1 s2.0 S0747717197901103 Main

J.
Symbolic Computation (1997) 24, 1–30
Formal Solutions and Factorization of Differential

Operators with Power Series Coefficients
MARK VAN HOEIJ†
Department of Mathematics, University of Nijmegen,
6525 ED Nijmegen, the Netherlands
The topic of this paper is formal solutions of linear differential equations with formal
power series coefficients. The method proposed for computing these solutions is based
on factorization of differential operators. The notion of exponential parts is introduced
to give a description of factorization properties and to characterize the formal solutions.
The algorithms will be described and their implementation is available.
c 1997 Academic Press Limited
1. Introduction
Factorization of differential operators is a powerful computer algebra tool for handling
ordinary linear differential equations. It can be applied to compute formal solutions and
to study the structure of a differential equation. A differential equation
y (n) + an−1 y (n−1) + · · · + a1 y 0 + a0 y = 0
corresponds to a differential operator
f = ∂ n + an−1 ∂ n−1 + · · · + a0 ∂ 0
acting on y. Here the coefficients ai are elements of the differential field k((x)) and ∂
is the differentiation d/dx. The field k is the field of constants. It is assumed to have
characteristic 0. The differential operator f is an element of the non-commutative ring
k((x))[∂]. This is an example of an Ore ring (Ore, 1933).
Sections 6 and 8 contain the main results of this paper. These results are expressed
using the notion of exponential parts. The exponential parts will be studied in Section 6
from the viewpoint of factorization, and in Section 8 from the viewpoint of formal so-
lutions. They form the key ingredient for our factorization algorithm for k(x)[∂] in (van
Hoeij, 1996, Chapter 3). Another application is found in Section 9. Here the question
is: when is a given vector space a solution space of a certain differential operator? This
question can easily be answered using the direct sum splitting in Section 8.
The algorithms in this paper are given in Sections 4, 5 and 8.4. From an algorithmic
† E-mail: hoeij@sci.kun.nl
0747–7171/97/010001 + 30 $25.00/0 sy970110

c 1997 Academic Press Limited
2 M. van Hoeij
point of view the factorization in k((x))[∂] is the central problem because the other
algorithms in this paper require this tool. We will discuss it in the rest of this section.
Note that in general elements of k((x)) consist of infinitely many terms. Only a finite
number of them can be computed. This means that a factorization can only be determined
up to some finite accuracy. The notion of accuracy will be formalized later. Increasing
the accuracy of a factorization will be called lifting a factorization.
From Malgrange (1979) we know that an element of k((x))[∂] which has only one
slope in the Newton polygon (cf. Section 3.3) and which has an irreducible Newton
polynomial (cf. Section 3.4) is irreducible in k((x))[∂]. Malgrange (1979) shows that in
the following two cases a differential operator f ∈ k((x))[∂] is reducible in this ring and
how a factorization can be computed:
1. An operator with a broken Newton polygon (i.e. more than one slope).
2. An operator with one slope > 0 where the Newton polynomial is reducible and not
a power of an irreducible polynomial.
In our method these two cases of factorization and the factorization of regular singular
operators are called coprime index 1 factorizations. Coprime index 1 means that the
factorization can be lifted by the usual Hensel lifting (cf. any book on computer algebra)
procedure. For a definition of the coprime index see Section 2.
Example 1.1.
1 3 2 1 1
f = ∂4 + 2
∂ + 4 ∂2 + 6 ∂ + 8 .
x x x x
The Newton polynomial is T 4 +T 3 +2T 2 +T +1. This polynomial can be factored over Q
as (T 2 +1)(T 2 +T +1). Because T 2 +1 and T 2 +T +1 in Q[T ] are coprime (i.e. the gcd is 1)
we can conclude from Malgrange (1979) that f is reducible in Q((x))[∂]. A factorization
of f = LR is obtained in two steps. The first step is to compute the factorization up to
accuracy 1 (definitions follow later, this integer 1 is related to the coprime index). This
accuracy is obtained when we have the Newton polynomials T 2 + 1 and T 2 + T + 1 of L
and R (here T 2 +1 and T 2 +T +1 can be interchanged to obtain a different factorization).
The next step is to lift the factorization up to the desired accuracy. Because T 2 + 1 and
T 2 + T + 1 are coprime this lifting can be done by the usual Hensel lifting procedure. In
each lift step the extended Euclidean algorithm is used. Note that in this example the
reducibility of f can be concluded from very few coefficients of f in k; the coefficients
which determine the Newton polynomial are sufficient.
Now there remains one hard case of factorization in k((x))[∂]. Here f has one slope
s 6= 0 and the Newton polynomial is of the form P d , where P is an irreducible polynomial
over k and d is an integer > 1. In this case it is more difficult to decide if f is reducible
or not. A factorization of f will have coprime index > 1.
Example 1.2.
2 + x4 2 8 1 + 20x2
f = ∂4 + 4
∂ − 5∂ + .
x x x8
The Newton polynomial of f is T 4 + 2T 2 + 1 = (T 2 + 1)(T 2 + 1). Because the two factors
T 2 + 1 and T 2 + 1 are not coprime we cannot apply Hensel lifting to find a factorization
Formal Solutions and Factorization of Differential Operators with Power Series Coefficients 3
over Q((x))[∂]. Malgrange provides a factorization method in Q((x))[∂] for this case. We
want to find a factorization in Q((x))[∂]. In this example f is reducible in Q((x))[∂].
However, f + 1/x6 (replace the coefficient 20 by 21) is irreducible in Q((x))[∂]. In the
previous example adding 1/x6 would have no influence on the reducibility of f because
the reducibility could already be decided from the Newton polynomial. We see that this
example is more complicated because more coefficients of f are relevant for deciding
reducibility. We shall proceed as follows:
• Compute a first-order right-hand factor ∂ − r of f where r ∈ k((x)). We use a

variant on the method in Malgrange (1979) for this.
• Compute an operator R ∈ k((x))[∂] of minimal order such that ∂ −r is a right-hand
factor of R.
• Perform a division to find a factorization f = LR.
For some applications, like factorization in k(x)[∂], we need to compute the factors L
and R up to a high accuracy. The method sketched for computing L and R is not very
suitable for this because it is slow. We will use this slow method to compute L and R
up to a certain accuracy (up to the coprime index) and then use a different method to
lift the factorization. Coprime index > 1 means that the usual Hensel lifting does not
work because the Newton polynomials of L and R have gcd 6= 1. For this case we give a
variant on the Hensel lifting method in Section 4.
The factorization of a differential operator f is done recursively. If f can be factored
f = LR then the factorization algorithm is applied to the factors L and R (or only
to R when we are only interested in right-hand factors) until f is factored in irreducible
factors. This causes a difficulty; if a factorization is required with a given accuracy it is
not clear how accurate the intermediate factorizations should be. To solve this problem
we use lazy evaluation in our implementation. This is a computer algebra trick which
makes exact computation in k((x)) possible. It does not use truncations of some finite
accuracy. Instead, an expression a ∈ k((x)) is denoted as the name and arguments of a
procedure that computes coefficients of a. These coefficients are automatically computed
and stored when they are needed. This method of computing in k((x)) is very efficient
because coefficients which are not used will not be computed.
The use of factorization for computing formal solutions is beneficial for the efficiency in
case the solutions involve algebraic extensions, cf. the comments after algorithm formal
solutions in Section 8.4.
2. Valuations and the Coprime Index

S
A discrete valuation on a ring D is a map v : D → Z {∞} such that for all a and b
in D we have: v(ab) = v(a)+v(b), v(a+b) ≥ min(v(a), v(b)) and v(a+b) = min(v(a), v(b))
if v(a) 6= v(b). v(0) = ∞. An example: D is the field of p-adic numbers Qp or D is a
polynomial ring Qp [x] over the p-adic numbers. Define the valuation v(a) of a ∈ Qp [x] as
the largest integer n such that a ∈ pn Zp [x]. Another example: s ∈ Q and D = k((x))[y]
where k is a field. Write
P s = n/d where n and d are integers, gcd(n, d) = 1 and d > 0.
Now the valuation vs ( i,j ai,j xi y j ) is defined as the minimum id − jn for which ai,j 6= 0.
A third example: k is a field, s ∈ Q, s ≥ 0 and D = k((x))[δ]. Here δ is defined as
x∂ ∈ k((x))[∂], cf. Section 3.2. Write s = n/d where n and d are integers, gcd(n, d) = 1
4 M. van Hoeij
P
and d > 0. Now the valuation vs ( i,j ai,j xi δ j ) is defined as the minimum id − jn for
which ai,j 6= 0.
S with a chain of additive subgroups · · · ⊃ D−1 ⊃ D0 ⊃ D1 · · ·
A filtered ring is a ring D
such that: 1 ∈ D0 , D = n∈Z Dn and Dn Dm ⊂ Dn+m for all integers n and m. The
chain (Dn )n∈Z is called a filtration of D. The associated graded ring gr D is defined as
⊕n Dn /Dn+1 . The symbol map σ : D → gr D is defined as: σ(0) = 0, σ(f ) = f + Dn+1 if
f ∈ Dn \ Dn+1 . For more information about filtrations see Björk (1979). A valuation v
defines a filtration on a ring D as follows
Dn = {f ∈ D | v(f ) ≥ n}.
For positive integers a the set D0 /Da has the structure of a ring.
For a ring D with a valuation v we can define a truncation σa with accuracy a for
non-zero elements f of D and positive integers a as follows
σa (f ) = f + Dv(f )+a ∈ Dv(f ) /Dv(f )+a .
The symbol map is σ1 .
Suppose f ∈ D can be written as f = LR where L, R ∈ D. For invertible elements
u ∈ D we have f = LR = (Lu)(u−1 R). We will call the ordered pair L, R equivalent
with the pair Lu, u−1 R. Let t be a positive integer. Then the ordered pair L, R is called
coprime with index t if for all a ≥ t the pair σa+1 (L), σa+1 (R) is determined up to the
above equivalence by σa (L), σa (R) and σa+t (f ). Assume t is minimal, then t is called the
coprime index of L, R. If L, R is not coprime for any integer t then the coprime index
is ∞.
For our examples Qp [x], k((x))[y] and k((x))[δ] the notion of equivalence for pairs L, R
can be avoided by restricting ourselves to monic elements f , L and R. Then we can define
the coprime index of the factorization f = LR as the smallest positive integer t for which
the following holds: for all integers a ≥ t and monic elements L0 and R0 of D, if
σa (L0 ) = σa (L) and σa (R0 ) = σa (R) and σa+t (L0 R0 ) = σa+t (f )
then
σa+1 (L0 ) = σa+1 (L) and σa+1 (R0 ) = σa+1 (R).
Example 2.1. Suppose we want to factor f = x2 + x + 3 ∈ D = Q3 [x]. First we look at

the truncation σ1 (f ) = x2 + x ∈ D0 /D1 which factors as x(x + 1) ∈ D0 /D1 . Because x
and x + 1 have gcd 1 in D0 /D1 ' F3 [x] we can apply Hensel lifting to find a factorization
f = LR in D. To determine L and R up to some accuracy a we only need to know f up
to accuracy a. So the coprime index is 1 in this example.
Example 2.2. f1 = x4 −x2 −2 = L1 R1 = (x2 +1)(x2 −2) ∈ Q3 [x] and f2 = x4 −x2 −20 =
L2 R2 = (x2 + 4)(x2 − 5) ∈ Q3 [x]. Now f1 and f2 are the same up to accuracy 2 (i.e.
are congruent modulo 32 ) but the factorizations L1 , R1 and L2 , R2 are different up to
this accuracy. It follows that to determine the factorization of f1 up to some accuracy a
it is not sufficient to know σa (f1 ). This means that the coprime index of L1 , R1 is > 1.
We cannot apply ordinary Hensel lifting to find a factorization of f1 because σ1 (L1 ) and
σ1 (R1 ) have gcd 6= 1.
The name coprime index is explained from the case k((x))[y]. In this ring L, R have
finite coprime index if and only if L and R are coprime in the usual sense (i.e. gcd(L, R) =
1). It is shown in van Hoeij (1996, Chapter 5) that the coprime index of a factorization
f = LR in k((x))[δ] is always finite.
3. Preliminaries
This section summarizes the concepts and notations we will use in this paper. Defini-
tions will be brief; references to more detailed descriptions are given.
3.1. the field k((x))
k is a field of characteristic 0, k is its algebraic closure. k((x)) is the field of formal Lau-
rent series in x with finite pole order and coefficients in k. k((x)) is the algebraic
S closure
of k((x)). It is (cf. Bliss, 1966) contained in the algebraically closed field n∈N k((x1/n )),
the field of Puiseux series with coefficients in k.
A ramification of the field k((x)) is a field extension k((x)) ⊂ k((r)) where r is algebraic
over k((x)) with minimum polynomial rn − ax for some non-zero a ∈ k and positive
integer n (cf. Sommeling, 1993). If a = 1 this is called a pure ramification.
For r ∈ k((x)) (not necessarily with minimum polynomial rn − ax) we call the smallest
integer n for which r ∈ k((x1/n )) the ramification index ram(r) of r. If L is a finite
algebraic extension of k((x)) then the ramification index of L is the smallest n for which
L ⊂ k((x1/n )).
k((x)) is a differential field with differentiation d/dx. If k((x)) ⊂ L is an algebraic
extension then d/dx can be extended in a unique way to L. All finite algebraic extensions
k((x)) ⊂ L are of the following form:
L = l((r))
where k ⊂ l is a finite extension and l((x)) ⊂ l((r)) is a ramification (cf. Sommeling, 1993,
Proposition 3.1.5).
3.2. the ring k((x))[δ]
Define δ = x∂ ∈ k((x))[∂]. We have δx = xδ +x in k((x))[δ]. Since k((x))[∂] = k((x))[δ]

we can represent differential operators in the form f = an δ n + · · · + a0 δ 0 . This form has
several advantages. The multiplication formula
X X X X
xi Pi (δ) xj Qj (δ) = xn Pi (δ + j)Qj (δ)
i j n i+j=n
and the definition of the Newton polygon (cf. Section 3.3) are easier for operators with
this syntax. The operators we consider are usually monic. This means an = 1. The order
of a differential operator f is the degree of f as a polynomial in δ.
f is called the least common left multiple of a sequence of differential operators f1 , . . . , fr
if all fi are right-hand factors of f , the order of f is minimal with this property, and f
is monic. Notation: f = LCLM(f1 , . . . , fr ) (cf. Singer,P 1996). The solution space of f is
spanned by the solutions of f1 , . . . , fr . So V (f ) = V (fi ) where V (f ) stands for the
solution space of f . In order to speak about the solutions of differential operators a dif-
ferential extension of k((x)) is required that contains a fundamental system of solutions
of f1 , . . . , fr . For this we can use the so-called universal extension that we will denote
as V . This V is constructed as follows (this construction is obtained from Hendriks and
6 M. van Hoeij
van der Put (1995), our V is called R in Hendriks and van der Put (1995, Lemma 2.1.1)).
Define the set
[
E= k[x−1/n ].
n∈N
First view Exp(e) and log(x) as variables and define the free k((x))-algebra W in these
variables W = k((x))[{Exp(e) | e ∈ E}, log(x)]. Then define the derivatives Exp(e)0 =
e
x Exp(e) and define the derivative of log(x) as 1/x. This turns W into a differential ring.
We can think of Exp(e) as
Z
e
Exp(e) = exp dx
x
d
because x dx acts on Exp(e) as multiplication by e. Now define V as the quotient ring
V = W/I where the ideal I is generated by the following relations:
Exp(e1 + e2 ) = Exp(e1 ) Exp(e2 ) for e1 , e2 ∈ E
and
Exp(q) = xq ∈ k((x)) for q ∈ Q.
Note that this ideal is closed under differentiation. Hence V is a differential ring. It is
proven in Hendriks and van der Put (1995) that V is an integral domain and that k is
the set of constants of V . We denote the set of solutions of f in V as V (f ). This is a
k-vector space. Since every f ∈ k((x))[δ] has a fundamental system of solutions in V (cf.
(Hendriks and van der Put, 1995)) it follows that
dim(V (f )) = order(f ).
The substitution map Se : k((x))[δ] → k((x))[δ] is a k((x)) homomorphism defined by
Se (δ) = δ + e for e ∈ k((x)). Se is a ring automorphism. The following is a well-known
relation between the solution spaces:
V (f ) = Exp(e) · V (Se (f )).
The algorithm “Riccati solution” in Section 5.1 introduces algebraic extensions over
k((x)). This requires computer code for algebraic extensions of the constants k ⊂ l. But
we can avoid writing code for ramifications. Given a field extension k((x)) ⊂ k((r)) where
rn = ax for some a ∈ k we will use the following ring isomorphism
θa,n : k((r))[δ] → k((x))[δ]
defined by θa,n (r) = x and θa,n (δ) = n1 δ. This map enables us to reduce computations
in k((r))[δ] to computations in k((x))[δ].
3.3. the Newton polygon
The Newton polygon of a monomial xi y j in the commutative polynomial ring k((x))[y]

is defined as the set {(j, b) ∈ R2 | i ≤ b}. The Newton polygon N (f ) of a non-zero
polynomial f ∈ k((x))[y] is defined as the convex hull of the union of the Newton polygons
of the monomials for which f has a non-zero coefficient (cf. Bliss, 1966, p. 36). The main
property is N (f g) = N (f ) + N (g) for f and g in k((x))[y]. A rational number s is called
a slope of f if s is the slope of one of the edges of the polygon N (f ). If s is a slope of f g
then s is a slope of f or s is a slope of g.
For the non-commutative case f ∈ k((x))[δ] definitions of the Newton polygon are
given in Malgrange (1979), Tournier (1987) and Sommeling (1993, p. 48). N (xi δ j ) is
defined as {(a, b) ∈ R2 | 0 ≤ a ≤ j, i ≤ b} and N (f ) is again defined as the convex hull of
the union of the Newton polygons of the monomials that appear in f . This definition is
slightly different from the commutative case. As a consequence all slopes are ≥ 0. This
is needed to ensure N (f g) = N (f ) + N (g). If f has only one slope s = 0 then f is called
regular singular.
3.4. the Newton polynomial
Let s ≥ 0 be a rational number. We have defined a valuation vs and a truncation σa

for non-zero elements of k((x))[δ] in Section 2. σa depends on s and will from now on be
denoted as σa,s .
If s > 0 then σ1,s (L)σ1,s (R) = σ1,s (LR) = σ1,s (R)σ1,s (L) for all L and R in k((x))[δ].
If s = 0 then σ1,s (L)σ1,s (R) = σ1,s (LR) = S−vs (L) (σ1,s (R)) · Svs (R) (σ1,s (L)).
So σ1,s is commutative (i.e. is the same for LR and RL) if s > 0. If s = 0 then σ1,s is
commutative up to substitutions S−vs (L) and Svs (R) which map δ to δ plus some integer.
To σ1,s (f ) for f ∈ k((x))[δ] corresponds a certain polynomial, the so-called Newton
polynomial Ns (f ) (the definition is given after the example) of f for slope s. The Newton
polynomial is useful for factorization in k((x))[δ] because if f = LR then σ1,s (L)σ1,s (R) =
σ1,s (f ). So a factorization of f induces a factorization of the Newton polynomial.
Example 3.1. Consider the following differential operator

f = 7x−5 + 2x−6 δ + 2x−5 δ + 3x−5 δ 2 − 3x−5 δ 3 + 5x−4 δ 3 + x−4 δ 5
+2x−2 δ 5 + 2x−3 δ 6 + 3x−2 δ 7 + 2x−1 δ 8 + δ 9 .
In figure 1 we have drawn every monomial xi δ j which appears in f by placing the
coefficient of this monomial on the point (j, i). This gives a set of points (j, i). For
all points (j, i) for which xi δ j has a non-zero coefficient in f we can draw the rectangle
with vertices (0, i), (j, i), (j, ∞) and (0, ∞). The Newton polygon is the convex hull of the
union of all these rectangles. It is the part of the plane between the points (0, ∞), (0, −6),
(1, −6), (5, −4), (9, 0) and (9, ∞). In the commutative case (i.e. if we had written y instead
of δ in f ) the definition of the Newton polygon is slightly different and the point (0, −6)
would have been (0, −5) in this example. But for k((x))[δ] the Newton polygon is defined
in such a way that there are no negative slopes.
The slopes of f are 0, 1/2 and 1. The Newton polynomials are N0 (f ) = 2T , N1/2 (f ) =
T 2 −3T +2 and N1 (f ) = T 4 +2T 3 +3T 2 +2T +1. Here T is used as a variable. Ns (f ) will
be defined for all non-negative s ∈ Q. However, we will only use the Newton polynomial
for those values s which are a slope in the Newton polygon because for other values the
Newton polynomial is trivial (i.e. degree 0).
Write s = n/d where n and d are integers, gcd(n, d) =S1 and d > 0. The valuation vs
gives a filtration (Di ), i ∈ Z. σ1,s (f ) is an element of D = i∈Z Di /Di+1 . A multiplication
is defined for elements of D. An addition is only defined for a, b ∈ D which are element
of the same Di /Di+1 .
D0 and k[xn δ d ] are equal modulo D1 . There is a k-linear bijection
Ns0 : Di /Di+1 → k[T ]
which is also a ring isomorphism if i = 0. If i = 0 then Ns0 is defined by Ns0 (xn δ d ) = T .
8 M. van Hoeij
(1,0) (2,0)
1
j
(0,-1) 2
(0,-2) 2 3
5 1
7 2 3 -3
0 2
Figure 1. Newton polygon of Example 3.1.
For every i ∈ Z there is a unique pair of integers ni , di such that the map φi : D0 /D1 →
Di /Di+1 defined by φ(a) = xni δ di a is a bijection. The integers ni , di can be determined
from the conditions 0 ≤ di < d and vs (xni δ di ) = i. Now Ns0 (a) for a ∈ Di /Di+1 can
be defined as Ns0 (φ−1 0
i (a)). Ns is also defined for non-zero elements of f ∈ k((x))[δ]
as Ns (σ1,s (f )). In our example N00 (f ) = 2T , N1/2
0 0
(f ) = T 2 − 3T + 2 and N10 (f ) =
T 9 + 2T 8 + 3T 7 + 2T 6 + T 5 .
For slope s = 0 we define the Newton polynomial N0 (f ) as N00 (f ). From the multipli-
cation formula in Section 3.2 the following property follows for L, R ∈ k((x))[δ]
N0 (LR) = ST =T +v0 (R) (N0 (L))N0 (R).
Here ST =T +v0 (R) (N0 (L)) means N0 (L) with T replaced by T + v0 (R). For our example f
we get N0 (f · f ) = 4(T − 6)T .
For slope s > 0 we have the following property for L, R ∈ k((x))[δ]
Ns0 (LR) = T p Ns0 (L)Ns0 (R).
Here p is either 0 or 1, depending on the slope s and the valuations vs (L) and vs (R).
Let i = vs (L) and j = vs (R). Then φi (1) · φj (1) = xni +nj δ di +dj mod Di+j+1 . This is
either equal to φi+j (1) or xn δ d φi+j (1) mod Di+j+1 , depending on whether di + dj is
smaller than d or not. In the first case p = 0, in the second case p = 1. For our example
0 0
N1/2 (f ·f ) = T ·(N1/2 (f ))2 and N10 (f ·f ) = (N10 (f ))2 . Now define Ns (f ) as Ns0 (f ) divided
by T to the power the multiplicity of the factor T in Ns0 (f ). Then
Ns (LR) = Ns (L)Ns (R)
for s > 0 and for all L, R ∈ k((x))[δ].
Note that our definition does not correspond to the usual definition of the Newton
polynomial. It corresponds to the definition of the reduced characteristic polynomial in
Barkatou (1988). The roots of N0 (f ) in k are called the exponents of f . If f ∈ k((x))[δ]
is regular singular (i.e. f has only one slope s = 0, or equivalently degree(N0 (f )) =
order(f )) and all exponents of f are integers then f is called semi-regular.
Property 3.1. If f = LR then the Newton polynomial of the right-hand factor Ns (R)
divides Ns (f ). However, for a left-hand factor this need not hold. But if s > 0 or if
v0 (R) = 0 (for example if R is regular singular and monic) then Ns (f ) = Ns (L)Ns (R)
so in such cases Ns (L) divides Ns (f ).
4. The Lift Algorithm

Suppose f ∈ k((x))[δ] is monic and that f = LR is a non-trivial factorization, where L
and R are monic elements of k((x))[δ]. Let s ≥ 0 be a rational number. Recall that there
is a valuation vs on D = k((x))[δ], a filtration (Dn,s ), n ∈ Z and a truncation map
σa,s depending on s. In this section we will assume that L and R have been computed
up to some accuracy a. How to compute this σa,s (L) and σa,s (R) will be the topic of
Sections 5 and 7. In this section we deal with the question how to compute σa+1,s (L)
and σa+1,s (R) from σa,s (L), σa,s (R) and f in an efficient way. The goal is an algorithm
with the following specification:
Lift Algorithm.
Assumption. f = LR where f, L, R are monic elements of k((x))[δ].
Input. a ≥ 1, s, σa,s (L), σa,s (R) and f .
Output. Either σa+1,s (L) and σa+1,s (R) or “failed”, where “failed” can only occur if
t > a where t is the coprime index.
We use this algorithm to lift a factorization. If the output is “failed” then we will use
the less efficient method in Section 7 to lift the factorization. Note that since a ≥ 1 this
can only happen if the coprime index is > 1.
Suppose t ≤ a. We will use indeterminates for those coefficients of σa+t,s (L) and
σa+t,s (R) which are not yet known. Then the equation σa+t,s (LR) = σa+t,s (f ) gives
a set of equations in these unknowns (more details on how to find these equations are
given below). t ≤ a is needed to ensure that all these equations are linear. Coprime
index t means that σa+1,s (L) and σa+1,s (R) can be uniquely determined from these
linear equations.
Unless the coprime index is 1, our algorithm usually does not know the coprime index
in concrete situations. Then the lift algorithm will use a guess for the coprime index.
If the lift algorithm is called for the first time, it takes t = 2. Otherwise it takes the
guess for t that was used in the previous lift step. Then it will try, by solving linear
equations, if there is a unique solution for σa+1,s (L) and σa+1,s (R) from σa,s (L), σa,s (R)
and σa+t,s (f ). If so, t remains unchanged and the accuracy of the factorization increases;
the output of the lift algorithm is σa+1,s (L) and σa+1,s (R). If the solution for σa+1,s (L)
and σa+1,s (R) is not unique (there is at least one solution because of the assumption
that the factorization f = LR exists) the number t will be increased by 1. If t is still ≤ a
then we can use recursion with our increased guess t for the coprime index. Otherwise, if
10 M. van Hoeij
t > a, the output of the lift algorithm is “failed”, and we will have to use the less-efficient
method in Section 7 to lift the factorization. Note that the efficiency of our lift algorithm
depends on the coprime index, if this number is very high then it may not provide any
speed-up over the method from Section 7.
A truncation σa,s (R) = R + Dvs (R)+a is stored as an element R0 ∈ k[x, 1/x, δ] with no
terms in Dvs (R)+a . Now write
X
r= rij xi δ j
i,j
where the sum is taken over all i, j such that vs (R) + a ≤ vs (xi δ j ) < vs (R) + a + t and
j ≤ order(R). Here rij are indeterminates. We set rij = 0 for j = order(R), i 6= 0, and
set rij = 1 for j = order(R), i = 0. Similarly write L0 and l. Now we look for values
for the lij and rij such that R0 + r and L0 + l approximate R and L up to accuracy
a + 1. If the coprime index is t, the accuracy is at least a + 1 if the following holds:
σa+t,s ((L0 + l)(R0 + r)) = σa+t,s (f ), or equivalently
(L0 + l)(R0 + r) ≡ f mod Dvs (f )+a+t .
(L0 + l)(R0 + r) = L0 R0 + lR0 + L0 r + lr. To determine lR0 mod Dvs (f )+a+t it suffices to
have R0 up to accuracy t because vs (l) + vs (R0 ) ≥ vs (f ) + a. Similarly σt,s (L0 ) suffices
to compute L0 r mod Dvs (f )+a+t . vs (lr) ≥ vs (f ) + a + a ≥ vs (f ) + a + t so lr vanishes
modulo Dvs (f )+a+t . Hence
f ≡ L0 R0 + lσt,s (R0 ) + σt,s (L0 )r mod Dvs (f )+a+t .
By equating the coefficients of the left-hand side to the coefficients of the right-hand side
(the coefficients of all monomials of valuation < vs (f )+a+t) we find the linear equations
in lij and rij . To determine these equations we must multiply l by σt,s (R0 ), (= σt,s (R)
because R0 equals R up to accuracy a and t ≤ a) which is the lowest block of R with
slope s and width t in the Newton polygon of R. Similarly we must compute σt,s (L0 )r.
Usually the most time consuming part is the multiplication L0 R0 modulo Dvs (f )+a+t .
One approach is the following. Compute L0 R0 in k[x, 1/x, δ] and store the result together
with L0 and R0 . In the next lift step a similar multiplication must be performed, but
then L0 and R0 are slightly changed. Suppose we must compute the product (L0 +e1 )(R0 +
e2 ) in the next lift step. Here L0 and R0 are large expressions and e1 and e2 are small.
Using the previous multiplication L0 R0 we can speed up this multiplication by writing
(L0 + e1 )(R0 + e2 ) = L0 R0 + e1 R0 + L0 e2 + e1 e2 . The result of this multiplication is again
stored for use in the next lift step.
In this approach L0 R0 has been computed exactly. This is not efficient since we only
need it up to accuracy a+t, i.e. modulo Dvs (f )+a+t . Computing modulo Dvs (f )+a+t is not
as convenient as computing modulo a power of x when using the multiplication formula
in Section 3.2. We compute L0 R0 modulo a suitable power of x such that L0 R0 can still
be determined modulo Dvs (f )+a+t . Unless the slope s is zero, however, a few more terms
of the product L0 R0 than needed have been computed then. These terms are stored to
speed up the multiplication the next time that the lift algorithm is called.
5. Coprime Index 1 Factorizations

The lifting process for coprime factorizations can be done by solving linear equations.
However, for coprime index 1 solving linear equations can be avoided. In this case we must
solve a system (see Section 4) of the form lσ1,s (R) + σ1,s (L)r = g where g is computed by
multiplying the so-far computed truncations (called L0 and R0 in Section 4) of L and R
and subtracting this product from f . As in Section 3.4 this equation can be converted
to an equation lR0 + rL0 = g for certain l, r, L0 , R0 , g ∈ k[T ] and l, r unknown. Such an
equation can be solved by the Euclidean algorithm.
Consider the example f in Section 3.4. f has slopes 0, 1/2 and 1 in this example. In
Malgrange (1979) a method is given to compute a right-hand factor f1 with only slope 0
and order 1, a right-hand factor f2 with slope 1/2 and order 4 and a right-hand factor f3
with slope 1 and order 4. The Newton polynomial of f2 is the same as the Newton
polynomial N1/2 (f ) of f for slope 1/2. It is 2 − 3T + T 2 = (T − 1)(T − 2). Because
gcd(T − 1, T − 2) = 1 this f2 is again reducible in Q((x))[δ], cf. Malgrange (1979). It
has a right-hand factor g1 of order 2 and slope 1/2 with Newton polynomial T − 1 and
a right-hand factor g2 with Newton polynomial T − 2. So to obtain g1 two factorization
were needed. In one application, our algorithm for factorization in Q(x)[∂], we are mainly
interested in one of the irreducible right-hand factors of f in Q((x))[δ]. That is why we
want to be able to compute g1 directly without using an intermediate factorization to
compute f2 . This is done by the following algorithm.
Algorithm Coprime Index 1 Factorizations.
Input. f ∈ k((x))[δ], f monic
Output. All monic coprime index 1 factorizations f = LR in k((x))[δ] such that R does
not have a non-trivial coprime index 1 factorization.
Note. The definition of coprime index depends on the valuation that is chosen on k((x))[δ].
Here only the valuations vs that are defined in Section 2 are considered.
for all slopes s of f do
g := Ns (f )
Compute a prime factorization of g in k[T ], g = cg1e1 · · · grer ,
where gi are the different monic irreducible factors and c ∈ k.
if s = 0 then
M := {g1 , . . . , gr }
N := M \ {g | g(T ) = h(T + i), h ∈ M, i ∈ N, i > 0}
else
N := {g1e1 , . . . , grer }
end if
for h in N do
Write h = T p + hp−1 T p−1 + · · · + h0 T 0 .
Write s = n/d with d > 0 and gcd(n, d) = 1 (if s = 0 then n = 0, d = 1)
R0 := δ pd + hp−1 x−n δ (p−1)d + hp−2 x−2n δ (p−2)d + · · · + h0 x−pn δ 0 .
Now R0 has Newton polynomial h. We want to lift R0 to a right-hand
factor R such that R0 is R modulo Dvs (R0 )+1 .
L := an operator such that σ1,s (f ) = σ1,s (L0 R0 ).
0
L0 is uniquely determined by requiring that L0 has no

monomials of valuation > vs (L0 ).
f , L0 , R0 with the lift algorithm gives a factorization f = LR
end do
end do
We need to prove the following.
12 M. van Hoeij
1. L0 and R0 lift to a unique coprime index 1 factorization f = LR.

2. The right-hand factors R do not allow a non-trivial coprime index 1 factorization.
3. All such coprime index 1 factorizations f = LR (f , L and R monic) are obtained
this way.
Suppose σa,s (L0 R0 ) = σa,s (f ), meaning that the factorization has been lifted up to
accuracy a. If no lift steps were done yet, then a = 1. Now we look for l ∈ Dvs (L0 )+a and
r ∈ Dvs (R0 )+a such that σa+1,s ((L0 +l)(R0 +r)) = σa+1,s (f ) and order(r) < order(R0 ). To
prove statement 1 we have to show that l, r exist and that σa+1,s (L0 +l) and σa+1,s (R0 +r)
are uniquely determined. This means that l mod Dvs (L)+a+1 ∈ D (cf. Section 3.4) and
r mod Dvs (R)+a+1 ∈ D are uniquely determined. Then L0 and R0 are replaced by L0 + l
and R0 + r and the accuracy of the approximations L0 and R0 for L and R has increased
by 1. l and r must satisfy the following equation in D
σ1,s (L)r + lσ1,s (R) = f − L0 R0 mod Dvs (f )+a+1 .
By applying Ns0 we obtain the following equation in k[T ] if s = 0
ST =T +a (L0 )r0 + l0 R0 = g
and
L 0 r0 + l 0 R 0 = g or T L0 r0 + l0 R0 = g
if s > 0. Here l0 = Ns0 (l mod Dvs (L)+a+1 ), r0 = Ns0 (r mod Dvs (R)+a+1 ), L0 = Ns0 (L),
R0 = Ns0 (R) and g = Ns0 (f − L0 R0 mod Dvs (f )+a+1 ). Note that vs (R) is 0 if s = 0. The
requirement order(r) < order(R) means degree(r0 ) < degree(R0 ). To prove statement 1
we now have to show that l0 , r0 ∈ k[T ] exist and are uniquely determined. For this it
is sufficient to show that gcd(T L0 , R0 ) = 1 if s > 0 and gcd(ST =T +a (L0 ), R0 ) = 1 if
s = 0. First the case s > 0. R0 is the factor h of the Newton polynomial in the algorithm.
L0 R0 = Ns0 (f ) = T i Ns (f ) for some integer i. The set N of factors h of Ns (f ) is chosen in
such a way in the algorithm that gcd(h, Ns (f )/h) = 1. Also gcd(h, T ) = 1 because Ns (f )
does not contain a factor T by definition and h is a factor of Ns (f ). So gcd(T L0 , R0 ) = 1.
Now for the case s = 0: we have L0 R0 = Ns (f ) because vs (R) = 0 (see the multiplication
formula for N0 in Section 3.4). R0 is the factor h of Ns (f ) in the algorithm. We have to
show that gcd(ST =T +a (L0 ), R0 ) = 1. The set N containing these factors h was chosen in
such a way that this holds for all a ≥ 1.
To prove the second statement we distinguish two cases. Suppose s = 0: then the
Newton polynomial of R is irreducible. Hence R must be irreducible because a factoriza-
tion of R gives a factorization of the Newton polynomial. Now suppose s > 0: then the
Newton polynomial is of the form pi where p is irreducible and i is an integer. If i > 1
then it is not clear whether R is reducible or not. Suppose R can be factored R = R1 R2 .
Then the Newton polynomials of R1 and R2 are both factors of pi . So the gcd of these
Newton polynomials is not equal to 1. Coprime index 1 would mean that σa+1,s (R1 ) and
σa+1,s (R2 ) can be uniquely determined from σa,s (R1 ), σa,s (R2 ) and σa+1,s (f ). To deter-
mine σa+1,s (R1 ) and σa+1,s (R2 ) requires solving an equation l0 Ns (R1 ) + r0 Ns (R2 ) = g
in k[T ]. Such an equation has a unique solution if and only if the gcd of the Newton
polynomials Ns (R1 ) and Ns (R2 ) is 1. So a possible factorization R = R1 R2 cannot be a
coprime index 1 factorization, which proves statement 2.
Suppose f = LR is a monic factorization satisfying statement 2. Now we need to show
that the algorithm finds this factorization. R can have only one slope s, otherwise it could
be factored by the given algorithm (which contradicts the assumption that statement 2
holds). First consider the case s = 0. Then Ns (R) must be an irreducible polynomial,
otherwise R can be factored by the algorithm. So Ns (R) must be an element of the set M
in the algorithm. It cannot be an element of {g | g(T ) = h(T + i), h ∈ M, i ∈ N, i > 0}
because then gcd(ST =T +a (L0 ), R0 ) = 1 does not hold for all a ≥ 1 which was shown to be
a necessary and sufficient condition for having coprime index 1 if s = 0. So Ns (R) ∈ N .
This means that σ1,s (R) and hence also σ1,s (L) are the same as σ1,s (R1 ) and σ1,s (L1 ) for
a factorization L1 , R1 of f given by the algorithm. Because the coprime index is 1 this
factorization L1 , R1 is completely determined by σ1,s (R1 ), σ1,s (L1 ) and f . Hence these
two factorizations L1 R1 and LR are the same and so the third statement holds. In the
same way the case s > 0 is proven. 2
Remark 5.1. The given method can also be applied for factorization in the ring L[δ]
where L is a finite extension of k((x)), because:
• the method is not different for algebraic extensions of the constants k ⊂ l;

• ramifications over l((x)) can be handled using the map θa,n in Section 3.2;
• all finite field extensions of k((x)) are obtained as an algebraic extension of the
constants followed by a ramification, cf. Section 3.1.
Consider again the example f in Section 3.4 and let k = Q. The given algorithm
produces a right-hand factor R1 with slope 0, order 1 and Newton polynomial T , a right-
hand factor R2 with slope 1/2, order 2 and polynomial T − 1, a right-hand factor R3
with slope 1/2, order 2 and polynomial T − 2 and a right-hand factor R4 with slope 1,
order 4 and Newton polynomial (T 2 + T + 1)2 . Now R1 , R2 and R3 are irreducible
in Q((x))[δ] because their Newton polynomials are irreducible. But it is not yet clear
whether R4 is irreducible or not. The second example in section 1 remains unfactored as
well. Reducible operators f that remain unfactored by the given factorization algorithm
are of the following form: f has one slope s > 0 and Ns (f ) is a power > 1 of an irreducible
polynomial. The given algorithm will compute only a trivial factorization L = 1, R = f
for this case. If such an operator is reducible then a factorization must have coprime index
> 1. In Section 6 the notion of exponential parts will be introduced. Using exponential
parts a description of the irreducible elements of k((x))[δ] will be given.
If f has one slope s > 0, s ∈ N and the Newton polynomial is a power of a polynomial
of degree 1, then compute Scx−s (f ) where c is the root of the Newton polynomial (see
also case 4 of the algorithm in Section 5.1). Then apply the factorization algorithm to
Scx−s (f ) and find a factorization of f by applying S−cx−s to the factors of Scx−s (f ). For
all other cases (i.e. s 6∈ N or degree(Ns (f )) > 1) we apply the method in Section 7. The
factorization obtained that way lifts rather slowly, i.e. it costs much time to compute
more terms. The lifting will be speeded up using the lift method of Section 4 whenever
that is possible (when its output is not the message “failed”).
A differential operator can have infinitely many different factorizations. For example ∂ 2
which equals 1/x2 times δ 2 −δ has ax+b as solutions, where a and b are constants. Hence
it has ∂ − (ax + b)0 /(ax + b) = ∂ − a/(ax + b) as right-hand factors. Note that algorithm
coprime index 1 factorizations produces only a finite number of different factorizations. In
the semi-regular case (cf. Section 3.4) it computes only one unique factorization, although
like the example ∂ 2 shows other factorizations could exist as well.
14 M. van Hoeij
5.1. computing first-order factors over k((x))
An element r of some differential extension of k((x)) is by definition a Riccati solution

of f ∈ k((x))[δ] if δ − r is a right-hand factor of f . The reason this is called a Riccati
solution is that they are solutions of the so-called Riccati equation. This is a non-linear
differential equation. The Riccati equation of f ∈ k((x))[δ] can be found by computing
a right-hand division of f by δ − u, where u is an indeterminate. The remainder of this
right-hand division is the Riccati equation. It is a polynomial in u and the derivatives
of u. It vanishes precisely when we substitute for u an element r such that δ − r is a right-
hand factor of f . The Riccati solutions are of the form xy 0 /y where y is a solution of f . In
the usual definition the Riccati solutions are the logarithmic derivatives y 0 /y of solutions
of f . The definition in this paper differs by a factor x because we work with δ = x∂
instead of ∂. In this paper only Riccati solutions in k((x)) are considered. In general
there exist more Riccati solutions in larger differential fields. The implementation does
not determine the Riccati equation itself because this can be a large expression. Instead
we use factorization to find Riccati solutions. Computing first-order right-hand factors
of f is the same as computing Riccati solutions.
The following algorithm is similar to the Rational Newton algorithm (cf. Barkatou,
1988) which is a version of the Newton algorithm (cf. Tournier, 1987; Della Dora et
al., 1982) that computes formal solutions using a minimal algebraic extension of the
constants field k. A difference between the Rational Newton algorithm and the following
algorithm Riccati solution is that we use factorization of differential operators. So the
order of the differential operator decreases during the computation.
Algorithm Riccati solution.
Input. f ∈ k((x))[δ].
Output. A first-order right-hand factor in k((x))[δ].
1. If order(f ) = 1 then the problem is trivial.

2. If one of the following holds:
(a) f is regular singular and the N0 (f ) is reducible.
(b) The Newton polygon has more than one slope.
(c) f has one slope s > 0 and Ns (f ) is not a power ≥ 1 of an irreducible polynomial.
then compute a coprime index 1 factorization and apply recursion to the right-hand
factor.
3. If f has one slope s and the Newton polynomial Ns (f ) is of the form pe with p
irreducible, e ≥ 1 and p is of degree d > 1. Then extend k by one root r of p.
Now compute a right-hand factor of order order(f )/d with (T − r)e as Newton
polynomial using a coprime index 1 factorization as in the algorithm in Section 5.
This is a coprime index 1 factorization because the gcd of (T − r)e and pe /(T − r)e
(this is the Newton polynomial of the left hand factor) is 1. Now apply recursion
to the right-hand factor.
4. If f has one slope s > 0, s ∈ N and Ns (f ) is a power of a polynomial of degree 1,
then compute Scx−s (f ) where c is the root of Ns (f ). Use recursion (this recursion
is valid because the slopes of Scx−s (f ) are smaller than the slope of f ) to find a
first order factor of Scx−s (f ). Then apply S−cx−s .
5. If f has one slope s > 0, s 6∈ N and the Newton polynomial is a power of a
polynomial of degree 1, then write s = n/d with gcd(n, d) = 1, n > 0. Now we will
apply a ramification of index d. Instead of extending the field k((x)) we apply the
isomorphism θa,d : k((r))[δ] → k((x))[δ] of Section 3.2. First we need to compute
a suitable value a ∈ k. θa,d (x) = θa,d (rd /a) = xd /a. Write the Newton polynomial
of f as (T − c)e , where c ∈ k and e ∈ N. Then the Newton polynomial of θa,d (f )
equals a constant times (T d − dd can )e . Now choose a equal to cp , p ∈ Z, such that
dd can is a dth power of an element b ∈ k. This is done by choosing p such that pn+1
is a multiple of d. Then the Newton polynomial (T d − dd can )e equals (T d − bd )e
and can be factored as (T − b)e g e with gcd((T − b)e , g e ) = 1. Now use a coprime
index 1 factorization as in Section 5 with (T − b)e as Newton polynomial for the
right-hand factor. This provides a right-hand factor R of order e = order(f )/d. Now
−1
use recursion on R to find a first order factor and apply θa,d .
Note that there are two cases where a field extension of k((x)) is applied. One case
was an extension of k of degree d, and the other case was a ramification of index d. Both
these cases were extensions of k((x)) of degree d. In both cases the algorithm finds a
right-hand factor of order order(f )/d over this algebraic extension. In the three other
cases field extensions were not needed. We can conclude
Lemma 5.1. Every f ∈ k((x))[δ] has a Riccati solution which is algebraic over k((x)) of
degree ≤ order(f ).
6. Exponential Parts
A commutative invariant is a map φ from k((x))[δ] to some set such that φ(f g) = φ(gf )
for all f, g ∈ k((x))[δ]. An example is the Newton polygon, i.e. N (f g) = N (gf ) for
all non-zero f and g. However, there are more properties of differential operators that
remain invariant under changing the order of multiplication. We want a commutative
invariant which contains as much information as possible. Sommeling (1993) defines
normalized eigenvalues and characteristic classes for matrix differential operators. The
topic of this section is the analogue of normalized eigenvalues for differential operators in
k((x))[δ]. We will call these exponential parts. The exponential parts are useful for several
topics. They appear as an exponential integral in the formal solutions (this explains the
name exponential part). They describe precisely the algebraic extensions over k((x))
needed to find the formal solutions. The exponential parts are also used in our method
of factorization in the ring k(x)[∂] in van Hoeij (1996, chapter 3). For factorization in
k((x))[δ] the exponential parts will be used to describe the irreducible elements (cf.
Theorem 6.2).
Differential operators (in this paper that means elements of k((x))[δ] or k((x))[δ]) can
be viewed as a special case of matrix differential operators. So our definition of exponential
parts could be viewed as a special case of the definition of normalized eigenvalues in
Sommeling (1993). A reason for giving a different definition is that the tools for computing
with matrix differential operators are not the same as for differential operators. Important
tools for matrix differential operators are the splitting lemma and the Moser algorithm.
The tools we use for differential operators are the substitution map and the Newton
polynomial. That is why we want to have a definition of exponential parts expressed
in these tools. Because then the definition allows the computation of exponential parts
using a variant of the “algorithm Riccati solution”, namely the “algorithm semi-regular
16 M. van Hoeij
parts” in Section 8.4. A second reason for our approach is that it allows the definition of
semi-regular parts of differential operators.
Let L be a finite extension of k((x)). Since L ⊂ k((x1/n )) for some integer n we can
write every r ∈ L as r = e + t with e ∈ E and t ∈ x1/n k[[x1/n ]]. Now e is called the
principal part pp(r) of r ∈ L. Using the following lemma we can conclude e ∈ k((x))[r] ⊂
L.
Lemma 6.1. Let n ∈ Q and r ∈ k((x)) be equal to rn xn plus higher-order terms. Then
rn xn is an element of the field k((x))[r].
Proof. Write r = rn xn +rm xm plus higher-order terms, where m ∈ Q, m > n. We want

to prove that there exists an s ∈ k((x))[r] of the form rn xn plus terms higher than xm .
Then we can conclude rn xn ∈ k((x))[r] by repeating this argument and using the fact
that the field k((x))[r] is complete (cf. Bourbaki, 1953, Chapter I, Section 2, Theorem 2).
We can find this s as a Q-linear combination of r and x dx dr
.2
Definition 6.1. Let f ∈ k((x))[δ], e ∈ E and n = ram(e). Let P = N0 (Se (f )), the New-
ton polynomial corresponding to slope 0 in the Newton polygon of Se (f ) ∈ k((x1/n ))[δ].
Now µe (f ) is defined as the number of roots (counted with multiplicity) of P in n1 Z and
µe (f ) is defined as the number of roots (counted with multiplicity) of P in Q.
Recall that ram(e) denotes the ramification index of e. Note that we have only defined
the Newton polynomial for elements of k((x))[δ], P not for ramifications of k((x)). Define
N0 (f ) for f ∈ k((x1/n ))[δ] as follows. Write f = i xi/n fi with fi ∈ k[δ]. Then N0 (f )
is (written as a polynomial in δ instead of T ) defined as fi where i is minimal such that
fi 6= 0.
We define an equivalence ∼ on E as follows: e1 ∼ e2 if e1 − e2 ∈ n1 Z where n is the
ramification index of e1 . Note that the ramification indices of e1 and e2 are the same if
e1 − e2 ∈ Q. If e1 ∼ e2 then µe1 (f ) = µe2 (f ) for all f ∈ k((x))[δ] so we can define µe for
e ∈ E/∼. Similarly µe (f ) is defined for e ∈ E/Q.
Definition 6.2. The exponential parts of an operator f ∈ k((x))[δ] are the elements
e ∈ E/∼ for which µe (f ) > 0. The number µe (f ) is the multiplicity of e in f .
Lemma 6.2. Let f = LR where f , L and R are elements of k((x1/n ))[δ]. Let Nf be the
number of roots of N0 (f ) in n1 Z, counted with multiplicity. Similarly define NL and NR .
Then Nf = NL + NR .
The proof of this lemma is not difficult; we will skip it. Note that if n = 1 then Nf =
µ0 (f ).
Lemma 6.3. If f = LR where f , L and R are elements of k((x))[δ] and e in E or in

E/∼ then µe (f ) = µe (L) + µe (R).
If f = LR where f , L and R are elements of k((x))[δ] and e in E or in E/Q then
µe (f ) = µe (L) + µe (R).
nZ
1
Proof. If n is the ramification index of e, then µe (f ) is the number of roots in of
N0 (Se (f )). Now the first statement follows using the previous lemma and the fact that
Se (f ) = Se (L)Se (R). The proof for µ is similar. 2
Theorem 6.1. Let f be a non-zero element of k((x))[δ], then the sum of the multiplic-
ities of all exponential parts is:
X
µe (f ) = order(f ).
e∈E/∼
Let f be a non-zero element of k((x))[δ], then

X
µe (f ) = order(f ).
e∈E/Q
Proof. If order(f ) = 1 then both statements hold. If f is reducible then we can use
induction and Lemma 6.3 so then both statements hold. In k((x))[δ] every operator of
order > 1 is reducible (see also the algorithm in Section 5.1 which computes a first-order
right-hand factor in k((x))[δ]) so the second statement holds.
To prove the first statement we need to show that the sum of the multiplicities is the
same for µ over all e ∈ E/∼ and µ over all e ∈ E/Q. Suppose e is an element of E/Q. The
sum of µe (f ) taken over all e ∈ E/∼ such that e ≡ e mod Q is equal to µe (f ) because
they are both equal to the number of rational roots of the same Newton polynomial. So
we can see that the sum of the multiplicities µ is the same as sum of the multiplicities µ
by grouping together those exponential parts of f that are congruent modulo Q. 2
6.1. semi-regular part
An operator f ∈ k((x))[δ] is called semi-regular over k((x)) if f has only one exponen-
tial part which is equal to 0 ∈ E/∼. A semi-regular operator is a regular singular operator
with only integer roots of the Newton polynomial. In other words µ0 (f ) = order(f ). An
operator f ∈ k((x))[δ] = k((x))[∂] is regular (or: non-singular) if f can be written as
a product of an element of k((x)) and a monic element of k[[x]][∂]. A regular operator
is regular singular and the roots of the Newton polynomial are 0, 1, . . . , order(f ) − 1.
So a regular operator is semi-regular. We can generalize the notion of semi-regular for
algebraic extensions k((x)) ⊂ L.
Definition 6.3. f ∈ L[δ] is called semi-regular over L if it is regular singular and all
roots of N0 (f ) are integers divided by the ramification index of L.
For a ramification rn = ax an isomorphism θa,n : k((r))[δ] → k((x))[δ] was given in

Section 3.2. Now f ∈ k((r))[δ] is semi-regular over k((r)) if and only if θa,n (f ) ∈ k((x))[δ]
is semi-regular over k((x)).
Definition 6.4. Let f ∈ k((x))[δ]. Then the semi-regular part Re of f for e ∈ E is the
monic right-hand factor in k((x))[e, δ] of Se (f ) of order µe (f ) which is semi-regular over
k((x))[e].
Re can be computed by a coprime index 1 factorization of Se (f ) as in Section 5 using

slope s = 0. The Newton polynomial (called h in the algorithm) is the largest factor
18 M. van Hoeij
of N0 (Se (f )) for which all roots are integers divided by the ramification index. Since
such coprime index 1 factorizations for a given Newton polynomial are unique (see the
comments after algorithm coprime index 1 factorizations) it follows that Re is uniquely
defined. Note that if the ramification index n is > 1 then in fact our algorithm does
not compute with Se (f ) but with θa,n (Se (f )) for some constant a (cf. the remark in
Section 5.1). Then we have to compute the highest-order factor of θa,n (Se (f )) of which
the roots of the Newton polynomial are integers, instead of integers divided by n.
S−e (Re ) is a right-hand factor of f . Note that if e1 ∼ e2 then S−e1 (Re1 ) = S−e2 (Re2 ).
Hence the operators S−e1 (Re1 ), . . . , S−ep (Rep ) in the following lemma are up to a per-
mutation uniquely determined by f .
Lemma 6.4. Let f be an element of k((x))[δ]. Let e1 , . . . , ep ∈ E be a list of represen-

tatives of all exponential parts in E/∼ of f . Then
f = LCLM(S−e1 (Re1 ), . . . , S−ep (Rep )).
Remark 6.1. A similar statement (expressed in the terminology of D-modules) is given

in Malgrange (1979, Corollaire 4.3.1). There is, however, a subtle but important difference
namely that in our lemma the operators Ri are semi-regular instead of regular singular. A
different notion of exponential parts corresponds to this difference as well; in Malgrange
(1979, Corollaire 4.3.1) a notion appears which, in our terminology, can be viewed as
elements of E/k instead of our E/∼. One often distinguishes the two notions irregular
singular and regular singular. In this paper we propose to drop the notion of regular
singular as much as possible and only to make a distinction between semi-regular and
not semi-regular, and measure the “non-semi-regularity” using the exponential parts in
E/∼. The motivation for doing this is to generalize algorithms that work for regular
singular operators to the irregular singular case. In van Hoeij (1996) the benefits of this
approach are shown.
Proof. Let R = LCLM(S−e1 (Re1 ), . . . , S−ep (Rep )). Conjugation over k((x)) only per-
mutes S−e1 (Re1 ), . . . , S−ep (Rep ). Hence R is invariant under conjugation over k((x)) and
so R ∈ k((x))[δ]. S−ei (Rei ) is a right-hand factor of R, so Rei is a right-hand factor of
Sei (R). So N0 (Rei ) is a factor of N0 (Sei (R)), hence µei (R) ≥ degree(N0 (Rei )) = µei (f )
because all roots of N0 (Rei ) are integers divided by the ramification index. Then by
Theorem 6.1 we can conclude order(R) ≥ order(f ). R is a right-hand factor of f because
the S−ei (Rei ) are right-hand factors of f . Hence f = R. 2
This provides a method to compute a fundamental system of solutions of f . The

solutions of f = LCLM(S−e1 (Re1 ), . . . , S−ep (Rep )) are spanned by the solutions of
S−e1 (Re1 ), . . . , S−ep (Rep ). The solutions of S−e1 (Re1 ) are
R obtained by multiplying the
solutions of Re1 by Exp(e1 ) (recall that Exp(e1 ) = exp( ex1 dx), cf. Section 3.2). Con-
sequently, when all ei and Rei have been computed, then the problem of finding the
solutions of f is reduced to solving semi-regular differential operators, which is a much
easier problem (cf. Section 8.1).
Define Re for e ∈ E and f ∈ k((x))[δ] as the largest regular singular factor of Se (f )
for which all roots of the Newton polynomial are rational numbers. Now we can show in
the same way for f ∈ k((x))[δ] that
f = LCLM(S−e1 (Re1 ), . . . , S−eq (Req )) (6.1)
where e1 , . . . , eq ∈ E is a list of representatives for all e ∈ E/Q for which µe (f ) > 0.
6.2. irreducible elements of k((x))[δ]
If r ∈ k((x)) is a Riccati solution of f ∈ k((x))[δ] then the principal part e = pp(r) ∈

k((x))[r] modulo ∼ is an exponential part of f . Conversely, if µe (f ) > 0 then f has
a Riccati solution re ∈ k((x))[e] of which the principal part is e modulo ∼. Though
there may be infinitely many such Riccati solutions, we can compute one such re in a
canonical way. The algorithm in Section 5 provides (although infinitely many different
factorizations could exist) only one unique factorization of semi-regular operators (namely
the one that has coprime index 1). This way we can compute a unique right-hand factor
δ − re of S−e (Re ) by computing a first-order factor of Re and applying S−e . If e1 ∼ e2
then re1 = re2 . So re is defined for exponential parts e ∈ E/∼ of f .
Suppose e1 ∈ E is algebraic over k((x)) of degree d and suppose µe1 (f ) > 0. Suppose
e1 , . . . , ed ∈ k((x)) are the conjugates of e1 over k((x)). If L is a Galois extension of
k((x)) then conjugation over k((x)) is an automorphism of L[δ]. So µei (f ) = µej (f ) for
all i, j. We can find unique right-hand factors δ − rei ∈ k((x))[ei , δ] ⊂ k((x))[δ] of f as
just described. Then R = LCLM(δ − re1 , . . . , δ − red ) is a right-hand factor of f . Because
conjugation is an automorphism the rei are all conjugates of re1 over k((x)). So the set
{δ − re1 , . . . , δ − red } is invariant under conjugation which implies that R is invariant
under conjugation over k((x)). Hence R ∈ k((x))[δ]. In general
X
order(LCLM(f1 , . . . , fn )) ≤ order(fi )
i
because the order of an operator is equal to the dimension of the solution space, and
the solution space of LCLM(f1 , . . . , fn ) is spanned by the solutions of f1 , . . . , fn . So
order(R) ≤ d. Since µei (R) = µe1 (R) > 0 for all i = 1, . . . , d we can conclude by
Theorem 6.1 that order(R) ≥ d if all ei represent different exponential parts. For this
we must prove ei − ej 6∈ Q if i 6= j. Suppose ei − ej ∈ Q. We now have to prove
that ei = ej . The Galois group G of k((x))[e1 , . . . , ed ] over k((x)) acts transitively on
e1 , . . . , ed . Hence γ(ei ) = ej for some γ ∈ G. If γ(ei ) = ei + (ej − ei ) where (ej − ei ) ∈ Q
then γ #G (ei ) = ei + (#G)(ej − ei ). Here #G denotes the number of elements of G.
However, for any finite group G and element γ ∈ G the equation γ #G = 1 holds so
γ #G (ei ) = ei . Hence (#G)(ej − ei ) = 0 and ei = ej . Now we can conclude order(R) = d.
We have partly proven the following.
Theorem 6.2. f ∈ k((x))[δ] has an exponential part e which is algebraic over k((x))
of degree d if and only if f has an irreducible right-hand factor R ∈ k((x))[δ] of order d.
Note. In a different terminology (normalized eigenvalues, characteristic classes and D-

modules) this result is found in Sommeling (1993) as well.
Proof. Given an exponential part of degree d over k((x)) we have already shown how
to construct R as LCLM(δ − re1 , . . . , δ − red ). Now we must show that R is irreducible in
k((x))[δ]. Suppose R has a non-trivial right-hand factor R1 of order d1 < d. By induction
we can conclude that R1 has an exponential part e which is algebraic over k((x)) of
degree d1 . Lemma 6.3 shows that e is an exponential part of R. Then e, e1 , . . . , ed are
d + 1 different exponential parts of R contradicting Theorem 6.1. So R is irreducible.
20 M. van Hoeij
Now suppose f has an irreducible right-hand factor R of order d. The exponential

parts of R are exponential parts of f by Lemma 6.3. We will show that all exponential
parts of R are conjugated over k((x)) and algebraic of degree d over k((x)). Let e1 be
an exponential part of R algebraic of degree p over k((x)). So the conjugates e1 , . . . , ep
are exponential parts of R and by our construction we find an irreducible factor R1 of R
of order p. Since R is irreducible we have R1 = R and hence p = d. Now e1 , . . . , ed
are d different exponential parts of R. Because of Theorem 6.1 there cannot be more
exponential parts, so all exponential parts of R are conjugated with e1 . 2
7. Coprime Index > 1 Factorization

How can one compute an irreducible factor of a polynomial f ∈ Q[y]? A method is to
compute a root r and the minimum polynomial of r. This is not the usual factorization
method for the ring Q[y]. But for the ring k((x))[δ] this idea supplies a method for the
cases we have not yet treated. The role of the root is played by a Riccati solution. The
analogue of the minimum polynomial for a Riccati solution r is the least-common left
multiple of δ − r and its conjugates. A minimum polynomial is the product of y − r and
its conjugates. One does not need to compute the conjugates to determine this product.
The same holds for the least common left multiple. To see this write the LCLM as an
operator R with undetermined coefficients R = an δ n + · · · + a0 δ 0 . Now the statement
that δ − r is a right-hand factor of R translates into a linear equation in a0 , . . . , an . This
is an equation over k((x))[r]. We know that all conjugated equations (which we do not
compute) hold as well. Then this system of equations can be converted to a system over
k((x)). We show how this can be done in a slightly more general situation. Suppose α is
algebraic of degree d over a field K and we have an equation b0 α0 + · · · + bd−1 αd−1 = 0
(in our application K = k((x)), α = r and the bi are k((x))-linear expressions in ai ).
The system formed by this linear equation and all its conjugates is equivalent with b0 =
b1 = · · · = bd−1 = 0. The reason is that the transition matrix (which is a Vandermonde
matrix) between these two systems of linear equations is invertible.
This method for computing R is not very efficient for two reasons. The right-hand
factor R is computed by solving linear equations over k((x)) which is rather complicated.
The computation of these linear equations involves an algebraic extension over k((x)).
So we prefer to lift R with the algorithm in section 4 whenever possible.
Example 7.1.
2 2 9 1
f = δ 4 + 2δ 3 − δ + + ∈ k((x))[δ].
x 4x x2
√
The exponential parts are e1 = √1x + 2−1 in E/∼ and the conjugates e2 , e3 , e4 of e1 over
√
Q((x)). If −1 6∈ k then e1 is algebraic
√ of degree 4 over k((x)) and then f is irreducible
in k((x))[δ]. Now assume that −1 ∈ k. Then e is algebraic of degree 2 over k((x)) and
hence f has an irreducible right-hand factor R ∈ k((x))[δ] of order 2. To e1 corresponds
the following right-hand factor in k((x))
√ √ √
−1/2 −1 0 27 3 −1 1/2 1587 4141 −1 1
r =δ−x − x + − − x + + x + ···.
2 80 40 12800 12800
Write R = δ 2 + a1 δ + a0 where a0 , a1 ∈ k((x)) are to be determined. Dividing R by r
results in a remainder of the form a0 b0,0 + a1 b0,1 + b0,2 + x1/2 (a0 b1,0 + a1 b1,1 + b1,2 ) for
some bi,j ∈ k((x)). By equating this to zero, the following linear equations are obtained:
a0 bi,0 + a1 bi,1 + bi,2 = 0, i = 1, 2. Solving these equations over k((x)) gives
√
1 √ 573 3661 −1 1
R=δ + 2
− −1 x + − 0
+ x + · · · δ1
2 6400 6400
√ √
−1 2 −1 37 0 12291 −1 48663 1
+ −x + − − x + − + x + · · · δ0 .
5 40 64000 64000
It is not efficient to compute many coefficients of a0 , a1 in this way. It suffices to deter-
mine R in this way up to accuracy 2 (i.e. to determine the coefficient of x0 in a1 and the
coefficient of x−1 in a0 ). Then the higher terms can be computed more efficiently by the
lift algorithm in Section 4.
8. Formal Solutions of Differential Equations

8.1. solutions of semi-regular equations
Let f ∈ k((x))[δ] be a semi-regular operator of order n ≥ 1. Then we can apply

Section 5 to factor f = L(δ −r) where r is an element of Z+x·k[[x]]. Sr (f ) = Sr (L)δ. We
can recursively computeR a fundamental system of solutions a1 , . . . , an−1 ∈ k((x))[log(x)]
of Sr (L). Define si = axi dx for i = 1, . . . , n − 1 and sn = 1. Then s1 , . . . , sn is a
fundamental system of solutions of Sr (f ). These si are elements of k((x))[log(x)] because
ai /x ∈ k((x))[log(x)] and every element of k((x))[log(x)] has an anti-derivative in this
ring. By requiring that the coefficients of x0 log(x)0 in s1 , . . . , sn−1 are 0 the si are
uniquely defined. RTo obtain the solutions of f we multiply the solutions of Sr (f ) by
t = Exp(r) = exp( xr dx). This t ∈ k((x)) can be computed efficiently as follows. If r is
written as m ∈ Z plus an element of x · k[[x]] then t can be written as xm + tm+1 xm+1 +
tm+2 xm+2 + · · ·. The the fact that t is a solution of δ − r gives a linear equation for tm+1 ,
after solving it we find an equation for tm+2 , etc.
The same method can also be used for an element f of L[δ] which is semi-regular
over L, where L is an algebraic extension of k((x)), for the same reason as in the remark
in Section 5.1. This way a uniquely defined basis of solutions s1 , . . . , sn ∈ L[log(x)] can
be computed. By Theorem 8.1 in Section 8.3 (first apply the theorem to k((x))[δ], then
generalize using the remark in Section 5.1) it follows that f is semi-regular over L if and
only if f has a fundamental system of solutions in L[log(x)].
8.2. the canonical basis of solutions
Let e1 , . . . , er ∈ E be representatives for the exponential parts of f . Computing ei and

the corresponding semi-regular parts Rei can be done by the algorithm in Section 8.4.
Note that the algorithm only computes the ei up to conjugation over k((x)). This means
that the formal solutions will also be computed up to conjugation over k((x)), i.e. if a
number of solutions are conjugated then only one of them will be computed.
The semi-regular Rei ∈ k((x))[ei , δ] has a basis of solutions si,j ∈ k((x))[ei , log(x)]. So
according to Section 6.1 we get a basis of solutions of the form
y = Exp(ei )si,j where ei ∈ E and si,j ∈ k((x))[ei , log(x)] (8.1)
R ei
(recall that Exp(ei ) ∈ V stands for exp( x dx)). In the LCLM factorization in Lemma 6.4
22 M. van Hoeij
the S−ei (Rei ) are uniquely determined. Furthermore a unique basis of solutions for semi-
regular operators was defined in the previous section. As a consequence, the basis of
solutions obtained in this way is uniquely defined. We will call this basis the canonical
basis of solutions.
For a solution in the form (8.1) si,j is called the semi-regular part of (8.1) and ei is
called the exponential part of (8.1). The exponential part of (8.1) is an exponential part of
the operator as well. The semi-regular part si,j is a solution of the semi-regular part Rei .
Note that from a given y in the form (8.1), ei can be determined modulo ∼ (without
further restrictions on si,j one cannot determine ei ∈ E from y because when replacing
for example ei by ei − 1 and si,j by x · si,j in y we obtain an equivalent expression).
A few introductory comments on the next section. Every f ∈ k((x))[δ] is an element
of some L[δ] where L is a finite extension of k((x)). By a suitable transformation θa,d
as in the remark in Section 5.1 the problem of finding solutions of f can be reduced to
finding solutions of an operator θa,d (f ) ∈ l((x))[δ]. The solutions of f can be obtained
from the solutions of θa,d (f ). But the elements of the basis of solutions that we find for
f are not necessarily in the form (8.1) (in other words: are not necessarily an element of
some Ve ) but are element of some V e , definitions follow in the next section.
√ √
Example 8.1. δ − x/(2 + 2 x). Apply θ1,2 to obtain 12 δ − 12 x/(1 + x). A basis for
the solutions is 1 + x. This is of the √ form (8.1) with e = 0. Now apply an inverse
transformation to find the solution 1 + x of f . This is not of the form (8.1) but it is a
sum of two terms of the form (8.1), one with e = 0 L and one with e = 1/2. This example
shows that the direct sum decomposition V (f ) = Ve (f ) in theorem 8.1 in the next
section which holds for f ∈ k((x))[δ] need not hold for f ∈ k((x))[δ]. For f ∈ k((x))[δ] a
less precise statement is given in Theorem 8.1, corresponding to the less precise version µ
of exponential parts.
8.3. the solution space and exponential parts
Definition 8.1. Define for e ∈ E the set

V e = Exp(e) · k((x))[log(x)] ⊂ V
and

Ve = Exp(e) · (k · k((x))[e])[log(x)] ⊂ V e
If e1 ∼ e2 then Ve1 = Ve2 so Ve is also defined for e ∈ E/∼. Similarly V e is defined for
e ∈ E/Q. Define
\ \
Ve (f ) = Ve V (f ) and V e (f ) = V e V (f ).
Note that k · k((x))[e] = k · k((x1/n )) where n = ram(e). The reason for writing
k · k((x1/n )) instead of k((x1/n )) is that in general (namely if k 6= k) the field k((x1/n ))
is not a subfield of k((x)).
Theorem 8.1. For non-zero f ∈ k((x))[δ]

M
V (f ) = Ve (f ) and dim(Ve (f )) = µe (f )
e
where the sum is taken over all e ∈ E/∼. For non-zero f ∈ k((x))[δ]
M
V (f ) = V e (f ) and dim(V e (f )) = µe (f )
e
where the sum is taken over all e ∈ E/Q.
This theorem enables us to give an alternative definition of exponential parts and their
multiplicities µe (f ) in terms of the solution space of f .
Proof. Let f ∈ k((x))[δ]. Each element of the basis of solutions in the previous T section
is an element of some Ve where e is an exponential part of f . So the sum of the Ve V (f )
contains a complete basis of solutions of f . In this basis of solutions, µe (f ) elements are
in the form (8.1), i.e. µe (f ) elements are in Ve (f ). Hence
X
V (f ) = Ve (f ) and dim(Ve (f )) ≥ µe (f )
e
where the sum is taken over all exponential parts of f . It follows from P the following
Lemma
P 8.1 that this is a direct sum. Then order(f ) = dim(V (f )) = e dim(Ve (f )) ≥
µ
e e (f ) = order(f ) hence the ≥ must be an equality. The second statement follows in
the same way. 2
Lemma 8.1.
M M
V = Ve and V = Ve
e∈E/∼ e∈E/Q
L
Proof. Let n ∈ N. Then k((x)) = q Exp(q) · (k · k((x1/n ))) where the sum is taken
over all q ∈ Q with 0 ≤ q < 1/n. So for e ∈ E/Q
M
Ve = Ve
e
where the sum is taken over all e ∈ E/∼ such that e = e mod Q. This reduces the first
direct sum to the second one. Because of the relations Exp(e1 ) Exp(e2 ) = Exp(e1 + e2 )
every
P element of V can be written as a polynomial in the Exp(e) of degree 1. So V =
e V e . We will show that this is a direct sum which finishes the proof of this lemma.
PLet e1 , . . . , ed ∈ E be different modulo Q. Let si ∈ k((x))[log(x)] and s =

i Exp(ei )si = 0. To prove that the sum is direct we need to show that all si are
zero. Assume that not allPsi = 0 and that d > 1 is minimal with this property. Then
0
all si 6= 0. Now x dx ds
= i Exp(ei )(ei si + xsi ). Suppose the vectors (s1 , . . . , sd ) and
(e1 s1 + xs01 , . . . , ed sd + xs0d ) are linearly independent over k((x))(log(x)). Then we can
find a linear combination in which at least one (but not all) of the components van-
ishes. This contradicts the fact that d is minimal (multiply with a suitable element of
k((x))[log(x)] to eliminate log(x) from the denominator). So these two vectors must be
linearly dependent over k((x))(log(x)). It follows that
e1 s1 + xs01 e2 s2 + xs02
= ∈ k((x))(log(x))
s1 s2
so
e2 − e1 = xs01 /s1 − xs02 /s2 = xb0 /b
24 M. van Hoeij
where b = s1 /s2 ∈ k((x))(log(x)). But e2 − e1 ∈ E and e2 − e1 6∈ Q which contradicts

Lemma 8.2. 2
Lemma 8.2. Let b ∈ k((x1/n ))(log(x)). Suppose that the logarithmic derivative c = xb0 /b
is an element of k((x)). Then c ∈ n1 Z + x1/n · k[[x1/n ]].
Proof. Write b = p/q with p, q ∈ k((x1/n ))[log(x)]. Write p = pl log(x)l + · · · and

q = qm log(x)m + · · · where pl , qm ∈ k((x1/n )). The dots stands for an element of
k((x1/n ))[log(x)] of lower degree as a polynomial in log(x).
xb0 xp0 xq 0 xp0 log(x)l + · · · xqm
0
log(x)m + · · ·
c= = − = l −
b p q pl log(x)l + · · · qm log(x)m + · · ·
x(p0l qm − qm 0
pl ) log(x)l+m + · · ·
= ∈ k((x)).
pl qm log(x)l+m + · · ·
Then x(p0l qm − qm
0
pl )/(pl qm ) must be the same element c of k((x)). Write r = pl /qm ∈
k((x )). Then c = x(p0l qm − qm
1/n 0
pl )/(pl qm ) = xr0 /r ∈ n1 Z + x1/n · k[[x1/n ]]. 2
8.4. coprime index 1 LCLM factorization
Lemma 8.3. Let f1 , . . . , fd ∈ k((x))[δ], e ∈ E/∼ and f = LCLM(f1 , . . . , fd ). Then

X
max µe (fi ) ≤ µe (f ) ≤ µe (fi ).
i
i
In particular every exponential part e of f is an exponential part of at least one of the fi .
Proof. These inequalities

T P follow fromP the dimensions of Ve (f ) and Ve (fi ) in the following
equation: Ve (f ) = Ve ( i V (fTi )) = i Ve (fi ). The second equality holds because the
V (fi ) are direct sums of V (fi ) Ve1 taken over all e1 ∈ E/∼. 2
Lemma 8.4. Let f ∈ k((x))[δ]

P be monic and let f1 , . . . , fd ∈ k((x))[δ] be right-hand
factors of f . Suppose that i order(fi ) = order(f ) and that the fi have no exponential
parts in common. Then
• f = LCLM(f1 , . . . , fd )
• If e ∈ E/∼ and µe (f ) > 0 then there is precisely one fi such that Ve (f ) ⊂ V (fi ).
• For this e and fi the semi-regular part Re of f is the semi-regular part of fi as well.
Proof. Using the previous lemma, the fact that the fi have no exponential P part in com-
mon, and Theorem 6.1, we can conclude that order(LCLM(f1 , . . . , fd )) = order(fi ),
and this equals order(f ) by the assumption in this lemma. Since all fi , and hence this
LCLM, are right-hand factors of f the first statement follows. If e is an exponential part
of f then for precisely one i we have µe (fi ) > 0. Then µe (fi ) = µe (f ) because of the
previous lemma and because the µe of the other fj are zero. For the second statement
note that Ve (fi ) ⊂ Ve (f ), because fi is a right-hand factor of f . Since µe (fi ) = µe (f )
the dimensions are the same. Hence Ve (f ) = Ve (fi ) ⊂ V (fi ). The third statement follows
because V (S−e (Re )) = Ve (f ) ⊂ V (fi ), hence S−e (Re ) is a right-hand factor of fi and
so Re is a right-hand factor of Se (fi ). 2
Lemma 8.5. Let f, g ∈ k((x))[δ] and suppose gcd(Ns (f ), Ns (g)) = 1 holds for all s ∈ Q,
s > 0. Suppose gcd(Ns (f ), ST =T +n (Ns (g))) = 1 holds for s = 0 and all n ∈ Z. Then f
and g have no exponential parts in common.
Proof. For every exponential part e of f there exists a Riccati solution re of f such
that e is the principal part of re modulo ∼, cf. Section 6.2. Now the proof follows from
the next lemma. 2
Lemma 8.6. LetS r ∈ k((x)) be a Riccati solution of f ∈ k((x))[δ]. Suppose r, viewed as

an element of n k((x1/n )), can be written as rs xs plus higher-order terms, where s ∈ Q,
s ≤ 0 and rs 6= 0 if s < 0. Write s = n/d with n, d ∈ Z, gcd(n, d) = 1 and d > 0. Then
−s is a slope of f and rsd is a root of the Newton polynomial N−s (f ).
Proof. δ − r is a right-hand factor of f . If the ramification index of r is 1 the lemma

can easily be proved using the fact that the slopes of factors of f are slopes of f and the
Newton polynomials of right-hand factors of f are factors of the Newton polynomials of f ,
cf. Section 3.4. However, we have not defined the Newton polygon and Newton polynomial
over ramifications of k((x)). Choose d0 ∈ N such that θ1,d0 (δ − r) ∈ k((x))[δ]. Then d
must divide d0 . Now θ1,d0 (δ − r) is a right-hand factor of θ1,d0 (f ). The slope of θ1,d0 (δ − r)
is −sd0 so θ1,d0 (f ) has this slope as well. Hence f has a slope −s. The Newton polynomial
of θ1,d0 (δ −r) is d10 T −rs . If N−s (f ) = c(T p +ap−1 T p−1 +· · ·+a0 T 0 ) where c is a constant
then N−sd0 (θ1,d0 (f )) is a constant times T pd +d0 ap−1 T (p−1)d +· · ·+d0 a0 T 0 . So d10 T −rs
d pd
is a factor of this Newton polynomial hence rs is a root of T + ap−1 T p−1 + · · · + a0 T 0 .

d p
Now we can write algorithm LCLM factorization as follows. Take algorithm coprime
index 1 factorizations in Section 5. Replace the lines
if s = 0 then
M := {g1 , . . . , gr }
N := M \ {g | g(T ) = h(T + i), h ∈ M, i ∈ N, i > 0}
else
by the lines
if s = 0 then
M := {g1 , . . . , gr }
M 0 := M \ {g | g(T ) = h(T + i), h ∈ M, i ∈ N, i > 0}
M 00 := Q
{{n | ∃i∈Z gn (T + i) = h} | h ∈ M 0 }
N := { i∈h giei | h ∈ M 00 }
else
The resulting algorithm produces a number of factorizations. The sum of the or-
ders of the right-hand factors is equal to the order of f . The different right-hand fac-
tors f1 , . . . , fd have no exponential parts in common because of Lemma 8.5. Hence
f = LCLM(f1 , . . . , fd ). This variant on the algorithm in section 5 produces an LCLM
factorization, i.e. it produces a number of right-hand factors f1 , . . . , fd such that f =
LCLM(f1 , . . . , fd ). The orders of the fi need not be minimal because we only apply the
“easy” (i.e. coprime index 1) factorization method.
26 M. van Hoeij
Algorithm semi-regular parts.

Input. f ∈ k((x))[δ]
Output. Representatives e1 , . . . , ed ∈ E for all exponential parts up to conjugation over
k((x)) and the corresponding semi-regular parts Rei ∈ k((x))[ei , δ].
1. Same as case 1 in the algorithm Riccati solution. This is also a special case of 6
below after a suitable substitution map Se .
2. If algorithm LCLM factorization produces a non-trivial (i.e. d > 1) LCLM fac-
torization f = LCLM(f1 , . . . , fd ) then apply recursion to the right-hand factors
f1 , . . . , fd .
3. If the condition of case 3 in the algorithm Riccati solution holds, and furthermore
the slope of f is non-zero, then proceed as in case 3 in the algorithm Riccati solution;
apply recursion to the right-hand factor.
4. Same as case 4 in the algorithm Riccati solution, apply recursion to Scx−s (f ).
5. Same as case 5 in the algorithm Riccati solution, apply recursion on R.
6. If f has one slope s = 0 and the Newton polynomial has the following form Ns (f ) =
g·ST =T +i1 (g) · · · ST =T +in (g) where n ≥ 0 and ij are integers, and g is an irreducible
polynomial. Let r ∈ k be a root of g. Extend the field k with r (note that g can
have degree 1 in which case r ∈ k). Define h = T · (T + i1 ) · · · (T + in ). This is the
largest factor of N0 (Sr (f )) which has only integer roots. Now use a coprime index 1
factorization (cf. algorithm coprime index 1 factorizations in Section 5) to compute
a right-hand factor R of Sr (f ) that has Newton polynomial h.
The right-hand factors R that this algorithm produces in case 6 are the semi-regular
parts of f (actually such R is an image of a semi-regular part under certain maps θa,d
that were used in case 5). The corresponding exponential parts are obtained by keeping
track of the substitution maps Se and ramification maps θa,d that were performed. The
recursion in case 2 of the algorithm is valid because of Lemma 8.4.
In cases 3 and 5 of the algorithm a field extension over k((x)) is applied (also in case 6
if degree(g) > 1 but the argument is almost the same for this case). Suppose the degree
of this field extension is d. Then the algorithm computes a right-hand factor f1 of f and
uses recursion on this right-hand factor. Let f1 , . . . , fd ∈ L[δ] be the conjugates of f1 over
k((x)) where L is some finite extension of k((x)). Lemmas 8.5 and 8.4 were formulated
for k((x))[δ] instead of L[δ], but they are still applicable when using the less precise
notion of exponential parts µ. We must replace the condition “for all n ∈ Z” by “for all
n ∈ Q” in Lemma 8.5 in order for this lemma to hold for the case of µ instead of µ. So
our algorithm would produce all exponential parts and semi-regular parts if we would
use recursion on not only f1 but also on f2 , . . . , fd . However, this could introduce very
large algebraic field extensions (worst-case d factorial) which could make the algorithm
too slow to be useful. If we would use recursion on f2 , . . . , fd we will only find conjugates
of the exponential parts and semi-regular parts that are obtained from f1 . So there is no
need to do the recursion on f2 , . . . , fd because the result of that computation can also
be obtained as the conjugates (which are not computed, however) of the output of the
recursion on f1 .
Algorithm formal solutions.
Input. f ∈ k((x))[δ].
Output. A basis of solutions, up to conjugation over k((x)).
Step 1. This is the main step: apply algorithm semi-regular parts.
Step 2. Compute the solutions si,j of Rei as in Section 8.1.

Step 3. Return the set of Exp(ei )si,j .
Our method for computing formal solutions cannot avoid the use of field extensions
over k((x)) because these field extensions appear in the output. It does, however, delay the
use of algebraic extensions as long as possible. The use of algorithm LCLM factorization
reduces the problem of finding solutions of f to operators of smaller order. This way the
order of the operator is as small as possible at the moment that an algebraic extension
is introduced, and so the amount of computation in algebraic extensions is minimized.
Lazy evaluation is used to minimize the number of operations in the field of constants.
9. A Characterization of the Solution Spaces

The symbol log(x) is viewed as an element of a differential extension of k((x)) which
satisfies the equation y 0 = 1/x. The corresponding linear differential equation is y 00 +
1 0
x y = 0. We do not view log(x) as a function on an open subset of the complex plane,
but as a formal expression which is defined by the property that the derivative is 1/x.
From this viewpoint it is clear that the k((x))-homomorphism
Slog : k((x))[log(x)] → k((x))[log(x)]
defined by Slog (log(x)) = log(x) + 1 is a differential automorphism, because the deriva-
tive of log(x) + 1 is also 1/x, and hence all differential properties of log(x) + 1 and
log(x) are the same. This automorphism can be extended to the ring V by defining
Slog (Exp(e)) = Exp(e). If f ∈ V [δ] and y ∈ V is a solution of f then Slog (y) is a solu-
tion of Slog (f ). Note that the differential Galois group G of the Picard–Vessiot extension
k((x)) ⊂ k((x))(log(x)) contains more elements than just Slog . However, we will see that
it is sufficient to consider only Slog . This is explained from the fact that G is equal to
the Zariski closure of the group generated by Slog .
Let f ∈ k((x))[δ]. The questions of this section are: what are the possible right-hand
factors of f in k((x))[δ], or in k((x))[δ], what are the semi-regular and regular right-hand
factors. Every right-hand factor R corresponds to a subspace of solutions V (R) ⊂ V (f ).
But not every linear subspace W ⊂ V (f ) corresponds to a right-hand factor of f because
we do not look for right-hand factors in V [δ] but only in smaller rings like k((x))[δ]. So
the question now is the following. Given a finite dimensional k-vector space W ⊂ V ,
when is W the solution space of either:
1. a semi-regular operator in k((x))[δ];

2. a regular operator in k((x))[δ];
3. any operator in k((x))[δ];
4. any operator in k((x))[δ]?
Example 9.1. Let log(x) be a basis of W . Now there cannot be any f ∈ k((x))[δ] such
that W = V (f ). Because then Slog (log(x)) would be a solution of Slog (f ) = f . So f
has log(x) and Slog (log(x)) − log(x) = 1 as solutions. Hence the dimension of V (f ) is at
least 2.
Lemma 9.1. Let W be an n dimensional k-subspace of V . Then W = V (f ) for some

28 M. van Hoeij
semi-regular f ∈ k((x))[δ] if and only if W has a basis b1 , . . . , bn ∈ k((x))[log(x)] and

Slog (W ) = W .
Proof. Let f ∈ k((x))[δ] be semi-regular. Then it follows from Section 8.1 that V (f )
has a basis of solutions in k((x))[log(x)]. Furthermore Slog (V (f )) = V (Slog (f )) = V (f ).
Now suppose Slog (W ) = W and suppose b1 , . . . , bn ∈ k((x))[log(x)] is a basis of W as
a k-vector space. We want to construct a semi-regular operator f such that V (f ) = W .
Let b be an element of W of minimal degree d as a polynomial in log(x). Suppose d > 0.
Then Slog (b) − b ∈ W has degree d − 1 which contradicts the minimality of d. Hence
d = 0, so b is an element of k · k((x)). Then b ∈ l · k((x)) where l is some finite extension
of k. After multiplication by a constant we may assume that one of the coefficients of b
is 1. Then, by taking the trace over the field extension k ⊂ l, we may assume b ∈ k((x))
and b ∈ W (use here that W has a basis of elements in k((x))[log(x)], hence the trace
over k of an element b ∈ W is an element of W ). Now b 6= 0 because the trace of the
coefficient 1 is not 0. Because b ∈ V (f ) for the operator f that we want to construct
it follows that R = δ − xb0 /b must be a right-hand factor of f . This operator R is a
k-linear map from V to V . The kernel is the solution space of R and it has dimension 1.
Because the kernel is a subspace of W it follows that dim(R(W )) = n − 1. It is easy
to check that R(W ) satisfies the conditions of this lemma, hence by induction there is
a semi-regular operator L ∈ k((x))[δ] such that V (L) = R(W ). Now define f = LR.
This is a semi-regular operator in k((x))[δ] because L, R ∈ k((x))[δ] are semi-regular.
f (W ) = L(R(W )) = {0} and dim(W ) = order(f ) so V (f ) = W . 2
From the remark in Section 5.1 it follows that the lemma is also valid when k((x)) is
replaced by a finite extension L of k((x)).
Lemma 9.2. Let W be an n dimensional k-subspace of V . Then W = V (f ) for some

regular f ∈ k((x))[δ] if and only if W has a basis b1 , . . . , bn ∈ k[[x]] and all non-zero
elements of W have valuation < n.
Proof. If f ∈ k((x))[δ] is regular it is known by the Cauchy theorem that there exists
a basis b1 , . . . , bn ∈ k[[x]] of solutions such that bi is xi−1 modulo xn . It is easy to
compute these bi as follows. The equation f (bi ) = 0 (writing f as an element of k[[x]][∂]
is more convenient for this) gives a linear equation in the coefficient of xn in bi , a linear
equation for the coefficient of xn+1 , etc. From these equations the coefficients of bi can
be computed.
To prove the reverse statement let b1 , . . . , bn ∈ k[[x]] be a basis of W and suppose
that all non-zero elements of W have valuation (i.e. the smallest exponent of x which
has a non-zero coefficient) smaller than n. Then, after a basis transformation, we may
assume that bi is xi−1 modulo xn . Now define R1 ∈ k[[x]][∂] as R1 = ∂ − b01 /b1 . Define
for 1 ≤ d < n the operator Rd+1 ∈ k[[x]][∂] as follows: define yd+1 = Rd (bd+1 ). Note that
0
v(Ri (bd+1 )) = d − i for 1 ≤ i ≤ d. So v(yd+1 ) = 0 and hence ∂ − yd+1 /yd+1 ∈ k[[x]][∂].
0
Now define Rd+1 = (∂ − yd+1 /yd+1 )Rd . Now f = Rn is a monic element of k[[x]][∂],
hence regular, with V (f ) = W . 2
From the lemma we see that right-hand factors of regular operators need not be regular.
Suppose for example that 1, x, x2 is a basis of solutions of f . Then the right-hand factor
given by the basis of solutions 1, x2 is not regular. But the right-hand factor with the
basis 1, x + x2 is regular. An LCLM of regular operators is not necessarily regular either.

For certain purposes (not for all) semi-regular is a more convenient notion than regular
because factors, products, LCLM’s and symmetric products of semi-regular operators are
semi-regular.
LIf W T⊂ V is a solution space of a differential operator f ∈ k((x))[δ] then W =
e (W V e ) because of Theorem 8.1. Furthermore Slog (W ) must equal W because f is
invariant under Slog . This proves one part of the following lemma.
Lemma 9.3. Let W be a finite dimensional

L T k-subspace of V . Then W = V (f ) for some
f ∈ k((x))[δ] if and only if W = e (W V e ) = Slog (W ) where the sum is taken over
all e ∈ E/Q.
L T
Proof. Assume W 6= {0}Tis finite dimensional and W = e (W V e ) = Slog (W ). Let
e ∈ E such that We = W V e 6= {0}. Note that Slog (V e ) = V e hence We is invariant
under Slog . We has a basis of the form Exp(e) · si , i = 1, . . . , t where si ∈ k((x))[log(x)]
so si ∈ L[log(x)] for some finite extension L of k((x)). Using Lemma 9.1 it follows that
there exists an operator Re ∈ L[δ] which has si , i = 1, . . . , t as a basis of solutions. So
S−e (Re ) has Exp(e)si , i = 1, . . . , t as a basis of solutions and so S−e (Re ) must be a
right-hand factor of the operator f that T we want to construct. Choose a representative
e ∈ E for every e ∈ E/Q for which W V e 6= {0}. Construct the corresponding S−e (Re )
and define f as the LCLM of these S−e (Re ). Then V (f ) = W . 2
Lemma 9.4. Let W be a finite dimensional k-subspace of V . Then W = V (f ) for some

f ∈ k((x))[δ] if and only the conditions of the previous lemma hold, and furthermore W is
invariant under the action of the Galois group of the algebraic extension k((x)) ⊂ k((x)).
Proof. if τ is a k((x))-automorphism of k((x)) then τ can be extended to V by set-

ting τ (log(x)) = log(x) and τ (Exp(e)) = Exp(τ (e)). Now for any f ∈ k((x))[δ] we
have V (τ (f )) = τ (V (f )) because conjugation commutes with differentiation. So if f ∈
k((x))[δ] then V (f ) = τ (V (f )) which proves one part of the lemma. Now suppose W =
V (f ) for some monic f ∈ k((x))[δ] and suppose that W = τ (W ). Now order(f − τ (f )) <
order(f ) and W ⊂ V (f − τ (f )) so dim(V (f − τ (f ))) > order(f − τ (f )) and hence
f − τ (f ) must be 0. So if W is invariant under the Galois group of the algebraic exten-
sion k((x)) ⊂ k((x)) then f is invariant as well, hence f ∈ k((x))[δ]. 2
P
Every y ∈ V is a finite sum y = e be P with be ∈ Ve . Define W as the closure under
Galois actions and under Slog of the set e k · be . Now W satisfies the conditions of
the previous lemma, hence for every y ∈ V there is a g ∈ k((x))[δ] \ {0} such that
y ∈ V (g). From this it follows that for any non-zero f ∈ k((x))[δ] the map f : V → V
is surjective. This is seen as follows. If the kernel of g is not contained in the image of f
then the dimension of the kernel of gf would be smaller than the sum of the dimensions
of the kernels of g and f . In other words, order(gf ) < order(g) + order(f ) which is a
contradiction. Hence V (g) ⊂ f (V ) for every g and so f is surjective, f (V ) = V .
References
Barkatou, M.A. (1988). Rational Newton Algorithm for computing formal solutions of linear differential
—
equations, Proceedings of ISSAC’88, ACM Press.
30 M. van Hoeij
Björk, J.E., (1979). Rings of Differential Operators, North-Holland.

—
Bliss, G.A. (1966). Algebraic Functions, Dover.
—
Bourbaki, N. (1953). Espaces vectoriels topologiques, Paris.
—
Coddington, E., Levinson, N. (1955). Theory of ordinary differential equations, MacGraw-Hill.
—
Duval, A. (1983). Lemmes de Hensel et factorisation formelle pour les opérateurs aux différences, Funkcial
—
Ekvac, 26, p. 349–368.
Duval, D. (1989). Rational Puiseux expansions, Compos. Math. 70, No. 2, p. 119–154.
—
Della Dora, J., di Crescenzo, Cl., Tournier, E., (1982). An algorithm to obtain formal solutions of a
—
linear homogeneous differential equation at an irregular singular point, Proceedings of Symposium
EUROCAM’82 (Lecture Notes in Computer Science) 144, 273–280.
van den Essen, A., Levelt, A.H.M., (1992). An Explicit Description of all Simple k[[x]][∂]-Modules, Con-
—
temporary Mathematics, 130, p. 121–131.
Hendriks, P.A., van der Put, M. (1995). Galois action on solutions of a differential equation, J. Symbolic
—
Comput. 19, 559–576.
van Hoeij, M. (1996). Factorization of Linear Differential Operators, Ph.D. thesis, University of Nijmegen,
—
http://www-math.sci.kun.nl/math/compalg/diffop/.
Levelt, A.H.M. (1975). Jordan decomposition for a class of singular differential operators, Arkiv för Matem-
—
atik 13, 1-27, .
Levelt, A.H.M. (1990). Differential Galois theory and tensor products, Indagationes Mathematicae 1,
—
439–450.
Malgrange, B. (1979). Sur la réduction formelle des équations différentielles à singularités irrégulières,
—
Manuscript.
Ore, O. (1933). Theory of non-commutative polynomial rings, Ann. of Math. 34, 480–508.
—
van der Put, M. (1995). Singular complex differential equations: an introduction, Nieuw A. Wisk., (4),
—
13, 451–470.
Robba, P. (1980). Lemmes de Hensel pour les operateurs différentiels. Application a la reduction formelle
—
des equations différentielles, L’Enseign. Math. 26, 279–311.
Singer, M.F. (1996). Testing Reducibility of Linear Differential Operators: A Group Theoretic Perspective,
—
Appl. Algebra in Engrg. Comm. Comput. 7, 77–104.
Sommeling, R. (1993). Characteristic classes for irregular singularities, Ph.D. thesis, University of
—
Nijmegen.
Tournier, E. (1987). Solutions formelles d’équations différentielles, Thèse d’Etat, Faculté des Sciences de
—
Grenoble.
Originally received 12 December 1995

Accepted 19 March 1997

1 s2.0 S0747717197901103 Main

Uploaded by

Copyright:

Available Formats

1 s2.0 S0747717197901103 Main

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 s2.0 S0747717197901103 Main

Uploaded by

Copyright:

Available Formats

J.

Symbolic Computation (1997) 24, 1–30

Formal Solutions and Factorization of Differential

0747–7171/97/010001 + 30 $25.00/0 sy970110

• Compute a first-order right-hand factor ∂ − r of f where r ∈ k((x)). We use a

2. Valuations and the Coprime Index

Example 2.1. Suppose we want to factor f = x2 + x + 3 ∈ D = Q3 [x]. First we look at

3.1. the field k((x))

3.2. the ring k((x))[δ]

Define δ = x∂ ∈ k((x))[∂]. We have δx = xδ +x in k((x))[δ]. Since k((x))[∂] = k((x))[δ]

3.3. the Newton polygon

The Newton polygon of a monomial xi y j in the commutative polynomial ring k((x))[y]

3.4. the Newton polynomial

Let s ≥ 0 be a rational number. We have defined a valuation vs and a truncation σa

Example 3.1. Consider the following differential operator

Figure 1. Newton polygon of Example 3.1.

for s > 0 and for all L, R ∈ k((x))[δ].

4. The Lift Algorithm

5. Coprime Index 1 Factorizations

L0 is uniquely determined by requiring that L0 has no

1. L0 and R0 lift to a unique coprime index 1 factorization f = LR.

• the method is not different for algebraic extensions of the constants k ⊂ l;

5.1. computing first-order factors over k((x))

An element r of some differential extension of k((x)) is by definition a Riccati solution

1. If order(f ) = 1 then the problem is trivial.

Proof. Write r = rn xn +rm xm plus higher-order terms, where m ∈ Q, m > n. We want

Lemma 6.3. If f = LR where f , L and R are elements of k((x))[δ] and e in E or in

Let f be a non-zero element of k((x))[δ], then

6.1. semi-regular part

For a ramification rn = ax an isomorphism θa,n : k((r))[δ] → k((x))[δ] was given in

Re can be computed by a coprime index 1 factorization of Se (f ) as in Section 5 using

Lemma 6.4. Let f be an element of k((x))[δ]. Let e1 , . . . , ep ∈ E be a list of represen-

Remark 6.1. A similar statement (expressed in the terminology of D-modules) is given

This provides a method to compute a fundamental system of solutions of f . The

where e1 , . . . , eq ∈ E is a list of representatives for all e ∈ E/Q for which µe (f ) > 0.

6.2. irreducible elements of k((x))[δ]

If r ∈ k((x)) is a Riccati solution of f ∈ k((x))[δ] then the principal part e = pp(r) ∈

Note. In a different terminology (normalized eigenvalues, characteristic classes and D-

Now suppose f has an irreducible right-hand factor R of order d. The exponential

7. Coprime Index > 1 Factorization

8. Formal Solutions of Differential Equations

Let f ∈ k((x))[δ] be a semi-regular operator of order n ≥ 1. Then we can apply

8.2. the canonical basis of solutions

Let e1 , . . . , er ∈ E be representatives for the exponential parts of f . Computing ei and

8.3. the solution space and exponential parts

Definition 8.1. Define for e ∈ E the set

Theorem 8.1. For non-zero f ∈ k((x))[δ]

where the sum is taken over all e ∈ E/Q.

PLet e1 , . . . , ed ∈ E be different modulo Q. Let si ∈ k((x))[log(x)] and s =

where b = s1 /s2 ∈ k((x))(log(x)). But e2 − e1 ∈ E and e2 − e1 6∈ Q which contradicts

Proof. Write b = p/q with p, q ∈ k((x1/n ))[log(x)]. Write p = pl log(x)l + · · · and

8.4. coprime index 1 LCLM factorization

Lemma 8.3. Let f1 , . . . , fd ∈ k((x))[δ], e ∈ E/∼ and f = LCLM(f1 , . . . , fd ). Then

In particular every exponential part e of f is an exponential part of at least one of the fi .

Proof. These inequalities

Lemma 8.4. Let f ∈ k((x))[δ]

Lemma 8.6. LetS r ∈ k((x)) be a Riccati solution of f ∈ k((x))[δ]. Suppose r, viewed as

Proof. δ − r is a right-hand factor of f . If the ramification index of r is 1 the lemma

is a factor of this Newton polynomial hence rs is a root of T + ap−1 T p−1 + · · · + a0 T 0 .