereiitt Euler equrrtiu)uas and natural conditions
t hey cannot he the same furìctiont. I luw then do we obtain a soiu-tioni to ihe original prohulinui?
The answer to this qnuist unni Iic irr the study of convergence of
tire solutions, which will mow be developed.
This study will
show that the flitz soluttinir Uz associated with Io approaches 0G while u0 approaches the exact solution u, both approaches being uniform.
Tire first derivatives itt each ease approach in the
ineaii.
Let us make the essential condii ions of J anni the suinte, =
uJ0 = o.
Then u0 is admissible to 'u riunii, sinarilruiiv, u is admissible to It;. Oir substitution anid by nirikiuig use of tire fact that älg= = O for ali admissible futtictions it is possible to show that
L fn/ui0 2
1(u) - 1(in) = f g
-
dx 0,Ï0(u) - IG(ilG)
=f G
(n/uu dx 0. 2 dx dxMARCH 1971 / 59
X Xi (IO) dU g__i? dx X = - (2)If we now further i'it riet ourselves to approximate functions
G > (I, we have a minimum ninuciple. It is interesting that the ftiincliouirul I; of equation I)) is ri nuinuimuimn principle aiuti yet is
i nrplied I y thefuinuctin >tutul 1,; tf equi at ioni (M) which
is riot it mi
ni-mii ni priru ripie. Ni ile :tIu u th uit le of equal inri (9)15 not tItesame
nnuiuuniiuinn principle as I, uf equal 'iii (4).
= O,
('iii?
= t). (3)
TItt' basic teqiinlit tes of eilt loti (13) are a st itt erneut of the
utili nomi pri itt-i pies, t lie i t it egut ts bei tig tite sou it id variations of
thet'e--.lie(-t ive fitmietionals.
We next add tltee
two inequalities
ttt obt titi
J9(u1) - ¡12(u0) -4- 1(ti,)
t0(u)
i
(due
(lU'\2 L (due(/ti\2
= -
tii-
-
iu'.r+
2\ dx
dxj
2\
dx dx L¡j
/ì\2
g ii 0G I (IX > 0. 2\dz
dxj
We expa t al arid rea rialigeCIII 1:1 t ion(14)to obtai ti
j
(g - (;)
(duo)
dxjL
(q -
Q)u0dx 2+
¡ 1L
(( - g)
(dy
/1L
- q)udx
I Lg (du0 dx 0. (13)2J0
\dx
(IX/Next, we rearrange equation (15),
i
("
/du02de 2\
rL
-
(gG)(
Jdx
(qQ)(u0--u)dx
2\
dxdx /
¡-'L 1d h, ¡
f_2_Li()
dx0. (16)
2J0
\dx
dx,
We develop two basic inequalities by making tise of equations (3) and (3).
lue -
nul =IftItiQ
I----
\dx
(Ix/
/u\
-Jdx
< L" [
çL
(duo,/,,\2
dxl
-
L.J\dx
il.r/
Jr
rLld1i0
du \2
JXI
Ij____P)
dx[J0 \dx
dx/
-
f1.
(q -
Q)(u0 - u0)dx<[fL
1/2 L (q Q)2dx][f (u0 -
uQ)2dx]-
o (17)Also, the right-hat id ntot iii ter of eq u a t lt t t (16)
can be bouinidej
below with the knowledge that.g miu> (i. Thins
i
('
(du0
du\2
gv--- I
(IX 2\dx
dx/
I / J ( (/(1t(/ (.0 0. (21)2J
\dx
dx/
On combining eulliat jolis (1G),( 19), (20), and (21), we obtain
I
'
(due
du)2
dx ì(due;du)2
tlx2 j0
\dx
dx2 j0
dx dxf(g - G)
(--
dx1L
(q Q)(u0
-.[f1_
G)2+Q)2dx]1/2
r r (du0
du'\-Xl
I1 --
I dx[J0 \dx
cLr/+ [f
L(q -
Q)2d1/
[f (u0 -
tio)2dx]'.
(22)Next, we consider outiv (lu-
left and right members of
t (tein-equality of eqtration (22) and niake use of equation (18) to obtain tif
fu (du-
-Adx
lit' 2\
dx drJ .[
f (g - G)2
(pia
+
)2
(1.r][r-('
(/u)2
th Er
i'
r
r1
¡du2, a0\2
11/.+ '
Li
t' - Q i dx]
[j
-
)
(IX]rL
r
rL
la
du \ 2 o o (IxJ
(U -
u0)2dx L \ dxq ) dx dxdx,
lu getterai, tite fiuuirtittisit(;( u satisfy tlitfereut ditTen-ittial
. /
2 eqiiat lotis; thus the expu-e-ittui of tue mean difference uf t lie< ji
(-°
i
dx. (13) derivatives in equation (2) isnot zero ioud we can divido t hiough
\ dx
dx /
by it.
Also, since n0 atidti sat i'fv (lie respective clifferctttia!
/
(hiC(Ill \
2By returtiitig to equation (16), rearranging tite ieft-hnd lute- equations we know that
+
y")
is bounded above by 2If.grais, unid niakiitg tise of tite Se.hwartz-Iltiitiakovsky intequalit.y
liv using these factswe obtain tite firmai iiiequalit y tif t his
in-we obtain the following expressions:
vest igation,
11L
(g - G)
(2
-
iix[1L(th10
du)2
d.r]
<[.i13
1L
(gG)2dÏ]'
=f (g - G) (
+
(
-
) dx
+
2Ir
1L(q Q)2dx]/2 (24) <i[
r
(g - G)2 (
+
ax2LJo
dx dxg, [j0
(23)The inequality f equation
(24) elemoiistrates the meant cou-vergence of to as the fuinictionis G, Q approach g,
q, r
dx dx
(19) spectively.
By equation (17) we cani show uittifornt convergence of the displacements.
Discussion
Two fitta! expressiu)ns at-e needed to discuss the quest loti raised
(20) in the Introduction. These are obttiuted by ulsitug the triangle inequalities
ELf'
[j
i-J /
MARCH 1971
liarisactious of the ASME
<
1dudug
dx-
(ix dxdx
I- ti
(,r, -
1101 -fiI; - u,l,
r ¡'
f(l(TR duQ 2 /Jj.!.
du\
'/2Li0
!,,dx -
ej;)
IIX][
¡- -i;)
dx]
L (idu,\2
2/3+[Í
dxI dx
dx/
]
(25)13v contlittting equations(7)atol (24) with (25), we obtain
jUR
uI
()hí2
+
[(.
j.L
(g - G)2dx )'/z+ 2L
(f (q -
Q)2i.r)]
wli it'h t bese ilpprt Ix iitt:tt ions are pari i cri larly mua ,rtttttt is shell
liti tlysis using vu ri ti jot ial pri n ripies. 'l'i re fu net h u tais associated
wit h shell problems are very complex, rer1ltirittg sittiplihittition of tine geometrical expressions itt t,rder to lie feasible for
program-ming.
The author thanks tite reviewer for suggest iltg that he indicate t lie existence of a ttrtrnber of pitblished papers ott tite subject of
hitnite-element couvergetice. Ile also expresses appreciatiott for
tite prompt reply oit the review. Acknowledgment
This work was supported by NASA N(ffi 46-001-008 S2 68-16. Tite author wiiìes to thank P. E. Grafton of the Boeing Company Aerospace Group fuir suggesting the research topic.
References
i Martin, H. C., Introduction to Matrix .tfethods of Strnct,iral
Analysis, McGraw-Hill, New York, 1966.
2 l'rzemieziietki, J. S., Theory f ,tfati'ix Striietiii'al Analyitt, McGraw-Hill, New York, 196S.
3 Guyan, R. J., 'Redurtiot, of Stillness and Mass Matrices," AJAA Journal, Vol. 3, No. 2, Feb. 11)65 j , 380.
4 MeLay, R. WC., 'An Investigatioti Lutto the Titeory of tite
Dis-liittiernent Method of Analysis for I.iue'rr Elasticity," Ihi) thesis, Department of Etigiticering Mer1itntic... Uutiversitv of "uVisconsin,
M:rdison, Wis., Aug. 1963.
5 Key, S. M., ('rnttvergeiuce Ittvc'stigation of the I)irect Stiff-ncr_Sn Methoil," PhD e-i., [)ettriiuietii of Aeronautics nuit
Astro-1trirttít'-, University of \Vashittgtoii, Serti Iv, \Vttsh., iInir. 1966.
Tong, I'., :triul l'ian, T. IL. IL., ''Tie ('onvergence of the
Finite-1']etiieri t Method itt Sitiviitg Li trt':t t 11 ;n r ir I 'rol ,lenis, '' fnternaiional Joirrno/ of Solûls and S!rnrlures, 'ol. 3, ltui7, pp. 865--879.
7 Birkoff, O., $n'}tultz, M. LI., iii Varga, R. S., 'Piecewise
lierntite interi,ol:iti,ri it, Otte and Two \':ti'iabies With Aptiications to I'artial Differenti:,! Eu:ttiotto," ,\'nim ,ische Mathemat'i/r, Voi. 1),
1968, p. 232 -256.
S Oliveira, E. it. A., ''Tlteoretic:tl i"oundations of tite Finite-Flettiert Met hod,'' I',tri',al ¡0/0/ .100 d i of .Soiitis and ,S'lr' wi ri's,
4, 1968, pp. 112U t t52.
9 Ii ulnte, B. L.. ''irrt ert ii: ti titi iv hit z A pproxi nation.'' Jour-nalo.fMat/rernatics a,'!.lJ,'e/nar,je.', Vol LS. 1968, pit. 337-342.
It) Johnson, M. \\'., J,'., 'riti MeLay, It. W'., "Convergence of the
Firnite-1'1ettient Metl>d jut lit, l'iteory f Elasticity,'' JOUIINAL OF
Ai't't.ii, MECHANICS, \'i,l. 35,TItANS.A*\EE, Vii. 90, Series E., No.
2, Joue 1968, iii. 274 278.
Ii Fix, G., and St rang, (t., ''Fourier Artal','sis of tite
Fittite-Ele-meint Methods in ltiiz-(i:tlerkin 'I'heory,'' .S/r,dies in App/ted Math-e-no/leS, \'ol. 48, No. 3. 191W, trI,. 26,'- 273.
r r'-
(dURdu\
1''
fj[\l/2
I I I
- I
(lXiI
-LJo
\dx
dx/
J\2
Journal of Applied Mechanics
MARCH 1971 /61
+
L [eu
j.L
(g - G)hdx
+ 2L
(1L(q
-
Q)2dx)']
(26)
Equation (26) illustrates the uniform convergence of the Ritz
soluti, it U, to the exact solution Q and mean convergence of the first dcrivativei. Note that the qttili1 y tif tite solution depends on bot it the ¡tiimber of elemen i s it t e t tush n and on t he q tt;ilit y
of the approximate fioul jotis G tii,l ÇL We also see I hat tito funet i ut dGjdx need not converge t vEx. This property itf
i lie va 'ial i tital principle is used whet i we idealize each elemet ti in an antilysis t.o be constant in t lie ervs-setI it liai area and Young's
moduli s. In this case dG'dx = I) in each element.
Equal ion (26) suggest s t hut a cit isit citi formulation reip tires the qitility of the approximations of ('.', Q to lie uf theMil1i( order
as thai uf the straiti appriximul ivttt. l'itt siniplv, if strains :ite
approximated as piecewise ci;nt titi funv'i iuv t hen it is sufficient that G,Q also be piecewise cotisi uit finid ions.
Thi si udy of a rest ii'ied prtui luto sii uws t hat many approxi-niatiov rs made ott using variai ii teil pri i cip us are rig( rous, ai-thotigi tt hey do ifltl appear sit si tltel'li tiny. The knowledge of
titis is of practical import tinco to the analyst ; lie cati go so far as
to define fuori ions numet'itally in i he eqttatiotts of the Rit z