Return to Article Details Chebyshev-like methods and quadratic equations

Chebyshev-like methods and Quadratic Equations

J. A. Ezquerro, J. M. Gutiérrez, M. A. Hernández and M. A. Salanova
Abstract.

We show the classical Kantorovich technique to study the convergence of a new uniparametric family of third order iterative processes defined in Banach spaces. We obtain information about this family from the study of the same family in the real case. Besides we obtain, for a value of the parameter, a method which has order four when it is applied to quadratic equations.

Key words and phrases:
Iterative processes, Chebyshev method, quadratic equations.
1991 Mathematics Subject Classification:
47H10, 65J15.
Supported in part by a grant of the University of La Rioja.

1. Introduction

Several scientific problems can be written in the form

(1) F(x)=0.

In order to generalize as much as possible, let F be a nonlinear operator defined in an open convex domain Ω of a Banach space X with values in another Banach space Y. There are a lot of research works concerning the numerical solution of (1) by means of iterative processes, mainly by using Newton’s method [8], [11].

But there are other iterative processes to solve (1). One of them is Chebyshev method [4], [5]:

(2) xn+1=xn[1+12LF(xn)]F(xn)1F(xn),

where I is the identity operator on X and LF(x) is the linear operator on X given by

LF(x)=F(x)1F′′(x)F(x)1F(x),xX,

assuming that F(x)1 exists.

We notice that the only inverse operator we have to evaluate in each step of Chebyshev method is F(xn)1. This same inverse must be also calculate in each step of Newton’s method. However, the expression of other third order iterative

processes (Halley, super-Halley [7]) involves the computation of other different inverse operators. So, we can see Chebyshev method as the third order method with less computational cost.

In this paper we study a uniparametric family of third order iterations that only needs the computation of the inverse F(xn)1 in each step:

(3) xn+1=xn[1+12LF(xn)+αLF(xn)2]F(xn)1F(xn),α[0,12].

This family incudes the Chebyshev method as a particular case (α=0). We prove that the methods of (3) have got a faster convergence than Chebyshev method. Secondly, we find the value of the parameter α for which the convergence is four when the method is applied to quadratic equations. Finally, we compare the methods of the family (3), showing that for α=12 we obtain the best method, from the standpoint of the speed of convergence and the error estimates.

The origin of the family (3) is a Gander’s result [6] on third order iterative processes in the real case. In this work it is established that for a real function f of real variable, the iteration in the form:

xn+1=xnH(Lf(xn))f(xn)f(xn),

where

Lf(x)=f(x)f′′(x)f(x)2,

and H is a function satisfying H(0)=1,H(0)=12 and |H′′(0)|<+, are cubically convergent. In this paper we consider a function H with a finite Taylor’s expansion and its generalization to Banach spaces. Of course, there are other functions H that give rise to third order methods. But, for instance, the functions that originate the Halley or the super-Halley methods, have got a non finite Taylor expansion. Hence the computational cost is higher.

2. A study of the convergence

To prove the convergence of the iterative processes given in (3), we assume that F satisfies the classical Kantorovich conditions [8]. These kind of conditions have been used by different authors in the study of third order iterative processes [1], [2], [4], [13]. So, throughout this paper we assume the following conditions:

  • (i)

    There exists x0Ω where Γ0=F(x0)1 is defined.

  • (ii)

    Γ0(F′′(x)F′′(y))kxy,x,yΩ,k0.

  • (iii)

    Γ0F(x0)a,Γ0F′′(x0)b.

  • (iv)

    The equation

    (4) p(t)k6t3+b2t2t+a=0

    has got a negative root and two positive roots r1 and r2,(r1r2) when k>0. If k=0, then (4) has got two positive roots r1 and r2(r1r2). Equivalently,

    ab2+4kbb2+2k3k(b+b2+2k),if k>0,

    or ab12 if k=0.

First, we analyse the real sequence {tn} defined for

(5) t0=0,tn+1=G(tn)=tn[1+12Lp(tn)+αLp(tn)2]p(tn)p(tn),n0,

when α[0,12] and p is the polynomial (4).

The real sequences (5) allow us to establish the convergence of the sequences (3) defined in Banach spaces. So, under the hypothesis (𝐢)(𝐢𝐯), the sequences (5) and (3) are well defined and converge to r1 and to x, a solution of (1), respectively. Moreover

xn+1xntn+1tn,xxnr1tn,n0,

that is, {tn} majorizes {xn} (see [8], [12]).

Lemma 2.1.

Let p be the polynomial defined in (4). Then if 0α0.5, the sequences {tn} given in (5) are well defined, are increasing and converge to r1.

Proof.

It is easy to prove that tn+1tn because p is a decreasing convex polynomial in [0,r1].

On the other hand, let G be the function defined in (5). Then

G(t)=Lp(t)2[3(12α)+5αLp(t)(12+2αLp(t))Lp(t)].

As Lp(t)0 for t[0,r1], we have G(t)0, for t[0,r1], 0α12. Thus tnr1. Consequently, {tn} converges to r1.

Notice that the sequences {tn} converge cubically to r1 as a consequence of the Gander’s result given in the introduction. Now, we study the sequences {xn},α[0,12], defined in Banach spaces.

Lemma 2.2.

For α[0,12] and n0 we have:

  • [In]

    There exists Γn=F(xn)1.

  • [IIn]

    ΓnF(x0)p(t0)p(tn).

  • [IIIn]

    Γ0F′′(xn)p′′(tn)p(t0).

  • [IVn]

    Γ0F(xn)p(tn)p(t0).

  • [Vn]

    xn+1xntn+1tn.

Proof.

We use an inductive process. For n=0, [I0]–[IV0] follow immediately from the hypothesis. To prove [V0] we have

LF(x0)Γ0F′′(x0)Γ0F(x0)p(t0)p′′(t0)p(t0)2Lp(t0),

and then,

x1x0 (1+12LF(x0)+αLF(x0)2)Γ0F(x0)
(1+12Lp(t0)+αLp(t0)2)p(t0)p(t0)=t1t0.

Let us assume now that [In]–[Vn] are true for a given n. Then

Γ0F(xn+1)IΓ0F′′(x0)(xn+1x0) =x0xn+1Γ0[F′′(x)F′′(x0)]𝑑x
k2xn+1x02.

Consequently,

IΓ0F(xn+1) k2xn+1x02+bxn+1x0
k2tn+12+btn+1
=1+p(tn+1)<1,

and, by the Banach lemma on inversion of operators, there exists [Γ0F(xn+1)]1 (and therefore F(xn+1)1) and besides,

ΓnF(x0)1p(tn)=p(t0)p(tn).

So we have shown [In+1] and [IIn+1].

To see [IIIn+1], we notice that

Γ0F′′xn+1 =Γ0[F′′(xn+1)F′′(x0)]+Γ0F′′(x0)
kxn+1x0+bp′′(tn+1).

To show [IVn+1] we have, from (3) and the Taylor’s formula,

F(xn+1) =F(xn)+F(xn)(xn+1xn)+12F′′(xn)(xn+1xn)2
+xnxn+1[F′′(x)F′′(xn)](xn+1x)𝑑x
=12F′′(xn)(ΓnF(xn))2αF′′(xn)ΓnF(xn)LF(xn)ΓnF(xn)
+12F′′(xn)(xn+1xn)2+xnxn+1[F′′(x)F′′(xn)](xn+1x)𝑑x
=(12α)F′′(xn)ΓnF(xn)LF(xn)ΓnF(xn)
+αF′′(xn)ΓnF(xn)LF(xn)2ΓnF(xn)
+18F′′(xn)LF(xn)ΓnF(xn)LF(xn)ΓnF(xn)
+α2F′′(xn)LF(xn)ΓnF(xn)LF(xn)2ΓnF(xn)
+α22F′′(xn)LF(xn)2ΓnF(xn)LF(xn)2ΓnF(xn)
+xnxn+1[F′′(x)F′′(xn)](xn+1x)𝑑x.

The inductive procedure leads us to

LF(xn)Lp(tn),ΓnF(xn)p(tn)p(tn).

As 0α12, we deduce

Γ0F(xn+1)k6(tn+1tn)3
p(tn)p(t0)[(12α)Lp(tn)2+αLp(tn)3+18Lp(tn)3+α2Lp(tn)4+α22Lp(tn)5].

Repeating the same process with the polynomial p, we obtain

p(tn+1)=k6(tn+1tn)3+
+p(tn)[(12α)Lp(tn)2+αLp(tn)3+18Lp(tn)3+α2Lp(tn)4+α22Lp(tn)5],

and consequently,

(6) Γ0F(xn+1)p(tn+1)p(t0).

Finally, to see [Vn+1], we proceed as in the case n=0.

The following result gives us conditions on the convergence of the methods or the family (3).

Theorem 2.3.

Assume that conditions (i)–(iv) hold and

B¯=B(x0,r1)¯={xX;xx0r1}Ω.

Then:

  • (a)

    The sequences (3) are well-defined for α[0,12], lie in B (interior of B¯) and converge to a solution x of the equation (1).

  • (b)

    xis the only solution of (1) in B(x0,r2)Ω.

  • (c)

    We have the following error bounds:

    xxnr1tn.
Proof.

The convergence follows immediately from Lemmas 2.1 and 2.2. Besides, x is the solution of (1) as a consequence of (6). To show the uniqueness, we assume that y is another solution of (1) in B(x0,r)¯ for some r>0. Then, following Argyros and Chen [4], we deduce

0=F(y)F(x)=01F(x+t(yx))𝑑t(yx).

We have to see that 01F(x+t(yx))𝑑t has got an inverse. So, we take into account that

IΓ001F(x+t(yx))𝑑t=
=Γ001x0x+t(yx)F′′(z)𝑑z𝑑t
=Γ001x0x+t(yx)[F′′(x0)+(F′′(z)F′′(x0))]𝑑z𝑑t,

and if γ(t)=xx0+t(yx), then

IΓ001F(x+t(yx))𝑑t
01(b+k2γ(t))γ(t)𝑑t
<01(b+k2[tr+(1t)r1])[tr(1t)r1]𝑑t
=k6r22+(k6r1+b2)(r2+r1).

Now, we define the polynomial

q(r)=k6r2+(k6r1+b2)r+(k6r12+k2r11).

If q(r)0, then

tΓ001F(x+t(yx))𝑑t<1

and the inverse 01F(x+t(yx))𝑑t exists. Consequently, we deduce the uniqueness of the solution x in B(x0,r)Ω.

Notice that q(0)<0 and q(r1)=p(r1)<0 and then, the uniqueness holds in B(x0,r)Ω with r>r1. Moreover, by using Cardano’s formulas, for k>0 we have r1+r2=r03bk and r1r2=6akr0, where r0,r1 and r2 are the roots of the equation (4). So q(r2)=2a+br02r0<0. If k=0, then q(r2)=0. In both cases, the uniqueness in B(x0,r2)Ω holds (if k>0 the uniqueness holds in B(x0,r3)Ω where r3 is the positive root of q(r)=0).

Finally, by Lemma 2.2 and for m0, we have

xn+mxntn+mtn.

By letting m, we conclude the result. ∎

3. Error estimates

This section is devoted to the study of the error estimates for the real sequences (5). As a result we obtain error bounds for the sequences defined in Banach spaces.

We distinguish two situations:

  • If the polynomial p given in (4) is quadratic, we follow the technique developed by Ostrowski [9], to obtain error bounds.

  • If p is a cubic polynomial, we use the technique introduced in [7]. This provides error estimates instead of error bounds.

Let p be the polynomial given in (4) with k=0:

p(t)=b2t2+a.

Assume that p has two positive roots r1 and r2(r1r2). Let {tn} be the sequence defined in (5) and an=r1tn,bn=r2tn, n0. Then

p(tn)=b2anbn,p(tn)=b2(an+bn).

By (5), we have

an+1=r1tn+1=an3an3+4an2bn+5anbn2+2(12α)bn3(an+bn)5

and

bn+1=r2tn+1=bn3bn3+4bn2an+5bnan2+2(12α)an3(an+bn)5.

If r1<r2, let us write θ=r1r2<1 and μn=anbn. Then μn+1=μn3h(μn) where

h(x)=x3+4x2+5x+2(12α)2(12α)x3+5x2+4x+1,0x1.

If α=12, notice that

μn4μn+1=μn4μn2+4μn+55μn2+4μn+15μn4.

So μn5μn14513[513θ]4n and μnμn14θ4n. Hence

(r2tn)θ4nr1tn(r2tn)513[513θ]4n.

Consequently,

(r2r1)θ4n1θ4nr1tn513(r2r1)[513θ]4n1513[513θ]4n.

The second inequality holds if 5θ3<1.

On the other hand, as k is decreasing in α, the best error bound is attained for α=12 and the worse for α=0 (Chebyshev method).

Then, the error bounds obtained for the method of the family (5)(0<α<12) are better than the bonds given by Chebyshev method and worse than the ones given by the method (5) with α=12.

The definition of R-order of convergence given by Potra [10] allows us to establish that the methods of the family (5) and hence the methods of (3), have at least R-order three for 0α12 and four for α=12.

Finally, we analyze the case r1=r2. Then an=bn and

an+1 =an3α8,
r1tn =r1[3α8]n.

We see that the convergence of the methods of (5) is linear, but it is faster for increasing values of the parameter α.

Let us now consider the polynomial p, defined in (4), with k>0. Assume that p has two positive roots r1 and r2(r1r2) and a negative root, r0, that is

p(t)=k6(r1t)(r2t)(r0+t).

First we study the case r1<r2. We denote again an=r1tn,bn=r2tn and define

Q(tn)=bn3an+1an3bn+1=(r1G(tn))(r2tn)3(r2G(tn))(r1tn)3,

with G defined in (5).

As G(r1)=r1, G(r1)=G′′(r1)=0, we have for t close enough to r1

Q(t)(r2r1)2limtr1r1G(t)(r1t)3 =G′′′(r1)6(r2r1)2
=3(12α)p′′(r1)2+p′′′(r1)p(r1)6p(r1)2(r2r1)2
=(r2r1)(r0+r1)+2(12α)(r02r1r2)2(r0+r1)2=λ.

By letting n, we have tnr1 and it follows

anbn(an1bn1)3λ(λr1r2)3n1λ,

and if λθ<1, then

r1tn(r2r1)(λθ)3nλ(λθ)3n,n0.

When r1=r2, we have

Q~(tn)=an+1an=r1G(tn)r1tn.

Then, when t approaches r1,

Q~(t)Q~(r1)=3α8,

and

r1tnr1(3α8)n.

If p is a cubic polynomial, we obtain better error estimates for increasing

values of α[0,12], because λ is decreasing as a function of α. We obtain again the best error estimates for the method defined in (3) with α=12.

Now we are ready to give the following result.

Lemma 3.1.

With the previous notations, let θ=r1r2. Then for the methods of the family (5) applied to the polynomial (4) with k=0, we have the following error bounds:

  • When r1<r2 and α=12.

    (r2r1)θ4n1θ4nr1tnr2r11513[513θ]4n513[513θ]4n,if 53θ<1.
  • When r1<r2 and 0α<12:

    (r2r1)θ3n1θ3nr1tn(r2r1)[2θ]3n2[2θ]3n,if 2θ<1.
  • When r1=r2:

    r1tn=r1[3α8]n.

    If we apply the methods of (5) to a polynomial (4) with k>0, the error estimates are:

  • When r1<r2:

    r1tn(r2r1)(λθ)3nλ(λθ)3n,if λθ<1,

    with λ=(r2r1)(r0+r1)+2(12α)(r0+2r1r2)2(r0+r1)2.

  • When r1=r2:

    r1tnr1(3α8)n.

4. Examples

We give two examples to illustrate the previous results.

Example 1.

Let us consider the space X=C[(0,1)] of all continuous functions defined in the interval [0,1] with the norm

x=maxs[0,1]|x(s)|,

and the equation f(x)=0, where

F(x)(s)=x(s)s+1201scos(x(t))𝑑t,xC([0,1]),s[0,1].

With the previous notations and for x0=x0(s)=s, we calculate the first and second Fréchet derivatives of F to obtain

a=b=sin12sin1+cos1,k=12sin1+cos1.

So, in this case, the polynomial (4) is

p(t)=16(2sin1+cos1)[t3+3(sin1)t26(2sin1+cos1)t+6sin1],

which has two positive roots

r1=0.6095694860276291,r2=1.70990829134757.

Then, by Theorem 2.3, we know that F(x)=0 has a solution in B(x0,r1)¯ and this solution is unique in B(x0,r2).

Moreover, we have the following error estimates for 1011xx2 when α=0 and α=12:

Ifα=0, 1011xx21011(r1t2)48601696.08024329.

If α=12, 1011xx21011(r1t2)16958.95309621804. ∎

Observe that the previous results do agree quite well with our previous analysis which showed that better error estimates are obtained for the method (3) with α=12.

Example 2.

Let us consider again the space X=C[(0,1)], where now we define the equation F(x)=0, with

F(x)(s)=1x(s)+x(s)401ss+tx(t)𝑑t,xC([0,1]),s[0,1].

This is a quadratic equation that is known as Chandrasekhar’s equation [3].

For x0=x0(s)=1, we have a=0.2652, b=0.5303 and k=0. In that case the polynomial (4) has two roots:

r1=0.287042,r2=3.48512.

So Chandrasekhar’s equation has a solution that lies in B(x0,r1)¯ and is unique in B(x0,r2).

In Tables 1 and 2 we show some error expressions.

xxnr1tn

for two methods of the family (3) and other well known third order iterative processes, like the Halley method ({un}) or the super-Halley method ({vn}).

n r1tn(α=0) r1tn(α=12)
0 0.2870424876072915 0.2870424876072915
1 0.0035775864318520 0.0007359920056491
2 8.92413109 4.481221014
3 1.389791025 6.164361055
Table 1. Methods of (3) for α=0 and α=12
n r1un r1vn
0 0.2870424876072915 0.2870424876072915
1 0.0017877932422075 0.0001471713056166
2 5.5775831010 1.4339891017
3 1.6965261029 1.2927591069
Table 2. Halley and super-Halley methods

As we can see the method corresponding to α=12 provides better error estimates than the Chebyshev and the Halley methods, but worse than the super-Halley method. The cause is that the super-Halley method also has convergence of order four when it is applied to quadratic equations. ∎

References

  • [1] G. Alefeld, On the convergence of Halley’s method, Amer. Math. Monthly, 88 (1981), 530–536.
  • [2] M. Altman, Concerning the method of tangent hyperbolas for operator equations, Bull. Acad. Pol. Sci., Serie Sci. Math., Ast. et Phys., 9(1961), 633–637.
  • [3] I.K. Argyros, Quadratic equations and aplications to Chandrasekhar’s and related equations, Bull. Austral. Math. Soc., 38 (1988), 275–292.
  • [4] I.K. Argyros and D. Chen, Results on the Chebyshev method in Banach spaces, Proyecciones, 12, 2 (1993), 119–128.
  • [5] V. Candela and A. Marquina, Recurrence relations for rational cubic methods II: the Chebyshev method, Computing, 45 (1990), 355–367.
  • [6] W. Gander, On Halley’s iteration method, Amer. Math. Monthly, 92 (1985), 131–134.
  • [7] J.M. Gutiérrez and M.A. Hernández, A family of Chebyshev-Halley type methods in Banach spaces, Bull. Austral. Math. Soc., 55 (1997), 113–130.
  • [8] L.V. Kantorovich and G.P. Akilov, Functional Analysis, Pergamon Press, 1982.
  • [9] A.M. Ostrowski, Solution of Equations in Euclidean and Banach Spaces, Academic Press, 1963.
  • [10] F.A. Potra and V. Ptak, Nondiscrete Induction and Iterative Processes, Pitman Pub., 1983.
  • [11] L.B. Rall, Computational Solution of Nonlinear Operator Equations, Robert E. Krieger Publishing Company, Inc., 1979.
  • [12] W.C. Rheinboldt. A unified convergence theory for a class of iterative processes, SIAM J. Numer. Anal., 5 (1968), 42–63.
  • [13] T. Yamamoto, On the method of tangent hyperbolas in Banach spaces, J. Comput. Appl. Math. 21 (1988), 75–86.

Received February 18, 1998

Universidad de la Rioja

Dpto. de Matemáticas y Computación

C/Luis de Ulloa s/n, 26004, Logroño

Spain