Special Matrices and Their Applications in Numerical Mathematics: Second Edition
5/5
()
About this ebook
Author Miroslav Fiedler, a Professor at the Institute of Computer Science of the Academy of Sciences of the Czech Republic, Prague, begins with definitions of basic concepts of the theory of matrices and fundamental theorems. In subsequent chapters, he explores symmetric and Hermitian matrices, the mutual connections between graphs and matrices, and the theory of entrywise nonnegative matrices. After introducing M-matrices, or matrices of class K, Professor Fiedler discusses important properties of tensor products of matrices and compound matrices and describes the matricial representation of polynomials. He further defines band matrices and norms of vectors and matrices. The final five chapters treat selected numerical methods for solving problems from the field of linear algebra, using the concepts and results explained in the preceding chapters.
Related to Special Matrices and Their Applications in Numerical Mathematics
Titles in the series (100)
Infinite Series Rating: 4 out of 5 stars4/5A History of Mathematical Notations Rating: 4 out of 5 stars4/5History of the Theory of Numbers, Volume II: Diophantine Analysis Rating: 0 out of 5 stars0 ratingsLaplace Transforms and Their Applications to Differential Equations Rating: 5 out of 5 stars5/5Topology for Analysis Rating: 4 out of 5 stars4/5Gauge Theory and Variational Principles Rating: 2 out of 5 stars2/5First-Order Partial Differential Equations, Vol. 1 Rating: 5 out of 5 stars5/5Calculus Refresher Rating: 3 out of 5 stars3/5The Calculus Primer Rating: 0 out of 5 stars0 ratingsLinear Algebra Rating: 3 out of 5 stars3/5Differential Forms with Applications to the Physical Sciences Rating: 5 out of 5 stars5/5Dynamic Probabilistic Systems, Volume II: Semi-Markov and Decision Processes Rating: 0 out of 5 stars0 ratingsA Catalog of Special Plane Curves Rating: 2 out of 5 stars2/5An Introduction to Lebesgue Integration and Fourier Series Rating: 0 out of 5 stars0 ratingsFourier Series and Orthogonal Polynomials Rating: 0 out of 5 stars0 ratingsMathematics for the Nonmathematician Rating: 4 out of 5 stars4/5Advanced Calculus: Second Edition Rating: 5 out of 5 stars5/5Methods of Applied Mathematics Rating: 3 out of 5 stars3/5First-Order Partial Differential Equations, Vol. 2 Rating: 0 out of 5 stars0 ratingsAnalytic Inequalities Rating: 5 out of 5 stars5/5Geometry: A Comprehensive Course Rating: 4 out of 5 stars4/5Calculus: An Intuitive and Physical Approach (Second Edition) Rating: 4 out of 5 stars4/5Chebyshev and Fourier Spectral Methods: Second Revised Edition Rating: 4 out of 5 stars4/5Applied Functional Analysis Rating: 0 out of 5 stars0 ratingsCounterexamples in Topology Rating: 4 out of 5 stars4/5Numerical Methods Rating: 5 out of 5 stars5/5An Adventurer's Guide to Number Theory Rating: 4 out of 5 stars4/5Mathematical Foundations of Statistical Mechanics Rating: 4 out of 5 stars4/5Fourier Series Rating: 5 out of 5 stars5/5Optimization Theory for Large Systems Rating: 5 out of 5 stars5/5
Related ebooks
Vector Spaces and Matrices Rating: 0 out of 5 stars0 ratingsAn Introduction to Linear Algebra and Tensors Rating: 1 out of 5 stars1/5Matrices and Transformations Rating: 4 out of 5 stars4/5Methods of Applied Mathematics Rating: 3 out of 5 stars3/5The Theory of Matrices in Numerical Analysis Rating: 4 out of 5 stars4/5Kronecker Products and Matrix Calculus with Applications Rating: 0 out of 5 stars0 ratingsMatrix Theory Rating: 0 out of 5 stars0 ratingsDifferential Geometry Rating: 5 out of 5 stars5/5Modern Multidimensional Calculus Rating: 0 out of 5 stars0 ratingsComplex Variable Methods in Elasticity Rating: 0 out of 5 stars0 ratingsA First Course in Functional Analysis Rating: 0 out of 5 stars0 ratingsPrinciples of Combinatorics Rating: 0 out of 5 stars0 ratingsAdvanced Calculus Rating: 0 out of 5 stars0 ratingsLinear Algebra and Its Applications Rating: 3 out of 5 stars3/5Matrix Vector Analysis Rating: 0 out of 5 stars0 ratingsInvitation to Dynamical Systems Rating: 5 out of 5 stars5/5Advanced Trigonometry Rating: 2 out of 5 stars2/5The Method of Trigonometrical Sums in the Theory of Numbers Rating: 0 out of 5 stars0 ratingsVector and Tensor Analysis with Applications Rating: 4 out of 5 stars4/5Differential Geometry Rating: 0 out of 5 stars0 ratingsIntroduction to Matrices and Vectors Rating: 0 out of 5 stars0 ratingsOrdinary Differential Equations and Stability Theory: An Introduction Rating: 0 out of 5 stars0 ratingsMatrix Theory and Applications for Scientists and Engineers Rating: 0 out of 5 stars0 ratingsMatrices and Linear Algebra Rating: 4 out of 5 stars4/5Hilbert Space Methods in Partial Differential Equations Rating: 4 out of 5 stars4/5Linear Algebra Rating: 3 out of 5 stars3/5Elementary Differential Equations with Linear Algebra Rating: 0 out of 5 stars0 ratingsDeterminants and Matrices Rating: 3 out of 5 stars3/5Advanced Mathematics for Engineers and Scientists Rating: 4 out of 5 stars4/5
Mathematics For You
Geometry For Dummies Rating: 5 out of 5 stars5/5Calculus Made Easy Rating: 4 out of 5 stars4/5Algebra - The Very Basics Rating: 5 out of 5 stars5/5Quantum Physics for Beginners Rating: 4 out of 5 stars4/5Algebra I Workbook For Dummies Rating: 3 out of 5 stars3/5Mental Math Secrets - How To Be a Human Calculator Rating: 5 out of 5 stars5/5Basic Math & Pre-Algebra For Dummies Rating: 4 out of 5 stars4/5The Golden Ratio: The Divine Beauty of Mathematics Rating: 5 out of 5 stars5/5Painless Geometry Rating: 4 out of 5 stars4/5Practice Makes Perfect Algebra II Review and Workbook, Second Edition Rating: 4 out of 5 stars4/5The Thirteen Books of the Elements, Vol. 1 Rating: 0 out of 5 stars0 ratingsThe Everything Guide to Algebra: A Step-by-Step Guide to the Basics of Algebra - in Plain English! Rating: 4 out of 5 stars4/5The Little Book of Mathematical Principles, Theories & Things Rating: 3 out of 5 stars3/5Game Theory: A Simple Introduction Rating: 4 out of 5 stars4/5Precalculus: A Self-Teaching Guide Rating: 4 out of 5 stars4/5Is God a Mathematician? Rating: 4 out of 5 stars4/5Relativity: The special and the general theory Rating: 5 out of 5 stars5/5ACT Math & Science Prep: Includes 500+ Practice Questions Rating: 3 out of 5 stars3/5The Math Book: From Pythagoras to the 57th Dimension, 250 Milestones in the History of Mathematics Rating: 3 out of 5 stars3/5Real Estate by the Numbers: A Complete Reference Guide to Deal Analysis Rating: 0 out of 5 stars0 ratingsThe Elements of Euclid for the Use of Schools and Colleges (Illustrated) Rating: 0 out of 5 stars0 ratingsFlatland Rating: 4 out of 5 stars4/5My Best Mathematical and Logic Puzzles Rating: 5 out of 5 stars5/5Summary of The Black Swan: by Nassim Nicholas Taleb | Includes Analysis Rating: 5 out of 5 stars5/5
Reviews for Special Matrices and Their Applications in Numerical Mathematics
1 rating0 reviews
Book preview
Special Matrices and Their Applications in Numerical Mathematics - Miroslav Fiedler
Index
1
Basic concepts of matrix theory
This introductory chapter is essentially a survey of well known definitions and general principles from matrix theory which will be useful in the sequel.
1.1 Matrices
A matrix of type m-by-n or, equivalently, an m × n matrix, is a rectangular array of ran numbers (usually real or complex) arranged in m rows and n columns (m, n positive integers):
We call the number alk the entry of the matrix in the ith row and the kth. column. The set of m × n matrices with real entries will be denoted by Rm×n. Analogously, the set of m × n matrices with complex entries will be denoted by Cm×n. In some cases, entries can be polynomials, variables, functions, etc.
In this terminology, matrices with only one column (thus, n = 1) will be called column vectors and matrices with only one row (thus, m = 1) row vectors. In such case, we write Rm instead of Rm×1 and vectors will always be column vectors, unless we specify otherwise.
It is advantageous to denote the matrix by a single symbol, say, A, or [alk], C, etc.
Two matrices A = [aik], B = [bik] are equal, written A = B, if and only if they are both m × n matrices for some positive integers m and n, and aik = bik for i = 1, …, m, j = 1, …, n.
The importance of matrices is that one can introduce operations generalizing operations with numbers.
Matrices of the same type can be added: If A = [aik], B = [bik] then A + B is the matrix [aik + bik]. The operation of addition is thus entrywise. We admit also multiplication of a matrix by a number (real, complex, a parameter, etc.): If A = [aik] and if α is a number (also called scalar), then αA is the matrix [αaik] of the same type as A.
An m × n matrix A = [aik] can be multiplied by an n × p matrix B = [bkl] as follows: it is the m × p matrix C = [cil], where
It is important to notice that the matrices A and B can be multiplied (in this order) only if the number of columns of A is the same as the number of rows in B. Also, the entries of A and B should be multiplicable. In general, the product AB is not equal to BA, even if the multiplication of both products is possible. On the other hand, the multiplication fulfils the associative law
as well as (in this case, two) distributive laws:
and
whenever multiplications are possible.
An m × n matrix is called square if m = n. We then say that it is square of order m.
Of basic importance are the zero matrices of which all entries are zeros, and the identity matrices; these are square matrices, and have ones in the main diagonal and zeros elsewhere. Thus
are identity matrices of order one, two and three. We denote zero matrices simply by 0, the identity matrices by 1, sometimes with a subscript denoting the order.
The identity matrices of appropriate orders have the property that
hold for any matrix A.
Suppose now A = [aik] is an m × n matrix. Choose a subset M1 = {i1, i2, …, ir} of the set of row indices M = {1, 2, …, m}, and a subset N1 {k1, k2, …, ks} of the set of column indices N = {1, 2, …, n}. In addition, we assume that the integers ij and kl are ordered in such a way that i1 < i1 < … < ir and k1 < k2 < … < ks. The matrix
is called a submatrix of A. We denote this submatrix by A(M1, N1). If A is a square matrix with M = N, then every square submatrix A(M1, M1) (with M1 = N1) is called a principal submatrix. The principal submatrix will be sometimes denoted also by A(M1).
The most important type of submatrices are those whose sets of indices consist of consecutive integers. Such submatrices are called blocks. For example, in the matrix
the submatrices
are blocks. A matrix that is in this way subdivided by vertical
and horizontal
lines into blocks is said to be partitioned as indicated. The above matrix can thus be written as
the matrices Pik have the property that those with first indices equal have the same number of rows and (those with) second indices equal have the same number of columns.
An important property of partitioned matrices is that they behave to some extent as if the individual blocks were numbers. It is only necessary to preserve order in the calculation of products and to assure that the blocks have proper dimensions. Clearly, the addition of two matrices conformally partitioned into blocks can be performed in this manner. As regards multiplication, the following theorem holds (Exercise 1.1).
Theorem 1.1.1 Let
be matrices partitioned in such a way that for each i, i = l, …, r, and each k, k = 1, …, t, the number of columns in Pij equals the number of rows in Qjk, whenever j = 1, …, s.
Then the product PQ can be partitioned into blocks Rik, i = 1, …, r, k = 1, …, t, where
Example 1.1 Let us write the product
with two-by-two blocks. Then
i.e.
as can be verified by direct multiplication.
Let A = [aik] be an m × n matrix. The n × m matrix C = [cpq] whose entries are cpq = aqp, q = 1, …, m, p = 1, …, n is called the transpose of A; we denote it AT.
The operation of transposition has two important properties:
whenever AB is defined, and further
for any matrix A.
This notation will be used rather often, especially for column vectors. A column vector u with entries (components) u1, …, un will usually be written as [u1, …, un]T.
We turn now our attention to special cases of square matrices. A square matrix A = [aik] of order n is diagonal if aik = 0 for i ≠ k, i, k = 1, …, n; the diagonal matrix with the diagonal entries a11, …, ann will be often written as diag (a11, …, ann). The matrix A is a lower triangular matrix if aik = 0 for i < k, i, k = 1, …, n, and it is an upper triangular matrix if aik = 0 for i < i, k = 1, …, n. For example, the first two of the matrices
are lower triangular, the third is upper triangular, and the fourth matrix is a diagonal matrix which may be also written as diag (2, –1, 3). Clearly, the following statement holds.
Theorem 1.1.2 The product of two lower triangular matrices of the same order is again a lower triangular matrix, and the product of two upper triangular matrices of the same order is upper triangular. The product of two diagonal matrices of the same order is again a diagonal matrix.
A further important concept is that of a nonsingular matrix. We say that a square matrix A is nonsingular if there exists a matrix B such that AB = BA = 1. The matrix B is then called the inverse of A and is denoted by A–1. The following theorems can easily be proved.
Theorem 1.1.3 Given a nonsingular matrix A, its inverse is determined uniquely. Furthermore, we have (A–1)–1 = A.
Theorem 1.1.4 If A and B are nonsingular matrices of the same order, then their product AB is nonsingular as well and
Hence, the inverse of the product is the product of the inverses in the reverse order.
Theorem 1.1.5 If A is nonsingular, then its transpose AT is nonsingular as well and
Hence, the inverse of the transpose is the transpose of the inverse.
1.2 Determinants
Determinants play a crucial role in distinguishing whether a square matrix is singular or nonsingular.
The determinant det A of an n × n matrix A = [aik] is the number
The summation is over all permutations P of the integers 1, …, n and the factor σ(P) is the sign of the permutation P. It is equal to +1 if the permutation P is even, and to – 1 if P is odd. The permutation
is even if it has an even number of inversions, i.e. such pairs i, j for which i < j whereas ki > kj. An odd permutation has an odd number of inversions.
We state two consequences of (1.5). Clearly, if l is an identity matrix, then
The properties of permutations imply that for any square matrix A
Let A = [aik] be an m × n matrix. Let M = {1, 2, …, m}, N = {1, 2, …, n} and let A(M1, M2) be a square submatrix of A (M1 ⊂ M, M2 ⊂ N, |M1| = |M2|).¹ Then, we are able to consider the determinant det A(M1, M2) of the submatrix; such a determinant will be called a minor of A. If A is a square matrix, then together with the above minor there exists the minor det A(M \ M1, N \ M2).² It is the determinant of the submatrix whose rows consists precisely of the rows of A (in the same order as they occur in A) that are not included in the submatrix A(M1, M2), and whose columns consist precisely of the columns of A that are not included in A(M1, M2), again in the same order as they are in A. The number (–1)s(M1)+s(M2) det A(M \ M1, N \ M2), where s(M1) is the sum of the row indices belonging to M1 and s(M2) is the sum of the column indices belonging to M2, is then called the cofactor of the minor det A(M1, M2). It is advantageous to denote the cofactor of det A(M1, M2) by codet A(M1, M2). If A is a square matrix, then the determinants of principal submatrices play an important role. It is customary to call them principal minors.
Example 1.2 Let
Then for M1 = {1, 3}, M2 = {2, 3} we have
det A(M \ M1, N \ M2) = det[a21] = a21, s(M1) = 4, s(M2) = 5. Hence, codet A(M1, M2) = –a21.
The following important theorem gives an expansion of the determinant of A by some of its rows (or columns).³
Theorem 1.2.1 (Laplace expansion) Let M ⊂ N = {1, 2, …, n} be a fixed set of indices which has m elements. Let A = [aij], i, j = 1, …, n, be a square matrix of order n. Then
where the summation is over all the subsets Mj of the set of column indices that have m elements.
Remark. Since
the statement of the theorem may be rewritten as
where the summation is over all the sets Mj ⊂ N that have m elements.
Applying Theorem 1.2.1 to the transpose of A and turning back to the original matrix, we obtain a similar formula giving the expansion of the determinant by its columns:
where M is an arbitrary subset of N and the summation is over all the sets Mj ⊂ N that have the same number of elements as M.
As a corollary of the Laplace theorem we have the formulae (for A = [aik] of order n)
where Aik is the cofactor of the entry (the expansion of det A by its ith row), and
(the expansion of det A by its kth column).
An immediate consequence of (1.10) and (1.11) is the next theorem.
Theorem 1.2.2 Let A and B be square matrices of equal orders which differ only in the entries of one row or column. Let C be the matrix whose entries in this row or column are sums of the corresponding entries of A and B, the other entries being the same as in A and B. Then
Assuming that A and B differ only in the first column and denoting the columns of A by A1, A2, …, An, and the columns of B by B1, A2, …, An, we can write the above formula as
Repeatedly using this result, we establish a formula for the determinant of the sum of two matrices, A and B:
Here, the summation is over i1 = 1, 2, i2 = 1, 2, …, in = 1, 2, and =
.
If we now use (1.9) to expand each determinant on the right-hand side by all its columns that are from A we arrive at the formula
where the summation is over all pairs of subsets Mi, Mj, Mi ⊂ N, Mj ⊂ N, which have the same number of elements (|Mi| = |Mj|).
A further consequence of (1.10) and (1.11) is the following theorem.
Theorem 1.2.3 If a row or a column of a square matrix A contains only zero entries, then det A = 0.
If two rows of A are equal, then applying the Laplace expansion 1.2.1 to these rows one obtains the following theorem.
Theorem 1.2.4 If a square matrix A has two identical rows, then det A = 0. The same is true if A has two identical columns.
Let us apply the theorem to the matrix B obtained from a square matrix A = [aik] of order n by replacing (for j ≠ i) the jth row of A by its ith. row (without changing the ith row). Expanding det B by its jth row we have, according to (1.10)
The same procedure applied to columns gives
The following important theorem can be proved by induction on r using the expansion (1.8) and Theorem 1.2.3.
Theorem 1.2.5 Let A = [Aik], i, k = 1, …, r, be a partitioned square matrix whose diagonal blocks are square matrices and the blocks below diagonal are zero blocks: Aik = 0 for i > k (such a matrix is said to be upper block triangular), i.e.,
Then,
Remark. A similar formula holds for lower block triangular and for block diagonal matrices.
The next theorem presents the celebrated Binet-Cauchy formula.
Theorem 1.2.6 Let P be an m × n matrix and let Q be an n × m matrix, m < n. Set M = {1, 2, …, m}, N = {1, 2, …, n}. Then
where the summation is over all the subsets Ni of N that have m elements.
We do not give the proof here⁴ but we state an important corollary for m = n.
Theorem 1.2.7 If A and B are square matrices of equal orders, then
1.3 Nonsingular matrices. Inverse matrices
We now turn to nonsingular matrices. First, we prove the following fundamental theorem.
Theorem 1.3.1 A square matrix A = [aik] is nonsingular if and only if det A ≠ 0. In this case, the inverse A–1 = [αik] has the entries
where Aki is the cofactor of aki in the matrix A.
Proof. Suppose A has the inverse A–1. Then
and by Theorem 1.2.7 and (1.6), we have
Therefore,
Conversely, suppose that det A ≠ 0. Define the matrix B = [αik] by αik = Aki/ det A. Then it follows that
as well as
by (1.10).
Further, if i ≠ j, then
according to (1.13).
Equations (1.19) and (1.20) may also be written in the form
A similar argument yields the relation
Thus, B is the (unique) inverse of A
Before proceeding further, we restate (1.18) as a theorem.
Theorem 1.3.2 If A is a nonsingular matrix, then
Theorem 1.3.1 enables us, at least theoretically, to find the solution of a system of linear equations written as
in matrix notation, provided A is a nonsingular square matrix. Indeed, supposing x = [x1, …, xn]T is the column solution vector and left-multiplying the system by
we get
Here, Mj is the matrix obtained from A by replacing its jth column by b. This method of solving systems of linear equations is called Cramer’s rule. A consequence of this formula is the next theorem.
Theorem 1.3.3 A system of n linear equations in n unknowns with a nonsingular matrix of coefficients has a unique solution.
Theorem 1.3.4 Let A be a square matrix. If a matrix B satisfies the relation
where l is the identity matrix, then B = A–1.
Remark. Thus, if we want to prove that B is the inverse of A it is sufficient to verify (1.21). It is not necessary to show also that BA = l.
Proof. By Theorems 1.2.7, 1.3.1, and by (1.6), A is nonsingular. The first column of B in (1.21) can be taken for the solution of the system
But the first column of A–1 satisfies the equation as well. By Theorems 1.3.3, the first columns of B and A–1 are identical. The same can be proved for the remaining columns, and thus B = A
Theorem 1.3.5 Let A be a square matrix. If the system Ax = 0 has a solution in which not all the unknowns are zero, then det A = 0.
Proof. The system in question always has a (trivial) solution x = 0. If A were nonsingular, then
Square matrices whose determinant is zero are called singular. Therefore, Theorem 1.3.5 gives a sufficient condition for A to be a singular matrix. Later we shall see that this condition is necessary as well (Theorem 1.5.3).
The following theorem is an immediate consequence of the definition of a singular matrix and Theorem 1.2.7.
Theorem 1.3.6 If one of the factors of a product of square matrices is singular, then the product is singular as well.
Note also the following corollary of Theorem 1.2.5.
Theorem 1.3.7 A block triangular matrix (lower or upper) is nonsingular if and only if all its diagonal blocks are nonsingular. In particular, a (lower or upper) triangular matrix is nonsingular if and only if all its diagonal entries are different from zero.
We are now in a position to study the properties of the inverse of a nonsingular lower (or upper) triangular matrix.
Theorem 1.3.8 If A = [aik] is a nonsingular lower (upper) triangular matrix, then its inverse A–1 is lower (upper) triangular as well.
Proof. In view of Theorem 1.3.4 it is sufficient to show that there is a lower triangular matrix B satisfying
The entries of such a matrix B, if it exists, fulfil the condition
as well as
Hence,
However, from these equations the entries can be calculated successively by using i – j = 1, 2, …, n – 1. Indeed, the right-hand side of (1.23) contains only the bkj’s with k – j less than i – j
If B is the lower triangular matrix whose entries are the calculated bij’s, then it does hold AB = l, i.e. B = A–1. The corresponding statement for an upper triangular matrix is obtained by taking transposes and employing
Remark. Note also the following two important consequences of (1.22) and (1.23):
(1) If A = [aik] is a lower triangular matrix satisfying aii > 0 for every i and aij ≤ 0 for all i, j, i ≠ j, then all the entries of A–1 are nonnegative. (Naturally, the same is true for upper triangular matrices as well.)
(2) If A is a nonsingular lower block triangular matrix, the proof applies without changes, except for writing
(li are identity matrices) instead of (1.22) and considering aik and bik as blocks. Therefore, the theorem also holds for lower (or upper) block triangular matrices. Another corollary is the next theorem.
Theorem 1.3.9 Lower triangular matrices of order n constitute a subset of square matrices which is closed⁵ with respect to addition, multiplication, scalar multiplication and inversion (of nonsingular matrices). The same holds for upper triangular matrices.
Moreover, carrying out the operations with triangular matrices we find that the kth diagonal entry of the result (k fixed, 1 ≤ k ≤ n) is, according to the matrix operation being performed, the sum, product or reciprocal of the kth diagonal entries of the operands and is not affected by other entries.
1.4 Schur complement. Factorization
We now introduce a further concept. Let
be a partitioned matrix and A11 its nonsingular block. The Schur complement [A/A11] of the block A11 in A is then defined [34] to be the matrix
(Here, A need not be square.)
Theorem 1.4.1 Let A = [Aij], i, j = 1, …, r, be a partitioned matrix. Let the submatrices
be nonsingular. Then there exist a lower block triangular matrix B having identity matrices as diagonal blocks, an upper block triangular matrix C having identity diagonal blocks, and a block diagonal matrix D such that
Moreover, the matrices B, D, and C are determined uniquely by the conditions of the theorem; the diagonal blocks D1, …, Dr of D satisfy
and D1, …, Dr–1 are nonsingular.
Proof. First we prove the uniqueness assuming that D1, …, Dr–1 are nonsingular. Consider two possible factorizations of A, (and C are nonsingular, we may write
By –1BD is upper block triangular. Hence both matrices are block diagonal, the former being equal to D–1B C–1 are identity matrices). Thus, D –1B = [Yik], i.e. Yjj = lj, Yik = 0 for i < k, and further
for i > k, we have
Since Di, i = 1, …, r–1 are nonsingular, we get Yik = 0 whenever i > k, Zik = 0 whenever i < k–1B = l C–1, and thus B , C .
The existence of the factorization (1.26), and the validity of (1.27) and (1.28) will be proved by induction on r. For r = 1, both (1.26) and (1.27) hold. Let A have r > 1 block rows and assume the existence of (1.26) for matrices having r obtained from A by deleting its last block row and the last block column, has r with the diagonal blocks D1, …, Dr.
Now write A in the partitioned form
and try to find B, C, and D,
such that A = BDC. This requirement gives
The right-hand side is
and it is therefore sufficient to satisfy the equations
are nonsingular as well and the first two equations may be solved for C1r and Br1. We thus obtain
The last of the equations is then satisfied by
i.e., by Dr = [A]. Therefore, the factorization (1.26) exists and (1.28) holds also for k = r.
At the same time we showed that D1, …, Dr
The theorem has several corollaries.
Theorem 1.4.2 Let A be a square matrix which is partitioned in the form
where A11 is nonsingular. Then
Hence A is nonsingular if and only if the Schur complement [A/A11] of the submatrix A11 in A is nonsingular. In such case, the inverse of the Schur complement [A/A11] is the submatrix of A–1 in the position complementary to that of A11.
In other words, if A–1 is partitioned as was A above,
then
Proof. Using the formulae (1.27) and (1.28) of Theorem 1.4.1 with r = 2, we get
But det A = det B det D det C by (1.26) and (1.17), where det B = det C = 1, det D = det D1 det D2 because of (1.15). Altogether, we have det A = det A11 det [A/A11], which proves (1.29).
Now suppose that A, and thus also [A/A11] = D2, is nonsingular. The factorization (1.26) implies A–1 = C–1D–1B–1. But B–1, C–1, and D–1 are of the form
so that
This implies B22 = D–12 which is (
If a square matrix A = [aik] of order n has all principal minors det A(Nk), Nk = {1, 2, …, k}, k = 1, …, n, different from zero, we call it strongly nonsingular. An immediate corollary of Theorem 1.4.1 is the next statement on strongly nonsingular matrices.
Theorem 1.4.3 If A = [aik] is a strongly nonsingular matrix of order n, then there is a unique factorization A = BDC, where B and C are lower and upper triangular matrices, respectively, with ones along the main diagonal, and D is a diagonal matrix with diagonal entries d1, …, dn. Moreover, we have d1 = a11, dk = det A(Nk)/det A(Nk–1), k = 2, …, n.
We shall now show another property of the Schur complement.
Theorem 1.4.4 [8] Let
be a partitioned matrix whose submatrices A11 and
are both nonsingular. Then
Proof. By Theorem 1.4.1, there exists a unique factorization A = BDC, where B is a lower block triangular matrix having three block rows and identity diagonal blocks, C is upper block triangular with analogous structure, and D is a block diagonal matrix with diagonal blocks D1, D2, D3, where
On the other hand, A may be partitioned as
and factored into factors with two block rows. According to (, where
2 may be factored according to (1.26),
so that
It is easy to verify that the matrix
written as a partitioned matrix with three block rows, has identity diagonal blocks and this is true of the matrix
as well. Hence (1.33) is only another way of writing the factorization A = BDC3 =