Advances in Applied Mathematics 57 (2014) 44–59

Contents lists available at ScienceDirect

Advances in Applied Mathematics

www.elsevier.com/locate/yaama

Computation of the unipotent radical of the

differential Galois group for a parameterized
second-order linear differential equation
Carlos E. Arreche
Mathematics Department, The Graduate Center of the City University of
New York, 365 Fifth Avenue, New York, NY 10016, United States

a r t i c l e i n f o a b s t r a c t

Article history: We propose a new method to compute the unipotent radical

Received 10 February 2014 Ru (H) of the differential Galois group H associated to a
Accepted 11 March 2014 parameterized second-order homogeneous linear differential
Available online 23 April 2014 equation of the form
MSC:
12H20
34M15
∂2
34M03 Y − qY = 0,
20H20 ∂x2
13N10
33F10
37K20 where q ∈ F (x) is a rational function in x with coefficients
in a Π-field F of characteristic zero, and Π is a commuting
Keywords: set of parametric derivations. The procedure developed by
Parameterized Picard–Vessiot theory Dreyfus reduces the computation of Ru (H) to solving a
Parameterized linear differential creative telescoping problem, whose effective solution depends
equation on the assumption that the maximal reductive quotient
Linear differential algebraic group
H/Ru (H) is a Π-constant linear differential algebraic group.
Unipotent radical
Creative telescoping
When this condition is not satisfied, one can effectively
compute a new set of parametric derivations Π  such that
the associated differential Galois group H  has the property
that H  /Ru (H  ) is Π  -constant, and such that Ru (H) is
defined by the same differential equations as Ru (H  ). Thus the
computation of Ru (H) is reduced to the effective computation
of Ru (H  ). We expect that an elaboration of this method
will be successful in extending the applicability of some
recent algorithms developed by Minchenko, Ovchinnikov,

E-mail address: [email protected].

http://dx.doi.org/10.1016/j.aam.2014.03.001
0196-8858/© 2014 Elsevier Inc. All rights reserved.
 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59 45

and Singer to compute unipotent radicals for higher order

equations.
© 2014 Elsevier Inc. All rights reserved.

1. Introduction

Consider a linear differential equation of the form


n−1
δxn Y + ri δxi Y = 0, (1.1)
i=0

where ri ∈ K := F (x), the field of rational functions in x with coefficients in a Π-field F ,

δx denotes the derivative with respect to x, and Π := {∂1 , . . . , ∂m } is a set of commuting
derivations on F . Letting Δ := {δx } ∪ Π, we consider K as a Δ-field by setting ∂j x =
0 for each j. The parameterized Picard–Vessiot theory developed in [6] associates a
parameterized Picard–Vessiot (PPV) group to such an equation. In analogy with classical
Galois theory and the Picard–Vessiot theory developed by Kolchin [17], the PPV-group
measures the Π-algebraic relations amongst the solutions to (1.1). The differential Galois
groups that arise in this theory are linear differential algebraic groups, which are the
differential-algebraic analogues of linear algebraic groups: they are subgroups of GLn
that are defined by the vanishing of systems of polynomial differential equations in the
matrix entries. The study of linear differential algebraic groups was pioneered in [5]. The
parameterized Picard–Vessiot theory of [6] is a special case of an earlier generalization
of Kolchin’s differential Galois theory for strongly normal extensions, which is developed
in [22], as well as the differential Galois theory for linear difference-differential equations
with parameters presented in [14].
This article proposes a new method to compute the unipotent radical Ru (H) of the
PPV-group H corresponding to a second-order parameterized linear differential equation
of the form

δx2 Y − qY = 0, (1.2)

where q ∈ F (x) =: K, and F is a Π-field. In [11], Dreyfus applies results from [10]
to develop algorithms to compute H (see also [1] for a detailed discussion of Dreyfus’
results, in the setting of one parametric derivation). This algorithm is extended to arbi-
trary second-order linear differential equations over K in [3], by reinterpreting a classical
change-of-variables procedure in Galois-theoretic terms. In [11, §2.1], the computation
of Ru (H) is reduced to the computation of a finite set of linear differential operators in
F [Π], and a method is given for their computation that is known to halt under the as-
sumption that the maximal reductive quotient H/Ru (H) is Π-constant (cf. [24, Alg. 1]).
46 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59

We circumvent this obstruction by modifying the set of parametric derivations. De-

noting by D := F · Π the vector space over F spanned by Π, we compute a commuting
F -basis Π  for the Lie subspace L ⊆ D consisting of derivations ∂  ∈ D such that the
matrix entries of H/Ru (H) are ∂  -constant. We let the Π  -linear differential algebraic
group H  denote the PPV-group of (3.1), obtained by replacing Π with the new set of
parametric derivations Π  throughout the previous discussion. Theorem 3.2 states that
the maximal reductive quotient H  /Ru (H  ) is Π  -constant, and that Ru (H) is defined by
the same differential equations as Ru (H  ). Heuristically, any derivation ∂ ∈ D witnessing
the fact that H/Ru (H) is not Π-constant may be safely ignored during the computa-
tion of Ru (H), and this computation is thus reduced to the effective computation of
Ru (H  ) prescribed by Dreyfus’ procedure. This has the consequence, counterintuitive in
light of the fact that the known algorithms [7,11,24] for computing Ru (H) all require
that H/Ru (H) be Π-constant, that the computation of Ru (H) should be easier to per-
form when H/Ru (H) fails to be Π-constant, since in this case there are less derivations
appearing in the defining equations for Ru (H).
In Section 4, the main result (Theorem 3.2) is applied in the computation of the
PPV-group of a concrete parameterized linear differential equation (4.1) over K. An
application of Theorem 3.2 already appears in [2]. Let Gγ denote the PPV-group associ-
x
ated to the incomplete Gamma function, which is defined by γ(t, x) := 0 st−1 e−s ds for
Re(t) > 0, and extended analytically to a multivalued meromorphic function on C × C.
It satisfies:

∂2γ ∂γ
−p = 0, (1.3)
∂x2 ∂x

where p := 1−t−x
x . This is an example of a parameterized linear differential equation over
C(x, t), where δx := ∂x ∂
, and Π := { ∂t
∂
} is the parametric derivation. In [15] it was shown
that γ(t, x) does not satisfy any polynomial differential equations over C(x, t, ex , log(x))
∂
with respect to ∂t . This result was necessary in [6, Ex. 7.2] for the computation of Gγ ;
specifically, to conclude that Ru (Gγ ) = Ga , the additive group. In [2, Thm. 3.2], whose
proof follows the ideas of [14, Cor. 3.4.1], the differential transcendence of the solutions to
(1.3) was characterized in terms of two conditions on the coefficient p. Letting G (resp.,
G ) denote the PPV-group (resp., the non-parameterized PV-group) corresponding to
(1.3) for an arbitrary p ∈ C(x, t), the first condition is equivalent to the statement that
G/Ru (G) is not ∂t ∂
-constant, and the second condition says that Ru (G ) = Ga . When
the first condition is satisfied, one proves as in [14, Lem. 3.6(2)] that Ru (G) is either 0
or Ga . Since G is Zariski-dense in G by [6, Prop. 3.6(2)], it follows that Ru (G) = Ga
precisely when Ru (G ) = Ga . In other words, whenever G/Ru (G) is not ∂t ∂
-constant,
then Ru (G) is defined by the same (non-differential) equations as Ru (G ). Since G is


the “PPV-group” for (1.3) obtained by replacing Π = { ∂t ∂

} with Π  = ∅, Theorem 3.2
includes [2, Thm. 3.2] as a special case. In fact, the strategy for the proof of Theorem 3.2
is similar to the one followed in [2, Thm. 3.2]. But in order to carry it out in the more
general setting of several parametric derivations, it is necessary to strengthen some of the
 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59 47

technical results used in the proof of [2, Thm. 3.2]. This is done in Lemmas 3.5 and 3.7
(see also Remarks 3.6 and 3.6).
An algorithm to compute the unipotent radical Ru (G) of the PPV-group G cor-
responding to an nth-order parameterized linear differential equation over K is given
in [24], under the familiar assumption that G/Ru (G) is Π-constant. We expect that
an elaboration of the methods presented in this paper will be successful in extending
the procedure of [24] to compute Ru (G) in cases where G/Ru (G) is not necessarily
Π-constant. Algorithms to compute telescopers for rational functions, algebraic func-
tions, and hyper-exponential functions, are given in [8,9], and [4], respectively. The
notion of parallel telescoping investigated in [7] leads to algorithms [7, §5] to compute
PPV-groups in the setting of several principal derivations and one parametric derivation
(see Section 2.1). The creative telescoping problems solved by these algorithms lie at
the core of the algorithms presented in [11] (see also [1]) to compute the PPV-group
of (1.2). The hyper-exponential assumption, which is necessary in order to apply the
algorithms of [7,8] to compute Ru (H), coincides in this case with the requirement that
H/Ru (H) be Π-constant (see [7, §4] and Remark 3.1). We refer to [7, §1] for more details
and references concerning the connection between creative telescoping problems and the
computation of PPV-groups.

2. Preliminaries

See [19,26] for more details concerning the following definitions. A Δ-ring is a ring A
equipped with a finite set Δ := {δ1 , . . . , δm } of commuting derivations (that is, δi (ab) =
aδi (b) + δi (a)b and δi δj = δj δi for each a, b ∈ A and 1  i, j  m). We often omit the
parentheses, and write δa for δ(a). For Π ⊆ Δ, we denote the subring of Π-constants
of A by AΠ := {a ∈ K | δa = 0, δ ∈ Π}. When Π = {δ} is a singleton, we write Aδ
instead of AΠ . If A = K happens to be a field, we say that (K, Δ) is a Δ-field. Every
field is assumed to be of characteristic zero.
The ring of differential polynomials over K (in m differential indeterminates) is de-
noted by K{Y1 , . . . , Ym }Δ . Algebraically, it is the free K-algebra in the countably infinite
set of variables {θYi | 1  i  m, θ ∈ Θ}, where

  
Θ := δ1r1 · · · δnrn  ri ∈ Z0 for 1  i  n

is the free commutative monoid on the set Δ. The ring K{Y1 , . . . , Ym }Δ carries a natural
structure of Δ-ring, given by δi (θYj ) := (δi · θ)Yj .
We say p ∈ K{Y1 , . . . , Ym }Δ is a linear differential polynomial if it belongs to the
K-vector space spanned by the θYj , for θ ∈ Θ and 1  j  m. The K-vector space of
linear differential polynomials will be denoted by K{Y1 , . . . , Ym }1Δ .
The ring of linear differential operators K[Δ] is the K-span of Θ. Its (non-
commutative) ring structure is defined by composition of additive endomorphisms
48 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59

of K. The canonical identification of (left) K-vector spaces K[Δ]  K{Y }1Δ given by
 
θ aθ θ ↔ θ aθ θY will be assumed implicitly in what follows.
If M is a Δ-field and K ⊆ M is a subfield such that δ(K) ⊂ K for each δ ∈ Δ, we say
K is a Δ-subfield of M and M is a Δ-field extension of K. If y1 , . . . , yn ∈ M , we denote
the Δ-subfield of M generated over K by all the derivatives of the yi by

K y1 , . . . , yn
Δ ⊆ M.

A Δ-field K is Δ-closed if every system of polynomial differential equations defined

over K that admits a solution in some Δ-field extension of K already has a solution
in K. This notion is discussed at length in [16] (see also [6,29]).

2.1. Linear differential algebraic groups and parameterized Picard–Vessiot theory

We recall some standard facts from the parameterized Picard–Vessiot theory [6] (see
also [12,14,22]) and the theory of linear differential algebraic groups [5] (see also [20,25]).
Let F be a Π-field, where Π := {∂1 , . . . , ∂m }, and let K := F (x) be the field of rational
functions in x with coefficients in F . Let Δ := ({δx }∪Π), and consider K as a Δ-field by
setting δx x = 1, K δx = F , and ∂i x = 0 for each i. We will sometimes refer to δx as the
main derivation, and to Π as the set of parametric derivations. Consider the following
linear differential equation with respect to the main derivation, where ri ∈ K for each
0  i  n − 1:


n−1
δxn Y + ri δxi Y = 0. (2.1)
i=0

Definition 2.1. A Δ-field extension M ⊇ K is a parameterized Picard–Vessiot extension

(or PPV-extension) of K for (2.1) if:

(i) There exist n distinct, F -linearly independent elements y1 , . . . , yn ∈ M such that


δxn yj + i ri δxi yj = 0 for each 1  j  n.
(ii) M = K y1 , . . . , yn
Δ .
(iii) M δx = K δx .

We define the parameterized Picard–Vessiot group (or PPV-group) as the group of

Δ-automorphisms of M over K, and we denote it by GalΔ (M/K). The F -linear span of
all the yj is the solution space S.

If F is Π-closed, it is shown in [6] that a PPV-extension and PPV-group for (2.1) over
K exist and are unique up to K-Δ-isomorphism. Although this assumption allows for a
simpler exposition of the theory, several authors [12,30] have shown that, in many cases
of practical interest, the parameterized Picard–Vessiot theory can be developed without
 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59 49

assuming that F is Π-closed. In any case, we may always embed F in a Π-closed field
[16,29].
The action of GalΔ (M/K) is determined by its restriction to S, whence a choice of
F -basis for S determines an embedding GalΔ (M/K) → GLn (F ). It is shown in [6] that
this embedding identifies the PPV-group with a linear differential algebraic group.

Definition 2.2. (See [5,20].) Let F be a Π-closed field. We say that a subgroup G ⊆
GLn (F ) is a linear differential algebraic group if G is defined as a subset of GLn (F )
by the vanishing of a system of polynomial differential equations in the matrix entries,
with coefficients in F . We say that G is Π-constant if it is conjugate to a subgroup of
GLn (F Π ).

The study of linear differential algebraic groups was pioneered in [5], where the differ-
ential algebraic subgroups of the additive group Ga (F ) and the multiplicative group
Gm (F ) were classified in terms of finite sets of linear differential operators (see [5,
Prop. 11, Prop. 31 and its Corollary]). The differential algebraic subgroups of SL2 (F )
were classified in [28]. When the PPV-group GalΔ (M/K) is Π-constant, it is proved in [6,
Prop. 3.9(1)] that (2.1) is completely integrable [6, Defn. 3.8], and that M is a Picard–
Vessiot-extension (or PV-extension) of K for (2.1), in the non-parameterized sense of
[17]. The algorithms developed in [13] drastically reduce the number of conditions that
one has to check in order to decide whether (2.1) is completely integrable.
There is a parameterized Galois correspondence between the differential algebraic
subgroups Γ of GalΔ (M/K) and the intermediate Δ-fields K ⊆ L ⊆ M , given by
Γ → M Γ and L → GalΔ (M/L). Under this correspondence, an intermediate Δ-field
L is a PPV-extension of K (for some linear differential equation with respect to δx ) if
and only if GalΔ (M/L) is normal in GalΔ (M/K); in which case the homomorphism
GalΔ (M/K)  GalΔ (L/K), defined by σ → σ|L , is surjective with kernel GalΔ (M/L).
See [6, Thm. 3.5] and [12, §8.1] for more details.

3. Main result

We let K := F (x) denote the Δ-field defined in Section 2.1: F = K δx is Π-closed

field, δx x = 1, ∂x = 0 for each ∂ ∈ Π, and Δ := {δx } ∪ Π. Consider a second-order
parameterized linear differential equation

δx2 Y − qY = 0, (3.1)

where q ∈ K. Let M be a PPV-field of K for (3.1), and denote by H := GalΔ (M/K)

the corresponding PPV-group.
50 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59

3.1. Dreyfus’ algorithm

In [11] (see also [1]), Dreyfus develops a procedure to compute the PPV-group H corre-
sponding to (3.1). We begin with a brief summary of this procedure in the non-reductive
case (see [23,24]).
When H is not reductive [23, Defn. 2.2.6], it is proved in [21] that there exists an
F -basis {η, ξ} for the solution space S of (3.1) such that δx η = uη for some u ∈ K, and
δx ( ηξ ) = η −2 . The embedding H → SL2 (F ) is defined by

σ(η) = aσ η, σ(ξ) = a−1

σ ξ + bσ η,

and there exist differential algebraic subgroups A  Gm (F ) and B  Ga (F ) such that

the image of H → SL2 (F ) determined by this choice of basis is

  
 
aσ bσ  a b 
 σ ∈ H =  a ∈ A, b ∈ B . (3.2)
0 a−1
σ 0 a−1

If we let L denote the PPV-field corresponding to

δx Y − uY = 0, (3.3)

then the unipotent radical (see [23, Defn. 2.2.5]) B  Ru (H) coincides with the
PPV-group GalΔ (M/L), and the maximal reductive quotient A  H/Ru (H) is canoni-
cally isomorphic to GalΔ (L/K).
We refer to [11] for the computation of A (see also [4,8,21,23,27]). By the classification
result of [5, Prop. 11], B is described by a finite set of linear differential operators
p1 , . . . , ps ∈ F [Π], and it is shown in [11, §2] that they satisfy pi (η −2 ) ∈ δx (L) for
1  i  s. When A ⊆ Gm (F Π ), it is proved in [24] that M is of finite algebraic
transcendence degree over K, whence there exist bounds on the orders of the pi . But
if A  Gm (F Π ) and B = 0, then M is of infinite algebraic transcendence degree over
L (see [2,24], and cf. Corollary 3.3). We are not aware of a priori bounds on the orders
of the pi in this case (cf. [24, p. 13]), which raises the problem of deciding whether all
the pi have already been found, or whether it is still necessary to do more prolongations
(see [23–25]).
In the setting of one parametric derivation Π = {∂}, the problem of computing
Ru (H) when H/Ru (H) fails to be ∂-constant was solved completely in [1,2]. In this case,
it follows from [14, proof of Lem. 3.6(2)] that either Ru (H) = Ga (F ) or Ru (H) = 0, and
there are no other possibilities. In light of [24], this has the counterintuitive consequence
that the computation of Ru (H) is actually easier when H/Ru (H) is not ∂-constant (see
[2, Thm. 3.2]), since in this case the parametric derivation ∂ is barred from appearing
in the defining equations for Ru (H). Theorem 3.2 describes the generalization of this
phenomenon to the setting of several parametric derivations.
 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59 51

3.2. Computation of the unipotent radical

Let D := F · Π, the F -linear span of Π. Consider the set L of derivations ∂ ∈ D such

that every element of A  H/Ru (H) is constant with respect to ∂:

L := {∂ ∈ D | ∂a = 0, ∀a ∈ A}. (3.4)

A computation shows that L is a Lie subspace of D, i.e., an F -subspace that is closed

under the Lie bracket on derivations. By [5, Prop. 39] and [20, Prop. 0.6], there exists a
commuting F -basis Π  := {∂1 , . . . , ∂k } for L.
Now let Δ := {δx } ∪ Π  , and consider K as a Δ -field. Then, the Δ -field M  :=
K η, ξ
Δ is a PPV-extension of K for (3.1), and a Δ -subfield of M . We identify the
PPV-group H  := GalΔ (M  /K) with a Π  -subgroup of SL2 (F ) by means of the same
basis {η, ξ}, and define A and B  as in (3.2).

Remark 3.1. Let us briefly describe how to compute Π  . Observe that, for every ∂ ∈ D
and σ ∈ H,

∂η ∂η ∂aσ ∂η
σ = + and δx = ∂u. (3.5)
η η aσ η

Hence, the parameterized Galois correspondence implies that

  ∂η 
   
L= ∂∈D ∈ K = ∂ ∈ D  ∂u ∈ δx (K) . (3.6)
η

By writing a derivation ∂ = i ci ∂i with undetermined coefficients, and applying Her-
mite reduction to ∂u (see [4,8]), the condition ∂u ∈ δx (K) becomes an F -linear condition
on the coefficients ci . Thus the computation of a (possibly non-commuting) basis Π  for
L is reduced to linear algebra. The proof of [20, Prop. 0.6] gives an algebraic recipe
to produce a commuting basis Π  for L from the (possibly non-commuting) basis Π  .
This recipe was generalized and applied towards an algorithm to decide isomonodromy
in [13]. Corollary 3.3 below gives a simple and effective test to decide whether or not
Ru (H)  Ga (F ).

Theorem 3.2 (Main result). The reductive quotient H  /Ru (H  ) is Π  -constant, and the
linear differential operators {pi }si=1 ⊂ F [Π  ] defining Ru (H  ) are also defining operators
for Ru (H), under the natural inclusion F [Π  ] ⊆ F [Π].

Proof. That A is Π  -constant follows from Remark 3.1: since ∂η 

η ∈ K for each ∂ ∈ Π ,
 
we have that a = 0 for each a ∈ A . We will prove that B = B in a series of lemmas. By
∂a

Lemma 3.4, we have that B ⊆ B  . By Lemma 3.5, there is a finite set {pi }si=1 ⊂ F [Π  ]
such that B coincides with the set of those b ∈ Ga (F ) such that pi (b) = 0 for each
1  i  s. By Lemma 3.7, pi (b ) = 0 for each b ∈ B  and 1  i  s, whence B  ⊆ B. 2
52 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59

Corollary 3.3. (See [14, Lem. 3.6(2)] and [2, Prop. 4.4].) Suppose that Ru (H) = {0}.
Then,

Ru (H)  Ga (F ) ⇐⇒ L = {0}.

Proof. If L = {0}, then H  is the PV-group corresponding to (3.1), and B  is an algebraic

subgroup of Ga (F ). By Lemma 3.4, B  = {0}, because B ⊆ B  and B = {0}. Therefore,
B  = Ga (F ) = B, by Theorem 3.2.
Supposing instead that L = {0}, we have that Π  = ∅. If A is finite, the fact that
B  = Ga (F ) follows from [27, Prop. 3.3]. If A is infinite, since A is Π  -constant by

Theorem 3.2, the classification of [5, §IV.1] says that A = Gm (F Π ). That B  = Ga (F )
follows from [27, pp. 159–160] in this case (see also [2, proof of Prop. 4.4] and [24, Alg. 1]).
By Theorem 3.2, B = Ga (F ). 2

The following three lemmas were used in the proof of Theorem 3.2.

Lemma 3.4. The restriction homomorphism H → H  : σ → σ|M  induces an inclusion

Ru (H) → Ru (H  ).

Proof. The actions of H and H  on M and M  are completely determined by their

restrictions to the same solution space S = F ·η ⊕F ·ξ, whose definition is independent of
the chosen set of parametric derivations. Hence, the restriction homomorphism H → H 
is injective, and it is clear from the definitions that Ru (H) is then mapped (injectively)
into Ru (H  ). 2

The fact that B is the unipotent radical of (3.2), and not just any differential algebraic
subgroup of Ga (F ), allows us to produce a set of defining operators for B from F [Π  ] ⊆
F [Π], which sharpens the classification result of [5, Prop. 11] in this very particular
case. The following structural result, which was inspired by the results of [28] cited in its
proof, holds true for any linear differential algebraic group G of the form (3.2), whether
or not it happens to be a PPV-group over K.

Lemma 3.5. (See [14, Lem. 3.6(2)], [28, Thm. II.1.3 and Thm. II.1.4].) There exist
finitely many linear differential operators p1 , . . . , ps ∈ F [Π  ] ⊆ F [Π] such that
  
B = b ∈ F  pi (b) = 0, 1  i  s .

Proof. By [20, Prop. 0.7] the F -basis Π  for L can be extended to a commuting F -basis
Π̃ := {∂1 , . . . , ∂m
} for all of D. We denote by Θ̃ (resp., Θ ) the free commutative monoid
generated by Π̃ (resp., Π  ). Consider the orderly ranking on F {Y }Π̃ determined by the
lexicographic order on Θ̃ defined by setting δi  δj if j  i. In other words, to compare
two elements θ, θ in Θ̃, we first compare their total orders, and then the exponents of
∂1 , . . . , ∂m

, in that order.
 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59 53

By [28, Thm. II.1.3(b) and Thm. II.1.4], there is a characteristic set {p1 , . . . , ps } for
the defining ideal of B (with respect to this ranking) such that pi (aY ) = api (Y ) for
each a ∈ A and 1  i  s. Therefore, to show that {pi }si=1 ⊂ F [Π  ], it suffices to prove
that if p ∈ F [Π̃] does not belong to the image of F [Π  ] under the natural embedding
F [Π  ] ⊆ F [Π], then there exists an element a ∈ A such that p(aY ) − ap(Y ) = 0.
So suppose that p ∈ F [Π̃] and p ∈ / F [Π  ], and let cθ θY be the monomial in p of
highest rank such that cθ = 0 and θ contains a derivation
  
∂  ∈ Π̃\Π  = ∂k+1 
, . . . , ∂m .

Assume that ∂ is the derivation of lowest rank appearing effectively in θ, and let θ̃ denote
the element of Θ̃ obtained from θ by decreasing the order of ∂ by 1. Since θ (aY ) = aθ Y
for every a ∈ A and θ ∈ Θ , the leader of p(aY ) − ap(Y ) is cθ ∂ (a)θ̃Y whenever a ∈ A.
Since ∂ ∈
/ L, there exists a ∈ A such that ∂ (a) = 0, whence p(aY ) − ap(Y ) = 0. 2

Remark 3.6. When A is Π-constant, we may take Π  = Π, and Lemma 3.5 coincides with
[5, Prop. 11]. In case that Π = {∂} is a singleton and Π  = ∅, Lemma 3.5 is equivalent
to [14, Lem. 3.6(2)].

The previous result shows that B can be defined as a subset of Ga (F ) using derivations
from Π  only. The following lemma rules out the possibility that B could somehow be
defined by more Π  -differential equations than B  is.

Lemma 3.7. (See [6, Prop. 3.6(2)].) If p ∈ F [Π  ] is such that p(b) = 0 for every b ∈ B,
then p(b ) = 0 for every b ∈ B  . In other words, B ⊆ B  is Π  -dense.

Proof. Suppose that p ∈ F [Π  ] is such that p(b) = 0 for each b ∈ B. Then by [11, §2.1,
p. 7], we have p(η −2 ) ∈ δx (L). Moreover, since p ∈ F [Π  ],

p η −2 ∈ K η
Δ =: L ,

the fixed field of Ru (H  ). We will show that in fact p(η −2 ) ∈ δx (L ). Again by [11, §2.1,
p. 7], this will imply that p(b ) = 0 for each b ∈ B  . Assume that η is algebraically
transcendental over K, since otherwise A  μk , the group of kth roots of unity (see [5,
Prop. 31]), in which case H = H  and in particular B = B  .
By [6, Prop. 3.9] (cf. Remark 3.1), the fact that A is Π  -constant implies that

∂j η
vj := ∈K (3.7)
η

for each ∂j ∈ Π  , and therefore L = K(η) is algebraically isomorphic to the field of
rational functions in η with coefficients in K. It also follows from (3.7) that

−2vj = η 2 ∂j η −2 ∈ K. (3.8)
54 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59

Let us prove by induction that η 2 θ (η −2 ) ∈ K for each θ ∈ Θ , the free commutative

monoid on Π  . The base case is (3.8). Assuming that η 2 θ (η −2 ) =: vθ ∈ K, then
 
η 2 ∂j θ η −2 = η 2 ∂j vθ η −2 = ∂j vθ − 2vj vθ ∈ K

proves the induction step, and our claim. Hence, η 2 p(η −2 ) ∈ K for every p ∈ F [Π  ].
Since
 
∂1 η ∂m η
L := K η
Δ = K(η) ∂1 η, . . . , ∂m η
Δ = K(η) ,..., ,
η η Δ

it follows that L is algebraically generated as a field extension of L = K(η) by

 
∂j η 
θ  θ ∈ Θ, 1  j  m , (3.9)
η

where Θ is the free commutative monoid on Π. By [23, Cor. 5.1.2] and [24, Prop. 3.2] (see
also [24, §3.2.1]), if we consider L and K as δx -fields, then L is a (non-parameterized)
PV-extension of K, and the algebraic transcendence degree of L over K is finite. Hence,
we may choose a finite set β1 , . . . , βs of F -linearly independent generators for L over
L from the set (3.9). It follows from (3.5) that δx βi ∈ K for each 1  i  s. By the
Kolchin–Ostrowski theorem [18], the βi are then algebraically independent over L . We
define

N := K(β1 , . . . , βs ),

and observe that L = N (η). Since A is abelian, the subgroup GalΔ (L/N )  A is normal
and consequently N is a PPV-extension of K, by the parameterized Galois correspon-
dence [6, Thm. 3.5]. Again by [23, Cor. 5.1.2] and [24, Prop. 3.2], the δx -field N is a
(non-parameterized) PV-extension of the δx -field K. Since

δx η
=u∈K and δx βi ∈ K
η

for each 1  i  s, the Kolchin–Ostrowski theorem [18] implies that η is algebraically

transcendental over N . The corresponding PV-ring is

P := K[β1 , . . . , βs ] ⊂ N. (3.10)

Let f ∈ L such that δx (f ) = p(η −2 ). We claim that there exist elements g ∈ N and
c ∈ F such that f = gη −2 + c. To see this, let h ∈ K be such that

δx f = p η −2 = hη −2 ,
 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59 55

and write the partial fraction decomposition of f considered as a rational function in η,

where the coefficients ci , ek , and gj,k belong to N̄ , a fixed algebraic closure of N :
  gj,k
ci η i + = f. (3.11)
i
(η − ek )j
j,k

Let e0 = 0, and apply δx on both sides of (3.11) to obtain (cf. [1, Lem. 2.1]):

  δx gj,k − jugj,k jgj,k (δx ek − uek ) h

(δx ci + iuci )η i + + = δx f = 2 . (3.12)
i
(η − ek )j (η − ek ) j+1 η
j,k

Comparing the coefficients of η i in (3.12) shows that δx c0 = 0 and that δx ci = −iuci ,

which implies that ci = aη i for some a ∈ F , and therefore ci = 0 for i > 0 and c0 ∈ F .
Now fix k > 0, so that ek = 0, and let j > 0 be the smallest integer such that gj,k = 0.
Comparing the coefficients of (η − ek )−j in (3.12), we obtain that δx gj,k = jugj,k , which
implies that gj,k = aη j for some 0 = a ∈ F . This is impossible, and therefore there is
no such j, and only k = 0 appears in the sum (3.11). We obtain that δx gj,0 = jugj,0 for
j = 2 by comparing the coefficients of η −j in (3.12), which again implies that gj,0 = 0
whenever j = 2. Therefore,

f = g2,0 η −2 + c0 ,

where c0 ∈ F and g2,0 ∈ N̄ is algebraic over N . Since

g2,0 = η 2 (f − c0 ) ∈ L = N (η),

the fact that η is algebraically transcendental over N implies that g2,0 ∈ N .

Having shown that f = gη −2 + c for some g ∈ N and c ∈ F , let us now show that the
element g actually belongs to K, which implies that f ∈ L . Since

(δx g − 2ug)η −2 = δx f = hη −2

for some h ∈ K, it follows that

δx g − 2ug = h. (3.13)

We begin by showing that g ∈ K[β1 , . . . , βS ] must be a polynomial expression in


the βi . Indeed, it follows from (3.13) that the K-vector space j K · δxj g ⊂ N is finite-

dimensional over K. By [26, Cor. 1.38], the finite-dimensionality of j K · δxj g over K
is a necessary and sufficient condition for g ∈ N to belong to the PV-ring P defined in
(3.10).
To show that g ∈ K, we proceed by contradiction. Suppose that rI β I is a monomial
in g, considered as a polynomial in the βi , with 0 = |I| maximal and 0 = rI ∈ K. Since
56 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59

the coefficient of β I in the right-hand side of (3.13) is 0, we see that δx rI = 2urI , which
implies that rI = aη 2 for some 0 = a ∈ F , a contradiction. Hence, no such monomial
rI β I appears in g, which means that g ∈ K and gη −2 + c = f ∈ L , concluding the proof
of the lemma. 2

Remark 3.8. In case Π  = ∅, or in other words when the Lie subspace L defined in (3.4)
is {0}, then H  is the (non-parameterized) PV-group for (3.1), and Lemma 3.7 reduces
to a special case of [6, Prop. 3.6(2)].

4. An example

We let K = F (x) denote the Δ-field defined in the previous section, where Π :=
{∂1 , ∂2 }, ∂j := ∂t∂j for j = 1, 2, and F denotes a Π-closed field containing Q(t1 , t2 )
(see [16,29]). In this section, we will apply Theorem 3.2 to compute the PPV-group H
corresponding to the parameterized linear differential equation

x2 + (2 − 2t1 t2 )x + t21 t22 − 3t1 t2 + 2

δx2 Y − Y = 0. (4.1)
x2

The Riccati equation

x2 + (2 − 2t1 t2 )x + t21 t22 − 3t1 t2 + 2

δx u + u2 = =: q
x2
admits the solution

t1 t2 − 1 − x
u := .
x

Therefore, by [21] there is a basis {η, ξ} for the solution space of (4.1) such that δx η = uη
and δx ( ηξ ) = η −2 , and by [11, §2.1] there exist differential algebraic subgroups A 
Gm (F ) and B  Ga (F ) such that H is given by (3.2). Since ∂1 u = tx2 and ∂2 u = tx1 , we
have that

∂12 u = 0 = ∂22 u and t1 ∂1 u − t2 ∂2 u = 0. (4.2)

Therefore, the F [Π]-span of {∂1 u, ∂2 u} and the F -span of {∂1 u, ∂2 u, ∂1 ∂2 u} coincide

modulo δx (K). By [11, §2.1],
  ∂ a


 1 ∂2 a ∂1 a ∂2 a
A= a ∈ Gm (F )  t1 = t2 ; ∂1 = 0 = ∂2 .
a a a a

Since ∂1 u ∈ / δx (K), the Lie subspace of derivations L ⊂ F · Π defined in (3.4), or

equivalently in (3.6), has dimension at most 1 over F . Hence, by (4.2) L coincides with
F · (t1 ∂1 − t2 ∂2 ), the F -span of ∂1 := t1 ∂1 − t2 ∂2 , and we may take Π  := {∂1 }.
 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59 57

By Theorem 3.2, to compute the unipotent radical Ru (H) = B it suffices to compute

Ru (H  ) =: B  , where H  denotes the PPV-group of (4.1) relative to the new set of
parametric derivations Π  = {∂1 }. It follows from (4.2) that the system

δx Y − uY = 0
(4.3)
∂1 Y = 0

is isomonodromic [13] (or completely integrable, in the terminology of [6, Defn. 3.8]).
Therefore, by [6, Prop. 3.9] L = K(η) is a (non-parameterized) PV-extension of K for
(4.3), and (cf. Theorem 3.2)
   
GalΔ L /K  H  /Ru H   A = Gm F ∂1 .

By [2, Lem. 4.3] and [6, Prop. 2.6(2)], to see that B  = 0, it suffices to show that there
is no f ∈ K such that δx f + 2uf = 1. We prove this by contradiction, along the lines of
[2, proof of Cor. 3.3]. Assume that f ∈ K satisfies

δx f + 2uf = 1. (4.4)

First, note that f cannot be δx -constant, whence it must a have a pole somewhere in
P1 (F ). But f cannot have a pole outside of {0, ∞}, for otherwise the left-hand side of
(4.4) would have a pole. If f had a pole at 0, the residue of 2u at 0 would have to be an
integer, which is clearly false. Therefore, f can only have a pole at ∞, i.e., f must be a
polynomial in x. Moreover, f must be divisible by x, because otherwise the left-hand side
of (4.4) would have a pole at 0. But then the degree of the polynomial on the left-hand
side of (4.4) is equal to the degree of f , which is at least 1, since f is not constant. This
contradiction concludes the proof that there is no solution in K for (4.4), and therefore
that B  = 0. Since

ξ ξ
δx ∂1 = ∂1 η −2 = 0 =⇒ ∂1 ∈ F = δx (F · x) ⊂ δx (K),
η η

it follows from [11, §2.1, p. 7] that B  = Ga (F ∂1 ). Therefore, by Theorem 3.2,

 
a b  a, b ∈ F ; a = 0; t1 ∂a1 a = t2 ∂a2 a ;
H  .
0 a−1 ∂1 ( ∂a1 a ) = 0 = ∂2 ( ∂a2 a ); t1 ∂1 b = t2 ∂2 b

Acknowledgments

This work owes very much to Alexey Ovchinnikov’s essential suggestions and cor-
rections, as well as his patience, guidance and advice over the course of this project.
Richard Churchill’s generosity with his time and ideas has helped me better understand
numerous points in Picard–Vessiot theory and Kovacic’s algorithm. Phyllis Cassidy and
58 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59

William Sit have been invaluable guides in almost every aspect of differential algebra
with which I have struggled. Also indispensable were many helpful discussions about
this work in particular, and so much else, with: the members of the Kolchin Seminar
in New York, Raymond Hoobler, Michael Singer, Sergey Gorchinskiy, William Keigher,
Jacques-Arthur Weil, Thomas Dreyfus, Michael Wibmer, James Freitag, and Omar de
León Sánchez. Manuel Kauers and the referee did excellent editorial work on behalf of
the journal. I sincerely thank all of them for all their help and for their kind interest in
my work.
This material is based upon work supported by the National Science Foundation
(NSF) Graduate Research Fellowship Program under grant 40017-04-06 and by NSF
grant CCF-0952591.

References

[1] C.E. Arreche, Computing the differential Galois group of a one-parameter family of second order
linear differential equations, preprint, arXiv:1208.2226 (2012).
[2] C.E. Arreche, A Galois-theoretic proof of the differential transcendence of the incomplete Gamma
function, J. Algebra 389 (2013) 119–127, http://dx.doi.org/10.1016/j.jalgebra.2013.04.037.
[3] C.E. Arreche, Computing the differential Galois group of a parameterized second-order linear dif-
ferential equation, To appear in Proceedings of ISSAC, 2014, arXiv:1401.5127.
[4] A. Bostan, S. Chen, F. Chyzak, Z. Li, G. Xin, Hermite Reduction and Creative Telescoping for
Hyperexponential Functions, Proceedings of ISSAC, 2013, pp. 77–84.
[5] P.J. Cassidy, Differential algebraic groups, Amer. J. Math. 94 (1972) 891–954.
[6] P.J. Cassidy, M.F. Singer, Galois theory of parameterized differential equations and linear differential
algebraic groups, in: D. Bertrand, B. Enriquez, C. Mitschi, C. Sabbah, R. Schaefke (Eds.), IRMA
Lect. Math. Theor. Phys., vol. 9, Eur. Math. Soc. Publishing House, Zürich, 2007, pp. 113–155,
http://dx.doi.org/10.4171/020-1/7.
[7] S. Chen, R. Feng, Z. Li, M.F. Singer, Parallel telescoping and parameterized Picard–Vessiot theory,
To appear in Proceedings of ISSAC, 2014, arXiv:1401.4666.
[8] S. Chen, M. Kauers, M.F. Singer, Telescopers for Rational and Algebraic Functions via Residues,
Proceedings of ISSAC, 2012, pp. 130–137.
[9] S. Chen, M.F. Singer, Residues and telescopers for bivariate rational functions, Adv. in Appl. Math.
49 (2012) 111–133, http://dx.doi.org/10.1016/j.aam.2012.04.003.
[10] T. Dreyfus, A density theorem for parameterized differential Galois theory, Pacific J. Math. (2014),
in press arXiv:1203.2904.
[11] T. Dreyfus, Computing the Galois group of some parameterized linear differential equa-
tion of order two, Proc. Amer. Math. Soc. 142 (2014) 1194–1207, http://dx.doi.org/10.1090/
S0002-9939-2014-11826-0.
[12] H. Gillet, S. Gorchinskiy, A. Ovchinnikov, Parameterized Picard–Vessiot extensions and Atiyah
extensions, Adv. Math. 238 (2013) 322–411.
[13] S. Gorchinskiy, A. Ovchinnikov, Isomonodromic differential equations and differential Tannakian
categories, J. Math. Pures Appl. (2014), http://dx.doi.org/10.1016/j.matpur.2013.11.001, in press.
[14] C. Hardouin, M.F. Singer, Differential Galois theory of linear difference equations, Math. Ann. 342
(2008) 333–377, http://dx.doi.org/10.1007/s00208-008-0238-z.
[15] J. Johnson, G. Reinhart, L. Rubel, Some counterexamples to separation of variables, J. Differential
Equations 121 (1995) 42–66.
[16] E. Kolchin, Constrained extensions of differential fields, Adv. Math. 12 (1974) 141–170,
http://dx.doi.org/10.1016/S0001-8708(74)80001-0.
[17] E.R. Kolchin, Algebraic matric groups and the Picard–Vessiot theory of homogeneous linear ordinary
differential equations, Ann. of Math. 49 (1948) 1–42.
[18] E.R. Kolchin, Algebraic groups and algebraic dependence, Amer. J. Math. 90 (1968) 1151–1164.
[19] E.R. Kolchin, Differential Algebra and Algebraic Groups, Academic Press, New York, 1976.
[20] E.R. Kolchin, Differential Algebraic Groups, Pure Appl. Math., vol. 114, Academic Press, Orlando,
FL, 1985.
 C.E. Arreche / Advances in Applied Mathematics 57 (2014) 44–59 59

[21] J.J. Kovacic, An algorithm for solving second order linear homogeneous differential equations,
J. Symbolic Comput. 2 (1986) 3–43, http://dx.doi.org/10.1016/S0747-7171(86)80010-4.
[22] P. Landesman, Generalized differential Galois theory, Trans. Amer. Math. Soc. 360 (2008)
4441–4495, http://dx.doi.org/10.1090/S0002-9947-08-04586-8.
[23] A. Minchenko, A. Ovchinnikov, M.F. Singer, Reductive linear differential algebraic groups and
the Galois groups of parameterized linear differential equations, Int. Math. Res. Not. (2014),
http://dx.doi.org/10.1093/imrn/rnt344, in press.
[24] A. Minchenko, A. Ovchinnikov, M.F. Singer, Unipotent differential algebraic groups as param-
eterized differential Galois groups, J. Inst. Math. Jussieu (2014), http://dx.doi.org/10.1017/
S1474748013000200, in press.
[25] A. Ovchinnikov, Differential Tannakian categories, J. Algebra 321 (10) (2009) 3043–3062,
http://dx.doi.org/10.1016/j.jalgebra.2009.008.
[26] M. van der Put, M.F. Singer, Galois Theory of Linear Differential Equations, Grundlehren Math.
Wiss., Springer-Verlag, New York, 2003.
[27] M.F. Singer, Linear algebraic groups as parameterized Picard–Vessiot Galois groups, J. Algebra 373
(2013) 151–161, http://dx.doi.org/10.1016/j.jalgebra.2012.09.037.
[28] W.Y. Sit, Differential algebraic subgroups of SL(2) and strong normality in simple extensions, Amer.
J. Math. 97 (1975) 627–698.
[29] D. Trushin, Splitting fields and general differential Galois theory, Sb. Math. 201 (2010) 1323–1353,
http://dx.doi.org/10.1070/SM2010v201n09ABEH004114.
[30] M. Wibmer, Existence of ∂-parameterized Picard–Vessiot extensions over fields with algebraically
closed constants, J. Algebra 361 (2012) 163–171, http://dx.doi.org/10.1016/j.jalgebra.2012.03.035.