.
MÀîrÎÈ*rerlcÀ_
äÉvûË D|Á.NALYSE NUIÍÉRIQUEEl
DE'r'HÉoRrE DE rr'AppRoxurrÀTroNI,'ANALYSII NUIIEIIIQUB BT LA THÉORIE DIì
L'APPROXIMA1.ION¡,r¡ii;:i:, f6¡¡10 7j'Nd;:1ì,
l}/fir 'pp;,67-79
,. .. îhe
generalbilinea¡ prçrprqmilg problern can be formulated
as Tollows:(1)
maximize{/(x,
V):
ex+
dyf
x?'Cy},:,
subject
to linear
constrantsI t
Ì ,,, i
,whe're'A,
B,
C ârei x m, s x lt, tn x
n,-'luatriccs
resPectivelliand
a,ll, c, d, x, y
are vectorsof the
appropriatc clirnellsioll.
-, Bilinear ,programming
is a
generalizatiouof the li'ear
prograrnmiug rrrat lor thefirst
time was formu.latedin
1968by
.+.r.,r.,rrlN,rr. il]. He
gavéàr
programming rvhichihttr gar;" .
e local optimnm
itr a firiite
number remarkedthat.
oireof the
Altman'sïor
theoptimality
and. not necessary.his criterion'and
corrstructsa finite
he bilinear programrning..t' ,Arr,interesting and
comprehensive strrd.yon bilinear
prograrnrrring rs'also douein tg]-bv
v,r¡iD.,\i, ,\.I
ON BILÏNEAR PROGRAMMING
blems.
is
nseclto
establish sim- maxl-plogfammlug
pro a local and globall! .
'
\,' :.
:\.il.r by
r.
MARU9CTACI
I 1clu5-Napoca¡1.
Introduction'Ax:4, x>0
,Ry : lt, y2
o, (2)i''
(3)
i,,, ,l,If i $'ìe consider: t'he simplex
tablêau
i,iilii ,,i;
ttit .,,,, it lr, t' ,i iL; . , , -] t.r li. rr..,';,,, 1
tlren
afler a
Jordan elimination step r,r'e getthe
tableaui,
x jt'
.,uþ
,.'lu
1trhere C',
b', d', [i',
à' are formed.by
the eliments girreuin
(7) and.B'
is theif 'the pivot
elemeilt,is,an'elenrreúLof tlie:uråitrix
R li, /t) .r:i': : :ìl rri , ',ì i ,:, ,' : .'.'J Ì.ì i l:llÌiiri, , ,,i, e,,= C.f $ti i ', . i:, , i, ,¡ \\,^¡;, I0:
e
0d,
0, R:
0
q.
p b a
OÇ¡
ON BILINBAR PNOERÀITUI¡¡C 69
f:
elirni-
c/"
Ít
0 (ì
0 (8)
ch o11e
Ð' 0
(:
þ' B'
d'
AO
xJt
rl--j.
rvhl daTI
ic:
iltddåtio¡sø|, rule
::t: :.' ,i
, ,1,,,1i,. ,
i
L -1i
U:+i
¡í
'Ll's
-
¡tIL I,
Now r,ve shall specify
the
characteristicsof a
Jordaneliminalion
ste¡:ln a bilineai
programming.Thus,
let us
consider(4) ,f(x, y) :ä,p,*
øf äu,ro+
P+ Ðå
c¡¡.ï¡!¡,,t i. ' .1, i_
(6)
,Lt,h:Db¡¡!¡
1. þn,, h=.1,2,
. ..,
s,!:t
and let
bþq+
O bethe pivot
elêiten't,'tfiLoiafter
substituting, ;;,, ,i yo:'Ïl_,Ðijri,+ìo_ao)''
:oþc
\
i+clr.l ¡ ,,i, ,. : ,t
,1i1.,, rrr,in
(4), we obtan11¡
t '"t|: :" ir i1':
.f (x,
yi
") - Ð c{ct lD* ai*
-,i d'ot¿o*P'+'Ð ,, (Ðci¡t¡ -f
cisapI
Bi)-
iÍ
1
(5)
2
fr
(7) 6t¡
z¡,=tfiUi¡*)'#
øn:,' :ft1!' l', 2,i .t.i,'t r,rryhere,,,
i:
L,1,; r
':" ¡:i:
'. ,;r; f i:i
, fi7,''':;t
'lri :,i
.f i. :,
..,,,?0r,
,' l íi(,.'
.r, ii
ir,i l, li¡i r ii',,
'
' ifii¡l.i\t ir t.'.í',: ':
ípfqgtflmmlng
:t'4 t
! iÍ.i,t a\ii'il
¡!' i '.,til ì:,;i' :iì
!'Ê,
Jorrl4g ell¡nhntiop. ;iur r, t, i
I'j
i ,r l
: ii
li 1,.'(.i,: .i ,
'ii)
:6 oN BILINIAR:P\QGRAMMINö 7t
îo
obtain a b.f.s. we shall uset
ation stepsl deçcribed ut th-":sectiã" Z.io
simplifythe uo that A
an¿B are
ofthe full rank. ï'hen startinþ from
.
,' t.,,' \, I,r i. ¡l .\\);":! r,r'i l
and assuming (witùòut loss of
geierality) that
thepivot
fromthe
ffustr
and s column ofA
ancl B respectìvely,,
(Jordan elimination steps) we getthe tablepu i
',\^0, *i1* Þ -À;,,1tli + Q'+ nì,,ä,
cl¡x¡!¡'
a b
-c
- %r!1
f:
(1 1)
P
li1
a
0
rl 0
a1
1,11,
ol
tlteø 'Plocal
/:.1 I
-v
0
b.f
.s.
(xo,yo),
aherebil,ine øù þro gr anr'núng
P lQ
and-!,
11
i¡i i '
of
tk:e!,{(xo,
xo)>0,
xs&
Proof.
From
(11)it is
seen tha (b"Lemma 1. If
1ì
:
(¿1,0),
yo.f(x, y)
-(3)
a1
xr- jt:
(10)
"ili¡i.
,'ì¡ ":l 1 í' li î-
t.:'i
¿ ¡,,i ,lì.:,lt
rl--
0.
'':l
elements were taken
ihenafterrfsJ.s.
,',1 i\til",l',-. '' ' "'
70
r.MARuscrAc
4Remark 1.
lf
wetake a pivot
element,in ,the rm.atrixA
then,itrSteadof tableau (8)
rve considerthe
tableau'xyl
and after
a
Jordan elimination stepwith pivot
elemetrt a'¡o*
0 u'eget
thetableaU - , ¡,i. : :i:
iir, i:,.ír',,l ,ir'.
,i''r'XL,,.Zþ'..frr, Y
1; , ,.,,;, i,i t, t.t:l
3. Optinrality
e. r i . .. ", : i
A pair (r, y) = R'x R" is
calledba
1e- soltrtig¡1of the bili- ,r""r ptägtr*tiriä'g (l) -
(g)if x
andy arc
siblc soltrtionof
(2) audl3l
respectivelv.'-'
-T-h'. tottowing theorem results immediatqly,,from
thc: theory,of ilinearcr0
0d,
A
0
Í:
r_
a'
d
0 0 H c
0 0 B
0 (,"r
'; 1/'
A'
(ìfensible (see
fll)
r. If
(x(1)-(3),
þrogrønl'bil,inea.r tlt'c
"f
rra o trul'í'on' so aslcb d.:
oþtíutal'
s an' ,S tkere'.
y*
tken
trk
problenc TIIEOREI,I programming ilØng
Now the
additionalrule
consists itr.t L-
ll:
ri ¡l ¡ 't ".1
. j .
tl1:df T'.,
,,ìand
so/(xd,,yr) ;-,/(¡0, yo)< 0
f.otx,)O, li > 0 sufficièntly small,
im- plies.(i)-(íi)'
¡'' "'Ñow,'let
1xo,yo) be ä
degerrerate b.f.s.
x0:
(aI, 0), 'yo:
(b1, 0), arrdlet us d,enote
':" Io: {i
æt{1,z,...,/}Io} :o}
THËoR-rM
3.
Degeneløte,bl,s' 1xt,P) is
ø local møx'imutttof f on
Adf
øød"'o'¡úy'åf
',,
, I(i) p> 0, q ) 0',, , ,,Ì
t' 'P*oÍ..' (J) If '(*9,
yo)is a
b.f.s.thät is a locâl
mqximunl, then"f(x, v) -.f(xo,y9 <
oin a certain
neighborhoodof
(xo, yo). Consideringxi:
(n.,0, ..,,
fri,,..,0), xr- t> 0, i e I,
we have
As/(xr,
yo)-,f(xo, yo)< 0, it follorvs
thatþn>0,i e
io,i... p>
OSince
for *' î X it is
necessarythat
YheI,¡+øju<0
sirnilarlv we
get,'To""!
r, i
bloo 0,j
æJ.
.
, ' From
(14)it is
seenthat
' .f(xt,y,) -.,/(xo, yo)<
0 ,/(xt, yu) --.f (xo, yo):
ON SILINEAR. PROGRAMMING v3
tl
I
I
(i)-(ü)
botd0, ielo -þ¡t, i ê
Io!;, t, , ] ll',ir;l
rri, ' i,(ä)¿ .¿li
,( 0,
Y(,i',tj) .e:
IonX Jl; :
, :wherc
Ð Ð cl¡'x,'!i" Ð
P;x, - Ð'qi !¡ ¿
:¡*l j-s*l i-t*l j*s+¡
,f(t, y) -,/(xo,
yo) <i0,
, íri ,for
eachxrÞ
0,i: r |-1, ..., nt,
anó.!¡
Þ',0,i :
s-l
7,'.., n, saffí'
ciently
small,i.e,
(x0, yo)is
a local:rnaximumfor /in O: X x Y
where,'X:{x eRnlAx:,a, x>0J,Y={i eR'llfy:b, y>0}'
THEoRE;v-
2. Lat
1xn; Vo), xo:
(at,0),
yn:
(b1,0)
l:eI
naxd'cgemeratcb.f.s.then (xo,.yo) 1salo9øI 1!x!.xi'Ìn!,t::of,f
?l,,ni!ryao(?,,,iÍ, ,,:,::,,i,r,,r,.
(i) p70,q>0., | ;,,i;,:.r,
I(ä) ,,i'< 0,
V(i,j)æ Io x Jo, i :
" ; 'il ' :;' ll wherc,¡,¡
I = {r+
1, ...,
tn},/ - {t +
1,...,n}
Io:
{d* Ilþ,,= 0}, Ju:
{.Ju Jlq¡ *
0}.Proof. (e). Frorn
(11) rve have(13)
./(x,y) -f(xo,
yo): -Ðp,x,-L r¡!¡ * Ð Ð
cl¡r¡)'¡:
:,Ð"(D,i, t, -
Þ,)*,* ,Ð.[ Ð'1, *, - q¡)v¡+ Ð. l,Ð,
'1,ti)*,'
; 'From
(13)it is
clearthat (i)-(ii)
irnpliesthat .f(*, {) -.f(xn,
yo)<
0for
eachxr>
0,yiÞ 0
sufficierrtly small,i.e,
(x.o,(=+)
Let
(xo,¡f) be a local
maximum andI
t (12)
72
a local
maxilnum.yo) is
Ð r,l D
:t*l ti-¡*r
(14) "f(x', yi)
1(xg, yu): .'r
:;..Now,ifp>0,tI >
consider
ÉJ', àJ i
æ Io¡j
í*l,o,j
cl¡
tr, i'þ p
(ctr¡ú
-
q,,)t,lclr't
-
p,)t,Therefore
"/(x,
y) -./(xo,
Vo):,
j n.lo '
'r. rr,r¡nÚsc¡rc',r,lrl.;i , i .:_t i
,r:i , ,rlÌ j,
cI
v¡ - ,oi) n i' ,Ð,,r,1,:D*, 'f *, --
zqo)0, the¡r
fro¡n
(12),it
follon's.that
it4
( 16)
r,.,I. IMARUSCI,!|C
6 7 76
Consider
the
inequalities. :, ', i , ,, :,,1i. .. :
ON BILINIAR PROGRAMMING
I
{r = n.l
{1 = o"'
i
'i (l*)i r,
e'r4 rh.'á2. Let
(xo, yb) ba o d'eg'ùoerøia b.¡.s',If
th'bre 't,s.ø,f'¡'.0,i
=
Iq,or tlrcyci¡
blu.> O,j *
J¿, thçnfot'cucryts\,, (li Vjì
is .not'fe,asîl)tLsolutiÒn, where
' ''
., i,,',,
,,, ,i
, .i
i.;r,*u -,,(i11,
0,
. ,. r,
x,,, ,...,r 9),.,fu: t,)
0,(15) , . r, i:
\tr, :
(b1,0,
, t . ,yu, ..
. . , 0)t; 'lu i I >
O.Proof. Consid"r x""defined
in (ts¡'and let'ø|,'= 0,ti L Ïr;r.ph"ir,l,
: -ol,t <0, Vr> 0, that rleans
xuÉ X, Ort
-Ql=
Similarly it
can be provedthat
ypé Y, Vt > 0.{' ;r il 'i
'¡Now
the
sufficiencyof the
conditiorip(i)-(ü).¿,iollowS
directly,.frornI,emma'2 and
(14).\ 5. .
Global
¡r1uxir1¡rm ,
f)l ttrc lriliney
¿liogramrnirrgAssurnc
tha!
(a9, y1),xo: (a1,,0)
i.qa local maximum
of/ on l).
Iherr it follows that (í)-(ii) or' (i)- (ii),, hold. Fo/ .r': (a',
A)...,
x¡,...,0), tr :
(b1,0,..., !¡, ...i;'0)
we have¡/¡ ç
1xrj¿'t tT
.,1, I r í.
since function
/
touchits
r¡ráxirnurnirr a
b,f.s.,ít
follou,sthat
i';,¡,11i .,,
rnax r{y'(x;,,y)l(x; V),er,Xl.X y1},-/(xo,
yo);;
,ì:i,ri whereÐ,ç
,*,,,ili,
'
'.:" .t-,',.¡.¡¡l 'j rrr liì
;¡: ;, i ¡t:l
(17)
5- ',, --
î7r
xi
X1:XO
'o t']l
I
./(x',
x't -
ntilJ.{l >,gll(xr, yi) - f
(xo, yo):
0},Yt:Yn
)-\\LJj=J
".
.That !s rvhy, in
ord.erto
deter ninea rew local
ruaxirrrun, rve shallfi'd
a localmaxirnq'r
9f/ o^n_{x\XÐ x (\y'), i.".
uããingto
ilrel'iiiat
constraints (2)-(3) the following lwo:
(18) ' 'Ð ,i! , ,; 'Dnó ,'t
with this
ner'vbilinear
programrning rne
procêed.i,ililuri u¡ti1
the ptoblern becornes inconsistentlri t f i ii I I
6r, Ileppripfion
,,of the
algorithmFrorr
above u'e cönclud.ewithrthe'folloiving:
algcirithmfór he
globalrning
probleru.u
(10)find a
b.f,s.f (i)
holds tb.en go tci step S, other_osen
in a
colurnnfor
u,hlchþ¡ 10
i '¡ l,
....
s(1þ3.
'r'ests(ii) or (ii)r. rf (ii) or
(ii),, holclthe'
goto step 4.
other-tvrse,
if
, ;: ,
_:" -"
,
t)¡')
0,, (d,j) e II x I2
i$.9,two. J.s. by chosing'ilre' þivcit elelnents
and go
to
Step2. i ., i:
ii in the coluurni
aÀdj
respectively*) 'flte tc¡nts corrcsponcliug
to
x.!: I
otyf : -l
,are rrrissilg nr (17).li! ':l Ìr
'll,¡
i,''j "i
' ,1:
ôtt¡
¡L
Define lli
J/.rr
:
uri\r {t> jlf(xt,
vr)¡,fi+9,
yo):0},
wlrere
x, : jt¡ : t.
,r, "
If
there isj e./
such llnatf(xi, yt)
,f(xo, yo):0 has no
positivc solution,then
ónc täkesii : *.o. Similariy, if
-thereis i e /
suchthat
Í(x', yj) -./("0,
Yo):0
hasno positive solution thet
yf':=rrf co' ,'t
:(i)
donot hold
lthereis -
1 and-4 _ _B f 4
negativeother
two ¡.s.
and weget the
tableau-xt-i,à _!t _!a I
-16
0
-4
-12
0 0
'oì 10
I0,,
000
10000 01000 00010 00101
2 o .)
I
s-6
0000 20
00o
2 2 2 o
-16
2 4010 100 0,-o
0001
0 0
I
3 0,0
0-ll
-8
2
-0
2 2 2 I 8 0 0
I
0 0 0
I
01 10 00 00 ll 100
5
I
I
6 (,
00
4-8
öñ nr¡.1ñ.ren Þhocnal"iMiñc
Í=
t:
0t
1l
1
-t¡- lt I
-*2-'frg,--
0
I
.>
Àfte¡
otherthree
J.s.we get the
tableau''.:i-
*L- tu2-
!z': tt-
1_
2, Since we do
0
,5 -6
*l*
tul-
It- tt=
',,. ,
,ii
- qi-
xø_ ,t(a_ rçt_!e
0- 0-
JL_
0-
Steþ
4.
Deterrninexf
andyf from (16), ,í,,, 1,,r;, ,
,:,, ; ,st-eþ
5. Add the
inequalities(r7) to the initiar
constraints and got.
Step
1.,
The .algorithmis
terminatedrvh:u the
bilinear programmiogp"obl"rl
is inconsistent.
, i7. Example
îlo illustrate the
algorithrrr rve,solve the rollowing exarnple:
maximizèf(x,y): -xa- llxr- *r, I 4tr+2xry,- xtls* 6xny,*Sro1,n"
xz
-l x¿:2
'y, lyø -2
.' ::, ¡ : t;
1z * y¿:2
ix, 7
0,y¡
Þ0,
i, :j: 1,2,
3, 4.Steþ
L The initial
tableauis:
i,,-
xt-
xz-
fia- x¿- ),t -!,¿ -!d -!¿
0
1
0 0
0
0 0 0 B
1
0 0 0
0 0 0
I
0 0
tr
1
0 0
i
0 0 0 0 0
0 0 0
_()
subject to
I
I to
2 2 2 2
After a J. s. we
gott ,1,1,.:
I. MARU$CIAC
^tL
0.:
0- 0- 0-
0Q0 000
)-
xs2-t
0650
whjch sholvs
that the new
problemis
inconsistent(xr: -xu -
1<
0'foi
everyxu2
0).. Theiefois
¡0: (2,2,0,0) ;
y0: (2, 2, 0, O) is the optimal
solution and/(x0'
vo): 0'.
(Iu'iu. ,,8øbeç-tsotyøi", clwi-Naþocø
Ul A:1 t rrr a n, l{., Bilinear þrogratnnti.ng.8u11. Acarl. Polo[. Sci. Sér. Sci. Math. Astr:ofom,
. phys.,
t?l Sokir¡r'.st íjøhstat'eM.Al,'tmana,,tsilineitcoeþrograntniro' uanie". (33), 91-98 (1975)'
[3] V an <1 al, À., , Èkonomska Analiza 4, Nr. 1-2, 21-41 (1970)'
0
1 1
0 0 0
100 000 000 001 010 0 0-6 11 I 0
00 0 -11-10
oN tstl]NÉÀR pnOc;n¡løvrÑfi 79
f:
Received 30. llf. 1978.
îi
1
-8
2
1
3 I 2
-19 -25
RËITDRENCES
(Jniaersitøieo Bøbeç-BoIYai
Faculløtea de Møtematicã Str. Rogàl'niceønu' I 3400 Cluj-Naþoco
Steþ
l. After
oneJ.s.
rveget the
tableau-%r-xs-jt-!s
It:
tuÙ-
*4 - JL-
Jb
5-l 62
Sleþ 3.
Since y'n-
11,þs: l, 4¿,:4, 8s=g.it follows tlnat
(xo' yo)'xo
=
(2,2,0,0);
Vo=
(2,?,0'Q),is a local maximúm aird /(x0'
yo)! 8 -- 8'=
0'"steþ4.wehave i I 'i i r¡ i'
l@: fl-.f(xo, v9 : lt(t -3) ':
o;i'"' t),,:3
Í(*{,"yr) -,(i.,
Vo): -19t + 6ltz'\0,'tsa ; :
tO¡6,I
Similarly
we obtain'ä
=='19/6, andyiì: +co.
liìl
'I'herefore
the
inequslities (17)are ,
'Steþ
5. The
new simplex tabjeauis the
following0 8
1 a
11
10
'
xT: min
{3,19/6}:
3- L MÀRtr$clAëwhere
fü
2
7,q
tt
2
6
r ir..t ¡,i i ''ì
2
llli ' ' l' ;; tl
: +oo
/c¡
: -5,
il i (, ' ir
y1),-/(io,
yq):'-51,- F:ö,
\.' \ ¡r,:'.
r,¡ i..: ill'r¡ì)'11:i
,r$
2 I
-19 xè 3i'and"i62)
-: 19.it - ij Siuce
Í(x',
: i 1l
-frt-1h -¡t¿--!s , 1 0,Q,,Q,,,i
4.
1; 0 0,' 0
I0,09,1 ,0,0 1
A0 0 0-6 01 00
tuL-
u2
,-,tu6-
lt-
Nõ-_