Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Analytic Equations: 4.1 Power Series

Download as pdf or txt
Download as pdf or txt
You are on page 1of 21

Chapter 4

Analytic equations

In the preceding chapter we allowed the dependent variable (y or u) to be


complex, but required the independent variable (x or t) to be real. We
found that the solutions could be constructed in terms of elementary func-
tions. In the present chapter, not only the dependent variable but also the
independent variable is allowed to be complex. The derivatives are there-
fore complex derivatives, so the functions differentiated must be analytic.
We shall again be able to construct the solutions, although now they will
take the form of power-series, and may represent new, unfamiliar functions.
We provide in §4.1 statements, without proof, of the few facts regarding
analytic functions that we require in the present chapter. However, in this
and the next chapter, some familiarity with methods of complex analysis,
as described in books on complex-function theory, is desirable.

4.1 Power Series


Consider the power series

X
ak z k (4.1)
k=0

where the coefficients {ak } are prescribed complex numbers and z is a com-
plex variable. Any such power series has a radius of convergence R. R may
be zero, in which case the series fails to converge except in the uninteresting
case when z = 0. If R is not zero, then the series converges for each z with
|z| < R and diverges for any z with |z| > R. It is also possible to have
R = ∞, meaning that the series converges for all z in the complex plane.
The reasoning leading to these conclusions is the following.

93
94

Suppose that thePseries (4.1) converges for some value of z, say z = z0 .


The convergence of k k
ak z0 implies that ak z0 → 0 as k → ∞ and therefore

−k
that |a
k| < M |z0 | k for some sufficiently large M . Therefore if |z| < |z0 |
k k
P
then ak z ≤ M ρ with ρ = |z/z0 | < 1. The series
ak z is therefore
absolutely convergent, by comparison with the convergent series MP ρk .
P
The radius of convergence R is now the largest number r such that ak z k
converges if |z| < r. The series (4.1) therefore represents a complex function
within the disk of radius R, say

X
a (z) = ak z k .
k=0

What is more, the function a is analytic for |z| < R, and its derivative is
obtained by term-by-term differentiation of the series:

X
a0 (z) = kak z k−1 ,
k=0

the radius of convergence of the series on the right-hand side of preceding


equation being the same, R, as that of the series for a.
There is a sometimes convenient formula for the radius of convergence
of the series (4.1):
R−1 = lim sup |an |1/n . (4.2)

The feature of analytic functions that is used in the present chapter is


that any analytic function can be represented by a power-series expansion.
More precisely, suppose a is an analytic function of the complex variable
z = x + iy in a domain D of the complex plane, and let z1 be any point of
D. Then

X
a (z) = ak (z − z1 )k , (4.3)
k=0
where the series has a nonzero radius of convergence R, i.e., converges for
|z − z1 | < R. The radius of convergence R is the distance from z1 to the
boundary of D.

4.2 Linear, analytic equations


A linear, analytic equation is one for which the coefficient functions are an-
alytic and therefore possess convergent power-series expansions, as in equa-
tion (4.3) above. The simple conclusion will be that the solutions also pos-
sess convergent power-series expansions. This provides a powerful method
95

for constructing solutions. Even if the coefficients are real and we are ulti-
mately interested in real solutions of these equations, it is useful and illumi-
nating to allow for solutions in the form of complex functions of a complex
independent variable. To emphasize this we employ standard, complex no-
tation z for the independent variable and w for the dependent variable, and
consider the differential equation
Lw ≡ w(n) + a1 (z) w(n−1) + · · · + an−1 w0 + an w = 0. (4.4)
We have written the homogeneous equation but, as usual, we shall also be
interested in solutions of the inhomogeneous equation.
The definition of linear dependence and independence is the same as
it was for real equations, with only the obvious changes: we need to al-
low complex coefficients, and we need to consider the linear dependence or
independence, not on an interval, but on a domain in the complex plane:

Definition 4.2.1 The functions w1 , w2 , . . . , wm are linearly dependent in


the domain D of the complex plane if there exist complex constants c1 , c2 , . . . , cm ,
not all zero, such that
c1 w1 (z) + c2 w2 (z) + · · · + cm wm (z) = 0
at each point z of D.

As in the case of real equations, we shall find that any solution of equa-
tion (4.4) is a linear combination of a basis of n linearly independent solu-
tions.
By a shift of the independent variable (z 0 = z − z1 ) we may assume
without loss of generality that the expansions of the coefficients may be
written

X
aj (z) = ajk z k , j = 1, 2, . . . , n (4.5)
k=0
where each of the n series converges if |z| < r. Normally we take for r the
smallest of the radii of convergence of the coefficient functions.
The important features of the theory are already in evidence if n = 2,
so we now turn to this.

4.3 Power-series expansions for n = 2


The basic equation (4.4) can be written
w00 + p (z) w0 + q (z) w = 0 (4.6)
96

where p and q possess power-series expansions like that of equation (4.5),


with coefficients {pk } and {qk } respectively. If there is any solution w which
has a derivative in a neighborhood of the origin, i.e., is analytic there, it too
has a power-series expansion

X
w (z) = wk z k . (4.7)
k=0

Example 4.3.1 Consider equation w00 − w = 0 and further impose the


initial conditions w (0) = 1, w0 (0) = 1. In the expansion (4.7) we must
then have w0 = 1, w1 = 1. Substitution of that expansion in the differential
equation then gives the recursion formula wk+2 = wk / (k + 1) (k + 2). It is
easy to check that wk = 1/k!: it’s true for k = 0, 1 and follows, by
Pinduction,
from the recursion formula for the general case. Thus w (z) = ∞ k
k=0 z /k!,
which we recognize as exp z. 
In the preceding example, the coefficients were p = 0 and q = 1; these
are analytic in the entire plane, and we find that the solutions are like-
wise analytic in the entire plane. In the general case when the coefficients
have power-series expansions convergent with some radius of convergence
R (which could be infinite, as in this example), the procedure is to write a
power-series expansion for w with coefficients to be determined, and substi-
tute this series into equation (4.4). This requires the multiplication of two
power series to obtain the product power series. The formula for such a
product is

X ∞
X ∞
X n
X
an z n bn z n = cn z n where cn = ak bn−k . (4.8)
n=0 n=0 n=0 k=0

Suppose now we assume Pprovisionally that the equation (4.4) possesses a


k
power-series solution w = wk z . Putting off the question whether we can
expect it to converge, we proceed formally and determine the coefficients
{wk }, as follows. Substitution of the power-series expression into equation
(4.4) results in the formula

X ∞
X ∞
X ∞
X ∞
X
k−2 k−1 j k
k (k − 1) wk z + kwk z pj z + wk z qj z j = 0.
k=0 k=0 j=0 k=0 j=0

By collecting coefficients of like powers of z we obtain, on the left-hand side


of the preceding equation, the following power series:
∞ k k
( )
X X X
(k + 1) (k + 2) wk+2 + pk−l (l + 1) wl+1 + qk−l wl z k = 0.
k=0 l=0 l=0
97

If this is to hold in any neighborhood of the origin, it must be that each


coefficient vanishes separately. This gives the recursion formula
k
X k
X
(k + 1) (k + 2) wk+2 + pk−l (l + 1) wl+1 + qk−l wl = 0, (4.9)
l=0 l=0

for k = 0, 1, . . .. If w0 and w1 are given, this determines all subsequent


coefficients. In particular, if w0 = w1 = 0, then all subsequent cofficients
also vanish. Therefore if the solution with vanishing initial data is analytic
in a domain containing the origin, it vanishes identically there, by a standard
theorem in complex-function theory. From this we can infer, as in Chapter 2,
that if w1 and w2 are linearly independent solutions in D, then any solution
is a linear combination of them. More generally, since w0 and w1 represent
the values of w and its first derivative at z = 0, this recursion formula
shows that the solution is uniquely determined by its initial values, at least
formally.
The Wronskian plays a similar role for solutions of the homogeneous
equations in the present, complex case as it did in the previous, real
case. However, for more general functions, not necessarily satisfying
such an equation, it plays a somewhat different role. For example,
Theorem 4.3.1 If u and v are analytic functions in a domain D of
the complex plane, and if their Wronskian vanishes identically on D,
then they are linearly dependent on D.
Proof: Assume that v does not vanish identically on D or the conclu-
sion is trivially true. Let v(z0 ) 6= 0 for some z0 ∈ D; then v does not
vanish on a neighborhood d ⊂ D. We may therefore write the equation
expressing the vanishing of their Wronskian as
v0
u0 − u = 0, z ∈ d.
v
Since u(z) = v(z) is a solution, u(z) = cv(z) is the unique solution of
this equation on d for some choice of c. Thus the function u(z) − cv(z)
is analytic on D and vanishes on an open subset of D, and must, by a
standard theorem in complex-function theory, therefore vanish on all
of D. 
This contrasts with the real case. For example, the functions
 3
3 x if x > 0
u(x) = x and v(x) =
−x3 if x ≤ 0

are linearly independent on the real line, but their Wronskian vanishes
identically there.
98

To turn the formal solution into a genuine solution, we need to know


that the series converge.

Example 4.3.2 Consider the equation


4 2
w00 − w0 + w = 0.
1−z (1 − z)2
The coefficients p and q both have power-series expansions with radius of
convergence R = 1. The equation has a pair of linearly independent solutions
w1 = (1 − z)−1 and w2 = (1 − z)−2 . These also have power-series expansions
with the same radius of convergence, R = 1. 

Example 4.3.3 Consider the equation


2
w00 − w = 0.
(1 + z)2
Here p = 0 and q has a power series with radius of convergence R = 1. This
equation has the linearly independent pair of solutions w1 = (1 + z)2 and
w2 = (1 + z)−1 . The latter has a power series expansion with R = 1 but the
former has a power series with infinite radius of convergence. 

Example 4.3.4 Now consider the equation


1
w00 + w0 = 0.
1−z
It has solutions w1 = 1 and w2 = 1 − 2z + z 2 . Here the coefficient q has
a power series with radius of convergence R = 1, and both solutions have
power series with infinite radius of convergence. 

In the preceding examples the coefficients have power series convergent


for |z| < 1. It is possible, as we see from these examples, that this gives the
maximal disk of convergence for any power-series solution, or that there may
be one or even two independent solutions that have power series convergent
beyond this disk. We see in the next section that any power-series solution
converges with at least the radius of convergence of coefficient functions.

4.4 The method of majorants


The method for establishing the convergence of the power series with coeffi-
cients given by the recursion formula (4.9) is remarkably simple. It is based
on the idea of a majorizing series:
99

P∞ k
P∞ k
Definition 4.4.1 The series k=o Ak z majorizes the series k=0 ak z if
Ak ≥ |ak |, k = 0, 1, . . ..
It is seen from the definition that the coefficients of the majorizing series
are real and non-negative. The basic fact about two power series related
in this way is that if the majorizing series converges for |z| < r, so also
does the other series. The proof of this remark may be basedP on the Cauchy
convergence criterion: the infinite series of complex numbers cn converges
if and only if, given any positive number , there is an integer N such that

Xn
ck <  provided n > m ≥ N.



k=m

Suppose the majorizing series has radius of convergence R and let ρ be any
real number in the interval (0, R). The majorizing series then converges for
any z such that |z| ≤ ρ. For such z we have
m
X m
X m
X
k k
| ak z | ≤ |ak ||z| ≤ Ak ρk .
n n n

Given  > 0 we can choose N so that, for values of n and m greater than N
(and n > m), the last sum Pon the right is less than . Hence the radius of
convergence of the series ak z k cannot be less than ρ and, since ρ is any
positive number less than R, the radius of convergence cannot be less than
R.
Consider now the differential equation (4.6) and suppose its coefficients
p and q have power-series expansions

X ∞
X
q (z) = qk z k and p (z) = pk z k
k=0 k=0

which are majorized, respectively, by the power series



X ∞
X
Q (z) = Qk z k and P (z) = Pk z k .
k=0 k=0

Denote by w0 and w1 the initial data for the solution w of equation (4.6).
Lemma 4.4.1 Let W0 ≥ |w0 | and W1 ≥ |w1 |. Then the solution W (z) of
the initial-value problem
W 00 = P (z) W 0 + Q (z) W, W (0) = W0 , W 0 (0) = W1
majorizes the solution w (z) of the initial-value problem consisting of equa-
tion (4.6) together with the initial data w0 , w1 .
100

Proof: The recursion relation for the solution of the majorant initial-value
problem is the same, except for a change of sign, as that given by equation
(4.9) above. Comparison of the successive coefficients in the two cases, and
a simple induction, provides the conclusion. 
Suppose now that the lesser of the radii of convergence of the power
series for p and q is R, and let r be any positive number with r < R. Then
choose
P∞ ρ so that r < ρ < R. There is some M > 0 such that the series
−k
M k=0 ρ z k , convergent for |z| < ρ, majorizes each of the series for p and
q. If the series with coefficients M ρ−k majorizes that for q, so also does the
series with coefficients M 2 (k + 1) ρ−k if M ≥ 1; we are free to assume this.
Then the majorizing functions are
∞ ∞
X zk M X zk M2
P (z) = M = , Q (z) = M 2 (k + 1) k = .
0
ρk 1 − z/ρ
0
ρ (1 − z/ρ)2

The equation
M M2
W 00 = W0 + W
1 − z/ρ (1 − z/ρ)2
has the solution W = A (1 − z/ρ)α where
 q 
α = 1 − ρM − (1 − ρM )2 + 4ρ2 M 2 < 0.

For this solution W (0) = A and W 0 (0) = −αA/ρ. If A is chosen large


enough these are positive numbers dominating the initial data w0 and w1 .
The solution of the majorizing problem now is seen to have radius of conver-
gence ρ. It follows that the power-series solution of the initial-value problem
converges for |z| = r. But r was any number less than R, the least radius of
convergence of the coefficient series for p and q. Therefore the power-series
solution for w must also converge with radius of convergence at least equal
to R. We summarize this.

Theorem 4.4.1 Any power-series solution of equation (4.6) has radius of


convergence at least equal to R, where R is the lesser of the radii of conver-
gence of the power series for the two coefficient functions p and q.

PROBLEM SET 4.4.1

1. Let w0 + p (z) w = 0 be a first-order equation with p (z) = ∞ k


P
k=0 pk z
for |z| < R. Obtain the formal power-series solution of the equation.
101

2. Refer to Example 4.3.1 and impose instead the initial data w (0) =
1, w0 (0) = −1. Identify the solution with a known, elementary func-
tion. The same with the initial data w (0) = 1, w0 (0) = 0 and
w (0) = 0, w0 (0) = 1.
3. Prove the correctness of the relation (4.8).
4. Prove P
the statement in the text that the radius of convergence P of the
series ak z k is at least as great as that of a majorizing series Ak z k .
5. In Example 4.3.3 find the coefficients {qk } and write the recursion
formula (4.9) explicitly for this example. Carry
P outk the calculation of
the coefficients {ak } of the expansion w = ak z via this formula,
0
for the initial data w(0) = a0 = 1, w (0) = a1 = 2, and obtain the
function w1 (z) of Example 4.3.3 in this way.
6. Rewrite the equation of Example 4.3.3 as (1 + z)2 w00 −2w = 0. Obtain
a recursion formula by substituting the series for w in this equation
instead. Solve the initial-value problem for the initial data given in
the preceding problem using the recursion formula that you obtain this
way.
7. Find the recursion formula for the coefficients of the power-series so-
lution of the equation
1
w00 + w = 0.
1 + z2
What is the radius of convergence for general initial conditions?
8. Find a lower bound for radius of convergence for power-series solutions
to the equation
1 + Az + Bz 2 w00 + (C + Dz) w0 + Ew = 0.


Here A, . . . , E are real constants. Do not work out the recursion for-
mula. Do rely on Theorem 4.4.1.
9. Put A = C = 0 in the preceding problem. Find a condition on B, D, E
for this equation to possess a poynomial solution (a non-trivial condi-
tion, please: do not set w equal to zero).

The next three problems refer to Legendre’s equation:


 
d 2 dw

1−z + λw = 0, (4.10)
dz dz
102

where λ is a constant.

ak z k about
P
10. Find the recursion formula for power-series solutions
the origin in the form

k(k + 1) − λ
ak+2 = ak (4.11)
(k + 1)(k + 2)

and give conditions on a0 , a1 (i) ensuring that the solution is an even


function of z and (ii) that the solution is an odd function of z.

11. Find a condition on λ for Legendre’s equation to have a polynomial


solution.

12. Suppose, as in Problem 11, that one solution u(z) of Legendre’s equa-
tion is bounded at z = 1. Assuming u(1) 6= 0, show that the second
solution v(z) becomes infinite like ln(1 − z) there.

13. Airy’s equation is


w00 + zw = 0.

Obtain the power-series expansion under the initial data w(0) = 1, w0 (0) =
0. What is its radius of convergence?

14. Consider the equation v 00 + µ − z 2 v = 0, where µ is a constant. Find



2
the equation satisfied by u if v = e−z /2 u; this is called Hermite’s equa-
tion. For what values of µ does Hermite’s equation have polynomial
solutions?

15. Chebyshev’s equation is

1 − z 2 u00 − zu0 + λu = 0.


The Chebyshev polynomials are solutions of this equation for certain


values of λ. What are they?

16. In the equation of Problem 15 make the substiution z = cos θ so that


u (z) = v (θ) and find the equation satisfied by v. What are its solu-
tions when λ is such that a Chebyshev polynomial exists?
103

4.5 Higher-order equations and systems


The nth-order equation (4.4) can be treated in exactly the same fashion as
the second-order equation. The basic conclusion is that if R denotes the
least of the radii of convergence of the coefficient functions a1 , . . . , an , then
any power-series solution of the equation has a radius of convergence not
less than R. The reasoning in the following section establishes this.

4.5.1 The system of first-order equations


The equation (4.4) can be reformulated as a system of n first-order equations,
exactly as in §2.3. The result is a system that may be written in vector-
matrix form as
W 0 = A (z) W (4.12)

where now W is a function of the complex variable z and has values in


complex n-dimensional space, and A (z) is an n × n matrix with coefficients
that are analytic functions of z. This matrix has the special form of a
companion matrix as in equation (2.36), but this special form will play no
role in the considerations below; we may as well – and shall – assume we are
given the system with no knowledge of the matrix A except that its entries
are analytic functions of z.
P A power-series solution of equation (4.12) may be sought in the form
Wk z k , where now Wk is an n-component constant vector.PSince the co-
efficient matrix A also has a power-series expansion A (z) = Ak z k , where
Ak is a complex n × n matrix, the recursion relation for the coefficients Wk
is found to be
Xk
(k + 1) Wk+1 = Aj Wk−j . (4.13)
j=0

This determines each Wk+1 provided all its predecessors W0 , . . . , Wk are


known. In particular, if W0 is specified, then all subsequent coefficients are
determined by this relation.
The convergence of this expansion can again be verified via Cauchy’s
method of majorants. We outline P the steps without filling in all the details.
Suppose each entry a (z) = (a ) z l of the matrix A is majorized by a
jk l jk l
series l (bjk )l z l ; form the matrix B having these entries. Then consider
P
the equation
V 0 = BV, (4.14)
104

and the corresponding recursion formula


k
X
(k + 1) Vk+1 = Bj Vk−j . (4.15)
j=0

Suppose the initial vector V0 satisfies the condition V0i ≥ |W0i |, i = 1, . . . , n,


i.e., the components of V0 are at least equal in magnitude to those of W0 ,
term by term. It then follows by induction that P fori each k, k = 0, 1, . . . ,
V
Pki ≥ |W ki | , i = 1, . . . , n. Thus the series Vki z majorizes the series
Wki z i . For this reason the system (4.14) is said to majorize the system
(4.12).
Wk z k is
P
The strategy for establishing the convergence of the series
now, just as in the preceding section, to find a majorizing system and a
solution of it whose initial data likewise dominate those prescribed for the
system (4.12). If R is the least of the radii of convergence of the power-
series for the n2 functions aij (z) of the matrix A, then for any number ρ
with 0 < ρ < R,P each of those is majorized on a disk of radius r < ρ by the
power series M (z/ρ)K = M (1 − z/ρ)−1 for some sufficiently large M .
We may therefore take for B the matrix
 
1 1 ... 1
z −1 
 1 1 ... 1 
  
M 1−  .. .. . . ..  .
ρ  . . . . 
1 1 ... 1

We seek a solution of the majorizing system (4.14) in the form of a vector


V all of whose components are the same, i.e.,
 
1
 1 
V (z) = v (z)  .  ,
 
 .. 
1

where v is complex-valued function of z, i.e., a scalar. For initial data we


may choose
n
X
v (0) = |W0i | ;
i=1

this ensures that the initial data for the majorant system appropriately
dominate those for the system (4.12). It is easily checked that with this
105

choice of initial data the majorizing system reduces to the solution of the
one-dimensional system
 −1 n
0 z X
v = Mn 1 − v, v (0) = |W0i | ,
ρ
i=1

whose solution is  n
−M nρ X
z
v (z) = 1− |W0i | .
ρ
i=1

The power-series expansion of this function has radius of convergence ρ. It


follows that the power series for W converges for |z| < r < ρ. But ρ and r
can be arbitrarily close to R, so this power series converges for |z| < R. We
have proved

Theorem 4.5.1 In equation (4.12) suppose the entries of the matrix A have
power series expansions about the origin with least radius of convergence R.
Then, for arbitrary initial data W0 , the solution to that equation exists and
has a power-series expansion with radius of convergence not less than R.

4.5.2 The fundamental matrix solution


Theorem 4.5.1 asserts the existence of a power-series solution for the vector,
or column matrix, W (z) whose initial value W0 is prescribed. The unique-
ness of this solution is clear: if W0 = 0 then Wk = 0 for all k = 0, 1, 2, . . ..
By linearity, if two solutions have the same initial data then their difference
has zero initial data and they are the same.
There are n linearly independent choices of such an initial-value vector,
(1) (2) (n)
according to the notions of linear independence in C n . Let W0 , W0 , . . . , W0
be such a choice. Then the corresponding n solutions of the differential
equation W (1) (z), W (2) (z), . . . , W (n) (z) are necessarily linearly independent
on the domain D where the solutions are defined (cf. 4.2.1). These n so-
lutions, evaluated at any point z∗ of the domain of their definition, are
linearly independent as vectors in C n . If they were not, there would be a
linear combination of them that is zero. By uniqueness of solutions, this
linear combination would remain zero in a neighborhood of z∗ and therefore
in the entire domain in which it is analytic.
Thus there is a fundamental matrix solution Φ(z) satisfying the matrix
differential equation
φ0 (z) = A(z)φ(z) (4.16)
106

with initial value Φ(0) which is a nonsingular matrix. This matrix Φ is


nonsingular throughout the domain where it remains analytic (at least a
disk of radius R). As in the real case, the matrix Ψ(z) = Φ(z)K, where
K is constant matrix, satisfies equation (4.16) if Φ does. Therefore, if K is
nonsingular, then Ψ is likewise a fundamental matrix solution, and any two
fundamental matrix solutions are related in this way.
These remarks represent a direct extension of results of the real case to
the complex case.

4.5.3 The inhomogeneous solution


The inhomogeneous equation can be treated like the homogeneous equation,
via series solution as in §4.5.1. In place of the system (4.12) we now have
the system
W 0 = A (z) W + α (z) (4.17)
where α is a given analytic function of z taking values in the complex n-
dimensional space C n ; in other words, each component of α is an analytic
function. We now denote by R the minimum of the radii of convergence
not only for the entries aij of the matrix A but also for the components αi
of the column vector α. Each component of the latter is majorized by the
power series for M (1 − z/ρ) for any ρ < R, some sufficiently large M , and
all z in the disk of radius r provided r < ρ. Denote by β the majorizing
vector whose components are all identical and equal to M (1 − z/ρ). Then
the system

V 0 = BV + β, (4.18)
with the same choice of B as in the preceding section, will lead to the first-
order equation

z −1 z −1
   
0
v = nM 1 − v+M 1− .
ρ ρ
This has the solution
 −M nρ
z 1
v =γ 1− − .
ρ n
The initial data must again dominate those for the problem (4.17). This
can be achieved by choosing the constant γ large enough. The conclusion
is the same: any power-series solution of the system (4.17) has radius of
convergence at least equal to R.
107

Alternatively, the inhomogeneous equation may be treated as we did the


real case, by the variation of parameters formula. Introducing a fundamental-
matrix solution Φ(z), we can verify that the formula
Z z
W (z) = Φ(z)C + Φ(z)Φ−1 (ζ)R(ζ) dζ, (4.19)
0

where C is an arbitrary, complex, constant vector, represents the most gen-


eral solution to the inhomogeneous problem.

4.6 The nonlinear, analytic equation


For nonlinear equations the method of majorants continues to be effective for
proving the existence of solutions. The case of a single, first-order equation
illustrates the technique.
Consider the nonlinear initial-value problem

w0 = f (z, w), w(z0 ) = w0 , (4.20)

where w is sought as the complex-valued function of the complex variable z


that is defined and satisfies the equation in a neighborhood of the point z0 ,
and takes on the value w0 when z = z0 . The equation is assumed analytic
in the sense that the function f is represented by the power series

X
f (z, w) = akl (z − z0 )k (w − w0 )l (4.21)
k,l=0

which converges at least for some values of z 6= z0 and w 6= w0 . Without loss


of generality we may simplify the problem slightly by choosing z0 = w0 = 0,
since a shift of the variable achieves this. The problem then becomes

w0 = f (z, w), w(0) = 0 (4.22)

where

X
f (z, w) = akl z k wl , (4.23)
k,l=0

and it is assumed that the indicated series converges for at least one value
(z∗ , w∗ ) where neither z∗ nor w∗ vanishes. From this latter requirement it
follows that there exist real, positive numbers ρ, σ such that the series

X
|akl |ρk σ l
k,l
108

converges, and there is a constant M such that

|akl | ≤ M ρ−k σ −l , k, l = 0, 1, 2, . . . (4.24)

The reasoning is the same as that leading to the analogous conclusion in


§4.1.
We seek a power-series solution

X
w(z) = wj z j = w1 z + w2 z 2 + w3 z 3 + . . . .
j=1

The procedure for generating the coefficients in this series is the same as
that for linear equations: substitute in equation (4.22), collect the terms
of like powers of z, and set the resulting coefficient equal to zero. There
are two important rules to verify in this procedure: the first is that it de-
termines the coefficients {w}∞ 1 recursively – i.e., wk should depend only
on w1 , w2 , . . . , wk−1 ; the second is that the expression for wk should be a
polynomial in w1 , . . . , wk−1 and the coefficients {aij }, with only positive
coefficients. Let’s examine the first few of these expressions:

w1 = a00 ,
w2 = (1/2) (a10 + a01 w1 ) ,
w3 = (1/3) a20 + a11 w1 + a02 w12 + a01 w2 ,


w4 = (1/4) a30 + a21 w1 + a11 w2 + a01 w3 + a12 w12 + 2a02 w1 w2 + a03 w13 .


These first four coefficients conform to the rules, and it is not difficult to
verify them generally1 .
Now define the function
X
F (z, W ) = M (z/ρ)k (W/σ)l ; (4.25)
k=0,l=0

the series is absolutely convergent if |z| < ρ and |w| < σ, so the function F
is defined in that domain of C 2 . Moreover, it majorizes the corresponding
series for f (z, w). The initial-value problem

W 0 = F (z, W ), W (0) = 0 (4.26)

can be solved formally in exactly the same way that the formal series for
w(z) was obtained. Because of the second of the two rules noted above, it
1
A fully detailed verification can be found in Hille, chapter 2.
109

follows that |wk | ≤ Wk for k = 1, 2, . . .. Therefore, if the series for W has


radius of convergence R, then that for w has a radius of convergence of at
least R. But the series for F is easily recognized as the product of the series
for the functions (1 − z/ρ)−1 with that of (1 − W/σ)−1 :

z −1 W −1
   
F (z, W ) = M 1− 1− .
ρ σ

With this choice of F , we can integrate equation (4.26) as follows:

W − (1/2σ)W 2 = −(M ρ) ln(1 − z/ρ),

where we have used the initial condition to evaluate the constant term. A
single-valued determination of the logarithm is obtained by introducing a
branch cut along the real axis from z = ρ to infinity. Solving this quadratic
for W gives
(1/2)
W = σ ± σ 2 + 2σρ ln(1 − z/ρ)

.
With a determination of the square-root function that makes it positive
when its argument is real and positive, we need to choose the minus sign in
order to satisfy the initial condition:
(1/2)
W = σ − σ 2 + 2M σρ ln(1 − z/ρ)

.

For |z| < ρ this function is analytic as long as the argument of the square
root does not lie on the negative, real axis. The smallest value of z for which
this fails occurs with z real, say z = r where

σ + 2M ρ ln(1 − r/ρ) = 0,

or
r = ρ (1 − exp(−σ/2M ρ)) . (4.27)
The majorizing series therefore converges with radius of convergence r. This
proves

Theorem 4.6.1 The initial-value problem defined by equations (4.22) and


(4.23) has a solution that is analytic in a disk of radius R ≥ r, where r is
given by equation (4.27) in which the constants M, ρ, σ are those of equation
(4.24).

The estimate (4.27) for the radius of convergence is typically smaller


than the actual radius of convergence of the series solution, but we can no
110

longer expect the maximal result that was found in the linear case. In fact,
the application of this estimate includes the linear case, and it is not hard
to see that it provides an unnecessarily conservative estimate in this case.
This technique can be extended to prove the existence of analytic solu-
tions to a system of n analytic equations: if
n
X
wk0 = fj (z, w1 , w2 , . . . , wn ), wi (z0 ) = ai
j=1

and the functions fj can be expressed in convergent power series in their


variables provided |z| < ρ, |wi | < σi , i = 1, 2, . . . , n. The bound M is the
maximum of bounds Mi taken on by the series
X
|ai,k1 ,k2 ,...,kn |ρj σ1k1 · · · σnkn = Mi .

We omit the details. Since the nth order equation

w(n) = g(z, w, w0 , . . . , w(n−1) )

can be reduced to an nth-order system, the analytic character of its solutions


is a corollary of the corresponding result for the system.

PROBLEM SET 4.6.1

1. Generalize Theorem 4.3.1 to n functions analytic in a domain D ⊂ C.

2. Consider the equation w0 = w2 with initial data w(z0 ) = w0 , and


transform it to the form of equation (4.22). Compare the radius of
convergence as estimated by eqution (4.27) with the exact radius of
convergence of a power-series solution.

3. For any autonomous equation, i.e., one of the form w0 = f (w), provide
an estimate for the radius of convergence of power-series solutions that
is independent of the parameter ρ.

4. Consider the linear initial-value problem w0 = a(z)w + b(z), w(0) = 0.


Suppose that the functions a and b are analytic in a disk centered at
the origin and of radius R. Compare the estimate of equation (4.27)
with that which you know from the linear theory.
314
Bibliography

[1] G. Birkhoff and G.-C. Rota. Ordinary Differential Equations. John


Wiley and Sons, Inc., New York, 1978.

[2] P.F. Byrd and M. D. Friedman. Handbook of elliptic integrals for engi-
neers and physicists. Springer, Berlin, 1954.

[3] Jack Carr. Applications of Centre Manifold Theory. Springer-Verlag,


New York, 1981.

[4] Earl A. Coddington and Norman Levinson. Theory of Ordinary Differ-


ential Equations. New York: McGraw-Hill, 1955.

[5] J. Guckenheimer and P. Holmes. Nonlinear Oscillations, Dynamical


Systems, and Bifurcation of Vector Fields. Springer-Verlag, New York,
1983.

[6] Einar Hille. Lectures on Ordinary Differential Equations. London:


Addison-Wesley Publishing Company, 1969.

[7] M. Hirsch and S. Smale. Differential Equations, Dynamical Systems,


and Linear Algebra. Academic Press, New York, 1974.

[8] E.L. Ince. Ordinary Differential Equations. Dover Publications, New


York, 1956.

[9] S. MacLane and G. Birkhoff. Algebra. New York: Macmillan, 1967.

[10] Jerrold E. Marsden, Anthony J. Tromba, and Alan Weinstein. Ba-


sic Multivariable Calculus. New York: Springer-Verlag:W.H. Freeman,
1993.

[11] I.G. Petrovski. Ordinary Differential Equations. Prentice-Hall Inc.,


Englewood Cliffs, 1966.

315
316

[12] Walter Rudin. Principles of Mathematical Analysis. McGraw-Hill, New


York, 1964.

You might also like