TSWJ The Scientific World Journal 1537-744X 2356-6140 Hindawi Publishing Corporation 10.1155/2014/560931 560931 Research Article An Inversion-Free Method for Finding Positive Definite Solution of a Rational Matrix Equation http://orcid.org/0000-0002-6905-8951 Soleymani Fazlollah 1 Sharifi Mahdi 1 Karimi Vanani Solat 1 Khaksar Haghani Farhad 1 Kılıçman Adem 2 Saberi Nik Hassan 1 Department of Mathematics Islamic Azad University, Shahrekord Branch Shahrekord Iran srbiau.ac.ir 2 Department of Mathematics and Institute for Mathematical Research Universiti Putra Malaysia 43400 Serdang Malaysia upm.edu.my 2014 1882014 2014 15 07 2014 26 07 2014 19 8 2014 2014 Copyright © 2014 Fazlollah Soleymani et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

A new iterative scheme has been constructed for finding minimal solution of a rational matrix equation of the form X+A*X-1A=I. The new method is inversion-free per computing step. The convergence of the method has been studied and tested via numerical experiments.

1. Introduction

In this paper, we will discuss the following nonlinear matrix equation: (1)X+A*X-1A=I, where A is an n×n nonsingular complex matrix, I is the unit matrix of the appropriate size, and XCn×n is an unknown Hermitian positive definite (HPD) matrix that should be found. It was proved in  that if (1) has an HPD solution, then all its Hermitian solutions are positive definite and, moreover, it has the maximal solution XL and the minimal solution XS in the sense that XSXXL for any HPD solution X.

A lot of papers have been published regarding the iterative HPD solutions of such nonlinear rational matrix equations in the literature due to their importance in some practical problems arising in control theory, dynamical problems, and so forth (see [2, 3]).

The most common iterative method for finding the maximal solution of (1) is the following fixed-point iteration : (2)X0=I,Xk+1=I-A*Xk-1A.

The maximal solution of (1) can be obtained through XL=I-YS, where YS is the minimal solution of the dual equation Y+AY-1A*=I.

In 2010, Monsalve and Raydan in  proposed the following iteration method (also known as Newton’s method) for finding the minimal solution: (3)X0=AA*,Xk+1=Xk(2I-A-*(I-Xk)A-1Xk), which is an inversion-free scheme. Since, A-1 should be computed only once in contrast to the matrix iteration (2). Note that A-*=A-1*, and similar notations are used throughout.

Remark 1.

We remark that there are several other well-known iterative methods for solving (1) rather than Newton’s method (3). To the best of our knowledge, the procedure of extending higher-order iterative methods for finding the solution of (1) has not been exploited up to now. Hence, we hope that this interlink among the fields of root-finding and solving (1) may lead to discovering novel and innovative techniques.

The rest of this paper is organized as follows. In Section 2, we develop and analyze a new inversion-free method for finding roots of a special map F. In Section 3, we provide some numerical comparisons by employing some experiments in machine precision. Some concluding remarks will be drawn in Section 4.

2. A New Iterative Method

An equivalent formulation of (1) is to find an HPD matrix X such that F(X)=0. Toward this goal, we write F(X)X+A*X-1A-I=0. Furthermore, we have (4)F(X)=A*X-1A+X-I=A*X-1A-(I-X)=A*X-1A-(A-*-XA-*)A*=A*X-1A-A(A-1A-*-A-1XA-*)A*=(A-1XA-*)-1-A(A-1A-*-A-1XA-*)A*. Now, using a change of variable as Z=A-1XA-*, we could simplify (4) as follows: (5)G(Z)=Z-1-A(A-1A-*-Z)A*=0.

In order to obtain an iterative method for finding the minimal HPD solution of (1), it is now enough to solve the well-known matrix equation Z-1-B=0, at which B=A(A-1A-*-Z)A*. One of such ways to challenge this matrix inversion problem is via applying the Schulz-type iteration methods (see, e.g., ). Applying Chebyshev’s method  yields (6)X0=AA*,Xk+1=Xk[(3I-A-*(I-Xk)A-1Xk)3I-A-*(I-Xk)×A-1Xk(3I-A-*(I-Xk)A-1Xk)].

We remark that there is a tight relationship between iterative methods for nonlinear systems and the construction of higher-order methods for matrix equations ([12, 13]).

The matrix iteration (6) requires A-1 to be computed only once at the beginning of the iteration and this makes the iterative method fall in the category of inversion-free algorithms for solving (1).

In the meantime, it is easy to show that the zeros of the map G(Z)=Z-1-B are equal to the zeros of the map F(X)=X-1-H, wherein H=A-*(I-X)A-1. To be more precise, we are finding the inverse of the matrix H which matches the minimal HPD solution of (1).

Remark 2.

Following Remark 1, we applied Chebyshev’s method for (1) in this work and will study its theoretical behavior. The extension of the other well-known root-finding schemes for finding the minimal solution of (1) will remain for future studies.

Lemma 3.

The proposed method (6) produces a sequence of Hermitian matrices using the Hermitian initial matrix X0=AA*.

Proof.

The initial matrix AA* is Hermitian, and Hk=A-*(I-Xk)A-1. Thus, H0=A-*A-1-A-*AA*A-1 is also Hermitian; that is, H0*=H0. Now using inductive argument, we have (7)(X1)*=(X0[3I-H0X0(3I-H0X0)])*=(3X0-3X0H0X0+[X0H0X0H0X0])*=3X0-3X0H0X0+X0H0X0H0X0=X1. By considering (Xl)*=Xl, (lk) we now show that (8)(Xl+1)*=(Xl[3I-HlXl(3I-HlXl)])*=(3Xl-3XlHlXl+[XlHlXlHlXl])*=3Xl-3XlHlXl+XlHlXlHlXl=Xl+1. Note that Hl=(Hl)* has been used in (8). Now the conclusion holds for any l+1. Thus, the proof is complete.

Theorem 4.

By considering that A and Xk are nonsingular matrices, the sequence {Xk} generated by (6) is convergent to the minimal solution using the initial matrix X0=AA*.

Proof.

Let us consider Hk=A-*(I-Xk)A-1. We therefore have (9)I-HkXk+1=I-[A-*(I-Xk)A-1]×(Xk[3I-A-*(I-Xk)A-1Xk×(3I-A-*(I-Xk)A-1Xk)])=(I-HkXk)3. Taking a generic matrix operator norm from both sides of (9), we obtain (10)I-HkXk+1I-HkXk3.

On the other hand, Chebyshev’s method for matrix inversion problem is convergent if the initial approximation reads I-HX0<1. That is to say, I-[A-*(I-X)A-1]X0<1. This together with the initial matrix X0=AA* gives (11)A-*XA*<1, which is true when X is the minimal HPD solution of (1).

Note that since (12)0<XS=I-A*XS-1A=A*[A-*A-1-XS-1]A=A*[X0-1-XS-1]A, we obtain that X0-1>XS-1; thus X0<XS. And subsequently using mathematical induction, it would be observed that {Xk} tends to XS.

The only problem that happens in this process is the fact that the convergence order is q-linear. In fact, although Chebyshev’s method for matrix inversion has third local order of convergence, this rate will not be preserved for finding the minimal HPD solution of (1).

The reason is that the matrix H, which we must compute its inverse by Chebyshev’s method, is dependent on the X itself. That is to say, the unknown is located in the essence of the matrix H=A-*(I-X)A-1.

Theorem 5.

The sequence of matrices produced by (6) satisfies the following error inequality: (13)Xk+1-XSλkXk-XS, where λk=-XS-1δk+XkXS2δk+3Yk-XkXS-1Yk-YkXS-1Xk+YkA-*A-1Xk and Yk=XkA-*A-1Xk.

Proof.

First since limkXk=XS and by using (6) we have (14)Xk+1-XS=Xk[3I-A-*(I-Xk)A-1Xk×(3I-A-*(I-Xk)A-1Xk)]-XS=[3Xk-3XkA-*(I-Xk)A-1Xk+XkA-*×(I-Xk)A-1XkA-*(I-Xk)A-1Xk]-XS=[3Xk-3XkA-*((I-XS)-(Xk-XS))A-1Xk+XkA-*((I-XS)-(Xk-XS))A-1XkA-*×((I-XS)-(Xk-XS))A-1Xk]-XS=[3Xk-3XkXS-1Xk+3XkA-*(Xk-XS)A-1Xk+Xk(XS-1Xk)2-XkXS-1XkA-*(Xk-XS)×A-1Xk-XkA-*(Xk-XS)A-1XkXS-1Xk+XkA-*×(Xk-XS)A-1XkA-*(Xk-XS)A-1Xk]-XS=-[XS-3Xk+3XkXS-1Xk-Xk(XS-1Xk)2]+3XkA-*(Xk-XS)A-1Xk+Xk(XS-1Xk)2-XkXS-1XkA-*(Xk-XS)A-1Xk-XkA-*(Xk-XS)A-1XkXS-1Xk+XkA-*×(Xk-XS)A-1XkA-*(Xk-XS)A-1Xk. Note that we have used the fact that (I-XS)=A*XS-1A. Relation (14) yields (15)δk+1=-δkXS-1δk+XkXS2δk2+3XkA-*(δk)A-1Xk-XkXS-1XkA-*δkA-1Xk-XkA-*δkA-1XkXS-1Xk+XkA-*δkA-1XkA-*δkA-1Xk=λkδk, wherein δk=Xk-XS and λk=-XS-1δk+XkXS2δk+3XkA-*A-1Xk-XkXS-1XkA-*A-1Xk-XkA-*A-1XkXS-1Xk + XkA-*A-1XkA-*A-1Xk. We remark that δkA-1Xk=A-1Xkδk.

Consequently, one has the error inequality (13). This shows the q-linear order of convergence for finding the minimal HPD solution of (1). We thus have (16)0<limsupkXk+1-XSXk-XS<1, which is guaranteed since (17)0<limsupkλk<1.

3. Numerical Comparisons

In this section, we mainly investigate the performance of the new method (6) for matrix equation (1). All experiments were run on a Pentium IV computer, using Mathematica 8 . We report the number of required iterations (Iter) for converging. In our implementations, we stop all considered methods when the infinity norm of two successive iterates is less than given tolerance.

Note that recently Zhang in  studied a way to accelerate the beginning of such iterative methods for finding the minimal solution of (1) via applying multiple Newton’s method for matrix inversion. This technique could be given by (18)Xk+1=Xk((t+1)I-tHkXk), for any 1t2. Subsequently, we could improve the behavior of the new method (6) using (18) as provided in Algorithm 1.

<bold>Algorithm 1: </bold>A hybrid method for computing the minimal HPD solution of (<xref ref-type="disp-formula" rid="EEq1">1</xref>).

(1) Choose X0=AA*

(2) for k=0,1,,l, use (18)

(3) end for

(4) set X0=Xl

(5) for k=0,1, until convergence (Xk+1-Xk<ϵ), use (6)

(6) end for

We compare Algorithm 1, denoted by PM, with (2) denoted by M1, (3) denoted by M2, and the method proposed by El-Sayed and Al-Dbiban  denoted by M3, which is a modification of the method presented by Zhan in , as follows: (19)X0=Y0=I,Yk+1=(I-Xk)Yk+I,Xk+1=I-A*Yk+1A.

Example 1 (see [<xref ref-type="bibr" rid="B18">18</xref>]).

In this experiment, we compare the results of different methods for finding the minimal solution of (1) when the matrix A is defined by (20)A=(0.370.130.12-0.300.340.120.11-0.170.29), and the solution is (21)XS=(0.215981-0.09604060.101305-0.09604060.331082-0.1544870.101305-0.1544870.241782).

The results are given in Figure 1(a) in terms of the number of iterations when the stopping criterion is Xk+1-Xk10-8.

Number of iterations against accuracies for experiment 1 (a) and experiment 2 (b).

Note that PM and M2 converge to XS, whereas all the other schemes converge to XL; thus we use other schemes to find the maximal solution of the dual equation D(X)=X+AX-1A*-I in our written codes so as to have fair comparisons. Here t=2 has been chosen for PM (with l=19). This l=19 for the number of iterations in the inner finite loop of Algorithm 1 has been considered in the numerical report.

Furthermore, we have chosen this number empirically. In fact, varying l shows us that we even can obtain better or worse results than the reported ones in different examples.

In Example 1, we have used t=2. In fact we have chosen this value for t since we are solving an operator equation in essence. To be more precise, we wish to consider the solution of the operator equation to be of multiplicity 2. This consideration makes the algorithm converge faster at the initial phase of the process and when we are enough close to the solution, then we flash back to the ordinary methods, that is, treat the solution as a simple zero (solution) of the operator equation.

Example 2 (see [<xref ref-type="bibr" rid="B19">19</xref>]).

Applying the stopping criterion Xk+1-Xk10-12, we compare the behavior of various methods for the following test matrix: (22)A=(0.1-0.15-0.25980760.150.2125-0.06495190.2598076-0.06495190.137), with the solution as (23)XS=(0.112684-0.00001301610.000142799-0.00001301610.0784090.01983070.0001427990.01983070.101129). The results are illustrated in Figure 1(b), wherein t=1.2 has been chosen for PM (with l=1).

4. Conclusions

We have studied the fact that the minimal HPD solution of (1) is equivalent to the roots of a nonlinear map. This special map has been solved by the well-known Chebyshev method as a matrix inversion problem.

The developed method requires the computation of one matrix inverse at the beginning of the process and it is hence an inversion-free method. The convergence and the rate of convergence have been studied for this scheme. Furthermore, using a proper acceleration technique from the literature, we have further speeded up the process of finding the HPD solution of (1).

Conflict of Interests

The authors declare that there is no conflict of interests regarding the publication of this paper.

Acknowledgments

The authors express their sincere thanks to the referees for the careful and details reading of the manuscript and very helpful suggestions that improved the manuscript substantially. The authors also gratefully acknowledge that this research was partially supported by the University Putra Malaysia under the GP-IBT Grant Scheme having project number GP-IBT/2013/9420100.

Engwerda J. C. Ran C. M. A. Rijkeboer A. L. Necessary and sufficient conditions for the existence of a positive definite solution of the matrix equation X+A*X-1A=Q Linear Algebra and its Applications 1993 186 255 275 El-Sayed S. M. Ran A. C. M. On an iteration method for solving a class of nonlinear matrix equations SIAM Journal on Matrix Analysis and Applications 2001/02 23 3 632 645 10.1137/S0895479899345571 MR1896810 2-s2.0-0036018627 Li J. Solutions and improved perturbation analysis for the matrix equation X-A*X-pA=Q(p>0) Abstract and Applied Analysis 2013 2013 12 575964 10.1155/2013/575964 Engwerda J. C. On the existence of a positive definite solution of the matrix equation X+A*X-1A=I Linear Algebra and its Applications 1993 194 91 108 10.1016/0024-3795(93)90115-5 MR1243822 2-s2.0-33846609846 Monsalve M. Raydan M. A new inversion-free method for a rational matrix equation Linear Algebra and Its Applications 2010 433 1 64 71 10.1016/j.laa.2010.02.006 MR2645065 2-s2.0-77950865355 Ullah M. Z. Soleymani F. Al-Fhaid A. S. An efficient matrix iteration for computing weighted Moore-Penrose inverse Applied Mathematics and Computation 2014 226 441 454 10.1016/j.amc.2013.10.046 MR3144324 Soleymani F. Tohidi E. Shateyi S. Haghani F. K. Some matrix iterations for computing matrix sign function Journal of Applied Mathematics 2014 2014 9 425654 10.1155/2014/425654 MR3232916 Soleymani F. Sharifi M. Shateyi S. Haghani F. A class of Steffensen-type iterative methods for nonlinear systems Journal of Applied Mathematics 2014 2014 9 705375 10.1155/2014/705375 MR3198398 Soleymani F. Stanimirović P. S. Shateyi S. Haghani F. K. pproximating the matrix sign function using a novel iterative method Abstract and Applied Analysis 2014 2014 9 105301 10.1155/2014/105301 MR3240521 Soleymani F. Sharifi M. Shateyi S. Khaksar Haghani F. An algorithm for computing geometric mean of two Hermitian positive definite matrices via matrix sign Abstract and Applied Analysis 2014 2014 6 978629 10.1155/2014/978629 Traub J. F. Iterative Methods for the Solution of Equations 1964 Englewood Cliffs, NJ, USA Prentice Hall MR0169356 Iliev A. Kyurkchiev N. Nontrivial Methods in Numerical Analysis: Selected Topics in Numerical Analysis 2010 LAP LAMBERT Academic Publishing Soleymani F. Lotfi T. Bakhtiari P. A multi-step class of iterative methods for nonlinear systems Optimization Letters 2014 8 3 1001 1015 10.1007/s11590-013-0617-6 MR3170582 ZBL1286.93068 2-s2.0-84874392490 Trott M. The Mathematica Guide-Book for Numerics 2006 New York, NY, USA Springer Zhang L. An improved inversion-free method for solving the matrix equation X+AXαA=Q Journal of Computational and Applied Mathematics 2013 253 200 203 El-Sayed S. M. Al-Dbiban A. M. A new inversion free iteration for solving the equation X+A*X-1A=Q Journal of Computational and Applied Mathematics 2005 181 1 148 156 10.1016/j.cam.2004.11.025 MR2146416 2-s2.0-19444385682 Zhan X. Computing the extremal positive definite solutions of a matrix equation SIAM Journal on Scientific Computing 1996 17 5 1167 1174 10.1137/S1064827594277041 MR1404867 ZBL0856.65044 2-s2.0-0000593006 Guo C. H. Lancaster P. Iterative solution of two matrix equations Mathematics of Computation 1999 68 228 1589 1603 10.1090/S0025-5718-99-01122-9 MR1651757 2-s2.0-0033443128 El-Sayed S. M. An algorithm for computing positive definite solutions of the nonlinear matrix equation X+A*X-1A=I International Journal of Computer Mathematics 2003 80 12 1527 1534 10.1080/00207160310001603316 MR2031600 2-s2.0-28344437075