QUASICONVEXITY, QUASIMONOTONICITY AND APPLICATI0NS IN VARIATI0NAL INEQUALITIES - OPTIMIZATION
NGUYEN THI HONG LINH
Trang nhan đề
Mục lục
Mở đầu
Chương1: Quasiconvexity and quasimonotonicity.
Chương2: Applications in variational inequalities and optimization.
Chương3: Conclusions and recommendations.
Tài liệu tham khảo
18 trang |
Chia sẻ: maiphuongtl | Lượt xem: 1768 | Lượt tải: 0
Bạn đang xem nội dung tài liệu Luận văn Quasiconvexity, quasimonotonicity and applicati0ns in variati0nal inequalities - Optimization, để tải tài liệu về máy bạn click vào nút DOWNLOAD ở trên
Chapter 2
APPLICATIONS IN VARIATIONAL INEQU AL-
ITIES - OPTIMIZATION
2.1. SOLUTION EXISTENCE OF QUASIMONOTONE VARIA-
TIONAL INEQUALITIES
The purposeof the sectionis to provethe existenceof solutionsof the Stam-
pachiavariationalinequalityfor a quasimonotonemultivaluedoperator.
Let X bea Banachspace,K bea nonempty,convexsubsetof X andthemulti-
valuedopertorT : K -t 2x*.
The StampachiaVariational InequalitiesProblem (SVI)
Find x E K :Vy E K,3x* E T(x) : (x*,y - x)2:00
The Minty Variational InequalitiesProblem (MVI)
Find x E K :Vy E K, Vy*E T(y) : (y*,y - x)2:o.
We denoteby S(T,K) thesetof thesolutionsof theStampachiavariationalin-
equality
x E 5(T, K) {:} or E K and Vy E K,3x* E T(x) : (x*,y - x) 2: 0
and by 5str(T, K) the set of strongsolutionsof the sameinequality
orE 5str(T,K) {:} or E K and3x*E T(x),Vy E K: (x*,y - x) 2:o.
Also,wedenotebyM(T, K) thesetofsolutionsoftheMintyvariationalinequality
x E M(T,K) {:} x E K and Vy E K,Vy* E T(y): (x*,y - x) 2:00
Finally, wecall x E K a local solutionof the Minty variationalinequalityif there
existsaneighborhoodU ofx suchthatx E M(T, KnU). WedenotebyLM(T, K)
thesetof theselocalsolutions.ClearlyM(T, K) c LM(T, K).
23
Definition 27. Given a convexsetK c X and an operatorT : K + 2x* with
nonemptyvalues,T is calleduppersign-continuousonK if, for everyx,y E K,
thefollowingimplicationholds:
(Vt E]O,1[,infx*ET(xi)(x*,y - x) ~ 0) =}SUP:r;*ET(;r)(X*,y- x) ~0,
where
Xt = (1- t)x+ty.
For example,if T is upper hemicontinuous( i.e., the restrictionof T to every
linesegmentof K is theuscwith respecto thew*-topologyin X*) thenT is
uppersign-continuous.Any strictly positivefunctionis uppersign-continuous.
Proposition 28. LetT : X + 2x* bea properlyquasimonotoneoperatorwhose
domaintheclosedandconvexsetK. If K weaklycompactthentheMintyvaria-
tionalinequalityhasa solution.
Proof.DefinethemultivaluedmappingG : K +2x*\0by
G(x):= {yE K: (x*,y- x) ~ O,Vx*E T(x)}.
For every Xl, x2, ..., xn E K and y E CO{XI,X2,...,xn}, properly quasimonotone
impliesthat y E U7~1G(Xi)' In addition, for eachx E K, G(x) is alsoweakly
compact.The well-knownKy Fan lemmaimpliesthat nxEK G(x) is a solutionof
the Minty variationalinequlity. -
Proposition 29. See[lj. LetK beanonempty,convexsubsetoftheBanachspace
X andletT : K + 2x* bequasimonotone.Then,oneof thefolowingassertions
holds
(i) T is properlyquasimonotone,
24
(ii) LM(T, K) =I- 0.
In addition,if K is weaklycompact,henLM(T, K) =I- 0 in bothcases.
Proof. Supposethat T is notproperlyquasimonotone.Then, thereexistXl, ...,XnE
K, X*; E T(X.i),i = 1,...,n, and X E CO{.TI'...,xn}suchthat
(X\, X - X'i) > 0,i = 1,...,n
By continuityof thefunctionalsx;, thereexistsaneighborhoodU ofx suchthat,
foranyy E K n U, onehas,
(X\, Y - Xi) >o.
By quasimonotonicity,
Vy*ET(y),(y*,y-x) ~O.
Thus,x E LM(T, K) sincethepreviousinequalityholdsfor everyy E K n U.
It remainsto showthat LM(T, K) =I- 0wheneverK is weaklycompactand T is
properlyquasimonotone.But undersuchassumtion,it is known Proposition 28
thatM(T, K) i=0;sinceM(T, K) c LM(T, K), it followsthatLM(T, K) i=0. .
Proposition 30. See[1].LetK beanonemptyconvexsubsetoftheBanachspace
X'
X and letT : K -* 2 beanoperator.
(i) If T is pseudomonotone,thenLM(T, K) =M(T,K).
(ii) If for everyx E K thereexistsa convexneighborhoodVx ofx andanupper
sign-continuousoperatorSx : Vxn K -* 2x' with nonempty,w*-compactvalues
satisfingSx(Y)C T(y),Vy E Vxn K, thenLM(T, K) c S(T, K).
(iii) Additionally to theassumptionsof (ii), if theoperatorsSx areconvexvalued,
thenLM(T,K) c S(T,K) =M(T,K).
25
Proof.(i) Letx beanelementofLM(T, K). Then,thereexistsaneighborhoodU
ofx suchthatx E M(T, K nU). Foranyy E K, thereexistsz = x +t(y - x),t E
]0,1[,suchthatz E K n U. Then,foranyz* E T(z),
(z*,y - z) = [(1- t)jt](z*,z - x) 2:o.
By pseudomonotonicity,
(y*,y - x) 2:0,Vy*E T(y).
Thereforex is an elementof M(T, K).
(ii) Let x beanelementof LM(T, K). Thus,thereexistsa neighborhoodU ofx
v
suchthat .TE M(Sx,K n V n U). Lety E K n Vx. SinceK n Vxis convex,there
existsy' E]x,y] forwhich[x,y']c (K n Vxn U) andthus
infuE[x,Y'jinfu*ESx(U)(u*, u - x) 2:o.
By theuppersign-continuityofSx,
SUpX*ESx(x)(x*, Y - x) 2:o.
But Sx(.T)is w*-compactandwededucethat
infYEvxnKmaxx*ESx(x)*,y - x) 2:o. (8)
whichmeanthat, for all y E Vxn K, thereexistsx* E SAx) c T(x) suchthat
(x*,y - x) 2:O.Therefor,x is anelementofS(T, K) since,usingtheconvexityof
K, onecanproveeasilythattheaboverelationholdsforanyy E K
(iii) This is a consequenceof the Sion minimaxtheoremappliedto the relation
(8). .
In particular, if T itself is uppersign-continuousand has nonemptyconvex,and
26
w*-compactvalues,then wecantakein the lemmaVx=K, Sx=T.
Remark From Proposition 29 and Proposition 30, with the assumptionK is
weaklycompactset, T is pseudomonotoneoperatorand upper -sign continuous
thenS(T, K) =I=- 0. The assumptionpseudomonotoneis strongerthan quasimono-
tone, in the later theorem, the authors provethe existenceos solution of the
Stampachiavariationalinequalityfor a quasimonotonemultivaluedoperator.
Theorem 31. See(l). LetK bea nonemptyconvexsubsetofX. Further,letT :
K -7 2x* bea quasimonotoneoperatorsuchthatthefollowingcoercivitycondition
holds
3p> 0,Vx E K n B(O,p),3y E K withIlyll< Ilxll
suchthatVx*E T(x), (x*,x - y) 2:O.
Supposethatthereexistspi >p suchthatKn13(O,pi) is nonemptyweaklycompact.
Moreover,supposethat,for everyx E K, thereexistsaconvexneighborhoodVxofx
andun uppersign-continuousoperatorS;J; : VTnK -7 2x*withnonempty,convex,
w*-compactvaluessatisfyingSx(y)c T(y),Vy E Vxn K. ThenSstr(T,K) =I=- 0.
Proof.The setK~ := K n 13(0,pi) is nonempty,convexandweaklycompact.
Accordingto PRO, LM(T, K~)=I=- 0. By Proposition30,thesetSstr(T,K~)=I=- 0.
ChooseXoE Sstr(T,K~). Then
3x~E T(x) :Vy E K~,(x~,y- xo) 2:O.
Accordingto coercivitycondition,thereexistsYoE K n B(O,pi)suchthat
Vx~E T(xo), (x*,Xo- Yo)2:0,
(if 11.1',011< pi we can takeYo= xo). It fowllowsthat
(x~,Yo- xo) = O.
27
Now,foreveryy E K, thereexistst E [0,1[suchthat (1- t)y+tyoE K~;hence,
(x~,(1- t)y+tyo- xo)~ O.
It followsthat
(x~,y - xo)~ O.
I.e.
XoE Sstr(T,K).
.
Notethat,in theTheorem31, theconditiononthecompactnessof K n 13(0,p')
is satisfiedautomaticallyif K is weaklycompactor X is reflexiveand K is closed;
the coercivity condition is also satisfiedautomaticallyif K bounded. Finally,
the conditionon the existenceof Sx is sasfiedif T itself is upper sign-continous
with nonempty,convex,w*-compactvalues.Thus, Theorem31generalizedcorre-
spondingresultsfor pseudomonotoneoperator,quasimonotoneoperatorwhereK
is assumedto containinnerpoint, denselypseudomonotoneoperator.
2.2. APPLICATIONS IN OPTIMIZATION
(See[13].)In the sequelX is a normedvectorspace. We denoteby N(C,.T)
the normal coneat x E C to a convexsubsetC of X givenby
N(C,x):= {x*E X*: Vu E C, (.T*,U- x) ::; O}.
It is thepolarconeofthe tangentconeT (C,x) whichis theclosureof ffi.+(C - x).
Definition 32. Thelowersubdifferential,or Plastriasubdifferentialof afunction
f : X -+ ffi. on a BanachspaceX at somepointx of its domaindomf := {x E
28
X : f(.r) E JR.}is theset
eJ<f(x) := {x*E X* : \:IxE Sf(x), f(x) - f(x) ~ (x*,x - x)},
whereSf (x) :=Sf (f, f (x)) is thestrictsublevelsetoff atX.
Definition 33. Thefollowingvariant,calledtheinfradifferentialor Gutierrezsub-
differential:
o~f(x) :={x*E X* : \:IxE Sf(x), f(x) - f(x) ~ (x*,x - x)},
If no pointof the levelsetLf(x) := f-l(f(X)) is a localminimizerof f, we
have(}<f(x)= o~f(x). This equalityalsoholdswhenf is radiallycontinuous
(i.e. continuousalong lines) and semistrictlyquasiconvexin the sensethat when
f(.r) < f(y) onehasf((l - t)x + ty) < f(y) foranyt E (0,1);in particular,this
equalityholdsfor convexcontinuousfunctions.
Definition 34. f is a Plastriafunctionat x if its strictsublevelsetSf (x) is
convexandsuchthat
N(Sf(x),x) =JR.+o<f(x). (9)
Definition 35. f is a Gutierrezfunctionatx if its sublevelsetSf(x) zsconvex
andsuchthat
N(Sf(x),x) =JR.+o~f(x). (10)
Sinceo<f(x) and o~f(x) areshadyin the sensethat they are stableun-
der multiplicationby any t E [1,00),relations(9) and (10)areequivalentto
N(Sf(x),x) = [O,l](}<f(x)andN(Sf(x),x) = [0,1](}~f(x) respectively.These
conditionsbeing rather stringent,it may be usefulto replacef by its extension
29
by +00outsidesomeball. Howeverwe providethreecriteria. The first onedeals
with convextransformablefunctions,an importantclassof quasiconvexfunctions.
Proposition 36. See[i3}. Supposef is aproperconvexfunctionandx E domf :=
f-1(JR) issuchthatf(x) > inff(X) andJR+(domf -x) =X. Thenf is a Gutierrez
function and a Plastria function:
N(Sr(x),x) = JR+o~f(x) =JR+of(x)=JR+o<f(x)=N(Sf(x),x).
Moregenerally,if f := hog, where9 : JR-. JRoois a convexfunction andh :
T-. JRoois an increasingfunction onsomeintervalT of JRoocontainingg(X), with
h(+00)=+00,thenf is a GutierrezfunctionandaPlastriafunctionatxprovided
f(x) >inff(X). Moreover,JR+o<f(x)=JR+o<g(x)= JR+o~g(x)= JR+o~f(x).
The secondoneis a slight improvementof previousresultsin [22],[16].It uses
the now classicalnotion of calmness:f : X -.JR is said to be calmwith ratec at
w E X if f(w) is finite and if
'ixEX f(x) - f(w) ~ -c Ilx - wll.
Such a condition is obviouslysatisfiedif f is Lipschitzian with rate c or if f is
Stepanovian(orstable)with ratecat w in thesensethat If(x) - f(w)1 :::;c Ilx - wll
for anyx EX.
Proposition 37. See[i3}. Assumethatf is radially continuouson X and calm
withratec E JR+at eachpointof thelevelsetL.r(x) :=f-l(f(x)) andthatS.r(x)
andSf (x) areconvex.Then
N(Sf(x),x)\cBx=o<f(x)\cBx,
N(S.r(x),x)\cBx=o~f(x)\cBx
30
If moreoverN (S1(x),x) -1={O},thenf is a Plastriafunctionwhileif N (Sj (x),x) -1=
{a}and a Gutierrezfunction at x.
The condition N(Sj(x),x) -1={a} (or N(S1(x),x) =I-{a}) is a rather mild
conditionwhenX is finitedimensional.However,whenX is infinitedimensional,
it mayoccurthata closedconvexsetC =I- X is suchthatN (C,x) = X* forsome
xE C.
The third criterionusesa differentiabilityassumptionand bringssomesupple-
mentto [19,Prop. 15].
Proposition 38. See{13}.Supposef is quasiconvex,differentiableat x with a
nonnul derivative.If jJ<f(x) is nonempty,thenf is a Plastriafunctionatx and
thereexistssomer 2: 1 suchthat8<f(x) = [r,oo)1'(x).If 8~f(x) is nonempty,
thenf is a Gutierrezfunctionatx andthereexistssome"82:1 suchthat8~f (x) =
["8,00)1'(x).
Thefollowingexampleshowsthatonemayhaver > 1.
Example. Let X =IR and for c <0 let f begivenby f(x) =C3for x E (-00, c),
f(x) =X3 for x 2:c. Then, for x =1,wehave8<f(x) =8~f(x) = [r,oo)with
r = max(3,1+c+C2).
Lemma 39. See{13}.Let (9i)iEI beafinite family of quasiconvexGutierrezfunc-
tions at somex E X. For i E I, letCi := 9;1((-00,0]). Assume9i is u.s.c. atx,
9i(x) = 0 for eachi E I andeither
(a) there existsomek E I and somez E Ck such that 9i(Z) < 0 for each
i E 1\{k}(Slatercondition),or
(b) Ci is closedfor eachi E I andIR+
(IJ.- IICi) = X I, whereIJ. is theiEI
31
diagonalof X I.
Then,g : =maXiEI gi is a Gutierrezfunctionatx andonehas
~+a::;g(x)=L ~+a::;gi(X).
iEI
(11)
Optimality conditions for constrained problems
In the presentsectionweconsiderthe minimizationproblem
(C) minimizef(x) subjectto x E C
wheref :X -+ ~ is a functiononthen.v.s.X andC is a convexsubsetofX.
Proposition 40. See [18]. Suppose f is an U.S.c. Plastria function at x and x is
a solution to (C) but is not a local minimizer of j. Then one has
0 E a<j(x) +N(C, x). (12)
Proof.Sincef is quasiconvexandU.s.c.,thestrictsublevelsetSj(x) is openand
convex;it is nonemptysincex is notaminimizerof j. Sincex is asolutionto (C),
this sublevelset is disjoint from C. Thus, the Hahn-Banachseparationtheorem
yieldssomec E ~ and u* in the unit sphereof X* suchthat
(u*,x-x) 2::c2::(u*,w-x) 'l/wE Sf (x),'l/xE C. (13)
Takingx = x, weseethat c :::;O.Moreover,sincex is not a localminimizerof
f, thereexistsa sequence(wn)-+ x such that Wn E Sf (x) foreachn. Therefore
c = O.Thenwehaveu* E N(Sf(x), x) =~+a<j(x) and sinceu* =J0, we can
find x* E a<f (x) and r E ~+ suchthat x* =ru*. On theotherhand,thefirst
inequalityof (13) meansthat -u* E N( C, x). Thus, x* +ru* = 0 and (12) is
satisfied. .
32
Now let us givea sufficientcondition. Observethat no assumptionis required
on f besidesfinitenessatx.
Proposition 41. See[18).Supposethatf :X ---t IRU{oo} is an arbitraryfunction
andf is finiteatx andsatisfiesrelation(12). Thenx is a solutionto (C).
Proof.Let x* E fJ<f(x) be suchthat -x* E N(C,x). Assumethat x is not a
solutionto (C): thereexistssomex E C suchthatf(x) < f(x). Thenonehas,by
thedefinitionsof [)<f (x) andN (C,x),
0> f(x) - f(x) 2::(x*,x - x),
(x*,.r- x) 2::0,
a contradiction. .
A slightsupplementto theprecedingresultscanbegiven.It dealswith strict
solutionsto (C), i.e. pointsx E C suchthat f(x) < f(x) for each.r E C\{.r}.
For thesufficientconditionweassumethatC is strictlyconvexat x in thesense
that (x*,x - x) <a foreveryx E C\ {x}andx* E N(C,x)\ {a}.Observethatif
N(C,x)\{O} is nonempty(in particularif C is a convexsubsetof a finitedimen-
sionalspace)andif C is strictlyconvexatx, thenx is anextremalpointofC (i.e.
C\ {x}is convex).
Proposition 42. See[18).Givenafunctionf :X ---t IRU {oo}finiteatx anda
subsetC ofX whichis strictlyconvexatx, thefollowingrelationimpliesthatx is
a strictsolutionto (C) or a globalminimizeroff onX :
aE [):50f(x)+N(C,x). (14)
Conversely,whenX isfinitedimensional,C is aconvexsubsetofX notreduced
to {x},x is an extremalpointof C andf is a Gutierrezfunctionatx, relation
33
(14)is necessaryin orderthatx bea strictsolutionto (C) or a globalminimizer
off onx.
Proof.Supposerelation(14)holdsandC isstrictlyconvexatx. If x isnotaglobal
minimizerof f onX thereexistssomex* E [)~f(x) suchthat -x* E N(C,x) and
x* i- O. Then, if x E C\{x} is such that f(x) :::;f(x) we have (x*,x - x) :::;
f(x) - f(x) :::;0 sincex* E [)~f(x) and (-x*,.'E-x) < 0 since-x* E N(C,x)\{O},
a contradiction.Thusx is a strictsolutionto (C).
Whenx is a strictsolutionto (C), thesetsC\ {x}andSj(x) aredisjoint.If
moreoverf is a Gutierrezfunctionatx andx is anextremalpointofC butis not
a globalminimizerof f onX, andC i- {x},thesesetsareconvexandnonempty.
Thus, whenX is finite dimensional,a separationtheoremyieldssomec E IR and
u* in the unit sphereof X* suchthat
(u*,x-x) :::::c:::::(u*,w-x) 'i/wE Sj(x), 'i/xE C\ {x}. (15)
Sincex can be arbitrarily closeto x, we havec :::;O. On the other hand, since
we cantakew = x, we havec ::::: 0, hencec = O.Thus -u* E N(C,x) and
u* E N(Sr(x), x) =IR+[)~f(x) sincef is a Gutierrezfunctionat x. Sinceu* i- 0,
onecanfindr > 0 andx* E [)~f(x) suchthatu* = rx* and-x* E N(C,x),
so that relation(14)holds. Whenx is a globalminimizerof f on X, wehave
0 E [)~f(x)n(-N(C,x)). .
Necessary condition for the mathematical programming problem
Let us considernow the casethe constraintset C is definedby a finite family
of inequalities,so that problem (C) turns into the mathematicalprogramming
34
problem
(M) minimizef (x) subjectto x E C := {xEX: gl (x) :::;0, ..., gn(x) :::;O}.
Let us first considerthe caseof a singleconstraint.
Lemma 43. See[13J.Letx bea solutionto (M) in whichgl =... =gn =9 and
x is nota localminimizeroff. Assumethatf is a Plastriafunctionatx,that9
is U.S.c.atx anda GutierrezfunctionatX. Theng(x)= 0 andthereexistssome
y E ~+suchthat
0 E (;<f(x)+y8s'g(x).
Proof.By Proposition40,thereexistsx* E 8<f(x) suchthat -x* E N(C,x). If
g(x) < 0, since9 is U.S.c.atx, x belongsto theinteriorof C, hencex is a local
minimizerof f, andourassumptiondiscardsthatcase.Thusg(x)= 0,andsince
9 is a Gutierrezfunctionat x, wehaveN(C,x) = ~+8s'g(x).Thus thereexists
Y E ~+suchthat -x* E y8s'g(x). .
Now let us turn to the generalcase.We will usethe Lemma39.
Proposition 44. See[13J.Letx bea solutionto (M) andx is nota localmin-
imizerof f. Assumethatf is a Plastriafunctionatx, gl,...,gn are Gutierrez
functionsatx andU.S.c.atX. Let I := {i EN: gi(X)= O},Ci :=g,;,-l((-oo,0])
andassumethatoneof theassumptions(a) or (b)ofLemma39is satisfied.Then,
thereexistsomeY1,...,YnE ~+ suchthat
0 E 8<f(x)+Y18s'gl(x)+ ...+Yn8s'gn(x),
for i = 1,...,n, Yig'i(x) = O.
35
Proof. Let h := maxlsisngi,let D := h-1((-00,0]). Let I := l(x) := {i E
{1, ...,17,}: g'i(X) = h(x)}. Then, for i E {1,...,17,}\1 the point x belongsto the in-
terior of Ci := gjl(( -00,0]), sothat for anyx E C := g-I(( -00,0]) andany t >0
smallenoughwehavex+t(x- x) E D. It followsthatN(D,x) = N(C,x).By
Proposition40thereexistssomex* E a<f(x)suchthat-x* E N(D, x) = N(C,x)..
Let ussetg := maXiEIgi. Theng is u.s.c.at x andis a Gutierrezfunctionat x
by Lemma11. Then, by relation(11),thereexistYi E JR;.+,y7E aSgi(x) suchthat
-x* =YIY~+ ...+YnY~andtheresultis proved. -
Proposition 36 showsthat the precedingstatementencompassesthe classical
result for convexmathematicalprogramming.
A link with the classicalKarush, Kuhn and Tucker Theoremis delineatedin
the next statement.
Corollary 45. Seeli8}. Supposetheassumptionsof theprecedingpropositionare
satisfiedand thatf, gl, ...,gn aredifferentiableatx withnon null derivatives.Then
thereexistsomeAI, ...,An E JR;.+suchthat
f'(x) +Alg~(X)+...+Ang~(x)=0,
Aigi(x) = 0 for i = 1,...,17,.
Proof. By Proposition 38andtheprecedingresult,thereexistsomer 2:1,Yi E JR;.+
and someY7E aSgi(x) for i = 1,...,17,suchthat
r f'(x) +YIY~+... +YnY~= 0;
alsoy.:= Sig~(X)forsomeSi2:1.SettingAi=r-l S'iYi,we get the result. -
Let us give a simple sufficientconditionfor the mathematicalprogramming
problem(M).
36
Proposition 46. See[13). Ifx E C is suchthatthereexistY-iE JR+for i = 1,...,n
suchthatthefollow1:ngconditionsaresatisfied,thenx is a solutiontoproblem(M):
0 E a<f(x)+Yla:S;gl(x)+ ... +Yna:S;gn(x),
gl(x) ::;0,...,gn(x)::;0,
Ylgl(X)=0,...,Yngn(x)=O.
Proof.Supposeon thecontrarythatthereexistssomex E C suchthat f(x) <
f(x). Letx* E a<f(x), x7E Y-ia:S;g'i(X)fori = 1,...,n besuchthat
-* + -* + -* 0x YIXI ...+ Ynxn= .
Let l(x) := {i E {I, ...,n} : g.;(x)= O}.Thenfor i E l(x), by thedefinitionsof
a<f(x), a:S;gi(X)wehave,sincef(x) <f(x), g'i(X)::;0=g'i(X),
(x*,.1:- x) ::;f(x) - f(x),
(x,~,x - x) ::;gi(X)- gi(X) i = 1,...n.
Multiplying eachsideof the last inequalityby Yi and addingthe obtainedsidesof
the obtainedrelationsto the first onewe get,sinceYi =0 if i E {I, ...,n}\l(x),
n n
0 =(x*,x - x)+L Yi(x:,x - x) ::;f (x) - f (x)+L Yi (gi(x) - gi(x))
i=l i=l
::;f(x) - f(x),
acontradiction.-
37
2.3. RELATION BETWEEN THE MINTY VARIATIONAL IN-
EQUALITY AND OPTIMIZATION
Given X be a Banach space,f : K c X ~ R U {oo},cp: R ~ R andthe
operatorT :K c X ~ X*.
Weconsidertheopimizationproblem(M):
minf(x),xEK
and the Minty variatinalinequalityW.r.tcp(cp-MVI) :
Find x E K: (y*,y-x) ~ cp(llx-yll),Vy E K,Vy* E T(y).
Whencp(t)- 0,Vt E dom(cp),(cp-MVI)Minty variatinalinequalityW.r.tcpbecomes
Minty variationalinequality.
Theorem 47. LetX bea Banachspace& reliable,f is l.s.candcp-quasiconvex
andcpis continuousandholdVAE (0,1): cp(Allull)~ Acp(llull),VuE X.
If K = N whereN is open,convexneighborhoodfx orK = X thenthefollowing
statementsareequivalent:
(i) x is anoptimalsolutionof (M);
(ii) .r is a cp-Mintysolutionof &f(x) withcpo
Proof.(i) =?(ii). Let x beanoptimalsolution(M).
Assumethatx is a strictminimumof (M) i.e.Vx E K :x =I- x : f(x) > f(x).
Applying the mean valuedtheoremto segment[x,x], we get :Jc E [x,xl, Cn ~
C,cn*E &f(cn) suchthat
liminf(cn*,x- x) >0,n
liminf(cn*,C- cn)> O.n
38
With d =c+t(x - .f), weget
(cn*,d- cn)>O.
Taked=x onehas
(Cn*,x - cn) >o.
Sincef is <p-quasimonotoneimplyingthat=?8f is <p-quasimonotone.It follows
'\Ix*E 8f(x) : (x*,x - cn)2:<p(llx- cnll),
Since<pis continuous,wehave
(x*,x - c) - <p(llx- cll) 2:O.
Thereis BE (0,l]suchthat( x - x)B=x-c. So,onehas
B(x*,x - x) - <p(llx- xllB)2:0,
B(x*,x- x) 2:<p(llx- xlIB)2:B<p(llx- xii).
(X*,.T- x) 2:<p(llx- xii).
Supposethatx is nota strictminimum,setthefunctionh : K --7 R U {oo}such
that
h(x) ={ ~(x)
if x =x,
if x =I- x where a <f(x).
Sincefisc, <p-quasiconvex=?h is lsc,<p-quasiconvex
'\Ix=I-x: (x*,x - x) 2:<p(llx- xll),'\Ix*E 8h(x)=8f(x),
'\IxE K : (x*,x - x) 2:<p(lIx- xii),'\Ix*E 8f(x).
(ii) =?(i) If x is notoptimalsolution(M), wehave
3x E K : f(x) <f(x).
39
We get
3c E [x,x], Cn ---+ C,Cn* E af( cn): (cn*,d - cn) = (cn*, C - cn) +(cn*,X - x) >0,
with d = C+ t(x - x).
Since K convex, open neighborhood of x, for n so large
[x,x] c K, CnE K.
Taked = .1:,oneget
(Cn*, X - cn) >0 =? (cn*, Cn- x) <0 : a contracdition with (ii) .
REMARK. If cp- 0 i.e f quasiconvex then the solutions of (M) and the solutions
of (MVI) arethe same([10]).
40