Hindawi Publishing Corporation Journal of Mathematics Volume 2013, Article ID 207176, 8 pages http://dx.doi.org/10.1155/2013/207176
Research Article The Inverses of Block Toeplitz Matrices
Xiao-Guang Lv and Ting-Zhu Huang
School of Mathematical Sciences, University of Electronic Science and Technology of China, Chengdu, Sichuan 611731, China
Correspondence should be addressed to Xiao-Guang Lv; [email protected]
Received 30 December 2012; Accepted 26 March 2013
Academic Editor: Peter Grabner
Copyright Β© 2013 X.-G. Lv and T.-Z. Huang. 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.
We study the inverses of block Toeplitz matrices based on the analysis of the block cyclic displacement. New formulas for the inverses of block Toeplitz matrices are proposed. We show that the inverses of block Toeplitz matrices can be decomposed as a sum of products of block circulant matrices. In the scalar case, the inverse formulas are proved to be numerically forward stable, if the Toeplitz matrix is nonsingular and well conditioned.
1. Introduction is nonzero, then the first and the last columns of the inverse β1 of the Toeplitz matrix are sufficient to reconstruct π .In[6], Let an inverse formula can be obtained by the solutions of two π0 πβ1 β β β π1βπ equations (the so-called fundamental equations), where each [ ] [ π1 π0 d π2βπ] right-hand side of them is a shifted column of the Toeplitz π=[ . . ] (1) [ . dd . ] matrix. Later, Ben-Artzi and Shalom [7], Labahn and Shalom . . [8], Huckle [9], Ng et al. [10], and Heinig [11]havestudied [ππβ1 ππβ2 β β β π0 ] the Toeplitz matrix inverse formulas when the (1, 1)st entry be an πΓπblock Toeplitz matrix with blocks of size πΓπ. of the inverse of a Toeplitz matrix is zero. In [12], Cabay π We use the shorthand π=(ππβπ)π,π=1 for a block Toeplitz and Meleshko presented an efficient algorithm (NPADE) for matrix. The block Toeplitz systems ππ₯ =π arise in a numerically computing PadeapproximantsinaweaklystableΒ΄ variety of applications in mathematics, scientific computing, fashion. As an application of NPADE, it has been shown that and engineering, for instance, image restoration problems itcanbeusedtocomputestably,inaweaksense,theinverse in image processing, numerical differential equations and of a Hankel or Toeplitz matrix. integral equations, time series analysis, and control theory [1β When π>1, additional problems are encountered in 3]. If we want to solve more than one block Toeplitz linear obtaining the inverse formula of a block Toeplitz matrix. A system with the same coefficient matrix, then we usually well-known formula of Gohberg and Heinig can construct β1 solvefourorsospecialblocklinearsystemsinorderto π ,providedthatthefirstandlastcolumnstogetherwith determine the block Toeplitz inverse formula that expresses the first and last rows of the inverse are known13 [ ]. In [14], β1 π as the sum of products of block upper and block lower a set of new formulas for the inverse of a block Hankel Toeplitz matrices. For example, Van Barel and Bultheel [4] or block Toeplitz matrix is given by Labahn et al. The gave an inverse formula for a block Toeplitz matrix and then formulas are expressed in terms of certain matrix Padeforms,Β΄ derived a weakly stable algorithm to solve a block Toeplitz which approximate a matrix power series associated with system of linear equations. The special structure of block the block Hankel matrix. We refer the reader to [15, 16]for Toeplitz matrices has resulted in some closed formulas for the computation of Pade-HermiteΒ΄ and simultaneous PadeΒ΄ their inverses. systems in detail. In [7], Ben-Artzi and Shalom have proved In the scalar case, Gohberg and Semencul [5]haveshown that each inverse of a Toeplitz matrix can be constructed that if the (1, 1)st entry of the inverse of a Toeplitz matrix π via three of its columns, and thus, a parametrization of the 2 Journal of Mathematics set of inverses of Toeplitz matrices is obtained. Then they equations with the same coefficient matrix π,wedecompose generalized these results to block Toeplitz matrices; see [17]. the inverses of block Toeplitz matrices as a sum of products In [18], Gemignani has shown that the representation of of block circulant matrices. From the point of view of the β1 π relies upon a strong structure-preserving property of computation, the formulas involving block circulants instead the Schur complements of the nonsingular leading principal of block upper or lower Toeplitz matrices are more attractive. submatrices of a certain generalized Bezoutian of matrix For example, in the scalar case, matrix-vector product by polynomials. a circulant matrix is roughly twice as fast as matrix-vector In this paper, we focus our attention to the inverses of product by a triangular Toeplitz matrix of the same size. block Toeplitz matrices with the help of the block cyclic In fact, the efficient multiplication of a Toeplitz matrix and displacement. In [19], Ammar and Gader have shown that a vector is achieved by embedding the Toeplitz matrix in the inverse of a Toeplitz matrix can be represented as sums of a circulant matrix of twice the size. The derivation of the products of lower triangular Toeplitz matrices and circulant formulas we present is based on the factor cyclic displacement matrices. The derivation of their results is based on the of block matrices. Later on, we give some corollaries about the idea of cyclic displacement structure. In [20], Gohberg and representations of the inverse matrices. The results generalize Olshevsky also obtained new formulas for representation of theGohberg-Olshevsky,Heinig-Rost,andBen-Artzi-Shalom matrices and their inverses in the form of sums of products of formulas. In the scalar case, we consider the stability of factor circulant, which are based on the analysis of the factor theToeplitzmatrixinverseformulas.Weshowthatifthe cyclic displacement of matrices. The results in applications to Toeplitz matrix is nonsingular and well conditioned, the Toeplitz matrices generalized the Gohberg-Semencul, Ben- inverse formulas are numerically forward stable. Artzi-Shalom, and Heinig-Rost formulas. Motivated by a We summarize briefly the content of our paper. Section 2 number of related results on Toepltiz inverse formulas, we reviews some fundamental concepts such as the factor cyclic study the representation of the inverses of block Toeplitz displacement structure of a block matrix and every square matrices. Since block Toeplitz matrices have similar displace- block matrix can be written as a sum of products of block ment structure as Toeplitz matrices, all results about Toeplitz circulant matrices. In Section 3, we decompose the inverses matrices extend quite naturally to block Toeplitz matrices. of block Toeplitz matrices as a sum of products of block At first, for purposes of presentation, we adopt some circulant matrices by solving four equations with the same notations that will be used throughout the paper. We denote coefficient matrix π. Some corollaries are also obtained in the πΓπidentity matrix by πΌπ. πΈπ (π=0,1,...,πβ1)means this section. Wediscuss the stability of the inverse formulas of amatrixofdimensionππ Γπ in which the (π + 1)-th block scalar Toeplitz matrices in Section 4. Some remarks are given is the identity matrix πΌπ, and the remaining blocks are all in Section 5. zeros. π½ is called an πΓπblock reverse identity matrix, if π½ has the identity matrices on the block antidiagonal and zero elsewhere. Let π΄=(π΄ππ ) be a block matrix or a block vector 2. The Factor Block Cyclic Displacement πΓπ π΄σΈ =(π΄ ) with block size of . ππ is said to be the block The inverse formulas derived in the next section depend on π΄ πΆ(π, π) π transpose of .By we denote the block -circulant some results of the factor block cyclic displacement of a block π=[π,...,π ]σΈ with the first column 0 πβ1 with a block size of matrix. The focus of this section lies upon the definitions and πΓπ;thatis,thematrixofformisdefinedby the properties of the factor block cyclic displacement of a block matrix. We start with the definition of the factor block π ππ β β β β β β ππ 0 πβ1 1 cyclic displacement. [ . ] [ . ] Let ππ be the block π-cyclic lower shift matrix defined by [ π1 π0 πππβ1 d . ] [ ] [ . . ] . [ . π1 dd . ] (2) [ ] 0 β β β β β β 0 ππΌ [ . ] . dddπππβ1 [ ] [πΌ0dd 0 ] [ππβ1 β β β β β β π1 π0 ] [ ] [ . ] [0πΌdd . ] . (3) In this work, new formulas for the inverses of block [. . ] [. ddd . ] Toeplitz matrices are proposed. As is well known, any block . . 0β β β 0 πΌ 0 Toeplitz matrix π has a property of block persymmetry, that [ ] σΈ is, π½π π½=π,whereπ½ is the block reverse identity matrix. Unfortunately, the matrix inverse does not hold this property. The block π-cyclic displacement of an πΓπblock matrix π΄ This can explain why the Gohberg-HeΛΔ±nig formula [13] needs with blocks size of πΓπis defined as the first and last columns together with the first and last rows of the inverse for constructing the inverse. For a block π π π σΈ Toeplitz matrix ,let π and π be its block circulant and βπ (π΄) =π΄βπππ΄π1/π,π=0.ΜΈ (4) block skew-circulant parts, respectively. The representation π=ππ β πΌ+ππ β πΌshows that there will exist block circulant matrices π΄π (π = 1, 2) and π΅π (π = 1, 2) such that The number πΌ=rank β1(π΄) is referred to as block π-cyclic β1 π =π΄1π΅1 +π΄2π΅2.Inthispaper,bysolvingfourlinear displacement rank of the block matrix π΄.Sinceallblock Journal of Mathematics 3
Toeplitz matrices have a special structure, it is not difficult Proof. From computing the block π-cyclic displacement of to obtain that matrix π΄ defined by (4), it then follows that
ππβ1 βπππβ1 β β β π1βπ βππ1 β (π΄) =β (πΆ ( , π)) [ 1 ] π π lr [ π β π π β β β π ] [ 1 1βπ ] πΌ σΈ [ π ] π 1 β (π) = [ ] + β ( β πΆ(π ,π)β πΆ(β , ) ) π [ . . . ] πβπ π π π π [ . . . ] π=1 (10) [ 1 ] π β π π β β β π πΌ σΈ πβ1 π β1 π 1 [ ] = β βπ (πΆ π(π ,π)β πΆ(βπ, ) ). πβππ=1 π ππ βπ σΈ [ 1 ] π [ π β π ] π π΄ [ ] [ 1 1βπ ] Note that the -cyclic displacement of any block matrix [πβ1 βπππβ1] [ π ] π =πΈ β [ ] + [ ] β πΈσΈ , is a linear operator and two block -circulant matrices can 0 [ . ] [ . ] 0 . [ . ] commute; we obtain [ 1 ] [ π1βπ βππ1 ] πΌ σΈ ππβ1 β πβ1 π 1 [ π ] β (π΄) = β β (πΆ (π ,π))β πΆ(β , ) . π πβπ π π π π (11) (5) π=1 π π where ππ may be an arbitrary πΓπmatrix. Therefore, the π- By computing the block -cyclic displacement of block - πΆ(π ,π) π cyclic displacement rank of a block Toeplitz matrix is at most cyclic matrix π for all ,wehave 2π. π πΌ 1 σΈ By straightforward computations, it is easy to see that the β (π΄) = β β (πΆ (π ,π))β πΆ(β , ) β (π΄) = 0 π΄ π π π π π equality π holds if and only if is a block -circulant πβππ=1 π matrix. For any invertible block matrix π΄ there exists a simple π πΌ 1 σΈ interrelation between the block -cyclic displacement of the = β π β πΈσΈ β πΆ(β , ) π΄β1 π π΄ π 0 π (12) inverse matrix and the block -cyclic displacement of . π=1 π This connection is given by πΌ σΈ σΈ 1 β1 β1 β1 σΈ β β πΈ0 β πΜπ β πΆ(βπ, ) , βπ (π΄ )=π΄ βπ (π΄) πππ΄ π . 1/π (6) π=1 π
From the previous equation, we know that the block π-cyclic πΜσΈ π = 1,...,πΌ β1 where π for are the first row of the displacement rank of the inverse matrix π΄ isthesameas σΈ block π-circulant matrices πΆ(ππ,π),respectively.SinceπΈ0 β that of π΄.Moreover,thepreviousequationalsoallowsusto σΈ σΈ πΆ(βπ,1/π) =β , it is easy to observe that the sum of the make use of the block π-cyclic displacement technique for the π πΌ βπΌ π βσΈ calculation of the inverse matrix. We will study the inverses first terms in the previous equality is equal to π=1 π π. of block Toeplitz matrices with the help of the block cyclic In the following, we need to show that the sum of the last πΌ terms in the previous equality is equal to the zero matrix. displacement. We need the following result. π β1 σΈ Using the fact that ππ =ππΌand ππ =π1/π,wecheck Theorem 1. If the block π-cyclic displacement of π΄ is given as πβ1 the sum π β1 β1 β1 βπ βππ (π βπππ ππ )ππ =0. (13) πΌ σΈ σΈ π=0 π΄βπππ΄π1/π = β ππβπ, (7) π=1 Thus we obtain where ππ and βπ (π = 1,...,π) are πΓ1block vectors with πβ1 π β1 β1 β1 βπ blocks of size πΓπ,then βππ (π βπππ ππ )ππ π=0 π πΌ 1 σΈ π΄= ( ,π)+ ( β (π ,π)β (β , ) ) πβ1 πΌ C lr πβπ C π C π π (8) π σΈ βπ π=1 = β β ππππβπππ π=0 π=1 (14) or πΌ 1 σΈ πΌ σΈ = β πΆ(π ,π)β πΆ(β , ) π 1 π π π π΄=C (ππ, π) + ( β C (ππ,π)β C(βπ, ) ), (9) π=1 πβπ π=1 π =0. where C(ππ, π) is a block π-circulant matrix with the same last π΄ (ππ, π) π πΌ σΈ σΈ πΌ block row as that of and C is the block -circulant By comparing βπ=1 πΈ0 β πΜπ β πΆ(βπ,1/π) with βπ=1 πΆ(ππ,π)β π΄ σΈ πΌ σΈ σΈ matrix with the same last block column as that of . πΆ(βπ,1/π) ,weobtainβπ=1 πΈ0 β πΜπ β πΆ(βπ,1/π) =0. 4 Journal of Mathematics
Finally, we discuss why πΆ(lr,π) is the block π-circulant Substituting (16)and(18) into the previous equality and with the same last block rows as the matrix π΄.Weknowthat computing, we obtain the last block row of the matrix πΆ(ππ,π) is independent of πΌ the number π.Inviewofβπ=1 πΆ(ππ,π)β πΆ(βπ,1/π)=0,the π₯0 last block rows of the matrix π΄ and C (lr,π)are the same. [ ] [ π₯1 ] Equation (8) is now completely proved. A similar argu- β (πβ1)=π[ ] π [ . ] ment shows that (9)canbeobtainedfrom(7). . [π₯πβ1] 3. Inverses of Block Toeplitz Matrices 1 1 β1 σΈ β (ππβ1 β πβ1,...,π1 β π1βπ,ππ)π π1/π In this section, we focus our attention to new formulas on the π π inverses of block Toeplitz matrices. It is easy to find that any π’0 block Toeplitz matrix has a property of block persymmetry, [ ] π½πσΈ π½=π π½ [ π’1 ] that is, ,where is the block reverse identity matrix. βπ[ ] β (0,...,0,πΌ )β πβ1πσΈ [ . ] π 1/π Unfortunately, the matrix inverse does not have this property. . Λ This can explain why the Gohberg-HeΔ±nig formula [13] needs [π’πβ1] the first and last block columns together with the first and last blockrowsoftheinversetoconstructit.UnliketheGohberg- π₯0 [ ] Semencul formula [5], it only needs the first and the last [ π₯1 ] (1, 1) = [ ] (π’Μ ,ππ’Μ ,...,ππ’Μ ) columnsoftheinverseifthe st entry of the inverse is [ . ] πβ1 0 πβ2 nonzero. . π₯ Let [ πβ1] σΈ π(π₯ ,π₯ ,...,π₯ ) =πΈ, (15) π’0 0 1 πβ1 0 [ ] [ π’1 ] (π₯Μ , π₯Μ ,...,π₯Μ )π=πΈσΈ , β [ ] (π₯Μ ,ππ₯Μ ,...,ππ₯Μ ). 0 1 πβ1 πβ1 (16) [ . ] πβ1 0 πβ2 . σΈ [π’πβ1] π(π’0,π’1,...,π’πβ1) (21) 1 1 σΈ (17) =(ππ,π1 β π1βπ,...,ππβ1 β πβ1) , π π Theorem 2. Let π be a block Toeplitz matrix given by (1).If π =0ΜΈ 1 1 (15)β(18) are solvable for some and an arbitrary matrix (π’Μ , π’Μ ,...,π’Μ ) π=(π β π ,...,π β π ,π ) . π π 0 1 πβ1 πβ1 π β1 1 π 1βπ π π of order ,then (18) (a) π is invertible; For convenience, we will write π₯, π₯,Μ π’ and π’Μ as β1 σΈ σΈ (b) π can be denoted as π₯=(π₯0,π₯1,...,π₯πβ1) , π₯=(Μ π₯Μ0, π₯Μ1,...,π₯Μπβ1) , (19) π’=(π’,π’ ,...,π’ )σΈ , π’=(Μ π’Μ , π’Μ ,...,π’Μ )σΈ . π 0 1 πβ1 0 1 πβ1 πβ1 = ( (π₯, π) β (π½π’,Μ π) πβπ C C From (5), (6), (15), and (17), by simple computations we (22) obtain the following equality: πβπ βC (π’ β πΈ0,π)β C (π½π₯,Μ π)) , β1 β1 β1 σΈ π βπ (π )= βπ βπ (π) πππ π1/π
π₯0 [ ] [ π₯1 ] π( =π)ΜΈ =β[ ] where is an arbitrary number. [ . ] . Proof. (a) Assume that there exists nonzero vector π of [π₯πβ1] dimension ππ such that ππ =0.Thenbycomputing(16) (βπ ,π βππ ,...,π βππ)π π=0 β (βππ,πβ1 βπππβ1,...,π1βπ βππ1) (20) and (18), we have π β1 πβ1 1βπ 1 π and (πΌπ,0,...,0)πππ=0. These imply βπ(π)πππ=0. β π πβ1πσΈ σΈ π 1/π Since βπ(π) = ππ βπ ππ1/π,weobtainππππ=0.A π similar argument shows that ππππ = 0 (π= 2,3,...,πβ1). π’0 [ ] πππ π = 0 (π= 0,1,...,πβ1) [ π’1 ] Premultiplying the equalities π β [ ] β (πΌ ,0,...,0)β π πβ1πσΈ . (π₯Μ , π₯Μ ,...,π₯Μ ) [ . ] π π 1/π by 0 1 πβ1 , we can conclude that all elements of the . vector are equal to zero. Therefore, the invertibility of π is now π’ [ πβ1] proved. Journal of Mathematics 5
(b) According to Theorem 1,weget writtenasasumofproductsofblockskew-circulantmatrices and block circulant matrices. σΈ β1 π 1 π =πΆ(lr,π)+ (πΆ (π₯, π) 1/πβ πΆ(ππ π’,Μ ) πβπ π π₯ βπ₯ β β β βπ₯ { 0 πβ1 1 {[ ] σΈ {[ . ] 1 1 { π₯1 π₯0 d . Μ β1 [ ] +πΆ (π’, π) 1/πβ πΆ(ππ π₯, ) ). π = {[ . ] π 2 {[ . ddβπ₯ ] {[ . πβ1] (23) { . {[π₯πβ1 . π₯1 π₯0 ] πΆ( ,π) π lr is the block -circulant matrix with the same last π’Μ π’Μ β β β π’Μ πβ1 πβ1 0 πβ2 block row as that of .Itisnotdifficulttoseethat [ ] πΆ( ,π) = πΆ(ππ π₯,Μ 1/π)σΈ =πΆ(π½π₯,Μ π) [ . ] lr 1/π . After computations [π’Μπβ2 π’Μπβ1 d . ] and arrangement, we obtain the formula (22)ontheinverses Γ [ . ] [ . ddπ’Μ ] of block Toeplitz matrices. [ . 0 ] . [ π’Μ0 . π’Μπβ2 π’Μπβ1] Remark 3. In fact, the inverse formula (22)isageneralization (25) of the formula of Gohberg and Olshevsky [20], where they π’0 β2πΌπ βπ’πβ1 β β β βπ’1 consider scalar Toeplitz matrices. Since the matrix inverse [ ] [ . ] does not hold the property of persymmetry, in the block case, [ π’1 π’0 β2πΌπ d . ] β1 β [ ] for constructing π we need to solve four linear equations [ . ] [ . ddβπ’πβ1 ] with the same coefficient matrix π. Labahn et al. also gave . this conclusion in terms of certain matrix Padeforms[Β΄ 14]. [ π’πβ1 . π’1 π’0 β2πΌπ] π₯Μ π₯Μ β β β π₯Μ Remark 4. For the inverse formula (22), we take an interest πβ1 0 πβ2 } π=1 π=0 [ ]} in two particular cases. When and ,theinverse [ . ]} [π₯Μπβ2 π₯Μπβ1 d . ]} formula (22)canbewrittenas Γ [ ] . [ . ]} [ . ddπ₯Μ0 ]} π₯ 0 β β β 0 } { 0 . } {[ ] [ π₯Μ . π₯Μ π₯Μ ]} {[ . ] 0 . πβ2 πβ1 {[ π₯1 π₯0 d . ] πβ1 = [ ] {[ . ] {[ . dd 0 ] { This is a generalization of the formula of Ammar and Gader . [21], where they considered the Hermitian Toeplitz matrix. {[π₯πβ1 . π₯1 π₯0] σΈ σΈ σΈ π’Μ π’Μ β β β π’Μ Let π¦=(π¦0,π¦1,...,π¦πβ1) ,π§ = (π§0,π§1,...,π§πβ1) , π¦Μ = πβ1 0 πβ2 σΈ [ . ] (π¦Μ , π¦Μ ,...,π¦Μ ) and π§Μ =(π§Μ0, π§Μ1,...,π§Μπβ1) be the solutions [π’Μ π’Μ d . ] 0 1 πβ1 [ πβ2 πβ1 . ] ππ¦ =πΈ ,ππ§ =πΈ , π¦ΜσΈ π=πΈσΈ Γ [ ] of the equations π π+1 πβπβ1, [ . ] ΜσΈ σΈ [ . ddπ’Μ0 ] and π§ π=πΈπβπβ2 for some integer π(0β€ π β€ πβ1), . π¦ π¦Μ π’Μ . π’Μ π’Μ respectively. In addition, πβ1 and 0 are both nonsingular, [ 0 . πβ2 πβ1] π¦Μβ1 βπβ1 π¦Μ π =βπβ1 π π¦ π¦β1 0β€πβ€πβ2 (24) 0 π=1 π βπ+π π=1 βπ πβ1 πβ1 for and Μβ1 Μβ1 πβ1 Μ β1 πβ1 β1 π’0 βπΌπ 0β β β 0 (1/π)π¦0 + π¦0 βπ=1 π¦ππβπ+π =(1/π)π¦πβ1 + βπ=1 πβππ¦πβ1π¦πβ1 [ . ] for π=πβ1. [ . ] πβ1 [ π’1 π’0 βπΌπ d . ] π<πβ1 π =π +(1/π)β π π¦ π¦β1 = β [ ] For ,wehave π 0 π=1 βπ πβ1 πβ1 [ . ] π +(1/π)π¦Μβ1 βπβ1 πΜ π [ . dd 0 ] 0 0 π=1 π βπ+π.Thenweobtain . [ π’πβ1 . π’1 π’0 βπΌπ] 1 ππ’ =π (π π¦βπ§)π¦β1 π₯Μ π₯Μ β β β π₯Μ π π πβ1 πβ1 0 πβ2 } [ . ]} [π₯Μ π₯Μ d . ]} [ πβ2 πβ1 . ] ππ¦πβ1 Γ [ . ]} . [ ] [ . ddπ₯Μ ]} 1 [ 0 ] [ . 0 ]} = π [ ] π¦β1 } [ . ] πβ1 . π . [ π₯Μ0 . π₯Μπβ2 π₯Μπβ1]} [ 0 ] So we obtain that the inverse of a block Toeplitz matrix can 0 be represented as a sum of products of block lower triangular [ ] 1 [ π¦0 ] 1 Toeplitz matrices and block circulant matrices. This is a + π [ ] π¦β1 β ππ§π¦β1 [ . ] πβ1 πβ1 generalization of the formula of Ammar and Gader [19]. And π . π when π=1and π=β1,theinverseformula(22)canbe [π¦πβ2] 6 Journal of Mathematics
π π(π βπ)π¦ β1 πβ1 πβ1 β1 0 π 0 πβ1 π¦Μ β π¦Μ πβπ+π = β πβππ¦πβ1π¦ for 0β€πβ€πβ2and [ ] [ ] 0 π=1 π π=1 πβ1 [ π1 ] 1 [ βπβπ+1π¦πβ1 ] (1/π)π¦Μβ1 + π¦Μβ1 βπβ1 π¦Μ π =(1/π)π¦β1 + βπβ1 π π¦ π¦β1 = [ ] + [ ] 0 0 π=1 π βπ+π πβ1 π=1 βπ πβ1 πβ1 [ . ] [ . ] π=πβ1 π . π . for ,then is invertible and [ππβ1] [ βπβ1π¦πβ1 ] β1 1 β1 σΈ π = (πΆ (π₯, π)π¦ β πΆ(π½Μ0 (πππ¦βΜ π§),π))Μ β1 1 β1 1 β1 πβπ Γπ¦πβ1 + πΈπ+1π¦πβ1 β πΈπ+1π¦πβ1 (28) π π β1 βπΆ((πππ¦βπ§)π¦πβ1,π)β πΆ(π½π₯,Μ π)) . ππ [ 1 ] [π β π ] Corollary 5 shows how to construct the inverse of a block [ 1 βπ+1] [ π ] Toeplitz matrix, provided that the first block column and last = [ ] , [ . ] block row, along with two successive block columns and block [ . ] [ 1 ] rows of the inverse, are given. More precisely, the choice of π β π π=πβ1gives that πΈπ+1 =πΈ0 and πΈπβπβ2 =πΈπβ1.Thismeans πβ1 π β1 [ ] that π§=π₯and π§=Μ π₯Μ. As a consequence, we have the following 1 result. π’ΜσΈ π= π¦Μβ1 (π¦ΜσΈ π β π§ΜσΈ )π π 0 π Corollary 6. Let π be the block Toeplitz matrix given by σΈ 1 β1 (1).If (15) and (16), ππ¦ =πΈ πβ1,andπ¦Μ π=πΈ0σΈ are = π¦Μ0 (0,...,0,ππ¦Μ0)π Μ π solvable, in addition, π¦πβ1 and π¦0 are both nonsingular, β1 πβ1 πβ1 β1 π¦Μ β π¦Μ πβπ+π =β πβππ¦πβ1π¦ ,thenπ is invertible and 1 1 0 π=1 π π=1 πβ1 + π¦Μβ1 (π¦Μ ,...,π¦Μ ,0)β π¦Μβ1π§ΜσΈ π 0 1 πβ1 1 π π πβ1 = (πΆ (π₯, π)π¦ β πΆ(π½Μβ1 (πσΈ π¦βΜ π₯),π))Μ πβπ 0 π =(ππβ1,...,π1,π0) (29) βπΆ ((π π¦βπ₯)π¦β1 ,π)β πΆ(π½π₯,Μ π)) . 1 π πβ1 + π¦Μβ1 (βπ¦Μ π ,...,βπ¦Μ π ,ππ¦Μ (π βπ)) π 0 0 β1 0 βπ+1 0 π 0 Corollary 6 is a generalization of the well-known formula 1 1 of Gohberg and HeΛΔ±nig. In [13], the authors have shown that + π¦Μβ1πΈσΈ β π¦Μβ1πΈσΈ π 0 πβπβ1 π 0 πβπβ1 the first and last block columns together with the first and last block rows of the inverse are sufficient to construct the inverse β1 1 1 π . In addition, we decompose the inverses of block Toeplitz =(π β π ,...,π β π ,π ). πβ1 π β1 1 π βπ+1 π matrices as a sum of products of block circulant matrices. (26) 4. Stability Analysis β1 For π=πβ1,wehaveππ βπ0 =(1/π)π¦πβ1 + (1/π) βπβ1 π π¦ π¦β1 =(1/π)π¦Μβ1 +(1/π)π¦Μβ1 βπβ1 π¦Μ π In the scalar case, the inverse formula (22)inSection 3 is π=1 βπ πβ1 πβ1 0 0 π=1 π βπ+π. just the well-known formula of Gohberg and Olshevsky; see A similar argument shows that Theorem 3.1 in [20]. Hence, solving two linear systems are sufficient to reconstruct the inverse of the Toeplitz matrix ππ β1 σΈ σΈ [ 1 ] π .Letπ₯=(π₯0,π₯1,...,π₯πβ1) and π’=(π’0,π’1,...,π’πβ1) be [π β π ] [ 1 βπ+1] the solutions of (15)and(17), respectively. In the scalar case, [ π ] ππ’ = [ ] , formula (22)canbewrittenas [ . ] [ . ] [ 1 ] (27) β1 π ππβ1 β πβ1 π = (πΆ (π₯, π) β πΆ (π’,π) [ π ] πβπ (30) 1 1 πβπ π’ΜσΈ π=(π β π ,...,π β π ,π ). βπΆ (π’ β πΈ ,π)β πΆ(π₯,π)), πβ1 π β1 1 π βπ+1 π π 0
β1 π =0ΜΈ π( =π)ΜΈ As a result, we obtain that π’ = (1/π)(πππ¦βπ§)π¦ and where and is an arbitrary number. For conve- πβ1 β1 σΈ β1 σΈ σΈ nience, we rewrite π as π’Μ =(1/π)π¦0 (π¦Μ ππ β π§Μ ) are the solutions of (17)and(18), respectively. This result together with Theorem 2 yields the π πβ1 β (πΆ πΆ βπΆ πΆ ) . following corollary. πβπ 11 12 21 22 (31) Corollary 5. π Let be the block Toeplitz matrix given by (1). In this section, we want to show that the evaluation of some ππ¦ =πΈ ,ππ§ =πΈ , π¦ΜσΈ π=πΈσΈ If (15) and (16), π π+1 πβπβ1, scalar inverse formulas in Section 3 is forward stable. An σΈ σΈ and π§Μ π=πΈπβπβ2 are solvable for some integer π(0 β€ algorithm is called forward stable if for all well-conditioned πβ€n β1),inaddition,π¦πβ1 and π¦Μ0 are both nonsingular, problems, the computed solution π₯Μ is close to the true Journal of Mathematics 7
solution π₯ in the sense that the relative error βπ₯ β π₯βΜ 2/βπ₯β2 is subtracting the matrices, and πΊ represents the error of the small. In the matrix computation, roundoff errors occur. Let multiplication by π/(π β. π) For the error matrices ΞπΆ11, π,π 2 π΄, π΅ βπΆ and πΌβπΆ;ifweneglecttheπ(π ) terms, then for ΞπΆ12, ΞπΆ21,andΞπΆ22,wehave any floating-point arithmetic with machine precision π,there σ΅© σ΅© σ΅© σ΅© σ΅©ΞπΆ σ΅© β€ Μπσ΅©πΆ σ΅© β€ Μπβππβπ₯β , are σ΅© 11σ΅©πΉ σ΅© 11σ΅©πΉ 2 (πΌπ΄) =πΌπ΄+πΈ, σ΅© σ΅© Μσ΅© σ΅© Μβ fl σ΅©ΞπΆ12σ΅©πΉ β€ πσ΅©πΆ12σ΅©πΉ β€ π ππβπ’β2, β σ΅¨πβπσ΅¨ (36) βπΈβπΉ β€π|πΌ|βπ΄βπΉ β€π π |πΌ|βπ΄β2, σ΅© σ΅© σ΅© σ΅© σ΅¨ σ΅¨ σ΅©ΞπΆ21σ΅© β€ Μπσ΅©πΆ21σ΅© β€ Μπβπ(σ΅¨ σ΅¨ +πβπ’β2), πΉ πΉ σ΅¨ π σ΅¨ fl (π΄+π΅) =π΄+π΅+πΈ, (32) σ΅© σ΅© Μσ΅© σ΅© Μβ σ΅©ΞπΆ22σ΅©πΉ β€ πσ΅©πΆ22σ΅©πΉ = π ππβπ₯β2. βπΈβπΉ β€πβπ΄+π΅βπΉ β€πβπβπ΄+π΅β2, It follows that fl (π΄π΅) =π΄π΅+πΈ, βπΈβπΉ β€ππβπ΄βπΉβπ΅βπΉ. βπΈβ2 β€ βπΈβπΉ See [22]. According to the floating-point arithmetic, we have σ΅© σ΅© σ΅© σ΅© σ΅© σ΅© σ΅© σ΅© the following result. β€ππ(σ΅©πΆ11σ΅©πΉσ΅©πΆ12σ΅©πΉ + σ΅©πΆ21σ΅©πΉσ΅©πΆ22σ΅©πΉ) σ΅¨ σ΅¨ Theorem 7. Let π be a nonsingular Toeplitz matrix given by 2 σ΅¨πβπσ΅¨ β€πππβπ₯β2 (σ΅¨ σ΅¨ +2πβπ’β2), (1) with π=1and let it be well conditioned then the inversion σ΅¨ π σ΅¨ formula (30) is forward stable for π =0ΜΈ and a given number σ΅© β1σ΅© π( =π)ΜΈ βπΉβ β€πβπσ΅©π σ΅© , . 2 σ΅© σ΅©2 (37)
Proof. Assume that we have computed the solutions π₯Μ, π’Μ βπΊβ2 β€ βπΊβπΉ in the inverse formula (30)whichareperturbedbythe π σ΅© σ΅© σ΅© σ΅© σ΅© σ΅© σ΅© σ΅© normwise relative errors bounded by Μπ β€π (σ΅©πΆ σ΅© σ΅©πΆ σ΅© + σ΅©πΆ σ΅© σ΅©πΆ σ΅© ) πβπ σ΅© 11σ΅©πΉσ΅© 12σ΅©πΉ σ΅© 21σ΅©πΉσ΅© 22σ΅©πΉ βπ₯Μβ β€ βπ₯β (1+Μπ) , βπ’Μβ β€ βπ’β (1+Μπ) . 2 2 2 2 (33) σ΅¨ σ΅¨ ππππ σ΅¨πβπσ΅¨ β€ βπ₯β2 (σ΅¨ σ΅¨ +2πβπ’β2). Let π=max{1, π} and π=max{1, π}.Then,wehave πβπ σ΅¨ π σ΅¨ σ΅© σ΅© β σ΅©πΆ11σ΅©πΉ β€ ππβπ₯β2, After adding all these error bounds, we have σ΅© σ΅© σ΅© σ΅© β σ΅©Μβ1 β1σ΅© π 2 σ΅©πΆ12σ΅©πΉ β€ ππβπ’β2, σ΅©π βπ σ΅© β€ (2Μππ + ππ +ππ) σ΅© σ΅©2 πβπ σ΅© σ΅© πβπ 2 σ΅¨ σ΅¨ σ΅©πΆ σ΅© β€ βπ( ) +ππ2βπ’β2 σ΅¨πβπσ΅¨ σ΅© 21σ΅©πΉ 2 Γπβπ₯β2 (σ΅¨ σ΅¨ +2πβπ’β2) π (34) σ΅¨ π σ΅¨ (38) σ΅¨ σ΅¨ σ΅¨πβπσ΅¨ β€ βπ σ΅¨ σ΅¨ + βππβπ’β , πβππσ΅© β1σ΅© σ΅¨ σ΅¨ 2 + σ΅©π σ΅© . σ΅¨ π σ΅¨ πβπσ΅© σ΅©2 σ΅© σ΅© σ΅©πΆ σ΅© β€ βππβπ₯β . σ΅© 22σ΅©πΉ 2 Note that π₯ and π’ are the solutions of (15)and(17), βπ₯β β€βπβ1β βπ’β β€βπβ1β βπβ π₯Μ π’Μ respectively; then 2 2 and 2 2 2, By using the perturbed solutions , ,theinverseformula π=(π,π β(1/π)π ,...,π β(1/π)π )σΈ (30) can be expressed as where π 1 1βπ πβ1 β1 .Thus, the relative error is β1 π σ΅© σ΅© Μ Μ Μ Μ Μ σ΅© β1 β1σ΅© π = fl ( (πΆ11πΆ12 + πΆ21πΆ22)) σ΅©πΜ βπ σ΅© πβπ σ΅© σ΅© π 2 σ΅© σ΅© 2 β€ (2Μππ + ππ +ππ) σ΅©πβ1σ΅© πβπ π σ΅© σ΅©2 = ( ((πΆ +ΞπΆ )(πΆ +ΞπΆ ) fl πβπ 11 11 12 12 σ΅¨ σ΅¨ β σ΅¨πβπσ΅¨ σ΅© β1σ΅© σ΅© σ΅© π ππ Γπ(σ΅¨ σ΅¨ +2πσ΅©π σ΅© σ΅©πσ΅© )+ . σ΅¨ π σ΅¨ σ΅© σ΅©2 2 πβπ (35) +(πΆ21 +ΞπΆ21)(πΆ22 +ΞπΆ22)) ) (39) π π βπβ1β βπβ =πβ1 + (ΞπΆ πΆ +πΆ ΞπΆ +ΞπΆ πΆ As is well conditioned, thus, 2 is finite. Obviously, 2 πβπ 11 12 11 12 21 22 is finite. Therefore, the formula (30) presented in this section is forward stable. +πΆ21ΞπΆ22 +πΈ+πΉ)+πΊ. For example, there are two particular cases of the inverse Here, πΈ is the matrix containing the error which results from formula (30). When π=1and π=β1,theinverseformula β1 computing the matrix products, πΉ contains the error from π = (1/2)(πΆ(π₯, β1) β πΆ(π’, 1) 0β πΆ(π’2πΈ ,β1)β πΆ(π₯,1)), 8 Journal of Mathematics
β1 which denotes π as a sum of products of skew-circulant [7] A. Ben-Artzi and T. Shalom, βOn inversion of Toeplitz and close and circulant matrices, is numerically forward stable for a to Toeplitz matrices,β Linear Algebra and its Applications,vol.75, well-conditioned and nonsingular Toeplitz matrix π.Infact, pp.173β192,1986. this is the formula of Ammar and Gader in [21]. And when [8] G. Labahn and T. Shalom, βInversion of Toeplitz matrices π=1and π=0, this is the formula of Ammar and with only two standard equations,β Linear Algebra and its β1 Gader in [19]. The inverse formula π =(πΆ(π₯,0)β πΆ(π’,1)β Applications,vol.175,pp.143β158,1992. πΆ(π’0 βπΈ ,0)β πΆ(π₯,1))in which skew-circulant matrices are [9] T. Huckle, βComputations with Gohberg-Semencul-type for- replaced by upper triangular factors is also forward stable, if mulas for Toeplitz matrices,β Linear Algebra and its Applications, π is nonsingular and well conditioned. vol. 273, pp. 169β198, 1998. We also find that even when π=βand a given number [10] M. K. Ng, K. Rost, and Y.-W. Wen, βOn inversion of Toeplitz π =0ΜΈ ,theinverseformula(30) is still numerically forward matrices,β Linear Algebra and its Applications,vol.348,pp.145β stable. A similar argument can show that the inverse formulas 151, 2002. in Corollaries 5 and 6 for the scalar case are all numerically [11] G. Heinig, βOn the reconstruction of Toeplitz matrix inverses forward stable when π is nonsingular and well conditioned. from columns,β Linear Algebra and its Applications,vol.350,pp. 199β212, 2002. [12] S. Cabay and R. Meleshko, βA weakly stable algorithm for PadeΒ΄ 5. Concluding Remarks approximants and the inversion of Hankel matrices,β SIAM Journal on Matrix Analysis and Applications,vol.14,no.3,pp. In this paper, with the help of the block cyclic displace- 735β765, 1993. ment, the inverses of block Toeplitz matrices are discussed. [13]I.C.GohbergandG.HaΛΔ±nig, βInversion of finite Toeplitz New formulas for inversion of block Toeplitz matrices are matrices consisting of elements of a noncommutative algebra,β obtained, which can be decomposed in a sum of products RevueRoumainedeMathematiquesΒ΄ Pures et AppliqueesΒ΄ ,vol.19, of block circulant matrices. In the scalar case, we analyze pp. 623β663, 1974. the stability of the inverse formulas. In addition, it should [14] G. Labahn, D. K. Choi, and S. Cabay, βThe inverses of block be remarked that any mention of matrix PadeformsforΒ΄ Hankel and block Toeplitz matrices,β SIAMJournalonComput- representing inverses of Toeplitz matrices could also be given ing,vol.19,no.1,pp.98β123,1990. in terms of solving block linear systems. Also the proofs of [15]S.Cabay,A.R.Jones,andG.Labahn,βComputationofnumer- theinversesformulasin[14] could also be extended to prove ical Pade-HermiteΒ΄ and simultaneous Padesystems.I.NearΒ΄ the inverse formulas in this work. inversion of generalized Sylvester matrices,β SIAM Journal on Matrix Analysis and Applications,vol.17,no.2,pp.248β267, 1996. Acknowledgments [16] S. Cabay, A. R. Jones, and G. Labahn, βComputation of This research is supported by NSFC (61170311), 973 Program numerical Pade-HermiteΒ΄ and simultaneous PadeΒ΄ systems. II. A weakly stable algorithm,β SIAMJournalonMatrixAnalysisand (2013CB329404), and Sichuan Province Sci. & Tech. Research Applications,vol.17,no.2,pp.268β297,1996. Project (2012GZX0080). The authors are grateful to the [17] A. Ben-Artzi and T. Shalom, βOn inversion of block Toeplitz unknown referee for pertinent comments, especially for matrices,β Integral Equations and Operator Theory,vol.8,no.6, suggesting four references, which contributed substantially to pp. 751β779, 1985. this revised work. [18] L. Gemignani, βSchur complements of Bezoutians and the inversion of block Hankel and block Toeplitz matrices,β Linear References Algebra and its Applications,vol.253,pp.39β59,1997. [19] G. Ammar and P. Gader, βA variant of the Gohberg-Semencul [1] R. H.-F. Chan and X.-Q. Jin, An Introduction to Iterative formula involving circulant matrices,β SIAM Journal on Matrix Toeplitz Solvers,vol.5ofFundamentals of Algorithms,Society Analysis and Applications,vol.12,no.3,pp.534β540,1991. for Industrial and Applied Mathematics (SIAM), Philadelphia, [20] I. Gohberg and V. Olshevsky, βCirculants, displacements and Pa,USA,2007. decompositions of matrices,β Integral Equations and Operator [2]M.K.Ng,Iterative Methods for Toeplitz Systems, Numerical Theory,vol.15,no.5,pp.730β743,1992. Mathematics and Scientific Computation, Oxford University [21] G. Ammar and P.Gader, βNew decompositions of the inverse of Press, New York, NY, USA, 2004. atoeplitzmatrix,βinProceedings of the International Symposium [3] X.-Q. Jin, Developments and Applications of Block Toeplitz MTNS-89: Signal Processing, Scattering and Operator Theory, Iterative Solvers,vol.2ofCombinatorics and Computer Science, and Numerical Methods, vol. 3, pp. 421β428, Birkhaauser,Β¨ Kluwer Academic, Dordrecht, The Netherlands, 2002. Boston, Mass, USA, 1990. [4] M. Van Barel and A. Bultheel, βA lookahead algorithm for [22] G. H. Golub and C. F. Van Loan, Matrix Computations,vol. the solution of block Toeplitz systems,β Linear Algebra and its 3ofJohns Hopkins Series in the Mathematical Sciences,Johns Applications,vol.266,pp.291β335,1997. Hopkins University Press, Baltimore, Md, USA, 2nd edition, [5]I.C.GohbergandA.A.Semencul,βTheinversionoffinite 1989. Toeplitz matrices and their continual analogues,β Matematich- eskie Issledovaniya,vol.7,no.2,pp.201β223,1972. [6] G. Heinig and K. Rost, Algebraic Methods for Toeplitz-Like Matrices and Operators,vol.19ofMathematical Research, Birkhaauser,Β¨ Berlin, Germany, 1984. Advances in Advances in Journal of Journal of Operations Research Decision Sciences Applied Mathematics Algebra Probability and Statistics Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014
The Scientific International Journal of World Journal Differential Equations Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014
Submit your manuscripts at http://www.hindawi.com
International Journal of Advances in Combinatorics Mathematical Physics Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014
Journal of Journal of Mathematical Problems Abstract and Discrete Dynamics in Complex Analysis Mathematics in Engineering Applied Analysis Nature and Society Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014
International Journal of Journal of Mathematics and Mathematical Discrete Mathematics Sciences
Journal of International Journal of Journal of Function Spaces Stochastic Analysis Optimization Hindawi Publishing Corporation Hindawi Publishing Corporation Volume 2014 Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014