MULTIPLE BINOMIAL SUMS

ALIN BOSTAN, PIERRE LAIREZ, AND BRUNO SALVY

Abstract. Multiple binomial sums form a large class of multi-indexed se-

quences, closed under partial summation, which contains most of the sequences
obtained by multiple summation of products of binomial coefficients and also all
the sequences with algebraic generating function. We study the representation
of the generating functions of binomial sums by integrals of rational functions.
The outcome is twofold. Firstly, we show that a univariate sequence is a
multiple binomial sum if and only if its generating function is the diagonal of
a rational function. Secondly, we propose algorithms that decide the equality
of multiple binomial sums and that compute recurrence relations for them.
In conjunction with geometric simplifications of the integral representations,
this approach behaves well in practice. The process avoids the computation of
certificates and the problem of the appearance of spurious singularities that
afflicts discrete creative telescoping, both in theory and in practice.

Introduction
The computation of definite sums in computer algebra is classically handled by
the method of creative telescoping initiated in the 1990s by Zeilberger (Zeilberger
1990, 1991a; Wilf and Zeilberger 1992). For example, it applies to sums like
 3
n n k   n X n  2 
4k

X X X n X i+j 4n − 2i − 2j
(1) 2k

,   or .
j j 2n − 2i
k=0 k k=0 j=0 i=0 j=0
P
In order to compute a sum k u(n, k) of a bivariate sequence u, this method
computes an identity of the form
ap (n)u(n + p, k) + · · · + a0 (n)u(n, k) = v(n, k + 1) − v(n, k).
Provided that it is possible to sum both sides over k and that the sequence v
vanishes at the endpoints of the domain of summation, the left-hand side—called
a telescoper—gives a recurrence for the sum. The sequence v is then called the
certificate of the identity.
In the case of multiple sums, this idea leads to searching for a telescoping identity
of the form

(2) ap (n)u(n + p, k1 , . . . , km ) + · · · + a0 (n)u(n, k1 , . . . , km ) =

v1 (n, k1 + 1, k2 , . . . , km ) − v1 (n, k1 , . . . , km ) + · · ·


+ vm (n, k1 , . . . , km + 1) − vm (n, k1 , . . . , km ) .
Again, under favorable circumstances the sums of the sequences on the right-hand
side telescope, leaving a recurrence for the sum on the left-hand side.

Date: June 15, 2016.

2010 Mathematics Subject Classification. 05A10 (33F10 68W30).
Key words and phrases. Binomial sum, multiple sum, symbolic computation, diagonal, integral
representation.
1
2 A. BOSTAN, P. LAIREZ, AND B. SALVY

This high-level presentation hides practical difficulties. It is important to check

that the sequences on both sides of the identities above are defined over the whole
range of summation (Abramov 2006; Abramov and Petkovšek 2005). More often
than not, singularities do appear. To the best of our knowledge, no algorithm
based on creative telescoping manages to work around this difficulty; they all let
the user handle it. As a consequence, computing the certificate is not merely a
useful by-product of the algorithm, but indeed a necessary part of the computation.
Unfortunately, the size of the certificate may be much larger than that of the final
recurrence and thus costly in terms of computational complexity.
The computation of multiple integrals of rational functions has some similarities
with the computation of discrete sums and the method of creative telescoping
applies there too. It may also produce extra singularities in the certificate, but
in the differential setting this is not an issue anymore: for the integrals we are
interested in, the integration path can always be moved to get around any extra
singularity. Moreover, we have showed (Bostan et al. 2013a; Lairez 2016) that
integration of multivariate rational functions over cycles can be achieved efficiently
without computing the corresponding certificate and without introducing spurious
singularities. In that case, the algorithm computes a linear differential equation
for the parameterized integral. It turns out that numerous multiple sums can be
cast into problems of rational integration by passing to generating functions. The
algorithmic consequences of this observation form the object of the present work.

Content. In §1, we define a class of multivariate sequences, called (multiple)

binomial sums, that contains the binomial coefficient sequence and that is closed
under pointwise addition, pointwise multiplication, linear change of variables and
partial summation. Not every sum that creative telescoping can handle is a binomial
sum: for example, among the three sums in Eq. (1), the second one and the third
one are binomial sums but the first one is not, since it contains the inverse of a
binomial coefficient; moreover, it cannot be rewritten as a binomial sum (see §1.2).
Yet many sums coming from combinatorics and number theory are binomial sums.
In §2, we explain how to compute integral representations of the generating function
of a binomial sum in an automated way. The outcome is twofold. Firstly, in §3, we
work further on these integral representations to show that the generating functions
of univariate binomial sums are exactly the diagonals of rational power series. This
equivalence characterizes binomial sums in an intrinsic way which dismisses the
arbitrariness of the definition. All the theory of diagonals transfers to univariate
binomial sums and gives many interesting arithmetic properties. Secondly, in §4, we
show how to use integral representations to actually compute with binomial sums
(e.g. find recurrence relations or prove identities automatically) via the computation
of Picard-Fuchs equations. The direct approach leads to integral representations
that involve far too many variables to be efficiently handled. In §5, we propose
a general method, that we call geometric reduction, to reduce tremendously the
number of variables in practice. In §6, we describe some variants with the purpose
of implementing the algorithms, and finally, in §7, we show how the method applies
to some classical identities and more recent ones that were conjectural so far.
All the algorithms that are presented here are implemented in Maple and are
available at https://github.com/lairez/binomsums.
 MULTIPLE BINOMIAL SUMS 3

Example. The following proof of an identity of Dixon (1891),

2n  3
k 2n (3n)!
X
(3) (−1) = (−1)n 3 ,
k n!
k=0
illustrates well the main points of the method. The strategy is as follows: find an
integral representation of the generating function of the left-hand side; simplify
this integral representation using partial integration; use the simplified integral
representation to compute a differential equation of which the generating function
is solution; transform this equation into a recurrence relation; solve this recurrence
relation.
First of all, the binomial coefficient nk is the coefficient of xk in (1+x)n . Cauchy’s

integral formula ensures that

(1 + x)n dx
  I
n 1
= ,
k 2πi γ xk x
where γ is the circle x ∈ C |x| = 21 . Therefore, the cube of a binomial coefficient


can be represented as a triple integral

 3
(1 + x1 )2n (1 + x2 )2n (1 + x3 )2n dx1 dx2 dx3
I
2n 1
= .
k (2πi)3 γ×γ×γ xk1 xk2 xk3 x1 x2 x3
As a result, the generating function of the left-hand side of Equation (3) is
∞ 2n  3
def
X X 2n
y(t) = tn (−1)k
n=0
k
k=0
∞ X 2n 3
!n  k
−1
I X
1 Y
2 dx1 dx2 dx3
= t (1 + x i )
(2iπ)3 γ 3 n=0 i=1
x 1 x 2 x 3 x1 x2 x3
k=0
!n   2n+1
1
I X ∞ Y3 1 − x1−1x 2 x 3 dx1 dx2 dx3
= t (1 + xi )2
(2iπ)3 γ 3 n=0 i=1 1 + x1 x12 x3 x1 x2 x3
 Q3 
1
I x1 x2 x3 − t i=1 (1 + xi )2 dx1 dx2 dx3
=   .
(2iπ)3 γ 3 x2 x2 x2 − t Q3 (1 + x )2 1 − t Q3 (1 + x )2
1 2 3 i=1 i i=1 i

The partial integral with respect to x3 along the circle |x3 | = 12 is the sum of the
residues of the rational function being integrated at the poles whose modulus is
less than 12 . When |t| is small and |x1 | = |x2 | = 12 , the poles coming from the
Q3
factor x21 x22 x23 − t i=1 (1 + xi )2 all have a modulus that is smaller than 12 : they
are asymptotically proportional to |t|1/2 . In contrast, the poles coming from the
Q3
factor 1 − t i=1 (1 + xi )2 behave like |t|−1/2 and have all a modulus that is bigger
than 12 . In particular, any two poles that come from the same factor are either both
asymptotically small or both asymptotically large. This implies that the partial
integral is a rational function of t, x1 and x2 ; and we compute that
I
1 x1 x2 dx1 dx2
y(t) = .
(2iπ)2 γ×γ x21 x22 − t(1 + x1 )2 (1 + x2 )2 (1 − x1 x2 )2
This formula echoes the original proof of Dixon (1891) in which the left-hand side
of (3) is expressed as the coefficient of (xy)4n in ((1 − y 2 )(1 − z 2 )(1 − y 2 z 2 ))2n .
Using any algorithm that performs definite integration of rational functions (Chyzak
4 A. BOSTAN, P. LAIREZ, AND B. SALVY

2000; Koutschan 2010; Lairez 2016) reveals a differential equation satisfied by y(t):
t(27t + 1)y 00 + (54t + 1)y 0 + 6y = 0.
Looking at the coefficient of tn in this equality leads to the recurrence relation
3(3n + 2)(3n + 1)un + (n + 1)2 un+1 = 0,
P2n
3
where un = k=0 (−1)k 2n k . Since u0 = 1, it leads to a proof of Dixon’s identity
by induction on n. The treatment above differs in one important aspect from what
follows: the use of genuine integrals and explicit integration paths rather than
formal residues that will be introduced in §2.
Comparison with creative telescoping. As mentioned above, the computation
of multiple binomial sums can be handled by the method of creative telescoping.
The amount of work in this direction is considerable and we refer the reader to
surveys (Chyzak 2014; Koutschan 2013). In the specific context of multiple sums,
the most relevant works are those of Wegschaider (1997), Chyzak (2000), Apagodu
and Zeilberger (2006), and Garoufalidis and Sun (2010). We show on the example
of Dixon’s identity how the method of creative telescoping and the method of
generating functions differ fundamentally even on a single sum.

3
Let un,k = (−1)k 2n
k . This bivariate sequence satisfies the recurrence relations
(2n + 2 − k)3 (2n + 1 − k)3 un+1,k − 8(1 + n)3 (1 + 2n)3 un,k = 0
(4)
and (k + 1)3 un,k+1 + (2n − k)3 un,k = 0.
With these relations as input, Zeilberger’s algorithm finds the sequence
P (n, k)
vn,k = un,k ,
2(2n + 2 − k)3 (2n + 1 − k)3
where P (n, k) = k 3 (9k 4 n − 90k 3 n2 + 348k 2 n3 − 624kn4 + 448n5
+ 6k 4 − 132k 3 n + 792k 2 n2 − 1932kn3 + 1760n4 − 48k 3 + 594k 2 n
− 2214kn2 + 2728n3 + 147k 2 − 1113kn + 2084n2 − 207k + 784n + 116),
that satisfies
(5) 3(3n + 2)(3n + 1)un,k + (n + 1)2 un+1,k = vn,k+1 − vn,k .
Whatever the way the sequence vn,k is found, it is easy to check the telescopic
relation (5): using the recurrence relations for un,k , each of the four terms in (5)
rewrites in the form R(n, k)un,k , for some rational function R(n, k). However,
for some specific values of n and k, the sequence vn,k is not defined, due to the
denominator.
P2n
To deduce a recurrence relation for an = k=0 un,k , it is desirable to sum the
telescopic relation (5), over k, from 0 to 2n + 2. Unfortunately, that would hit the
forbidden set where vn,k is not defined. We can only safely sum up to k = 2n − 1.
Doing so, we obtain that
2n−1
X 2n−1
X
3(3n + 2)(3n + 1) un,k + (n + 1)2 un+1,k = vn,2n − vn,0 ,
k=0 k=0
and then

3(3n + 2)(3n + 1)(an − un,2n ) + (n + 1)2 (an+1 − un+1,2n+2 − un+1,2n+1 − un+1,2n )

= n3 (8n5 + 52n4 + 146n3 + 223n2 + 185n + 58).
 MULTIPLE BINOMIAL SUMS 5

It turns out that 3(3n + 2)(3n + 1)un,2n + (n + 1)2 (un+1,2n+2 + un+1,2n+1 + un+1,2n )
evaluates exactly to the right-hand side of the above identity, and this leads to
Dixon’s identity.
In this example, spurious singularities clearly appear in the range of summation.
Thus, deriving an identity such as Dixon’s from a telescopic identity such as (5) is
not straightforward and involves the certificate. A few works address this issue for
single sums (Abramov and Petkovšek 2005; Abramov 2006), but none for the case
of multiple sums: existing algorithms (Wegschaider 1997; Chyzak 2000; Garoufalidis
and Sun 2010) only give the telescopic identity without performing the summation.
A recent attempt by Chyzak et al. (2014) to check the recurrence satisfied by Apéry’s
sequence in the proof assistant Coq has shown how difficult it is to formalize this
summation step. Note that because of this issue, even the existence of a linear
recurrence for such sums can hardly be inferred from the fact that the algorithm of
creative telescoping always terminates with success.
This issue is rooted in the method of creative telescoping by the fact that
sequences are represented through the linear recurrence relations that they satisfy.
Unfortunately, this representation is not very faithful when the leading terms of
the relations vanish for some values of the indices. The method
of generating
functions avoids this issue. For example, the binomial coefficient nk is represented
unambiguously as the coefficient of xk in (1 + x)n (to be understood as a power series

when n < 0), rather than as a solution to the recurrence relations (n−k) nk = n n−1


k
and k nk = n n−1



k−1 .

Related work. The method of generating functions is classical and has been largely
studied, in particular for the approach described here by Egorychev (1984), see
also (Egorychev and Zima 2008). Egorychev’s method is a general approach to
summation, but not quite an algorithm. In this work, we make it completely effective
and practical for the class of binomial sums.
The special case when the generating functions are differentially finite (D-finite)
has been studied by Lipshitz (1989). From the effectivity point of view, the starting
point is his proof that diagonals of D-finite power series are D-finite (Lipshitz 1988).
The argument, based on linear algebra, is constructive but does not translate into
an efficient algorithm because of the large dimensions involved. This led Wilf and
Zeilberger (1992, p. 596) to comment that “This approach, while it is explicit in
principle, in fact yields an infeasible algorithm.” Still, using this construction of
diagonals, many closure properties of the sequences under consideration (called
P-recursive) can be proved (and, in principle, computed). Then, the representation
of a convergent definite sum amounts to evaluating a generating series at 1 and this
proves the existence of linear recurrences for the definite sums of all P-recursive
sequences. Abramov and Petkovšek (2002) showed that in particular the so-called
proper hypergeometric sequences are P-recursive in the sense of Lipshitz. The proof is
also constructive, relying on Lipshitz’s construction of diagonals to perform products
of sequences.
While we are close to Lipshitz’s approach, three enhancements make the method
of generating functions presented here efficient: we use more efficient algorithms for
computing multiple integrals and diagonals that have appeared in the last twenty
years (Chyzak 2000; Koutschan 2010; Bostan et al. 2013a; Lairez 2016); we restrict
ourselves to binomial sums, which makes it possible to manipulate the generating
functions through rational integral representations (see §2.2 and §2.3); and a third
6 A. BOSTAN, P. LAIREZ, AND B. SALVY

decisive improvement comes with the geometric reduction procedure for simplifying
integral representations (see §5).
Creative telescoping is another summation algorithm developed by Zeilberger
(1991b) and proved to work for all proper hypergeometric sequences (Wilf and
Zeilberger 1992). This method has the advantage of being applicable and often
efficient in practice. However, as already mentioned, it relies on certificates whose size
grows fast with the number of variables (Bostan et al. 2013a) and, more importantly,
whose summation is not straightforward, making the complete automation of the
method problematic. For proper hypergeometric sums, a different effective approach
developed by Takayama (1995) does not suffer from the certificate problem. It
consists in expressing the sum as the evaluation of a hypergeometric series and
reducing its shifts with respect to a non-commutative Gröbner basis of the contiguity
relations of the series, reducing the question to linear algebra in the finite-dimensional
quotient.
The class of sums we consider is a subclass of the sums of proper hypergeometric
sequences. We give an algorithm that avoids the computation of certificates in that
case, and relies on an efficient method to deal with the integral representations of
sums. The same approach has been recently used by Bostan et al. (2013b) on various
examples, though in a less systematic manner. Examples in §7 give an idea of the
extent of the class we are dealing with. It is a subclass of the balanced multisums,
shown by Garoufalidis (2009) to possess nice asymptotic properties. More recently,
a smaller family of binomial multisums was studied by Garrabrant and Pak (2014):
they are further constrained to be diagonals of N-rational power series.
Acknowledgments. We thank Marko Petkovšek for orienting us in some of the
literature. This work has been supported in part by FastRelax ANR-14-CE25-0018-01
and by the research grant BU 1371/2-2 of the Deutsche Forschungsgemeinschaft.

1. The algebra of binomial sums

1.1. Basic objects. For all n, k ∈ Z, the binomial coefficient nk is considered in

this work as defined to be the coefficient of xk in the formal power series (1 + x)n .
In other words,
   
n n(n − 1)(n − 2) · · · (n − k + 1) n
= for k > 0 and = 0 for k < 0.
k k! k
P0 b
For all a, b ∈ Z, we define the directed sum k=a as
P
b
b  k=a uk if a 6 b,


X 0 def
uk = 0 if a = b + 1,

k=a − Pa−1 u if a > b + 1,

k=b+1 k
Pb
in contrast with the usual convention that k=a uk = 0 when a > b. This implies
the following flexible summation rule for directed sums:
b c c
X 0 X 0 X 0
uk + uk = uk , for all a, b, c ∈ Z,
k=a k=b+1 k=a

and also the geometric summation formula

b
X 0 ra − rb+1
rk = for any r 6= 1,
1−r
k=a
 MULTIPLE BINOMIAL SUMS 7

valid independently of the relative position of a and b.

Let K be a field of characteristic zero, and let d > 1. We denote by Sd the K-
algebra of sequences Zd → K, where addition and multiplication are performed
component-wise. Elements of Zd are denoted using underlined lower case letters,
such as n. The algebra Sd may be embedded in the algebra of all functions ZN → K
¯
by sending a sequence u : Zd → K to the function ũ defined by
ũ(n1 , n2 , . . . ) = u(n1 , . . . , nd ).
Let S be the union of the Sd ’s in the set of all functions ZN → K. For u ∈ S
and n ∈ Zd , the notation un represents un,0,0,... . These conventions let us restrict
¯ ¯ ¯
or extend the number of indices as needed without keeping track of it.

Definition 1.1. The algebra of binomial sums over the field K, denoted B, is the
smallest subalgebra of S such that:
(a) The Kronecker delta sequence n ∈ Z 7→ δn , defined by δ0 = 1 and δn = 0
if n 6= 0, is in B.
(b) The geometric sequences n ∈ Z 7→
C n , for all C ∈ K \ {0}, are in B.
(c) The binomial sequence (n, k) 7→ nk (an element of S2 ) is in B.
(d) If λ : Zd → Ze is an affine map and if u ∈ B, then n ∈ Zd 7→ uλ(n) is in B.
¯ ¯
(e) If u ∈ B, then the following directed indefinite sum is in B:
m
X 0
(n, m) ∈ Zd × Z 7→ un,k .
¯ k=0
¯

Let us give a few useful examples. All polynomial sequences Zd → K are in B,

because B is an algebra and because the sequence n 7→ n1 = n is in B, thanks to

points (c) and (d) of the definition.

Let (Hn )n∈Z be the sequence defined by Hn = 1 if n > 0 and Hn = 0 if n < 0. It
P0 n
is a binomial sum since Hn = k=0 δk . As a consequence, we obtain the closure
of B by usual (indefinite) sums since
m m
X X 0
un,k = Hm un,k .
¯ ¯
k=0 k=0

By combining the rules (d) and (e) of the definition, we also obtain the closure
of B under sums whose bounds depend linearly on the parameter: if u ∈ B and if λ
and µ are affine maps : Zd → Z, then the sequence
µ(n)
X ¯ 0
(6) n ∈ Zd 7→ un,k
¯ ¯
k=λ(n)
¯
is a binomial sum.
See also Figure 1 for an example of a classical binomial sum.

1.2. Characterization of binomial sums. A few simple criteria make it possible

to prove that a given sequence is not a binomial sum. For example:
• The sequence (n!)n>0 is not a binomial sum. Indeed, the set of sequences
that grow at most

exponentially is closed under the rules that define binomial
sums. Since nk 6 2|n|+|k| , every binomial sum grows at most exponentially;
but this is not the case for the sequence (n!)n>0 .
8 A. BOSTAN, P. LAIREZ, AND B. SALVY

     
k k k
j j j

 3
k
j
(e)
  m  3
n X 0 k

k j=0
j

n 7→ n + k (d) (d) m 7→ k
    k  3
n n+k X 0 k

k k j
j=0

   k  3
n n + k X0 k
k k j=0
j

(e)
m    k  3
X 0 n n + k X0 k
δn
k k j=0
j
k=0

(e) (d) m 7→ n
n    k  3
X 0 n n + k X0 k
Hn
k k j=0
j
k=0

n    k  3
X n n + k X0 k
k k j=0
j
k=0

Figure 1. Construction of a binomial sum. The last step replaces the

directed sum by a usual sum in order to have nonzero values for n > 0
only.

• The sequence of all prime integers is not a binomial sum because it does not
satisfy any nonzero linear recurrence relation with polynomial coefficients
(Flajolet et al. 2005), whereas every binomial sum does, see Corollary 3.6.
• The sequence (1/n)n>1 is not a binomial sum. To prove this, we can
easily reduce to the case where K is a number field and then study the
denominators that may appear in the elements of a binomial sum. One may
introduce new prime divisors in the denominators only by multiplying with
a scalar or with rule (b), so that the denominators of the elements of a given
binomial sum contain only finitely many prime divisors. This is clearly not
the case for the sequence (1/n)n>1 .
 MULTIPLE BINOMIAL SUMS 9

By the same argument, the first sum of Eq. (1) is not a binomial sum.
In-
deed, by creative telescoping, it can be shown to equal (2n + 1)4n+1 / 2n+2
n+1 +
1/3 and thus all prime numbers appear as denominators.
• The sequence (un )n>0 defined by u0 = 0, u1 = 1 and by the recurrence (2n +
1)un+2 − (7n + 11)un+1 + (2n + 1)un = 0 is not a binomial √ sum. This follows
√
from the asymptotic estimate
p un ∼ C · (4/(7 − 33)) · n 75/44 , with C ≈
n

0.56, and the fact that 75/44 is not a rational number (Garoufalidis 2009,
Theorem 5).
These criteria are, in substance, the only ones that we know to prove that
a given sequence is not a binomial sequence. Conjecturally, they characterize
univariate binomial sums. Indeed, we will see that the equivalence between univariate
binomial sums and diagonals of rational functions (Theorem 3.5) leads, among
many interesting corollaries, to the following reformulation of a conjecture due to
Christol (1990, Conjecture 4): “any sequence (un )n>0 of integers that grows at most
exponentially and that is solution of a linear recurrence equation with polynomial
coefficients is a binomial sum.”

2. Generating functions
To go back and forth between sequences and power series, one can use convergent
power series and Cauchy’s integrals to extract coefficients, as shown in the intro-
duction with Dixon’s identity, or one can use formal power series. Doing so, the
theory keeps close to the actual algorithms and we avoid the tedious tracking of
convergence radii. However, this requires the introduction of a field of multivariate
formal power series that makes it possible, by embedding the rational functions
into it, to define what the coefficient of a given monomial is in the power series
expansion of an arbitrary multivariate rational function. We choose here to use the
field of iterated Laurent series. It is an instance of the classical field of Hahn series
when the value group is Zn with the lexicographic order. We refer to Xin (2004) for
a complete treatment of this field and we simply gather here the main definitions
and results.
2.1. Iterated Laurent series. For a field A, let A((t)) be the field of univariate
formal Laurent power series with coefficients in A. For d > 0, let Ld be the field of
iterated formal Laurent power series K((zd ))((zd−1 )) · · · ((z1 )). It naturally contains
the field of rational functions in z1 , . . . , zd . For n = (n1 , . . . , nd ) ∈ Zd , let z ¯n
¯ ¯
denote the monomial z1n1 · · · zdnd , and for f ∈ Ld , let [z ¯n ]f denote the coefficient
n ¯ n
of z ¯n in f , that is the coefficient of zd d in the coefficient of zd−1 in [. . . ] in the
d−1

¯
coefficient of z1n1 of f . An element of f ∈ Ld is entirely characterized by its coefficient
function n ∈ Zd 7→ [z ¯n ]f . On occasion, we will write [z1n ]f , this means [z1n z20 · · · zd0 ]f ,
¯ ¯
e.g., [1]f means [z10 · · · zd0 ]f ; more generally we write [m]f , where m is a monomial.
In any case, the bracket notation always yields an element of K.
Let ≺ denote the lexicographic ordering on Zd . For n, m ∈ Zd we write z ¯n ≺ z m ¯
¯ ¯ ¯ ¯
if m ≺ n (mind the inversion: a monomial is larger when its exponent is smaller).
¯ ¯
In particular z1 ≺ · · · ≺ zd ≺ 1. With an analytic point of view, we work in an
infinitesimal region where zd is infinitesimally small and where zi is infinitesimally
small compared to zi+1 . The relation z ¯n ≺ z m n
¯ means that |z ¯ | is smaller than |z ¯ |
m
¯ ¯ ¯ ¯
in this region.
For f ∈ Ld , the support of f , denoted supp(f ), is the set of all n ∈ Zd such
¯
that [z ¯n ]f is not zero. It is well-known (e.g. Xin 2004, Prop. 2-1.2) that a function ϕ :
¯
10 A. BOSTAN, P. LAIREZ, AND B. SALVY

Zd → K is the coefficient function of an element of Ld if and only if the support of ϕ

is well-ordered for the order ≺ (that is to say every subset of the support of ϕ has a
least element). The valuation of a non zero f ∈ Ld , denoted v(f ), is the smallest
element of supp(f ) ⊂ Zd for the ordering ≺. Since supp(f ) is well-ordered, v(f )
does exist. The leading monomial of f , denoted lm(f ), is z v(f ) ; it is the largest
¯
monomial that appears in f .
For 1 6 i 6 d, the partial derivative ∂i = ∂/∂zi with respect to the variable zi ,
defined for rational functions, extends to a derivation in Ld such that [z ¯n ]∂i f =
¯
(ni + 1)[z ¯n ](f /zi ) for any f ∈ Ld .
¯
Let w1 , . . . , wd be monomials in the variables z1 , . . . , ze . When f is a rational
function, the substitution f (w1 , . . . , wd ) is defined in a simple way. For elements
of Ld , it is slightly more technical. Let ϕ : Zd → Ze be the additive map defined
by ϕ(n) = v(w¯n ) (recall that v denotes the valuation). Conversely, this map entirely
¯ ¯
determines the monomials w1 , . . . , wd . When ϕ is strictly increasing (and thus
injective) with respect to the lexicographic ordering, we define f ϕ as the unique
element of Le such that
( −1
n ϕ [z ϕ (¯n) ]f if n ∈ ϕ(Zd ),
[z ¯ ]f = ¯ ¯
¯ 0 otherwise.
The map f ∈ Ld 7→ f ϕ ∈ Le is a field morphism. In particular, for a monomial z ¯n ,
ϕ ¯
we check that (z ¯n ) = z ϕ(¯n) . When f is a rational function, f ϕ coincides with the
¯ ¯
usual substitution f (w1 , . . . , wd ).
Another important construction is the sum of geometric sequences. Let f ∈ Ld
be a Laurent power series with lm(f ) ≺ 1. The set of all n ∈ Zd such that
¯
there is at least one k ∈ N such that [z ¯n ]f k 6= 0 is well-ordered (Neumann 1949,
¯d
Theorem 3.4). Moreover, for any n ∈ Z , the coefficient [z ¯n ]f k vanishes for all but
¯ ¯
finitely many k ∈ N (Neumann 1949, Theorem 3.5). The following result is easily
deduced.
Lemma 2.1. Let f ∈ Ld and let z ¯n be a monomial. If lm(f ) ≺ 1, then [z ¯n ]f k = 0
¯ ¯
for all but finitely many k ∈ N and moreover
1 X
[z ¯n ] = [z ¯n ]f k .
¯ 1−f k>0
¯

In what follows, there will be variables z1 , . . . , zd , denoted z1`d (and sometimes

under different names) and we will denote z1 ≺ · · · ≺ zd the fact that we consider
the field Ld with this ordering to define coefficients, residues, etc. The variables are
always ordered by increasing index, and t1`d ≺ z1`e denotes t1 ≺ · · · ≺ td ≺ z1 ≺
· · · ≺ ze . An element of Ld ∩ K(z1`d ) is called a rational Laurent series, and an
element of KJz1`d K ∩ K(z1`d ) is called a rational power series.
Example 1. Since Ld is a field containing all the rational functions in the vari-
ables z1 , . . . , zd , one may define the coefficient of a monomial in a rational func-
tion. However, it strongly depends on the ordering of the variables. For example,
in L2 = K((z2 ))((z1 )), the coefficient of 1 in z2 /(z1 + z2 ) is 1 because
z2 1
= 1 − z1 + O(z12 ),
z1 + z2 z2
whereas the coefficient of 1 in z1 /(z1 + z2 ) is 0 because
z1 1
= z1 + O(z12 ).
z1 + z2 z2
 MULTIPLE BINOMIAL SUMS 11

2.2. Binomial sums and coefficients of rational functions. Binomial sums

can be obtained as certain extractions of coefficients of rational functions. We will
make use of sequences of the form u : n ∈ Zd 7→ [1] (R0 R1n1 · · · Rdnd ), where R0`d
¯
are rational functions of ordered variables z1`r . They generalize several notions.
For example if R0 ∈ K(z1 , . . . , zr ) and Ri = 1/zi , then u is simply the coefficient
function of R0 ; and if d = 1, R0 ∈ K(z1 , . . . , zr ) and R1 = 1/(z1 · · · zr ), then u is
the sequence of the diagonal coefficients of R0 .

Theorem 2.2. Every binomial sum is a linear combination of finitely many se-
quences of the form n ∈ Zd 7→ [1] (R0 R1n1 · · · Rdnd ), where R0`d are rational functions
¯
of ordered variables z1`r , for some r ∈ N.

Proof. It is clear that δn = [1]z n , that C n = [1]C n and that nk = [1](1 + z)n z −k ,

for all n, k ∈ Z. Thus, it is enough to prove that the vector space generated by the
sequences of the form [1] (R0 R1n1 · · · Rdnd ) is a subalgebra of S which is closed by the
rules defining binomial sums, see §1.1.
0
First, it is closed under product: if R0`d and R0`d are rational functions in the
0
variables z1`r and z1`r0 respectively, then
d
! d
! d
!
0 ni
Y Y Y
ni 0 0 0 ni
[1] R0 Ri [1] R0 Ri = [1] R0 R0 (Ri Ri ) ,
i=1 i=1 i=1
0
with the ordering z1`r ≺ z1`r 0, for example. Then, to prove the closure under
change of variables, it is enough to reorder the factors:
d Pe ! 
d
! e d
!nj 
Y aij nj +bi Y Y Y a
[1] R0 Ri j=1 = [1]  R0 Ribi Ri ij .
i=1 i=1 j=1 i=1

Only the closure under partial sum remains. Let u be a sequence of the form
d
!
Y
d k ni
(n, k) ∈ Z × Z 7→ [1] T R0 Ri ,
¯ i=1
where T and R0`d are rational functions. If T = 1, then
m d
! d
!
X 0 Y (1 + v)R0
Y
un,k = [1] (m + 1)R0 Rini = [1] (1 + v)m Rini ,
¯
i=1
v i=1
k=0

where v is a new variable, because [v](1 + v)m+1 = m + 1. If T 6= 1, then

m d
! d
!
X 0 R0 Y ni R0 T m Y ni
un,k = [1] R − [1] T Ri ,
¯ 1 − T i=1 i 1−T i=1
k=0
which concludes the proof. 

The previous proof is algorithmic and Algorithm 1 p. 14 proposes a variant

which is more suitable for actual computations—see §2.4. It is a straightforward
rewriting procedure and often uses more terms and more variables than necessary.
The geometric reduction procedure, see §5, handles this issue.
12 A. BOSTAN, P. LAIREZ, AND B. SALVY

Example 2. Following step by step the proof of Theorem 2.2, we obtain that
n   n
(1 + z1 )(1 + z2 )2
 
X n n+k 1 + z2
= [1]
k k 1 + z1 − z1 z2 z1 z2
k=0
z1 z2 n
− [1] ((1 + z1 )(1 − z2 )) .
1 + z1 − z1 z2
We notice that the second term is always zero.

Corollary 2.3. For any binomial sum u : Zd → K, there exist r ∈ N and a rational
function R(t1`d , z1`r ), with t1`d ≺ z1`r , such that for any n ∈ Zd ,
( ¯
u n if n 1 , . . . , n d > 0,
[t¯n z¯0 ]R = ¯
¯ ¯ 0 otherwise.

(See Proposition 3.10 below for a converse.)

Proof. It is enough to prove the claim for a generating set of the vector space of all
binomial sums. In accordance with Theorem 2.2, let u : Zd → K be a sequence of
the form n ∈ Zd 7→ [1] (S0 S1n1 · · · Sdnd ), where S0`d are rational functions of ordered
¯
variables z1`r . Let R(t1`d , z1`r ) be the rational function
d
Y 1
R = S0 .
i=1
1 − ti Si

Lemma 2.1 implies that for any n ∈ Zd

X¯
[t¯n z¯0 ]R = [t¯n z¯0 ] S0 · (t1 S1 )k1 · · · (td Sd )kd .


¯ ¯ k ,...,k >0
¯ ¯
1 d

Since the variables ti do not appear in the Si ’s, the coefficient in the sum is not zero
only if n = k. In particular, if n is not in Nd , then [t¯n z¯0 ]R = 0. And if n ∈ Nd , then
¯ ¯ ¯ ¯ ¯ ¯
[t¯n z¯0 ]R = [t¯n z¯0 ] (S0 · (t1 S1 )n1 · · · (td Sd )nd ) = [1] (S0 S1n1 · · · Sdnd ) = un ,
¯ ¯ ¯ ¯ ¯
which concludes the proof. 

2.3. Residues. The notion of residue makes it possible to represent the full gen-
erating function of a binomial sum. It is a key step toward their computation.
For f ∈ Ld and 1 6 i 6 d, the formal residue of f with respect to zi , denoted reszi f ,
is the unique element of Ld such that
(
[z ¯n ](zi f ) if ni = 0,
[z ¯n ](res f ) = ¯
¯ zi 0 otherwise.
This is somehow the coefficient of 1/zi in f , considered as a power series in zi
(though f is not a Laurent series in zi since it may contain infinitely many negative
powers of zi ). When f is a rational function, care should be taken not to confuse the
formal residue with the classical residue at zi = 0, which is the coefficient of 1/zi in
the partial fraction decomposition of f with respect to zi . However, the former can
be expressed in terms of classical residues, see §5, and like the classical residues, it
vanishes on derivatives:

Lemma 2.4. reszi ∂i f = 0 for all f ∈ Ld .

If α is a set of variables {zi1 , . . . , zir }, let resα f denote the iterated residue
reszi1 · · · reszir f . It is easily checked that this definition does not depend on the
 MULTIPLE BINOMIAL SUMS 13

order in which the variables appear. This implies, together with Lemma 2.4, the
following lemma:
P

Lemma 2.5. resα v∈α ∂v fv = 0, for any family (fv )v∈α of elements of Ld .

The following lemma will also be useful:

Lemma 2.6. Let α, β ⊂ {z1`d } be disjoint sets of variables. If f ∈ Ld does not
depend on the variables in β and if g ∈ Ld does not depend on the variables in α,
then


res f res g = res (f g).
α β α∪β

Proof. resα∪β f g = resα resβ f g = resα f resβ g = resα f resβ g . 
The order on the variables matters. For example, let F (x, y) be a rational
function, with x ≺ y. The residue resx F is a rational function of y. Indeed, if −n
is the exponent of x in lm(F ), which we assume to be negative, then
 n−1 
1 ∂ n
res F = x F .
x (n − 1)! ∂xn−1 x=0
In contrast, the residue with respect to y (or any other variable which is not the
smallest) is not, in general, a rational function. It is an important point because we
will represent generating series of binomial sums—which need be neither rational
nor algebraic—as residues of rational functions.
1
Example 3. Let F = xy(y2 +y−x) . Assume that y ≺ x, that is to say, we work in
the field K((x))((y)). We compute that resy F = − x12 because
1 (y + 1)
F =− 2
+ 2 2
x y x (y + y − x)
and the second term does not contain any negative power of y.
On the other hand, if we assume that x ≺ y, then resy F is not a rational function
anymore. Using Proposition 5.2, we can compute that
1 2 1
res F = − 2 + 2 1/2
= − + 3 − 10x + 35x2 + O(x3 ).
y x x + 4x − x(1 + 4x) x
Residues of rational functions can be used to represent any binomial sum; it is
the main point of the method and it is the last corollary of Theorem 2.2. Following
Egorychev we call them integral representations, or formal integral representations,
to emphasize the use of formal power series and residues rather than of analytic
objects.
Corollary 2.7 (Integral representations). For any binomial sum u : Zd → K, there
exist r ∈ N and a rational function R(t1`d , z1`r ), with t1`d ≺ z1`r , such that
X
un t¯n = res R.
¯¯ z1`r
n∈Nd
¯
In other words, the generating function of the restriction to Nd of a binomial sum
is a residue of a rational function.
Proof. By Corollary 2.3, there exist r ∈ N and a rational function R̃(t1`d , z1`r )
such that un = [t¯n z¯0 ]R̃ for all n ∈ Nd . Let R be R̃/(z1 · · · zr ). Then
¯ ¯ ¯ X ¯ X
un t¯n = [t¯n z¯0 ]R̃ t¯n = res R.



d
¯
d
¯ ¯ ¯ z1`r
n∈N n∈N
¯ ¯
14 A. BOSTAN, P. LAIREZ, AND B. SALVY

Algorithm 1. Constant term representation of a binomial sum

Input: A binomial sum u : Zd → K given as an abstract syntax tree
n1 nd
Output: P (n1`d ; z1`r ) a linear combination of expressions of the form nα¯ R0 R1 · · · Rd
d ¯
where α ∈ N and R0`d are rational functions of z1`r .
¯
Specification: un = [1]P (n1`d ; z1`r ) for any n ∈ Zd .
¯ ¯
function SumToCT(u)
if un = δn1 then return z1n1
¯
else if un = an1
for some a ∈ K then return an1
¯
else if un = nn12 then return (1 + z1 )n1 /z1n2
¯
else if un = vn + wn then return SumToCT(v) + SumToCT(w)
¯ ¯ ¯
else if un = vn wn then
¯ ¯ ¯
P (n1`d ; z1`r ) ← SumToCT(v)
Q(n1`e ; z1`s ) ← SumToCT(w)
return P (n1`d ; z1`r )Q(n1`e ; zr+1`r+s )
else if un = vλ(n) for some affine map λ : Zd → Ze then
¯ ¯
P (n1`d ; z1`r ) ← SumToCT(v)
return P (λ(n1`d ); z1`r )
P0 n1
else if un1 ,n2 ,... = k=0 vk,n2 ,... then
P (n1`d ; z1`r ) ← SumToCT(v)
Compute Q(n1`d ; z1`r ) s.t. Q(n1 + 1, n2`d ; z1`r ) − Q(n1`d ; z1`r ) = P
. See Lemma 2.8.
return Q(n1 + 1, n2`d ; z1`r ) − Q(0, n2`d ; z1`r )

In §3, we prove an equivalence between univariate binomial sums and diagonals

of rational functions that are a special kind of residues.

2.4. Algorithms. The proofs of Theorem 2.2 and its corollaries are constructive,
but in order to implement them, it is useful to consider not only sequences of
the form [1](R0 R¯n ) but also sequences of the form [1](nα n
¯ R0 R¯ ), for some rational
¯ d ¯ ¯
functions R0`d and α ∈ N . They bring two benefits. First, polynomial factors often
¯
appear in binomial sums and it is convenient to be able to represent them without
adding new variables. Second, we can always perform sums without adding new
variables, contrary to what is done in the case T = 1 of the proof of Theorem 2.2.

Lemma 2.8. For any A = 6 0 in some field k, and any α ∈ N, there exists a
polynomial P ∈ k[n] of degree at most α + 1 such that for all n ∈ Z,
nα An = P (n + 1)An+1 − P (n)An .

Proof. It is a linear system of α + 2 equations, one for each monomial 1, n, . . . , nα+1 ,

in the α + 2 coefficients of P . If A 6= 1, the homogeneous equation P (n + 1)A = P (n)
has no solutions, so the system is invertible and has a solution. If A = 1, then
the equation P (n + 1)A = P (n) has a one-dimensional space of solutions but the
equation corresponding to the monomial nα+1 reduces to 0 = 0, so the system again
has a solution. 

Algorithm 1 implements Theorem 2.2 in this generalized setting. The input data,
a binomial sum, is given as an expression considered as an abstract syntax tree, as
depicted in Figure 1. The implementation of Corollaries 2.3 and 2.7 must also be
adjusted to handle the monomials in n.
¯
 MULTIPLE BINOMIAL SUMS 15

Algorithm 2. Integral representation of a binomial sum

Input: A binomial sum u : Zd → K given as an abstract syntax tree
Output: A rational
P function R(t1`d , z1`r )
n
Specification: n∈Nd
u n t ¯ = resz1`r R, where t1`d ≺ z1`r
¯ ¯

function SumToRes(u)
Xm
n1 nd
nα¯ Rk,0 Rk,1 · · · Rk,d ← SumToCT(u)
k

k=1
¯
m
1 X
return Rk,0 Fα1 (t1 Rk,1 ) · · · Fαd (td Rk,d )
z1 . . . zr
k=1
where F0 (z) = 1
1−z and Fα+1 (z) = zFα0 (z)

To this effect, for α ∈ N, let

 α
def
X
α n d 1
Fα (z) = n z = z · .
dz 1−z
n>0
n1 nd
Now let us consider a sequence u : n ∈ Zd 7→ [1] (nα¯ S0 S1 · · · Sd ), where S0`d are
¯ ¯ d
rational functions of ordered variables z1`r and α ∈ N Let
¯
d d
Y 1 Y
R̃ = S0 and R = S0 Fαi (ti Si ),
i=1
1 − ti Si i=1
where t1`d ≺ z1`r . By definition of the Fα ’s, we check that
 α1  αd
∂ ∂
R = t1 · · · td · R̃,
∂t1 ∂td
so that [t¯n z¯0 ]R = nα n 0 d
¯ [t¯ z¯ ]R̃, for n ∈ Z , by definition of the differentiation in Lr . In
¯ ¯ ¯ ¯ ¯ ¯
the proof of Corollary 2.3, we checked that [t¯n z¯0 ]R̃ = [1]S0 S1n1 · · · Sdnd , for n ∈ Nd ,
¯ ¯ ¯
and 0 otherwise. Therefore, un = [t¯n z¯0 ]R, for n ∈ Nd . This gives an implementation
¯ ¯ ¯
of Corollaries 2.3 and 2.7 in the generalized setting. Algorithm 2 sums up the
procedure for computing integral representations.

2.5. Analytic integral representations. When K is a subfield of C, then formal

residues can be written as integrals.
Proposition 2.9. Let R(t1`d , z1`e ) be a rational function whose denominator does
not vanish when t1 = · · · = td = 0. There exist positive real numbers s1`d and r1`e
such that on the set (t1`d ) ∈ Cd ∀i, |ti | 6 si , the power series resz1`e R ∈ CJt1`d K
converges and I
1
res R = R(t1`d , z1`e )dz1`e ,
z1`e (2πi)e γ
where γ = {z ∈ Ce | ∀1 6 i 6 e, |zi | = ri }.
Proof. When R is a Laurent monomial, the equality follows from Cauchy’s integral
formula. By linearity, it still holds when R is a Laurent polynomial.
In the general case, let R be written as a/f , where a and f are polynomials. We
may assume that the leading coefficient of f is 1 and so f decomposes as lm(f )(1 − g)
where g is a Laurent polynomial with monomials ≺ 1. The hypothesis that f does not
vanish when t1 = · · · = td = 0 implies that lm(f ) depends only on the variables z1`e
and that g contains no negative power of the variables t1`d . This and the fact
that all monomials of g are ≺ 1 imply that there exist positive real numbers s1`d
16 A. BOSTAN, P. LAIREZ, AND B. SALVY

and r1`e such that |g(t1`d , z1`e )| 6 12 if |ti | 6 si and |zi | = ri . For example, we
can take si = exp(− exp(N/i)) and ri = exp(− exp(N/(d + i))), for some large
enough N , because
exp(− exp(N/i))p = o (exp(− exp(N/j))q ) , N → ∞,
for any p, q > 0 and i < j. On the one hand
ag k
X  
res R = res ,
z1`e z1`e lm(f )
k>0

by Lemma 2.1, and on the other hand, if |ti | 6 si , for 1 6 i 6 d then

I X I ag k
R(t1`d , z1`e )dz1`e = dz1`e .
γ lm(f )
k>0 γ

where γ = {z ∈ Ce | ∀1 6 i 6 e, |zi | = ri }, because the sum k

P
k>0 g converges
uniformly on γ, since |g| 6 21 . And the lemma follows from the case where R is a
Laurent polynomial. 

3. Diagonals
Let R(z1`d ) = n∈Nd an z ¯n be a rational power series in K(z1 , . . . , zd ). The
P
¯ ¯¯
diagonal of R is the univariate power series
def
X
diag R = an,...,n tn .
n>0

Diagonals have be introduced to study properties of the Hadamard product of power

series (Cameron and Martin 1938; Furstenberg 1967). They also appear in the theory
of G-functions (Christol 1988). They can be written as residues: with t ≺ z2`d , it
is easy to check that
 
1 t
(7) diag R = res R , z2 , . . . , zd .
z2 ,...,zd z2 · · · zd z2 · · · zd
Despite their simplistic appearance, diagonals have very strong properties, the first
of which is differential finiteness:

Theorem 3.1 (Christol 1985, Lipshitz 1988). Let R(z1`d ) be a rational power
series in K(z1 , . . . , zd ). There exist polynomials p0`r ∈ K[t], not all zero, such
that pr f (r) + · · · + p1 f 0 + p0 f = 0, where f = diag R.

We recall that a power series f ∈ KJtK is called algebraic if there exists a nonzero
polynomial P ∈ K[x, y] such that P (t, f ) = 0.

Theorem 3.2 (Furstenberg 1967). A power series f ∈ KJtK is algebraic if and only
if f is the diagonal of a bivariate rational power series.

Diagonals of rational power series in more than two variables need not be algebraic,
as shown by  
X (3n)! 1
n
t = diag .
n!3 1 − z1 − z2 − z3
n>0
However, the situation is simpler modulo any prime number p. The following result
is stated by Furstenberg (1967, Theorem 1) over a field of finite characteristic;
since reduction modulo p and diagonal extraction commute, we obtain the following
statement:
 MULTIPLE BINOMIAL SUMS 17

Theorem 3.3 (Furstenberg 1967). Let f ∈ QJtK be the diagonal of a rational power
series with rational coefficients. Finitely many primes may divide the denominator of
a coefficient of f . For all prime p except those, the power series f (mod p) ∈ Fp JtK
is algebraic.
Example 4. It is easy to check that the series f = n>0 (3n)! n
P
n!3 t satisfies
− 14
f ≡ (1 + t) mod 5,

1
2 −6
f ≡ 1 + 6t + 6t mod 7,
− 1
f ≡ 1 + 6t + 2t2 + 8t3 10 mod 11, etc.

The characterization of diagonals of rational power series remains largely an

open problem, despite very recent attempts (Christol 2015). A natural measure
of the complexity of a power series f ∈ QJtK is the least integer n, if any, such
that f is the diagonal of a rational power series in n variables. Let N (f ) be this
number, or N (f ) = ∞ if there is no such n.1 It is clear that N (f ) = 1 if and
only if f is rational. According to Theorem 3.2, N (f ) = 2 if and only if f is
algebraic and not rational. It is easy to find power series f with N (f ) = 3, for
P (3n)! n

example N n n!3 t = 3 because it is the diagonal of a trivariate rational power
series but it is not algebraic. However, we do not know of any power series f such
that N (f ) > 3.
Conjecture 3.4 (Christol 1990). Let f ∈ ZJtK be a power series with integer
coefficients, positive radius of convergence and such that pr f (r) + · · · + p1 f 0 + p0 f = 0
for some polynomials p0`r ∈ Q[t], not all zero. Then f is the diagonal of a rational
power series.
In this section, we prove the following equivalence:
Theorem 3.5. A sequence u : N → K is a binomial sum if and only if the generating
function n>0 un tn is the diagonal of a rational power series.
P

In §3.1, we prove that the generating function of a binomial sum is a diagonal

and in §3.3, we prove the converse. When writing that a sequence u : N → K
is a binomial sum, we mean that there exists a binomial sum v : Z → K whose
support in contained in N and which coincides with u on N. Theorem 3.5 can be
stated equivalently without restriction on the support: a sequence u : Z → K is a
binomial sum if and only if the generating functions n>0 un tn and n>0 u−n tn
P P

are diagonals of rational power series.

Note that Garrabrant and Pak (2014, Theorem 1.3) proved a similar, although
essentially different, result: the subclass of binomial multisums they consider corre-
sponds to the subclass of diagonals of N-rational power series.
Theorem 3.5 and the theory of diagonals of rational functions imply right away
interesting corollaries.
Corollary 3.6. If u : Z → K is a binomial sum, then there exist polynomials p0`r
in K[t], not all zero, such that pr (n)un+r + · · · + p1 (n)un+1 + p0 (n)un = 0 for
all n ∈ Z.
As mentioned in the introduction, this is a special case of a more general result
for proper hypergeometric sums that can alternately be obtained as a consequence
of the results of Lipshitz (1989) and Abramov and Petkovšek (2002).
1This notion is closely related to the grade of a power series (Allouche and Mendès France 2011).
18 A. BOSTAN, P. LAIREZ, AND B. SALVY

un tn of a
P
Corollary 3.7 (Flajolet and Soria 1998). If the generating function n
sequence u : N → K is algebraic, then u is a binomial sum.
Corollary 3.8. Let u : N → Q be a binomial sum. Finitely many primes may
divide the denominators of values of u. For all primes p except those, the generating
function of a binomial sum is algebraic modulo p.
Moreover, Christol’s conjecture is equivalent to the following:
Conjecture 3.9. If an integer sequence u : N → Z grows at most exponentially
and satisfies a recurrence pr (n)un+r + · · · + p0 (n)un = 0, for some polynomials p0`r
with integer coefficients, not all zero, then u is a binomial sum.
The proof of Theorem 3.5 also gives information on binomial sums depending on
several indices, in the form of a converse of Corollary 2.3.
Proposition 3.10. A sequence Nd → K is a binomial sum if and only if there
exists a rational function R(t1`d , z1`r ), with t1`d ≺ z1`r , such that un = [t¯n z¯0 ]R
¯ ¯ ¯
for all n ∈ Zd .
¯
Sketch of the proof. The “only if” part is Corollary 2.3. Conversely, let R(t1`d , z1`r )
be a rational function and let un = [t¯n z¯0 ]R, for n ∈ Zd . We assume that un = 0
¯ ¯ ¯ ¯ ¯
if n ∈/ Nd . Using the same technique as in §3.1, we show that there exist a
¯
rational power series S(z1`d+r ) and monomials w1`d in the variables z1`d+r such
that un = [w¯n ]S. Then, with the same technique as in §3.3, we write un as a
¯ ¯ ¯
binomial sum. 
In other words, binomial sums are exactly the constant terms of rational power
series, where the largest variables, for the order ≺, are eliminated.
3.1. Binomial sums as diagonals. Corollary 2.7 and Equation (7) provide an
expression of the generating function of a binomial sum (of one free variable) as the
diagonal of a rational Laurent series, but more is needed to obtain it as the diagonal
of a rational power series and obtain the “only if” part of Theorem 3.5.
Let u : N → K be a binomial sum. In this section, we aim at constructing a
rational power series S such that n>0 un tn = diag S. By Corollary 2.3, there
P

exists a rational function R̃(z0`r ) = fa such that un = [z0n z10 · · · zr0 ]R̃ for all n ∈ Z.
Recall the notation f ϕ , for f ∈ Lr+1 and ϕ an increasing endomorphism of Zr+1 ,
introduced in §2.1. For example, if f = z1 + z2 ∈ L2 and ϕ(n1 , n2 ) = (n1 , n1 + n2 )
then f ϕ = z1 z2 + z2 = z2 (1 + z1 ).
Lemma 3.11. For any polynomial f ∈ K[z0`r ] there exists an increasing endomor-
phism ϕ of Zr+1 such that f ϕ = C z m ¯ (1 + g), for some m ∈ N
r+1
, C ∈ K \ {0} and
¯ ¯
g ∈ K[z0`r ] with g(0, . . . , 0) = 0.
Proof. Let z¯a and z¯b be monomials of f such that z¯a ≺ z¯b . Let i be the smallest
¯ ¯ ¯ ¯
integer such that ai 6= bi . By definition ai > bi . For k ∈ N, let ϕk : Zr+1 → Zr+1
be defined by
ϕk : (n0 , . . . , nr ) ∈ Zr+1 7→ (n0 , . . . , ni , ni+1 + kni , . . . , nr + kni ) ∈ Zr+1 .
It is strictly increasing and if k is large enough then ϕk (a) > ϕk (b) componentwise,
¯ ¯
that is (z¯b )ϕk divides (z¯a )ϕk . We may apply repeatedly this argument to construct
¯ ¯
an increasing endomorphism ϕ of Zr+1 such that the leading monomial of f ϕ divides
all the monomials of f ϕ , which proves the Lemma. 
 MULTIPLE BINOMIAL SUMS 19

Let ϕ be the endomorphism given by the lemma above applied to the denominator
of R̃. For 0 6 0 6 r, let wi = ziϕ . Then
aϕ
R̃(w0`r ) = R̃ϕ = ,
C zm¯ (1 + g)
¯
because ·ϕ is a field morphism, as explained in §2.1. And by definition of R̃, we
aϕ
have un = [w0n ]R̃(w0`r ) for all n ∈ Z. Let R be the rational power series C(1+g) , so
n m
that un = [w0 ](R/z ¯ ). We now prove that we can reduce to the case where m = 0.
¯ ¯
If m 6= 0, let i be the smallest integer such that mi 6= 0. The specialization R|zi =0
¯
is a rational power series and R − R|zi =0 = zi T for some rational power series T .
For all n > 0, the coefficient of w0n in R|zi =0 /z m ¯ is zero because the exponent of zi
¯
in w0 is nonnegative while the exponent of zi in every monomial in R|zi =0 /z m
n
¯
T ¯
is −mi < 0. Thus un = [w0n ] zm ¯ /zi
, and we can replace R by T and subtract 1 to
the first nonzero coordinate ¯of m, which makes m decrease for the lexicographic
¯ ¯
ordering. Iterating this procedure leads to m = 0 and thus un = [w0n ]T for some
¯
rational power series T .
Let us write w0 as z0a0 · · · zrar , with a0`r ∈ Nr+1 . If all the ai ’s were equal to 1,
then n un tn would be the diagonal of T . First, we reduce to the case where all
P

the ai ’s are positive. If some ai is zero, then

a a
[w0n ]T = [z0a0 · · · zi−1
i−1 i+1
zi+1 · · · zrar ]T |zi =0
and we can simply remove the variable zi .
Second, we reduce further to the case where all the ai ’s are 1. Let us consider
the following rational power series:
1 X X
U= ··· T (ε0 z0 , . . . , εr zr ),
a0 · · · ar a0 ar
ε0 =1 εr =1

where the εi ranges over the ai -th roots of unity. By construction, if m is a monomial
in the ziai , then [m]U = [m]T . In particular un = [w0n ]U .
We may consider T and U as elements of the extension of the field K(z0a0 , . . . , zrar )
by the roots of the polynomials X ai − zi , for 0 6 i 6 r. By construction,
the rational function U is left invariant by the automorphisms of this exten-
sion. Thus U ∈ K(z0a0 , . . . , zrar ). Let S(z0`r ) be the unique rational function
such that U = S(z0a0 , . . . , zrar ). It is a rational power series and un = [z0n · · · zrn ]S.
n
P
Thus n>0 un t = diag S, which concludes the proof that binomial sums are
diagonals of rational functions, the “only if” part of Theorem 3.5.

3.2. Summation over a polyhedron. To prove that, conversely, diagonals of

rational power series are generating functions of binomial sums, we prove a property
of closure of binomial sums under certain summations that generalize the indefinite
summation of rule (e) of Definition 1.1. Let u : Zd+e → K be a binomial sum and
let Γ ⊂ Rd+e be a rational polyhedron, that is to say
\
Γ= x ∈ Rd+e λ(x) > 0 ,
λ∈Λ

for a finite set Λ of linear maps Rd+e → R with integer coefficients in the canonical
bases. This section is dedicated to the proof of the following:
20 A. BOSTAN, P. LAIREZ, AND B. SALVY

Proposition 3.12. If for all n ∈ Zd the set {m ∈ Re | (n, m) ∈ Γ} is bounded, then

¯ ¯ ¯ ¯
the sequence X
v : n ∈ Zd 7→ un,m 1Γ (n, m)
¯ m∈Ze
¯ ¯ ¯ ¯
¯
is a binomial sum, where 1Γ (n, m) = 1 if (n, m) ∈ Γ and 0 otherwise.
¯ ¯ ¯ ¯
Recall that H : Z → K is the sequence defined by Hn = 1 if n > 0 and 0
otherwise; it is a binomial sum, see §1.1. Thus 1Γ is a binomial sum, because 1Γ (n) =
Q ¯
λ∈Λ Hλ(n) and each factor of the product is a binomial sum, thanks to rule (d) of
¯
the definition of binomial sums. However, the summation defining v ranges over an
infinite set, rule (e) is not enough to conclude directly, neither is the generalized
summation of Equation (6).
For m ∈ Re , let |m|∞ denote max16i6e |mi |. For n ∈ Zd , let B(n) be
¯ ¯ ¯ ¯
def e
B(n) = sup {|m|∞ | m ∈ R and (n, m) ∈ Γ} ∪ {−∞} .
¯ ¯ ¯ ¯ ¯
By hypothesis on Γ, B(n) < ∞ for all n ∈ Zd .
¯ ¯
Lemma 3.13. There exists C > 0 such that B(n) 6 C(1 + |n|∞ ), for all n ∈ Zd .
¯ ¯
Proof. By contradiction, let us assume that such a C does not exist, that is there
exists a sequence pk = (nk , mk ) of elements of Γ such that |mk |∞ / |nk |∞ and |nk |∞
¯ ¯ ¯ ¯ ¯
tend to ∞. Up to¯considering a subsequence, we may assume that mk / |mk |∞ has a
e ¯ ¯
limit ` ∈ R , which is nonzero. For all α ∈ [0, 1] and k > 0, the point p0 + α(pk − p0 )
is in Γ, because Γ is convex. Let u > 0 and let αk = u/ |mk |∞ . If k is¯ large ¯enough,
¯
¯
then 0 6 αk 6 1. Moreover
p0 + αk (pk − p0 ) → (n0 , m0 + u`).
¯ ¯ ¯ k→∞ ¯ ¯
Yet Γ is closed so (n0 , m0 + u`) ∈ Γ. By definition |m0 + u`|∞ 6 B(n0 ) < ∞. This
¯ ¯ ¯ ¯
is a contradiction because ` 6= 0 and u is arbitrarily large. 

Thus, for all n ∈ Zd ,

¯ X
(8) vn = un,m 1Γ (n, m).
¯
m∈Ze
¯ ¯ ¯ ¯
¯
|m|6C(1+|n|∞ )
¯ ¯
For 1 6 i 6 d, let wi,+ be the sequence
i−1 d
def
Y Y
wni,+ = Hni (Hni −nj −1 Hni +nj −1 ) (Hni −nj Hni +nj ).
¯
j=1 j=i+1

After checking that for any a, b ∈ Z, Ha+b Ha−b = 1 if a > |b| and 0 otherwise,
we see that wni,+ = 1 if ni = |n|∞ and |nj | < |ni | for all j < i ; and 0 otherwise.
¯
Likewise, let wi,− be the sequence
i−1 d
def
Y Y
wni,− = H−ni −1 (H−ni −nj −1 H−ni +nj −1 ) (H−ni −nj H−ni +nj ),
¯
j=1 j=i+1

so that wni,− = 1 if ni = −|n|∞ < 0 and |nj | < |ni | for all j < i ; and 0 otherwise.
¯
The 2d sequences wi,+ and wi,− are binomial sums that partition 1: for all n ∈ Zd
¯
Xd Xd
1= wni,+ + wni,− .
¯ ¯
i=1 i=1
 MULTIPLE BINOMIAL SUMS 21

By design, the sum in Equation (8) rewrites as

d C(1+ni ) C(1+ni )
X X X
vn = wni,+ ··· un,m 1Γ (n, m)
¯
i=1
¯ ¯ ¯ ¯ ¯
m1 =−C(1+ni ) me =−C(1+ni )

d C(1−ni ) C(1−ni )
X X X
+ wni,− ··· un,m 1Γ (n, m),
i=1
¯ ¯ ¯ ¯ ¯
m1 =−C(1−ni ) me =−C(1−ni )

which concludes the proof of Proposition 3.12, because now the summation bounds
are affine functions with integer coefficients of n.
¯
From a practical point of view, summations over polyhedra can be handled in a
very different way, see §6.3.

3.3. Diagonals as binomial sums. We now prove the second part of Theorem 3.5:
the diagonal of a rational function is the generating function of a binomial sum.
Let R(z1`d ) be a rational power series and let us prove that the sequence defined
by un = [z1n · · · zdn ]R is a binomial sum. Since the binomial sums are closed under
linear combinations, it is enough to consider the case where the numerator of R is a
monomial and where the constant term of its denominator is 1. (It cannot be zero
since R is a power series.) Thus R has the form
z m0
R= P¯e ¯
,
ak z m
1+ ¯
k
k=1
¯
where the mk ’s have nonnegative coordinates and mk 6= 0 if k 6= 0. Let y0`e be new
¯ ¯
variables and let S be the rational power series
def y X k1 + · · · + ke 
S(y0`e ) = Pe 0 = y0 ak11 · · · ake e y1k1 · · · yeke .
1 + k=1 ak yk k 1 , . . . , k e
k∈Ne |
¯
{z }
Ck
¯

The coefficient sequence Ck of this power series is a binomial sum because the
¯
multinomial coefficient is a product of binomial coefficients:
      
k1 + · · · + ke k1 + · · · + ke k2 + · · · + ke ke−1 + ke
= ··· .
k1 , . . . , k e k1 k2 ke−1
Let Γ ⊂ R × Re be the rational polyhedron defined by
n e o
def
X
Γ = (n, k) ∈ R × Re k1 > 0, . . . , ke > 0 and m0 + ki mi = (n, . . . , n) .
¯ ¯ i=1
¯
By construction R(z1`d ) = S(z ¯ , . . . , z ¯ ), and the image of a monomial y0 y1k1 · · · yeke
m0 me
¯ ¯
is a diagonal monomial z1n · · · zdn if and only if (n, k) ∈ Γ. Thus
X ¯
[z1n · · · zdn ]R = Ck 1Γ (n, k).
k∈Ze
¯ ¯
Thanks to the positivity conditions on the mk ’s, it is obvious that Γ satisfies the
¯
finiteness hypothesis of Proposition 3.12. Thus the sequence un is a binomial sum.

4. Computing binomial sums

Computing may have different meanings. When manipulating sequences like
binomial sums, one may want, for example, to compute recurrence relations that
they satisfy, to decide their equality, to compute their asymptotic behavior or to find
simple closed-form formulas for them. The representation of the generating series
22 A. BOSTAN, P. LAIREZ, AND B. SALVY

of binomial sums as residues or diagonals of rational functions gives an interesting

tool to tackle these goals for binomial sums, though some questions remain open.
While the customary tool to compute recurrence relations satisfied by binomial
sums is creative telescoping, integral representations of binomial sums give another
approach: given a binomial sum we first compute an integral representation (this
is fast and easy, see Algorithm 1), and next we compute a differential equation
satisfied by the integral, which translates immediately into a recurrence relation
for the binomial sum. The decision problem A = B can be solved by computing
recurrence relations for A − B and checking sufficiently many initial conditions,
see §4.4 for more details.
A recurrence relation is also a good starting point for deriving the asymptotic
behavior of a univariate binomial sum (e.g. Mezzarobba and Salvy 2010). However,
some constants that determine the asymptotic behavior can be computed numerically
but it is not known how to decide their nullity, which makes it difficult to catch
the subdominant behavior. A more direct method is possible in many cases once
the generating function has been written as the diagonal of a rational power series
(Pemantle and Wilson 2013; Salvy and Melczer 2016).
The problem of simplifying binomial sums is still largely open: for example,
given the left-hand side of Strehl’s identity (see §7.2.1), it is not known how to
discover automatically the right-hand side. The only known automatic simplification
procedures consist in computing a recurrence relation and applying the algorithm of
Petkovšek (1992) to find hypergeometric solutions or the algorithm of Abramov and
Petkovšek (1994) to find D’Alembertian solutions. See §7.2 and §7.3 for numerous
examples.
The computation of a recurrence relation for a binomial sum is done in two steps.
Firstly, an integral representation is found and then a Picard-Fuchs equation is
computed.

4.1. Picard-Fuchs equations. Let L be a field of characteristic zero and let R(z1`r )
be a rational function with coefficients in L, written as R = P F where P and F
are polynomials. Let AF be the localized ring L[z1`r , F −1 ]. It is known that
the L-linear quotient space
 
def
HF = AF / ∂z∂ 1 AF + · · · + ∂z∂ r AF
is finite-dimensional (Grothendieck 1966). Let us assume that there is a derivation ∂
defined on L. It extends naturally, with ∂zi = 0, to a derivation on AF that
∂
commutes with the derivations ∂z i
, so that ∂ defines a derivation on the L-linear
space HF . Since HF is finite-dimensional, there exist c0`m ∈ L, not all zero, such
that
cm ∂ m R + · · · + c1 ∂R + c0 R ∈ ∂z∂ 1 AF + · · · + ∂z∂ r AF .
Now let us assume that L is the field of rational functions K(t1`d ) and that ∂ is the
derivation ∂t∂ i for some i. Then the operator ∂ commutes with the operator resz1`r ,
as do the multiplications by elements of L, and Lemma 2.5 implies that
m
∂ ∂
cm ∂t m f + · · · + c1 ∂t f + c0 f = 0,
i i

where f = resz1`r R, with t1`d ≺ z1`r . Differential equations that arise in this way
are called Picard-Fuchs equations.
Recall that Ld is the field of iterated Laurent series introduced in §2.1. A
series f (t1`d ) ∈ Ld is called differentially finite if the K(t1`d )-linear subspace
 MULTIPLE BINOMIAL SUMS 23

of Ld generated by the derivatives ∂ n1 +···+nd f /∂tn1 1 · · · ∂tnd d , for n1`d > 0, is finite-
dimensional. In particular, a univariate Laurent series f ∈ L1 is differentially finite
if and only if there exist m > 0 and polynomials p0`m ∈ K[t], not all zero, such
that pm f (m) + · · · + p1 f 0 + p0 f = 0. In that case, we say that f is solution of
a differential equation of order m and degree maxk deg pk . The above argument
implies the following classical theorem of which Lipshitz (1988) gave an elementary
proof.
Theorem 4.1. For any rational function R(t1`d , z1`r ) with t1`d ≺ z1`r , the
residue resz1`r R is differentially finite.
In previous work (Bostan et al. 2013a, Theorem 12; Lairez 2014, §I.35.3), we
proved the following quantitative result about the size of Picard-Fuchs equations
and the complexity of their computation. We also described an efficient algorithm
to compute Picard-Fuchs equations (Lairez 2016).
P
Theorem 4.2. Let R ∈ K(t, z1`r ) be a rational function, written as R = F , with P
and F polynomials. Let
N = max(degz1`r P + r + 1, degz1`r F ) and dt = max(degt P, degt F ).
Then resz1`r R, with t ≺ z1`r , is solution of a linear differential equation of order at
most N r and degree at most ( 58 N 3r + N r ) exp(r)dt . Moreover, this differential equa-
tion can be computed with O(exp(5r)N 8r dt ) arithmetic operations in K, uniformly
in all the parameters.
4.2. Power series solutions of differential equations. When a power series
satisfies a given linear differential equation with polynomial coefficients, one only
needs a few initial conditions to determine entirely the power series.
Let L ∈ K[t]h∂t i be a linear differential operator in ∂t with polynomial coefficients
in t. There exists a unique n ∈ Z and a unique bL ∈ K[a] such that L(ta ) =
bL (a)ta+n + o(ta+n ) for all a ∈ Z and t → 0. The polynomial bL is the indicial
polynomial of L at t = 0. For more details about the indicial polynomial, see
Ordinary Differential Equations (Ince 1944). It is easy to check that if f ∈ KJtK is
annihilated by L, then its leading monomial tn satisfies bL (n) = 0.
Proposition 4.3. Let f ∈ KJtK. If L(f ) = 0 and if [tn ]f = 0 for all n ∈ N such
that bL (n) = 0, then f = 0.
It is worth noting that the indicial equation of a Picard-Fuchs equation is very
special:
Theorem 4.4 (Katz 1970). If L is a Picard-Fuchs equation, then the degree of bL
equals the order of L and all the roots of bL are rational.
The data of a differential operator L and elements of K for each nonnegative
integer root of bL (that we will call here sufficient initial conditions) determines
entirely an element of KJtK. It is an excellent data structure for manipulating power
series (Salvy and Zimmermann 1994). For example, it lets one compute efficiently
the coefficients of the power series n un tn : the differential equation translates into
P

a recurrence relation
pr (n)un+r + . . . + p1 (n)un+1 + p0 (n)un = 0
for some polynomials p0`r ∈ K[n] and the sufficient initial conditions translate into
initial conditions for the recurrence, exactly where we need them.
24 A. BOSTAN, P. LAIREZ, AND B. SALVY

4.3. Equality test for univariate differentially finite power series. Let f ∈
KJtK be a power series given by a differential operator L ∈ K[t]h∂t i such that
L(f ) = 0, and by sufficient initial conditions. Let M be another differential
operator. We may decide if M(f ) = 0 in the following way. Firstly, we compute
the right g.c.d. of M and L: this is the operator D of the largest order such
that M = M0 D and L = L0 D for some operators M0 and L0 in K(t)h∂t i. Then, it
is enough to compute the indicial equation bL0 and to compute [tn ]D(f ) for each
nonnegative integer root n of bL0 . We will find only zeros if and only if M(f ) = 0.
Indeed, M(f ) = 0 if and only if D(f ) = 0 and since L0 (D(f )) = 0, we may apply
Proposition 4.3 to check whether D(f ) = 0 or not.
Now, let g ∈ KJtK be another power series given by a differential operator M ∈
K[t]h∂t i and sufficient initial conditions. To decide if f = g, it is enough to check
that M(f ) = 0, with the above method, and to check that the coefficients of f
and g corresponding to the nonnegative integer roots of bM are the same.

4.4. Equality test for binomial sums. Let u : Z → K be a binomial sum. Up

to considering separately the binomial sums Hn un and Hn u−n , it is enough to
look at the case where un = 0 for n < 0. In this case u is entirely determined by
its generating function f = n>0 un tn . Using Algorithm 1 we obtain an integral
P

representation of f , and then, as explained in §4.1, we obtain a differential operator L

that annihilates f . Since u is a binomial sum given explicitly, we can compute
sufficient initial conditions. Given another binomial sum v : N → K, we can
check that u = v by computing a differential operator annihilating the generating
function g of v together with sufficient initial conditions and by checking that f = g
with the method explained in §4.3.
Let us now consider the multivariate case. As above, we can reduce the equality
test for binomial sums Zd → K to the equality test for binomial sums Nd → K. And
this equality test can be reduced to the equality test for binomial sums Nd−1 → K
as follows. Let u : Nd → K be a binomial sum. It is determined by its generating
function f (t1`d ) = n∈Nd un t¯n ∈ Ld . With Algorithm 1, we compute a rational
P
¯ ¯¯
function R(t1`d , z1`r ) such that f = resz1`r f . Let K0 be the field K(t1`d−1 ).
Following §4.1, we can compute an operator L ∈ K0 [td ]h∂td i such that L(f ) = 0.
This gives a differential equation for f considered as a power series in td with
coefficients in Ld−1 . The sufficient initial conditions to determine f are given by
the power series n∈Nd−1 un,k t¯n , where k ranges over the nonnegative integer roots
P
¯ ¯
of bL . These power¯ series are the generating functions of binomial sums in d − 1
variables. This reduces the equality test for binomial sums in d variables to the
equality test for binomial sums in d − 1 variables.

5. Geometric reduction
Putting into practice the computation of binomial sums through integral repre-
sentations shows immediately that the number of integration variables is high and
makes the computation of the Picard-Fuchs equations slow. However, the rational
functions obtained this way are very peculiar. For example, their denominator often
factors into small pieces. This section presents a sufficient condition under which
a residue resv F of a rational function F is rational. This leads to rewriting an
iterated residue of a rational function, like the ones given by Corollary 2.7, into
another one with one or several variables less. This simplification procedure is very
 MULTIPLE BINOMIAL SUMS 25

efficient on the residues we are interested in and reduces the number of variables
significantly.
Conceptually the simplification procedure is simple: in terms of integrals it boils
down to computing partial integrals in specific cases where we know that they
are rational. The rational nature of a partial integral depends on the integration
cycle and integration algorithms usually forget about this cycle. Instead, they
compute the Picard-Fuchs equations—see §4.1—that annihilate a given integral
for any integration cycle. In our setting, the integration cycle underlies the notion
of residue—see §2.5. This provides a symbolic treatment that we call geometric
reduction since it takes into account the geometry of the cycle and decreases the
number of variables for which the general integration algorithm above is actually
needed. The time required by the computation is dramatically reduced by this
symbolic treatment.

5.1. Rational poles. Let us consider variables v and z1`d , where v can appear
anywhere in the variable ordering. Let F (v, z1`d ) = a/f be a rational function.
In general, resv F is not a rational function—except if v is the smallest variable,
see §2.3.
Let ρ ∈ Ld be a power series in the variables z1`d . The classical residue of F
at v = ρ, denoted Resv=ρ F , is the coefficient of 1/(v − ρ) in the partial fraction
decomposition of F as an element of Ld (v). Contrary to the formal residue resv F ,
the classical residue is always in the field generated by ρ and the coefficients of F .
Similarly to the formal residues, the classical residues of a derivative ∂G/∂v are all
zero.
Classical residues can be computed in a simple way: if ρ is not a pole of F ,
then Resv=ρ F = 0; if ρ is a pole of order 1, then Resv=ρ F = ((v − ρ)F ) |v=ρ ; and
if ρ is a pole of order r > 1, then
 r−1
(v − ρ)r F

1 ∂
Res F = .
v=ρ (r − 1)! ∂ r−1 v v=ρ

In its simplest form, the geometric reduction applies when f factors over K(z1`d )
as a product of factors of degree 1:
Y
f = C(z1`d ) (v − ρ)nρ ,
ρ∈U

where U is a finite subset of K(z1`d ). Then the partial fraction decomposition of F

writes as !
X aρ X bρ,k
F = + + P (v),
v−ρ (v − ρ)k
ρ∈U k>1
where aρ ∈ K(z1`d ) is Resv=ρ F , where bρ,k ∈ K(z1`d ) and where P ∈ K(z1`d )[v].
The terms with multiple poles and the polynomial term are derivatives and thus
their formal residue with respect to v are zero. Hence
 
X aρ
res F = res .
v v v−ρ
ρ∈U

Let ρ ∈ U . Depending on the leading monomial lm(ρ) of ρ, as an element of Ld ,

two situations may occur. Either lm(ρ) ≺ v, in which case
∞
aρ aρ X  ρ n
= ,
v−ρ v n=0 v
26 A. BOSTAN, P. LAIREZ, AND B. SALVY

aρ
and resv v−ρ = aρ ; or lm(ρ)  v, in which case
∞  n
aρ aρ X v
=− ,
v−ρ ρ n=0 ρ
a
ρ
hence resv v−ρ = 0. Since the variable v does not appear in ρ, the equality lm(ρ) = v
cannot happen. In the end, we obtain that
X
(9) res F = [lm(ρ) ≺ v] Res F,
v v=ρ
ρ∈U

where the bracket is 1 if the inequality inside is true and 0 otherwise. In particular,
the right-hand side is a rational function that we can compute.
Example 5. Let d > 0 be an integer and let us consider the binomial sum
n   
X n dk
un = (−1)k .
k n
k=0
n
We will show that un = (−d) . This example is interesting because the geometric
reduction procedure alone is able to compute entirely the double integral representing
the generating function of u, whereas Zeilberger’s algorithm finds a recurrence
relation of order d − 1 (Paule and Schorn 1995), far from the minimal recurrence
relation un+1 + dun = 0.
Algorithm 1 computes that
X z2
un tn = res ,
z1 ,z2 (z2 − t(1 + z1 ))(z1 z2 + t(1 + z1 )(1 + z2 )d )
n>0

with t ≺ z1 ≺ z2 . Let F denote the rational function on the right-hand side. Each
factor of the denominator of F has degree 1 with respect to z1 . Thus Equation (9)
applies. The roots of the denominator are
z2 −1
ρ1 = − 1 and ρ2 = z2 ,
t 1 + t(1+z 2)
d

moreover lm(ρ1 ) = z2 /t  z1 and lm(ρ2 ) = t/z2 ≺ z1 . Thus resz1 F = Resz1 =ρ2 F

and we obtain
X 1
un tn = res res F = res .
z2 z1 z2 t(1 + z2 )d + z2 − t
n>0

If d > 2, the denominator of the latter rational function does not split into factors
of degree 1 and Equation (9) does not apply. However, the study of nonrational
poles can lead to a further reduction.
5.2. Arbitrary poles. Equation (9) extends to the general case. To describe this
generalization, we need an algebraic closure of Ld . Let Ld,N be the field
def 1/N 1/N 1/N 1/N
Ld,N = K((zd ))((zd−1 )) . . . ((z2 ))((z1 )).
1/N
It is the algebraic extension of Ld generated by the zi . Let Ld,∞ be the union of
all Ld,N , N > 0, and let K ⊗K Ld,∞ be the compositum of Ld,∞ and K, where K is
an algebraic closure of K. The following is classical (e.g. Rayner 1974).
Lemma 5.1 (Iterated Puiseux theorem). The field K ⊗K Ld,∞ is an algebraic
closure of Ld .
The field K ⊗K Ld,∞ is thus simply denoted Ld . The valuation defined on Ld
is extended to a valuation defined on Ld with values in the group Qd , ordered
 MULTIPLE BINOMIAL SUMS 27

lexicographically. The leading monomial lm(ρ) of an f ∈ Ld is also defined as z v(f ) .

¯
The argument of §5.1 applies and shows that
X
(10) res F = [lm(ρ) ≺ v] Res F,
v v=ρ
ρ pole of F

where this time, the poles are in Ld . A root ρ is called small if ρ = 0 or lm(ρ) ≺ v
and large otherwise.
1
Example 6. Let F = xy(y 2 +y−x) , with x ≺ y. With respect to y, the poles of F
are 0, and √ √
1 1 + 4x 1 1 + 4x
ρ1 = − + and ρ2 = − − .
2 2 2 2
Only 0 and ρ1 are small. Thus
1 2
res F = Res F + Res F = − + √ .
y y=0 y=ρ1 x2 2
x + 4x − x 1 + 4x
Equation (10) does not look as interesting as Equation (9) because the right-hand
side is algebraic and need not be a rational function. However, if all roots are large,
then the residue is zero, which is interesting. On the contrary, if they are all small,
then resv F is the sum of all the classical residues of F , which equals the residue at
infinity:  
def 1
res F = Res F = Res − 2 F |v←1/v ,
v v=∞ v=0 v
which is a rational function. Thus, in the case where the poles are either all small
or all large, Equation (10) shows that the residue resv F is rational and shows how
to compute it.
Actually it is enough to check that any two conjugate poles (two poles are
conjugate if they annihilate the same irreducible factor of the denominator of F )
Qr
are simultaneously large or small. Indeed, we can write f = k=1 fknk where f1`r
are irreducible polynomials in K(z1`d )[v], and the partial fraction decomposition
leads to
r
X ak
F =
fknk
k=1
for some polynomials a1`r . Equation (10) applies to each term of this sum. If Uk
denotes the set of all the roots of fk , then
r X
X ak
res F = [lm(ρ) ≺ v] Res nk ,
v v=ρ f
k=1 ρ∈Uk k

and we can apply the all large or all small criterion to each subsum separately.
Proposition 5.2. With the notations above, if for all k, there exists εk ∈ {0, 1}
such that [lm(ρ) ≺ v] = εk for all ρ ∈ Uk , then
r
X ak
res F = εk Res .
v v=∞ fknk
k=1

It remains to explain how to compute the leading monomials of the roots of a

polynomial f ∈ K(z1`d )[v]. If d = 1, then it is the classical method of Newton’s
polygon for the resolution of bivariate polynomial equations using Puiseux series (e.g.
Walker 1950; Cutkosky 2004). When d > 2, we may apply this method recursively
by considering Ld as a subfield of the field of Puiseux series with coefficients in Ld−1 .
Based on this idea, Algorithm 3 proposes an implementation which avoids all
28 A. BOSTAN, P. LAIREZ, AND B. SALVY

Algorithm 3. Computation of the all large or all small criterion

Input: S ⊂ Nd finite; an integer 1 6 k 6 d.
Output: A subset of {in, out}.
a z n ∈ K[z1`d ] be a polynomial.
P
Specification: Let P = n∈Nd n ¯
If S = {n ∈
d ¯ ¯ ¯
N | an 6= 0} then AllLargeOrAllSmall(S, k) contains out (resp. in) if
¯
and only if there exists a nonzero ρ ∈ Ld in which zk does not appear such
that P (z1 , . . . , zk−1 , ρ, zk+1 , . . . , zd ) = 0 and lm(ρ)  zk (resp. lm(ρ) ≺ zk ).

function AllLargeOrAllSmall(S, k)
if maxm∈S mk = minm∈S mk then return ∅
if k = 1 then return {out}
µ ← min
 m∈S m1
M ← (m2`n ) ∈ Nd−1 m ∈ S and m1 = µ
r ← AllLargeOrAllSmall(M, k − 1)
if maxm∈S mk > maxm∈M mk then r ← r ∪ {out}
if minm∈S mk < minm∈M mk then r ← r ∪ {in}
return r

technicalities. Algorithm 4 sums up the geometric reduction procedure with respect

to the variable v. Of course one should apply the procedure iteratively with every
variable until no further reduction is possible. The success of the reduction may
depend on the order in which we eliminate all the variables. It is possible to try
every possible order efficiently thanks to the following compatibility property: if
both GeomRed(GeomRed(R, u), v) and GeomRed(GeomRed(R, v), u) do not
fail, then they are both equal to resu,v (R).

Example 7. Continuing Example 5, let us consider resz F1 , where F1 = 1/(t(1 +

Pd−1 d
k
z)d +z−t), with t ≺ z. The denominator factors as zg, where g = 1+t k=0 k+1 z .
1
Thus, the leading monomial of any root of f is t− d−1 , which it not ≺ z. All the
roots of g are large, so only the pole z = 0 remains and resz F1 = Resz=0 F1 , by
1
Proposition 5.2. Thus resz F1 = 1+dt and it follows that
n   
k n dk
X
(−1) = (−d)n .
k n
k=0
If we computed the Picard-Fuchs equation of F1 , we would find a differential equation
of order d because the Picard-Fuchs operator annihilates all the periods, and the
periods associated to the roots of g are algebraic of degree d − 1.

6. Optimizations
6.1. Infinite sums. So far, we have only considered binomial sums in which the
P
bounds of the symbols are finite and explicit. It is possible and desirable to
consider infinite sums, or more exactly syntactically infinite sums which in fact
reduce to finite sums but whose summation bounds are implicit. This often leads to
simpler integral representations.
Pn
For example, in the sum k=0 nk , the upper
summation bound n is not really


P∞
useful when n > 0, we could as well write k=0 nk , which defines the same sequence.
It is possible to adapt Algorithm 1 to handle infinite sums as long as the underlying
summations in the field of iterated Laurent series are convergent. Recall that a
 MULTIPLE BINOMIAL SUMS 29

Algorithm 4. Geometric reduction

Input: F (z1`d ), a rational function; an integer 1 6 k 6 d.
Output: Fail or a rational function S(z1`k−1 , zk+1`d ).
Specification: If a rational function S is returned and then reszk F = S.

function GeomRed(F Pr , k)
Decompose F as i=1 ai /fini +P (v) where the fi ’s are irreducible polynomials
in K[z1`d ] and a1`r , P ∈ K(z1`k−1 , zk+1`d )[zk ].
S←0
for i from 1 to r do
τ ← AllLargeOrAllSmall ({exponents of the monomials of fi }, k)
if τ ⊂ {in} then S ← Reszk =∞ (ai /fini )
if τ = {in, out} then return Fail
return S

geometric sum n>0 f n converges in the field Ld if and only if lm(f ) ≺ 1, see
P

Lemma 2.1.
To compute an integral representation of binomial sums involving infinite sum-
mations, the principle is to proceed as in §2.2 except when an infinite geometric
P
sum n>0 f in Ld shows up, where we check that lm(f ) ≺ 1 so that Lemma 2.1
applies. If it does, then the summation is performed and the computation continues.
If it does not, then the binomial sum is simply rejected.
P∞
sum un = k=0 nk , for n > 0. Note that


Example 8. Consider the binomial
Pn
Algorithm 1 applied to k=0 nk returns

n  
z − z −n
 
X n
= [1] (1 + z)n .
k z−1
k=0
We proceed as in §2.2 except that infinite geometric sums in Ld are valid when
Lemma 2.1 applies. Firstly nk = [1](1 + z)n z −k . Then we consider the infinite


P∞
sum k=0 (1 + z)n z −k . Since 1/z  1, it does not converge, so the binomial sum is
P ∞ n

rejected. And indeed, when n < 0 the sum n=0 k has infinitely many non zero
n n



terms. If we change k into n−k , which is the same when n > 0, we obtain
∞ ∞  n  n
X X 1 1 1 1
un = [1](1 + z)n z k−n−1 = [1] 1+ z k = [1] 1+ ,
z z z(1 − z) z
k=0 k=0
where this time the sum converges.

6.2. Building blocks. Besides the binomial coefficient nk = [1](1 + z)n z −k , we

have found useful to have additional building blocks to extend Algorithms 1 and 2
that contruct integral representations. Without enlarging the class of binomial
sums, one can add any sequence of the form n 7→ [1]R0 R1n1 · · · Rdnd , where R0`d are
¯
rational functions, as a new building block. Judicious extra building blocks may
give simpler integral representations or speed up the computation. For example,
when working with Motzkin numbers
∞      
def n−k n 2k + 2 2k + 2
X
Mn = (−1) − ,
k k+1 k+2
k=0
30 A. BOSTAN, P. LAIREZ, AND B. SALVY

k k k

n n n

n 0 n 00
n




(a) k (b) k
(c) k

Figure 2. Supports of some variants of the binomial coefficient.

one may add the new building block

n
(1 − x)(1 + x)2 1 + x + x2

Mn = [1] .
x x
In the applications below, we made use of alternative definitions of the binomial
coefficient and their corresponding representations:
 0
xk−n y −k
 
n def n
(11) = Hn Hk = [1] ,
k k 1−x−y
 00  
n def n 1
(12) = = [1] .
k n−k (1 − z)k+1 z n−k
They differ from the binomial coefficient, as defined in §1, only when n < 0 and
agree on their non zero values (see also Figure 2).

6.3. Summation over a polyhedron. As in §3.2, let Γ ⊂ Rd+e > be a rational

polyhedron and u : Zd+e → K be a binomial sum. Let us consider the sequence
X
v : n ∈ Nd 7→ un,m 1Γ (n, m),
¯ m∈Ze
¯ ¯ ¯ ¯
¯
where 1Γ (n, m) = 1 if (n, m) ∈ Γ and 0 otherwise. Proposition 3.12 shows that v is
¯ ¯ ¯ ¯
a binomial sum under an additional finiteness hypothesis on Γ, and according to
Corollary 2.7, there exists a rational function R(t1`d , z1`r ) such that
X
vn t¯n = res R.
¯¯ z1`r
dn∈N
¯
It is possible to circumvent the construction used in the proof of Proposition 3.12
and compute directly such a rational function R given two ingredients: firstly, the
generating function of Γ
def
X
ϕΓ (t1`d , s1`e ) = 1Γ (n, m)t¯n sm
¯,
d+e
¯ ¯ ¯ ¯
(n,m)∈Z
¯ ¯
which is known to be a rational function (e.g. Brion 1988); and secondly, a represen-
tation of the binomial sum u as
un,m = [1]RT1n1 · · · Tdnd S1m1 · · · Seme
¯ ¯
 MULTIPLE BINOMIAL SUMS 31

for some rational functions R, T1`d and S1`e ∈ K(z1`r ). Then, with t1`e ≺ z1`d ,
X X
vn t¯n = [1]R 1Γ (n, m)(t1 T1 )n1 · · · (td Td )nd S1m1 · · · Seme
¯¯ ¯ ¯
d
n∈N d+e
(n,m)∈Z
¯ ¯ ¯
R · ϕΓ (t1 T1 , . . . , td Td , S1 , . . . , Sd )
= res ,
z1`r z1 · · · zd
provided that the sums converge in Ld+e . This method is interesting because it is
known how to compute efficiently compact representations of the rational function ϕΓ
(Barvinok 2008).

7. Applications
7.1. Andrews-Paule identity. We detail the proof of the following identity:
n X n  2    2
X i+j 4n − 2i − 2j 2n
(13) = (2n + 1) .
i=0 j=0
j 2n − 2i n

This identity appeared first as a problem in the American Mathematical Monthly

(Blodgett 1990) and was subsequently proved by Andrews and Paule (1993) and
Wegschaider (1997) using various tools from the method of creative telescoping. It
attracted attention because the theory of creative telescoping was unable to give a
complete automated proof at that time.
Let un denote the left-hand side. It can be written as an infinite sum
∞ X ∞  0 2  0
X i+j (2n − 2i) + (2n − 2j)
un = ,
i=0 j=0
j 2n − 2i

0
0
where nk is the natural binomial defined by nk = Hn Hk nk , see §6.2. Of course, we

could have stuck to the former definition and used finite sums, but while the natural
binomials introduce two variables instead of one in the integral representation—see
Equation (11)—the integral representation obtained after the geometric reduction
step is often simpler when using the natural binomial.
So we obtain the following integral representation:
X
un tn = res (z2 z2 −t)(z4 z6 −z2 )(z3 z5 −z2 )(1−z
z1 z2
3 −z4 )(1−z5 −z6 )(1−z1 −z2 )
,
z1`6 1 2 1 2
n>0

with the ordering t ≺ z1 ≺ z2 ≺ z3 ≺ z4 ≺ z5 ≺ z6 . As expected, each binomial

coefficient brings two extra variables so we end up with six variables in addition to
the parameter t.
Geometric reduction applies successively with respect to the variables z1 , z3 , z4
−1/2 1/2 1/2
and z5 . For example, the poles w.r.t. the variable z1 , are {±tz2 }, {±z4 z6 }
and {1 − z2 }, gathered by conjugacy classes. The first pair of conjugate poles
is ≺ z1 whereas the second one is  z1 . The rational root 1 − z2 is  z1 . Thus, by
application of Proposition 5.2,
X (1 − z2 )z23
un tn = res .
z2`6 (t − (1 − z2 ) z ) (1 − z3 − z4 ) (z 2 − z3 z5 ) (1 − z5 − z6 ) (t − z 2 z4 z6 )
2 2
n>0 2 2 2

In the end, repeated application of Algorithm 4 leads to

X (z6 − 1)z63
un tn = res .
z4 ,z6 (t − z (z6 − 1)2 )((z4 − 1)t − z4 z 2 (z 2 + z4 − 1))
2
n>0 6 6 6
32 A. BOSTAN, P. LAIREZ, AND B. SALVY

Using the algorithm of Lairez (2016), we obtain (in about one second) a differential
operator annihilating n>0 un tn :
P

16t4 (256t2 + 736t + 81)(16t − 1)2 ∂t6

+ 16t3 (16t − 1)(86016t3 + 256256t2 + 20976t − 1053)∂t5
+ 4t2 (36601856t4 + 113760256t3 + 6103168t2 − 908088t + 14823)∂t4
+ 16t(22691840t4 + 75716608t3 + 6677824t2 − 459552t + 3645)∂t3
+ (305827840t4 + 1109626112t3 + 139138736t2 − 4247073t + 9720)∂t2
+(60272640t3 +244005120t2 +42117840t−374625)∂t +691200t2 +3369600t+996300
The roots of the indicial equation are 0, 1, − 12 et 12 . This differential operator
corresponds to the Picard-Fuchs equation associated to the integral representation,
but it is not the minimal-order operator annihilating n>0 un tn . This may happen
P

because the differential operator that we compute annihilates more that simply the
residue in which we ar interested: it annihilates every period of the integral of the
rational function inside the residue. Of course, this is not a issue as long as we do
obtain a differential equation.
Concerning the right-hand side, we find the integral representation
 2
X 2n t + u2 (u2 − 1)u1 (u1 − 1)
(2n + 1) tn = res 2
,
n z 1 2 (t − u2 (u2 − 1)u1 (u1 − 1))
,z
n>0

which cannot be simplified further with the geometric reduction. We compute

(in about 0.1s) the annihilating operator: t(16t − 1)∂t2 + (48t − 1)∂t + 12. The
Andrews-Paule identity follows with the equality test described in §4.2.
In this case the right-hand side is a hypergeometric sequence, so it can be
discovered automatically: the differential equation of order 6 leads to a recurrence
relation of order 4 for un from which the algorithm of Petkovšek (1992) finds
the hypergeometric solutions and the initial conditions are enough to identify the
right-hand side.
7.2. Several known identities. This section shows the integral representations
and the Picard-Fuchs equations appearing in the proofs of known identities. The
integral representations have been obtained with the method presented in this article
and the variants presented in §6. Note that the computation of the annihilating
operator never takes longer than 4 seconds.
n  2  2 X n    k  3
X n n+k n n+k X k
7.2.1. Strehl (1994). = .
k k k k j=0
j
k=0 k=0
This identity relates the Apéry numbers (left) and Franel numbers (the inner
sum in the right-hand side).
1
g.f.l.h.s2 = res
z1`3 (1 − z1 )(1 − z2 )(1 − z3 )z1 z2 z3 − t(z1 + z2 z3 − z1 z2 z3 )
1
g.f.r.h.s3 = res
z1`3 (1−z1 )(1−z2 )(1−z3 )z1 z2 z3 −t(1−z3 −z2 (1−(2+z1 (1−z2 )(1−z3 ))z3 ))

ann. op.4 = t2 (t2 − 34t + 1)∂t3 + 3t(2t2 − 51t + 1)∂t2 + (7t2 − 112t + 1)∂t + t − 5
2Generating function of the left-hand side
3Generating function of the right-hand side
4Annihilating operator of both right and left-hand sides
 MULTIPLE BINOMIAL SUMS 33

7.2.2. Graham et al. (1989, p. 33) and Wegschaider (1997, §5.7.6).

      X  4
XX n n n + s n + r 2n − r − s n
(−1)n+r+s =
r s s r n k
r>0 s>0 k>0

1
g.f.l.h.s = res
(1 − z1 )(1 − z2 )(1 − z3 )z1 z2 z3 + (z2 − z3 )(z1 − z3 )t
z1`3

1
g.f.r.h.s = res
z1`3 (1 − z1 )(1 − z2 )(1 − z3 )z1 z2 z3 + (1 − z1 − z2 )z3 t − (1 − z1 )(1 − z2 )t

ann. op. = t2 (4t+1)(16t−1)∂t3 +3t(128t2 +18t−1)∂t2 t+(444t2 +40t−11)∂t +60t+2

7.2.3. Dent. The following identity is due to Dent and used as an example by
Wegschaider (1997, p. 90):
n1X
+2nX
2      
k 2n2 + n1 − k n1 n1 + n2
(−1)j = 2n1 , with n1 , n2 > 0.
j 2n2 − j k−j n1
k=0 j>0

1
g.f.l.h.s = g.f.r.h.s =
1 − 2t1 − t2
Here, the generating series is rational and geometric reduction performs the entire
computation, there is no need to compute a Picard-Fuchs equation.
2n  3
k 2n (3n)!
X
7.2.4. Dixon (1891). (−1) = (−1)n 3
k n!
k=0

(1 − z2 )(1 − z1 )z1 z2
g.f.l.h.s = res
z1`2 z12 z22 (1
− z2 )2 (1 − z1 )2 − (1 − z1 − z2 )2 t
1
g.f.r.h.s = res
z1`3 t + z1 z2 (1 − z1 − z2 )

ann. op. = t(27t + 1)∂t2 + (54t + 1)∂t + 6

7.2.5. Moriarty (see Egorychev 1984, p. 11).
m      
k k n n+k n n+m
X
n m
(−4) = (−1) 4
m n+k 2k n+m 2m
k=n
Because of the division by n + m, the right-hand side is not obviously a binomial
sum. However, it becomes obvious after observing that
     
n n+k n+k−1 1 n+k−1
= + .
n+k 2k 2k 2 2k − 1

1 (1 − t1 )(1 + t1 )
g.f.l.h.s = g.f.r.h.s =
2 t21 + 4t1 t2 + 2t1 + 1
Here again it is a rational power series and the geometric reduction finds it.
7.2.6. Davletshin et al. (2015, Theorem 1.2).
 
∞   nX 1 /2   X n1  
X n 2 m − 1 m − 1
1+ 2q−1  + 
q=1
q m=1
q − 1 m=1
q − 1
∞      
X
q−1 n2 n1 bn1 /2c
= 2 +
q=1
q q q
34 A. BOSTAN, P. LAIREZ, AND B. SALVY

t1 t2 (1 + t1 − t2 − t1 t2 − 2t21 − 2t21 t2 )
g.f.l.h.s = g.f.r.h.s =
(1 − t2 )(1 − t1 )(1 − t2 − t21 t2 − t21 )(1 − t1 − t2 − t1 t2 )
The summation bound n1 /2 and the integer part bn1 /2c may look problematic,
until we observe that
n/2   X ∞  
X m−1 m−1
= Hn−2m
m=1
q−1 m=1
q−1
  X∞  
bn/2c k
and = (δ2k−n + δ2k+1−n ) ,
q q
k=0
using the binomial sums δ and H defined in §1.

7.2.7. Davletshin et al. (2015, Theorem 1.1).

 
∞   n1 /2   X ∞ ∞ q
X n2  X m−1 X Y
2+ + ··· δk1 +···+kq −m (ki + 1)
q=1
q m=1
q−1 i=1
k1 =1 kq =1
   
n1 + 2n2 bn1 /2c + n2
= +
2n2 n2
t1 t2 (1 − t2 − 2t21 + t31 − 2t21 )
g.f.l.h.s = g.f.r.h.s =
(1 − t2 )(1 − t1 )(1 − t2 − t21 )(1 − t2 − 2t1 + t21 )
Again, the left-hand side does not have the appearance of a binomial sum until
we remark that
∞ ∞ q  q 
X X Y t(2 − t) 1
··· δk1 +···+kq −m (ki + 1) = [1] .
i=1
(1 − t)2 tm
k1 =1 kq =1

7.2.8. Chapman and Tauraso (2016).

n   3m   
X n−k X n + 1 − 2k m−k
(−4)−k (−2)−j = 0, n, m > 0,
k − 1 j=1 j−1 3m − j
k=1

where we consider the binomial coefficient as defined in §1.1 and not one of the
variants of §6.2. This is an important clarification because negative values may
appear in the upper arguments of the binomial coefficients. The geometric reduction
is enough to prove this equality, no integration step is required.

7.3. Proof of some conjectures.

7.3.1. Le Borgne’s identity. The following identity for Baxter’s numbers arises as a
conjecture in an unpublished work by Yvan Le Borgne. With the methods presented
here we can prove it automatically.

1 + Fn−1,−1 + 2Fn0,0 − Fn0,1 + Fn1,0 − 3Fn1,1 + Fn1,2 − Fn3,1 + 3Fn3,2

n+2 n+2
n


n+2

X m m+1 m+2
3,3 4,2 4,3 5,2
− Fn − 2Fn + Fn − Fn = n+2 n+2


,
m=0 1 2
n−1
X d−a
X
d−a−c n

 n+d+1−2a−2c+2b n+d+1−2a−2c+2b


where Fna,b =



c d−a−c n−a−c+b − n+1−a−c+b .
d=0 c=0
The sum in the right-hand side does not have the appearance of a binomial sum
(even if it is such), but the identity becomes clearly an identity between binomial
 MULTIPLE BINOMIAL SUMS 35

sums after multiplication by n+2

n+2

1 2 . Alternatively, one can use the following
equality to write the right-hand side as a binomial sum:



n n n


n+2 n+2


n+2

k
k+1 k+2

m m+1 m+2  n n

n 
n+2 n+2


= det  k−1 k
k+1 ,
n

n n


1 2
k−2 k−1 k

7.3.2. Identities from Brent et al. (2014). We have been able to prove the following
identities conjectured by Brent et al. (2014). The left-hand sides involve absolute
values of nonlinear polynomials. They are nevertheless binomial sums for two
reasons. The first one is that all the nonlinear polynomials under consideration
split into linear or positive factors. For example |i3 − j 3 | = |i − j|(i2 + ij + j 2 )
and |i − j| = (i − j)(Hi−j − Hj−i ) is a binomial sum.
The second reason, which we used for the computation, is that we can eliminate
the absolute values by using the symmetries of the sums. For example
X  2n  2n  n X i−1   
3 3
X 2n 2n
|i − j | = 4 i3 .
i,j
n + i n + j i=0 j=−i
n + i n + j

The integral representations obtained are too lengthy to be presented here, they
can be found online5. Since the right-hand sides are hypergeometric sequences, they
can be computed from the recurrence relations satisfied by the left-hand sides with
the algorithm of Petkovšek (1992).

X  2n  2n  2n2 (5n − 2) 4n
 
3 3
(Brent et al. 2014, Eq. 5.7) |i − j | =
i,j
n+i n+j 4n − 1 2n

(Brent et al.2014, 
Eq. 5.8)
2n2 (43n3 − 70n2 + 36n − 6) 4n
X 2n  
2n 5 5
|i − j | =
i,j
n+i n+j (4n − 1)(4n − 3) 2n

(Brentet al. 
2014, Eq.5.9)  
X 2n 2n 2n2 (531n5 −1960n4 +2800n3 −1952n2 +668n−90) 4n
|i7 − j 7 | = (4n−1)(4n−3)(4n−5)
i,j
n + i n + j 2n

(Brent et al. 2014, Eq. 5.12)

 2
X  2n  2n 
2 2 2n3 (n − 1) 2n
|ij(i − j )| =
i,j
n+i n+j 2n − 1 n

(Brent et al. 2014, Eq. 5.14)

 2
X  2n  2n 
3 3 2 2 2n4 (n − 1)(3n2 − 6n + 2) 2n
|i j (i − j )| =
i,j
n+i n+j (2n − 3)(2n − 1) n

7.4. Computational limitations. The method is mainly limited by the integra-

tion step. When the integral representation has more than four variables (in
addition to the parameter and after the geometric reduction), then computation of
the Picard-Fuchs equation becomes challenging. For example, in Strehl’s second

5https://github.com/lairez/binomsums
36 A. BOSTAN, P. LAIREZ, AND B. SALVY

identity
n  3  3 ∞    ∞  2  2  
X n n+k X n n+k X k 2i 2i
= ,
k k k k i=0
i i k−i
k=0 k=0
the integral representation of the generating function of each side has five variables
and a parameter. For the left-hand side, we obtain
1
res .
z1`5 (z1 z3 z4 z5 +z2 z3 z4 z5 −z1 z2 z3 z4 z5 −z3 z4 z5 +z1 z2 )t+z1 z2 z3 z4 z5 (1−z1 )(1−z2 )(1−z3 )(1−z4 )(1−z5 )

The integral representation of the right-hand side is more complicated ant still has
five variables, in addition to the parameter. The computation requires several hours
with the current algorithms. Without the geometric reduction, it involves nine
variables and a parameter.

References
Abramov, S. A. (2006). “On the summation of P -recursive sequences”. In:
Proceedings of the international symposium on symbolic and algebraic
computation. ISSAC ’06. ACM Press, pp. 17–22. doi:
10.1145/1145768.1145779.
Abramov, S. A. and M. Petkovšek (1994). “D’Alembertian solutions of linear
differential and difference equations”. In: Proceedings of the international
symposium on symbolic and algebraic computation. ISSAC ’94. ACM Press,
pp. 169–174. doi: 10.1145/190347.190412.
– (2002). “On the structure of multivariate hypergeometric terms”. In: Advances in
Applied Mathematics 29.3, pp. 386–411. doi:
10.1016/S0196-8858(02)00022-2.
Abramov, S. A. and M. Petkovšek (2005). “Gosper’s algorithm, accurate
summation, and the discrete Newton-Leibniz formula”. In: Proceedings of the
international symposium on symbolic and algebraic computation. ISSAC ’05.
ACM Press, pp. 5–12. doi: 10.1145/1073884.1073888.
Allouche, J.-P. and M. Mendès France (2011). “Hadamard grade of power series”.
In: Journal of Number Theory 131.11, pp. 2013–2022. doi:
10.1016/j.jnt.2011.04.011.
Andrews, G. E. and P. Paule (1993). “Some questions concerning
computer-generated proofs of a binomial double-sum identity”. In: Journal of
Symbolic Computation 16.2, pp. 147–153. doi: 10.1006/jsco.1993.1038.
Apagodu, M. and D. Zeilberger (2006). “Multi-variable Zeilberger and
Almkvist-Zeilberger algorithms and the sharpening of Wilf-Zeilberger theory”.
In: Advances in Applied Mathematics 37.2, pp. 139–152. doi:
10.1016/j.aam.2005.09.003.
Barvinok, A. (2008). Integer points in polyhedra. Zurich Lectures in Advanced
Mathematics. European Mathematical Society (EMS), Zürich. doi:
10.4171/052.
Blodgett, R. J. (1990). “Problem E3376”. In: American Mathematical Monthly 97.3,
p. 240.
Bostan, A., P. Lairez, and B. Salvy (2013a). “Creative telescoping for rational
functions using the Griffiths-Dwork method”. In: Proceedings of the international
symposium on symbolic and algebraic computation. ISSAC ’13. ACM Press,
pp. 93–100. doi: 10.1145/2465506.2465935.
 REFERENCES 37

Bostan, A., S. Boukraa, G. Christol, S. Hassani, and J.-M. Maillard (2013b). “Ising
n-fold integrals as diagonals of rational functions and integrality of series
expansions”. In: Journal of Physics. A. Mathematical and Theoretical 46.18,
pp. 185202, 44. doi: 10.1088/1751-8113/46/18/185202.
Brent, R. P., H. Ohtsuka, J.-A. H. Osborn, and H. Prodinger (2014). Some
binomial sums involving absolute values. arXiv: 1411.1477.
Brion, M. (1988). “Points entiers dans les polyèdres convexes”. In: Annales
Scientifiques de l’École Normale Supérieure. Quatrième Série 21.4, pp. 653–663.
Cameron, R. H. and W. T. Martin (1938). “Analytic continuation of diagonals and
Hadamard compositions of multiple power series”. In: Transactions of the
American Mathematical Society 44.1, pp. 1–7. doi: 10.2307/1990100.
Chapman, R. and R. Tauraso (2016). “Problem 11914”. In: American Mathematical
Monthly 123.5, p. 613.
Christol, G. (1985). “Diagonales de fractions rationnelles et équations de
Picard-Fuchs”. In: Study group on ultrametric analysis, 12th year, 1984/85, No.
1. Paris: Secrétariat Math., Exp. No. 13, 12.
– (1988). “Diagonales de fractions rationnelles”. In: Séminaire de Théorie des
Nombres, Paris 1986–87. Vol. 75. Progr. Math. Boston, MA: Birkhäuser Boston,
pp. 65–90.
– (1990). “Globally bounded solutions of differential equations”. In: Analytic
number theory (Tokyo, 1988). Vol. 1434. Lecture Notes in Math. Berlin:
Springer, pp. 45–64. doi: 10.1007/BFb0097124.
– (2015). “Diagonals of rational fractions”. In: Eur. Math. Soc. Newsl. 97,
pp. 37–43.
Chyzak, F. (2000). “An extension of Zeilberger’s fast algorithm to general
holonomic functions”. In: Discrete Mathematics 217.1-3. Formal power series and
algebraic combinatorics (Vienna, 1997), pp. 115–134. doi:
10.1016/S0012-365X(99)00259-9.
– (2014). The ABC of Creative Telescoping: Algorithms, Bounds, Complexity.
Mémoire d’habilitation à diriger les recherches.
Chyzak, F., A. Mahboubi, T. Sibut-Pinote, and E. Tassi (2014). “A
Computer-Algebra-Based Formal Proof of the Irrationality of ζ(3)”. In: 5th
International Conference on Interactive Theorem Proving. Vienna, Austria. doi:
10.1007/978-3-319-08970-6_11.
Cutkosky, S. D. (2004). Resolution of singularities. Vol. 63. Graduate Studies in
Mathematics. American Mathematical Society. doi: 10.1090/gsm/063.
Davletshin, M., G. Egorychev, and V. Krivokolesko (2015). New applications of the
Egorychev method of coefficients of integral representation and calculation of
combinatorial sums. arXiv: 1506.03596.
Dixon, A. C. (1891). “On the sum of the cubes of the coefficients in a certain
expansion by the binomial theorem”. In: Messenger of mathematics 20,
pp. 79–80.
Egorychev, G. P. (1984). Integral representation and the computation of
combinatorial sums. Vol. 59. Translations of Mathematical Monographs.
Translated from the Russian by H. H. McFadden, Translation edited by Lev J.
Leifman. Providence, RI: American Mathematical Society.
Egorychev, G. P. and E. V. Zima (2008). “Integral representation and algorithms
for closed form summation”. In: Handbook of algebra. Vol. 5. Vol. 5. Handb.
38 REFERENCES

Algebr. Elsevier/North-Holland, Amsterdam, pp. 459–529. doi:

10.1016/S1570-7954(07)05009-7.
Flajolet, P. and M. Soria (1998). Coefficients of algebraic series. Tech. rep. 3504.
Algorithms seminar 1997-1998. Inria, pp. 27–30.
Flajolet, P., S. Gerhold, and B. Salvy (2005). “On the Non-Holonomic Character of
Logarithms, Powers, and the nth Prime Function”. In: Electr. J. Comb. 11.2.
Furstenberg, H. (1967). “Algebraic Functions over Finite Fields”. In: Journal of
Algebra 7.2, pp. 271–277. doi: 10.1016/0021-8693(67)90061-0.
Garoufalidis, S. (2009). “G-functions and multisum versus holonomic sequences”.
In: Advances in Mathematics 220.6, pp. 1945–1955. doi:
10.1016/j.aim.2008.11.012.
Garoufalidis, S. and X. Sun (2010). “A new algorithm for the recursion of
hypergeometric multisums with improved universal denominator”. In: Gems in
experimental mathematics. Vol. 517. Contemp. Math. Providence, RI: Amer.
Math. Soc., pp. 143–156. doi: 10.1090/conm/517/10138.
Garrabrant, S. and I. Pak (2014). Counting with irrational tiles. arXiv: 1407.8222.
Graham, R. L., D. E. Knuth, and O. Patashnik (1989). Concrete mathematics. A
foundation for computer science. Addison-Wesley Publishing Company.
Grothendieck, A. (1966). “On the de Rham cohomology of algebraic varieties”. In:
Institut des Hautes Études Scientifiques. Publications Mathématiques 29,
pp. 95–103.
Ince, E. L. (1944). Ordinary Differential Equations. Dover Publications, New York.
Katz, N. M. (1970). “Nilpotent connections and the monodromy theorem:
Applications of a result of Turrittin”. In: Institut des Hautes Études
Scientifiques. Publications Mathématiques 39, pp. 175–232.
Koutschan, C. (2010). “A fast approach to creative telescoping”. In: Mathematics in
Computer Science 4.2-3, pp. 259–266. doi: 10.1007/s11786-010-0055-0.
– (2013). “Creative telescoping for holonomic functions”. In: Computer Algebra in
Quantum Field Theory. Ed. by C. Schneider and J. Blümlein. Texts &
Monographs in Symbolic Computation. Springer, pp. 171–194.
Lairez, P. (2014). “Périodes d’intégrales rationnelles : algorithmes et applications”.
PhD thesis. École polytechnique. Tel: 01089130.
– (2016). “Computing periods of rational integrals”. In: Mathematics of
Computation 85.300, pp. 1719–1752. doi: 10.1090/mcom/3054.
Lipshitz, L. (1988). “The diagonal of a D-finite power series is D-finite”. In:
Journal of Algebra 113.2, pp. 373–378. doi: 10.1016/0021-8693(88)90166-4.
– (1989). “D-Finite Power Series”. In: Journal of Algebra 122.2, pp. 353–373.
Mezzarobba, M. and B. Salvy (2010). “Effective bounds for P-recursive sequences”.
In: Journal of Symbolic Computation 45.10, pp. 1075–1096. doi:
10.1016/j.jsc.2010.06.024.
Neumann, B. H. (1949). “On ordered division rings”. In: Transactions of the
American Mathematical Society 66, pp. 202–252. doi:
10.1090/S0002-9947-1949-0032593-5.
Paule, P. and M. Schorn (1995). “A Mathematica Version of Zeilberger’s Algorithm
for Proving Binomial Coefficient Identities”. In: Journal of Symbolic
Computation 20.5-6, pp. 673–698. doi: 10.1006/jsco.1995.1071.
Pemantle, R. and M. C. Wilson (2013). Analytic Combinatorics in Several Variables.
Vol. 140. Cambridge University Press. doi: 10.1017/CBO9781139381864.
 REFERENCES 39

Petkovšek, M. (1992). “Hypergeometric solutions of linear recurrences with

polynomial coefficients”. In: Journal of Symbolic Computation 14.2-3,
pp. 243–264. doi: 10.1016/0747-7171(92)90038-6.
Rayner, F. J. (1974). “Algebraically closed fields analogous to fields of Puiseux
series”. In: Journal of the London Mathematical Society. 2nd ser. 8, pp. 504–506.
doi: 10.1112/jlms/s2-8.3.504.
Salvy, B. and S. Melczer (2016). “Symbolic-numeric tools for analytic combinatorics
in several variables”. In: Proceedings of the international symposium on symbolic
and algebraic computation. ISSAC ’16. ACM Press. doi:
10.1145/2930889.2930913.
Salvy, B. and P. Zimmermann (1994). “Gfun: a Maple package for the manipulation
of generating and holonomic functions in one variable”. In: ACM Transactions
on Mathematical Software 20.2, pp. 163–177. doi: 10.1145/178365.178368.
Strehl, V. (1994). “Binomial identities. Combinatorial and algorithmic aspects”. In:
Discrete Mathematics 136.1-3. Trends in discrete mathematics, pp. 309–346. doi:
10.1016/0012-365X(94)00118-3.
Takayama, N. (1995). “An algorithm for finding recurrence relations of binomial
sums and its complexity”. In: Journal of Symbolic Computation 20.5-6. Symbolic
computation in combinatorics ∆1 (Ithaca, NY, 1993), pp. 637–651. doi:
10.1006/jsco.1995.1069.
Walker, R. J. (1950). Algebraic Curves. Princeton University Press.
Wegschaider, K. (1997). “Computer Generated Proofs of Binomial Multi-Sum
Identities”. MA thesis. RISC, J. Kepler University.
Wilf, H. S. and D. Zeilberger (1992). “An algorithmic proof theory for
hypergeometric (ordinary and “q”) multisum/integral identities”. In: Inventiones
Mathematicae 108.3, pp. 575–633. doi: 10.1007/BF02100618.
Xin, G. (2004). “The ring of Malcev-Neumann series and the residue theorem”.
PhD thesis. Brandeis University. arXiv: 0405133.
Zeilberger, D. (1990). “A holonomic systems approach to special functions
identities”. In: Journal of Computational and Applied Mathematics 32.3,
pp. 321–368. doi: 10.1016/0377-0427(90)90042-X.
– (1991a). “A Maple program for proving hypergeometric identities”. In: SIGSAM
Bulletin 25.3, pp. 4–13. doi: 10.1145/122514.122515.
– (1991b). “The method of creative telescoping”. In: Journal of Symbolic
Computation 11, pp. 195–204. doi: 10.1016/S0747-7171(08)80044-2.

(A. Bostan) Inria Saclay Île-de-France, Bâtiment Alan Turing, 1 rue Honoré d’Estienne
d’Orves, 91120 Palaiseau, France
E-mail address: [email protected]

(P. Lairez) Technische Universität Berlin, Fakultät II, Sekretariat 3-2, Straße des 17.
Juni 136, 10623 Berlin, Germany
E-mail address: [email protected]

(B. Salvy) Inria, LIP (U. Lyon, CNRS, ENS Lyon, UCBL), France
E-mail address: [email protected]