REVUE D'ANALYSE NUMÉRIQUE ET DE THÉORIE DE L'APPR,OXIMATION Tome 29, No 2, 2000, pp. 119-122
A NEW CONVERGENCE THEOREM FOR THE STEFFENSEN
METHOD IN BANACH
SPACEAND APPLICATIONS
IOANNIS K. ARGYROS
Abstract. We approximate a locally unique solution of a nonlinear operator eqlation in a Banach space using the Stefiensen method. A new semilocal convergence theorem is provided usi[g Lipschitz conditions on the second F]échet-derivative of the operator involved.
Earlier results have used Lipschitz conditions only on the first divided difference. This way our conditions are different from earlie¡ ones. Hence, they have theoretical and practical value.
A
numerical example is also provided to show that our results apply to solve a nonlinear equation, where earlier ones fail.Key word,s and, phrases: Banach space, Steffensen method, divided difference, F!échet- derivative, Newton-Kantorovich hypothesis.
1.
INTRODUCTION_ .J" this stud¡
lve are concernedwith the
probremof
approximating alocally unique solution
z*
of the equation(1) F(ø) :s,
where
tr'is
a twice FYéchet-differentiable operator defined on a convex subsetD
of. a Banach space .Ewith
values on itself.'We use the Steffensen method
(2) rntr : rn - lr,,B (r,) ; Fl-t F @,) (ro e D) (" ì
0)to
generate a sequence{*"} (, )
0) convergingto r*.
Here,[r,a;F]
d.enotesa divided difference of order one
at
the pointsr,u € D,
which is an elementof L(8,.8), the
space of bounded linear operatorsfrom
-Ðinto itself. B
is a continuous operator defined onD with
vâluesin .8,
usually relatedto f'
byB(")--r-F(") ("e D).
Sufficient convergence conditions
for the
Steffensen method were givenin
[1], [2], L4l,[5]'
[6], [9]-[11] usingmainly
Lipschirz-rype condirions on the AMS subject classification: 65805, 47HLT, 4gDt5L20 Ioannis K. Arsvros
first
divided differenceof F. In our
study, we useLipschitz
conditions on the second tr'réchet-derivativeto
obtain a u.ew semilocal convergence theoremfor the
Steffensenmethod. l'his
way, our convergence conditions differ from earlier ones. Therefore our results have theoretical as well as practical value.Finally, we complete this study by providing a numerical example to show
that
the Steffensen methodstarting
from aninitial
guess :ú0 convergesto r*,
whereas the same is
not
guaranteed by existing conditions [1], [2], [4], [5], [6], lsl-[11]2.
CONVERGENCE ANALYSISFrom now on we
will
setA(r) :
[r,B (r)
;F] r € D,
andfurther
denoteA("")
byA,
(rr.>
0)for
simplicity.Let
a, b, c, d,.R be given nonnegative constantswith
ce
[0, 1), rs e D
such
that Al' :
A(ro)-t € L (8, E), f
be an increasing real function, wh.ich is continuous and nonvanishing on [0,.R]. Define the polynomialp
byThat
completes the proof of Lemmal. ¡
Lpvtua 2.
Assu,me condi,ti,on(7)
holds, and(e)
.f(t) # 0, /(ú) <
p,(t)for
arr úe
[0, 11] .Then
iteration {t,r} (" 2 0)
giuenby (10) is
monoton,ically i,ncreasing anrl conuergesto
1.Proof. Define function g by
(10) su):t e(t)
r (t) Then by differentiating function g we get
(11)
s'(t): f (t)U(t) - p'(t))+ l'(tþg-.
r
þ)2It
followsfrom the proof of
Lemmal,
(B) and(g) that p,(t) <
0,p(f) >
0,f
(t)_< 0_, andl'(t) >
0for all
ú€ [0,r1].
Hence,by
(11) functiong
increases on[0,r1].
So,if t¡
e. [0,11] for some Æ, hen:ttc+7,andf¿-,.1 (i ,r- p(tk) - p(rt)
lftù:rr- r1¡:r,
A Theorem for the Steffensen Method T2T
p(tn)
I (t*)
That
completes the proof of Lemma2. !
Remark, 7.
It
can easi,ly be seen by (11) that cond,'ition(g)
can be replaced,by the ueaker
f (t) + 0, Í (t)(f (t) - p,(t)) + Í'(t)p(t) t
0for
aut e
[0,"r]
.We can nov/ prove the semilocal convergence theorem
for
Steffensen me-thod
(2).TnpoRprvr
7. Let F : D c E -+ E
bea
twice Fréchet-rtifferentíaile oyterator. Assume:(a)
there eristsr0 e D
suchthat Aot : A(*o)-r e L(E,E);
(b)
for allr e U(*o,A) : {reElllr-*oll <Ã}
theree,ist
constantsu,b,c
such that(12) Il¿1,
@,,(*) _ p,,("0))ll (
ollr _
"oll ,
(13) lltlrp,,("0)ll (
ó,(14) ll11', @' (") _ A("))ll <
";
(c)
condi,tions(7) and (9) are
sat,isfi,ed,for
some cont,inuous monoton,ically 'increas'ing and nonuanishi,ng functionf
on 10,11] such that(15) llt;'çq@)-
Ao)ll< f fll"-roll +1) < 1, f oratr reu(ro,rt)
2 ò
(3)
the constants
a,
B by(4)
p(t) :
lot' +Iur' - (r -
c)t+
d,,2(1-
c)b+Jæ+2a(1-c)'
t¡1t¡-
(r.:
(5) þ: (L- c)a-loo'-ï0c,,
and the
iteration {t"} (" l0),
by(6) ! ! p(t")
tn-tr : ," - Vö,
to:0
(n>
0) . We need the lemmas:LnttvlR 7.
The real polgnomi,al p has two positiue zerosrr,
12 with 11{
12and
a
negatiue zero-r3 ("¡ > 0) if
and only i,f(7) d<p.
Proof. Polynomial
p
hasa
negative zero -'t'3) sincep(0) - d > 0,
andp(t) < 0asú -+--oo.
Moreover,p'(0): -(1 -q) < 0,andp'(ú) ) 0as
ú -+
+oo.
Hence, there existsazero
of ptin (0,-),
which by the form ofp
is givenby
(4). Thus,p
has two positive zerosif
and onlyif
(B) p(a) <0,
which is equivalent
to
condition (7).t22 and (16)
(d)
the followi,ns hold:(17) (18)
(
ie)
Ioannis K. Arsyros
a
< lla;1r ("0)ll
;c
e
[0,1),rz(Ã,
U
(ro,R) ç n,
llr"+t-r"ll(ún+1
-úr¿llr"- r-ll < rt-tn
A Theorem for the Steffensen Method
4 5 r23
where 11 o,nd, T'2 are the positi,ue zeros
of
equati,onp(t) :0,
and polynomi,al p is giuen bA@.
Then, the Steffensen method,
{r"} (n )
0) generated by (2) i,s wetl defined, rema'ins i,nU (r.6,11)for alln)0,
and conaerges to asolut'ionr* e
U(rs,r1) of
equati,on-F(z) : g. If rt < rz the
solutionr*
'is un,iquein
U(ro,rz),
uhereas
if 11:
12,r*
is un'iquein
U(16,r).
Moreoaer, the following error bounds hold
for
aIIn )
0(20) and (21)
Proof
.
We first show linear operator ,4(r)
is invertible for allr e
U (ro,a),
wherea
is given bV(a). It
followsfrom
(15) the Banach lemma on invertible operators [4], [B], the estimatellz';, çq(") - Áo)ll < ¡ fl" - "oll) +
1<
1,that
A(")-t e L
(Ez,.E1) and(22) ll, f"l-'r,ll . -r 11"-,olt)-' < -/ (o)-'.
We must show
that
estimate (20) holdsfor all n > 0. First,
notethat
11 is defined,andbyusing
(2), (6), and (16)weget llrl-"oll : ll-{t¡("0)ll <
d,
: tt -
ú¡,which
shows (20)for n : 0. It
followsfrom
(22)that
linear operatorA(*t)-t e L(E,-E),
and hence12 cal
then be definedby (2).
Letr e lrs,r1]: {r:r:\rL +
(1-À)"0,
0<
À( 1}. By
Taylor's formula [3], l4], [7]for
a twice Fréchet-differentiable operator G onD,
we can write(23)
G(*) :
G (*o)+
G'(rs) (* -
*o)*ï"" ("0) (, - ,o)' *
+
j ,""
(a)- G"(,.)l (, -
a) da.æo
Using approximation (23) for G
(r) : A,rF @) (" € D),
we can get(24) AlrF @) : Ao'F ("0) + (* -
*o)+,4tt (F,
(*o)-
Ao)+
+;Alt
1 Ftt @ù @- *r)' *
* j or' (p"
(y)- F', (ro)) (* -
a) dy.1O
Let Àd:
s, thenby
using (B), (12)-(14), (16), approximarion (24) gives(25) ll4'¡(")ll <
(1-À) d+c^d+)a*a2+
fo,r3as:p(s)
(since
ï - ro:
À(rr - ro) : -À,4;1f
(ø¡)).Moreover,
by
(2), (22) and (25), we getil*r-zril ( llar",)-',4'll .ll.4t'F (*t)ll . -ffi :tz-tr.
similarl¡
\rye can show (20) forall n ) 0.
Estimate (20) and Lemma 2 implythat
steffensen method{""}
(?_2 0) is cauchyin
a'Banach space,E, urrd u, suchit
convergesto
somer* et (r0,")
(sinceu (*o,s)
is a closed set). Fyom (25) and thecontinuity of
.F' we getF (r.):
O.'f,"rtirermoïe,
estimíte
(Zt) f,o_lbyl immediatelyfrom
(20) by using standard majorization techniques'[B], [4],[8].
To show uniqueness,letz e
U(rs,r2) with f lr¡ :0.
UsingiZ+¡'t*
r :0,
we obtain,.xr-z : Al'(F'
("0)- Ao)(r-ro) f ï,OOto,,(*o)("-*o)2 +
(26) * I o.'
lp" (y)- F" (ro)l (, -
a) da.aDo
We-Setll,-"oll (r-r-úoif.zeU ("0, r), and,llz-roll :þ(rz_úo),0 <u<
7,if z
€.U(rs,rz).
Hence,by
(21) and (26), wà getfoi
altn)0:
ll'i_
""'ll <
,!. tn,if. z e t(r9,r1),
andllr-",11 { t"?i-tn),
if.z e U('*0,r2)-."In
either case, rñ/e gutJ]lg rn:
zt which yieldst* :
z.That
completes the proof of Theorem1. !
We now state a theorem by pavaloiu [g] for comparison:
TnnoRpru
2. LetB,F: D c E
-+E
becontinuous operatorswi,thdiuid,ed, di,fferences of order onefr,y; Bl
and,[*,A; Fl
respecti,aely. Assume:(u)
(27) B(r):r-F(")("eD);
A Theorern forthe Stellensen \4ethod L25
6 7
124 Ioannis K. Arevros
(28)
(2e )
(b)
There eï'istsro € D
suclt'thatls - [r¡, B (*ù; ¡]-1 e L(E'E)
andllroll ( h';
(")
rnax{llr1 -
roll, llr, - A (z¡)ll} (
ho;Set also
B(r): r-F(') (*€D) ''r'hen'
using(4)' (5)'
(12)-(16)'
(27)
-
(30) and (33), we obtainl¿r
:
1.1566 265,h2:?, no:
'5391566, h3:
'4157352)
'25'a:!..1566265,Ó:'"3855422,c:'0119947'cl:'2891566 o :
1.0155688, B:'5392822'
condition
(31) of Theorem 2 cloes nob hold.That
is, Theoretn 2 cannot guar- anteethat
Steffensen methoclstarting
from z6:
0 convergesto
a solution of"qrrution
Fþ;) :0,
wher-e frtnctionF
is gìr'enby
(32)'-However: all conditions of Theorem
l
are satisfiecl. Indeeclf.orrt"tlt"
above we havethat
condition (7) is satisfiecl.I)efl'e
funcrion,/
sothat 1+.i
(¿): l¿;t (A(t) - ¡o)l
!¿e 'D)' It
caneasily be secn thal, tlrc left-hand siclc inequalil,.y
in
(15) is satlsilccl as equallry' whereas the right-hand sicle is smaller than 1. Fur-thelrnoreit
is sirnple algebrato
shorvthat
condition(9)
holdsalso. (
oncliiions (18)and
(19) are needeclto
show uniquenessor
trrá sorutionr* \n (J(ro,r2). 'rhey
canccrtainly
beUniqueness is then guarantecd only
(1) : -.3f33065, it
followsthat
for pothesesof
Theorem 1 are satisfied'arting
fi'omø¡ : 0,
convergesto
aI'(r):0,
',r'ÌrereF
is givenb)'
(3?)ae'iith
the resultsin
[1], [2]' [5]' [6]' [10], [11].Remarlt2.Cond,i,tion(7)canbereplaced'bgastronger'hrt'teasiertocheck Newton-I{antorouich type iy7tothesi's
l4),
lynomr'alsp:'
b, ^ ":,'- ".'. , .,','
1ft bapt(t):ït' -
(1- c)ttct, f¡(t;) ;ad+t:
andbz
:
1- c. lt
can eas'ily lte seen that-the
"l nomin
theu1"r^'Vt"r,iof,üdr,í5),
prouid'erl'that
hllpothesis2t4ct<(L-")'
holds, ancl 14,
rs
(r+{r5)
are tlte nonnegatiue zerr¡s of the equation p1(¿):
0'S;;r;p
(¿)< pt"(i) for
atl ¿€
[0, r2f,
ue hauer'1(
r¿(
rs{ rz'
Remark
3.
The results obtai,ned''in Theorern 2 can be ertend'edfor
steJJensen- Ai,tken metltod(33)
!)n,*r: an-lBt(a,),8,(aù; Fl-t F (aò (n2 0)'
whe,e
Bt,Bz: D C E -+ E
are continuous operators relatedfo F [4]'
[10]'[11].
Si'rnply,"¡tt*" O(")-: lr,B (t:); Fl (r e D) to
obtairt'a
TheorernI'
h,old,i'ng
for
method (33)' (d)(30) lll*,rt;t3l-la,u; Blll (hzllr-ull 'forallr'v'ue
D;(") (31) o,nd
hs:
holt'Jt',<
4;1(f)
U(ro,2ho)çD'
Then,
stefJer.tsen ,metl¿od çteneratecÌ,by (2)
i,swell
defined, remains 'inU'(ro,2lt'L)for"oltrr>0,anrlcon'uergestoasolutionu*ofequat'i'onF'(r):g' Màràor"r,
ihe ¡ottowing error bounds holdfor
alln)'
0:ll'-: ""ll " #'
and
ll"- - ,,ll {
hTlrzllr"- *n,-tll ll"- - B ('"-r)ll
' u:here
hT
>lll,n,R(,"); Fl-'ll t', >
olo"*n(;î:;ill¡r,y;Blll
(
r¿¿( r fu
aIIr,v ee u(ro,2hs),
thert'r*
'ís tÍ¿eu,ni,qui solution of equati,on
(1) i'nU (r¡,2hs)'
we
provide an exampleto
showthat
rrnder the conditions of Theorem l, steflensen method converges to a soluiionr*
ofequation (1)' whereas the same isnot
guaranteed under the conditions of Theorem 2'Exampte. Let
E:
IR, IJ: [-1,1], ro : 0'
Define functionf'
onD
by(32) F (r) : å,'* i* -
ancl the clivided difference
lr,y; Fllry
lr,a;
þ'l: F
(')^- !-("), r,u e D,, I
uu-'e
-r-l- 51 64
126 Ioannis K. Arsvros 8
I
A Theorem for the Steffensen Method L27Remarlc
4.
Cond,i,ti,on (11r) can be replaced byll¿tt (r'
@)- A("))ll <
cot
ctll" - "oll for
some co2 0, ct 2
0and
allU(rs,R).
We can also setc: cl+qR.
Cond,ition (18) can be replaced, bya { R,
but uniquenessis
then guaranteed onlyin U(rs,a).
Remarlc
5.
The results obtained ,in Theorem1
can be ertend,ed, so as to hold a rnore general setti,ng as follows:(a)
Let cs, c1 be nonnegatiue constantsiut,
u2 be posi,ti,ue monoton,ically increas'ing.functionsof
one uariableon[0,R]
wi,th]g3rr
(ú):
]ïu2(t) :
gsuch that
lltl,
@,,(*) _ F,,(".))ll ( ,r
(l1"_ roll)
,and
ll¿tt (¡' (") - A("))ll <
cs*
c1u2|1"- "oll)
for
allr €U (ro,R).
(b) Functi,onp giuen by t
p(t): I
,lA-lu1(r)ar+]ut2-
(1- c0-cru2(ú))¿+
d,on ll,Rl,
0
has a un'ique zero e¡
€
[0,R],
andp(rt) <
0.Moreover, set
e¡:
11 , andR:12.Furthermore,
replace conditions (12), (Ia)by (u)'
(7) by (b), and polynomial p by function p above. Then, under the rest of the hypotheses, asit
can easily be seen from the proof, the conclusions of Theoremt
hold in this more general setting. Call such a result Theorem 1/,.Finally,
notethat for u1(t) : at,
co: ct
and u1(¿):
O(or
c1:
0),¿
€
[0, rR] functionp
reducesto
polynomialp
andrheorem
1//to
Theorem 1.[6] CÀTINA$, E., On Some Steffensen-tgpe ,iterat,iue methods for a class of nonl'i,near equa- úions, Revue d'analyse numerique et de theorie de l'approximation, 24, nos. L-2, pp.37-43, 1995.
[7] GRAVES, L, M., Riemann integration and Taylor's theorem in general analysis, Trars.
Amer. Math. Soc., 29, pp. t63-L77,1927.
[8] KANTOROVICH, L. V. and AKILOV, G. P., Functional Analysis, Pergamon Press, Oxford, 1982.
[9] PÄVÄLOIrJ,L, Sur Ia méthode de Steffensen pour Ia résolut'ion d,es équations- opéra- tionnelles non I'inéaires, Rev. Roum. Math. Pure et Appl.,
XIII,
no. 6, pp. 857-861, 1968.[10] PAVÄLOIU, I., Sur une généralisation d,e Ia m,ethod,e d,e Steffensen, Revue d'analyse numerique et de theorie de l'approximation, 21, no. 1, pp. 59-65, 1992.
[11] PÀ.VÀLOIU,
I.,
B,i,Iateral approrirnat,ions for the solutions of scalar equations, Ftevre d'analyse numerique et de theorie de l'approximation, 23, no. 1, pp. 95 100, 1994.Received January 27, 2000 Cameron Un'iuersity
D epartrnent of M athematics
Lauton, OK 73505, U.S.A.
E -mail : ioannisa@cameron. edu
REFERENCES
[1] ARGYROS, I. K., On the solution of nonl'inear equo,tions wi,th a nonilifferentiable term, Revue d'analyse numerique et de theorie de 1'approximation,22, no. 2, pp. l2b-18b, 1993.
[2] ARGYROS, I. K., An error analysis for the Steffensen method under generalized, Zabrejlco- Nguen-type assumptions, Revue d'analyse numerique et de theorie de L'approximatìon, 25, nos. 7-2, pp.t7-22, 1996.
[3] ARGYROS, I. K., Polynomial Operator Equat,ions
in
Abstract Spaces and, Applications, CRC Press LLC, Boca Raton, Florida, 1gg8.[4] ARGYROS,
L
K. and sZIDARovszKY, F., The theory and Applications of lteration Method,s, C.R.C. Press, Boca Raton, Ftorida, 1gg3.[5] BALAzs, M. and GOLDNER, G., on the approrimate solut'ion of equations in Hilbert
space by a Steffensen-type method,, Revue d'analyse numerique et de theorie de 1,ap- proximation, L7, no. 1, pp. 19-23, 1gB8