Open Access

A note on the Frobenius conditional number with positive definite matrices

Journal of Inequalities and Applications20112011:120

https://doi.org/10.1186/1029-242X-2011-120

Received: 12 June 2011

Accepted: 24 November 2011

Published: 24 November 2011

Abstract

In this article, we focus on the lower bounds of the Frobenius condition number. Using the generalized Schwarz inequality, we present some lower bounds for the Frobenius condition number of a positive definite matrix depending on its trace, determinant, and Frobenius norm. Also, we give some results on a kind of matrices with special structure, the positive definite matrices with centrosymmetric structure.

Keywords

Lower BoundCondition NumberPositive Real NumberOrthogonal MatrixMatrix Norm

1 Introduction and preliminaries

In this article, we use the following notations. Let n×nand n×nbe the space of n×n complex and real matrices, respectively. The identity matrix in n×nis denoted by I = I n . Let A T , Ā, A H , and tr(A) denote the transpose, the conjugate, the conjugate transpose, and the trace of a matrix A, respectively. Re(a) stands for the real part of a number a. The Frobenius inner product < ·, · > F in m×nis defined as < A,B > F = Re(tr(BHA)), for A,B m×n, i.e., < A,B > is the real part of the trace of B H A. The induced matrix norm is | | A | | F = < A , A > F = Re ( tr ( A H A ) ) = tr ( A H A ) , which is called the Frobenius (Euclidean) norm. The Frobenius inner product allows us to define the consine of the angle between two given real n × n matrices as
cos ( A , B ) = < A , B > F | | A | | F | | B | | F .
(1.1)
The cosine of the angle between two real n × n depends on the Frobenius inner product and the Frobenius norms of given matrices. A matrix A n×nis Hermitian if A H = A. An Hermitian matrix A is said to be positive semidefinite or nonnegative definite, written as A ≥ 0, if (see, e.g., [[1], p. 159])
x H A x 0 , x n ,
(1.2)

A is further called positive definite, symbolized A > 0, if the strict inequality in (1.2) holds for all nonzero x n . An equivalent condition for A n to be positive definite is that A is Hermitian and all eigenvalues of A are positive real numbers.

The quanity
μ ( A ) = | | A | | | | A - 1 | | , if A is nonsingular ; if A is singular .
(1.3)

is called the condition number of matrix μ(A) with respect to the matrix norm || · ||. Notice that μ(A) = ||A-1|| · ||A|| ≥ ||A-1A|| = ||I|| ≥ 1 for any matrix norm (see, e.g., [[2], p. 336]). The condition number μ(A) of a nonsingular matrix A plays an important role in the numerical solution of linear systems since it measures the sensitivity of the solution of linear system Ax = b to the perturbations on A and b. There are several methods that allow to find good approximations of the condition number of a general square matrix.

We first introduce some inequalities. Buzano [3] obtained the following extension of the celebrated Schwarz inequality in a real or complex inner product space (H, < ·, ·>).

Lemma 1.1 ([3]). For any a, b, x H, there is
| < a , x > < x , b > | 1 2 ( | | a | | | | b | | + | < a . b > | ) | | x | | 2 .
(1.4)
It is clear that for a = b, (1.4) becomes the standard Schwarz inequality
| < a , x > 2 | | | a | | 2 | | x | | 2 , a , x H
(1.5)

with equality if and only if there exists a scalar λ such that x = λa.

Also Dragomir [4] has stated the following inequality.

Lemma 1.2[4]. For any a,b,x H, and x ≠ 0, there is the following
< a , x > < x , b > | | x | | 2 - < a , b > 2 | | a | | | | b | | 2
(1.6)

Dannan [5] showed the following inequality by using arithmetic-geometric inequality.

Lemma 1.3[5]. For n-square positive definite matrices A and B,
n ( detA det B ) m n tr ( A m B m ) ,
(1.7)

where m is a positive integer.

By taking A = I, B = A-1, and m = 1 in (1.7), we obtain
n ( detI det A - 1 ) 1 n tr ( I A - 1 ) ,
(1.8)
n 1 detA 1 n tr ( A - 1 )
(1.9)

In [6], Türkmen and Ulukök proposed the following,

Lemma 1.4[6]. Let both A and B be n-square positive definite matrices, then
cos ( A , I ) cos ( B , I ) 1 2 cos ( A , B ) + 1 ] ,
(1.10)
cos ( A , A - 1 ) cos ( A , I ) cos ( A - 1 , I ) 1 2 [ cos ( A , A - 1 ) + 1 ] 1 ,
(1.11)
cos ( A , I ) 1 , cos ( A - 1 , I ) 1 ,
(1.12)

As a consequence, in the following section, we give some bounds for the Frobenius condition numbers by considering inequalities given in this section.

2 Main results

Theorem 2.1. Let A be a positive definite real matrix, α be any real number. Then,
2 n tr A 1 + α tr A 2 α ( detA ) ( 1 - α ) n - n μ F ( A ) ;
(2.1)
2 tr A 1 + α tr A α - 1 tr A 2 α - n μ F ( A ) .
(2.2)

where μ F (A) is the Frobenius conditional number of A.

Proof. Let X, A, B be positive definite real matrices. From Lemma 1.2, we have the following
< A , X > F < X , B > F | | X | | F 2 < A , B > F 2 | | A | | F | | B | | F 2 ,
(2.3)
i.e.,
tr ( A T X ) tr ( X T B ) | | X | | F 2 - tr ( A T B ) 2 | | A | | F | | B | | F 2 .
(2.4)
Let B = A-1, then (2.4) turns into
tr ( A T X ) tr ( X T A - 1 ) | | X | | F 2 - tr ( A T A - 1 ) 2 | | A | | F | | A - 1 | | F 2 .
(2.5)
Since both X and A are positive definite, we have
tr ( AX ) tr ( X A - 1 ) | | X | | F 2 - n 2 | | A | | F | | A - 1 | | F 2 = μ F ( A ) 2 ,
(2.6)

where μ F (A) is the Frobenius condition number of A.

By taking X = A α (α is an arbitrary real number) into (2.6), there exists
tr A 1 + α tr A - ( 1 - α ) tr A 2 α - n 2 μ F ( A ) 2 .
(2.7)
Thus, it follows that
tr A 1 + α tr A - ( 1 - α ) tr A 2 α - n 2 μ F ( A ) 2 ,
(2.8)
i.e.,
2 tr A 1 + α tr A α - 1 tr A 2 α - n μ F ( A ) .
(2.9)
From (1.9), by replacing A with A1, we get
0 < n 1 ( det A ) ( 1 - α ) n tr ( A - ( 1 - α ) ) .
(2.10)
Taking (2.10) into (2.8), we can write
n tr A 1 + α tr A 2 α 1 ( det A ) ( 1 - α ) n - n 2 μ F ( A ) 2 ,
(2.11)
i.e.,
2 n tr A 1 + α tr A 2 α ( detA ) ( 1 - α ) n - n μ F ( A ) .
(2.12)
In particular, let α = 1, and by taking it into (2.7), we have the following
tr A 2 trI tr A 2 - n 2 μ F ( A ) 2 ,
(2.13)
i.e.,
n μ F ( A ) .
(2.14)
Note, when α = 1 2 , (2.12) becomes
2 n tr A 3 2 t r A ( det A ) 1 2 n - n μ F ( A ) ,
(2.15)
Taking α = 1 into (2.12), we obtain that
2 tr A ( det A ) 1 n - n μ F ( A ) .
(2.16)

(2.15), (2.16) can be found in [6].

Example 2.2.
A = 2 0 0 2 - 1 .
Here trA = 2.5, detA = 1 and n = 2. Then, from (2.15) and (2.16), we obtain two lower bounds of μ F (A):
μ F ( A ) 2 n tr A 3 2 trA ( detA ) 1 2 n = 3 . 091168 , and μ F ( A ) 2 tr A ( det A ) 1 n = 3 .
Taking α = 1/4 into (2.1) and (2.2) from Theorem 2.1, another two lower bounds are obtained as follows:
μ F ( A ) 2 n tr A 1 + α tr A 2 α ( det A ) ( 1 - α ) n - n = 3 . 277585 , and μ F ( A ) 2 tr A 1 + α tr A α - 1 tr A 2 α - n = 4 . 006938 .

In fact, μ F (A) = 4.25. Thus, Theorem 2.1 is indeed a generalization of (2.15) and (2.16) given in [6].

Lemma 2.3. Let a1, a2,..., a n be positive numbers, and
f ( x ) = a 1 x + a 2 x + + a n x + a 1 - x + + a n - x .

Then, f(x) is monotonously increasing for x [0,+ ∞).

Proof. It is obvious that
f ( x ) = a 1 x ln a 1 + a 2 x ln a 2 + + a n x ln a n - a 1 - x ln a 1 - - a n x ln a n = ln a 1 ( a 1 x - a 1 - x ) + ln a 2 ( a 2 x - a 1 - x ) + ln a n ( a n x - a n - x ) .

When a i ≥ 1, and x [0, + ∞), ln a i 0 , a i x 1 , and a i - x 1 . Thus, ln a i ( a i x - a i - x ) 0 .

When 0 < a i < 1, and x [0, + ∞), we have ln a i < 0 , 0 < a i x 1 , and a i - x 1 . Thus, ln a i ( a i x - a i - x ) 0 . Therefore, f'(x) ≥ 0, and f(x) is increasing for x [0, ∞).

Theorem 2.4. Let A be a positive definite real matrix. Then
n n A 1 2 F tr A 1 2 μ F ( A ) .
(2.17)
Proof. Since A is positive definite, from Lemma 1.4, we have
cos ( A 1 / 2 , A 1 / 2 cos ( A 1 / 2 , I ) .
(2.18)
That is
n μ F ( A 1 2 ) = tr ( A 1 2 A - 1 2 ) | | A 1 2 | | F | | ( A 1 2 ) - 1 | | F tr A 1 2 n | | A 1 2 | | F .
(2.19)
Let all the positive real eigenvalues of A be λ1, λ2,..., λ n > 0. Then, the eigenvalues of A-1 are 1/λ1, 1/λ2,..., 1/λ n > 0, the eigenvalues of A2 are λ 1 2 , λ 2 2 , . . . , λ n 2 > 0 , and the eigenvalue of A-2 are λ 1 - 2 , λ 2 - 2 , . . . , λ n - 2 > 0 . Next, we will prove the following by induction.
i = 1 n λ i 2 i = 1 n 1 λ i 2 i = 1 n λ i i = 1 n 1 λ i .
(2.20)

In case n = 1, it is obvious that (2.20) holds.

In case n = 2 , i = 1 n λ i 2 i = 1 n 1 λ i 2 = 2 + ( λ 1 λ 2 ) 2 + ( λ 2 λ 1 ) 2 . From Lemma 2.3, 2 + (λ12)2 + (λ21)2 ≥ 2 + λ12 + λ21 = (λ1 + λ2)(1/λ1 + 1/λ2). Thus, (2.20) holds. Suppose that (2.20) holds, when n = k, i.e.,
i = 1 k λ i 2 i = 1 k 1 λ i 2 i = 1 k λ i i = 1 k 1 λ i .
(2.21)
In case n = k + 1,
( i = 1 k + 1 λ i 2 ) ( i = 1 k + 1 1 / λ i 2 ) = ( i = 1 k λ i 2 + λ k + 1 2 ) ( i = 1 k 1 / λ i 2 + 1 / λ k + 1 2 ) = ( i = 1 k λ i 2 ) ( i = 1 k 1 / λ i 2 ) + i = 1 k ( λ i / λ k + 1 ) 2 + i = 1 k ( λ k + 1 / λ i ) 2 + 1.
(2.22)
By Lemma 2.3, we get
i = 1 k + 1 λ i 2 i = 1 k + 1 1 λ i 2 i = 1 k λ i i = 1 k 1 λ i + i = 1 k ( λ i λ k + 1 ) + i = 1 k ( λ k + 1 λ i ) + 1 = i = 1 k + 1 λ i i = 1 k + 1 1 λ i .
(2.23)
Thus, when n = k + 1, (2.20) holds. On the other hand, | | A | | F = tr ( A 2 ) = i = 1 n λ i 2 , | | A - 1 | | F = tr A - 2 = i = 1 n λ i - 2 , | | A 1 2 | | F = trA = i = 1 n λ i and | | A - 1 2 | | F = tr A - 1 = i = 1 n λ i - 1 . Therefore,
| | A | | F 2 | | A - 1 | | F 2 | | A 1 2 | | F 2 | | A - 1 2 | | F 2 ,
(35)
i.e.,
μ F ( A ) μ F ( A 1 2 ) .
(2.25)
Taking (2.25) into (2.19), we obtain
n μ F ( A ) n μ F ( A 1 2 ) tr A 1 2 n | | A 1 2 | | F ,
(2.26)
i.e.,
n n | | A 1 2 | | F tr A 1 2 μ F ( A ) .
(2.27)

Remark 2.5. (2.17) can be extended to any 0 < α ≤ 1, i.e., n n | | A α | | F tr A α μ F ( A ) .

3 The Frobenius condition number of a centrosymmetric positive definite matrix

Definition 3.1 (see [7]). Let A = (a ij )p xq p×q. A is a centrosymmetric matrix, if
a i j = a p - i + 1 , q - j + 1 , 1 i p , 1 j q , o r J p A J q = A ,

where J n = (e n , en-1,..., e1), e i denotes the unit vector with the i-th entry 1.

Using the partition of matrix, the central symmetric character of a square centrosym-metric matrix can be described as follows (see [7]):

Lemma 3.2. Let A = (a ij )n×n(n = 2m) be centrosymmetric. Then, A has the form,
A = B J m C J m C J m B J m , P T A P = B - J m C 0 0 B + J m C ,
(3.1)

Where B,C m×m, P = 1 2 I m I m - J m J m .

Lemma 3.3. Let Abeann × n(n = 2m) centrosymmetric positive definite matrix with the following form
A = B J m C J m C J m B J m ,
where B,C m×m. Then there exists an orthogonal matrix P such that
P T A P = M N ,

where M = B - J m C, N = B + J m C and M-1, N-1are positive definite matrices.

Proof. From Lemma 3.2, there exists an orthogonal matrix P such that
H = P T A P = M N .

Since A is positive definite, then any eigenvalue of A is positive real number. Thus, the eigenvalues of H are all positive real numbers. That is to say, all eigenvalues of M and N are positive real numbers. Thus, M and N are both positive definite. It is obvious that M-1 and N-1 are positive definite matrices.

Lemma 3.4. Let A, B are positive definite real matrices. Then,
tr A ( detB ) 1 n | | A | | F | | B - 1 | | F .
(3.2)
Proof. Let X be positive definite. By Lemma 1.2, we have
< A , X > F < X , B > F | | X | | F 2 - < A , B > F 2 | | A | | F B | | F 2 .
(3.3)
Thus,
tr ( A T X ) tr ( X T B ) | | X | | F 2 - tr ( A T B ) 2 | | A | | F B | | F 2 .
(3.4)
By replacing X, B with I n and B-1, respectively, in inequality (3.4), we can obtain
tr ( A ) tr ( B - 1 ) n - tr ( A T B - 1 ) 2 | | A | | F | | B - 1 | | F 2 .
(3.5)
That is,
tr ( A ) tr ( B - 1 ) n - tr ( A T B - 1 ) 2 | | A | | F | | B - 1 | | F 2 .
(3.6)
From Schwarz inequality (1.5),
tr ( A T B - 1 ) = < A , B - 1 > F | | A | | F | | B - 1 | | F .
(3.7)
By taking (3.7) into (3.6), we have
tr ( A ) tr ( B - 1 ) n | | A | | F | | B - 1 | | F | | A | | F | | B - 1 | | F 2 + tr ( A T B - 1 ) 2 .
(3.8)
From (1.9),
n ( 1 detB ) 1 n tr B - 1 .
(3.9)
Therefore,
tr A ( det B ) 1 n | | A | | F | | B - 1 | | F .
(3.10)
Theorem 3.5. Let A be a centrosymmetric positive definite matrix with the form
A = B J m C J m C J m B J m , B , C m × m .
Let M = B - J m C, N = B + J m C. Then,
μ F ( A ) 2 tr M ( det M ) 1 m - m 2 + 2 tr N ( det N ) 1 m - m 2 + tr M ( det N ) 1 m 2 + tr N ( det M ) 1 m 2 .
(3.11)
Proof. By Lemma 3.2, there exists an orthogonal matrix P such that
H = P T A P = B - J m C B + J m C = M N .
Thus,
μ F ( A ) = | | A | | F | | A - 1 | | F = μ F ( H ) = | | H | | F | | H - 1 | | F = | | M | | F 2 + | | N | | F 2 | | M - 1 | | F 2 + | | N - 1 | | F 2 .
(3.12)
Therefore,
μ F 2 ( A ) = ( | | M | | F 2 + | | N | | F 2 ) ( | | M - 1 | | F 2 + | | N - 1 | | F 2 ) = | | M | | F 2 | | M - 1 | | F 2 + | | N | | F 2 | | N - 1 | | F 2 + | | M | | F 2 | | N - 1 | | F 2 + | | N | | F 2 | | M - 1 | | F 2 = μ F 2 ( M ) + μ F 2 ( N ) + ( | | M | | F | | N - 1 | | F ) 2 + ( | | N | | F | | M - 1 | | F ) 2 .
From Lemma 3.4,
tr M ( det N ) 1 m | | M | | F | | N - 1 | | F , tr N ( det M ) 1 m | | N | | F | | M - 1 | | F .
(3.14)
From (2.18),
2 tr M ( det M ) 1 m - m μ F ( M ) , and 2 tr N ( det N ) 1 m - m μ F ( N ) .
(3.15)
Thus,
μ F ( A ) 2 tr M ( det M ) 1 m - m 2 + 2 tr N ( det N ) 1 m - m 2 + tr M ( det N ) 1 m 2 + tr N ( det M ) 1 m 2 .
Example 3.6.
A = 5 0 0 3 0 5 3 0 0 3 5 0 3 0 0 5 , and P T A P = M N = 2 0 0 2 8 0 0 8 .
Here n = 4, m = 2, det (A) = 256, tr A = 20, tr M = 4, tr N = 16, det(M) = 4, det(N) = 64. From Theorem 3.5, a lower bound of μ F (A) is as follows:
μ F ( A ) 2 trM ( detM ) 1 m - m 2 + 2 trN ( detN ) 1 m - m 2 + trM ( detN ) 1 m 2 + trN ( detM ) 1 m 2 = 8 . 5 .
In fact
μ F ( A ) = μ F 2 ( M ) + μ F 2 ( N ) + ( | | M | | F | | N - 1 | | F ) 2 + ( | | N | | F | | M - 1 | | F ) 2 = 8 . 5 .
On the other hand, the lower bounds of μ F (A) in (2.15) and (2.16) provided by [6] are
μ F ( A ) 2 n tr A 3 2 tr A ( detA ) 1 2 n - n = 6 . 1823376 , and μ F ( A ) 2 tr A ( det A ) 1 n - n = 6 .

It can easily be seen that, in this example, the best lower bound is the first one given by Theorem 3.5.

Declarations

Acknowledgements

The authors would like to thank two anonymous referees for reading this article carefully, providing valuable suggestions and comments which have been implemented in this revised version. This study was supported by the National Science Foundation of China No. 11171013 and the Nation Science Foundation of China No. 60831001.

Authors’ Affiliations

(1)
LMIB, School of Mathematics and System Science, Beihang University, Beijing, P.R. China

References

  1. Zhang F: Matrix Theory: Basic Results and Techniques. Springer, New York; 1999.View ArticleGoogle Scholar
  2. Horn RA, Johnson CR: Matrix Analysis. Cambridge University Press, Cambridge; 1985.View ArticleGoogle Scholar
  3. Buzano ML: Gemeralizzazione della diseguaglianza di Cauchy-Schwarz. Rendiconti del Seminario Matematico Universita e Politecnico di Torino 1971, 31: 405–409. (1974) (Italian)MathSciNetGoogle Scholar
  4. Dragomir SS: Refinemens of Buzano's and Kurepa's inequalities in inner product spaces. Facta Universitatis 2005, 20: 65–73.MATHGoogle Scholar
  5. Danna FM: Matrix and operator inequalites. J Inequal Pure Appl Math 2001.,2(3): article 34Google Scholar
  6. Ramazan T, Zübeyde U: On the Frobenius condition number of positive definite matrices. J Inequal Appl 2010. Article ID 897279 (2010)Google Scholar
  7. Liu ZY: Some properties of centrosymmetric matrices. Appl Math Comput 2002, 141: 17–26.MathSciNetGoogle Scholar

Copyright

© Li et al; licensee Springer. 2011

This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.