You are on page 1of 7

Closed Form Solutions of Linear Odes having

Elliptic Function Coefficients


Reinhold Burger

George Labahn

Mark van Hoeij

Symbolic Computation Group


School of Computer Science
University of Waterloo
Waterloo, Ontario, Canada

Symbolic Computation Group


School of Computer Science
University of Waterloo
Waterloo, Ontario, Canada

Department of Mathematics
Florida State University
Tallahassee, Florida, USA

rfburger@scg.uwaterloo.ca glabahn@uwaterloo.ca
ABSTRACT
We onsider the problem of nding losed form solutions of
linear di erential equations having oe ients whi h are ellipti fun tions. For se ond order equations we show how to
solve su h an ode in terms of doubly periodi fun tions of the
se ond kind. The method depends on two pro edures, the
rst using a se ond symmetri power of an ode along with
a de ision pro edure for determining when su h equations
have ellipti fun tion solutions while the se ond involves the
omputation of exponential solutions.

Categories and Subject Descriptors


I.1.2 [Computing Methodologies: Symboli and Algebrai Manipulation|Algorithms

General Terms
Algorithms

Keywords
Linear odes; ellipti fun tions; Kova i 's algorithm; de ision
pro edures

1. INTRODUCTION
In this paper we onsider linear di erential equations of
the form
an (x)y

(n)

(x) +  + a1 (x)y0 (x) + a0 (x)y(x) = 0

(1)

where the oe ients ai (x) are doubly periodi having the


same periods. Doubly periodi fun tions are omplex-valued
fun tions having two independent periods, that is, two independent onstants T and T 0 su h that
f (x + T )

= f (x) and f (x + T 0 ) = f (x) for all

x:

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
ISSAC04, July 47, 2004, Santander, Spain.
Copyright 2004 ACM 1-58113-827-X/04/0007 ...$5.00.

hoeij@math.fsu.edu

Classi al examples of doubly periodi fun tions in lude the


Weierstrass } and }0 fun tions, the Ja obi sn, n and dn
fun tions and ratios of Theta fun tions [1. We will also
require that our oe ients are in fa t ellipti fun tions,
that is, that they do not have any essential singularities.
We remark that the term doubly periodi is often used when
ellipti is really meant, it being assumed that the fun tions
under dis ussion are also meromorphi .
Linear di erential equations with ellipti fun tion oe ients appear histori ally in many instan es ( f. [9). For example, solving Lapla e's equation in three dimensions in ellipsoidal oordinates using the separation of variables method
gives the lassi al Lame equation
y

00 (x) (n(n + 1)}(x) + B )y(x) = 0

(2)

where n is a positive integer and B is a onstant. Additional


interesting examples an be found in many texts, see for
example Forsyth [9, Halphen [10 and Kamke [14.
There are two lassi al representations of ellipti fun tions, the Weierstrass form and the Ja obi form. In either
ase, we an
p onvert (1) to a linear ode having oe ients
from K(z; w(z )) where w(z ) is a polynomial of degree 3
or 4 and K is a eld of onstants. Using methods from [16,
17 one obtains a de ision pro edure for determining when
(1) has solutions whi h are ellipti fun tions.
Pi ard's theorem [9, 15 says that when all solutions of
(1) are uniform or path-independent then the ode has solutions that are doubly periodi of the se ond kind. These
fun tions an be expressed in terms of the Weierstrass Zeta
and Sigma fun tions or the Ja obi Zeta fun tion [1. In this
paper we show how we an nd su h solutions in the ase
of all se ond order linear odes, even when all solutions are
not ne essarily uniform. In fa t we present a omplete algorithm for solving redu ible se ond order linear odes with
ellipti fun tion oe ients. We do this by nding ellipti
fun tion solutions of the se ond symmetri power of the original ode (in se tion 3), or else nding exponential solutions
of the symmetri produ t of the equation with its onjugate (in se tion 4). The methods have been implemented [7
and are available (or will shortly be available) in the Maple
omputer algebra system. We mention that our method is
not the rst algorithm for nding su h solutions. Indeed
the pro edures in [17 nd all Liouvillian solutions of (1), in
both the redu ible or irredu ible ases. Our ontribution is
to provide a new method whi h both e ient and omplete
for the se ond order redu ible ase of (1).

The remainder of the paper is organized as follows. In the


next se tion we give ba kground information for odes of the
form (1). Se tion 3 gives a method for solving the se ond
order problem in terms of doubly-periodi fun tions of the
se ond kind when there are ellipti fun tion solutions of an
asso iated third order equation while the following se tion
produ es solutions whi h are doubly-periodi fun tions of
the se ond kind via nding exponential solutions of an asso iated ode of fourth order. The paper ends with a on lusion
along with topi s for future resear h.

2. PRELIMINARIES
In this se tion we give some well-known fa ts about ellipti fun tions, doubly periodi fun tions of the se ond kind
and linear odes having ellipti fun tions as oe ients. Additional information and details an be found in the referen es [2, 3, 9, 13.

2.1

Elliptic Function Solutions of Arbitrary


Order Equations

There are two lassi al representations of ellipti fun tions, the Weierstrass form and the Ja obi form ( .f. [3).
Let }(x) = }(x; g2 ; g3 ) denote the Weierstrass } fun tion
where g2 and g3 are onstants
p whi h are determined by the
periods. Then }0 (x) = 4}(x)3 g2 }(x) g3 and every
ellipti fun tion an be represented as a rational fun tion
of } and }0 ( f. [2). In the ase of Ja obi forms there are
many representations for ellipti fun tions. For example, let
sn(x) = sn(x; k) where k is a onstant determined by the
periods. Then
p
0
sn (x) = n(x)  dn(x) = (1 sn2 (x))  (1 k2 sn2 (x))
and every ellipti fun tion an be represented as a rational
fun tion of sn and sn0 . A similar statement holds for all the
other 11 forms of the Ja obi ellipti fun tions ( .f. [2).
Theorem 2.1 below gives the prin ipal representations for
our problems. The result follows dire tly from Singer [17
along with the isomorphism (via a hange of variable from
x to f ):
p
(3)
K(x; w(x))[ d $ K(f; f 0 )[ d
dx

via
x

7! f;

w(x)

7! f 0 ;

df

d
dx

7! dfd

1 d
= 0 :
f dx

(4)

Theorem 2.1 (Singer [17). Let L be a linear di erential operator having oe ients in K(f; f 0 ) with K a eld
of onstants and where f satis es (f 0 )2 = w(f ) for some
polynomial w(z ) 2 K[z . Then one has a de ision pro edure
for nding solutions of L(y) = 0 in K(f; f 0 ).
2
Thus suppose our oe ients are in Weierstrass form and
that we are looking for solutions in the same form. Setting
f (x) = }(x; g2 ; g3 )1 , the Weierstrass } fun tion, allows us to
make use of the above formalism using w(z ) = 4z 3 g2 z g3 .
1

As in lassi al texts, we drop the g2 and


in ases where this is obvious.

g3

arguments for

2.2

Doubly-periodic Solutions of the Second


Kind

A se ond lass of fun tions whi h are almost doubly periodi also plays an important role in the study of linear odes
of the form (1). A fun tion F (x) is said to be doubly periodi
of the se ond kind if there exist two periods T , T 0 , and two
onstants s, s0 su h that
F (x + T )

= s  F (x);

F (x + T

0 ) = s0  F (x)

for all x in the omplex plane. The onstants s and s0 are


referred to as the period multipliers of F (x). These fun tions are not, in general, truly periodi , as the value of F (x)
hanges by a onstant fa tor ea h time x hanges by the
period.
Doubly periodi fun tions of the se ond kind are important from the following lassi al observation given by Pi ard
(see In e [13) in 1879.
Theorem 2.2 (Pi ard's Theorem). If the oe ients
of a homogeneous linear di erential equation are doubly periodi fun tions of the independent variable, and all solutions
of the equation are uniform (that is, path-independent), then
the equation possesses at least one solution whi h is a doubly
periodi fun tion of the se ond kind.
2
We also note that when F (x) is doubly periodi of the
0
se ond kind then the ratio G(x) = FF ((xx)) is doubly periodi .
0
For if T and T are the periods and s, s0 the orresponding
period multipliers of F then
G(x + T )

F 0 (x + T )
F (x + T )

sF 0 (x)
sF (x)

= G(x)

and similarly G(x + T 0 ) = G(x). Thus the logarithmi


derivative of any doubly periodi fun tion of the se ond kind
is doubly-periodi .
Conversely assume that F (x) is a solution of our equation
0
su h that FF ((xx)) = G(x), a doubly periodi fun tion. Then
the logarithmi derivative of F (Fx(+x)T ) is G(x + T ) G(x),
whi h is zero by assumption. Hen e there exists a onstant
s su h that F (x + T ) = sF (x). This also o urs for the
se ond period T 0 and hen e F (x) would be doubly-periodi
of the se ond kind. Thus nding doubly-periodi solutions
of the se ond kind is equivalent to the sear h of rst order
doubly-periodi right fa tors of the asso iated linear di erential operator.
We remark that, as mentioned earlier in the introdu tion, all fun tions are also assumed to not have any essential
singularities. Hen e the logarithmi derivative mentioned
above will in fa t be an ellipti fun tion.

2.3

Fundamental Bases

To understand how doubly-periodi solutions of the se ond kind arise in the study of (1), one needs to rst look at
how fundamental bases behave at the periods. Let the set
ff1 (x); f2 (x); : : : ; fn (x)g be a basis of solutions of (1) where
all the oe ients are doubly periodi fun tions with periods T , T 0 . Sin e the oe ients of (1) are doubly periodi
we have that f1 (x + T ), f2 (x + T ), . . . , fn (x + T ), are also

solutions of (1). Hen e ea h an be expressed as a linear


ombination of the basis fun tions, that is,
fj (x + T )

= a1j f1 (x) + a2j f2 (x) +  + anj fn (x)

for some onstants


this be omes

aij ,

(5)

independent of x. In matrix form

f~(x + T )

= f~(x)  A

(6)

where f~(x) = [f1 (x);  ; fn (x) and A = [aij nn : Similarly,
we also have a onstant matrix B = [bij whi h satis es the
matrix equation
0
f~(x + T ) = f~(x)  B:
(7)
The following gives some well known fa ts about the matri es A and B asso iated to a given fundamental system.
We in lude a proof for ompleteness.
Theorem 2.3. Let A and B be onstant matri es asso iated to a given fundamental basis ff1 (x); : : : ; fn(x)g of so-

lutions of the linear ode (1). Then


(i) A and B are nonsingular

(ii) If W (f~(x)) denotes the Wronskian of the fundamental


system then
W (f~(x + T ))
W (f~(x + T 0 ))
det(A) =
and
det(
B) =
:
~
~
W (f (x))

W (f (x))

(8)

(iii) The produ ts of the eigenvalues of A and B , taking


multipli ities
 are  R x+T 0 a (u) 
 R x+T ainto (a ount,
n 1
n 1 u) du and exp
exp
an (u)
an (u) du ,
x
x
respe tively.
(iv) Suppose F (x) is a solution of (1) with

F (x) = 1 f1 (x) +

 + n  fn (x):

Then F (x) is doubly-periodi of the se ond kind with


period multipliers s and s0 if and only if s and s0 are
eigenvalues of A and B , respe tively, ea h having ~ =
[ 1; : : : ; n t as a ommon eigenve tor.
Proof: Di erentiating both sides of (6) n 1 times with
respe t to x gives n 1 additional equations of the form
0
0
(n 1)
f~ (x + T ) = f~ (x)  A, : : : , f~
(x + T ) = f~(n 1) (x)  A:
Combining these into a single matrix equation, we have
[ f~(x + T ) j  j f~(n 1) (x + T ) t
= [ f~(x) j
(9)
 j f~(n 1) (x) t  A:
Sin e we have a fundamental system the Wronskian matrix
[ f~(x) j
  j f~(n 1) (x) t is nonsingular and hen e so
is A. A similar argument shows that B is also nonsingular
hen e we have (i). Part (ii) follows by taking determinants
of both sides of (9). Part (iii) follows from (ii) along with
Abel's identity [13, p. 75.
Suppose now that F (x) is a solution of (1) with
F (x) = 1 f1 (x) +  + n fn (x) = f~(x)  ~
(10)
where ~ = [ 1 ; : : : ; n t. Then
F (x + T ) = f~(x + T )  = f~(x)  A  ~

(11)

and

0 ) = f~(x + T 0 )  = f~(x)  B  ~:
(12)
Noti e that for any onstants s and s0 equations (11) and
F (x + T

(12) imply that


and

sF (x) = F (x + T )

()

() s0 f~(x)  ~ = f~(x)  B  ~:

s F (x)

= F (x + T 0 )

 

sf~(x) ~ = f~(x) A ~

(13)
(14)

Using a similar argument as before we see that equations


(13) and (14) are equivalent to

0  ~ = B  ~
so that sF (x) = F (x + T ) and s0 F (x) = F (x + T 0 ) if and
only if s and s0 are in fa t eigenvalues of the matri es A and

s ~ = A ~ and s

B,

respe tively, having a ommon eigenve tor ~.

Remark 2.4. The onditions of Pi ard's Theorem require


that all solutions of (1) be path-independent. As su h one
an traverse a path from x to x + T and then to x + T + T 0
and expe t to end at the same pla e as if one has traversed a
path from x to x + T 0 to x + T 0 + T . Using equations (6) and
(7) implies that A and B therefore ommute. An elementary result from linear algebra [8 implies that these matri es
then have eigenvalues s and s0 with a ommon eigenve tor.
Pi ard's Theorem thus follows from part (iv) of Theorem
2.3.
Remark 2.5. The multipliers are eigenvalues and thus
independent of the fundamental system.
Remark 2.6. On e one doubly periodi solution of the
se ond kind has been found we may use redu tion of order
to nd additional solutions. In the present ase, redu tion
of order for (1) results in a new ode also having doubly periodi oe ients, and all solutions uniform. Pi ard's Theorem therefore implies that this new ode also has at least one
solution whi h is doubly periodi of the se ond kind. This
may be repeated, to show that there is a basis of solutions of
(1) of the form

Z
y1 (x) = 1 (x); y2 (x) = 1 (x)

: : : ; yn (x) = 1 (x)

2 (x)



2 (x)dx;

n 1

n (x) (dx)


;

where the i (x) are doubly-periodi of the se ond kind (In e


[13, p. 376).

3. SOLVING 2ND ORDER EQUATIONS


VIA SYMMETRIC POWERS
In this se tion we are interested in determining when a
se ond order linear ode (1) has a basis of solutions whi h are
doubly-periodi of the se ond kind. From the omments at
the end of se tion 2.2 we know that this be omes equivalent
to sear hing for ellipti fun tion rst order fa tors of the
asso iated linear di erential operator. Thus, in the ase
of se ond order equations, we an look to apply the same

te hniques as used in modern versions of Kova i 's algorithm


[18. In parti ular, we an redu e our problem to nding
ellipti fun tion solutions of a related third order linear ode.
It is well known that using the transform
 Z

an 1 (x)
y(x) = exp
dx  u(x)
(15)
nan (x)
onverts an n-th order ode in y(x) into an ode in u(x) having the se ond highest term 0. In the ase of se ond order
equations this onverts

 00 (x) + a

a2 (x) y

into

(x)  y0 (x) + a0 (x)  y(x) = 0

00 (x) r(x)  u(x) = 0


 a (x) 0  a (x) 2 a (x)
u

(16)

Lemma 3.1. Assume that equation (1) has oe ient


1 (x) = 0 and that fy1 (x); : : : ; yn (x)g is a basis of solutions with ea h yi (x) doubly periodi of the se ond kind.
Then z (x) = y1 (x)  yn (x) is doubly periodi .
Let T and T 0 be the periods for the oe ients
of the linear ode and suppose that yi (x + T ) = si  yi (x),
0
yi (x + T 0 ) = si  yi (x), for i = 1; : : : ; n. Then from part (iii)
of Theorem 2.3 we have that s1  sn = 1. Therefore
z (x + T ) = y1 (x + T )  yn (x + T )
= (s1  sn ) y1 (x)  yn (x)
(18)
= z (x):

Proof:

Similarly, z (x + T 0 ) = z (x) for the se ond period T 0 and


hen e z (x) is doubly periodi .
2
Consider rst the ase of nding a basis of solutions of
a se ond order linear ode of the form (17). We show in
Corollary 3.2 and Theorem 3.3 that solutions of su h an
equation an be determined by looking for spe ial solutions
of
(19)

the se ond symmetri power of equation (17).


Corollary 3.2. Suppose that equation (17) has a basis of
solutions whi h are doubly-periodi of the se ond kind. Then
the se ond symmetri power (19) has at least one solution
z (x) whi h is doubly periodi .

Proof: It is well known ( .f. [14) that a basis for solutions of equation (19) is given by fy1 (x)2 ; y1 (x)y2 (x); y2 (x)2g
where fy1 (x); y2 (x)g is a basis for equation (17). The orollary thus follows dire tly from Lemma 3.1.
2
The following theorem shows how to build a basis for (17)
from a solution of (19).

y2 (x) =

z (x)

 exp( C2

dx
)
z (x)

where C is a onstant given by


2
0 2 2  z (x)  z 00(x) + 4  r(x)  z (x)2:
C = z (x)

while

= y20 (x)  y1 (x)


C
z (x)

y10 (x)
y1 (x)

y20 (x)
y2 (x)

y2 (x) y10 (x)

y20 (x)

y2 (x)

(23)

implies that

y10 (x)

y1 (x)

(24)

y10 (x)
y1 (x)

z 0 (x) C
2 z (x)

and

y20 (x)
y2 (x)

z 0 (x) + C
2 z (x)

0
whi h gives (21). In addition, di erentiating yy11 ((xx)) gives
 0 2
y100 (x)
y1 (x)
2  z (x)  z 00(x) 2  z 0 (x)2 + 2  C  z 0 (x)
=
y1 (x)
y1 (x)
4  z (x)2
so that

y100 (x)
y1 (x)

2  z (x)  z 00 (x) z 0(x)2 + C 2


:
4  z (x)2

00
Sin e yy11 ((xx)) = r(x) we get identity (22). Now, sin e C is
the Wronskian of y1 (x), y2 (x), C 6= 0 implies that y1 (x),
y2 (x) are linearly independent and form a basis for (17).
On the other hand, if C = 0, the formulas
(21) be ome
p
identi al, with one solution y1 (x) = z (x). A se ond solution, from standard
redu tion
p ofR order te hniques, is then
R
y2 (x) = y1 (x) y21(x) dx = z (x) z(1x) dx.
2
1

Remark 3.4. Theorem 3.3 appears to have been known


in the 1800's by Hermite [11, Brios hi [5, and Halphen
[10, at least in the ase of Lame's equation. Brios hi also
used the result to analyze Mathieu's equation
00
2
2
2
y (x) + (p k  v  os (x))  y(x) = 0
(25)

(so r(x) = k2 v2 os2(x) p) where the oe ients are only


periodi . Indeed, in this ase equation (22) is known as
Brios hi's Identity.

(21)

In the ase of se ond order linear odes having ellipti fun tion oe ients, Theorem 3.3 provides a simple method for
nding a general solution for the ode whenever two independent solutions, doubly periodi of the se ond kind, exist.
Classi ally these solutions are given in terms of the Weierstrass Zeta and Sigma fun tions whi h are de ned in terms
of } via
0 (x)
0
=  (x):
(26)
 (x) = }(x) and
(x)
These Weierstrass fun tions are not periodi but rather quasiperiodi ( f. [1).

(22)

positive integer,

Theorem 3.3. Let z (x) be a solution of (19) and set


Z
p
C
dx
)
(20)
y1 (x) = z (x)  exp(
2
z (x)

and

z 0 (x)
z (x)

Taking equations (23) and (24) together gives

an

000 (x) 4  r(x)  y0 (x) 2  r0 (x)  y(x) = 0;

Proof: Let z (x) = y1 (x)  y2 (x) and C be the Wronskian


of y1 (x); y2 (x). Note that C is a onstant by Abel's identity
[13. From z 0(x) = y10 (x)  y2 (x) + y1 (x)  y20 (x) we have that

(17)

0
with r(x) = 2a12 (x) + 2a12 (x)
a2 (x) .
The following is entral to the results in this se tion.

If C 6= 0, then y1 (x), y2 (x) are linearly independent and


form a basis of (17). If C = 0, then a basisp
for theR solution
spa e of (17) is given by y1 (x) and y2 (x) = z (x) z(1x) dx.

Example 3.5. Consider Lame's equation (2), with n any


B , g2 ,

and

g3

arbitrary onstants. We

sear h for solutions that are doubly periodi of the se ond


kind. The orresponding symmetri power is given by
000
0
0
z (x) 4 (n(n + 1)}(x) + B ) z (x) 2n(n +1)} (x)z (x) = 0
(27)

and this has a doubly periodi solution. Thus we an ompute solutions of (2) whi h are doubly periodi of the se ond
kind.
We an illustrate with some examples for n small. For
example, if n = 1, then solutions z (x) of (27) whi h are
rational in } and }0 are given by multiples of f} B g. Then
C 2 = z 0 (x)2 2z (x)z 00(x) + 4r(x)z (x)2 = 4B 3
g2 B g 3 ;
whi h is a onstant. If C 6= 0, two independent solutions are
R z0 (x) C 
1
y~1 (x) = exp
dx
 R2z10(x) p

3
= exp

1
2

} (x)

4B
g2 B g3
dx
}(x) B

R z0 (x)+C 
1
dx
y~2 (x) = exp
 R2z10(x) p
4B 3 g2 B
}
(
x
)+
= exp 21
}(x) B

g3

(28)


dx

We an also return the nal solution in a form found in


standard texts. Indeed, up to an additive onstant, we have
p 3


Z
4B g2 B g3
(x + a)
1 }0 (x)
2

}(x)

and
p
Z
1 }0 (x) + 4B 3
2

}(x)

dx

g2 B
B

g3

dx

= ln

= ln

x (a)

(x)(a)

(x + a)
(x)(a)

(29)

+x (a)
(30)

where a is a onstant su h that }(a) = B ( f. [1, 2). Hen e,


from (28), we have the solutions
 


 (x+a)
 (x+a)
y1 (x)

= exp ln

y2 (x)

 (x) (a)

x (a)

 (x) (a) e

x (a)

 (x a) x (a)
;
 (x) (a) e

where a is su h that }(a) = B .


For higher values of n, solutions for z (x) and C , and hen e
for y1 (x) and y2 (x), are still obtained. However, it be omes
more di ult to evaluate the integrals for in reasing values
of n so we just give z (x) and C 2 for n = 2 and 3:
1 2 1
1
2
g + B
B}(x) + }(x)
4 2 9
3
1 2
1
2
2
C
=
g2 g3 + g2 B + g3 B
3
3
7
4 5
3
g B +
B
27 2
81
1
1
1 3 1
n = 3 : z (x) =
g +
g B
B
g }(x)
4 3 15 2
225
4 2
2
1
2
3
+ B 2 }(x)
B}(x) + }(x)
75
5
1 3
9 2
9
2
2
C
=
g B+
g B
g2 g3 B
60 2
20 3
50
31 2 3
22
4
+
g B +
g B
1500 2
1875 3
4
14
5
7
g B +
B
5625 2
50625
(31)
n

= 2 : z (x) =

Using the omputer algebra system Maple, the solutions


generated from these solutions for z (x) and C
have all been veri ed to be solutions of the orresponding
instan e of Lame's equation.
2

y1 (x), y2 (x)

4. SOLVING 2ND ORDER EQUATIONS


VIA DIFFERENTIAL FACTORIZATION
If (17) has a doubly periodi of the se ond kind solution,
then by Theorem 3.3 su h a solution an be found when
there is a doubly periodi solution of (19). However, if (19)
has no doubly periodi solutions, then this does not ne essarily imply that there are no doubly periodi solutions of
the se ond kind for (17). For example, the linear ode
y

00 (x) (}(x)2

} (x))y(x) = 0

(32)

has a solution whi h is doubly periodi of the se ond kind


even though its se ond symmetri power has no doubly periodi solutions. In this se tion we give an alternate method
whi h will solve su h equations.
Re all from subse tion 2.2 that the logarithmi derivative of any doubly periodi fun tion of the se ond kind is
doubly-periodi . As su h any doubly-periodi solution of
the se ond kind of (1) gives a rst order right fa tor of our
original equation. Finding rst order fa tors of a linear operator is equivalent to nding exponential solutions of the
orresponding linear ode. In the ase of linear di erential
d
we an
operators in the domain K(x)[Dx where Dx = dx
nd exponential solutions with existing algorithms su h as
[4, 6, 12 while in our domain we an nd exponential solutions via the method from [17.
Assume now that our linear ode (1) has order 2. By using
a redu tion of the form (15) we an redu e the problem
of nding doubly-periodi solutions of the se ond kind to
nding right hand fa tors of
L

= Dx2

r(x)

where

r(x)

= a(}(x))+ b(}(x))}0 (x) (33)

with a(z ); b(z ) 2 K(z ).


Noti e thatpif one does a substitution of the form z = }(x)
d
then Dx = w(z )Dz with Dz = dz
where w(z ) 2 K[z
0
2
satis es } (x) = w(}(x)). pThus we an express L as a
di erential operator in K(z; w(z ))[Dz by
p
w0(z )
2
(34)
Dz a(z ) b(z ) w(z ):
L = w(z )Dz +
2
We may assume that b(z ) 6= 0 (otherwise L is in K(z )[Dz in
whi h ase we an use the Kova i algorithm instead of the
method given below). We an reate a fourth order operator
having rational oe ients by
^ := symmetri produ t(L; L) 2 K(z )[Dz
L
(35)
p
z ))[Dz is the onjugate of L (that is,
where L 2 K(z; w(p
p
w(z )). The
every o urren e of w(z ) is repla ed by
^ of minimal order
symmetri produ t produ es an operator L
su h that the produ t of any solution of L and any solution
of L is a solution of L^ .

operator. A doubly-periodi solution of the se ond kind for


our example is then given by

Theorem 4.1.

(a) Suppose Dx

is a right fa tor of L with


0
s(x) = u(}(x)) + v(}(x))} (x):

Then Dz

s(x)

=
=
=

(z )

is a right fa tor of L^ . Set

(z )

;
2
v(z )2 w(z ) + v0 (z )w(z ) + 12 v(z )w0(z );
1
(a0 (z ) + 4a(z )v(z ) 4t(z )v(z ) t0 (z )):
2b(z )

If a(z ) = u(z )2 + t(z ), then we have the fa torization


L

= (Dx + s(x))(Dx

s(x))

where s(x) = u(}(x)) + v(}(x))}0 (x).

(36)

Proof: Suppose that Dx s(x) is a right fa tor of L.


Changing oordinates of this fa tor via z = }(x) and simu(z ) p
plifying ensures that Dz
v(z ) + w(z) w(z ) is a right


u(z ) p
w
(
z
)
fa tor of L. As su h Dz
v(z )
is a right
w (z )
fa tor of L and so Dz 2v(z ) is a right fa tor of L^ , the
symmetri produ t of L and L. Thus for ea h right fa tor
^ we obtain the orresponding formula for v(z ).
of L
In order to obtain a formula for the omponent u(z ) rst
noti e that Dx2 r(x) = (Dx + s(x))(Dx s(x)) if and only
if
r(x)

= s0 (x) + s(x)2

whi h in turn o urs if and only if


+ v(z )2 w(z ) + v0 (z )w(z ) +

a(z )

u(z )

b(z )

u (z ) + 2u(z )v(z ):

1
0
v(z )w (z )
2
(37)
(38)

Taking derivatives of equation (37), multiplying equation


(38) by 2u(z ), taking di eren es of the two resulting equations and then substituting for u(z )2 using equation (37)
gives the formula for u(z ).
2
Example 4.2. Consider rst equation (32) from the start
of this se tion. In this ase the symmetri produ t has a right
fa tor of the form Dz . As su h both v(z ) = 0 and t(z ) = 0.
Further omputation gives u(z ) = z and hen e the linear
ode asso iated to (32) has a right fa tor Dx + }(x). This
in turn gives a doubly-periodi of the se ond kind solution of
(32) of the form
e

}(x)dx

(1+

}0 (x) )dx
}(x)

= ex }(x):

For both examples, a se ond independent solution for the ode


an be found using redu tion of order.
2

2v(z ) is a right fa tor of L^ .

(b) Suppose Dz
v(z )
t(z )
u(z )

= e(x) :

Example 4.3. Consider now the equation


g2
2 0
00
y (x) (6}(x) + 1
+
} (x))y(x) = 0:
2}(x) }(x)

As before, the orresponding symmetri power equation (17)


does not have a doubly-periodi solution. Forming the operators L and L, taking the symmetri produ t L^ and fa toring
then gives a right fa tor of the form Dz 2z . Using Theorem 4.1 we determine that v(z ) = z1 and u(z ) = 1. Hen e
}0 (x)
Dx (1+ }(x) ) is a right fa tor of our original se ond order

Remark 4.4. We an obtain a similar result for fa torizations of the form


= (Dz + s(z ))(Dz s(z ))
p
p
where r(z ) = a(z )+b(z ) w(z ) and s(z ) = u(z )+v(z ) w(z ).
Indeed, if Dz (z ) is a right fa tor of L 2 K(z )[Dz then
the onditions for su h a fa torization would be u(z ) = (2z)
and
(a0 (z ) + 4u(z )a(z ) 4u(z )3 6u(z )u0 (z ) u00 (z ))
v(z ) =
2b(z )w(z )
2
2
with a(z ) = u(z ) + v(z ) w(z ) + u0 (z ). Thus one an obtain a similar pro edure by rst hanging oordinate systems,
then normalizing
to remove the linear term of the operator
p
in K(x; w(z ))[Dz and nally nding right fa tors of the
resulting symmetri produ t in K(z )[Dz .
L

= Dz2

r(z )

Remark 4.5 (Completeness). Theorem 4.1 provides


a omplete method for nding doubly periodi solutions of the
se ond kind for a se ond order linear operator of the form
(34). Namely, ompute all exponential solutions of L^ whi h
in turn gives all rst order right fa tors Dz (z ) of L^ . Then
sear h for those (z ), if any, for whi h the ondition a(z ) =
u(z )2 + t(z ) in Theorem 4.1 holds. Those then produ e the
fa torization (36).
Remark 4.6 (Implementation). An algorithm for
nding all solutions doubly periodi of the se ond kind for
se ond order equations has been implemented and will be in
the oming version of Maple. In implementing our results
[7 we have taken advantage of the existen e of an e ient
implementation of the method from se tion 3. By rst trying
the methods from se tion 3 we do not need to implement the
omplete version of Theorem 4.1. We mention two reasons
here. First, be ause we have tried the method of symmetri
powers, we do not need to ompute all exponential solutions
of L^ . Instead we only sear h for those that are relatively
easy to ompute, namely those that do not involve algebrai
extensions of the onstants, whi h makes our implementation mu h faster. Se ondly, if there exist in nitely many
rst order right fa tors Dz (z ) of L^ , then all su h (z )
an be represented by nitely many (z ) that ontain parameters, and we have to nd the parameter values for whi h
a(z ) = u(z )2 + t(z ). While nding these parameters is not
hard to implement, we have not done so.
We an justify the simpli ations in our implementation
of Theorem 4.1 as follows. If we have two rst order fa tors,
then the method of se tion 3 will nd them. Hen e we only
need to worry about the ase where we have a unique rst
order fa tor. Sin e this fa tor does not require an extension of the eld of onstants, it is also the ase that the
fa tor Dz (z ) of L^ does not require extensions of the
eld of onstants. Hen e no solutions are in fa t lost by
not extending our oe ient eld. Similarly, suppose L^ has
in nitely many right fa tors. In this ase, where L has a
unique rst order right fa tor while L^ has in nitely many,

it an be shown that the se ond symmetri power will have


an ellipti fun tion solution, and hen e the method from se tion 3 will again work. Thus, our simpli ed implementation
for Theorem 4.1 ombined with the method in se tion 3 gives
a omplete algorithm for solutions whi h are doubly periodi
of the se ond kind for se ond order linear odes with ellipti
fun tion oe ients.

5. CONCLUSION
We have implemented an e ient algorithm for nding
ellipti fun tion solutions of arbitrary order linear odes of
the form (1) along with a pro edure based on the results of
se tion 3 to nd general solutions for (1) of the se ond order. The algorithms are in luded in version 9 of the Maple
omputer algebra system, used both inside dsolve, the differential equation solver, and as a stand alone fun tion dperiodi sols in the DEtools pa kage. The extensions reported
in se tion 4 have been added and will be in luded in the
next version of Maple. A report on these implementations
will be forth oming [7.
There are a number of topi s for future resear h. We
are interested in developing e ient and omplete methods for all solutions whi h are doubly-periodi of the se ond
kind, for odes of the form (1) whi h have orders higher than
2. In addition, we are interested in nding di erential fa torizations
of higher order linear di erential operators in
p
K[z; w(z ) with w(z ) 2 K[z . Finally, in the ase of se ond order odes we are interested in extending the te hniques
used in this paper to ompute general Liouvillian solutions
for (1) when there are no solutions doubly periodi of the
se ond kind.
A knowledgements: We would like to thank Anne Fredet
and the anonymous referees for their helpful omments on
this paper.

6. REFERENCES
[1 M. Abramowitz and I.A. Stegun, Handbook of
Mathemati al Fun tions, Dover Publi ations, New
York, (1972).
[2 N.I. Akhiezer, Elements of the Theory of Ellipti
Fun tions, Ameri an Mathemati al So iety, (1990).
[3 P.E. Appell and E. La our, Prin ipes de la Theorie
des Fon tions Elliptiques et Appli ations,
Gauthier-Villars et Cie, (1897).

[4 E. Beke, Die Irreduzibilitat der homogenen linearen


Di erentialglei hungen, Math. Ann. 45, p. 278-294,
(1894).
[5 F. Brios hi, Sur l'equation de Lame. Comptes Rendus
de l'A ademie des S ien es, 86, 313-315, (1878).
[6 M. Bronstein, Linear Ordinary Di erential
Equations: breaking through the order 2 barrier,
Pro eedings of ISSAC'92 (P. Wang ed.), 42-48,
(1992).
[7 R.F. Burger, G. Labahn and M. van Hoeij, An
algorithm to solve linear odes with ellipti fun tion
oe ients, Manus ript.
[8 W.L. Ferrar, Algebra; A Text-book of Determinants,
Matri es, and Algebrai Forms, 2nd edition, Oxford
University Press, (1957).
[9 A.R. Forsyth, Di erential Equations I{VI,
Cambridge University Press, Cambridge, England,
(1906).
[10 G.H. Halphen, Traite des Fon tions Elliptiques et de
Leurs Appli ations, Volume 2. Gauthier-Villars et
Fils, (1888).
[11 C. Hermite, Oeuvres, Volume 3. Gauthier-Villars,
(1912).
[12 M. van Hoeij, Fa torization of Di erential Operators
with Rational Fun tion Coe ients, Journal of
Symboli Computation 24 537-561 (1997).
[13 E. In e, Ordinary Di erential Equations, Dover
Publi ations, New York, (1956).
[14 E. Kamke, Di erentialglei hungen: Losungsmethoden
und Losungen, Chelsea Publishing Co, New York,
(1959).
[15 E. Pi ard, Sur une generalisation des fon tions
periodiques et sur ertaines equations di erentielles
lineaires, Comptes Rendus de l'A ademie des
S ien es, 89, 140-145 (1879).
[16 M.F. Singer, Liouvillian Solutions of n-th order
Homogeneous Linear Di erential Equations,
Ameri an Journal of Mathemati s, 103 661-682
(1981).
[17 M.F. Singer, Liouvillian Solutions of Linear
Di erential Equations with Liouvillian Coe ients,
Journal of Symboli Computation, 11 251-273
(1991).
[18 F. Ulmer and J.A. Weil, A Note on Kova i 's
Algorithm, Journal of Symboli Computation, 22
179-200 (1996).