SYMBOLIC INTEGRATION

TUTORIAL

Manuel Bronstein
INRIA Sophia Antipolis
[email protected]

ISSAC98, Rostock (August 1998)

and Differential Algebra Workshop, Rutgers
November 2000

Contents
1 Rational Functions
1.1 The full partial-fraction algorithm . . . . . . . . . . . . . . . . .
1.2 The Hermite reduction . . . . . . . . . . . . . . . . . . . . . . . .
1.3 The RothsteinTrager and LazardRiobooTrager algorithms . .
2 Algebraic Functions
2.1 The Hermite reduction . .
2.2 Simple radical extensions
2.3 Liouvilles Theorem . . .
2.4 The integral part . . . . .
2.5 The logarithmic part . . .

5
6
7
8

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

9
10
14
15
16
17

3 Elementary Functions
3.1 Differential algebra . . . . . . . . . .
3.2 The Hermite reduction . . . . . . . .
3.3 The polynomial reduction . . . . . .
3.4 The residue criterion . . . . . . . . .
3.5 The transcendental logarithmic case
3.6 The transcendental exponential case
3.7 The transcendental tangent case . .
3.8 The algebraic logarithmic case . . .
3.9 The algebraic exponential case . . .

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

20
21
22
23
24
26
27
28
29
31

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

An elementary function of a variable x is a function that can be obtained

from the rational functions in x by repeatedly adjoining a finite number ofnested
logarithms, exponentials, and algebraic numbers or functions. Since 1 is
elementary, the trigonometric functions and their inverses are also elementary
(when they are rewritten using complex exponentials and logarithms) as well as
all the usual functions of calculus. For example,
p
sin(x + tan(x3 x3 x + 1))
(1)
is elementary when rewritten as

1 tx1
(e
ex 1t )
2

where

1 e2 1(x x x+1)

.
t=
1 + e2 1(x3 x3 x+1)

This tutorial describes recent algorithmic solutions to the problem of integration

in finite terms: to decide in a finite number of steps whether a given elementary
function has an elementary indefinite integral, and to compute it explicitly if it
exists. While this problem was studied extensively by Abel and Liouville during the last century, the difficulties posed by algebraic functions caused Hardy
(1916) to state that there is reason to suppose that no such method can be
given. This conjecture was eventually disproved by Risch (1970), who described an algorithm for this problem in a series of reports [12, 13, 14, 15]. In
the past 30 years, this procedure has been repeatedly improved, extended and
refined, yielding practical algorithms that are now becoming standard and are
implemented in most of the major computer algebra systems. In this tutorial, we
outline the above algorithms for various classes of elementary functions, starting
with rational functions and progressively increasing the class of functions up to
general elementary functions. Proofs of correctness of the algorithms presented
here can be found in several of the references, and are generally too long and
too detailed to be described in this tutorial.
Notations: we write x for the variable of integration, and 0 for the derivation
d/dx. Z, Q, R and C denote respectively the integers, rational, real and complex
numbers. All fields are commutative and, except when mentioned explicitly
otherwise, have characteristic 0. If K is a field, then K denotes its algebraic
closure. For a polynomial p, pp(p) denotes the primitive part of p, i.e.p divided
by the gcd of its coefficients.

Rational Functions

By a rational function, we mean a quotient of polynomials in the integration

variable x. This means that other functions can appear in the integrand, provided they do not involve x, hence that the coefficients of our polynomials in x
lie in an arbitrary field K satisfying: a K, a0 = 0.

1.1

The full partial-fraction algorithm

This method, which dates back to Newton, Leibniz and Bernoulli, should not be
used in practice, yet it remains the method found in most calculus texts and is
often taught. Its major drawback is the factorization of the denominator of the
integrand over the real or complex numbers. We outline it because it provides
the theoretical foundations for all the subsequent algorithms. Let f R(x) be
our integrand, and write f = P + A/D where P, A, D R[x], gcd(A, D) = 1,
and deg(A) < deg(D). Let
D=c

n
Y

(x ai )ei

i=1

m
Y

(x2 + bj x + cj )fj

j=1

be the irreducible factorization of D over R, where c, the ai s, bj s and cj s are

in R and the ei s and fj s are positive integers. Computing the partial fraction
decomposition of f , we get
f =P +

ei
n X
X
i=1 k=1

fj

X X Bjk x + Cjk
Aik
+
k
(x ai )
(x2 + bj x + cj )k
j=1
k=1

where the Aik s, Bjk s and Cjk s are in R. Hence,

Z

Z
f=

P+

ei Z
n X
X
i=1 k=1

fj

XX
Aik
+
(x ai )k j=1

k=1

Bjk x + Cjk
.
(x2 + bj x + cj )k

Computing P poses no problem (it will for any other class of functions), and
for the other terms we have

Z
Aik
Aik (x ai )1k /(1 k) if k > 1
=
(2)
k
Ai1 log(x ai )
if k = 1
(x ai )
and, noting that b2j 4cj < 0 since x2 + bj x + cj is irreducible in R[x],
Z

Bj1 x + Cj1
Bj1
2C

b
B
2x
+
b
j1
j j1
j
=
log(x2 +bj x+cj )+ q
arctan q
(x2 + bj x + cj )
2
2
4c b
4c b2
j

and for k > 1,

Z
Bjk x + Cjk
(x2 + bj x + cj )k

(2Cjk bj Bjk )x + bj Cjk 2cj Bjk

(k 1)(4cj b2j )(x2 + bj x + cj )k1
Z
(2k 3)(2Cjk bj Bjk )
+
.
(k 1)(4cj b2j )(x2 + bj x + cj )k1

This last formula is then used recursively until k = 1.

An alternative is to factor D linearly over C: D =

use (2) on each term of
f =P +

q X
ei
X
i=1 j=1

Qq

ei
i=1 (x i ) ,

Aij
.
(x i )j

and then

(3)

Note that this alternative is applicable to coefficients in any field K, if we factor

D linearly over its algebraic closure K, and is equivalent to expanding f into
its Laurent series at all its finite poles, since that series at x = i K is
f=

Aiei
Ai2
Ai1
+ +
+
+
e
2
i
(x i )
(x i )
(x i )

where the Aij s are the same as those in (3). Thus, this approach can be
seen as expanding the integrand into series around all its poles (including ),
then integrating the series termwise, and then interpolating for the answer,
by summing all the polar terms, obtaining the integral of (3). In addition,
this alternative shows that any rational function f K(x) has an elementary
integral in the form
Z
f = v + c1 log(u1 ) + + cm log(um )
(4)
where v, u1 , . . . , um K(x) are rational functions, and c1 , . . . , cm K are
constants. The original Risch algorithm is essentially a generalization of this
approach that searches for integrals of arbitrary elementary functions in a form
similar to (4).

1.2

The Hermite reduction

The major computational inconvenience of the full partial fraction approach is

the need to factor polynomials over R, C or K, thereby introducing algebraic
numbers even if the integrand and its integral are both in Q(x). On the other
hand, introducing algebraic numbers may be necessary, for example it is proven

in [14] that any field containing an integral of 1/(x2 + 2) must also contain 2.
Modern research has yielded so-called rational algorithms that
compute as much of the integral as possible with all calculations being
done in K(x), and
compute the minimal algebraic extension of K necessary to express the
integral.
The first rational algorithms for integration date back to the 19th century, when
both Hermite [6] and Ostrogradsky [11] invented methods for computing the
v of (4) entirely within K(x). We describe here only Hermites method, since
it is the one that has been generalized to arbitrary elementary functions. The
basic idea is that if an irreducible p K[x] appears with multiplicity k > 1 in
7

the factorization of the denominator of the integrand, then (2) implies that it
appears with multiplicity k 1 in the denominator of the integral. Furthermore,
it is possible to compute the product of all such irreducibles for each k without
factoring the denominator into irreducibles by computing its squarefree factorm
ization, i.e. a factorization D = D1 D22 Dm
, where each Di is squarefree and
gcd(Di , Dj ) = 1 for i 6= j. A straightforward way to compute it is as follows:
m1
let R = gcd(D, D0 ), then R = D2 D32 Dm
, so D/R = D1 D2 Dm and
gcd(R, D/R) = D2 Dm , which implies finally that
D1 =

D/R
.
gcd(R, D/R)

Computing recursively a squarefree factorization of R completes the one for D.

Note that [23] presents a more efficient method for this decomposition. Let now
f K(x) be our integrand, and write f = P + A/D where P, A, D K[x],
m
gcd(A, D) = 1, and deg(A) < deg(D). Let D = D1 D22 Dm
be a squarefree
factorization of D and suppose that m 2 (otherwise D is already squarefree).
Let then V = Dm and U = D/V m . Since gcd(U V 0 , V ) = 1, we can use the
extended Euclidean algorithm to find B, C K[x] such that
A
= BU V 0 + CV
1m
and deg(B) < deg(V ). Multiplying both sides by (1 m)/(U V m ) gives
A
(1 m)BV 0
(1 m)C
=
+
UV m
Vm
U V m1
so, adding and subtracting B 0 /V m1 to the right hand side, we get


A
B0
(m 1)BV 0
(1 m)C U B 0
=

+
UV m
V m1
Vm
U V m1
and integrating both sides yields
Z
Z
B
(1 m)C U B 0
A
=
+
UV m
V m1
U V m1
so the integrand is reduced to one with a smaller power of V in the denominator.
This process is repeated until the denominator is squarefree, yielding g, h K(x)
such that f = g 0 + h and h has a squarefree denominator.

1.3

The RothsteinTrager and LazardRiobooTrager algorithms

Following the Hermite reduction, we only have to integrate fractions of the form
f = A/D with deg(A) < deg(D) and D squarefree. It follows from (2) that
Z
n
X
f=
ai log(x i )
i=1

where the i s are the zeros of D in K, and the ai s are the residues of f at
the i s. The problem is then to compute those residues without splitting D.
Rothstein [18] and Trager [19] independently proved that the ai s are exactly
the zeroes of
R = resultantx (D, A tD0 ) K[t]
(5)
and that the splitting field of R over K is indeed the minimal algebraic extension
of K necessary to express the integral in the form (4). The integral is then given
by
Z
m
X
X
A
=
a log(gcd(D, A aD0 ))
(6)
D
i=1
a|Ri (a)=0

Qm

ei
i=1 Ri

is the irreducible factorization of R over K. Note that

where R =
this algorithm requires factoring R into irreducibles over K, and computing
greatest common divisors in (K[t]/(Ri ))[x], hence computing with algebraic
numbers. Trager and Lazard & Rioboo [7] independently discovered that those
computations can be avoided, if one uses the subresultant PRS algorithm to
compute the resultant of (5): let ((R0 , R1 , . . . , Rk 6= 0, 0, . . .) be the subresultant
PRS with respect to x of D and A tD0 and R = Q1 Q22 . . . Qm
m be a squarefree
factorization of their resultant. Then,
X
a log(gcd(D, A aD0 )) =
a|Qi (a)=0

P
Pa|Qi (a)=0 a log(D)
a|Qi (a)=0 a log(ppx (Rki )(a, x))

if i = deg(D)
where deg(Rki ) = i, 1 ki n
if i < deg(D)

Evaluating ppx (Rki ) at t = a where a is a root of Qi is equivalent to reducing

each coefficient with respect to x of ppx (Rki ) modulo Qi , hence computing
in the algebraic extension K[t]/(Qi ). Even this step can be avoided: it is in
fact sufficient to ensure that Qi and the leading coefficient with respect to
x of Rki do not have a nontrivial common factor, which implies then that
the remainder by Qi is nonzero, see [10] for details and other alternatives for
computing ppx (Rki )(a, x).

Algebraic Functions

By an algebraic function, we mean an element of a finitely generated algebraic

extension E of the rational function field K(x). This includes nested radicals
and implicit algebraic functions, not all of which can be expressed by radicals.
It turns out that the algorithms we used for rational functions can be extended
to algebraic functions, but with several difficulties, the first one being to define
the proper analogues of polynomials, numerators and denominators. Since E
is algebraic over K(x), for any E, there exists a polynomial p K[x][y]
such that p(x, ) = 0. We say that E is integral over K[x] if there is a
9

polynomial p K[x][y], monic in y, such that p(x, ) = 0. Integral elements

are analogous to polynomials in that their value is defined for any x K (unlike
non-integral elements, which must have at least one pole in K). The set
OK[x] = { E such that is integral over K[x]}
is called the integral closure of K[x] in E. It is a ring and
Pma finitely generated K[x]-module. Let E be any element and p = i=0 ai y i K[x][y]
be such that p(x, ) = 0 and am 6= 0. Then, q(x, am y) = 0 where q =
Pm1
y m + i=0 ai ami1
y i is monic in y, so am y OK[x] . We need a canonim
cal representation for algebraic functions similar to quotients of polynomials for
rational functions.
Expressions
as quotients of integral functions are not unique,

for example x/x = x/ x. However, E is a finite-dimensional vector space over

K(x), so let n = [E : K(x)] and w = (w1 , . . . , wn ) be any basis for E over K(x).
By the above remark, there are a1 , . . . , an K(x) such that ai wi OK[x] for
each i. Since (a1 w1 , . . . , an wn ) is also a basis for E over K(x), we can assume
without loss of generality that the basis w isPcomposed of integral elements.
n
Any E can be written uniquely as = i=1 fi wi for f1 , . . . , fn K(x),
and putting the fi s over a monic common denominator D K[x], we get an
expression
A1 w1 + . . . + An wn
=
D
Pn
where A1 , . . . , An K[x] and gcd(D, A1 , . . . , An ) = 1. We call i=1 Ai wi
OK[x] and D K[x] respectively the numerator and denominator of with
respect to w. They are defined uniquely once the basis w is fixed.

2.1

The Hermite reduction

Now that we have numerators and denominators for algebraic functions, we can
attempt to generalize the Hermite reduction of the previous section, so let f E
n
be our
Pmintegrand, w = (w1 , . . . , wn ) OK[x] be a basis for E over K(x) and
let i=1 Ai wi OK[x] and D K[x] be the numerator and denominator of f
m
with respect to w, Let D = D1 D22 Dm
be a squarefree factorization of D and
suppose that m 2.
Let
then
V
=
D
and
U = D/V m , and we ask whether we
m
Pn
can compute B = i=1 Bi wi OK[x] and h E such that deg(Bi ) < deg(V )
for each i,
Z
Z Pn
B
i=1 Ai wi
=
+
h
(7)
UV m
V m1
and the denominator of h with respect to w has no factor of order m or higher.
This turns out to reduce to solving the following linear system
f1 S1 + . . . + fn Sn = A1 w1 + . . . + An wn

(8)

for f1 , . . . , fn K(x), where

Si = U V m

 w 0
i
V m1
10

for 1 i n .

(9)

Indeed, suppose that (8) has a solution f1 , . . . , fn K(x), and write fi = Ti /Q,
where Q, T1 , . . . , Tn K[x] and gcd(Q, T1 , . . . , Tn ) = 1. Suppose further that
gcd(Q, V ) = 1. Then, we can use the extended Euclidean algorithm to find
A, R K[x] such that AV + RQ = 1, and Euclidean division to find Qi , Bi
K[x] such that deg(Bi ) < deg(V ) when Bi 6= 0 and RTi = V Qi + Bi for each i.
We then have
0
 Pn
i=1 Bi wi
h = f
V m1
Pn
Pn
n
 w 0
0
X
i
i=1 Ai wi
i=1 Bi wi
=

(RT

V
Q
)
i
i
m1
UV m
V m1
V
i=1
Pn
Pn
n
 w 0 Pn B 0 w
X
R i=1 Ti Si
i
i
i=1 Ai wi

+
V
=
Q
i=1m1i
i
m1
UV m
UV m
V
V
i=1
Pn
Pn
Pn
Pn
0
0
(1 RQ) i=1 Ai wi
0
i=1 Qi wi
i=1 Qi wi
i=1 Bi wi
+

(m

1)V

=
m
m2
m1
m1
UV
V
V
Pn
Pn
Pn V
0
0
0
AA
w
((m

1)V
Q
+
B
)w
Q
w
i
i
i
i
i
i
i=1
=
i=1
+ i=1m2 i .
U V m1
V m1
V
Hence, if in
Pnaddition the denominator of h has no factor of order m or higher,
then B = i=1 Bi wi OK[x] and h solve (7) and we have reduced the integrand.
Unfortunately, it can happen that the denominator of h has a factor of order
m or higher, or that (8) has no solution in K(x) whose denominator is coprime
with V , as the following example shows.
Example 1 Let E = K(x)[y]/(y 4 + (x2 + x)y x2 ) with basis w = (1, y, y 2 , y 3 )
over K(x) and consider the integrand
f=

w4
y3
= 2 E.
x2
x

We have D = x2 , so U = 1, V = x and m = 2. Then, S1 = x2 (1/x)0 = 1,

 y 0
24(1x2 )y 3 +32x(1x)y 2 (9x4 +45x3 +209x2 +63x+18)y18x(x3 +x2 x1)
=
,
S2 = x2
x
27x4 + 108x3 + 418x2 + 108x + 27
 2 0
64x(1x)y 3 +9(x4 +2x3 2x1)y 2 +12x(x3 +x2 x1)y+48x2 (1x2 )
2 y
S3 = x
=
x
27x4 + 108x3 + 418x2 + 108x + 27
and
2

S4 = x

y3
x

0
=

(27x4 +81x3 +209x2 +27x)y 3 +18x(x3 +x2 x1)y 2 +24x2 (x2 1)y+96x3 (1x)

27x4 + 108x3 + 418x2 + 108x + 27

so (8) becomes

f1
0
f2 0
M =
f3
0
f4
1

11

(10)

where

0
M =

0
0

18x(x3 +x2 x1)

F
(9x4 +45x3 +209x2 +63x+18)
F
32x(1x)
F
24(1x2 )
F

48x2 (1x2 )
F
12x(x3 +x2 x1)
F
9(x4 +2x3 2x1)
F
64x(1x)
F

96x3 (1x)
F
24x2 (x2 1)
F
18x(x3 +x2 x1)
F
(27x4 +81x3 +209x2 +27x)
F

and F = 27x4 + 108x3 + 418x2 + 108x + 27. The system (10) admits the unique
solution f1 = f2 = 0, f3 = 2 and f4 = (x + 1)/x, whose denominator is not
coprime with V , so the Hermite reduction is not applicable.
The above problem was first solved by Trager [20], who proved that if w is an
integral basis, i.e. its elements generate OK[x] over K[x], then the system (8)
always has a unique solution in K(x) when m > 1, and that solution always
has a denominator coprime with V . Furthermore, the denominator of each wi0
must be squarefree, implying that the denominator of h is a factor of F U V m1
where F K[x] is squarefree and coprime with U V . He also described an
algorithm for computing an integral basis, a necessary preprocessing for his
Hermite reduction. The main problem with that approach is that computing
the integral basis, whether by the method of [20] or the local alternative [21],
can be in general more expensive than the rest of the reduction process. We
describe here the lazy Hermite reduction [5], which avoids the precomputation
of an integral basis. It is based on the observation that if m > 1 and (8) does
not have a solution allowing us to perform the reduction, then either
the Si s are linearly dependent over K(x), or
(8) has a unique solution in K(x) whose denominator has a nontrivial
common factor with V , or
the denominator of some wi is not squarefree.
In all of the above cases, we can replace our basis w by a new one, also made
up of integral elements, so that the K[x]-module generated by the new basis
strictly contains the one generated by w:
Theorem 1 ([5]) Suppose that m 2 and that {S1 , . . . , Sn } as given by (9)
are linearly
Pndependent over K(x), and let T1 , . . . , Tn K[x] be not all 0 and
such that i=1 Ti Si = 0. Then,
w0 =

n
U X
Ti wi OK[x] .
V i=1

Furthermore, if gcd(T1 , . . . , Tn ) = 1, then w0

/ K[x]w1 + + K[x]wn .
Theorem 2 ([5]) Suppose that m 2 and that {S1 , . . . , Sn } as given by (9)
are linearly independent over K(x), and let Q, T1 , . . . , Tn K[x] be such that
n
X
i=1

Ai wi =

1 X
T i Si .
Q i=1

12

Then,
n

w0 =

U (V / gcd(V, Q)) X
Ti wi OK[x] .
gcd(V, Q)
i=1

Furthermore, if gcd(Q, T1 , . . . , Tn ) = 1 and deg(gcd(V, Q)) 1, then w0

/
K[x]w1 + + K[x]wn .
Theorem 3 ([5]) Suppose that the denominator F of some wi is not squarefree, and let F = F1 F22 Fkk be its squarefree factorization. Then,
w0 = F1 Fk wi0 OK[x] \ (K[x]w1 + + K[x]wn ) .
The lazy Hermite reduction proceeds by solving the system (8) in K(x). Either
the reduction will succeed, or one of the above theorems
Pn produces an element
w0 OK[x] \ (K[x]w1 + + K[x]wn ). Let then
i=1 Ci wi and F be the
numerator and denominator of w0 with respect to w. Using Hermitian row
reduction, we can zero out the last row of

..

F
C1

Cn

C2

obtaining a matrix of the form

C

1,1

C1,2
C2,2
..
.

C1,n
C2,n
..

Cn,n

C2,1
.
.
.

Cn,1 Cn,2
0
0
Pn
with Cij K[x]. Let wi = ( j=1 Cij wj )/F for 1 i n. Then, w =
(w1 , . . . , wn ) is a basis for E over K and
K[x]w1 + + K[x]wn = K[x]w1 + + K[x]wn + K[x]w0
is a submodule of OK[x] , which strictly contains K[x]w1 + + K[x]wn , since it
contains w0 . Any stricly increasing chain of submodules of OK[x] must stabilize
after a finite number of steps, which means that this process produces a basis
for which either the Hermite reduction can be carried out, or for which f has a
squarefree denominator.
Example 2 Continuing example 1 for which the Hermite reduction failed, Theorem 2 implies that
w0 =

1
(2xw3 + (x + 1)w4 ) = 2xy 2 + (x + 1)y 3 x OK[x] .
x
13

Performing a Hermitian row reduction on

x
0

2x

x
x+1

yields

x
0

1
0

so the new basis is w = (1, y, y 2 , y 3 /x), and the denominator of f with respect
to w is x, which is squarefree.

2.2

Simple radical extensions

The integration algorithm becomes easier when E is a simple radical extension

of K(x),i.e. E = K(x)[y]/(y n a) for some a K(x). Write a = A/D where
A, D K[x], and let ADn1 = A1 A22 Akk be a squarefree factorization of
ADn1 . Writing i = nqi + ri for 1 i k, where 0 ri < n, let F =
Aq11 Aqkk , H = Ar11 Arkk and z = yD/F . Then,

n
ADn1
y n Dn
D
zn = y
=
= Ar11 . . . Arkk = H .
=
F
Fn
F
Since ri < n for each i, the squarefree factorization of H is of the form H =
m
with m < n. An integral basis is then w = (w1 , . . . , wn ) where
H1 H22 Hm
wi = Qm

j=1

z i1
b(i1)j/nc

for 1 i n

(11)

Hj

and the Hermite reduction with respect to the above basis is always guaranteed
to succeed. Furthermore, when using that basis, the system (8) becomes diagQm
bij/nc
onal and its solution can be written explicitly: writing Di = j=1 Hj
we
have

0
 w 0
z i1
i
m
=
U
V
Si = U V m
V m1
Di1 V m1



0
i 1 H0
Di1
V0
z i1
= UV m

(m 1)
n H
Di1
V
Di1 V m1
 


i 1 H0
Di1 0
= U V

(m 1)V 0 wi
n H
Di1

14

so the unique solution of (8) in K(x) is

fi =

Ai
 
H0
U V i1
n H

Di1 0
Di1

(m 1)V 0

for 1 i n

(12)

and it can be shown that the denominator of each fi is coprime with V when
m 2.
Example 3 Consider
Z

(2x8 + 1) x8 + 1
dx .
x17 + 2x9 + x

The integrand is
f=

(2x8 + 1)y
E = Q(x)[y]/(y 2 x8 1)
+ 2x9 + x

x17

so H = x8 + 1 which is squarefree, implying that the integral basis (11) is

(w1 , w2 ) = (1, y). The squarefree factorization of x17 + 2x9 + x is x(x8 + 1)2 so
U = x, V = x8 + 1, m = 2, and the solution (12) of (8) is
f1 = 0,

(2x8 + 1)/4
2x8 + 1
 =
.
f2 = 
7
x8
7
x (x8 + 1) 21 x8x
8 +1 8x

We have Q = x8 , so V Q = 1, A = 1, R = 1 and RQf2 = V /2 1/4,

implying that
 0
B
y
y
and h = f
=
B=
4
V
x(x8 + 1)
solve (7),i.e.
Z

Z 8
(2x8 + 1) x8 + 1
x8 + 1
x +1
dx =
+
dx
17
9
8
x + 2x + x
4(x + 1)
x(x8 + 1)

and the remaining integrand has a squarefree denominator.

2.3

Liouvilles Theorem

Up to this point, the algorithms we have presented never fail, yet it can happen
that an algebraic function does not have an elementary integral, for example
Z
xdx

1 x3
which is not an elementary function of x. So we need a way to recognize such
functions before completing the integration algorithm. Liouville was the first
to state and prove a precise theorem from Laplaces observation that we can
restrict the elementary integration problem by allowing only new logarithms
to appear linearly in the integral, all the other terms appearing in the integral
being already in the integrand.
15

Theorem 4 (Liouville [8, 9]) Let E be an algebraic extension of the rational

function field K(x), and f E. If f has an elementary integral, then there
exist v E, constants c1 , . . . , ck K and u1 , . . . , uk E(c1 , . . . , ck ) such that
f = v 0 + c1

u01
u0
+ + ck k .
u1
uk

(13)

The above is a restriction to algebraic functions of the strong Liouville Theorem,

whose proof can be found in [4, 14]. An elegant and elementary algebraic proof
of a slightly weaker version can be found in [17]. As a consequence, we can look
for an integral in the form (4), Liouvilles Theorem guaranteeing that there is
no elementary integral if we cannot find one in that form. Note that the above
theorem does not say that every integral must have the above form, and in fact
that form is not always the most convenient one, for example



Z
1
1 + x
dx
= arctan(x) =
.
log
1 + x2
2
1 x

2.4

The integral part

Following the Hermite reduction, we can assume that we have a basis w =

(w1 , . . . , wn ) of E P
over K(x) made of integral elements such that our integrand
n
is of the form f = i=1 Ai wi /D where D K[x] is squarefree. Given Liouvilles
Theorem, we now have to solve equation (13) for v, u1 , . . . , uk and the constants
c1 , . . . , ck . Since D is squarefree, it can be shown that v OK[x] for any
solution, and in fact v corresponds to the polynomial part of the integral of
rational functions. It is however more difficult to compute than the integral
of polynomials, so Trager [20] gave a change of variable that guarantees that
either v 0 = 0 or f has no elementary integral. In order to describe it, we need
to define the analogue for algebraic functions of having a nontrivial polynomial
part:
Pm wei say that E is integral at infinity if there is a polynomial p =
i=1 ai y K[x][y] such that p(x, ) = 0 and deg(am ) deg(ai ) for each i.
Note that a rational function A/D K(x) is integral at infinity if and only if
deg(A) deg(D) since it is a zero of Dy A. When E is not integral at
infinity, we say that it has a pole at infinity. Let
O = { E such that is integral at infinity} .
A set (b1 , . . . , bn ) E n is called normal at infinity if there
Pn are r1 , . . . , rn
K(x) such that every O can be written as =
i=1 Bi ri bi /C where
C, B1 , . . . , Bn K[x] and deg(C) deg(Bi ) for each i. We say that the differential dx is integral at infinity if x1+1/r O where r is the smallest
ramification index at infinity. Trager [20] described an algorithm that converts
an arbitrary integral basis w1 , . . . , wn into one that is also normal at infinity, so
the first part of his integration algorithm is as follows:
1. Pick any basis b = (b1 , . . . , bn ) of E over K(x) that is composed of integral
elements.
16

2. Pick an integer N Z that is not zero of the denominator of f with respect

to b, nor of the discriminant of E over K(x), and perform the change of
variable x = N + 1/z, dx = dz/z 2 on the integrand.
3. Compute an integral basis w for E over K(z) and make it normal at
infinity.
4. Perform
R the Hermite
R reduction on f using w, this yields g, h E such
that f dz = g + hdz and h has a squarefree denominator with respect
to w.
R
R
5. If hz 2 has a pole at infinity, then f dz and hdz are not elementary
functions.
R
6. Otherwise, hdz is elementary if and only if there are constants c1 , . . . , ck
K and u1 , . . . , uk E(c1 , . . . , ck ) such that
h=

c1 du1
ck duk
+ +
u1 dz
uk dz

(14)

The condition that N is not a zero of the denominator of f with respect to

b implies that the f dz is integral at infinity after the change of variable, and
Trager
R proved that
R if hdz is not integral at infinity after the Hermite reduction,
then hdz and f dz are not elementary functions. The condition that N is
not a zero of the discriminant of E over K(x) implies that the ramification
indices at infinity are all equal to 1 after the change of variable, hence that hdz
is integral at infinity if and only if hz 2 O . That second condition on N can
be disregarded, in which case we must replace hz 2 in step 5 by hz 1+1/r where r
is the smallest ramification index at infinity. Note that hz 2 O implies that
hz 1+1/r O , but not conversely. Finally, we remark that for simple radical
extensions, the integral basis (11) is already normal at infinity.
Alternatively, we can use the lazy Hermite reduction in the above algorithm:
in step 3, we pick any basis made of integral elements, then perform the lazy
Hermite reduction in step 4. If h K(z) after the Hermite reduction, then we
can complete the integral without computing an integral basis. Otherwise, we
compute an integral basis and make
R it normal at infinity between steps 4 and
5. This lazy variant can compute f dx whenever it is an element of E without
computing an integral basis.

2.5

The logarithmic part

Following the previous sections, we are left with solving equation (14) for the
constants c1 , . . . , ck and for u1 , . . . , uk . We must make at this point the following
additional assumptions:
we have an integral primitive element for E over K(z), i.e. y OK[z] such
that E = K(z)(y),

17

 [E : K(z)] = [E : K(z)], i.e. the minimal polynomial for y over K[z] is

absolutely irreducible, and
we have an integral basis w = (w1 , . . . , wn ) for E over K(z), and w is
normal at infinity.
A primitive element can be computed by considering linear combinations of the
generators of E over K(x) with random coefficients in K(x), and Trager [20]
describes an absolute factorization algorithm, so the above assumptions can be
ensured, although those steps can be computationally very expensive, except
in the case of simple radical extensions. Before describing the second part of
Tragers integration algorithm, we need to define some concepts from the theory
of algebraic curves. Given a finite algebraic extension E = K(z)(y) of K(z),
a place P of EPis a proper local subring of E containing K, and a divisor is
a formal sum
nP P with finite support, where the nP s are integers and the
P s are places. Let P be a place, then its maximal ideal P is principal, so let
p E be a generator of P . The order at P is the function P : E Z which
maps f E P
to the largest k Z such that f pk P . Given f E , the divisor
of f is (f ) = P (f )P where the sum is taken over all the places. It has finite
support since P (f )P6= 0 if and only if P is a pole or zero of f . Finally, we say
that a divisor =
nP P is principal if = (f ) for some f E . Note that
P
if is principal, then
nP = 0, but the converse is not generally true, except
if E = K(z). Tragers algorithm proceeds essentially by constructing candidate
divisors for the ui s of (14):
Pn
1. Let i=1 Ai wi be the numerator of h with respect to w, and D be its
(squarefree) denominator.
Pn
2. Write i=1 Ai wi = G/H, where G K[z, y] and H K[z].
3. Let F K[z, y] be the (monic) minimum polynomial for y over K(z), t
be a new indeterminate and compute
 
 

dD
,F
, D K[t] .
R(t) = resultantz ppt resultanty G tH
dz
4. Let 1 , . . . , s K be the distinct nonzero roots of R, (q1 , . . . , qk ) be a
basis for the vector space that they generate over Q, write i = ri1 q1 + +
rik qk for each i, where rij Q and let m > 0 be a common denominator
for all the rij s.
Ps
P
5. For 1 j k, let j = i=1 mrij l rl Pl where rl is the ramification
index of Pl and Pl runs over all the places at which hdz has residue rl i .
6. If there are nonzero integers n1 , . . . , nk such that nj j is principal for each
j, then let
k
1 X qj duj
u=h
m j=1 nj uj dz
18

where uj E(1 , . . . , s ) is such that nj j = (uj ). If u = 0, then

R
Pk
hdz = j=1 qj log(uj )/(mnj ), otherwise if either u 6= 0 or there is no
such integer nj for at least one j, then hdz has no elementary integral.
Note that this algorithm expresses the integral, when it is elementary, with the
smallest possible number of logarithms. Steps 3 to 6 requires computing in
the splitting field K0 of R over K, but it can be proven that, as in the case
of rational functions, K0 is the minimal algebraic extension of K necessary to
express the integral in the form (4). Trager [20] describes a representation of
divisors as fractional ideals and gives algorithms for the arithmetic of divisors
and for testing whether a given divisor is principal. In order to determine
whether there exists an integer N such that N is principal, we need to reduce
the algebraic extension to one over a finite field Fpq for some
P good prime
p Z. Over Fpq , it is known that for every divisor =
nP P such that
P
q 2g
nP = 0, M is principal for some integer 1 M (1 + p ) , where g is
the genus of the curve [22], so we compute such an M by testing M = 1, 2, 3, . . .
until we find it. It can then be shown that for almost all primes p, if M is not
principal in characteristic 0, then N is not principal for any integer N 6= 0.
Since we can test whether the prime p is good by testing whether the image in
Fpq of the discriminant of the discriminant of the minimal polynomial for y over
K[z] is 0, this yields a complete algorithm. In the special case of hyperelliptic
extensions, i.e. simple radical extensions of degree 2, Bertrand [1] describes a
simpler representation of divisors for which the arithmetic and principality test
are more efficient than the general methods.
Example 4 Continuing example 3, we were left with the integrand

x8 + 1
w2
=
E = Q(x)[y]/(y 2 x8 1)
x(x8 + 1)
x(x8 + 1)
where (w1 , w2 ) = (1, y) is an integral basis normal at infinity, and the denominator D = x(x8 + 1) of the integrand is squarefree. Its numerator is w2 = y, so
the resultant of step 3 is
resultantx (ppt (resultanty (y t(9x8 + 1), y 2 x8 1)), x(x8 + 1)) = ct16 (t2 1)
where c is a large nonzero integer. Its nonzero roots are 1, and the integrand
has residue 1 at the place P corresponding to the point (x, y) = (0, 1) and 1 at
the place Q corresponding to the point (x, y) = (0, 1), so the divisor 1 of step
5 is 1 = P Q. It turns out that 1 , 21 and 31 are not principal, but that
 4 
w2
1 (x4 /(1 + y))0
x
and

=0
41 =
1+y
x(x8 + 1) 4 x4 /(1 + y)
which implies that
Z

x8 + 1
1
dx = log
x(x8 + 1)
4
19

x4

1 + x8 + 1


.

Example 5 Consider
Z

xdx
.
1 x3

The integrand is
f=

xy
E = Q(x)[y]/(y 2 + x3 1)
1 x3

where (w1 , w2 ) = (1, y) is an integral basis normal at infinity, and the denominator D = 1 x3 of the integrand is squarefree. Its numerator is xw2 = xy, so
the resultant of step 3 is
resultantx (ppt (resultanty (xy + 3tx2 , y 2 + x3 1)), 1 x3 ) = 729t6
R
whose only root is 0. Since f 6= 0, we conclude from step 6 that f dx is not an
elementary function.
Example 6
Z

dx

.
x 1 x3

The integrand is
f=

y
E = Q(x)[y]/(y 2 + x3 1)
x x4

where (w1 , w2 ) = (1, y) is an integral basis normal at infinity, and the denominator D = x x4 of the integrand is squarefree. Its numerator is w2 = y, so
the resultant of step 3 is
resultantx (ppt (resultanty (y + t(4x3 1), y 2 + x3 1)), x x4 ) = 729t6 (t2 1) .
Its nonzero roots are 1, and the integrand has residue 1 at the place P corresponding to the point (x, y) = (0, 1) and 1 at the place Q corresponding to the
point (x, y) = (0, 1), so the divisor 1 of step 5 is 1 = P Q. It turns out
that 1 and 21 are not principal, but that


y1
y
1 ((y 1)/(y + 1))0
31 =
and

=0
4
y+1
xx
3 (y 1)/(y + 1)
which implies that
Z

dx
1

= log
3
3
x 1x

1 x3 1

.
1 x3 + 1

Elementary Functions

Let f be an arbitrary elementary function. In order to generalize the algorithms

of the previous sections, we need to build an algebraic model in which f behaves
in some sense like a rational or algebraic function. For that purpose, we need
to formally define differential fields and elementary functions.
20

3.1

Differential algebra

A differential field (K,0 ) is a field K with a given map a a0 from K into

K, satisfying (a + b)0 = a0 + b0 and (ab)0 = a0 b + ab0 . Such a map is called a
derivation on K. An element a K which satisfies a0 = 0 is called a constant,
and the set Const(K) = {a K such that a0 = 0} of all the constants of K is
a subfield of K.
A differential field (E,0 ) is a differential extension of (K,0 ) if K E and the
derivation on E extends the one on K. In that case, an element t E is a
monomial over K if t is transcendental over K and t0 K[t], which implies that
both K[t] and K(t) are closed under 0 . An element t E is elementary over K,
if either
t0 = b0 /b for some b K , in which case we say that t is a logarithm over
K, and write t = log(b), or
t0 = b0 t some b K , in which case we say that t is an exponential over
K, and write t = eb , or
t is algebraic over K.
A differential extension (E,0 ) of (K,0 ) is elementary over K, if there exist
t1 , . . . , tm in E such that E = K(t1 , . . . , tm ) and each ti is elementary over
K(t1 , . . . , ti1 ). We say that f K has an elementary integral over K if there
exists an elementary extension (F,0 ) of (K,0 ) and g F such that g 0 = f . An
elementary function of the variable x is an element of an elementary extension
of the rational function field (C(x), d/dx), where C = Const(C(x)).
Elementary extensions are useful for modeling any function as a rational or
algebraic function of one main variable over the other terms present in the
function: given an elementary integrand f (x)dx, the integration algorithm first
constructs a field C containing all the constants appearing in f , then the rational function field (C(x), d/dx), then an elementary tower E = C(x)(t1 , . . . , tk )
containing f . Note that such a tower is not unique, and in addition, adjoining
a logarithm could in fact adjoin only a new constant, and an exponential could
in fact be algebraic, for example
Q(x)(log(x), log(2x)) = Q(log(2))(x)(log(x))

and Q(x)(elog(x)/2 ) = Q(x)( x). There are however algorithms that detect all
such occurences and modify the tower accordingly [16], so we can assume that
all the logarithms and exponentials appearing in E are monomials, and that
Const(E) = C. Let now k0 be the largest index such that tk0 is transcendental
over K = C(x)(t1 , . . . , tk0 1 ) and t = tk0 . Then E is a finitely generated algebraic extension of K(t), and in the special case k0 = k, E = K(t). Thus, f E
can be seen as a univariate rational or algebraic function over K, the major
difference with the pure rational or algebraic cases being that K is not constant
with respect to the derivation. It turns out that the algorithms of the previous
sections can be generalized to such towers, new methods being required only for
the polynomial (or integral) part. We note that Liouvilles Theorem remains
valid when E is an arbitrary differential field, so the integration algorithms work
by attempting to solve equation (13) as previously.
21


1
Example 7 The
function (1) is the element f = (t t ) 1/2 of E = K(t)
where K = Q( 1)(x)(t1 , t2 ) with
p

3
t1 = x3 x + 1, t2 = e2 1(x t1 ) ,
and t = e((1t2 )/(1+t2 ))x 1
which is transcendental over K. Alternatively, it can also be written as the
element f = 2/(1 + 2 ) of F = K() where K == Q(x)(1 , 2 ) with


p
x + 2
3
3
1 = x x + 1, 2 = tan(x 1 ),
and = tan
2
which is a transcendental monomial over K. It turns out that both towers can
be used in order to integrate f .
The algorithms of the previous sections relied extensively on squarefree factorization and on the concept of squarefree polynomials. The appropriate analogue
in monomial extensions is the notion of normal polynomials: let t be a monomial over K, we say that p K[t] is normal (with respect to 0 ) if gcd(p, p0 ) = 1,
and that p is special if gcd(p, p0 ) = p,i.e. p | p0 in K[t]. For p K[t] squarefree,
let ps = gcd(p, p0 ) and pn = p/ps . Then, p = ps pn while ps is special and pn is
normal. Therefore, squarefree factorization can be used to write any q K[t]
as a product q = qs qn , where gcd(qs , qn ) = 1, qs is special and all the squarefree
factors of qn are normal. We call qs the special part of q and qn its normal part.

3.2

The Hermite reduction

The Hermite reductions we presented for rational and algebraic functions work
in exactly the same way algebraic extensions of monomial extensions of K,
as long as we apply them only to the normal part of the denominator of the
integrand. Thus, if D is the denominator of the integrand, we let S be the
m
special part of D, D1 D22 Dm
be a squarefree factorization of the normal part
m
of D, V = Dm , U = D/V and the rational and algebraic Hermite reductions
proceed normally, eventually yielding an integrand whose denominator has a
squarefree normal part.
Example 8 Consider
Z

x tan(x)
dx .
tan(x)2

The integrand is
f=

xt
K(t)
t2

where K = Q(x) and t0 = t2 + 1 .

Its denominator is D = t2 , and gcd(t, t0 ) = 1 implying that t is normal, so

m = 2, V = t, U = D/t2 = 1, and the extended Euclidean algorithm yields
A
= t x = x(t2 + 1) + (xt + 1)t = xU V 0 + (xt + 1)V
1m
22

implying that
Z

x
x tan(x)
dx =

tan(x)2
tan(x)

Z
xdx

and the remaining integrand has a squarefree denominator.

Example 9 Consider
p
Z
log(x)2 + 2x log(x) + x2 + (x + 1) x + log(x)
dx .
x log(x)2 + 2x2 log(x) + x3
The integrand is
f=

t2 + 2xt + x2 + (x + 1)y
E = K(t)[y]/(y 2 x t)
xt2 + 2x2 t + x3

where K = Q(x) and t = log(x). The denominator of f with respect to the basis
w = (1, y) is D = xt2 + 2x2 t + x3 whose squarefree factorization is x(t + x)2 .
Both x and t + x are normal, so m = 2, V = t + x, U = D/V 2 = x, and the
solution (12) of (8) is
f1 =

t2 + 2xt + x2
t2 + 2xt + x2
x+1
 = 2 .
=

, f2 = 
0
0 +1
x((t + 1))
x+1
x (t + x) 12 tt+x
(t0 + 1)

We have Q = 1, so 0V + 1Q = 1, A = 0, R = 1, RQf1 = f1 = V 2 /(x + 1)

and RQf2 = f2 = 0V 2, so B = 2y and
 0
B
1
h=f
=
V
x
implying that
p
Z
Z
log(x)2 + 2x log(x) + x2 + (x + 1) x + log(x)
2
dx
p
dx
=

+
x log(x)2 + 2x2 log(x) + x3
x
x + log(x)
and the remaining integrand has a squarefree denominator.

3.3

The polynomial reduction

In the transcendental case E = K(t) and when t is a monomial satisfying

degt (t0 ) 2, then it is possible to reduce the degree of the polynomial part of the
integrand until it is smaller than degt (t0 ). In the case when t = tan(b) for some
b K, then it is possible either to prove that the integral is not elementary, or
to reduce the polynomial part of the integrand to be in K. Let f K(t) be our
integrand and write f = P + A/D where P, A, D K[t] and deg(A) < deg(D).
Pe
Pd
Write P = i=0 pi ti and t0 = i=0 ci ti where p0 , . . . , pe , c0 , . . . , cd K, d 2,
pe 6= 0 and cd 6= 0. It is easy to verify that if e d, then
0

ae
ed+1
P =
t
+P
(15)
(e d + 1)cd
23

where P K[t] is such that P = 0 or degt (P ) < e. Repeating the above

transformation we obtain Q, R K[t] such that R = 0 or degt (R) < d and
Pd1
P = Q0 + R. Write then R = i=0 ri ti where r0 , . . . , rd1 K. Again, it is
easy to verify that for any special S K[t] with degt (S) > 0, we have
R=

rd1 S 0
1
+R
degt (S) cd S

where R K[t] is such that R = 0 or degt (R) < e 1. Furthermore, it can be

proven [4] that if R + A/D has an elementary integral over K(t), then rd1 /cd
is a constant, which implies that

Z 
Z
rd1
A
1
log(S) +
R+
R=
degt (S) cd
D
so we are left with an integrand whose polynomial part has degree at most
degt (t0 ) 2. In the case t = tan(b) for b K, then t0 = b0 t2 + b0 , so R K.
Example 10 Consider
Z

(1 + x tan(x) + tan(x)2 )dx .

The integrand is
where K = Q(x) and t0 = t2 + 1 .

f = 1 + xt + t2 K(t)

Using (15), we get P = f t0 = f (t2 + 1) = xt so

Z
Z
(1 + x tan(x) + tan(x)2 )dx = tan(x) + x tan(x)dx
ans since x0 6= 0, the above criterion implies that the remaining integral is not
an elementary function.

3.4

The residue criterion

Similarly to the Hermite reduction, the RothsteinTrager and LazardRioboo

Trager algorithms are easy to generalize to the transcendental case E = K(t) for
arbitrary monomials t: let f K(t) be our integrand and write f = P + A/D +
B/S where P, A, D, B, S K[t], deg(A) < deg(D), S is special and, following
the Hermite reduction, D isPnormal.
z be a new indeterminate, :

Let
P then
i
0 i
K[z] K[z] be given by
a
z
=
a
z
,
i
i
i i
R = resultantt (D, A zD0 )

K[z]

be the RothsteinTrager resultant, R = R1 R22 . . . Rkk be its squarefree factorization, Qi = gcdz (Ri , (Ri )) for each i, and
g=

k
X

a log(gcd t (D, A aD0 )) .

i=1 a|Qi (a)=0

24

Note that the the roots of each Qi must all be constants, and that the arguments
of the logarithms can be obtained directly from the subresultant PRS of D and
A zD0 as in the rational function case. It can then be proven [4] that
f g 0 is always simpler than f ,
the splitting field of Q1 . . . Qk over
R K is the minimal algebraic extension
of K needed in order to express f in the form (4),
if f has an elementary integral over K(t), then R | (R) in K[z] and the
denominator of f g 0 is special.
Thus, while in the pure rational function case the remaining integrand is a
polynomial, in this case the remaining integrand has a special denominator.
In that case we have additionally that if its integral is elementary, then (13)
has a solution such that v K(t) has a special denominator, and each ui
K(c1 , . . . , ck )[t] is special.
Example 11 Consider
Z

2 log(x)2 log(x) x2
dx .
log(x)3 x2 log(x)

The integrand is
2t2 t x2
K(t)
t3 x2 t

f=

where K = Q(x) and t = log(x) .

Its denominator is D = t3 x2 t, which is normal, and the resultant is



3
2 2x 3z 2
t + (2xz 1)t + x(z x)
R = resultantt (t x t,
x


1
x
3
2
3
2
= 4x (1 x ) z xz z +
4
4
which is squarefree in K[z]. We have
(R) = x2 (4(5x2 + 3)z 3 + 8x(3x2 2)z 2 + (5x2 3)z 2x(3x2 2))
so



1
Q1 = gcd z (R, R) = x2 z 2
4

and


gcd t


2x 3a 2
t + (2xa 1)t + x(a x) = t + 2ax
t + x t,
x
3

where a2 1/4 = 0, whence

g=

a log(t + 2ax) =

a|a2 1/4=0

25

1
1
log(t + x) log(t x) .
2
2

Computing f g 0 we find

 Z
Z
2 log(x)2 log(x) x2
1
log(x) + x
dx
dx
=
log
+
3
2
log(x) x log(x)
2
log(x) x
log(x)
and since degz (Q1 ) < degz (R), it follows that the remaining integral is not an
elementary function (it is in fact the logarithmic integral Li(x)).
In the most general case, when E = K(t)(y) is algebraic over K(t) and y
is integral over K[t], the criterion part of the above result remains valid: let
w = (w1 , . . . , w
Pnn) be an integral
Pnbasis for E over K(t) and write the integrand
f E as f = i=1 Ai wi /D+ i=1 Bi wi /S
Pnwhere S is special and, following the
Hermite reduction, D is normal. Write i=1 Ai wi = G/H, where G K[t, y]
and H K[t], let F K[t, y] be the (monic) minimum polynomial for y over
K(t), z be a new indeterminate and compute
R(z) = resultantt (ppz (resultanty (G tHD0 , F )), D) K[t] .

(16)

It can then be proven [2] that if f has an elementary integral over E, then
R | (R) in K[z].
Example 12 Consider
Z

(1/3)

log(1 + ex )
dx .
1 + log(1 + ex )

(17)

The integrand is
f =f =

y
E = K(t)[y]/(y 3 t)
t+1

where K = Q(x)(t1 ), t1 = ex and t = log(1+t1 ). Its denominator with respect to

the integral basis w = (1, y, y 2 ) is D = t + 1, which is normal, and the resultant
is
R = resultantt (ppz (resultanty (y zt1 /(1+t1 ), y 3 t)), t+1) =
We have
(R) =

t31
z 3 1 .
(1 + t1 )3

3t31
z3
(1 + t1 )4

which is coprime with R in K[z], implying that the integral (17) is not an elementary function.

3.5

The transcendental logarithmic case

Suppose now that t = log(b) for some b K , and that E = K(t). Then, every
special polynomial must be in K, so, following the residue criterion, we must
look for a solution v K[t], u1 , . . . , uk K(c1 , . . . , cn ) of (13). Furthermore,

26

Pd
the integrand f is also in K[t], so write f = i=0 fi ti where f0 , . . . , fd K and
Pd+1
fd 6= 0. We must have degt (v) d + 1, so writing v = i=0 vi ti , we get
Z

fd td + . . . f1 t + f0 = vd+1 td+1 + . . . + v1 t + v0 +

k
X

ci log(ui ) .

i=1

If d = 0, then the above is simply an integration problem for f0 K, which

can be solved recursively. Otherwise, differentiating both sides and equating
the coefficients of td , we get vd+1 0 = 0 and
fd = vd 0 + (d + 1)vd+1

b0
.
b

(18)

Since fd K, we can recursively apply the integration algorithm to fd , either

proving that (18) has no solution, in which case f has no elementary integral,
or obtaining the constant vd+1 , and vd up to an additive constant (in fact, we
apply recursively a specialized version of the integration algorithm to equations
of the form (18), see [4] for details). Write then vd = vd + cd where vd K is
known and cd Const(K) is undetermined. Equating the coefficients of td1
yields
b0
b0
fd1 dvd = vd1 0 + dcd
b
b
which is an equation of the form (18), so we again recursively compute cd and
vd1 up to an additive constant. We repeat this process until either one of
the recursive integrations fails, in which case f has no elementary integral, or
we reduce our integrand to an element of K, which is then integrated recursively. The algorithm of this section can also be applied to real arc-tangent
extensions,i.e. K(t) where t is a monomial satisfying t0 = b0 /(1 + b2 ) for some
b K.

3.6

The transcendental exponential case

Suppose now that t = eb for some b K, and that E = K(t). Then, every
nonzero special polynomial must be of the form atm for a K and m N.
Since
(atm )0
a0
t0
a0
=
+m =
+ mb0 ,
m
at
a
t
a
we must then look for a solution v K[t, t1 ], u1 , . . . , uk K(c1 , . . . , cn )
Pd
of (13). Furthermore, the integrand f is also in K[t, t1 ], so write f = i=e fi ti
where fe , . . . , fd K and e, d Z. Since (atm )0 = (a0 + mb0 )tm for any m Z,
Pd
we must have v = M b + i=e vi ti for some integer M , hence
Z X
d
i=e

fi ti = M b +

d
X
i=e

27

vi ti +

k
X
i=1

ci log(ui ) .

Differentiating both sides and equating the coefficients of each power of td , we

get
k
X
u0
f0 = (v0 + M b)0 +
ci i ,
ui
i=1
which is simply an integration problem for f0 K, and
fi = vi0 + ib0 vi

for e i d, i 6= 0 .

The above problem is called a Risch differential equation over K. Although

solving it seems more complicated than solving g 0 = f , it is actually simpler
than an integration problem because we look for the solutions vi in K only
rather than in an extension of K. Bronstein [2, 3, 4] and Risch [12, 13, 14]
describe algorithms for solving this type of equation when K is an elementary
extension of the rational function field.

3.7

The transcendental tangent case

/K
Suppose now that t = tan(b) for some b K,i.e. t0 = b0 (1 + t2 ), that 1
and that E = K(t). Then, every nonzero special polynomial must be of the
form a(t2 + 1)m for a K and m N. Since
a0
(t2 + 1)0
a0
(a(t2 + 1)m )0
=
+
m
=
+ 2mb0 t
a(t2 + 1)m
a
t2 + 1
a
we must look for v = V /(t2 + 1)m where V K[t], m1 , . . . , mk N, constants
c1 , . . . , ck K and u1 , . . . , uk K(c1 , . . . , ck ) such that
f = v 0 + 2b0 t

k
X

ci mi +

i=1

k
X
i=1

ci

u0i
.
ui

Furthermore, the integrand f K(t) following the residue criterion must be of

the form f = A/(t2 + 1)M where A K[t] and M 0. If M > 0, it can be
shown that m = M and that
 0 
   
c
0
2mb0
c
a
+
=
(19)
d0
2mb0
0
d
b
where at+b and ct+d are the remainders modulo t2 +1 of A and V respectively.
The above is a coupled differential system, which can be solved by methods
similar to the ones used for Risch differential equations [4]. If it has no solution,
then the integral is not elementary, otherwise we reduce the integrand to h
K[t], at which point the polynomial reduction either proves that its integral is
not elementary, or reduce the integrand to an element of K, which is integrated
recursively.

28

Example 13 Consider
Z

sin(x)
dx .
x

The integrand is
f=

2t/x
K(t)
t2 + 1

where K = Q(x) and t = tan

x
2

Its denominator is D = t2 + 1, which is special, and the system (19) becomes

 0 
  

c
0 1
c
2/x
+
=
d0
1 0
d
0
which has no solution in Q(x), implying that the integral is not an elementary
function.

3.8

The algebraic logarithmic case

The transcendental logarithmic case method also generalizes to the case when
E = K(t)(y) is algebraic over K(t), t = log(b) for b K and y is integral over
K[t]: following the residue criterion, we can assume that R | (R) where R is
given by (16), hence that all its roots in K are constants. The polynomial part
of the integrand is replaced by a family of at most [E : K(t)] Puiseux expansions
at infinity, each of the form
X
am m + . . . + a1 1 +
ai i
(20)
i0

where r = t1 for some positive integer r. Applying the integration algorithm

recursively to ar K, we can test whether there exist Const(K) and v K
such that
b0
ar = v 0 + .
b
If there are no such v and c for at least one of the series, then the integral
is not elementary, otherwise is uniquely determined by ar , so let 1 , . . . , q
where q [E : K(t)] be the distinct constants we obtain, 1 , . . . , s K be
the distinct nonzero roots of R, and (q1 , . . . , qk ) be a basis for the vector space
generated by the i s and i s over Q. Write i = ri1 q1 + + rik qk and
i = si1 q1 + + sik qk for each i, where rij , sij Q and let m > 0 be a common
denominator for all the rij s and sij s. For 1 j k, let
j =

s
X
i=1

mrij

rl Pl

q
X
i=1

msij

sl Ql

where rl is the ramification index of Pl , sl is the ramification index of Ql , Pl

runs over all the finite places at which hdz has residue rl i and Ql runs over all
the infinite places at which = i . As in the pure algebraic case, if there is a
29

j for which N j is not principal for any nonzero integer N , then the integral is
not elementary, otherwise, let n1 , . . . , nk be nonzero integers such that nj j is
principal for each j, and
k
1 X qj u0j
h=f
m j=1 nj uj
where f is the integrand and uj E(1 , . . . , s , 1 , . . . , q ) is such that nj j =
(uj ). If the integral of h is elementary, then (13) must have a solution with
v OK[x] and u1 , . . . , uk K, so we must solve
Pn
h=

n
n
k
X
X
X
Ai wi
u0
0
0
=
vi wi +
vi wi +
ci i
D
ui
i=1
i=1
i=1

i=1

(21)

for v1 , . . . , vn K[t], constants c1 , . . . , cn K and u1 , . . . , uk K where

w = (w1 , . . . , wn ) is an integral basis for E over K(t).
If E is a simple radical extension of K(t), and we use the basis (11) and the
notation of that section, then w1 = 1 and


D0
i 1 H0
i1 wi for 1 i n .
wi0 =
(22)
n H
Di1
This implies that (21) becomes
k
X
u0
A1
= v10 +
ci i
D
ui
i=1

which is simply an integration problem for A1 /D K(t), and



0
Di1
i 1 H0
Ai
0
= vi +

vi for 1 < i n
D
n H
Di1

(23)

(24)

which are Risch differential equations over K(t).

Example 14 Consider
p
Z
(x2 + 2x + 1) x + log(x) + (3x + 1) log(x) + 3x2 + x
p
dx .
(x log(x) + x2 ) x + log(x) + x2 log(x) + x3
The integrand is
f=

((3x + 1)t x3 + x2 )y (2x2 x 1)t 2x3 + x2 + x

E = K(t)[y]/(F )
xt2 (x3 2x2 )t x4 + x3

where F = y 2 x t, K = Q(x) and t = log(x). Its denominator with respect

to the integral basis w = (1, y) is D = xt2 (x3 2x2 )t x4 + x3 , which is
normal, and the resultant is
R

= resultantt (ppz (resultanty (((3x + 1)t x3 + x2 )y

(2x2 x 1)t 2x3 + x2 + x zD0 , F )), D)
= x12 (2x + 1)2 (x + 1)2 (x 1)2 z 3 (z 2)
30

We have
(R) =

36x3 + 16x2 28x 12

R
x(2x + 1)(x + 1)(x 1)

so R | (R) in K[z]. Its only nonzero root is 2, and the integrand has residue
2 at the place P corresponding to the point (t, y) = (x2 x, x). There is only
one place Q at infinity of ramification index 2, and the coefficient of t1 in the
Puiseux expansion of f at Q is
a2 = 1 2x +

x0
1
= (x x2 )0 +
x
x

which implies that the corresponding is 1. Therefore, the divisor for the logand
is = 2P 2Q. It turns out that = (u) where u = (x + y)2 E , so the new
integrand is
0
u0
(x + y)
(x + 1)y
.
h=f
=f 2
=
u
x+y
xt + x2
We have y 2 = t + x, which is squarefree, so (23) becomes
0 = v10 +

k
X
i=1

ci

u0i
ui

whose solution is v1 = k = 0, and (24) becomes

x+1
x+1
= v20 +
v2
xt + x2
2xt + 2x2
whose solution is v2 = 2, implying that h = 2y 0 , hence that
p
Z
(x2 + x + 1) x + log(x) + (3x + 1) log(x) + 3x2 + x
p
dx =
(x log(x) + x2 ) x + log(x) + x2 log(x) + x3
p
p
2 x + log(x) + 2 log(x + x + log(x)) .
In the general case when E is not a radical extension of K(t), (21) is
solvedP
by bounding degt (vi ) and comparing the Puiseux expansions at infinn
ity of i=1 vi wi with those of the form (20) of h, see [2, 12] for details.

3.9

The algebraic exponential case

The transcendental exponential case method also generalizes to the case when
E = K(t)(y) is algebraic over K(t), t = eb for b K and y is integral over K[t]:
following the residue criterion, we can assume that R | (R) where R is given
by (16), hence that all its roots in K are constants. The denominator of the
integrand must be of the form D = tm U where gcd(U, t) = 1, U is squarefree,
and m 0.
If m > 0, E is a simple radical extension of K(t), and we use the basis (11),
then it is possible to reduce the power of t appearing in D by a process similar
31

Pn
to the Hermite reduction: writing the integrand f = i=1 Ai wi /(tm U ), we ask
whether we can compute b1 , . . . , bn K and C1 , . . . , Cn K[t] such that
Pn
Z Pn
Z Pn
i=1 Ci wi
i=1 bi wi
i=1 Ai wi
+
=
.
tm U
tm
tm1 U
Differentiating both sides and multiplying through by tm we get
Pn
Pn
n
n
n
X
X
X
t i=1 Ci wi
i=1 Ai wi
=
.
b0i wi +
bi wi0 mb0
bi wi +
U
U
i=1
i=1
i=1
Using (22) and equating the coefficients of wi on both sides, we get
Ai
tCi
= b0i + (i mb0 )bi +
U
U
where
i =

for 1 i n

(25)

D0
i 1 H0
i1 K(t) .
n H
Di1

Since t0 /t = b0 K, it follows that the denominator of i is not divisible by t

in K[t], hence, evaluating (25) at t = 0, we get
Ai (0)
= b0i + (i (0) mb0 )bi
U (0)

for 1 i n

(26)

which are Risch differential equations over K(t). If any of them has no solution
in K(t), then the integral is not elementary, otherwise we repeat this process
until the denominator of the integrand is normal. We then perform the change
0
of variable t = t1 , which is also exponential over K since t = b0 t, and repeat
the above process in order to eliminate the power of t from the denominator of
the integrand. It can be shown that after this process, any solution of (13) must
have v K.
Example 15 Consider
Z

(1/3)

3(x + ex )

+ (2x2 + 3x)ex + 5x2

(1/3)

x(x + ex )

dx .

The integrand is
f=

((2x2 + 3x)t + 5x2 )y 2 + 3t + 3x

E = K(t)[y]/(y 3 t x)
xt + x2

where K = Q(x) and t = ex . Its denominator with respect to the integral basis
w = (1, y, y 2 ) is D = xt + x2 , which is normal, and the resultant is
R

= resultantt (ppz (resultanty (((2x2 + 3x)t + 5x2 )y 2 + 3t + 3x zD0 ,

y 3 t x)), D) = x8 (1 x)3 z 3 .
32

We have
(R) =

11x 8
R
x(x 1)

so R | (R) in K[z], its only root being 0. Since D is not divisible by t, let
0
2
3
t = t1 and z = ty. We have t = t and z 3 t xt = 0, so the integral
basis (11) is


z2
.
w = (w1 , w2 , w3 ) = 1, z,
t
Writing f in terms of that basis gives
2

f=

3xt + 3t + (5x2 t + 2x2 + 3x)w3

2

x2 t + xt
2

whose denominator D = t(x + x2 t) is divisible by t. We have H = t (1 + xt),

so D0 = D1 = 1 and D2 = t, implying that
1 = 0, 2 =

(1 3x)t 2
(2 3x)t 1
, and 3 =
.
3xt + 3
3xt + 3

Therefore the equations (26) become

2
1
0 = b01 + b1 , 0 = b02 + b2 , and 2x + 3 = b03 + b3
3
3
whose solutions are b1 = b2 = 0 and b3 = 3x, implying that the new integrand is

0
3xw3
3
h=f
=
x
t
hence that
Z
(1/3)
3(x + ex )
+ (2x2 + 3x)ex + 5x2
x(x +

(1/3)
ex )

(2/3)

dx = 3x(x + ex )

Z
+3

dx
.
x

In the general case when E is not a radical extension ofP

K(t), following the
n
m
where
Hermite reduction, any solution of (13) must have v =
i=1 vi wi /t
v1 , . . . , vm K[t]. We can compute v by bounding
deg
(v
)
and
comparing
the
i
t
Pn
Puiseux expansions at t = 0 and at infinity of i=1 vi wi /tm with those of the
form (20) of the integrand, see [2, 12] for details.
Once we are reduced to solving (13) for v K, constants c1 , . . . , ck K
and u1 , . . . , uk E(c1 , . . . , ck ) , constants 1 , . . . , s K can be determined at
all the places above t = 0 and at infinity in a manner similar to the algebraic
logarithmic case, at which point the algorithm proceeds by constructing the
divisors j and the uj s as in that case. Again, the details are quite technical
and can be found in [2, 12, 13].

33

References
[1] L. Bertrand. Computing a hyperelliptic integral using arithmetic in the
jacobian of the curve. Applicable Algebra in Engineering, Communication
and Computing, 6:275298, 1995.
[2] M. Bronstein. On the integration of elementary functions. Journal of
Symbolic Computation, 9(2):117173, February 1990.
[3] M. Bronstein. The Risch differential equation on an algebraic curve. In
S. Watt, editor, Proceedings of ISSAC91, pages 241246. ACM Press, 1991.
[4] M. Bronstein. Symbolic Integration I Transcendental Functions. Springer,
Heidelberg, 1997. 2nd Ed., 2004.
[5] M. Bronstein. The lazy Hermite reduction. Rapport de Recherche RR
3562, INRIA, 1998.
[6] E. Hermite. Sur lintegration des fractions rationelles. Nouvelles Annales
de Mathematiques (2eme serie), 11:145148, 1872.
[7] D. Lazard and R. Rioboo. Integration of rational functions: Rational computation of the logarithmic part. Journal of Symbolic Computation, 9:113
116, 1990.
[8] J. Liouville. Premier memoire sur la determination des integrales dont la
valeur est algebrique. Journal de lEcole Polytechnique, 14:124148, 1833.
[9] J. Liouville. Second memoire sur la determination des integrales dont la
valeur est algebrique. Journal de lEcole Polytechnique, 14:149193, 1833.
[10] T. Mulders. A note on subresultants and a correction to the Lazard
RiobooTrager formula in rational function integration. Journal of Symbolic Computation, 24(1):4550, 1997.
[11] M.W. Ostrogradsky. De lintegration des fractions rationelles. Bulletin de
la Classe PhysicoMathematiques de lAcademie Imperiale des Sciences de
St. Petersbourg, IV:145167,286300, 1845.
[12] R. Risch. On the integration of elementary functions which are built up
using algebraic operations. Research Report SP-2801/002/00, System Development Corporation, Santa Monica, CA, USA, 1968.
[13] R. Risch. Further results on elementary functions. Research Report RC2402, IBM Research, Yorktown Heights, NY, USA, 1969.
[14] R. Risch. The problem of integration in finite terms. Transactions of the
American Mathematical Society, 139:167189, 1969.
[15] R. Risch. The solution of problem of integration in finite terms. Bulletin
of the American Mathematical Society, 76:605608, 1970.
34

[16] R. Risch. Algebraic properties of the elementary functions of analysis.

American Journal of Mathematics, 101:743759, 1979.
[17] M. Rosenlicht. Integration in finite terms. American Mathematical Monthly,
79:963972, 1972.
[18] M. Rothstein. A new algorithm for the integration of exponential and
logarithmic functions. In Proceedings of the 1977 MACSYMA Users Conference, pages 263274. NASA Pub. CP-2012, 1977.
[19] B.M. Trager. Algebraic factoring and rational function integration. In
Proceedings of SYMSAC76, pages 219226, 1976.
[20] B.M. Trager. On the integration of algebraic functions. PhD thesis, MIT,
Computer Science, 1984.
[21] M. van Hoeij. An algorithm for computing an integral basis in an algebraic
function field. J. Symbolic Computation, 18(4):353364, October 1994.
[22] Andre Weil. Courbes algebriques et varietes Abeliennes. Hermann, Paris,
1971.
[23] D.Y.Y. Yun. On square-free decomposition algorithms. In Proceedings of
SYMSAC76, pages 2635, 1976.

35