数学杂志  2014, Vol. 34 Issue (6): 1149-1154   PDF    
扩展功能
加入收藏夹
复制引文信息
加入引用管理器
Email Alert
RSS
本文作者相关文章
崔晓梅
刘丽波
高寒
关于矩阵方程X+A*X-αA +B*X-βB=I的正定解
崔晓梅, 刘丽波, 高寒    
吉林化工学院理学院, 吉林 吉林 132022
摘要:本文研究了矩阵方程X + A*X-αA + B*X-βB=I在α, β∈(0, 1] 时的正定解.利用单调有界极限存在准则, 构造三种迭代算法, 获得了方程的正定解, 拓宽了此类方程的求解方法.数值算例说明算法的可行性.
关键词矩阵方程    正定解    迭代方法    
ON THE POSITIVE DEFINITE SOLUTION OF MATRIX EQUATION X + A*X- αA + B*X-βB = I
CUI Xiao-mei, LIU Li-bo, GAO Han    
College of Sciences, Jilin Institute of Chemical Technology, Jilin 132022, China
Abstract: The positive definite solutions of the matrix equation X + A*X- αA + B*X-βB = I, α, β∈(0, 1] are investigated in this paper. By using monotone bounded limit existence criteria, three iterative algorithms are constructed to obtain the positive definite solution which widen the solution of such equation. Numerical examples are given to illustrate the effectiveness of the methods.
Key words: matrix equation     positive definite solution     iterative methods    
1 引言

本文研究了非线性矩阵方程

$X+A^*X^{-\alpha}A+B^*X^{-\beta}B=I.$ (1)

$\alpha, \beta\in(0, 1]$的Hermite正定解, 其中$A, B$$n$阶非奇异复方阵, $A^*$$A$的共轭转置, $I$$n$阶单位阵.

形如

$X+ A^*X^{-\alpha}A=Q, X-A^*X^{-\alpha}A=Q, $ (2)

其中$Q$是正定矩阵.当$\alpha=1$时, 方程(2) 产生于控制理论, 动态规划, 随机过滤和统计学等领域[1-2].许多作者[3-4]针对方程(2) 中参数$\alpha$取不同值, 研究了解的存在性, 收敛速度, 性质及迭代算法等.近年来较复杂方程形式受到人们关注, 杜仲复[5]从矩阵分解角度研究了$X-A^*X^{-\alpha}A-B^*X^{-\beta}B=I$, 当$\alpha, \beta\in(0, 1]$时, 方程解存在的若干条件及迭代求解算法; 龙建辉, 何优美, 廖安平[6-8]等研究形如$X+\sum\limits_{i=1}^{m}A_{i}^*X^{-n}A_i=I$的方程, 其中文献[6]是本文当$\alpha=1, \beta=1$的特殊情况; Sarhan, 段雪峰[9-11]等对更为复杂的形式$X^r+\sum\limits_{i=1}^{m}A_i^*X^{\delta _i}A_i=Q$也取得了较好的理论成果.

$\|A\|$表示$A$的谱范数; $A>0(A\geq 0)$表示$A$是正定$($$)$正定矩阵, $A>B(A\geq B)$表示$A-B$是正定$($$)$正定矩阵.

2 主要结论

引理2.1 [12] 若$A>B>0(A\geq B>0)$, 则$A^\alpha>B^\alpha(A^\alpha\geq B^\alpha>0)$$\alpha\in(0, 1], $$A^\alpha<B^\alpha(0<A^\alpha\leq B^\alpha)$$\alpha\in[-1, 0).$

引理2.2 若方程(1) 有正定解$X$, 则$X\leq I$

$\max \left\{(AA^*)^{\frac{1}{\alpha}}, (BB^*)^{\frac{1}{\beta}}\right\}<X\leq I-A^*A-B^*B.$

 显然$X\leq I$.由引理2.1知$X^{-\alpha}\geq I, X^{-\beta}\geq I$, 则

$X=I-A^*X^{-\alpha}A-B^*X^{-\beta}B\leq I-A^*A-B^*B.$

$0<A^*X^{-\alpha}A<I$, 则

$0<A^*X^{\frac{-\alpha}{2}}X^{\frac{-\alpha}{2}}A<I, $

因此

$0<X^{\frac{-\alpha}{2}}AA^*X^{\frac{-\alpha}{2}}<I, $

进而

$0<AA^*<X^\alpha, $

再由引理2.1得

$(AA^*)^{\frac{1}{\alpha}}<X.$

同理可知$(BB^*)^{\frac{1}{\beta}}<X$.得证.

定理2.3 方程(1) 有正定解的充要条件是存在非奇异阵$W, Y, Z\in C^{n\times n}$使得$(W^T, Y^T, Z^T)^{T}$为列正交阵, 且$A=(W^*W)^{\alpha/2}Y$, $B=(W^*W)^{\beta/2}Z$, 此时$X=W^*W$为方程的解, 且所有的解可写成此形式, 进而可令$W$为三角阵.

 必要性 若方程有正定解$X$, 则$X=W^*W$, $W$为非奇异方阵, 特别的可选择Cholesky分解, $W$可为三角阵, 重写方程如下

$W^*W+A^*(W^*W)^{-\alpha}A+B^*(W^*W)^{-\beta}B=I, $

$ W^*W+A^*(W^*W)^{-\alpha/2}(W^*W)^{-\alpha/2}A+B^*(W^*W)^{-\beta/2}(W^*W)^{-\beta/2}B=I, \\ W^*W+A^*(W^*W)^{*-\alpha/2}(W^*W)^{-\alpha/2}A+B^*(W^*W)^{*-\beta/2}(W^*W)^{-\beta/2}B=I, \\ \begin{pmatrix}W\\(W^*W)^{-\alpha/2}A\\(W^*W)^{-\beta/2}B\end{pmatrix}^* \begin{pmatrix}W\\(W^*W)^{-\alpha/2}A\\(W^*W)^{-\beta/2}B\end{pmatrix} =I. $

$Y=(W^*W)^{-\alpha/2}A$, $Z=(W^*W)^{-\beta/2}B$, 则$A=(W^*W)^{\alpha/2}Y$, $B=(W^*W)^{\beta/2}Z$.

充分性 将$A, B, X$代入方程(1) 的左端即可验证$X=W^*W$为方程的解.

算法2.4

$\left\{\begin{aligned}&\quad X_0=Y_0=I, \\&Y_{n+1}=(I-X_n)Y_n+I, \\&X_{n+1}=I-A^*Y_{n+1}^{\alpha}A-B^*Y_{n+1}^{\beta} B \quad n=0, 1, 2, \cdots.\end{aligned}\right.$

定理2.5 若方程(1) 有一正定解, 且序列$\{X_n\}, \{Y_n\}$由算法$(2.4)$决定, 则$\{X_n\}$单调下降且收敛于正定解$X$.

 往证

$I=X_0\geq X_1\geq X_2\geq\cdots\geq X_n\geq X , I=Y_0\leq Y_1\leq Y_2\leq\cdots\leq Y_n\leq X^{-1}.$ (3)

$I\geq I-A^*A-B^*B\geq I-A^*X^{-\alpha}A-B^*X^{-\beta}B, $

$X_0\geq X_1\geq X.$

$\{Y_n\}$,

$I=Y_0=Y_1\leq X^{-1}.$

则不等式(3) 对$n=0, 1$成立.假设(3) 对$n=k$时成立, 即

$I=X_0\geq X_1\geq X_2\geq\cdots\geq X_k\geq X , I=Y_0\leq Y_1\leq Y_2\leq\cdots\leq Y_k\leq X^{-1}.$

如下证明式(3) 当$n=k+1$时成立, 则

$Y_{k+1}=(I-X_k)Y_k+I\geq(I-X_{k-1})Y_{k-1}+I=Y_k, $

亦有

$Y_{k+1}=(I-X_k)Y_k+I\leq(I-X)X^{-1}+I=X^{-1}, $

$Y_k\leq Y_{k+1}\leq X^{-1}.$

考虑$X_k$

$X_k-X_{k+1}=A^*(Y_{k+1}^{\alpha}-Y_k^{\alpha})A+B^*(Y_{k+1}^{\beta}-Y_k^{\beta})B, $

$Y_{k+1}\geq Y_k, $由引理2.1知$Y_{k+1}^{\alpha}> Y_k^{\alpha}$, $Y_{k+1}^{\beta}> Y_k^{\beta}$, 因此$X_k\geq X_{k+1}$.由

$X_{k+1}=I-A^*Y_{k+1}^{\alpha}A-B^*Y_{k+1}^{\beta}B\geq I-A^*X^{-\alpha}A-B^*X^{-\beta}B=X, $

$X_k\geq X_{k+1}\geq X.$

故当$n=k+1$时不等式(3) 也成立, 且$\lim\limits_{n\rightarrow\infty}X_n$$\lim\limits_{n\rightarrow\infty}Y_n$存在.对算法(2.4) 取极限得到$ Y=X^{-1}$

$X=I-A^*X^{-\alpha}A-B^*X^{-\beta}B.$

进而因对任意$n$$X_n\geq X$, 则$X$为所求正定解.

算法2.6

$\left\{\begin{aligned}&\quad X_0=\delta I, \\& X_{n+1}=I-A^*X_n^{-\alpha}A-B^*X_n^{-\beta}B, n=0, 1, 2, \cdots. \end{aligned}\right.$

定理2.7 若方程(1) 有一正定解, 且序列$\{X_n\}$由算法(2.6) 决定, 则当$\delta>1$时, $\{X_n\}$单调递减有下界; 当$0<\delta<1$$I-\delta^{-\alpha}A^*A-\delta^{-\beta}B^*B>\delta I$时, $\{X_n\}$单调递增有上界, $\{X_n\}$收敛于方程正定解$X$.

 对$\delta>1$, $0<\delta<1$两种情况证明.

情况1 当$\delta>1$时, 由迭代序列知$X_0=\delta I$,

$X_1=I-A^*X_0^{-\alpha}A-B^*X_0^{-\beta}B=I-\delta^{-\alpha}A^*A-\delta^{-\beta}B^*B<X_0, $

假设$X_k<X_{k-1}$, 往证$X_{k+1}<X_k$.

$X_{k+1}=I-A^*X_k^{-\alpha}A-B^*X_k^{-\beta}B<I-A^*X_{k-1}^{-\alpha}A-B^*X_{k-1}^{-\beta}B=X_k, $

由此知序列$\{X_n\}$单调递减.

如下证明$\{X_n\}$有下界. $n=0$时显然成立

$\begin{aligned}X_0-X&=\delta I-(I-A^*X^{-\alpha}A-B^*X^{-\beta}B)>0, \\X_1-X&=(I-A^*X_0^{-\alpha}A-B^*X_0^{-\beta}B)-(I-A^*X^{-\alpha}A-B^*X^{-\beta}B)\\&=A^*(X^{-\alpha}-X_0^{-\alpha})A+B^*(X^{-\beta}-X_0^{-\beta})B>0.\end{aligned}$

假设$n=k$时, 有$X_k>X$, 往证$n=k+1$时, $X_{k+1}>X$.

$\begin{aligned}X_{k+1}-X&=(I-A^*X_k^{-\alpha}A-B^*X_k^{-\beta}B)-(I-A^*X^{-\alpha}A-B^*X^{-\beta}B)\\&=A^*(X^{-\alpha}-X_k^{-\alpha})A+B^*(X^{-\beta}-X_k^{-\beta})B>0.\end{aligned}$

因此$\{X_n\}$是单调递减并以某正定阵$X$为下界的矩阵序列, $\{X_n\}$收敛于方程正定解$X$.

情况2$0<\delta<1$$I-\delta^{-\alpha}A^*A-\delta^{-\beta}B^*B>\delta I$时, 证明类似于$\delta>1$情况.

对方程(1) 考虑如下, 令$Y=X^{-1}$, 则方程(1) 有解$X$等价于下式有解$Y$:

$Y=I+Y^{1/2}A^*Y^{\alpha}AY^{1/2}+Y^{1/2}B^*Y^{\beta}BY^{1/2}.$

算法2.8

$\left\{\begin{aligned}&Y_0=0, \\&Y_n=I+Y_{n-1}^{1/2}A^*Y_{n-1}^{\alpha}AY_{n-1}^{1/2}+Y_{n-1}^{1/2}B^*Y_{n-1}^{\beta}BY_{n-1}^{1/2}, n=0, 1, 2, \cdots. \end{aligned}\right.$

定理2.9 设$0<\|A\|^2, \|B\|^2<\frac{2}{27}$, 则由算法(2.8) 决定的迭代序列收敛到方程(1) 的正定解$X^{-1}$, 且$X\in[\frac{2}{3}I, I]$.

 根据文[5, Theorem 8], 由算法2.8决定的迭代序列收敛到方程(1) 的唯一解$X=Y^{-1}\in[\frac{2}{3}I, \frac{3}{2}I]$, 再由引理2.2知$X\in[\frac{2}{3}I, I]$.

3 数值算例

如下数值例子对本文给出的迭代方法进行说明.所有的结果都在matlab7.10.0中运行得到.设残差

$R(X)=\| X+A^*X^{-\alpha}A+B^*X^{-\beta}B-I\|_\infty.$

实验停机的条件设为

$R(X)\leq1.0\times10^{-10}.$

例3.1 考虑方程(1) 其中

$\begin{equation*} A=\left( \begin{array}{ccc} 0.0100&-0.1500&-0.2590\\ 0.0150&0.2120&0.0640 \\ 0.0250&-0.0690&0.1380 \\ \end{array} \right), B=\left( \begin{array}{ccc} 0.1600&-0.0250&0.0200 \\ 0.0250&-0.2880&-0.0600 \\ 0.0040&-0.0160&-0.1200 \\ \end{array} \right). \end{equation*}$

任意取两组参数进行数值计算比较.取$\alpha=0.5$, $\beta=0.2, $分别用三种算法解得

$\begin{equation*} X=\left( \begin{array}{ccc} 0.9726&0.0112&-0.0034 \\ 0.0112&0.8366&-0.0655 \\ -0.0034&-0.0655&0.8873 \\ \end{array} \right). \end{equation*}$

$\alpha=0.8$, $\beta=0.4, $分别用三种算法解得

$\begin{equation*} X=\left( \begin{array}{ccc} 0.9723&0.0111&-0.0036 \\ 0.0111&0.8300&-0.0689 \\ -0.0036&-0.0689&0.8841 \\ \end{array} \right). \end{equation*}$

由以上结果可知, 本文的三种算法较准确的给出了方程的正定解. 表 1, 表 2对以上两组参数三种算法计算的结果进行了比较.

表 1 $\alpha=0.5, \beta=0.2$三种算法结果比较

表 2 $\alpha=0.8, \beta=0.4$三种算法结果比较

通过此例说明:三种算法可行有效.由表 1, 表 2可知算法2.4及算法2.8迭代次数相近, 算法2.6迭代较快, 优于前两种算法.

参考文献
[1] Anderson W N, Morley T D, Trapp G E. Positive solutions to $X=A-BX^{-1}B^*$[J]. Linear Algebra Appl., 1990, 134: 53–62. DOI:10.1016/0024-3795(90)90005-W
[2] Engwerda J C. On the existence of the positive deflnite solution of the matrix equation $X+A^TX^{-1}A=I$[J]. Linear Algebra Appl., 1993, 194: 91–108. DOI:10.1016/0024-3795(93)90115-5
[3] Vejdi I Hassanov. Positive deflnite solutions of the matrix equations $X ± A^*X^{-q}A=Q.*$[J]. Linear Algebra Appl., 2005, 404: 166–182. DOI:10.1016/j.laa.2005.02.024
[4] Wang Xingtao, Li Yuanming. On equations that are equivalent to the nonlinear matrix equation $X+A^*X^{-\alpha }A=Q$[J]. J. Comput. Appl. Math., 2010, 234: 2441–2449. DOI:10.1016/j.cam.2010.03.004
[5] 杜仲复. 矩阵方程$X-A^*X^{-\alpha }A-B^*X^{-\beta}B=I$[J]. 吉林大学学报(理学版), 2010, 1(48): 26–32.
[6] Long Jianhui, Hu Xiyan, Zhang Lei. On the Hermitian positive deflnite solution of the nonlinear matrix equation $X+A^*X^{-1}A+B^*X^{-1}B=I$[J]. Bull Braz. Math. Soc. New. Series, 2008, 39(3): 371–386. DOI:10.1007/s00574-008-0011-7
[7] He Youmei, Long Jianhui. On the Hermitian positive deflnite solution of the nonlinear matrix equation $X+\sum\limits_{i=1}^{m}A_{i}^*X^{-1}A_i=I$[J]. Appl. Math. Comput., 2010, 216: 3480–3485.
[8] 廖安平, 黄叶楠, 沈 金荣. 矩阵方程的正定解$X+\sum\limits_{i=1}^{m}A_{i}^*X^{-n}A_i=I$[J]. 长沙大学报 , 2009, 23(2): 1–4.
[9] Sarhan A M, Naglaa M El- Shazly, Enas M Shehata. On the existence of extremal positive deflnite solutions of the nonlinear matrix equation $X.r+\sum\limits_{i=1}^{m}A_i^*X^{\delta _i}A_i=I$[J]. Math. Computer Modelling, 2010, 51: 1107–1117. DOI:10.1016/j.mcm.2009.12.021
[10] Duan Xuefeng, Liao Anping, Tang Bin. On the nonlinear matrix equation $X-\sum\limits_{i=1}^{m}A_{i}^*X^{\delta_{i}}A_i=Q$[J]. Linear Algebra Appl., 2008, 429: 110–121. DOI:10.1016/j.laa.2008.02.014
[11] Duan Xuefeng, Liao Anping. On Hermitian positive deflnite solution of the matrix equation $X-\sum\limits_{i=1}^{m}A_{i}^*X^{r}A_i=Q$[J]. Journal of Computational and Applied Mathematics, 2009, 229: 27–36. DOI:10.1016/j.cam.2008.10.018
[12] Parodi M. La localisation des valeurs caracterisiques des matrices et ses application[M]. Paris: Gauthiervillars, 1959.