Rockafellar RT. A Dual Approach To Solving Nonlinear Programming Problems by Unconstrained Optimization
Rockafellar RT. A Dual Approach To Solving Nonlinear Programming Problems by Unconstrained Optimization
npaar
R. Tyrrell ROCKAFELLAR *
Un etsit) oJ t1tdshi,rt.t1, Sett l., lr'. lhnqbh, U S A.
Re.e1ved,1 Jlnuil} 19 7l
Re!ised.rxnus(ipfue.eived II July l9ll
Several recent lleorirhds tur solvins nonhr.ar lrosranming triblsn\ wirh cq!aliiJ,.. ,
\lranra have Dade use ol an augrenred penrlly Lasransirn lun.lio., etrcre lemr nrvotving
squats .l thc constJain! iunclions xrc add.d t. ihe .rdinir! Li-qrxnriu li fin later, the
cotespondin8 penalty LrgFlsir ror Droblems Nith lncquality c.nsrfuhr\ is des.ribcd. ard ir\
r.lalion\hip wirh drc th.ory ol duilily is erimnred. ln rh. conver .dd. lhe noditicd dlal
probicn .onsnts .r na:iDiznrr a dif]erentilble ..n.rve lun.rion (indn.dlr de red) s!bleur
to no .onnraints ar ill It is sho$n rrrrr iny ndrnizins sequen.e lor drc dral c,n be rrde r!
\ield. in a r.nerll ear,, rn aiynptoti.dly minn.iznrg :cqucn.e lu Lhe prin $hi.h tytr.ltty
.onverres ir lers rs rlpidlr
L Introduciion
* Supporied in tarl by Lne An F.rcc Ofir.e ol S.leniiiic Rescarch nnder sranr Af ?\fOSR.
72-2269.
A duDl tpptu.ch tu sah n18 onlineat ltottatntnitt( Prahlc ls
minimize to(r) + t
i=l
r.i,/l(-\) over )r €X , (1.3)
(2.3 )
(2.4)
i ,:p t if d(x) < -.."-',/Zr.
ovcr (ri,:) € -trX Rl]. The minimization in: cin bc canied out explicit-
1,v- and thc rcsidual problem is then (:.2). This dcrivation. since it ap-
pears to do violcrcc to possiblc conrerity properties of the .onstrairt
flllrctions. hardly suggests the strong propcrtics which Z/ trtrns out to
havc for conl'cx progranmirg.
Ol colrrse. in problerns with lnixed inequality aDd equalit,v con
stroints, the eqLLations crn be erpresscr:l is pairs of iDcqualities. or terms
like those in l:1.4) can be included in (2.2J. This gener.rlizatioD is ele
mentalv. and so to kccp thc notalion simpler we do nol cilr-v it ortt
explicitly in lhe present rliscr.rssion.
meorem 3.2. For ewry r > O, the functio 4 is co cat'e and satisfies
(3.s)
Thus the .lual probtems (D,) dtl llave the same optitkal satutiol1s d d
supremutn ar lhe ordi a4 dual (Dol llorcoret lastumhtg 8l * --)19'
is ercry\hert Jilite attd contir diJ|er(,ltiable on Ra SpcJJ 'tlb '
'outi.i g,O')
y
if for a gi'ren the itlliDw t defittitlg huPPens b be altained 4l s
poiitt x (nol necessatib unique), the
Proof. Lel
c.O') < Z.(-r, f ') < ,.(}, )) + (-r' f .). !Zf (r., ))
=c.0)+0' r)v"Z,(f,.-l,) (3.15)
satisf'hg &(r,) =.g,Lr. ), while ,(r') < C,0,') lor all r'. The tn o pr.operties
imply v ,(r,) = Ig,(_r). Bur since,4 is quadratic, we have
Thus ,40 ') equals the expression on rhe dght side of (2.3).
Ir vierv ofTheorem 3.2. we shall in the rest of this paper refer simpl,v
ta dull aptimal saluti.)ns and the dual .)ptimal rialre, since these are
independent of ,'. Note that a dual optirnal solution is necessrrily a /?o,-
tEgdtire \ector, since go0.)= - iff* 0. (B,v convention, dual opti-
nlal solutions are not said to exist when the functions & are al1 iden
ticallv - ) The dual optilnal vahe is, of course. generally lcss then or
equal to the primal optinal vallre (i.e.. the infimum irr (P)). If they are
equal, we say (P) is ,o//11d1.
A Kuhn Tucket \ector for (P) relative to the Lagrangian a- is a vec-
tor t such that
-< inf Zr(jr, t) = int h (P). (3.17)
x e-Y
This condition on i is known to hold if and only ift is an optimal
SubstitLrting (3.3) into the definition ofg,, we therefore have
then for the point ; at $hich the nxximum in (3.5J is attaincd (uniqlre
bi/ strict concavity) we have also
(rl4r) 2
snQ) )z Y' <
(3.13 )
< co(t) - (1i4l) z-Y2+\t'(!'-i) for al1-1,' € Rn .
in other words.
C rb)' ) < L r(J, y' \ < Lt (-r, -r) + 0' _r). trl,,(.r, ])
=c,O) + Cr' ).-J.yrLt(x,t) (3.r5)
for all L,'€ Rn. Thjs srvs th.r ! Y/. (r,r)e aC.0.), and lherefore
!'ya.(r,r') = vg,(1) as ciairneti. '
Carcltury 3.3. T/1e fLolction g,, unless it is itlentiLatty * , satirJies
fat all )' and r'
satisfytug /?0,) = g/(-'). wiile lLr')< g.0') lor all .f '. The two properrjes
imply v ,(r) = vg,O.). But sincc ft is qLradratic, we have
Cora lfur! 3.1. I<etttir e l.) lhe Lagla$ L I t > 0, o trc hdt lhe same
giat
Kuh Tuclc$ t)e(:tors und sacl(lle poitlts Thu\ (-,r) is 4 sdcldLe poi t oJ
L,ifuudanl),iJtllcott\il1.lt),Ku/111-Tutket.ottclitionsdtesati\fied.
ii) "r'-r > 0. li(-i) < 0, ir,/,(t) = 0 for i = 1. ..., n1:
(.1i) - minilnizes J + ti!
tj 1!,1, ot'er X
is (we.lkly) compact for somc real a greatcr than the infimlLm in (P)
and some €> 0. A Kuhn Tucker vector exists if, say. there is a strictiy
feasible solurion to (P) (i.e-. the Slater condition is satisfied) and the
infinum in (P) is not -; this can be weakened when some ol the
constraints a,e lineill.
Corollary 3..1 enrphasizcs properties which Z, tbr / > 0 has in com-
ll1on wiih thc ordin.rry Lagrangian a0. Thc next reslllt. howeler. gives a
strong propcrty whose well-known abscnce fbr Zc has been a sedous
impedirnent to computational approachcs based on duality
(concavc) lunction Z.(.i, r:) over aU .r € R)4. fhis shor,ls that (i. t) is a
sacldle point of 1,,.. lrlrd conscquently. .t is alr optirnlr] solutior.
Thc reason wh! d1e characrcrizaljor ir'I-heorem 3.5 lails fot r=0
is tlut the set of poiits where r0( . r-) Llttajus its nrinilnuln Drily include
besides rhc optimal solutions to (P). vadous points \lhich are not.ven
leasible.
The infinum of the quantity 1im supr,* lf.(irr)l over all slrch
sequences is lhe asfmpt()tic aptimal rdll? in (P), and an asymptotically
feasible sequence {-yt } for which this infimurn is attained is an as}mp-
toticallt' mitittizing sequence 1br (P). As is well known, thc asynlptotic
optimal valuc in (P) equals the dual optimal value if the latter is not -,
or if there exist asyrnptoticall-\/ feasible sequences at all (e.g., [7, I 1, 191)
Thus. in prrticular, if (P) is normal and possesses feasible solutions, a
sequence {ir* i in -tr is ns,vnptoticall,v minimizing if and only if (4.5)
holds and
, gt,L,(.-rk,.vk ) vg,Ol) 2 ( ur .
rnd hence
1., ! , S,,' .,-D. .,\S.. , \It.,, tt4rr,.:
,F R,,
In view of the definition ofFo ard the remarks prcceding the theorem,
lh , _1c 1. rh"r .' i.. n.r, noro .llrrit.t rt,,i g, ,ll-ence.
The nexi result sharpens a speciai casc ofTheorem 4.1.
j/ ={l l,.Jr'+1,.i€X.
dfi(,&))
o=^. =,]ff:,,
{ Afir.\r))- r,(r) )''
(The inequalit,v ,..?(-r,l); > 0 holds at all events for all:r, t and:, due
to the convexity of the functions I.)
Thcse assumptions imply. of course, by convexity thlt - is thc
,rlqre optimal solution to {P). and thrt t (bcing the onlyK hn Tucker
vector, in view of (ii)) is the unique oplimal solution to lhe dual prob-
lems (D,). Our rim is to use the second derivative inlor ation to
analyze ihe speed of convergence to t end -t inherent in the scheme of
Theorem 4.1.
sup g. g.(-],) { e ,
(5.2)
(s.6)
\rherc tlrc matrix AO) is defirccl b!
,1(r,)rr, = D w,vf({lr )) . (s.7)
to solve the equation v,Z.(-1, .r) = 0 for :f in terms ofl. at lcast locally
around}'-. Denoting the solution by tLl). we have
(5.10)
Lrrrtu).r)=0.
whcrc (5-5) holcls. In particular. since {O) is cortinuous ir } and
t(') =-r, the metrix v;2,(t0).-r) is positive definite for-t sLrfficienti,v
near t. and hence th; convex lunction a,(', t) attains its minimum
over -lf uniquely at {(.},). Summarizitlg then, we heve sholvn the exis-
ience of a neighborhood ol (.1, ]-) on which properties (ii), (iii) and (iv)
ho1d. We can choosc this neighborhood to imply (i) and to be for lhe
tbrm LrX f. where U is a compact convex neighborhood of --.
Now since tr,( , ),) is e continuous function attailling its mininluln
vallre g,(t) uniquely at t, therc is an a > 0 stLch thei
B-\' the continuity of Z. and g/. the same inequalily must hold if -t is
replaced by any l,€ trlo, wherc I/0 is a certain neighborhood oft in [/
The convcxity of l.(I,l)
in -r thcn lmplics that
vg"(_r) = vrl,(tLr), ]) .
( s.13 )
dl,x^-€()l)12<d| (s.t5)
dt-r& - €Crr)12 < qlc.(t) -c.0)l < c2q r-k -r12 (5.20)
and consequently
"rt least as rapid" rs the confergence ol {-rk }, provided only that the
numbers ar clecreasc fast enough- Thus, whatever algodthm one applies
to generate n naximizing seqrtence for the dLral proble (D,), one ca]l
hopc to gcnerate collespondingly good convergence toward a solution
of (P). Natur Lly, this docs not e\c1ude thc possibility thal. by generat-
ing {-r't} and {xl} by a special mcthod, elen sharper convergcnce
properties ol {.rl I might be guaranteed than would fo1low jLrst from ihc
properties ol it[].
We emphasize rgeinthat these resulls carry over in the obvious way
if (affinc) equality const rints arc explicitly introduced into the mode1.
References
[1] K.J, Aror r.d R.NI Solov, (;rdi.ni mefiods toi.o.sl[ined D,rnra. wilh v.akened
asshlrionl', ini Studies n] lineat dnd nanlinear PfaErdtnmins, Lds K. AtroN, L Hul
n'icz rnd H. Uzawa (Stantord Univ. Pre$. Stxnturd, Calil.. 1951).
tll K.J. Axow, li.J. Could xnd S.V. ilowe, A gcnoal saddle poinr resull tur connrain.d
oplimizarion". LNilLite ol Stulistics IInneo S.ries No. 7?il. D€pl nl Statislics. Univ ol
NorLn Carolina, Chapel Hill, N,C., (1911)
[]l A.V. I iacco dd
G.P. Niocormi.*, ,\'o,lucat Pro.lantnnlg: Se.lr.ntial Utlcansttuined
Mi"ihi2atb Tcch iqtes (.\\\lcy. NerYork, 1968).
l4l R. Flercler. A .lass ol lor no.llnear ptogrrnninr ril! lemrinlrion and con
'ncrhods
vergence propeitiei', in: hn.s* dnd na linear ptusanttins, Ed. J, Abdie lNord
H.ll!nd, Ansterdxn, i970) Fp. 151 175
{51 R. llelchu and Shnley ,\. Lill. A cld$ ol n€ihods loJ nonLinear prcsranmnrg iliConr
purarional .xpcricnce", in N)nlineo Prc!ruttnnq, Eds. J.B. Ros€n, O.L. Nlangxsarian
and K. Ritter {Arldemic Press. New YoJk. 1971I !p.67-91.
l6l E.G, Gotshrein. flr€ lrerry.f DudliO, ii Mafietnatical PraErunmins ani tsApplicdtio t,
Nrukd (1971) (in Ru$ian).
I I a.C. G.rshtein, Irc,/) o/ Catuet Prcgtuhhitls, Al\lS T rarslatior S*C s 19 7
t ( 2)
[8I P,C. Hlarho1.] and ].D. aLrys, a ncnmethod lot the oprnrtation ol n nonhrear tinclion
sublect to nonrined .onslrainls" , Conputt .lautu l t3 (1914) 174 18,1
[9] M.R, Hestenes, Multiplier and cradient delhodf', nrt Canpttikg dethads in optnniza'
tianprcbletns 2.l,as. L A. Zxdeh. L.1l. Neunadl. A.V. Brhrrhhrd (A.adcmlclre$,
Nes Yo.t, 1969) pp. 143 164.
[] Ol II.R. H*tenec, ruLnriplicr and Eadie.t neinodJ . n)unal of Optihizdtion fheary anl
Applicdt ians 4 l.1969l 3I)3 320.
[11]J.L. J.l,v and P.J Laurcnt, 'StabiliLy dd dnalily in coneex minmizationprobleml',
Rewe F/an1aie d hllbtnati')re et de Raherche Oparutianer€ R 2 (19?i)3 42.
ll2) P.r. Lulenl,Apptaxi ation et Optilnietian l.Hetn^t1n. Pdis, 19? 2).
I L ] I A. Miele. !. E. CragC, R. R. lvcr drd ,^.V. Levy, "Use ol lle augrncnled penaliy l un.ti.n in
nathenaticar lr.grxnnine ptohlen[,partl ,Jowul af Optiuizatb n Thed! u.l Appli'
.,tds8(1971)ll5 130.
I I4 ] A. Miclc, LE. drd A.V. Levy, "Use ol rh€ augnrcnicd pcnaLry lunction in n athe'm
Crase
iical proCnnrnins problems" , tst 11, Jou/Ml af Optimizatio n mear! ah.l Appli.ations A
(1971) 131 153.
A.tutldpptoach to sal ns no lineat ploEtanj"i.t ptublens I7-l
[ 15 ] A. NIic]c, t.]]. I .seley and Ll. Gasg, ' A nrodiication .l rhe ncfiod of nulriptiers aor
n! tlrematioal ?rogranninE lrobGnl', nr : L.r,i4r.s r/ optihizttion, Ed. A.\ . B a-
krnhna. (Academic Pres, New York, 1972) pF. 24t 260.
It6j A Miclc,P.]]. I{oseley, A.V. I err, and c.11. CoeIjrt, Or rhe mertrod otnutijllie$ tor
mrurcmaricar prcsrimnlne p lob]l s , Jauthal ol Apiniz ati.n Thary and Apptitdtiors
i0 (1972) l 31.
llrl M.J.D. toF.ll. A nernod lor nonlinear conshainrsin minnnizadon problens,, in: Opr:
aearirn, [d. R. Fletcher (Acadcni. Prcs, New York, 1969) pl, 28] t298
[8j R.L Rockrtel]lr, Conyex luncti.ns and dualil! nr oprimizllion FrobleDs and dynan
ics . \n: Mathenatiui s!stehs theot! d11t:/ ..an.ftns, l, Eds, H.\r. Kuhn andc.l. Szeg6
(Spdnlcr, Benin,19691pp. Il? 141.
I I 9l R.T. Rockaielhr. c,,,€] ,4,"ryJa (?rinEron Lrniv. rF$, princ.ron, \.J , I 9 70).
[ 20 ] R.T. Rockalcllar, 'Ordinary c..vex DrolraD s withouL I duaLiry gap.,, Jdldt o/ Oprm!
zdtia Ther-" dn.) Applnatio,r 7 {1971) 1,13 1.18.
[21] R.T. Ro.kalellai. New appli.rtions .t d uaLitr jn .onver proej@fr ns.,. pitlen vcsion
'\r !dd te\er( ,t-,,," r"t , ,.o.r.,oni,r. ,.J p..B.Jr,,,rt
Th. Ilasue. r970, and else{tr.tetin: }ta.ee.lints althc foufth.onluen.e an prcbdbiliry
treorr, hasov, Romani!, 19tl (Edituia Acadenici Rcplbticli Soctutisie Ronalia,
Buchar$1. 197i) pp. ?3 E1.
l22l R.T.Ro.kilellai,'lhenulrlplermcrhodotHedcnesandpoi,etaDlliedroconvexpro-
rirnr E /., ,/ Ot.'--D,o. tL,o,,a,t tpph a-:o/ 2\nr ta-J).