Functions of Matrices Nicholas J. Higham 2025 Scribd Download
Functions of Matrices Nicholas J. Higham 2025 Scribd Download
com
https://ebookgate.com/product/functions-of-matrices-
nicholas-j-higham/
OR CLICK HERE
DOWLOAD NOW
https://ebookgate.com/product/physics-of-the-piano-1st-edition-
nicholas-j-giordano/
ebookgate.com
https://ebookgate.com/product/dangerous-frames-nicholas-j-g-winter/
ebookgate.com
https://ebookgate.com/product/visual-perception-3rd-edition-nicholas-
j-wade/
ebookgate.com
https://ebookgate.com/product/totally-nonnegative-matrices-shaun-m-
fallat/
ebookgate.com
Diary of a Disaster British Aid to Greece 1940 1941 Higham
https://ebookgate.com/product/diary-of-a-disaster-british-aid-to-
greece-1940-1941-higham/
ebookgate.com
https://ebookgate.com/product/traffic-highway-engineering-fouth-
edition-nicholas-j-garber/
ebookgate.com
https://ebookgate.com/product/nist-handbook-of-mathematical-
functions-1-pap-cdr-edition-frank-w-j-olver/
ebookgate.com
https://ebookgate.com/product/120-banned-books-censorship-histories-
of-world-literature-second-edition-nicholas-j-karolides/
ebookgate.com
https://ebookgate.com/product/pediatric-hematology-methods-and-
protocols-1st-edition-nicholas-j-goulden/
ebookgate.com
ot104_HighamFM-B:Gockenbach 2/8/2008 2:47 PM Page 1
Functions
of Matrices
ot104_HighamFM-B:Gockenbach 2/8/2008 2:47 PM Page 2
ot104_HighamFM-B:Gockenbach 2/8/2008 2:47 PM Page 3
Functions
of Matrices
Theory and Computation
Nicholas J. Higham
University of Manchester
Manchester, United Kingdom
10 9 8 7 6 5 4 3 2 1
All rights reserved. Printed in the United States of America. No part of this book may
be reproduced, stored, or transmitted in any manner without the written permission of the
publisher. For information, write to the Society for Industrial and Applied Mathematics,
3600 Market Street, 6th Floor, Philadelphia, PA 19104-2688 USA.
Trademarked names may be used in this book without the inclusion of a trademark symbol.
These names are used in an editorial context only; no infringement of trademark is intended.
QA188.H53 2008
512.9'434--dc22
2007061811
is a registered trademark.
ot104_HighamFM-B:Gockenbach 2/8/2008 2:47 PM Page 5
To Françoise
ot104_HighamFM-B:Gockenbach 2/8/2008 2:47 PM Page 6
Contents
List of Tables xv
Preface xvii
2 Applications 35
2.1 Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.1.1 Exponential Integrators . . . . . . . . . . . . . . . . . . . . . . 36
2.2 Nuclear Magnetic Resonance . . . . . . . . . . . . . . . . . . . . . . . 37
2.3 Markov Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4 Control Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.5 The Nonsymmetric Eigenvalue Problem . . . . . . . . . . . . . . . . . 41
2.6 Orthogonalization and the Orthogonal Procrustes Problem . . . . . . 42
2.7 Theoretical Particle Physics . . . . . . . . . . . . . . . . . . . . . . . 43
2.8 Other Matrix Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.9 Nonlinear Matrix Equations . . . . . . . . . . . . . . . . . . . . . . . 44
2.10 Geometric Mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.11 Pseudospectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.12 Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
vii
viii Contents
3 Conditioning 55
3.1 Condition Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2 Properties of the Fréchet Derivative . . . . . . . . . . . . . . . . . . . 57
3.3 Bounding the Condition Number . . . . . . . . . . . . . . . . . . . . 63
3.4 Computing or Estimating the Condition Number . . . . . . . . . . . 64
3.5 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
14 Miscellany 313
14.1 Structured Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
14.1.1 Algebras and Groups . . . . . . . . . . . . . . . . . . . . . . . 313
14.1.2 Monotone Functions . . . . . . . . . . . . . . . . . . . . . . . 315
14.1.3 Other Structures . . . . . . . . . . . . . . . . . . . . . . . . . 315
14.1.4 Data Sparse Representations . . . . . . . . . . . . . . . . . . . 316
14.1.5 Computing Structured f (A) for Structured A . . . . . . . . . 316
14.2 Exponential Decay of Functions of Banded Matrices . . . . . . . . . . 317
14.3 Approximating Entries of Matrix Functions . . . . . . . . . . . . . . . 318
A Notation 319
Bibliography 379
Index 415
List of Figures
3.1 Relative errors in the Frobenius norm for the finite difference approx-
imation (3.22) to the Fréchet derivative. . . . . . . . . . . . . . . . . 68
6.1 The cardioid (6.45), shaded, together with the unit circle . . . . . . . 157
7.1 Convergence of the Newton iteration (7.6) for a pth root of unity . . 179
7.2 Regions of a ∈ C for which the inverse Newton iteration (7.15) con-
verges to a−1/p . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
9.1 Normwise relative errors for funm mod and condrel (exp, A)u. . . . . . 230
12.1 Normwise relative errors for Algorithm 12.6, MATLAB’s funm, and
Algorithm 12.7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
12.2 Same data as in Figure 12.1 presented as a performance profile. . . . 297
12.3 Normwise relative errors for Algorithm 12.6, Algorithm 12.7, Algo-
rithm 12.8, funm, and sine obtained as shifted cosine from Algo-
rithm 12.6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
xiii
List of Tables
5.1 Iteration functions fℓm from the Padé family (5.27). . . . . . . . . . . 116
5.2 Number of iterations for scaled Newton iteration. . . . . . . . . . . . 125
5.3 Newton iteration with spectral scaling for Jordan block J(2) ∈ R16×16 . 126
5.4 Newton iteration with determinantal scaling for random A ∈ R16×16
with κ2 (A) = 1010 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.5 Newton iteration with determinantal scaling for random A ∈ R16×16
with real eigenvalues parametrized by d. . . . . . . . . . . . . . . . . 127
xv
xvi List of Tables
11.1 Maximal values θm of kXk such that the bound (11.19) ensures krm (X)−
log(I +X)k does not exceed u = 2−53 , along with upper bound (11.20)
for κ(qm (X)) and upper bound (11.21) for φm , both with kXk = θm . 277
B.1 Constants αpq such that kAkp ≤ αpq kAkq , A ∈ Cm×n . . . . . . . . . . 327
Functions of matrices have been studied for as long as matrix algebra itself. Indeed,
in his seminal A Memoir on the Theory of Matrices (1858), Cayley investigated the
square root of a matrix, and it was not long before definitions of f (A) for general f
were proposed by Sylvester and others. From their origin in pure mathematics, ma-
trix functions have broadened into a subject of study in applied mathematics, with
widespread applications in science and engineering. Research on matrix functions in-
volves matrix theory, numerical analysis, approximation theory, and the development
of algorithms and software, so it employs a wide range of theory and methods and
promotes an appreciation of all these important topics.
My first foray into f (A) was as a graduate student when I became interested in
the matrix square root. I have worked on matrix functions on and off ever since.
Although there is a large literature on the subject, including chapters in several
books (notably Gantmacher [203, ], Horn and Johnson [296, ], Lancaster
and Tismenetsky [371, ], and Golub and Van Loan [224, ]), there has not
previously been a book devoted to matrix functions. I started to write this book in
2003. In the intervening period interest in matrix functions has grown significantly,
with new applications appearing and the literature expanding at a fast rate, so the
appearance of this book is timely.
This book is a research monograph that aims to give a reasonably complete treat-
ment of the theory of matrix functions and numerical methods for computing them,
as well as an overview of applications. The theory of matrix functions is beautiful and
nontrivial. I have strived for an elegant presentation with illuminating examples, em-
phasizing results of practical interest. I focus on three equivalent definitions of f (A),
based on the Jordan canonical form, polynomial interpolation, and the Cauchy inte-
gral formula, and use all three to develop the theory. A thorough treatment is given
of problem sensitivity, based on the Fréchet derivative. The applications described
include both the well known and the more speculative or recent, and differential
equations and algebraic Riccati equations underlie many of them.
The bulk of the book is concerned with numerical methods and the associated
issues of accuracy, stability, and computational cost. Both general purpose methods
and methods for specific functions are covered. Little mention is made of methods
that are numerically unstable or have exorbitant operation counts of order n4 or
higher; many methods proposed in the literature are ruled out for at least one of
these reasons.
The focus is on theory and methods for general matrices, but a brief introduction
to functions of structured matrices is given in Section 14.1. The problem of computing
a function of a matrix times a vector, f (A)b, is of growing importance, though as yet
numerical methods are relatively undeveloped; Chapter 13 is devoted to this topic.
One of the pleasures of writing this book has been to explore the many connec-
tions between matrix functions and other subjects, particularly matrix analysis and
numerical analysis in general. These connections range from the expected, such as
xvii
xviii Preface
divided differences, the Kronecker product, and unitarily invariant norms, to the un-
expected, which include the Mandelbrot set, the geometric mean, partial isometries,
and the role of the Fréchet derivative beyond measuring problem sensitivity.
I have endeavoured to make this book more than just a monograph about matrix
functions, and so it includes many useful or interesting facts, results, tricks, and
techniques that have a (sometimes indirect) f (A) connection. In particular, the book
contains a substantial amount of matrix theory, as well as many historical references,
some of which appear not to have previously been known to researchers in the area.
I hope that the book will be found useful as a source of statements and applications
of results in matrix analysis and numerical linear algebra, as well as a reference on
matrix functions.
Four main themes pervade the book.
Role of the sign function. The matrix sign function has fundamental theoretical
and algorithmic connections with the matrix square root, the polar decomposition,
and, to a lesser extent, matrix pth roots. For example, a large class of iterations for
the matrix square root can be obtained from corresponding iterations for the matrix
sign function, and Newton’s method for the matrix square root is mathematically
equivalent to Newton’s method for the matrix sign function.
Stability. The stability of iterations for matrix functions can be effectively defined
and analyzed in terms of power boundedness of the Fréchet derivative of the iteration
function at the solution. Unlike some earlier, more ad hoc analyses, no assumptions
are required on the underlying matrix. General results (Theorems 4.18 and 4.19)
simplify the analysis for idempotent functions such as the matrix sign function and
the unitary polar factor.
Schur decomposition and Parlett recurrence. The use of a Schur decomposition
followed by reordering and application of the block form of the Parlett recurrence
yields a powerful general algorithm, with f -dependence restricted to the evaluation
of f on the diagonal blocks of the Schur form.
Padé approximation. For transcendental functions the use of Padé approximants,
in conjunction with an appropriate scaling technique that brings the matrix argument
close to the origin, yields an effective class of algorithms whose computational building
blocks are typically just matrix multiplication and the solution of multiple right-hand
side linear systems. Part of the success of this approach rests on the several ways
in which rational functions can be evaluated at a matrix argument, which gives the
scope to find a good compromise between speed and stability.
In addition to surveying, unifying, and sometimes improving existing results and
algorithms, this book contains new results. Some of particular note are as follows.
• Theorem 1.35, which relates f (αIm + AB) to f (αIn + BA) for A ∈ Cm×n and
B ∈ Cn×m and is an analogue for general matrix functions of the Sherman–
Morrison–Woodbury formula for the matrix inverse.
• Theorem 4.15, which shows that convergence of a scalar iteration implies con-
vergence of the corresponding matrix iteration when applied to a Jordan block,
under suitable assumptions. This result is useful when the matrix iteration
can be block diagonalized using the Jordan canonical form of the underlying
matrix, A. Nevertheless, we show in the context of Newton’s method for the
matrix square root that analysis via the Jordan canonical form of A does not
always give the strongest possible convergence result. In this case a stronger
result, Theorem 6.9, is obtained essentially by reducing the convergence analysis
to the consideration of the behaviour of the powers of a certain matrix.
Preface xix
• Theorems 5.13 and 8.19 on the stability of essentially all iterations for the ma-
trix sign function and the unitary polar factor, and the general results in The-
orems 4.18 and 4.19 on which these are based.
• Theorems 6.14–6.16 on the convergence of the binomial, Pulay, and Visser iter-
ations for the matrix square root.
• An improved Schur–Parlett algorithm for the matrix logarithm, given in Sec-
tion 11.6, which makes use of improved implementations of the inverse scaling
and squaring method in Section 11.5.
The Audience
The book’s main audience is specialists in numerical analysis and applied linear al-
gebra, but it will be of use to anyone who wishes to know something of the theory
of matrix functions and state of the art methods for computing them. Much of the
book can be understood with only a basic grounding in numerical analysis and linear
algebra.
Acknowledgments
A number of people have influenced my thinking about matrix functions. Discussions
with Ralph Byers in 1984, when he was working on the matrix sign function and I was
investigating the polar decomposition, first made me aware of connections between
these two important tools. The work on the matrix exponential of Cleve Moler and
Charlie Van Loan has been a frequent source of inspiration. Beresford Parlett’s ideas
on the exploitation of the Schur form and the adroit use of divided differences have
been a guiding light. Charles Kenney and Alan Laub’s many contributions to the
matrix function arena have been important in my own research and are reported on
many pages of this book. Finally, Nick Trefethen has shown me the importance of the
Cauchy integral formula and has offered valuable comments on drafts at all stages.
I am grateful to several other people for providing valuable help, suggestions, or
advice during the writing of the book:
Rafik Alam, Awad Al-Mohy, Zhaojun Bai, Timo Betcke, Rajendra Bhatia,
Tony Crilly, Philip Davies, Oliver Ernst, Andreas Frommer, Chun-Hua Guo,
Gareth Hargreaves, Des Higham, Roger Horn, Bruno Iannazzo, Ilse Ipsen,
Peter Lancaster, Jörg Liesen, Lijing Lin, Steve Mackey, Roy Mathias,
Volker Mehrmann, Thomas Schmelzer, Gil Strang, Françoise Tisseur, and
Andre Weideman.
Working with the SIAM staff on the publication of this book has been a pleasure. I
thank, in particular, Elizabeth Greenspan (acquisitions), Sara Murphy (acquisitions),
Lois Sellers (design), and Kelly Thomas (copy editing).
Research leading to this book has been supported by the Engineering and Physical
Sciences Research Council, The Royal Society, and the Wolfson Foundation.
In this first chapter we give a concise treatment of the theory of matrix functions,
concentrating on those aspects that are most useful in the development of algorithms.
Most of the results in this chapter are for general functions. Results specific to
particular functions can be found in later chapters devoted to those functions.
1.1. Introduction
The term “function of a matrix” can have several different meanings. In this book we
are interested in a definition that takes a scalar function f and a matrix A ∈ Cn×n
and specifies f (A) to be a matrix of the same dimensions as A; it does so in a way
that provides a useful generalization of the function of a scalar variable f (z), z ∈ C.
Other interpretations of f (A) that are not our focus here are as follows:
• Elementwise operations on matrices, for example sin A = (sin aij ). These oper-
ations are available in some programming languages. For example, Fortran 95
supports “elemental operations” [423, ], and most of MATLAB’s elemen-
tary and special functions are applied in an elementwise fashion when given
matrix arguments. However, elementwise operations do not integrate well with
matrix algebra, as is clear from the fact that the elementwise square of A is not
equal to the matrix product of A with itself. (Nevertheless, the elementwise
product of two matrices, known as the Hadamard product or Schur product, is
a useful concept [294, ], [296, , Chap. 5].)
• Functions producing a scalar result, such as the trace, the determinant, the
spectral radius, the condition number κ(A) = kAk kA−1 k, and one particular
generalization to matrix arguments of the hypergeometric function [359, ].
• Functions mapping Cn×n to Cm×m that do not stem from a scalar function.
Examples include matrix polynomials with matrix coefficients, the matrix trans-
pose, the adjugate (or adjoint) matrix, compound matrices comprising minors
of a given matrix, and factors from matrix factorizations. However, as a special
case, the polar factors of a matrix are treated in Chapter 8.
Before giving formal definitions, we offer some motivating remarks. When f (t)
is a polynomial or rational function with scalar coefficients and a scalar argument,
t, it is natural to define f (A) by substituting A for t, replacing division by matrix
1
2 Theory of Matrix Functions
inversion (provided that the matrices to be inverted are nonsingular), and replacing
1 by the identity matrix. Then, for example,
1 + t2
f (t) = ⇒ f (A) = (I − A)−1 (I + A2 ) if 1 ∈
/ Λ(A).
1−t
Here, Λ(A) denotes the set of eigenvalues of A (the spectrum of A). Note that rational
functions of a matrix commute, so it does not matter whether we write (I − A)−1 (I +
A2 ) or (I + A2 )(I − A)−1 . If f has a convergent power series representation, such as
t2 t3 t4
log(1 + t) = t − + − + ···, |t| < 1,
2 3 4
we can again simply substitute A for t to define
A2 A3 A4
log(I + A) = A − + − + ···, ρ(A) < 1. (1.1)
2 3 4
Here, ρ denotes the spectral radius and the condition ρ(A) < 1 ensures convergence of
the matrix series (see Theorem 4.7). In this ad hoc fashion, a wide variety of matrix
functions can be defined. However, this approach has several drawbacks:
Z −1 AZ = J = diag(J1 , J2 , . . . , Jp ), (1.2a)
λk 1
..
λk .
Jk = Jk (λk ) = . ∈ Cmk ×mk , (1.2b)
.. 1
λk
1.2 Definitions of f (A) 3
Definition 1.2 (matrix function via Jordan canonical form). Let f be defined on
the spectrum of A ∈ Cn×n and let A have the Jordan canonical form (1.2). Then
where
′ f (mk −1) )(λk )
f (λk ) f (λk ) ...
(mk − 1)!
.. ..
f (Jk ) :=
f (λk ) . . .
(1.4)
..
. ′
f (λk )
f (λk )
1/2 1
A simple example illustrates the definition. For the Jordan block J = 0 1/2
and f (x) = x3 , (1.4) gives
f (1/2) f ′ (1/2) 1/8 3/4
f (J) = = ,
0 f (1/2) 0 1/8
[371, , Chap. 9]. Note that the values depend not just on the eigenvalues but also on the maximal
Jordan block sizes ni .
4 Theory of Matrix Functions
Finally, we explain how (1.4) can be obtained from Taylor series considerations.
In (1.2b) write Jk = λk I +Nk ∈ Cmk ×mk , where Nk is zero except for a superdiagonal
of 1s. Note that for mk = 3 we have
0 1 0 0 0 1
Nk = 0 0 1 , Nk2 = 0 0 0 , Nk3 = 0.
0 0 0 0 0 0
In general, powering Nk causes the superdiagonal of 1s to move a diagonal at a time
towards the top right-hand corner, until at the mk th power it disappears: Ekmk = 0;
so Nk is nilpotent. Assume that f has a convergent Taylor series expansion
f (j) (λk )(t − λk )j
f (t) = f (λk ) + f ′ (λk )(t − λk ) + · · · + + ···.
j!
On substituting Jk ∈ Cmk ×mk for t we obtain the finite series
f (mk −1) (λk )Nkmk −1
f (Jk ) = f (λk )I + f ′ (λk )Nk + · · · + , (1.5)
(mk − 1)!
since all powers of Nk from the mk th onwards are zero. This expression is easily seen
to agree with (1.4). An alternative derivation of (1.5) that does not rest on a Taylor
series is given in the next section.
Definition 1.2 requires the function f to take well-defined values on the spectrum
of A—including values associated
√ with derivatives, where appropriate. Thus in the
case of functions such as t and log t it is implicit that a single branch has been
chosen in (1.4). Moreover, if an eigenvalue occurs in more than one Jordan block
then the same choice of branch must be made in each block. If the latter requirement
is violated then a nonprimary matrix function is obtained, as discussed in Section 1.4.
Theorem 1.3. For polynomials p and q and A ∈ Cn×n , p(A) = q(A) if and only if
p and q take the same values on the spectrum of A.
Proof. Suppose that two polynomials p and q satisfy p(A) = q(A). Then d = p−q
is zero at A so is divisible by the minimal polynomial ψ. In other words, d takes only
the value zero on the spectrum of A, that is, p and q take the same values on the
spectrum of A.
Conversely, suppose p and q take the same values on the spectrum of A. Then
d = p − q is zero on the spectrum of A and so must be divisible by the minimum
polynomial ψ, in view of (1.6). Hence d = ψr for some polynomial r, and since
d(A) = ψ(A)r(A) = 0, it follows that p(A) = q(A).
Thus it is a property of polynomials that the matrix p(A) is completely determined
by the values of p on the spectrum of A. It is natural to generalize this property to
arbitrary functions and define f (A) in such a way that f (A) is completely determined
by the values of f on the spectrum of A.
Q
where φi (t) = f (t)/ j6=i (t − λj )nj . For a matrix with distinct eigenvalues (ni ≡ 1,
s = n) this formula reduces to the familiar Lagrange form
n
X Yn
t − λj
p(t) = f (λi )ℓi (t), ℓi (t) = . (1.9)
i=1 j=1
λi − λj
j6=i
where m = deg ψ and the set {xi }m i=1 comprises the distinct eigenvalues λ1 , . . . , λs
with λi having multiplicity ni . Here the f [. . .] denote divided differences, which are
defined in Section B.16. Another polynomial q for which f (A) = q(A) is given by
(1.10) with m = n and {xi }ni=1 the set of all n eigenvalues of A:
Remark 1.8. If f is given by a power series, Definition 1.4 says that f (A) is never-
theless expressible as a polynomial in A of degree at most n−1. Another way to arrive
at this conclusion is as follows. The Cayley–Hamilton theorem says that any matrix
1.2 Definitions of f (A) 7
satisfies its own characteristic equation: q(A) = 0,2 where q(t) = det(tI − A) is the
characteristic polynomial. This theorem follows immediately from the fact that the
minimal polynomial ψ divides q (see Problem 1.18 for another proof). Hence the nth
power of A, and inductively all higher powers, are expressible as a linear combination
of I, A, . . . , An−1 . Thus any power series in A can be reduced to a polynomial in A
of degree at most n − 1. This polynomial is rarely of an elegant form or of practical
interest; exceptions are given in (1.16) and Problem 10.13.
with a real coefficient matrix and right-hand side. We conclude that r has real
coefficients and hence f (A) = p(A) is real when A is real. This argument extends to
real n × n matrices under the stated condition on f . As a particular example, we can
conclude that if A is real and nonsingular with no eigenvalues on the negative real
axis then A has a real square root and a real logarithm. For a full characterization
of the existence of real square roots and logarithms see Theorem 1.23. Equivalent
conditions to f (A) being real for real A when f is analytic are given in Theorem 1.18.
Remark 1.10. We can derive directly from Definition 1.4 the formula (1.4) for a
function of the Jordan block Jk in (1.2). It suffices to note that the interpolation
conditions are p(j) (λk ) = f (j) (λk ), j = 0: mk − 1, so that the required Hermite
interpolating polynomial is
and then to evaluate p(Jk ), making use of the properties of the powers of Nk noted
in the previous section (cf. (1.5)).
The integrand contains the resolvent, (zI − A)−1 , which is defined on Γ since Γ
is disjoint from the spectrum of A.
This definition leads to short proofs of certain theoretical results and has the
advantage that it can be generalized to operators.
Theorem 1.12. Definition 1.2 (Jordan canonical form) and Definition 1.4 (Hermite
interpolation) are equivalent. If f is analytic then Definition 1.11 (Cauchy integral )
is equivalent to Definitions 1.2 and 1.4.
Proof. Definition 1.4 says that f (A) = p(A) for a Hermite interpolating poly-
nomial p satisfying (1.7). If A has the Jordan form (1.2) then f (A) = p(A) =
p(ZJZ −1 ) = Zp(J)Z −1 = Z diag(p(Jk ))Z −1 , just from elementary properties of ma-
trix polynomials. But since p(Jk ) is completely determined by the values of p on the
spectrum of Jk , and these values are a subset of the values of p on the spectrum of A,
it follows from Remark 1.5 and Remark 1.10 that p(Jk ) is precisely (1.4). Hence the
matrix f (A) obtained from Definition 1.4 agrees with that given by Definition 1.2.
For the equivalence of Definition 1.11 with the other two definitions, see Horn and
Johnson [296, , Thm. 6.2.28].
We will mainly use (for theoretical purposes) Definitions 1.2 and 1.4. The polyno-
mial interpolation definition, Definition 1.4, is well suited to proving basic properties
of matrix functions, such as those in Section 1.3, while the Jordan canonical form
definition, Definition 1.2, excels for solving matrix equations such as X 2 = A and
eX = A. For many purposes, such as the derivation of the formulae in the next
section, either of the definitions can be used.
In the rest of the book we will refer simply to “the definition of a matrix function”.
and so
t − v∗ u t−0
p(t) = f (0) + ∗ f (v ∗ u).
0 − v∗ u v u−0
Hence
f (0) uv ∗
f (A) = p(A) = − ∗ uv ∗ + f (0)I + f (v ∗ u) ∗
v u∗ v u
f (v u) − f (0)
= f (0)I + uv ∗ (1.13)
v∗ u − 0
= f (0)I + f [v ∗ u, 0] uv ∗ .
We have manipulated the expression into this form involving a divided difference
because it is suggestive of what happens when v ∗ u = 0. Indeed f [0, 0] = f ′ (0) and so
when v ∗ u = 0 we may expect that f (A) = f (0)I + f ′ (0)uv ∗ . To confirm this formula,
note that v ∗ u = 0 implies that the spectrum of A consists entirely of 0 and that
A2 = (v ∗ u)uv ∗ = 0. Hence, assuming A 6= 0, A must have one 2 × 2 Jordan block
corresponding to the eigenvalue 0, with the other n − 2 zero eigenvalues occurring in
1 × 1 Jordan blocks. The interpolation conditions (1.7) are therefore
is valid for all u and v. We could have obtained this formula directly by using the
divided difference form (1.10) of the Hermite interpolating polynomial r, but the
derivation above gives more insight.
We now show how the formula is obtained from Definition 1.2 when v ∗ u 6= 0 (for
the case v ∗ u = 0 see Problem 1.15). The Jordan canonical form can be written as
∗ ∗
v /(v u)
A = [ u X ] diag(v ∗ u, 0, . . . , 0) ,
Y
Hence
v ∗ /(v ∗ u) uv ∗
f (A) = [ u X ] diag(f (v ∗ u), f (0), . . . , f (0)) = f (v ∗ u) ∗ + f (0)XY.
Y v u
For a more general result involving a perturbation of arbitrary rank see Theorem 1.35.
1
p(t) = f (1) (t + 1) (t − i) (t + i) − f (−1) (t − 1) (t − i) (t + i)
4
+ if (i) (t − 1) (t + 1) (t + i) − if (−i) (t − 1) (t + 1) (t − i) . (1.18)
Thus f (A) = p(A), and in fact this formula holds even for n = 1: 3, since incorpo-
rating extra interpolation conditions does not affect the ability of the interpolating
polynomial to yield f (A) (see Remark 1.5). This expression can be quickly evaluated
in O(n2 log n) operations because multiplication of a vector by Fn can be carried out
in O(n log n) operations using the fast Fourier transform (FFT).
Because Fn is unitary and hence normal, Fn is unitarily diagonalizable: Fn =
QDQ∗ for some unitary Q and diagonal D. (Indeed, any matrix with minimal polyno-
mial ψ(t) has distinct eigenvalues and so is diagonalizable.) Thus f (Fn ) = Qf (D)Q∗ .
However, this formula requires knowledge of Q and D and so is much more compli-
cated to use than (1.18).
1.3. Properties
The sign of a good definition is that it leads to the properties one expects or hopes
for, as well as some useful properties that are less obvious. We collect some general
properties that follow from the definition of f (A).
Theorem 1.13. Let A ∈ Cn×n and let f be defined on the spectrum of A. Then
(a) f (A) commutes with A;
(b) f (AT ) = f (A)T ;
(c) f (XAX −1 ) = Xf (A)X −1 ;
(d) the eigenvalues of f (A) are f (λi ), where the λi are the eigenvalues of A;
1.3 Properties 11
Proof. Definition 1.4 implies that f (A) is a polynomial in A, p(A) say. Then
f (A)A = p(A)A = Ap(A) = Af (A), which proves the first property. For (b) we have
f (A)T = p(A)T = p(AT ) = f (AT ), where the last equality follows from the fact that
the values of f on the spectrum of A are the same as the values of f on the spectrum of
AT . (c) and (d) follow immediately from Definition 1.2. (e) follows from (c) when X is
nonsingular; more generally it is obtained from Xf (A) = Xp(A) = p(A)X = f (A)X.
For (f), f (A) = p(A) is clearly block triangular and its ith diagonal block is p(Aii ).
Since p interpolates f on the spectrum of A it interpolates f on the spectrum of each
Aii , and hence p(Aii ) = f (Aii ). (g) is a special case of (f). (h) is a special case of
(g), since Im ⊗ A = diag(A, A, . . . , A). Finally, we have A ⊗ B = Π(B ⊗ A)Π T for a
permutation matrix Π, and so
Theorem 1.14 (equality of two matrix functions). With the notation of Section 1.2,
f (A) = g(A) if and only if
f (j) (λi ) = 0, j = 0: ni − 1, i = 1: s.
Theorem 1.15 (sum and product of functions). Let f and g be functions defined on
the spectrum of A ∈ Cn×n .
(a) If h(t) = f (t) + g(t) then h(A) = f (A) + g(A).
(b) If h(t) = f (t)g(t) then h(A) = f (A)g(A).
12 Theory of Matrix Functions
Proof. Part (a) is immediate from any of the definitions of h(A). For part
(b), let p and q interpolate f and g on the spectrum of A, so that p(A) = f (A)
and q(A) = g(A). By differentiating and using the product rule we find that the
functions h(t) and r(t) = p(t)q(t) have the same values on the spectrum of A. Hence
h(A) = r(A) = p(A)q(A) = f (A)g(A).
The next result generalizes the previous one and says that scalar functional re-
lationships of a polynomial nature are preserved by matrix functions. For example
sin2 (A) + cos2 (A) = I, (A1/p )p = A, and eiA = cos(A) + i sin(A). Of course, gener-
alizations of scalar identities that involve two or more noncommuting matrices may
fail; for example, eA+B , eA eB , and eB eA are in general all different (see Section 10.1).
Theorem 1.17 (composite function). Let A ∈ Cn×n and let the distinct eigenvalues
of A be λ1 , . . . , λs with indices n1 , . . . , ns . Let h be defined on the spectrum of A (so
that the values h(j) (λi ), j = 0: ni − 1, i = 1: s exist) and let the values g (j) (h(λi )),
j = 0: ni − 1, i = 1: s exist. Then f (t) = g(h(t)) is defined on the spectrum of A and
f (A) = g(h(A)).
and all the derivatives on the right-hand side exist, f is defined on the spectrum of A.
Let p(t) be any polynomial satisfying the interpolation conditions
From Definition 1.2 it is clear that the indices of the eigenvalues µ1 , . . . , µs of h(A)
are at most n1 , . . . , ns , so the values on the right-hand side of (1.20) contain the
values of g on the spectrum of B = h(A); thus g(B) is defined and p(B) = g(B). It
now follows by (1.19) and (1.20) that the values of f (t) and p(h(t)) coincide on the
spectrum of A. Hence by applying Theorem 1.16 to Q(f (t), h(t)) = f (t) − p(h(t)) we
conclude that
f (A) = p(h(A)) = p(B) = g(B) = g(h(A)),
1.3 Properties 13
as required.
The assumptions in Theorem 1.17 on g for f (A) to exist are stronger than neces-
sary in certain cases where a Jordan block of Asplits
under evaluation of h. Consider,
for example, g(t) = t1/3 , h(t) = t2 , and A = 00 10 . The required derivative g ′ (0) in
Theorem 1.17 does not exist, but f (A) = (A2 )1/3 = 0 nevertheless does exist. (A
full description of the Jordan canonical form of f (A) in terms of that of A is given in
Theorem 1.36.)
Theorem 1.17 implies that exp(log A) = A, provided that log is defined on the
spectrum of A. However, log(exp(A)) = A does not hold unless the spectrum of
A satisfies suitable restrictions, since the scalar relation log(et ) = t is likewise not
generally true in view of et = et+2kπi for any integer k; see Problem 1.39.
Although f (AT ) = f (A)T always holds (Theorem 1.13 (b)), the property f (A∗ ) =
f (A)∗ does not. The next result says essentially that for an analytic function f
defined on a suitable domain that includes a subset S of the real line, f (A∗ ) = f (A)∗
holds precisely when f maps S back into the real line. This latter condition also
characterizes when A real implies f (A) real (cf. the sufficient conditions given in
Remark 1.9).
Theorem 1.18 (Higham, Mackey, Mackey, and Tisseur). Let f be analytic on an open
subset Ω ⊆ C such that each connected component of Ω is closed under conjugation.
Consider the corresponding matrix function f on its natural domain in Cn×n , the set
D = { A ∈ Cn×n : Λ(A) ⊆ Ω }. Then the following are equivalent:
(a) f (A∗ ) = f (A)∗ for all A ∈ D.
(b) f (A) = f (A) for all A ∈ D.
(c) f (Rn×n ∩ D) ⊆ Rn×n .
(d) f (R ∩ Ω) ⊆ R.
Proof. The first two properties are obviously equivalent, in view of Theorem 1.13 (b).
Our strategy is therefore to show that (b) ⇒ (c) ⇒ (d) ⇒ (b).
(b) ⇒ (c): If A ∈ Rn×n ∩ D then
Proof. See Horn and Johnson [296, , Thm. 6.2.27 (1)], and Mathias [412,
, Lem. 1.1] for the conditions as stated here.
14 Theory of Matrix Functions
For continuity of f (A) on the set of normal matrices just the continuity of f is
sufficient [296, , Thm. 6.2.37].
Our final result shows that under mild conditions to check the veracity of a matrix
identity it suffices to check it for diagonalizable matrices.
Theorem 1.20. Let f satisfy the conditions of Theorem 1.19. Then f (A) = 0 for all
A ∈ Cn×n with spectrum in D if and only if f (A) = 0 for all diagonalizable A ∈ Cn×n
with spectrum in D.
Proof. See Horn and Johnson [296, , Thm. 6.2.27 (2)].
For an example of the use of Theorem 1.20 see the proof of Theorem 11.1. The-
orem 1.13 (f) says that block triangular structure is preserved by matrix functions.
An explicit formula can be given for an important instance of the block 2 × 2 case.
Theorem 1.21. Let f satisfy the conditions of Theorem 1.19 with D containing the
spectrum of n−1 1
n−1 B c
A= ∈ Cn×n .
1 0 λ
Then
f (B) g(B)c
f (A) = , (1.21)
0 f (λ)
/ Λ(B) then g(B) = (B − λI)−1 (f (B) −
where g(z) = f [z, λ]. In particular, if λ ∈
f (λ)I).
Proof. We need only to demonstrate the formula for the (1,2) block F12 of f (A).
Equating (1,2) blocks in f (A)A = Af (A) (Theorem 1.13 (a)) yields BF12 + cf (λ) =
f (B)c + F12 λ, or (B − λI)F12 = (f (B) − f (λ)I)c. If λ ∈
/ Λ(B) the result is proved.
Otherwise, the result follows by a continuity argument: replace λ by λ(ǫ) = λ + ǫ, so
that λ(ǫ) ∈
/ Λ(B) for sufficiently small ǫ, let ǫ → 0, and use the continuity of divided
differences and of f (A).
For an expression for a function of a general block 2 × 2 block triangular matrix
see Theorem 4.12.
√
that is, solve X 2 = A. Taking f (t) = t, the interpolation
√ conditions in Defini-
tions 1.4 are (with s = 1, n1 = 1) simply p(1) = 1. The interpolating polynomial
is therefore either p(t) = 1 or p(t) = −1, corresponding to the two square roots of
1, giving I and −I as square roots of A. Both of these square roots are, trivially,
polynomials in A. Turning to Definition 1.2, the matrix A is already in Jordan form
with two 1 × 1 Jordan blocks, and the definition provides the same two square roots.
However, if we ignore the prescription at the end of Section 1.2.1 about the choice of
branches then we can obtain two more square roots,
−1 0 1 0
, ,
0 1 0 −1
in which the two eigenvalues 1 have been sent to different square roots. Moreover,
since A = ZIZ −1 is a Jordan canonical form for any nonsingular Z, Definition 1.2
yields the square roots
−1 0 1 0
Z Z −1 , Z Z −1 , (1.22)
0 1 0 −1
and these formulae provide an infinity of square roots, because only for diagonal Z
are the matrices in (1.22) independent of Z. Indeed, one infinite family of square
roots of A comprises the Householder reflections
cos θ sin θ
H(θ) = , θ ∈ [0, 2π].
sin θ − cos θ
Definitions 1.2, 1.4, and 1.11 yield primary matrix functions. In most applications
it is primary matrix functions that are of interest, and virtually all the existing the-
ory and available methods are for such functions. Nonprimary matrix functions are
obtained from Definition 1.2 when two equal eigenvalues in different Jordan blocks
are mapped to different values of f ; in other words, different branches of f are taken
for different Jordan blocks with the same eigenvalue. The function obtained thereby
depends on the matrix Z in (1.3). This possibility arises precisely when the function is
multivalued and the matrix is derogatory, that is, the matrix has multiple eigenvalues
and an eigenvalue appears in more than one Jordan block.
Unlike primary matrix functions, nonprimary ones are not expressible as polyno-
mials in the matrix. However, a nonprimary function obtained from Definition 1.2,
using the prescription in the previous paragraph, nevertheless commutes with the ma-
trix. Such a function has the form X = Z diag(fk (Jk ))Z −1 , where A = Z diag(Jk )Z −1
is a Jordan canonical form and where the notation fk denotes that the branch of f
taken depends on k. Then XA = AX, because fk (Jk ) is a primary matrix function
and so commutes with Jk .
But note that not all nonprimary matrix functions are obtainable
from the Jordan
canonical form prescription above. For example, A = 00 00 has the square root
X = 00 10 , and X is a Jordan block larger than the 1 × 1 Jordan blocks of A. This
example also illustrates that a nonprimary function can have the same spectrum as a
primary function, and so in general a nonprimary function cannot be identified from
its spectrum alone.
Nonprimary functions can be needed when, for a matrix A depending on a param-
eter t, a smooth curve of functions f (A(t)) needs to be computed and eigenvalues of
A(t) coalesce. Suppose we wish to compute square roots of
cos θ sin θ
G(θ) =
− sin θ cos θ
16 Theory of Matrix Functions
The only primary square roots of G(π) are ±iI, which are nonreal. While it is
nonprimary, G(π/2) is the square root we need in order to produce a smooth curve
of square roots.
An example of an application where nonprimary logarithms arise is the embed-
dability problems for Markov chains (see Section 2.3).
A primary matrix function with a nonprimary flavour is the matrix sign function
(see Chapter 5), which for a matrix A ∈ Cn×n is a (generally) nonprimary square
root of I that depends on A.
Unless otherwise stated, f (A) denotes a primary matrix function throughout this
book.
Theorem 1.22 (existence of matrix square root). A ∈ Cn×n has a square root if and
only if in the “ascent sequence” of integers d1 , d2 , . . . defined by
di = dim(null(Ai )) − dim(null(Ai−1 ))
Proof. See Cross and Lancaster [122, ] or Horn and Johnson [296, , Cor.
6.4.13].
To illustrate, consider a Jordan block J ∈ Cm×m with eigenvalue zero. We have
dim(null(J 0 )) = 0, dim(null(J)) = 1, dim(null(J 2 )) = 2, . . . , dim(null(J m )) = m,
and so the ascent sequence comprises m 1s. Hence Jk does not have a square root
unless m = 1. However, the matrix
0 1 0
0 0 0 (1.23)
0 0 0
has ascent sequence 2, 1, 0, . . . and so does have a square root—for example, the matrix
0 0 1
0 0 0 (1.24)
0 1 0
(which is the 3 × 3 Jordan block with eigenvalue 0 with rows and columns 2 and 3
interchanged).
1.6 Classification of Matrix Square Roots and Logarithms 17
Another important existence question is “If A is real does there exist a real f (A),
either primary or nonprimary?” For most common functions the answer is clearly yes,
by considering a power series representation. For the square root and logarithm the
answer is not obvious; the next result completes the partial answer to this question
given in Remark 1.9 and Theorem 1.18.
Proof. For the last part consider the real Schur decomposition, QT AQ = R (see
Section B.5), where Q ∈ Rn×n is orthogonal and R ∈ Rn×n is upper quasi-triangular.
Clearly, f (A) is real if and only if QT f (A)Q = f (R) is real, and a primary matrix
function f (R) is block upper triangular with diagonal blocks f (Rii ). If A has a
negative real eigenvalue then some Rii is 1 × 1 and negative, making f (Rii ) nonreal
for f the square root and logarithm.
The result of (b) is due to Culver [126, ], and the proof for (a) is similar; see
also Horn and Johnson [296, , Thms. 6.4.14, 6.4.15] and Nunemacher [451, ].
Theorem 1.23 implies that −In has a real, nonprimary square root and logarithm
for every even n. For some insight into part (a), note that if A has two Jordan blocks
J of the same size then its Jordan matrix has a principal submatrix of the form
J 0 0 I 2
0J = J 0 .
where f (Jk (λk )) is given in (1.4) and where jk = 1 or 2 denotes the branch of f ; thus
(1) (2)
Lk = −Lk . Our first result characterizes all square roots.
18 Theory of Matrix Functions
Theorem 1.24 (Gantmacher). Let A ∈ Cn×n be nonsingular with the Jordan canon-
ical form (1.2). Then all solutions to X 2 = A are given by
(j ) (j )
X = ZU diag(L1 1 , L2 2 , . . . , L(jp)
p )U
−1 −1
Z , (1.25)
where µ2k = λk , k = 1: p.
Now consider the matrix
(j ) (j )
L = diag(L1 1 , L2 2 , . . . , L(jp)
p ), (1.27)
(j )
where we choose the jk so that Lk k has eigenvalue µk for each k. Since L is a
square root of J, by the same argument as above L must have the Jordan canonical
form JX . Hence X = W LW −1 for some nonsingular W . From X 2 = A we have
W JW −1 = W L2 W −1 = ZJZ −1 , which can be rewritten as (Z −1 W )J = J(Z −1 W ).
Hence U = Z −1 W is an arbitrary matrix that commutes with J, which completes the
proof.
The structure of the matrix U in Theorem 1.24 is described in the next result.
Theorem 1.25 (commuting matrices). Let A ∈ Cn×n have the Jordan canonical
form (1.2). All solutions of AX = XA are given by X = ZW Z −1 , where W = (Wij )
with Wij ∈ Cmi ×mj (partitioned conformably with J in (1.2)) satisfies
0, λi 6= λj ,
Wij =
Tij , λi = λj ,
where Tij is an arbitrary upper trapezoidal Toeplitz matrix which, for mi < mj , has
the form Tij = [0, Uij ], where Uij is square.
Theorem 1.26 (classification of square roots). Let the nonsingular matrix A ∈ Cn×n
have the Jordan canonical form (1.2) with p Jordan blocks, and let s ≤ p be the number
of distinct eigenvalues of A. Then A has precisely 2s square roots that are primary
functions of A, given by
(j ) (j )
Xj = Z diag(L1 1 , L2 2 , . . . , L(jp)
p )Z
−1
, j = 1: 2s ,
Proof. The proof consists of showing that for the square roots (1.25) for which
ji = jk whenever λi = λk ,
(j ) (j ) (j ) (j )
U diag(L1 1 , L2 2 , . . . , L(jp)
p )U
−1
= diag(L1 1 , L2 2 , . . . , L(jp)
p ),
that is, U commutes with the block diagonal matrix in the middle. This commutativ-
ity follows from the explicit form for U provided by Theorem 1.25 and the fact that
upper triangular Toeplitz matrices commute.
Theorem 1.26 shows that the square roots of a nonsingular matrix fall into two
classes. The first class comprises finitely many primary square roots, which are “iso-
lated”, being characterized by the fact that the sum of any two of their eigenvalues
is nonzero. The second class, which may be empty, comprises a finite number of pa-
rametrized families of matrices, each family containing infinitely many square roots
sharing the same spectrum.
Theorem 1.26 has two specific implications of note. First, if λk 6= 0 then the two
upper triangular square roots of Jk (λk ) given by (1.4) with f the square root function
are the only square roots of Jk (λk ). Second, if A is nonsingular and nonderogatory,
that is, none of the s distinct eigenvalues appears in more than one Jordan block,
then A has precisely 2s square roots, each of which is a primary function of A.
There is no analogue of Theorems 1.24 and 1.26 for singular A. Indeed the Jordan
block structure of a square root (when one exists) can be very different from that
of A. The search for square roots X of a singular matrix is aided by Theorem 1.36
below, which helps identify the possible Jordan forms of X; see Problem 1.29.
Analogous results, with analogous proofs, hold for the matrix logarithm.
Theorem 1.27 (Gantmacher). Let A ∈ Cn×n be nonsingular with the Jordan canon-
ical form (1.2). Then all solutions to eX = A are given by
(j ) (j )
X = ZU diag(L1 1 , L2 2 , . . . , L(jp)
p )U
−1 −1
Z ,
where
(j )
Lk k = log(Jk (λk )) + 2jk πiImk ; (1.28)
log(Jk (λk )) denotes (1.4) with the f the principal branch of the logarithm, defined by
Im(log(z)) ∈ (−π, π]; jk is an arbitrary integer; and U is an arbitrary nonsingular
matrix that commutes with J.
(j )
where L1 1 is defined in (1.28), corresponding to all possible choices of the integers
j1 , . . . , jp , subject to the constraint that ji = jk whenever λi = λk .
If s < p then eX = A has nonprimary solutions. They form parametrized families
(j ) (j )
Xj (U ) = ZU diag(L1 1 , L2 2 , . . . , L(jp)
p )U
−1 −1
Z ,
Proof. Note first that a nonprimary square root of A, if one exists, must have
eigenvalues µi and µj with µi = −µj , and hence the eigenvalues cannot all lie in the
open right half-plane. Therefore only a primary square root can have spectrum in the
open right half-plane. Since A has no eigenvalues on R− , it is clear from Theorem 1.26
that there is precisely one primary square root of A whose eigenvalues all lie in the
open right half-plane. Hence the existence and uniqueness of A1/2 is established.
That A1/2 is real when A is real follows from Theorem 1.18 or Remark 1.9.
See Problem 1.27 for an extension of Theorem 1.29 that allows A to be singular.
Corollary 1.30. A Hermitian positive definite matrix A ∈ Cn×n has a unique Her-
mitian positive definite square root.
Proof. By Theorem 1.29 the only possible Hermitian positive definite square
root is A1/2 . That A1/2 is Hermitian positive definite follows from the expression
A1/2 = QD1/2 Q∗ , where A = QDQ∗ is a spectral decomposition (Q unitary, D
diagonal), with D having positive diagonal entries.
For a proof of the corollary from first principles see Problem 1.41.
This equality is trivial for monomials and follows immediately for general polynomials.
First we recap a result connecting the Jordan structures of AB and BA. We
denote by zi (X) the nonincreasing sequence of the sizes z1 , z2 , . . . , of the Jordan
blocks corresponding to the zero eigenvalues of the square matrix X.
Theorem 1.32 (Flanders). Let A ∈ Cm×n and B ∈ Cn×m . The nonzero eigenvalues
of AB are the same as those of BA and have the same Jordan structure. For the zero
eigenvalues (if any), |zi (AB) − zi (BA)| ≤ 1 for all i, where the shorter sequence
is appended with zeros as necessary, and any such set of inequalities is attained for
some A and B. If m 6= n then the larger (in dimension) of AB and BA has a zero
eigenvalue of geometric multiplicity at least |m − n|.
Theorem 1.33. Let A ∈ Cn×n and B ∈ Cm×m and let f be defined on the spectrum
of both A and B. Then there is a single polynomial p such that f (A) = p(A) and
f (B) = p(B).
Corollary 1.34. Let A ∈ Cm×n and B ∈ Cn×m and let f be defined on the spectra
of both AB and BA. Then
Af (BA) = f (AB)A. (1.30)
Proof. By Theorem 1.33 there is a single polynomial p such that f (AB) = p(AB)
and f (BA) = p(BA). Hence, using (1.29),
When A and B are square and A, say, is nonsingular, another proof of Corol-
lary 1.34 is as follows: AB = A(BA)A−1 so f (AB) = Af (BA)A−1 , or f (AB)A =
Af (BA).
As a special case of the corollary, when AB (and hence also BA) has no eigenvalues
on R− (which implies that A and B are square, in view of Theorem 1.32),
A(BA)1/2 = (AB)1/2 A.
22 Theory of Matrix Functions
In fact, this equality holds also when AB has a semisimple zero eigenvalue and the
definition of A1/2 is extended as in Problem 1.27.
Corollary 1.34 is useful for converting f (AB) into f (BA) within an expression, and
vice versa; see, for example, (2.26), the proof of Theorem 6.11, and (8.5). However,
when m > n, (1.30) cannot be directly solved to give an expression for f (AB) in
terms of f (BA), because (1.30) is an underdetermined system for f (AB). The next
result gives such an expression, and in more generality.
Theorem 1.35. Let A ∈ Cm×n and B ∈ Cn×m , with m ≥ n, and assume that BA
is nonsingular. Let f be defined on the spectrum of αIm + AB, and if m = n let f be
defined at α. Then
f (αIm + AB) = f (α)Im + A(BA)−1 f (αIn + BA) − f (α)In B. (1.31)
Proof. Note first that by Theorem 1.32, the given assumption on f implies that
f is defined on the spectrum of αIn + BA and at α.
Let g(t) = f [α + t, α] = t−1 (f (α + t) − f (α)), so that f (α + t) = f (α) + tg(t).
Then, using Corollary 1.34,
f (αIm + AB) = f (α)Im + ABg(AB)
= f (α)Im + Ag(BA)B
= f (α)Im + A(BA)−1 f (αIn + BA) − f (α)In B,
as required.
This result is of particular interest when m > n, for it converts the f (αIm + AB)
problem—a function evaluation of an m × m matrix—into the problem of evaluating
f and the inverse on n × n matrices. Some special cases of the result are as follows.
(a) With n = 1, we recover (1.16) (albeit with the restriction v ∗ u 6= 0).
(b) With f the inverse function and α = 1, (1.31) yields, after a little manipulation,
the formula (I + AB)−1 = I − A(I + BA)−1 B, which is often found in textbook
exercises. This formula in turn yields the Sherman–Morrison–Woodbury formula
(B.12) on writing A + U V ∗ = A(I + A−1 U · V ∗ ). Conversely, when f is analytic
we can obtain (1.31) by applying the Sherman–Morrison–Woodbury formula to the
Cauchy integral formula (1.12). However, Theorem 1.35 does not require analyticity.
As an application of Theorem 1.35, we now derive a formula for f (αIn +uv ∗ +xy ∗ ),
where u, v, x, y ∈ Cn , thereby extending (1.16) to the rank-2 case. Write
∗
v
uv ∗ + xy ∗ = [ u x ] ∗ ≡ AB.
y
Then
v∗ u v∗ x
C := BA = ∈ C2×2 .
y∗ u y∗ x
Hence
∗ ∗ −1
v∗
f (αIn + uv + xy ) = f (α)In + [ u x]C f (αI2 + C) − f (α)I2 . (1.32)
y∗
The evaluation of both C −1 and f (αI2 + C) can be done explicitly (see Problem 1.9
for the latter), so (1.32) gives a computable formula that can, for example, be used
for testing algorithms for the computation of matrix functions.
1.9 Miscellany 23
1.9. Miscellany
In this section we give a selection of miscellaneous results that either are needed
elsewhere in the book or are of independent interest.
The first result gives a complete description of the Jordan canonical form of f (A)
in terms of that of A. In particular, it shows that under the action of f a Jordan
block J(λ) splits into at least two smaller Jordan blocks if f ′ (λ) = 0.
Theorem 1.36 (Jordan structure of f (A)). Let A ∈ Cn×n with eigenvalues λk , and
let f be defined on the spectrum of A.
(a) If f ′ (λk ) 6= 0 then for every Jordan block J(λk ) in A there is a Jordan block
of the same size in f (A) associated with f (λk ).
(b) Let f ′ (λk ) = f ′′ (λk ) = · · · = f (ℓ−1) (λk ) = 0 but f (ℓ) (λk ) 6= 0, where ℓ ≥ 2,
and consider a Jordan block J(λk ) of size r in A.
(i ) If ℓ ≥ r, J(λk ) splits into r 1 × 1 Jordan blocks associated with f (λk ) in
f (A).
(ii ) If ℓ ≤ r − 1, J(λk ) splits into the following Jordan blocks associated with
f (λk ) in f (A):
• ℓ − q Jordan blocks of size p,
• q Jordan blocks of size p + 1,
where r = ℓp + q with 0 ≤ q ≤ ℓ − 1, p > 0.
Proof. We prove just the first part. From Definition 1.2 it is clear that f either
preserves the size of a Jordan block Jk (λk ) ∈ Cmk ×mk of A—that is, f (Jk (λk )) has
Jordan form Jk (f (λk )) ∈ Cmk ×mk —or splits Jk (λk ) into two or more smaller blocks,
each with eigenvalue f (λk ). When f ′ (λk ) 6= 0, (1.4) shows that f (Jk (λk ))−f (λk )I has
rank mk − 1, which implies that f does not split the block Jk (λk ). When f ′ (λk ) = 0,
it is clear from (1.4) that f (Jk (λk )) − f (λk )I has rank at most mk − 2, which implies
that f (Jk (λk )) has at least two Jordan blocks. For proofs of the precise splitting
details, see Horn and Johnson [296, , Thm. 6.2.25] or Lancaster and Tismenetsky
[371, , Thm. 9.4.7].
To illustrate the result, consider the matrix
0 1 0 0
0 0 1 0
A= ,
0 0 0 1
0 0 0 0
Clearly f (A) has Jordan form comprising two 1 × 1 blocks and one 2 × 2 block. We
have f ′ (0) = f ′′ (0) = 0 and f ′′′ (0) 6= 0. Applying Theorem 1.36 (b) with ℓ = 3,
r = 4, p = 1, q = 1, the theorem correctly predicts ℓ − q = 2 Jordan blocks of size
24 Theory of Matrix Functions
1 and q = 1 Jordan block of size 2. For an example of a Jordan block splitting with
f (X) = X 2 , see the matrices (1.23) and (1.24).
Theorem 1.36 is useful when trying to solve nonlinear matrix equations, because
once the Jordan form of f (A) is known it narrows down the possible Jordan forms of
A; see, e.g., Problems 1.30 and 1.51.
We noted in Section 1.4 that a nonprimary function of a derogatory A may com-
mute with A but is not a polynomial in A. The next result shows that all matrices
that commute with A are polynomials in A precisely when A is nonderogatory—that
is, when no eigenvalue appears in more than one Jordan block in the Jordan canonical
form of A.
Theorem 1.38. B ∈ Cn×n commutes with every matrix that commutes with A ∈
Cn×n if and only if B is a polynomial in A.
Proof. See Horn and Johnson [296, , Thm. 4.4.19].
The following result is useful for finding solutions of a nonlinear matrix equation
of the form f (X) = A.
Corollary 1.41. Suppose A, B ∈ Cn×n and AB = BA. Then for some ordering of
the eigenvalues of A, B, and AB we have λi (AopB) = λi (A)opλi (B), where op = +,
−, or ∗.
Proof. By Theorem 1.40 there exists a unitary U such that U ∗ AU = TA and
U BU = TB are both upper triangular. Thus U ∗ (A op B)U = TA op TB is upper
∗
Theorem 1.42 (McCoy). For A, B ∈ Cn×n the following conditions are equiva-
lent.
(a) There is an ordering of the eigenvalues such that λi (p(A, B)) = p(λi (A), λi (B))
for all polynomials of two variables p(x, y).
(b) There exists a unitary U ∈ Cn×n such that U ∗ AU and U ∗ BU are upper
triangular.
(c) p(A, B)(AB − BA) is nilpotent for all polynomials p(x, y) of two variables.
Theorem 1.43. A ∈ Cn×n is unitary if and only if A = eiH for some Hermitian H.
In this representation H can be taken to be Hermitian positive definite.
Proof. The Schur decomposition of A has the form A = QDQ∗ with Q unitary
and D = diag(exp(iθj )) = exp(iΘ), where Θ = diag(θj ) ∈ Rn×n . Hence A =
Q exp(iΘ)Q∗ = exp(iQΘQ∗ ) = exp(iH), where H = H ∗ . Without loss of generality
we can take θj > 0, whence H is positive definite.
Theorem 1.44. A ∈ Cn×n has the form A = eS with S real and skew-symmetric if
and only if A is real orthogonal with det(A) = 1.
Proof. “⇒”: IfP S is real and skew-symmetric then A is real, ATA = e−S eS = I,
and det(eS ) = exp( λi (S)) = exp(0) = 1, since the eigenvalues of S are either zero
or occur in pure imaginary complex conjugate pairs.
“⇐”: If A is real orthogonal then it has the real Schur decomposition A =QDQT
aj bj
with Q orthogonal and D = diag(Dii ), where each Dii is 1, −1, or of the form −b j aj
with a2j + b2j = 1. Since det(A) = 1, there is an even number of −1s, and so we can
aj bj
include the −1 blocks among the −b j aj
blocks. It is easy to show that
aj bj cos θj sin θj 0 θj
≡ = exp =: exp(Θj ). (1.33)
−bj aj − sin θj cos θj −θj 0
Note that another way of expressing Theorem 1.45 is that for any logarithm of a
nonsingular X, det(X) = exp(trace(log(X))).
Buchheim gave a derivation of the formula [84, ] and then generalized it to mul-
tiple eigenvalues using Hermite interpolation [85, ].
Weyr [614, ] defined f (A) using a power series for f and showed that the
series converges if the eigenvalues of A lie within the radius of convergence of the
series. Hensel [258, ] obtained necessary and sufficient conditions for convergence
when one or more eigenvalues lies on the circle of convergence (see Theorem 4.7).
Metzler [424, ] defined the transcendental functions eA , log(A), sin(A), and
arcsin(A), all via power series.
The Cauchy integral representation was anticipated by Frobenius [195, ], who
states that if f is analytic then f (A) is the sum of the residues of (zI − A)−1 f (z) at
the eigenvalues of A. Poincaré [473, ] uses the Cauchy integral representation,
and this way of defining f (A) was proposed in a letter from Cartan to Giorgi, circa
1928 [216, ].
The Jordan canonical form definition is due to Giorgi [216, ]; Cipolla [109,
] extended it to produce nonprimary matrix functions.
Probably the first book (actually a booklet) to be written on matrix functions is
that of Schwerdtfeger [513, ]. With the same notation as in Definitions 1.2 and
1.4 he defines
Xs nXi −1
f (j) (λi )
f (A) = Ai (A − λi I)j ,
i=1 j=0
j!
where the Ai are the Frobenius covariants: Ai = Z diag(gi (Jk ))Z −1 , where gi (Jk ) = I
if λi is an eigenvalue of Jk and gi (Jk ) = 0 otherwise, where A = Z diag(Jk )Z −1 is the
1.11 Notes and References 27
Jordan canonical form. This is just a rewritten form of the expression for f (A) given
by Definition 1.2 or by the Lagrange–Hermite formula (1.8). It can be restated as
s nX
X i −1
where the Zij depend on A but not on f . For more details on these formulae see Horn
and Johnson [296, , pp. 401–404, 438] and Lancaster and Tismenetsky [371, ,
Sec. 9.5].
The equivalence of all the above definitions of f (A) (modulo their different levels
of generality) was first shown by Rinehart [493, ] (see the quote at the end of the
chapter).
One of the earliest uses of matrices in practical applications was by Frazer, Duncan,
and Collar of the Aerodynamics Department of the National Physical Laboratory
(NPL), England, who were developing matrix methods for analyzing flutter (unwanted
vibrations) in aircraft. Their book Elementary Matrices and Some Applications to
Dynamics and Differential Equations [193, ] emphasizes the important role of
the matrix exponential in solving differential equations and was “the first to employ
matrices as an engineering tool” [71, ], and indeed “the first book to treat matrices
as a branch of applied mathematics” [112, ].
Early books with substantial material on matrix functions are Turnbull and Aitken
[579, , Sec. 6.6–6.8]; Wedderburn [611, , Chap. 8], which has a useful bibliog-
raphy arranged by year, covering 1853–1933; MacDuffee [399, , Chap. IX], which
gives a concise summary of early work with meticulous attribution of results; Ferrar
[184, , Chap. 5]; and Hamburger and Grimshaw [245, ]. Papers with useful
historical summaries include Afriat [5, ] and Heuvers and Moak [259, ].
Interest in computing matrix functions grew rather slowly following the advent of
the digital computer. As the histogram on page 379 indicates, the literature expanded
rapidly starting in the 1970s, and interest in the theory and computation of matrix
functions shows no signs of abating, spurred by the growing number of applications.
A landmark paper is Moler and Van Loan’s “Nineteen Dubious Ways to Compute
the Exponential of a Matrix” [437, ], [438, ], which masterfully organizes
and assesses the many different ways of approaching the eA problem. In particular,
it explains why many of the methods that have been (and continue to be) published
are unsuitable for finite precision computation.
The “problem solving environments” MATLAB, Maple, and Mathematica have
been invaluable for practitioners using matrix functions and numerical analysts de-
veloping algorithms for computing them. The original 1978 version of MATLAB
included the capability to evaluate the exponential, the logarithm, and several other
matrix functions. The availability of matrix functions in MATLAB and it competitors
has undoubtedly encouraged the use of succinct, matrix function-based solutions to
problems in science and engineering.
is given by Lancaster and Tismenetsky [371, , Chap. 9]. A classic reference is
Gantmacher [203, , Chap. 5]. Golub and Van Loan [224, , Chap. 11] briefly
treat the theory before turning to computational matters. Linear algebra and matrix
analysis textbooks with a significant component on f (A) include Cullen [125, 1972],
Pullman [481, ], and Meyer [426, ].
For more details on the Jordan canonical form see Horn and Johnson [295, ,
Chap. 3] and Lancaster and Tismenetsky [371, , Chap. 6].
Almost every textbook on numerical analysis contains a treatment of polynomial
interpolation for distinct nodes, including the Lagrange form (1.9) and the Newton
divided difference form (1.10). Textbook treatments of Hermite interpolation are
usually restricted to once-repeated nodes; for the general case see, for example, Horn
and Johnson [296, , Sec. 6.1.14] and Stoer and Bulirsch [542, , Sec. 2.1.5].
For the theory of functions of operators (sometimes called the holomorphic func-
tional calculus), see Davies [133, ], Dunford and Schwartz [172, ], [171, ],
and Kato [337, ].
Functions of the DFT matrix, and in particular fractional powers, are considered
by Dickinson and Steiglitz [151, ], who obtain a formula equivalent to (1.18).
Much has been written about fractional transforms, mainly in the engineering litera-
ture; for the fractional discrete cosine transform, for example, see Cariolaro, Erseghe,
and Kraniauskas [96, ].
Theorems 1.15–1.17 can be found in Lancaster and Tismenetsky [371, , Sec. 9.7].
Theorem 1.18 is from Higham, Mackey, Mackey, and Tisseur [283, , Thm. 3.2].
The sufficient condition of Remark 1.9 and the equivalence (c) ≡ (d) in Theorem 1.18
can be found in Richter [491, ].
Different characterizations of the reality of f (A) for real A can be found in Evard
and Uhlig [179, , Sec. 4] and Horn and Piepmeyer [298, ].
The terminology “primary matrix function” has been popularized through its use
by Horn and Johnson [296, , Chap. 6], but the term was used much earlier by
Rinehart [495, ] and Cullen [125, 1972].
A number of early papers investigate square roots and pth roots of (singular)
matrices, including Taber [561, ], Metzler [424, ], Frobenius [195, ],
Kreis [363, ], Baker [40, ], and Richter [491, ], and Wedderburn’s book
also treats the topic [611, , Secs. 8.04–8.06].
Theorem 1.24 is a special case of a result of Gantmacher for pth roots [203, ,
Sec. 8.6]. Theorem 1.26 is from Higham [268, ]. Theorem 1.27 is from [203, ,
Sec. 8.8].
Theorem 1.32 is proved by Flanders [188, ]. Alternative proofs are given by
Thompson [566, ] and Horn and Merino [297, , Sec. 6]; see also Johnson and
Schreiner [321, ].
We derived Theorem 1.35 as a generalization of (1.16) while writing this book;
our original proof is given in Problem 1.45. Harris [249, , Lem. 2] gives the result
for α = 0 and f a holomorphic function, with the same method of proof that we have
given. The special case of Theorem 1.35 with f the exponential and α = 0 is given
by Celledoni and Iserles [102, ].
Formulae for a rational function of a general matrix plus a rank-1 perturbation,
r(C + uv ∗ ), are derived by Bernstein and Van Loan [61, ]. These are more
complicated and less explicit than (1.31), though not directly comparable with it since
C need not be a multiple of the identity. The formulae involve the coefficients of r and
so cannot be conveniently applied to an arbitrary function f by using “f (A) = p(A)
Problems 29
Problems
The only way to learn mathematics is to do mathematics.
— PAUL R. HALMOS, A Hilbert Space Problem Book (1982)
1.1. Show that the value of f (A) given by Definition 1.2 is independent of the par-
ticular Jordan canonical form that is used.
1.2. Let Jk be the Jordan block (1.2b). Show that
f (mk −1) (−λk )
f (−λk ) −f ′ (−λk ) ... (−1)mk −1
(mk − 1)!
.. ..
f (−Jk ) = f (−λk ) . . . (1.34)
..
. ′
−f (−λk )
f (−λk )
1.3. (Cullen [125, , Thm. 8.9]) Define f (A) by the Jordan canonical form defi-
nition. Use Theorem 1.38 and the property f (XAX −1 ) = Xf (A)X −1 to show that
f (A) is a polynomial in A.
1.4. (a) Let A ∈ Cn×n have an eigenvalue λ and corresponding eigenvector x. Show
that (f (λ), x) is a corresponding eigenpair for f (A).
(b) Suppose A has constant row sums α, that is, Ae = αe, where e = [1, 1, . . . , 1]T .
Show that f (A) has row sums f (α). Deduce the corresponding result for column sums.
1.5. Show that the minimal polynomial ψ of A ∈ Cn×n exists, is unique, and has
degree at most n.
1.6. (Turnbull and Aitken [579, , p. 75]) Show that if A ∈ Cn×n has minimal
polynomial ψ(A) = A2 − A − I then (I − 31 A)−1 = 35 (A + 2I).
1.7. (Pullman [481, , p. 56]) The matrix
−2 2 −2 4
−1 2 −1 1
A=
0 0 1 0
−2 1 −1 4
Ich hatte jetzt das 27. Jahr meiner »Verbannung« auf meiner Insel
angetreten. Nie unterließ ich es, den Jahrestag meines Schiffbruchs
und meiner Ankunft auf der Insel in inbrünstigen Gebeten zu Gott zu
begehen. Seine Güte hatte mich bisher so wunderbar behütet, und
nun erfüllte mich die beglückende Hoffnung, wieder in die
Gesellschaft der Menschen zurückzukehren. Auch während der
letzten Zeit setzte ich meine Tagesarbeiten fort. Ich grub, pflanzte,
ergänzte meine Einzäunungen, sammelte Korn, Reis, Baumfrüchte
und Trauben ein; ich besorgte meine Ziegenherden, buk Brot und
Kuchen, verfertigte Kleider, Körbe und Töpfe. – Unterdessen war die
Regenzeit herangenaht, und ich mußte Bedacht darauf nehmen,
unser Boot sicher unterzubringen. Ich schaffte es daher so weit auf
den Strand, als die steigende Flut es erlaubte, und gebot Freitag,
daneben ein Becken zu graben, tief genug, um das Boot beständig
flott zu erhalten. Als die Flut dann zurückwich, führten wir einen
starken Damm auf, der das Becken verschloß und dem Eindringen
des Meeres vorbeugte. Um aber unser Fahrzeug gegen den Regen
zu schützen, bedeckten wir es mit einem Dach und erwarteten so
den Monat November oder Dezember, um die ersehnte Fahrt
anzutreten.
Mit Beginn der schönen Jahreszeit beeilten wir uns, die nötigen
Zurüstungen zur Reise zu treffen. Denn ich gedachte, vielleicht
schon in acht bis zwölf Tagen das Wasserbecken zu öffnen und das
Boot auslaufen zu lassen. Eines Morgens hatte ich Freitag nach dem
Meere hinabgeschickt, um eine Schildkröte zu fangen, weil wir
sowohl das Fleisch als auch die Eier dieses Tieres sehr wohl zu
schätzen wußten. Aber schon nach wenigen Minuten kam er eiligst
wieder zurück und übersprang den ersten Festungszaun.
»O Herr, Herr, o Jammer!«
»Was gibt's denn, was hast du?«
»Dort unten, dort unten! Eins, zwei, drei Kähne!« Freitag war so
erschrocken, daß er am ganzen Körper zitterte; er hatte sich
eingebildet, daß die Wilden nichts Geringeres beabsichtigten, als ihn
einzufangen, in Stücke zu zerhauen und aufzuessen. Ich suchte ihn
zu beruhigen, so gut ich konnte, und ihm begreiflich zu machen, daß
ich ja ganz in der nämlichen Gefahr schwebe wie er.
»Freitag«, sagte ich, »wir müssen mit ihnen um unser Leben
kämpfen; bist du bereit dazu?«
– »Jawohl, ich schieße auf sie; aber ihre Zahl ist groß.«
»Was thut das, Freitag? Unsre Gewehre werden einen Teil von
ihnen niederstrecken, und das Feuer und der Knall wird die andern
in die Flucht schlagen. Wenn ich dich aber mit meinem Leben
verteidige, willst du mir auch treulich zur Seite stehen und alles
thun, was ich dir sage?«
»Ja, Herr, ich will sterben, wenn du mir zu sterben befiehlst.«
Hierauf holte ich eine Flasche Rum, um Freitag in seiner mutigen
Stimmung zu erhalten; dann gebot ich ihm, die beiden gewöhnlichen
Jagdgewehre herbeizubringen, und ich selbst lud sie mit tüchtigen
Posten.
Hiernach stieg ich mit meinem Fernrohr auf die Warte, um zu
sehen, was an der Küste vorging. Da entdeckte ich nun, daß 21
Wilde in drei Kanoes gelandet waren, und zwar an der Südostküste,
was mich um so mehr wunder nahm, als ich noch nie an dieser
Stelle das geringste Anzeichen einer Landung der Kannibalen
bemerkt hatte. Der Ort, wo sie ausgestiegen waren, schien sehr
flach, der Strand niedrig; etwa 100 Schritte davon begann der Saum
eines dichten Gebüsches, welches sich ziemlich weit bis in die
Felsengruppen der inneren Insel hineinzog. Es deuchte mich, als ob
sie drei Gefangene bei sich hätten und auch diesmal aus keinem
andern Grunde an meine Insel gekommen wären, als wieder eines
ihrer Siegesfestmahle abzuhalten.
Zunächst lud ich nun vier Musketen mit sieben Kugeln, sowie
meine beiden Pistolen mit zwei Kugeln. Den Degen steckte ich in den
Gürtel und befahl Freitag, sein Beil, ein Pistol, zwei Musketen und
eine Flinte nebst Vorrat von Pulver und Blei zu ergreifen; ich selbst
aber nahm das andre Pistol und die übrigen Schießgewehre.
Außerdem steckten wir einige Brotkuchen und getrocknete Rosinen
zu uns, sowie ein Fläschchen Rum zur Stärkung unsrer
Lebensgeister. So gerüstet rückten wir aus. Auf einem Umweg von
ungefähr einer Viertelmeile bogen wir nach dem Rande des Gehölzes
ein, um hier, ungesehen von den Wilden, bis an die Bucht zu
gelangen und sie in Schußlinie vor uns zu haben.
Unter Beobachtung größter Vorsicht gelangten wir an das Ende
des Gehölzes und somit in die Nähe der Feinde, von denen mich nur
noch eine einzige Baumgruppe trennte. Ich befahl Freitag, auf einen
Baum zu steigen, um zu sehen, was die Wilden vornähmen. Er
kletterte sehr bald wieder herab und berichtete, er habe die Feinde
ganz deutlich gesehen; sie säßen rings um ein Feuer und verzehrten
das Fleisch eines ihrer Gefangenen; ein andrer liege dicht daneben
an Händen und Füßen gebunden und werde wahrscheinlich
demnächst an die Reihe des Verspeisens kommen. »Aber«, fügte
Freitag bedeutungsvoll hinzu, »es ist keiner von unserm Stamme,
sondern einer von den weißen bärtigen Männern, die sich in unserm
Vaterlande angesiedelt haben.« Dieser Bericht versetzte mich in Zorn
und Wut. Ich stieg nun mit meinem Fernglas ebenfalls auf einen
Baum und erkannte deutlich an Gesicht und Bekleidung in dem
gebundenen Manne einen Europäer.
Ein kleines Gebüsch zog sich von der Waldspitze noch ungefähr
100 Schritte nach links gegen den Strand hin, und ich konnte, durch
dasselbe gedeckt, den Wilden mich noch mehr nähern. Am Ende des
Buschwerks gelangte ich auf einen kleinen Sandhügel oder eine
Düne, von wo aus ich die jetzt nur noch in einer Entfernung von 80
Schritt lagernden Wilden aufs genaueste beobachten konnte. Es war
kein Augenblick mehr zu verlieren, denn eben bemerkte ich, wie sich
zwei der Kannibalen anschickten, des Europäers Hände und Füße
von den Fesseln zu befreien, um ihn dann am Feuer zu schlachten.
Ich sah mich nach Freitag um.
»Jetzt«, sagte ich zu ihm, »thue, wie ich dir sagen werde.«
»Ja, Herr! Befiehl!«
»So ahme genau das nach, was du mich thun siehst, und fehle
nicht!«
Mit diesen Worten legte ich eines der Jagdgewehre und eine der
Musketen auf den Boden. Freitag that dasselbe. Dann zielte ich auf
die beiden mit ihrem Schlachtopfer beschäftigten Wilden und gebot
Freitag, unter den übrigen Haufen zu feuern.
»Bist du fertig, Freitag?« – Freitag nickte zustimmend.
»Nun – dann Feuer!«
Zwei donnerähnliche Schüsse hallten hinaus auf Land und Meer. –
ebookgate.com