Vous êtes sur la page 1sur 9

A base-b extension of the binomial coefficient

Tanay V. Wakhare, Christophe Vignat

Quince Orchard High School, Gaithersburg, MD 20878, USA


E-mail: twakhare@gmail.com

Tulane University, New Orleans, LA 70118, USA


E-mail: cvignat@tulane.edu

arXiv:1607.02564v1 [math.NT] 9 Jul 2016

Received XX July 2016 in final form ????; Published online ????


doi:10.3842/JOURNAL.201*.1

Abstract. We study the properties of the base-b binomial coefficient defined by Jiu and
Vignat, introduced in the context of a digital binomial theorem. After introducing a general
summation formula we derive base-b analogues of the Stirling numbers of the second kind,
the Fibonacci numbers, and the classical exponential function.
Key words: binomial coefficient; Fibonacci numbers, bases, binomial theorems
2010 Mathematics Subject Classification: 05A10;11B65;11B73;11B39

Introduction

Recently, Jiu and Vignat have conducted work [5] concerning the base-b binomial coefficient. Letting n =
PNn 1
PNk 1 i
i
i=0 ni b and k =
i=0 ki b be the base-b expansions of n and k respectively, and defining N := max{Nn , Nk },
the base-b binomial coefficient is given as
 
N
1  
Y
n
ni
:=
,
(1.1)
k b
ki
i=0

and nk b = 0 if ki > ni . This also motivates the definition of a base b factorial, which we define as (n!)b :=
QNn 1
i=0 ni !.

P
The binomial theorem is a classic result stating that (X + Y )n = nk=0 nk X k Y nk . Nguyen subsequently
generalized this to the digital binomial theorem in [3, (26)], which states that
X
(X + Y )s2 (n) =
X s2 (k) Y s2 (nk) ,
(1.2)
0kb n

where sb (n) denotes the sum of the digits of n expressed in base b. The condition 0 k b n restricts the
summation over indices such that the each digit of k is less than the corresponding digit of n, so that ki ni for
all 0 i N 1. This notation, which has previously been referred to as digital dominance, will be adopted
throughout this paper. This is also equivalent to the condition that the addition of n k and k be carry-free in
base b, which can be written symmetrically as sb (k) + sb (n k) = sb (n).
The discovery of this digital binomial theorem spurred further extensions by Jiu and Vignat [5], Nguyen [3]
[4], and Mansour [1] [2], which led to the introduction of the base-b binomial coefficient.
The paper is organized as follows. In Section 2, we present a general summation formula, which is used in
Section 3 to derive theorems about a base-b analogue of the Stirling numbers of the second kind and in Section
4 to derive base-b analogues of the Fibonacci numbers. In Section 5 we introduce an analogue of the exponential
function involving the base-b factorial and derive some of its properties.
1

Sums over carry-free k

We can extend the methods of [5] to take sums over carry-free k, with a weighting by the base-b binomial
coefficient.
P
Theorem 1. Define S(~a) := nk=0 f (~a) where ~a is an arbitrary vector of variables. Then
N
1
Y

S (~ai ) =

ni
N
1 X
Y

N
1
Y

f (~ai ).

(2.1)

0kb n i=0

i=0 ki =0

i=0

f (~ai ) =

Proof . We define S(~a), then apply this sum to an arbitrary ni < b. We can then take a product over N distinct
ni s, corresponding to an N digit number. The second equality follows from the fact that
ni
N
1 X
Y

f (~ai ) =

i=0 ki =0

n0
X

k0 =0

n0
X

n1
X

k0 =0 k1 =0

n1
X

f (~a0 )

nN1

f (~a1 )

k1 =0
nN1

N
1
Y

kN1 =0

f (~aN 1 )

kN1 =0

f (~a0 )f (~a1 ) f (~aN 1 )

f (~ai ).

0kb n i=0


Corollary 2. Let S(n) :=
N
1
Y
i=0

S (ni ) =

Pn

k=0 f (n, k).

N
1
Y

Then

f (ni , ki ).

(2.2)

0kb n i=0

Proof . Take f (~a) = f (n, k) in (2.1).



P
Corollary 3. Let S2 (n) := nk=0 nk f (n, k). Then
N
1
Y
i=0

S2 (ni ) =

N
1 
Y

0kb n i=0


1
n   N
X
n Y
ni
f (ni , ki ).
f (ni , ki ) =
k b
ki

(2.3)

i=0

k=0


Proof . Take f (~a) = nk f (n, k) in (2.1). We can extend the sum over all 0 k n since
digitally dominated by n.

n
k b

= 0 if k is not


While the algebraic proof of this identity is very simple, it greatly simplifies and supersedes the proof of many
previously discovered identities relating to sums over digitally dominated k, and admits much more sweeping
generalizations. The previous work done on sums over carry-free k has centered on proving this identity for
special values of f (~a), often using the properties of infinite matrices. We now give the choices of f (~a) that yield
the central theorems in previous papers [1] [2] [3] [4] [5].

P
By taking f (n, k) = X k Y nk in (2.3) with constant X and Y , and noting that S2 (n) = nk=0 nk X k Y nk =
(X + Y )n , we recover the base-b binomial theorem[5, (Theorem 2)]. Taking the b = 2 case reduces to the digital
binomial theorem[3, (26)]. Namely,
n  
X
n
(X + Y )sb (n) =
X sb (k) Y sb (nk) .
(2.4)
k b
k=0

(X)k (Y )nk
k! (nk)! with
(X+Y )n
S(n) =
by
n!

We take f (n, k) =

constant X and Y in (2.1), where (x)k :=

(x+k)
(x)

is the Pochhammer

symbol. Noting that


the Vandermonde identity, we obtain an extension of the base-b binomial
theorem, the central theorem in [4, (Theorem 2)]. Namely,



N
1 
1 
Y
X NY
X + Y + ni 1
X + ki 1 Y + ni ki 1
=
.
(2.5)
ni
ki
ki
0kb n i=0

i=0

 y;r 
x(x+r)(x+(d1)r)
x;r
Taking f (~a) = f (x, y, r, n, k) = x;r
[1, (Definition 8)] and using
k
nk in (2.1) where d :=
d!
[1, (Lemma 9)] to find the sum of f over k gives [1, (Theorem 4)], which can be specialized to find q-analogues.
Namely,



1 
N
1 
Y
X NY
x i ; ri
y i ; ri
xi + y i ; r
.
(2.6)
=
ki
ni ki
ni
0kb n i=0

i=0

Taking f (~a) = f (n, k, x, y, s, p) = pk (x)snk (y) and f (~a) = pk (x)pnk (y) in (2.1) where s and p are normalized
Sheffer sequences and using [2, (Theorem 7)] to evaluate the sum of f over k gives [2, (Theorem 1)].
As a corollary of (2.6), we obtain a base-b analog of the Chu-Vandermonde identity. The Chu-Vandermone
identity states that, for integer m and n,
 

r  
X
n
m
n+m
=
.
(2.7)
k
rk
r
k=0

This is found as [7, (Page 8)], with the substitution n n + p and variables renamed to maintain consistency.
This can be extended to complex valued n and m. This has a base-b counterpart [5, (Theorem 8)] which can
be proved by taking r = 1 in (2.6) and applying (1.1). We note that if the addition of n and m in base b in

QN 1 ni +mi 
carry-free, then i=0
= n+m
. Changing variables to maintain consistent notation yields, for integral
ri
r
b
n and m such that their addition in base b in carry-free,


X n   m 
n+m
=
.
(2.8)
r
k b rk b
b
0kb r

As with the digital binomial theorem, an identity involving binomial coefficients has an obvious base-b analog.
However, we emphasize that the results do not immediately transfer, as can be seen by the restriction on n
and m in the base-b Chu-Vandermonde identity. However, the striking similarities between the digital binomial
theorem and binomial theorem ensure that some linear identities transfer almost identically. We list some
analogs of classic binomial identities derived from the binomial theorem, omitting the proofs since they follow
classical lines.
n  
X
n
k=0

n  
X
n
k=0
n 
X

= 2sb (n) .

(2.9)

sb (k) = sb (n)2sb (n)1 .

(2.10)


n
sb 2 (k) = sb (n)(sb (n) + 1)2sb (n)2 .
k b
k=0
 
n  2
X
2n
n
=
.
n b
k b

(2.11)

(2.12)

k=0

Stirling numbers of the second kind

The Stirling numbers of the second kind count the number of ways to partition a set of n elements into k
non-empty subsets. They have a natural base-b analog.
Theorem 4.
Sb (n, k) :=

N
1
Y
i=0

N
1
Y

 
k
1 X
sb (k)sb (j) k
S(n, ki ) =
(1)
(k!)b
j b
j=0

i=0

ji

!n

(3.1)

Proof . From [7, (Page 90)] they can be represented as


 
k
1 X
kj k
(1)
S(n, k) =
jn.
k!
j

(3.2)

j=0

Rearranging this relation and changing variables to maintain consistency with (2.3) yields
 
n
X
k n
(1)
k = (1)n n!S(, n).
k

(3.3)

k=0

Taking f (n, k) = (1)k k in (2.3) and noting that S2 (n) is nothing more than the left hand side of (3.2), we
obtain
1
n   N
N
1
X
Y
n Y
ni
(1)ki ki .
(3.4)
(1) ni !S(, ni ) =
k b
i=0

i=0

k=0

Simplifying the product on both sides yields


n  
N
1
X
Y
n
sb (n)
S(, ni ) =
(1)sb (k)
(1)
(n!)b
k b
i=0

k=0

N
1
Y

ki

i=0

(3.5)

Rearranging then gives


N
1
Y
i=0

 
n
1 X
sb (n)sb (k) n
S(, ni ) =
(1)
(n!)b
k b
k=0

N
1
Y

ki

i=0

(3.6)

which can be compared with (3.2). The parallels are immediately obvious. Renaming variables completes the
proof.

In general, any numbers with an explicit representation as a sum involving a binomial coefficient will have a
base-b analog. These base-b Stirling numbers can then be used to generalize certain results about the Stirling
d
numbers. A fundamental result concerning the differential operator := xD, D := dx
, found in [7, (Page 218)],
is
n =

n
X

S(n, k)xk D k .

(3.7)

k=0

Theorem 5. Define a multivariable differential operator N := x0 x1 xN 1 D0 D1 DN 1 where Di :=


Then
nN

k1 ,...,kN1 n

Sb (n, k)

N
1
Y

xki i Diki .

xi .

(3.8)

i=0

Proof . By the equality of mixed partial derivatives,


nN = (x0 D0 x1 D1 xN 1 DN 1 )n = (x0 D0 )n (x1 D1 )n (xN 1 DN 1 )n .

(3.9)

Expanding each term using (3.7) gives

n
n
X
X
kN1
k
nN =
S(n, k0 )xk00 D0k0
S(n, kN 1 )xNN1
1 DN 1
k0 =0

kN1 =0

N
1
Y

S(n, ki )xki i Diki

k1 ,...,kN1 n i=0

Sb (n, k)

N
1
Y

xki i Diki .

(3.10)

i=0

k1 ,...,kN1 n

In the second equality, we used the interchange of summation presented in (2.1).

Theorem 6.
Sb (n, k) =

X Y
j

|J|=j

kJ S(n 1, kJ )S(n 1, kJ 1),

(3.11)

where the product is over all partitions J of {1, . . . , n} and J = {1, . . . , n}\J
Proof . We begin with the Pascal-type recurrence [6, (4)] S(n, k) = S(n 1, k 1) + kS(n 1, k) and substitute
into the first equality in (3.1). Namely,
Sb (n, k) =

N
1
Y

S(n, ki ) =

N
1
Y
i=0

i=0

S(n 1, k 1) + kS(n 1, k),

from which the theorem follows.

(3.12)


The base-b Stirling number naturally occurs as a consequence of creating a multivariable generalization of an
existing single variable identity, a process which could be explored further in the future. Additionally, the base b
has no special significance here so long as it is larger than ki , suggesting that Sb (n, k) has a larger combinatorial
significance outside of the sum of digits function.

Fibonacci numbers

In this section, we introduce one further analogue of classic numbers with well-studied properties. The Fibonacci
numbers Fn are defined by the recurrence Fn+1 = Fn + Fn1 and the initial values F0 = 1 and F1 = 1.
Theorem 7. Denote by Fnb,j the base-b generalized Fibonacci numbers. Then
Fnb,j

:=

N
1
Y
i=0

Fni =


n 
X
nk
k=0

(4.1)

.
b

Furthermore,
b,j
b,j
Fnb,j = Fnb
j + Fn2bj .

(4.2)
5

Proof . By summing over shallow diagonals of Pascals triangle, we have the relation

n 
X
nk
Fn =
.
k

(4.3)

k=0


Taking f (n, k) = nk
in (2.2) and utilizing the relation (4.3) yields (4.1). From [5, (Remark 11)], for 0 j
k
N 1 we have the recurrence



 

n bj
n
n bj
.
(4.4)
+
=
k
k bj b
k b
b
Applying this recurrence to (4.1) yields


n 
n 
X
X
n k bj
n k bj
b,j
+
Fn =
k
k bj
b
b
k=0

k=0

j
nb
X

k=0

=
=

j
n2b
X

n 2bj (k bj )
k bj

k=0
b,j
Fn2b
j

n 2bj k
k
+

+
b

b,j
Fnb
j.

+
b

j
nb
X

k=0


n 
X
n k bj
k=0



n k bj
k
b

This completes the proof of the second part of the theorem.

Rather than the Stirling numbers, which have combinatorial significance, the Fibonacci numbers are well
studied objects in number theory. Checking base cases with j = 0 shows that F0b,0 = 1 and F1b,0 = 1, while
b,0
b,0
Fnb,0 = Fn2
+ Fn1
, showing that Fnb,0 = Fn and the base-b Fibonacci numbers reduce to the regular Fibonacci
numbers. Questions for the future include exploring whether these base-b Fibonacci numbers conserve other
properties of the Fibonacci numbers, and why we can represent them as independent of j while the recurrence
varies for different j.
b,j
b,j
We could also provide a generating function. Denote N = bj so that Fnb,j = FnN
+ Fn2N
with initial
conditions F0 , . . . , F2N 1 . This gives the generating function
P2N 1 b,j n PN 1 b,j n
X
Fn z n=0 Fn z
b,j n
.
(4.5)
Fn z = n=0
1 z N z 2N
n0

This gives also a Binet formula: the 2N roots of the equation z 2N z N 1 = 0 are
1

|| N e

k2
N

, || N ek N , 0 k N 1

(4.6)

with
1 +
=
2

1
=
2

(4.7)

and the general form for the Fibonacci numbers is


Fnb,j

N
1
X
k=0

k || N e

k2
N

+ k || N e

k2
N

(4.8)

where the 2N constants k , k are determined by the initial values F0 up to F2N 1 .


6

The base-b exponential

Analogously to the classical exponential function, we are interested in studying a modified exponential function
eb (x, w) :=

X
xsb (k)
k=0

(k!)b

wk .

(5.1)

We note that without the weighting from wk this series has a zero radius of convergence. We can relate this to
the classical exponential function, but require the following formula about the upper incomplete gamma function
(a, z)[8, (8.2.2)], where
Z
ta1 et dt.
(5.2)
(a, z) :=
z

From [8, (8.4.8)], we have


n
X
zk
k=0

= ez

k!

(n + 1, z)
.
n!

(5.3)

We can then prove a factorization theorem related to the base-b exponential, and provide very good approximations for it.
Lemma 8.
{k : 0 k b n, 0 ni b 1, 0 i < } = {k : 0 k }.

(5.4)

Proof . There is a bijection between both sets. To see this, fix a base b. Every natural number has a unique
representation in base b which occurs in the left-most set, while every number in the left-most set corresponds
to a unique natural number k.

Theorem 9.
eb (x, w) = exp x

i=0

bi

Y
i=0

(b, xwb )
(b 1)!

exp x

X
i=0

bi

exw+xw .

(5.5)

Proof . We now extend the summation over digitally dominated k in (2.1) to a sum over all natural numbers k.
This means by letting the number of digits in (2.1) tend to infinity while letting each ni = b 1 and applying
(5.4) we can convert our sum over digitally dominated k to a sum over natural numbers.
k
i
Fixing a base b and real x, w and setting f (~
ai ) = xki i! wb ki in (2.1) while taking N , we can simplify the
product on the right hand side by noting

Y
xk i
i=0

ki !

w b ki =

P i
xs (k)wk
1 P
x i=0 ki w i=0 b ki = b
.
(k!)b
(k!)b

(5.6)

Therefore, the right hand side simplifies to eb (x, w). Meanwhile, simplifying the sum on the left hand side by
applying (5.3), we obtain
ni
X
Y

i=0 ki =0

X
b1
 (b, xwbi )

Y
X
Y
i
i
(xwb )ki
exp xwb
xwb
=
= exp
f (~ai ) =
ki !
(b 1)!

i=0 ki =0

i=0

i=0

Y
i=0

(b, xwb )
(b 1)!

(5.7)

Equating these representations proves the first part of the theorem. The behavior of our base-b exponential then
P
bi
depends on the series
i=0 w and an infinite gamma product. For b = 2, we obtain the corollary
e2 (x, w) =



Y
i
1 + xwb .

(5.8)

i=0

We now prove the first approximation. Looking at the infinite product term on the right-hand side shows
that, even for small values of b, and |x| < 1 it can be approximated by the indicator function of the interval
[0, 1]
(
1, 0 w < 1
f (w) =
(5.9)
0, w > 1
This can be explained as follows: the function


i
b, xwb
w 7
(b)

(5.10)

is strictly decreasing over [0, 1] from 1 at w = 0, to 0 <

(b,x)
(b)

< 1. Moreover, as b increases, the ratio

(b,x)
(b)

increases to 1. For 0 < w < 1, wb is close to 0 for i large, so that using the asymptotic expansion [8, (8.7.3)]


1
z
b
(b, z) = (b) + z +
+ ...
(5.11)
b 1+b
we obtain


i
b, xwb
(b)

i+1

=1

i+1

xb w b
xb+1 wb +b
+
+ ...
(b + 1)
(b + 1) (b)
P

i+1
wb

(5.12)

Q 
is convergent so that i 1

As a consequence, for 0 w < 1,


Moreover, from


i+1
1
i
log b, xwb log (b) xb wb ,
b

i+1

xw b
(b+1)

and



i
b,xw b
i

(b)

are convergent.

(5.13)

we deduce, for 0 X < 1,

1X
X


1 X bi+1
1 wb
i
i+1


log b, xwb log (b) =

xb w b
w

b
b
b1w
i0

(5.14)

i0

which goes to 0 as b , so that the left-hand side goes to 0 as well. This yields the first approximation for
x [0, 1], which can in turn be approximated as
eb (x, w) exw+xw

(5.15)

since for x [0, 1],


ew

b2

ew .

(5.16)

8

Analogously to the classical identity ex ey = ex+y , we have a similar convolution identity for the base-b
exponential.
Theorem 10.
eb (x, w) eb (y, w) = eb (x + y, w),

(5.17)

where

ak xsb (k)

k=0

bl y sb (l) =

l=0

X
X

ak bnk xsb (k) y sb (nk) .

(5.18)

n=0 kb n

Proof . The theorem follows from taking ak =


(2.4).

wk
(k!)b

and bl =

wl
(l!)b

and applying the digital binomial theorem




The operation forms an analogue of multiplication for formal power series which instead applies to series
with involving powers of sb (k), where the convolution of two power series is defined as

ak xk

k=0

X
l=0

bl y l =

X
X

ak bnk xk y nk .

(5.19)

n=0 kn

We note that the operator must manually impose the condition k b n because of the basic result that
 
n
(n!)b
,
(5.20)
6=
k b (k!)b (n k)!b

since the left hand side is zero for k not digitally dominated by n, while the right hand side is well-defined and
in general non-zero for such k.

References
[1] T. Mansour and H. D. Nguyen,
http://arxiv.org/abs/1506.07945.

q-digital

binomial

theorem,

preprint,

2015,

[2] T. Mansour and H. D. Nguyen, A digital binomial theorem for Sheffer sequences, preprint, 2015,
http://arxiv.org/abs/1510.08529.
[3] H. D. Nguyen, A digital binomial theorem, preprint, 2014, http://arxiv.org/abs/1412.3181.
[4] H. D. Nguyen, A generalization of the digital binomial theorem, J. Integer Seq. 18 (2015), Article 15.5.7.
[5] L. Jiu and C. Vignat, On Binomial
http://arxiv.org/abs/1602.04149.

Identities

in

Arbitrary

Bases,

preprint,

2016,

[6] T. Agoh and K. Dilcher, Generalized convolution identities for Stirling numbers of the second kind, Integers:
Electronic Journal of Combinatorial Number Theory 8 (2008), Article #A25
[7] J. Riordan, Combinatorial Identities, R. E. Krieger Pub. Co., 1979.
[8] NIST Digital Library of Mathematical Functions. http://dlmf.nist.gov, Release 1.0.9 of 2014-08-29.
Online companion to [9].
[9] F. W. J. Olver and D. W. Lozier and R. F. Boisvert and C. W. Clark, editor. NIST Handbook of Mathematical Functions. Cambridge University Press, New York, NY, 2010. Print companion to [8].