THE PEANO-BAKER SERIES

arXiv:1011.1775v1 [math.CA] 8 Nov 2010

MICHAEL BAAKE AND ULRIKE SCHLAGEL

Abstract. This note reviews the Peano-Baker series and its use to solve the general linear
system of ODEs. The account is elementary and self-contained, and is meant as a pedagogic
introduction to this approach, which is well known but usually treated as a folklore result
or as a purely formal tool. Here, a simple convergence result is given, and two examples
illustrate that the series can be used explicitly as well.

1. Introduction
Consider the inhomogeneous linear initial value (or Cauchy) problem
(1)

x = Ax + b,

with x(t0 ) = x0 ,

on some interval I R that contains t0 , with time-dependent quantities x, A and b. Here,

A(t) denotes a family of matrices and b(t) a vector-valued function, both of dimension d over
R (or over C). For simplicity, we assume A and b to be continuous on I. When [A(t), A(s)] = 0
for all t, s I (where we use the notation [A, B] = AB BA), the unique solution of (1) is

Z t


Z t
 Z
A( ) d x0 +
(2)
x(t) = exp
exp A() d b( ) d ,
t0

t0

Rt

t0

x0 for the homogeneous case; see [14, Cor. 2.41].

Rt
Also, when A(t) = A is constant, one has exp t A( ) d = e (tt0 )A as usual; we refer to [2,
0
Secs. 11 and 12] for a general exposition.
The solution of (1) is still unique when [A(t), A(s)] 6= 0 for t 6= s, but the solution is then
given by the Peano-Baker series (PBS); see [12, Sec. 16.5], [7, Sec. 7.5] or [4, Sec. 1.3] for
background. This approach goes back to Peano (1888; see [15]), and was further developed
by Baker (1905; see [3]). Bakers paper also contains a more detailed account of the history
of this approach. The PBS is well-known in principle, but not widely known, and appears
mainly in engineering textbooks; compare [4, 13, 8] for examples. Quite frequently, it is used
in a purely formal manner (without convergence considerations), and it is also sometimes
claimed [17] that it is of little practical use for the solution. However, as one can learn from
the biographical sketches of Delone in [6], one should neither discard analytical tools nor
believe repeated claims without evidence. In fact, when new problems surface, many (old)
tools turn out to be more useful than expected.
It is thus the purpose of this short note to summarise the elementary properties of the
Peano-Baker series and to give a simple and self-contained account. In particular, we do
not restrict ourselves to the abstract fundamental system (or matrix solution), but discuss
its explicit series expansion with convergence. The need for it came up in the context of a

which simplifies to x(t) = exp

t0 A( ) d


MICHAEL BAAKE AND ULRIKE SCHLAGEL

problem in mathematical population genetics; see [18] for details. We add two little examples
to demonstrate that it can be used explicitly as well.
2. Homogeneous case: Volterra equation and convergence
Let A (t; t0 ) denote the fundamental system of the homogeneous part of (1), which is also
the flow of the linear system. It satisfies the Volterra integral equation
Z t
(3)
A (t; t0 ) = 1 + A( ) A ( ; t0 ) d , with A (t0 ; t0 ) = 1 .
t0

By means of a formal Picard iteration, this leads to

Z
Z t
Z t
1
(4)
A (t; t0 ) = 1 + A( ) d + A(1 )
A(2 ) d2 d1 + . . . =
where
In (t) :=

t0

t0

t0

t0

A(1 )

1
t0

A(2 )

1+

In (t),

n=1

n1

t0

A(n ) dn d2 d1 .

For convenience, we set I0 (t) = 1. By construction, we then have the recursion

Z t
(5)
In+1 (t) = A( ) In ( ) d
t0

for n N0 . Clearly, In (t0 ) = n,0 1, in line with A (t0 ; t0 ) = 1. Eq. (4) is known as the
Peano-Baker series (PBS) in control theory [8, p. 598], see also [9, Sec. IV.5], or as the
series expansion in the context of the time-ordered exponential in physics [11]. The PBS was
recently also extended to the setting of time scales [5].
Let us consider the individual terms of the PBS more closely.
Lemma 1. When A is continuous on I, the matrix functions In are differentiable and satisfy
I n (t) = A(t) In1 (t), for all t I and n N.
Proof. Fix n N, let h be a small number, and consider the difference
Z t+h
A( ) In1 ( ) d .
In (t + h) In (t) =
t

The matrix-valued integrand M := A In1 is (component-wise) continuous by assumption.

Consequently, we may employ the mean value theorem for integrals, independently for each


(h)

(h)
component, which results in h1 In (t + h) In (t) = Mi,j (t + i,j ) 1i,jd , where all i,j
[0, h], though they need not be equal. Still, this implies
In (t + h) In (t)
= A(t) In1 (t),
h0
h
lim

and our claim follows.

Let us show next (in modern terminology) that the PBS is nicely convergent in our finitedimensional setting; compare [12, Sec. 16.5] or [7, Sec. 2.11] for a slighlty different account,
[16] for further results in this direction, and [10, Ch. 4.3] for background on matrix Taylor

PEANO-BAKER SERIES

series. We begin by establishing a link to the usual exponential series. In the one-dimensional
case, this is [4, Cor. 1.3.1].
Lemma 2. If A is continuous on the interval I, with [A(t), A(s)] = 0 for all t, s I, one has
Z
n
1 t
A( ) d
In (t) =
n! t0
for all t I and n N0 . In particular, this applies when A is one-dimensional.
Proof. The claim is obviously true for n = 0 and n = 1. We can now employ induction via
the recursion (5), assuming the validity for some n N. This gives
Z t
Z
Z t
n
1  1
A(1 )
A(1 ) In (1 ) d1 =
A( ) d d1 ,
In+1 (t) =
n! t0
t0
t0
where the integrand on the right hand side can be rewritten as
Z
(n+1)
 Z 1
n
d  1
1
1
A(1 )
A( ) d
A( ) d
.
=
n!
(n + 1)! d1 t0
t0
This step employs the general chain rule, where the assumed commutativity is used. Inserting
this expression into the previous formula completes the induction step by an application of
the fundamental theorem of calculus.

Remark 1. When [A(t), A(s)] = 0 on I, the PBS (4) reduces to the well-known formula
Z


n
Z t
X
1  t
A( ) d
A( ) d
= exp
A (t; t0 ) =
n!
t0
t0
n=0

as a consequence of Lemma 2; see [14, Sec. 2.3] for a detailed exposition of this case. This
reduction is also mentioned in most of the sources cited so far; compare [4, Cor. 1.3.2].
Remark 2. A closer look at the proof of Lemma 2 shows that the condition [A(t), A(s)] = 0
may be replaced by the slightly weaker assumption that, for all t I, the matrix A(t)
Rt
commutes with the integral t A( ) d .
0
Also, it should be mentioned that Lemma 2 can alternatively be proved by direct calculations based on permutations of the integration variables followed by a suitable rearrangement
to cover the integration region [t0 , t]n . When combined with an induction argument, it suffices
to consider the permutations (12 . . . n), (1)(23 . . . n), . . . , (1)(2) . . . (n1, n). This approach is
slightly more general. As it is also less transparent, we skip further details.
Let now k.k denote any norm on R d (or on C d ), and define the compatible matrix norm
as usual by kAk := supkxk=1 kAxk. This implies the relations kAxk kAk kxk and, more
importantly, kABk kAk kBk.
Theorem 1. If kA(t)k is locally integrable on the interval I, the series representation (4) of
A (t; t0 ) is compactly convergent on I in the chosen matrix norm.


MICHAEL BAAKE AND ULRIKE SCHLAGEL

Proof. Let J I be compact, with t0 J. We show that the sequence of partial sums is
Cauchy on J. So, let m, n N with n > m and consider
n
m
n
n
X

X

X
X





Ik
Ik (t)
Ik (t) =
Ik (t)

k=0

k=0

n
X

Z tZ

k=m+1 t0

k=m+1

t0

k=m+1

k1

t0

A(1 ) A( ) d d2 d1 .
k
k

Observing A(1 ) A(k ) A(1 ) A(k ) , where all A(i ) are non-negative real
numbers, Lemma 2 implies that the last sum is majorised as
Z
n
X
k
1  t
A( ) d ,

k!
t0
k=m+1

which is the corresponding Cauchy

estimate
for the Taylor series of the ordinary exponential

Rt


function on R, evaluated at t A( ) d , which exists for all t J by assumption. Since this
0
series converges compactly, our claim follows.

3. Solution of the inhomogeneous problem

It is now obvious that A (t; t0 ) solves the homogeneous initial value problem (3). This
follows from a term-wise application of Lemma 1 to the PBS, which is fully justified by
Theorem 1. The determinant of A , which is a Wronskian and thus satisfies Liouvilles
theorem, reads
Z t
Z t








tr A( ) d ,
tr A( ) d = exp
(6)
det A (t; t0 ) = det A (t0 ; t0 ) exp
t0

t0

which never vanishes; see [2, Prop. 11.4] or [14, Thm. 2.23] for details. This means that A
has full rank and thus indeed constitutes a fundamental system of the homogeneous linear
system.
Since A (t; t0 ) is the unique solution of (3), the flow property implies the relation

(7)

A (t; s) A (s; t0 ) = A (t; t0 ),

1
which, due to A (t; t) = 1, also implies A (t; s)
= A (s; t). With the usual variation
of constants calculation, compare [2, Thm. 11.13], one can now easily derive the following
result.
Theorem 2. Let I be an interval, with t0 I, A a continuous matrix function on I, and b
also continuous on I. Then, the Cauchy problem (1) has the unique solution
Z t


A (t0 ; ) b( ) d
x(t) = A (t; t0 ) x0 +
t0

with A given by the PBS (4). When [A(t), A(s)] = 0 for all t, s I, the formula simplifies
to the explicit expression (2) with ordinary exponentials.


PEANO-BAKER SERIES

4. Examples
Let us demonstrate the explicit applicability of the PBS with two simple examples. Both
can also be solved by other means, but are perhaps still instructive.
Consider the matrix family A(t) = ( 10 at ) with t I, t0 = 0, and fixed parameter a. The
matrices commute for a = 1, but not otherwise. One finds
!
tn+1
tn
n
X

n
n!
(n+1)!
a1 ,
with

=
In (t) =
n
(at)n
0
=1
n!
so that the PBS gives
A (t; 0) =

!
et f (t)
0 eat

with

f (t) =

et eat (1 a)teat
.
(1 a)2

Note that f (t) simplifies to 21 t2 et for a = 1, in line with the then simpler ODE system. For
general t0 , one finds A (t; t0 ) = A (tt0 ; 0). The PBS differs both from the matrix exponential
and from the Mathias formula for upper triangular matrices; compare [10, Thm. 3.6].
As a second example, consider the ODE system
 

 
x
0 t
x
=
,
y
a 0
y
which leads to the Airy function via the 2nd order ODE y = at y; see [1, Ch. 10.4] for details.
With := a1/3 , the PBS leads to the formula


1
g(t)

2 f (t)

A (t; 0) =
2 g(t) f (t)
with
f (z) =

X
k=0

z 3k
3k ( 13 )k
(3k)!

and g(z) =

X
k=0

3k ( 23 )k

z 3k+1
(3k + 1)!

from [1, Eq. 10.4.3]. Here, ( 13 )k = 31 ( 13 + 1) ( 13 + k 1) as usual; see [1, Eq. 6.1.22]. Note
that
!
2
1 t2
a0
,
A (t; 0)
0 1
in line with the trivially solvable system for a = 0.

Acknowledgements
It is a pleasure to thank Peter Jarvis for discussions, and Wolf-J
urgen Beyn and Nikolai
Dolbilin for useful suggestions.

MICHAEL BAAKE AND ULRIKE SCHLAGEL

References
[1] M. Abramowitz und I.A. Stegun, Handbook of Mathematical Functions, Dover, New York (1965).
[2] H. Amann, Ordinary Differential Equations, de Gryuter, Berlin (1990).
[3] H.F. Baker, Note on the integration of linear differential equations, Proc. London Math. Soc. (2)
2 (1905) 293296.
[4] R.W. Brockett, Finite Dimensional Linear Systems, Wiley, New York (1970).
[5] J.J. Dacunha, Transition matrix and generalized matrix exponential via the Peano-Baker series,
J. Diff. Eq. Appl. 11 (2005) 12451264.
[6] B.N. Delone, The St. Petersburg School of Number Theory, AMS, Providence, RI (2005).
[7] R.A. Frazer, W.J. Duncan and A.R. Collar, Elementary Matrices and some Applications to Dynamics and Differential Equations, Cambridge University Press, Cambridge (1938).
[8] T.E. Fortmann and K.L. Hitz, An Introduction to Linear Control Theory, Dekker, New York
(1977).
[9] F.R. Gantmacher, Applications of the Theory of Matrices, Wiley, New York (1959).
[10] N.J. Higham, Functions of Matrices Theory and Computation, SIAM, Philadelphia (2008).
[11] J. Hoek, On Campbell-Baker-Hausdorff formulae and time-ordered exponentials, Nederl. Akad.
Wetensch. Proc. Ser. B 84 (1981) 4756.
[12] E.L. Ince, Ordinary Differential Equations, reprint of the 1926 edition, Dover, New York (1956).
[13] T. Kailath, Linear Systems, Prentice Hall, Englewood Cliffs (1980).
[14] W.G. Kelley and A.C. Peterson, The Theory of Differential Equations, 2nd ed., Springer, New
York (2010).
[15] G. Peano, Integration par series des equations differentielles lineaires, Math. Ann. 32 (1888)
450456.
[16] V. Prepelit
a, M. Doroftei and T. Vasilache, Peano-Baker series convergence for matrix valued
functions of bounded variation, Balkan J. Geom. Appl. 3 (1998) 111118.
[17] A. Rindos, S. Woolet, I. Viniotis and K. Trivedi, Exact methods for the transient analysis of nonhomogeneous continuous time Markov chains, in: 2nd International Workshop on the Numerical
Solution of Markov Chains, eds. W.J. Stewart, Kluwer, Dordrecht (1995), pp. 121133.
[18] U. Schlagel, Deterministische Rekombinations- und Selektionsdynamik, Diplomarbeit, Univ. Bielefeld (2008).
t fu
r Mathematik, Universita
t Bielefeld,
Fakulta
Postfach 100131, 33501 Bielefeld, Germany
Department of Mathematical and Statistical Sciences, University of Alberta,
Edmonton, Alberta T6G 2G1, Canada