On the influence of center-lipschitz conditions in the convergence analysis of multi-point iterative methods

Transcription

1 On the influence of center-lipschitz conditions in the convergence analysis of multi-point iterative methods Ioannis K. Argyros 1 Santhosh George 2 Abstract The aim of this article is to extend the local as well as the semilocal convergence analysis of multi-point iterative methods using center Lipschitz conditions in combination with our idea, of the restricted convergence region. It turns out that this way a finer convergence analysis for these methods is obtained than in earlier works without additional hypotheses. Numerical examples favoring our technique over earlier ones completes this article. AMS Subject Classification: 65F08, 37F50, 65N12. Key Words: Multi-point iterative methods; Banach space; local-semilocal convergence analysis. 1 Introduction Let X, Y be Banach spaces Ω X be a nonempty open set. By B(X, Y), we denote the space of bounded linear operators from X into Y. Let also U(w, d), be an open set centered at w X of radius d > 0 Ū(w, d) be its closure. Many problems from diverse disciplines such that Mathematics, Optimization, Mathematical Programming, Chemistry, Biology, Physics, Economics, Statistics, Engineering other disciplines [?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?], can be reduced to finding a solution x of the equation H(x) = 0, (1.1) where H : Ω Y is a continuous operator. Since,a unique solution x of equation (??) in a neighborhood of some initial data x 0 can be obtained only in special cases. Researchers construct iterative methods which generate a sequence converging to x. The most widely used iterative method is Newton s defined for each n = 0, 1, 2,... by x 0 Ω, x n+1 = x n H (x n ) 1 H(x n ). (1.2) 1 Department of Mathematical Sciences, Cameron University, Lawton, OK 73505, USA, iargyros@cameron.edu 2 Department of Mathematical Computational Sciences, National Institute of Technology Karnataka, India , sgeorge@nitk.edu.in by the author(s). Distributed under a Creative Commons CC BY license.

2 The order of convergence is an important concern when dealing with iterative methods. The computational cost increases in general especially when the convergence order increases. That is why researchers practitioners have developed iterative methods that on the one h avoid the computation of derivatives on the other h achieve high order of convergence. We consider the following multi-step iterative method defined for each n = 0, 1, 2,... by u n = v n (0) v n (1) = v n (0) H (v n (0) ) 1 H(v n (0) ) v n (2) = v n (1) H (v n (0) ) 1 H(v n (1) ) (1.3) u n+1 = v (k) n = v n (k 1) H (v n (0) ) 1 H(v (k 1) n ). The semi-local convergence of method (??) was given in [?]. It is well known that as the convergence order increases the convergence region decreases in general. To avoid this problem, we introduce a center-lipschitz-type condition that helps us determine an at least as small region as before containing the iterates {u n }. This way the resulting Lipschitz constants are at least as small. A tighter convergence analysis is obtained this way. The order of convergence was shown using Taylor expansions conditions reaching up to the k + 1 order derivative of H, although these derivatives do not appear in this method. As an academic example: Let X = Y = R, Ω = [ 5 2, 3 2 ]. Define ϕ on Ω by Then ϕ(x) = x 3 log x 2 + x 5 x 4 ϕ (x) = 3x 2 log x 2 + 5x 4 4x 3 + 2x 2, ϕ (x) = 6x log x x 3 12x x, ϕ (x) = 6 log x x 2 = 24x Obviously ϕ (x) is not bounded on Ω. So, the convergence of methods (??) is not guaranteed by the analysis in [?,?,?]. The rest of the article is organized as follows: Section 2 contains the conditions to be used in the semi-local convergence that follows in Section 3. Finally the numerical examples are given in the concluding Section 4. 2 Local convergence Let L 0 > 0, L > 0 L 1 1 be parameters. polynomial p by Define the scalar quadratic p(t) = (2L 0 + L)L 0 t 2 (4L 0 + 4L 0 L 1 + L)t + 2. (2.1) 2

3 The discriminant D of p is given by D = (4L 0 + 4L 0 L 1 + L) 2 8L 0 (2L 0 + L) = 16(L 0 L 1 ) 2 + L L 2 0L 1 + 8L 0 L 1 L > 0, so p has roots s 1 s 2 with 0 < s 1 < s 2 by the Descarte s rule of signs. Define also parameters ( L γ = 2(1 L 0 s 1 ) + 2L ) 0L 1 (1 L 0 s 1 ) 2 s 1 (2.2) r A = 2 2L 0 + L. (2.3) Notice that γ (0, 1], since p(s 1 ) = 0. The local convergence analysis of method (??) uses the conditions (A): (a1) H : Ω Y is a differentiable operator in the sense of Fréchet there exists x Ω such that H(x ) = 0 H (x ) 1 L(Y, X ). (a2) There exists L 0 > 0 such that for each x Ω Set Ω 0 = Ω B(x, 1 L 0 ). H (x ) 1 (H (x) H (x )) L 0 x x. (a3) There exist L = L(L 0 ) > 0 L 1 = L 1 (L 0 ) 1 such that for each x, y Ω 0 H (x ) 1 (H (y) H (x)) L x y H (x ) 1 H (x) L 1 x x. (a4) B(x, s 1 ) Ω. (a5) There exists s 3 s 1 such that s 3 < 2 L 0. Set Ω 1 = Ω B(x, 2 L 0 ). Based on the preceding conditions notations we can show a local convergence result for method (??). THEOREM 2.1 Under the conditions (A), further assume that u 0 B(x, s 1 ) {x }. Then, lim n u n = x, the following estimations hold v 1 n x L v 0 n x 2 2(1 L 0 v 0 0 x ), (2.4) v i n x γ i v 0 n x for each i = 2,..., k (2.5) u n+1 x γ k+n u 0 x. (2.6) Moreover, the point x is the unique solution of equation H(x) = 0 in the set Ω 1 given in (a5). 3

4 Proof. We use an induction based proof to show estimations (??)-(??). Let x B(x, s 1 ) {x }. By (a1) (a2), we obtain that H (x ) 1 (H (x) H (x )) L 0 x x < L 0 s 1 < 1. (2.7) It follows from the Banach lemma on invertible operators [?] (??) that H (x ) 1 L(Y, X ) H (x) 1 H (x ) 1 1 L 0 x x. (2.8) Then, x 0 = v0, 0 v0, 1..., v0 k are well defined by method (??) for n = 0. We can write v0 1 x = v0 0 x H (v0) 0 1 H(v0). 0 (2.9) Then, by using (a1), (a3), (??) (??), we get in turn that v 1 0 x = v 0 0 x H (v 0 0) 1 H(v 0 0) < H (v 0 0) 1 H (x ) 1 0 H (x ) 1 [H (x + θ(v 0 0 x )) H (v 0 0)]dθ(v 0 0 x ) L v0 0 x 2 2(1 L 0 v0 0 x ) Ls 1 2(1 L 0 s 1 ) v0 0 x < v 0 0 x < s 1, (2.10) which shows (??) for n = 0 v 1 0 B(x, s 1 ). Similarly by the second substep for n = 0, k = 2 we also get v 2 0 x = v 1 0 x H (v 0 0) 1 H(v 1 0) = v 1 0 x H (v 1 0) 1 H(v 1 0) +H (v 1 0) 1 [(H (v 0 0) H (x )) + (H (x ) H (v 1 0))] H (v 0 0) 1 H(v 1 0), (2.11) so by (a3), the definition of s 1 (??) (for x = v 1 0, v 0 0), we get in turn that v 2 0 x L v 1 0 x 2 2(1 L 0 v 1 0 x ) L 0 ( v0 1 x + v0 0 x ) + (1 L 0 v0 1 x )(1 L 0 v0 0 x ) ( L 2(1 L 0 s 1 ) + 2L ) 0L 1 (1 : 0 s 1 ) 2 s 1 v0 1 x γs 1 v 1 0 x < s 1, (2.12) 4

5 which shows (??) for n = 0 k = 2. Similarly, from so v0 i x = v0 i 1 x H (v0) 0 1 (H(v0 i 1 )) = v0 i 1 x H (v0 i 1 )H(v0 i 1 ) +H (v0 i 1 ) 1 [(H (v0) 0 H (x )) +(H (x ) H (v0 i 1 ))]H (v0) 0 1 H(v0 i 1 ) (2.13) v i 0 x γ v i 1 0 x γ i v 0 0 x < s 1, (2.14) x 1 x γ k+1 x 0 x < s 1, (2.15) which show (??) (??) for n = 1, i = 2, 3,..., k v i 0, x 1 B(x, s 1 ). Similarly, by induction we have in turn (as in (??)) leading to (as in (??)) (as in (??)) v 1 j x = v 0 j x H (v j 0 ) 1 H(v 0 j ) L v 0 j x 2 2(1 L 0 v 0 j x ) < s 1, (2.16) v i j x = v i 1 j x H (v 0 j ) 1 H(v 1 j ) +H (v i 1 j ) 1 [(H (v i 1 j ) H (x )) +(H (x ) H (vj 0 ))]H (vj 0 ) 1 H(v i 1 j ) v i j x γ i v 0 j x < s 1 (2.17) u n+1 x = v k n x γ k+n u 0 x < s 1, (2.18) which completes the induction for (??)-(??) also show that u n+1 B(x, s 1 ). It also follows from (??) that lim n u n = x, since γ [0, 1). Let x Ω 1 with H(x ) = 0. It then follows from the definition of Ω 1, (a2) T = 1 0 H (x + τ(x x ))dτ that H (x ) 1 (T H(x )) L so T 1 L(Y, X ). Then, from the estimate 0 x x dτ L 0 2 s 3 < 1, 0 = H(x ) H(x ) = T (x x ), (2.19) we get x = x. 5

6 REMARK 2.2 (a) In view of (a2), we can write H (x ) 1 H (x) = H (x ) 1 [(H (x) H (x )) + H (x )] 1 + H (x ) 1 (H (x) H (x )) 1 + L 0 x x, (2.20) so the second condition in (a3) can be dropped, we choose L 1 = 2, since x x s 1 < 1 L 0. (b) It follows from the definition of s 1 r A that s 1 < r A. That is the radius of convergence s 1 cannot be larger than the radius of convergence r A of Newton s method obtained by us [?,?,?,?,?]. (c) The local convergence of method (??) was not studied in [?]. But if it was call; s 1 the smallest positive solution of p(t) = 0, where p(t) = (2L 0 + L)L 0 t 2 (4L 0 + 4L 0 L1 + L)t + 2, (2.21) where L L 1 are the constants for the conditions in (a3) holding on Ω. But, we have L L (2.22) since Ω 0 Ω. Hence, we have L 1 L 1 (2.23) γ γ, (2.24) s 1 s 1. (2.25) Moreover, if strict inequality holds in (??) or (??), then, we have s 1 < s 1. Furthermore, by (??), our error bounds are more precise than the ones using L 0, L, L 1 γ. Hence, we have exped the applicability of method (??) in the local convergence case. In a similar way, we improve the semi-local convergence analysis of method (??) given in [?]. The work is given in the next section. 3 Semi-local convergence We need the following auxiliary result on majorizing sequences for method (??). LEMMA 3.1 Let K 0 > 0, K > 0, r 1 0 be parameters. Denote by δ the unique root in the interval (0, 1) of the polynomial ϕ given by ϕ(t) = 2K 0 t k+1 + K(t k + t k 1 2). Define the sequence {q n } for each n = 0, 1, 2,... i = 1, 2,..., k 1 by r0 0 = 0, rn 0 = r n, rn+1 1 = q n+1 + K(q n+1 q n + rn k 1 q n )(q n+1 rn k 1 ) (3.1) 1 2K 0 q n+1 6

7 r k n = q n+1, r i+1 n Moreover, suppose that = rn i + K(ri n q n + rn i 1 q n )(rn i rn i 1 ). 1 2K 0 q n 0 < K(q 1 + r k 1 0 ) 1 2K 0 q 1 δ < 1 2K 0 r 1 0. (3.2) Then, the sequence {q n } is increasing, bounded from above by q = r1 0 converges to its unique least upper bound q satisfying q 1 q q, 1 α r 1 n r i 1 n δ(r i 1 n r i 2 n ) δ kn+i 1 r 1 0, (3.3) r 1 n+1 r k n δ(r k n r k 1 n ) δ k(n+1) r 1 0 (3.4) q n = r 0 n r 1 n r 2 n... r k 1 n r k n = q n+1. (3.5) Proof. Replace t n, s i n, L 0, L, α in [?] by q n, r i n, K 0, K, δ. Next, we present the semi-local convergence analysis of method (??). THEOREM 3.2 Let H : Ω Y be a continuously differentiable operator in the sense of Fréchet [.,.; H] : Ω L(X, Y) be a divided difference of order one of H. Suppose there exist x 0 Ω K 0 > 0 such that H (x 0 ) 1 L(Y, X ). (3.6) H (x 0 ) 1 H(x 0 ) r 1 0, (3.7) H 1 (x 0 )([x, y; H] H (x 0 )) K 0 ( x x 0 + y x 0 ). (3.8) Set Ω 2 = Ω B(x 0, 1 2K 0 ). Moreover, suppose that for each x, y, z, w Ω 2 H (x 0 ) 1 ([x, y; H] [z, w; H]) K( x z + y w ), (3.9) the hypotheses of Lemma?? hold. Then, {u n } B(v 0, q ), lim n u n = u B(v 0, q ), H(u ) = 0. u u n q v n. (3.10) Moreover, u is the unique solution of equation H(x) = 0 in B(v 0, q ). Proof. Replace x n, y n, s i n, t n, α, L 0, L in [?] by u n, v n, r i n, q n, δ, K 0, K. REMARK 3.3 The condition H (x 0 ) 1 ([x, y; H] [z, w; H]) L( x z + y w ) (3.11) for each x, y, z, w Ω, some L > 0 is used in [?] instead of (??). But we have K 0 = L 0 (3.12) 7

8 K L, (3.13) δ α, (3.14) since Ω 2 Ω. Denote by q n, r i n the majorizing sequences used in [?] defined as sequences q n, r i n but with K 0 = L 0 K replaced by L. Then, we have by a simple induction argument, that q n q n (3.15) r i n r i n (3.16) 0 rn i q n r n i q n (3.17) 0 rn i rn i 1 r n i r n i 1 (3.18) 0 q n+1 rn k 1 q n r n k 1 (3.19) q q. (3.20) Moreover, if K < L, then (??-(??) hold as strict inequalities. Let us consider 1 the set Ω 3 = Ω B(x 1, 2K r1 0) provided that r0 1 < 1 2K. Moreover, suppose for each x, y, z, w Ω 3 H 1 (x 0 )([x, y; H] [z, w; H]) λ( x z + y w ). (3.21) Notice that Ω 3 Ω 2, so λ K. Then, Ω 3, (??), λ can replace Ω 2, (??), K respectively in Theorem??. Clearly, the corresponding to {t n } majorizing sequence call it { t n } is even tighter than {t n }. Hence, we have extended the applicability of method (??) in the semi-local convergence analysis too. These improvements are derived under the same conditions as in [?], since the computation of L is included in the computation of K as a special case. Examples where the new constants are smaller than the older ones can be found in the numerical section that follows in [?,?,?,?,?]. 4 Numerical examples We present the following examples to test the convergence criteria. Define the divided difference by [x, y; H] = 1 0 H (τx + (1 τ)y)dτ. EXAMPLE 4.1 Let X = Y = R 3, Ω = U(0, 1), x = (0, 0, 0) T define H on Ω by H(x) = H(x 1, x 2, x 3 ) = (e x1 1, e 1 2 x x 2, x 3 ) T. (4.1) 8

9 For the points u = (u 1, u 2, u 3 ) T, the Fréchet derivative is given by H (u) = eu (e 1)u Using the norm of the maximum of the rows (a 3 )-(a 4 ) since H (x ) = diag(1, 1, 1), we can define parameters for method (??) by L 0 = e 1, L 1 = L = e 1 e 1, L = L1 = e. Then, s 1 = The old radius is s 1 = EXAMPLE 4.2 Let X = Y = R, Ω = Ū(x 0, 1 ξ), x 0 = 1 ξ [0, 1 2 ). Define function H on Ω by H(x) = x 3 ξ. Then, we get by (??)-(??) (??) that for (i) k = 1, q 1 = r0 1 = 1 3 (1 ξ), L 0 = K 0 = 1 2 (3 ξ), L = 2 ξ, K = K 0. Notice that L 0 < K < L. The conditions of Lemma?? are satisfied for ξ I 1 = [ , 0.5) but the earlier conditions in [?] are satisfied for ξ I 2 = [ , 0.5), I 2 I 1. Moreover, if λ = 1 3(3 ξ) ( 2ξ2 + 5ξ + 6) then conditions of Lemma?? are satisfied for ξ I 3 = [ , 0.5). (ii)for k = 2, the conditions of Lemma?? are satisfied for ξ I 1 = [ , 0.7) but the earlier conditions in [?] are satisfied for ξ I 2 = [ , 0.7), I 2 I 1. Moreover, if λ = 1 3(3 ξ) ( 2ξ2 + 5ξ + 6) then conditions of Lemma?? are satisfied for ξ I 3 = [ , 0.7). 5 Conclusion Our idea of the convergence region in connection to the center Lipschitz condition were utilized to provide a local as well as a semilocal convergence analysis of method (??). Due to the fact that we locate a region at least as small as in earlier works [?] containing the iterates, the new Lipschitz parameters are also at least as small. This technique leads to a finer convergence analysis (see also Remark??, Remark?? the numerical examples). The novelty of the paper not only lies in the introduction of the new idea but also obtained using special cases of Lipschitz parameters appearing in [?]. Hence, no additional work to [?] is needed to arrive at these developments. This idea can be used to extend the applicability of other iterative methods appearing [?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?] along the same lines. References [1] S. Amat, I. K. Argyros, S. Busquier, M. A. Herńez-Veŕon. On two high-order families of frozen newton-type methods. Numerical Linear Algebra with Applications, 25(1),

10 [2] S. Amat, C. Berḿudez, M.A. Herńez-Veŕon, E. Martinez. On an efficient k-step iterative method for nonlinear equations. Journal of Computational Applied Mathematics, 302: , [3] S. Amat, S. Busquier, C. Berḿudez, S. Plaza, On two families of high order newton type methods. Applied Mathematics Letters, 25(12): , [4] S. Amat, S. Busquier, S. Plaza, Review of some iterative root-finding methods from a dynamical point of view. Scientia, 10(3):35, [5] I.K. Argyros, Computational theory of iterative methods, volume 15. Elsevier, [6] I. K. Argyros, On the semilocal convergence of a fast two-step newton method. Revista Colombiana de Matematicas, 42(1):15-24, [7] I. K. Argyros, S. George, N. Thapa, Mathematical Modeling For The Solution Of Equations And Systems Of Equations With Applications, Volume-I, Nova Publishes, NY, [8] I. K. Argyros, S. George, N. Thapa, Mathematical Modeling For The Solution Of Equations And Systems Of Equations With Applications, Volume- II, Nova Publishes, NY, [9] I. K. Argyros, A. Cordero, A. A. Magreñán, J. R. Torregrosa, Thirddegree anomalies of traubs method. Journal of Computational Applied Mathematics, 309: , [10] I. K. Argyros S. Hilout, Weaker conditions for the convergence of newtons method. Journal of Complexity, 28(3): , [11] I. K. Argyros S. Hilout, Computational methods in nonlinear analysis: efficient algorithms, fixed point theory applications. World Scientific, [12] I. K. Argyros, A. A. Magreñán, L. Orcos, J.A Sicilia, Local convergence of a relaxed two-step newton like method with applications. Journal of Mathematical Chemistry, 55(7): , [13] I. K.Argyros, A. A. Magréñan, A contemporary study of iterative methods, Elsevier (Academic Press), New York, [14] I. K.Argyros, A. A. Magréñan, Iterative methods their dynamics with applications, CRC Press, New York, USA, [15] A.Cordero, L. Guasp, J. R. Torregrosa, Choosing the most stable members of kous family of iterative methods. Journal of Computational Applied Mathematics,

11 [16] A. Cordero J. R. Torregrosa, Low-complexity root-finding iteration functions with no derivatives of any order of convergence. Journal of Computational Applied Mathematics, 275: , [17] A. Cordero, J. R. Torregrosa, P. Vindel, Study of the dynamics of thirdorder iterative methods on quadratic polynomials. International Journal of Computer Mathematics, 89(13-14): , [18] J.A. Ezquerro M.A. Herńez-Veŕon, How to improve the domain of starting points for steffensen s method. Studies in Applied Mathematics, 132(4): , [19] J. A. Ezquerro M. A. Herńez-Veŕon, Majorizing sequences for nonlinear fredholm hammerstein integral equations. Studies in Applied Mathematics. [20] J. A. Ezquerro, M. Grau-Śanchez, M. A. Herńez-Veŕon, M. Noguera, A family of iterative methods that uses divided differences of first second orders. Numerical algorithms, 70(3): , [21] J. A. Ezquerro, M.A.Herńez-Veŕon, A. I. Velasco, An analysis of the semilocal convergence for secant-like methods. Applied Mathematics Computation, 266: , [22] M.A. Herńez-Veŕon, E. Martínez, C. Teruel, Semilocal convergence of a k-step iterative process its application for solving a special kind of conservative problems. Numerical Algorithms, 76(2): , [23] L.V. Kantlorovich G.P. Akilov, Functional analysis pergamon press, [24] M. Moccari T. Lofti, Using majorizing sequences for the semi-local convergence of a high order multi-point method along with stability analysis, Intern.J. Non. Sci. Numer. Stimulation, [25] A. A. Magreñán, A.Cordero, J. M. Gutiérrez, J. R. Torregrosa, Real qualitative behavior of a fourth-order family of iterative methods by using the convergence plane. Mathematics Computers in Simulation, 105:49-61, [26] A. A. Magreñán I. K.Argyros, Improved convergence analysis for Newton-like methods. Numerical Algorithms, 71(4): , [27] A. A. Magreñán I. K. Argyros, Two-step newton methods. Journal of Complexity, 30(4): , [28] F.A. Potra V. Pták, Nondiscrete induction iterative processes, volume 103. Pitman Advanced Publishing Program,

12 [29] Rheinboldt, W.C., An adaptive continuation process for solving systems of nonlinear equations, Polish Academy of Science, Banach Ctr. Publ. 3 (1978), no. 1, [30] H.Ren I. K. Argyros, On the convergence of kingwerner-type methods of order free of derivatives. Applied Mathematics Computation, 256: , R volume 1. Cambridge University [31] W.T. Shaw, Complex analysis with Mathematical Press, [32] J. F. Traub, Iterative methods for the solution of equations. American Mathematical Soc.,