The Gamma and Beta Functions

Cambridge Books Online
http://ebooks.cambridge.org/
Special Functions
George E. Andrews, Richard Askey, Ranjan Roy
Book DOI: http://dx.doi.org/10.1017/CBO9781107325937
Online ISBN: 9781107325937
Hardback ISBN: 9780521623216
Paperback ISBN: 9780521789882
Chapter
1 - The Gamma and Beta Functions pp. 1-60
Chapter DOI: http://dx.doi.org/10.1017/CBO9781107325937.002
Cambridge University Press

The Gamma and Beta Functions
Euler discovered the gamma function, F(x), when he extended the domain of the
factorial function. Thus F(x) is a meromorphic function equal to (x — 1)! when x
is a positive integer. The gamma function has several representations, but the two
most important, found by Euler, represent it as an infinite integral and as a limit of
a finite product. We take the second as the definition.
Instead of viewing the beta function as a function, it is more illuminating to think
of it as a class of integrals - integrals that can be evaluated in terms of gamma
functions. We therefore often refer to beta functions as beta integrals.
In this chapter, we develop some elementary properties of the beta and gamma
functions. We give more than one proof for some results. Often, one proof gener-
alizes and others do not. We briefly discuss the finite field analogs of the gamma
and beta functions. These are called Gauss and Jacobi sums and are important in
number theory. We show how they can be used to prove Fermat's theorem that a
prime of the form An + 1 is expressible as a sum of two squares. We also treat a
simple multidimensional extension of a beta integral, due to Dirichlet, from which
the volume of an n-dimensional ellipsoid can be deduced.
We present an elementary derivation of Stirling's asymptotic formula for n! but
give a complex analytic proof of Euler's beautiful reflection formula. However, two
real analytic proofs due to Dedekind and Herglotz are included in the exercises. The
reflection formula serves to connect the gamma function with the trigonometric
functions. The gamma function has simple poles at zero and at the negative inte-
gers, whereas esc nx has poles at all the integers. The partial fraction expansions
of the logarithmic derivatives of F(JC) motivate us to consider the Hurwitz and
Riemann zeta functions. The latter function is of fundamental importance in the
theory of distribution of primes. We have included a short discussion of the func-
tional equation satisfied by the Riemann zeta function since it involves the gamma
function.
In this chapter we also present Kummer's proof of his result on the Fourier
expansion of log T(x). This formula is useful in number theory. The proof given
Downloaded from Cambridge Books Online by IP 128.122.253.212 on Wed Feb 18 18:52:10 GMT 2015.
http://dx.doi.org/10.1017/CBO9781107325937.002
Cambridge Books Online © Cambridge University Press, 2015
2 1 The Gamma and Beta Functions
uses Dirichlet's integral representations of log F(x) and its derivative. Thus, we
have included these results of Dirichlet and the related theorems of Gauss.
1.1 The Gamma and Beta Integrals and Functions

The problem of finding a function of a continuous variable x that equals n! when
x = n, an integer, was investigated by Euler in the late 1720s. This problem was
apparently suggested by Daniel Bernoulli and Goldbach. Its solution is contained
in Euler's letter of October 13,1729, to Goldbach. See Fuss [1843, pp. 1-18]. To
arrive at Euler's generalization of the factorial, suppose that x > 0 and n > 0 are
integers. Write
^ ± ^ (1.1.1)
(* + !)»
where (a)n denotes the shifted factorial defined by
(fl) B =a(a + l)---(a + i i - l ) forn > 0, (a) 0 = 1, (1.1.2)
and a is any real or complex number. Rewrite (1.1.1) as
Since
hm = 1,
x
we conclude that
nln
jc!=Jm^ * . (1.1.3)
Observe that, as long as x is a complex number not equal to a negative integer, the
limit in (1.1.3) exists, for
n\nx ( n V TT-
and
^ x(x -
1.1 The Gamma and Beta Integrals and Functions
Therefore, the infinite product
n( 1+ 7) 0+7
converges and the limit (1.1.3) exists. (Readers who are unfamiliar with infinite
products should consult Appendix A.) Thus we have a function
I I ( J C ) = lim ' (1.1.4)

fcôo (x + 1)*
defined for all complex x ^ — 1, —2, — 3 , . . . and Ti(n) = n\.
Definition 1.1.1 For all complex numbers x ^= 0, — 1, — 2 , . . . , the gamma func-

tion T(x) is defined by
k\kx~l
T(x)= lim ——-. (1.1.5)
kôc (x)k
An immediate consequence of Definition 1.1.1 is
F(JC + 1) =xT(x). (1.1.6)

Also,
l)=/i! (1.1.7)
follows immediately from the above argument or from iteration of (1.1.6) and use
of
F(l) = l. (1.1.8)
From (1.1.5) it follows that the gamma function has poles at zero and the negative
integers, but 1/ F(x) is an entire function with zeros at these points. Every entire
function has a product representation; the product representation of 1/F(JC) is
particularly nice.
Theorem 1.1.2
00
r /
)e-x/n\, (1.1.9)
r(x)
where y is Euler's constant given by
y= lim ( V l - l o g n ) . (1.1.10)
Proof.
1 *(* + ! ) . . . ( * + „ _ i)
= lim :
- V*7*
n=\
The infinite product in (1.1.9) exists because
and the factor e~*/n was introduced to make this possible. The limit in (1.1.10)
exists because the other limits exist, or its existence can be shown directly. One
way to do this is to show that the difference between adjacent expressions under
the limit sign decay in a way similar to l/n2. •
One may take (1.1.9) as a definition of T(x) as Weierstrass did, though the
formula had been found earlier by Schlomilch and Newman. See Nielsen [1906,
p. 10].
Over seventy years before Euler, Wallis [ 1656] attempted to compute the integral
/ J y/1 - x2dx = ±f^(l-x)l/2(l+x)l/2dx. Since this integral gives thearea of
a quarter circle, Wallis's aim was to obtain an expression for n. The only integral
he could actually evaluate was Jo xp{\ — x)qdx, where p and q are integers or
q = 0 and p is rational. He used the value of this integral and some audacious
guesswork to suggest that
4 A
(l.l.H)
Of course, he did not write it as a limit or use the gamma function. Still, this
result may have led Euler to consider the relation between the gamma function
and integrals of the form Jo xp(l — x)qdx where p and q are not necessarily
integers.
Definition 1.1.3 The beta integral is defined for Rex > 0, Re y > 0 by
B(x,y)= I tx~\\-t)y~ldt. (1.1.12)

Jo
1.1 The Gamma and Beta Integrals and Functions 5
One may also speak of the beta function B(x,y), which is obtained from the
integral by analytic continuation.
The integral (1.1.12) is symmetric in x and y as may be seen by the change of
variables u = 1 — t.
Theorem 1.1.4
Remark 1.1.1 The essential idea of the proof given below goes back to Euler
[1730, 1739] and consists of first setting up a functional relation for the beta
function and then iterating the relation. An integral representation for T(x) is
obtained as a byproduct. The functional equation technique is useful for evaluating
certain integrals and infinite series; we shall see some of its power in subsequent
chapters.
Proof. The functional relation we need is
B{x, y) = ^-^B(x, y + 1). (1.1.14)

y
First note that for Re x > 0 and Re y > 0,
l
r
B(x,y + 1)= / tx-l(l-t)(l-t)y-ldt
Jo
= B(x,y)-B(x + l,y). (1.1.15)
However, integration by parts gives
B(x,y + 1)= \-tx(l-ty] +- / tx(l-ty~ldt

L* Jo x Jo
= -B(x + l,y). (1.1.16)
x
Combine (1.1.15) and (1.1.16) to get the functional relation (1.1.14). Other proofs
of (1.1.14) are given in problems at the end of this chapter. Now iterate (1.1.14) to
obtain
x
B(x, y) = + y x + y+ B(X + 2) = ... =
x y n
B(x, y + n).
y(y + 1 ) (y)n
Rewrite this relation as
)nJo n
v —1 rn / J. \ n-\-y—1
(,+,)„„!„-
y />.(!_£-,
[ \ \ ( L \ du
n\n*+y-^ (y)nn Jo
Jo \ n
As n -> oo, the integral tends to /0°° tx~xe~ldt. This may be justified by the
Lebesgue dominated convergence theorem. Thus
F(y) f°° 1 ,
B(x, y) = ^ / tx-le-'dt. (1.1.17)
Set y = 1 in (1.1.12) and (1.1.17) to get
I = f f-Ut = B(x, 1) = -P^- (

x Jo r(x +1) Jo
Then (1.1.6) and (1.1.8) imply that /0°° tx-le-'dt = V(x) for R e x > 0. Now use
this in (1.1.17) to prove the theorem for Rejc > 0 and Re v > 0. The analytic
continuation is immediate from the value of this integral, since the gamma function
can be analytically continued. •
Remark 1.1.2 Euler's argument in [1739] for (1.1.13) used a recurrence relation
in x rather than in y. This leads to divergent infinite products and an integral that
is zero. He took two such integrals, with y and y = m, divided them, and argued
that the resulting "vanishing" integrals were the same. These canceled each other
when he took the quotient of the two integrals with y and y = m. The result was an
infinite product that converges and gives the correct answer. Euler's extraordinary
intuition guided him to correct results, even when his arguments were as bold as
this one.
Earlier, in 1730, Euler had evaluated (1.1.13) by a different method. He expanded
(1 — t)y~l in a series and integrated term by term. When y = n + 1, he stated the
value of this sum in product form.
An important consequence of the proof is the following corollary:
Corollary 1.1.5 For Re x > 0
r(jc) = / tx-xe-fdt. (1.1.18)

Jo
The above integral for T(x) is sometimes called the Eulerian integral of the
second kind. It is often taken as the definition of V(x) for Rex > 0. The Eulerian
integral of the first kind is (1.1.12). Legendre introduced this notation. Legendre's
F(x) is preferred over Gauss's function U(x) given by (1.1.4), because Theorem
1.1.4 does not have as nice a form in terms of n (x). For another reason, see Section
1.10.
The gamma function has poles at zero and at the negative integers. It is easy
to use the integral representation (1.1.18) to explicitly represent the poles and the
1.1 The Gamma and Beta Integrals and Functions 7
analytic continuation of T(x):
p\ pOO
F(x) = I tx-xe-tdt+ \ tx-le-ldt
Jo J\
tX le dt (1U9)
~ " -
The second function on the right-hand side is an entire function, and the first shows
that the poles are as claimed, with (— l)n/n\ being the residue at x = —n, n =
0,1,....
The beta integral has several useful forms that can be obtained by a change of
variables. For example, set t = s/(s + 1) in (1.1.12) to obtain the beta integral on
a half line,
(U.2O)
sY+y
Then again, take t = sin2 6 to get
T / 2 sin 2 - 1 0 cos2^"1 Odd = r ( * ) r ° ° . (1.1.21)

Jo 2T(x + y)
Put x = y = 1/2. The result is
_
2r(i) 2'
or
(1.1.22)
Since this implies [F(|)] 2 = 7r/4, we have a proof of Wallis's formula (1.1.11).
We also have the value of the normal integral
pOQ pOO pOO
/ e~x2dx = 2 e~x2dx = / r ^ V d f = T(l/2) = Jn. (1.1.23)

J-oo Jo Jo
Finally, the substitution t = (u — a)/(b — a) in (1.1.12) gives
rb
/ (b-u)x~l(u-a)y~ldu = (b-a)x+y-lB(x, y) = (b-a)x+ r(x+y)
Ja (1.1.24)
The special case a = — l,b = 1 is worth noting as it is often used:
(1.1.25)
-\ y)
A useful representation of the analytically continued beta function is
( j)( a
This follows immediately from Theorem 1.1.2. Observe that B(x, y) has poles at
x and y equal to zero or negative integers, and it is analytic elsewhere.
As mentioned before, the integral formula for Y(x) is often taken as the defini-
tion of the gamma function. One reason is that the gamma function very frequently
appears in this form. Moreover, the basic properties of the function can be devel-
oped easily from the integral. We have the powerful tools of integration by parts
and change of variables that can be applied to integrals. As an example, we give
another derivation of Theorem 1.1.4. This proof is also important because it can
be applied to obtain the finite field analog of Theorem 1.1.4. In that situation one
works with a finite sum instead of an integral.
Poisson [1823] and independently Jacobi [1834] had the idea of starting with
an appropriate double integral and evaluating it in two different ways. Thus, since
the integrals involved are absolutely convergent,
oo poo /»oo ppoo y l s

/ s - e' ds = T(x)V(y).
/ x x y x {s+t) x l
/ t - s - e- dsdt = / t - e-'dt /
Apply Jo
the change of variables s =Jouv and t = u(l Jo — v) to the double integral,
and observe that 0 < u < oo and 0 < v < 1 when 0 < s, t < oo. This change
of variables is suggested by first setting s + t = u. Computation of the Jacobian
gives dsdt = ududv and the double integral is transformed to
OO rl
/
e~uux+y~ldu / vx~\l - v)y~ldv = T(x + y)B(x, y).
Jo
A comparison of two evaluations of the double integral gives the necessary result.
This is Jacobi's proof. Poisson's proof is similar except that he applies the change
of variables t = r and s = ur to the double integral. In this case the beta integral
obtained is on the interval (0, oo) as in (1.1.20). See Exercise 1.
To complete this section we show how the limit formula for T (x) can be derived
from an integral representation of F(x). We first prove that when n is an integer
> 0 and Rex > 0,
n
-t)'dt= -
/' x(x
Jo
This is actually a special case of Theorem 1.1.4 but we give a direct proof by
induction, in order to avoid circularity in reasoning. Clearly (1.1.27) is true for
1.2 The Euler Reflection Formula
n = 0, and
f tx-\\-t)n+ldt= [ tx-\\-t)(\-t)ndt
Jo Jo
n\ n\
This proves (1.1.27) inductively. Now sett = u/n and let« -> oo. By the Lebesgue
dominated convergence theorem it follows that
f
Jo
tx~le~fdt = lim — forRejc > 0.
Thus, if we begin with the integral definition for F(JC) then the above formula can
be used to extend it to other values of x (i.e., those not equal to 0 , - 1 , — 2 , . . . ) .
Remark 1.1.3 It is traditional to call the integral (1.1.12) the beta function. A
better terminology might call this Euler's first beta integral and call (1.1.20) the
second beta integral. We call the integral in Exercise 13 Cauchy's beta integral.
We shall study other beta integrals in later chapters, but the common form of these
three is Jc[li(t)]p[l2(t)]qdt, where l\(t) and l2if) are linear functions of t, and
C is an appropriate curve. For Euler's first beta integral, the curve consists of a
line segment connecting the two zeros; for the second beta integral, it is a half line
joining one zero with infinity such that the other zero is not on this line; and for
Cauchy's beta integral, it is a line with zeros on opposite sides. See Whittaker and
Watson [1940, §12.43] for some examples of beta integrals that contain curves of
integration different from those mentioned above. An important one is given in
Exercise 54.
1.2 The Euler Reflection Formula

Among the many beautiful formulas involving the gamma function, the Euler
reflection formula is particularly significant, as it connects the gamma function
with the sine function. In this section, we derive this formula and briefly describe
how product and partial fraction expansions for the trigonometric functions can be
obtained from it. Euler's formula given in Theorem 1.2.1 shows that, in a sense,
the function 1/ F(JC) is half of the sine function.
Theorem 1.2.1 Euler's reflection formula:
*) = -r^—. (1.2.1)
Sin7TJt
Remark The proof given here uses contour integration. Since the gamma function
is a real variable function in the sense that many of its important characterizations
occur within that theory, three real variable proofs are outlined in the Exercises.
See Exercises 15, 16, and 26-27.
Since we shall show how some of the theory of trigonometric functions can be
derived from (1.2.1), we now state that sin x is here defined by the series
^ x + .
The cosine function is defined similarly. It is easy to show from this definition that
sine and cosine have period In and that eni = -l.SeeRudin[1976,pp. 182-184].
Proof. Set y = 1 - JC, 0 < x < 1 in (1.1.20) to obtain
f00 tx~l
r(*)r(l-*)= / r—dt. (1.2.2)
Jo 1 + *
J I+
To compute the integral in (1.2.2), consider the integral
where C consists of two circles about the origin of radii R and e respectively,
which are joined along the negative real axis from —R to —e. Move along the
outer circle in the counterclockwise direction, and along the inner circle in the
clockwise direction. By the residue theorem
zx
c 1-z
L dz = -2TTI, (1.2.3)
when zx~l has its principal value. Thus

f iRxeix®
iRe ff€ ttx~le
eixn ff~K iee
iexeix® ffR ttx~*e~
e i
-2jti= rxdO+ — dt+ I ^>d0+ — dt.
Let R —• oo and 6 -^ 0 so that the first and third integrals tend to zero and the
second and fourth combine to give (1.2.1) for 0 < JC < 1. The full result follows
by analytic continuation. One could also argue as follows: Equality of (1.2.1) for
0 < x < 1 implies equality in 0 < Rex < 1 by analyticity; for Rex = 0, x ^ 0
by continuity; and then for x shifted by integers using Y(x + 1) = xT(x) and
sin(x + n) = — sinjc. •
The next theorem is an immediate consequence of Theorem 1.2.1.
Theorem 1.2.2
n=\
1.2 The Euler Reflection Formula 11
OO / n ^
1
71 COt TTJt — - + 5Z (• v 1 ™
(
v
\1 !•l l
A , ^l.Z.DJ
k=—n
It ! . (-1)"
sin7rjc X ' x2 - n2 n Y \r
n=l A A-
n
7i tan n x lim
/ J h
(1.2.7)
k=—n :+±-Jc'
it sec 7i x = lim (-1)* (1.2.8)

+ x + -'
2 1
2
sin 7t x (X + ( • •")
n=—oo
Formula (1.2.4) follows from the product formula
proved in the previous section and from Theorem 1.2.1 intheformr(jc)r(l—x) =

—jcF(jc)r(—JC) = 7r/sin7rjc.
Formula (1.2.5) is the logarithmic derivative of (1.2.4), and (1.2.6) follows from
(1.2.5) since esc x = cot | — cot JC. The two formulas (1.2.7) and (1.2.8) are merely
variations of (1.2.5) and (1.2.6). Formula (1.2.9) is the derivative of (1.2.5). •
It is worth noting that (1.2.6) follows directly from (1.2.1) without the product
formula. We have
poo tx-\ P\ tx-\ POO tx-\
JCCSC7TX = / dt = / dt + / dt
Jo l + t Jo l + t Jx l + t
p\ tx-l_^_t-x
= / dt - ~l+rx) dt
Jo l+t
n+l
where
(tn+x +tn~x+l)dt
0 n +x + 1 /i — x +2
Thus (1.2.6) has been derived from (1.2.1).
Before going back to the study of the gamma function we note an important
consequence of (1.2.5).
Definition 1.2.3 The Bernoulli numbers Bn are defined by the power series ex-
pansion
r2k
^ J
It is easy to check that -^—^ + | is an even function. The first few Bernoulli
numbers are Bx = - 1 / 2 , B2 = 1/6, B4 = -1/30, B6 = 1/42.
Theorem 1.2.4 For each positive integer k,
n=\
Proof. By (1.2.10)
eix + e-ix 2ix ^ .,,
=
jccotx = ix-fx z^ *x + ~2~i^c =
1 ~ / ,(~ 1) B2k
and (1.2.5) gives the expansion

°° X2 oo oo x2k
xcotx = i + 2 y ^ -r—T—r = i — 2 y ^ y ^ —~——.
^ x2 - n2n2 ^ ^ «2fc7r2/:
Now equate the coefficients of xlk in the two series for x cot* to complete the
proof. •
Eisenstein [1847] showed that a theory of trigonometric functions could be sys-
tematically developed from the partial fractions expansion of cot x, taking (1.2.5)
as a starting point. According to Weil [1976, p. 6] this method provides the sim-
plest proofs of a series of important results on trigonometric functions orginally
due to Euler. Eisenstein's actual aim was to provide a theory of elliptic functions
along similar lines. A very accessible account of this work and its relation to mod-
ern number theory is contained in Weil's book. Weil refers to limôo X]-« ak a s
Eisenstein summation.
Theorem 1.2.2 shows that series of the form
1
-
^(x+ n)k'
where k is an integer, are related to trigonometric functions. As we shall see next,
the "half series"
oo 1
l
y
1.2 The Euler Reflection Formula 13
bears a similar relationship to the gamma function. In fact, one may start the study
of the gamma function with these half series.
Theorem 1.2.5
r'(l) = - y , (1.2.12)
r(x)
Proof. Take the logarithmic derivative of the product for 1 / F (x). This gives
-r(x) =
The case x = 1 gives (1.2.12). The other two formulas follow immediately. •
Corollary 1.2.6 Log F(x) is a convex function of x for x > 0.
Proof The right side of (1.2.14) is obviously positive. •
Remark The functional equation (1.1.6) and logarithmic convexity can be used
to derive the basic results about the gamma function. See Section 1.9.
We denote Fr (JC)/ F (x) by i/r (x). This is sometimes called the digamma function.
Gauss proved that i/r(x) can be evaluated by elementary functions when x is a
rational number. This result is contained in the next theorem.
Theorem 1.2.7
1 1 1
xls(x+n) =X - +Jt +—1 - + n = 1,2, 3 , . . . , (1.2.15)
X+ft— 1
, p\ it up 2nnp . *n\
cos i ogU sin I,
q i)
(1.2. 16)
where 0 < p < q;^ means that when q is even the term with index n = q/2 is
divided by 2. Here lq/2j denotes the greatest integer in q/2.
Proof The first formula is the logarithmic derivative of
T(x + n) = (x + n - 1)(% + n - 2) • • • xT{x).
We derive Gauss's formula (1.2.16) by an argument of Jensen [1915-1916] using
roots of unity. Begin with Simpson's dissection [1759]:
If fix) = EZo"nX\then
oo i k—1
kn m
"}Takn+mx + = -J2™-Jmf(uJx),
n=0 7=0
where w = e2ni^k is a primitive &th root of unity. This is a consequence of

£ * l j wJm = 0, m # O(modfc). Now by (1.2.13)
00
1
+ 1 p + nq
= lim V ( — - — ) t p + n q = : lim s(t)
by Abel's continuity theorem for power series. From the series — log(l — t)
Y1T=\ tn/n>an<^ Simpson's dissection with co = e2ni/q, we get
q-\
s(t) = -tp~q log(l - tq) + Yl^^ lo (1

§ ~ ^"^
= -tp~q log ]j^- - (tp~q - 1) log(l - 0

n=\
Letr -+ 1" to get
q-\
Replace p by q — p and add the two expressions to obtain
The left side is real, so it is equal to the real part of the right side. Thus
q
(1.2.17)
But
-jc) = - ^ i o g r ( x ) r ( i - j c ) = -7r
ax
1.3 The Hurwitz and Riemann Zeta Functions 15
So
\Kp/q) - \/f (I - (p/q)) = -n cotnp/q. (1.2.18)
Add this identity to (1.2.17) to get
, f P\ n np ^ 1^-y lixnp ( 2nn\

f I — ) = —y cot —- - log q + - > cos log 2 - 2 cos .
\qj 2 q 2j^x q \ q )
(1.2.19)
But cos 2n(q —ri)/q= cos 27tn/q, so the sum can be cut in half, going from 1
to [q/2j, where [xj denotes the greatest integer in x. Thus
\q/2\
( p\
— = —y
q)
it
2
up
cot
q
^—w 2nnp t /
log q + > cos
^ q
^ 2nn\
log 2 — 2 cos
V q )
n up , ^xr^' 2nnp A / . nn\
= —y cot log q + 2 > cos log 2 sin — . •
2
q frf q V q )
1.3 The Hurwitz and Riemann Zeta Functions

The half series
OO j
forx>0, (1.3.1)
called the Hurwitz zeta function, is of great interest. We have seen its connection
with the gamma function for positive integer values of s in the previous section.
Here we view the series essentially as a function of s and give a very brief discussion
of how the gamma function comes into the picture.
The case x = 1 is called the Riemann zeta function and is denoted by f (s). It
plays a very important role in the theory of the distribution of primes. The series
converges for Res > 1 and defines an analytic function in that region. It has a
continuation to the whole complex plane with a simple pole at s = 1. The analytic
continuation of f (s) up to Res > 0 is not difficult to obtain. Write the series for
f (s) as a Stieltjes integral involving [*J. Thus for Res > 1
[xjdx
The last integral converges absolutely for Re s > 0 and we have the required
continuation. The pole at s = 1 has residue 1 and, moreover,
= lim f, / .4H
n
I r I— r \
= lim ( V - - l o g n ) =y. (1.3.2)

m=\ /
The best way to obtain analytic continuation to the rest of the plane is from the
functional relation for the zeta function. We state the result here, since the gamma
function is also involved. There are several different proofs of this result and we
give a nice one due to Hardy [1922], as well as some others, in the exercises. In
Chapter 10 we give yet another proof.
Theorem 1.3.1 For all complex s,

((1)/2)
s). (1.3.3)
If 5 < 0, then 1 — s > 1 and the right side provides the value of t;(s). This
relation was demonstrated by Euler for integer values of s as well as for s = 1 /2 and
s = 3/2. He had proofs for integer values of s, using Abel means. An interesting
historical discussion is contained in Hardy [1949, pp. 23-26]. The importance of
t;(s) as a function of a complex variable in studying the distribution of primes was
first recognized by Riemann [1859].
The last section contained the result
t(2k) = — B2k7r2k.
S
(2*)!
The following corollary is then easy to prove.
Corollary 1.3.2
?(1 - 2k) = —B2k, ?(0) = — and ((-2k) =0 for k = 1, 2, 3 , . . . .
(1.3.4)
Corollary 1.3.3
(1.3.5)
, v , 2
Proof From the functional equation and the fact that
l-s V 2
1.3 The Hurwitz and Riemann Zeta Functions 17
we have
Now (1.3.2) implies that (s - l)f(s) = 1 + y (s - 1) + A(s - I) 2 H .So take

the logarithmic derivative of (1.3.6) to get
Set 51 = 1 and use Gauss's result in Theorem 1.2.7 with p = 1 and q = 2. This
proves the corollary. •
There is a generalization of the last corollary to the Hurwitz zeta function £ (JC , s).
A functional equation for this function exists, which would define it for all complex
s, but we need only the continuation up to some point to the left of Re s = 0. This
can be done by using the function f 0 ) . Start with the identity
00
?(*, s) - (Us) ~ sx$(s + D) = x~ + ]TV 5 [(1 + x/ny5 - (1 - sx/n)].

s
n=\
The sum on the right converges for Re s > — 1, and because £ (s) is defined for all
s, we have the continuation of f (*, s) to Res > — 1.
The following theorem is due to Lerch.
Theorem 1.3.4
s=0
s
Proof. The derivative of the equation t;(x + 1, s) = t;(x, s) — x with respect
to s at s = 0 gives
(1.3.8)
u
s=0 \ * / 5=0
For Re 5 > 1,
9JC 2
so
(139)
Now (1.3.8) and (1.3.9) together with (1.2.14) of Theorem 1.2.5 imply that
= C + logF(x).
ds
/ s=0
To determine that the constant C = — \ log lit, set x = 1 and use Corollary 1.3.3.
This completes the proof of Lerch's theorem. •
For a reference to Lerch's paper and also for a slightly different proof of Theorem
1.3.4, see Weil [1976, p. 60].
1.4 Stirling's Asymptotic Formula

De Moivre [1730] found that n\ behaves like Cnn+l^2e~n for large n, where C is
some constant. Stirling [1730] determined C to be V27r; de Moivre then used a
result of Stirling to give a proof of this claim. See Tweddle [1988, pp. 9-19]. This
formula is extremely useful and it is very likely that the reader has seen applications
of it. In this section we give an asymptotic formula for T{x) for Rex large, when
Imjc is fixed. First note that log T(x + n + 1) = Yl=\ l°g(k + x) + l o £ r ( * + 1).
We then employ the idea that an integral often gives the dominant part of the sum
of a series so that if the integral is subtracted from the series the resulting quantity
is of a lower order of magnitude than the original series. (We have already used
this idea in Equation (1.3.2) of the preceding section.) In Appendix D we prove the
Euler-Maclaurin summation formula, a very precise form of this idea when the
function being integrated is smooth. Two fuller accounts of the Euler-Maclaurin
summation formula are given by Hardy [1949, pp. 318-348] and by Olver [1974,
pp. 279-289].
Theorem 1.4.1
T{x) ~ \/7jzxx~xl1e~x as Rex —> oo.
Proof. Denote the right side of the equation

n-\
log r(jc + n) = ^2 l°S(k + *) + loS r(jc + 1)
by cn, so that
cn+i -cn =log(x + n).
By the analogy between the derivative and the finite difference we consider cn to
be approximately the integral of log(jc + ri) and set
cn = (n + x) \og(n +x)-(n+x) + dn.
1.4 Stirling's Asymptotic Formula 19
Substitute this in the previous equation to obtain
log(jc + n) = (TI + 1 + x)log(n + 1 + x) - (n + x) login + JC) + dn+l -dn-\.
Thus
1
4+1 - dn = 1 - in + x + 1) log 1 +
V n+x
Proceeding as before, take
and substitute in the previous equation to get
1 / 1
2
I2(n+x) $n+x)3J'
Now
therefore, limn_>oo(^n — eo) = K\ix) exists. Set

1
12(/I+JC) \ ( * + * ) :2) J'
where Kix) = K\ix) + eo. The term in + JC)"1 comes from completing the sum
in (1.4.1) to infinity and approximating the added sum by an integral. So we can
write
cn = in + x) login + x) - in + x) - - login + x)
where K(x) = log C(x). This implies that
12(n+x)
(1.4.2)
We claim C{x) is independent of x. By the definition of the gamma function

r(n + x) r(x) (x)n r(x) roo
hm ny = lim ?zJ = • = 1. (1.4.3)
r( + ) r()^(y)n r(y) T(x)
Now, from (1.4.2) and (1.4.3) we can conclude that
_V(n+x) C(x) . / x\n _ C(x)
1 = hm n = hm 1 + - \ e x = .
nôo r(n) C(0)nôo\ nJ C(0)
Thus C(x) is a constant and
F(JC) ~ Cx x ~ 1/2 ^~ x as Rex -> oo.
To find C, use Wallis's formula:

r 2(n$ 1
= hm p
(2)!
This gives C = \/2n and proves the theorem. Observe that the proof gives the
first term of an error estimate. •
We next state a more general result and deduce some interesting consequences.
A proof is given in Appendix D. For this we need a definition. The Bernoulli
polynomials Bn(x) are defined by
tn
ef — 1 ^—' n\ (1A4)
n=0
The Bernoulli numbers are given by Bn (0) = Bn for n > 1.
Theorem 1.4.2 For a complex number x not equal to zero or a negative real
number,
log
z \ /, /
7=1
l r°
•dt. (1.4.5)
2m Jo (x + 0 2 m
T/i^ v^/w^ of log x is the branch with log JC real when x is real and positive.
The expansion of log V(x) in (1.4.5) is an asymptotic series since the integral is
easily seen to be O(x~2m+l) for |argx\ < it - 8, 8 > 0.
1.4 Stirling's Asymptotic Formula 21
From this theorem the following corollary is immediately obtained.
Corollary 1.4.3 For 8 > 0 and |argx\ < n — 8,
F(x) ~ V2nxx~l/2e~x as \x\ -> oo.
Corollary 1.4.4 When x = a + ib, a\ < a < a2 and \b\ - • oo, then
\T{a + ib)\ = sFhx\b\a-ll2e-n^l2\\ + 0(l/|fc|)],
where the constant implied by O depends only on a\ and a2.
Proof. Take \b\ > I, a > 0. It is easy to check that the Bernoulli polynomial
B2 - B2(t) = t -t2. Thus \\B2 - B2(t)\ < \\t{\ - t)\ < \ for 0 < t < 1. So
(1.4.5) with m = 1 is
log T(a + ib)= (a + ib j log(a + ib) - (a + ib) + - log 2TT + R(X),
and
i r°° dt _ I r°° dt _ I _j \b\
~ 8 Jo k + ^ l 2 8^0 (a + t)2 + b2 S\b\ a'
Now
R e f a + ife - 2- ) l o g ( a + i b ) \ = (a--) 2 \og(a2 + b 2 ) 1 ' 2 - barctan -.

IA / J V / *
Also,
log(a2 +b2)l/2 = Uogb2 + ^log^l + ^)= log |*|
Moreover,
b a f^, iffe>0,
arctan - + arctan - = < 2 n
a b [-f, ifb<0.
This gives
b \ JZ a \
-b arctan - = -b\± - - - + O[—2
a \ 2 b \b
Putting all this together gives
log |r(a + ib)\ = ( a - ^ log |*| - | | * | + i log2^ + O (±- ).
The condition a > 0 is removed by a finite number of uses of the functional

equation (1.1.6) and the corollary follows. Observe that the proof only uses a =
o(\b\) rather than a bounded. •
Corollary 1.4.5 For |argjt| < n — 8, 8 > 0,
Corollary 1.4.4 shows that T{a + ib) decays exponentially in the imaginary
direction. This can be anticipated from the reflection formula, for
cosh**'
or
e7ib _i_ e —Tib

• 27te~nlbl as b -> ±oo,
or
as b - • ±oo.
n-+n
Similarly,
4
-ib sin 7rZ?/
and
as ±oo.
Since T(x) increases rapidly on the positive real axis and decreases rapidly in the
imaginary direction, there should be curves going to infinity on which a normalized
version of T(x) has a nondegenerate limit. Indeed, there are. See Exercise 18.
1.5 Gauss's Multiplication Formula for T(mx)

The factorization
together with the definition of the gamma function leads immediately to Legendre's
duplication formula contained in the next theorem.
Theorem 1.5.1
1.5 Gauss's Multiplication Formula for F(mx) 23
This proof suggests that one should consider the more general case: the factor-
ization of (a)mn, where m is a positive integer. This gives Guass's formula.
Theorem 1.5.2
T{ma){2nim-X)'2 = m ^ ^ F ^ F fa + -J-Y • • F (a + ^ - = - ^ . (1.5.2)
Proof. The same argument almost gives (1.5.2). What it gives is (1.5.2) but with
(27r)"r i m-5 replacedby

F( - ] . . . F ( — - ) =: P. (1.5.3)
\mj \ m J
To show that (1.5.3) is true, we show that
By the reflection formula
mj \ mj sin —
So it is enough to prove
m_, JZ . 2n . (m — l)n
2 sin — sin — • • • sin = m.
mm m
Start with the factorization
x —
— = ]J(x - exp(2kni/m)).
X
Let x —> 1 to obtain

m-\
m = TT(1 —Qxp(2k7ti/m))
k=\
i n 2it (m — 1)TT
= 2 sin — sin sin .
mm m
This proves (1.5.3). •
Remark 1.5.1 A different proof of (1.5.1) or (1.5.2) that uses the asymptotic
formula for F(JC) and the elementary property T(x + 1) = xF(x) is also possible.
In fact it is easily verifed that
* F(l/2)F(2x)
satisfies the relation g(x + 1) = g(x). Stirling's formula implies that g(x) ~ 1 as
x -> cxDsothatlim^ôo^^+w) = 1 when n is an integer. Since g(x-\-n) — g{x)
we can conclude that g (x) = 1. A similar proof may be given for Gauss's formula.
This is left to the reader.
An elegant proof of the multiplication formula using the integral definition of
the gamma function is due to Liouville [1855]. We reproduce it here.
The product of the gamma functions on the right side of (1.5.2) is
pOO pOO pOO
/ e-t'x^dxj e-x*x$+il/m)-ldx2--- e
Jo Jo Jo
poo poo poo
= / ... / e-U^2+-+xm)xa-lxa+a/m)
Jo Jo Jo
Introduce a change of variables:
_ z'H _ _
X\ — , X2 — X2, - • • » Xm — Xm.
The Jacobian is easily seen to be
mzm~l
x2 x3
and the integral can be written
/ •••/ exp - ( x2 + x3 + • • • + xm + )
JO Jo I V X2X3'"XmJj
KX2-"Xm) X2X3'"Xm
m
Set t = x2 + x3 -\ 1- xm + z /(x2x3 • - • xm), and rewrite the integral as
poo poo poo
/ ••• / e-'zma-Xx{Vm)-Xxflm)-1 • • • x«n-l)lm)-ldzdx2 ...dxM.
Jo Jo Jo
(1.5.4)
First compute
poo
poo poo m—\
poo in L
~
I
Jo Jo ,._-,
Clearly,
pOO pOO rn — L , ,
dl m-\ / / -t TT {j/m)-\dX2"'dXm
= —YYIZ1'1~V I ''' I e~l
Jn
Jo Jc\
Jo _T J X
2 ' ' ' Xm
J =
Now introduce a change of variables,
-Xm),X3 =X3,...,Xm =Xm,
1.5 Gauss's Multiplication Formula for F(rax) 25
and
' 1 — *^3 • *^4 I *** I in • Jv\ I <J / v^3 ** fit \/*
The Jacobian is
and ^ is given by
dl Z"00 r
°°
dz Jo
m-\
TT A
1 1 * V+l
7
/»oo
/»OO /»oo
/»OO m—i
/ / -fi T TX OVm)-1x ( ( m - l ) / / n ) - l ,X , ,
=—m / ••• / ^ ll 7+i i ^ 3''' dxmdx\
Jo Jo -_9
J —^
= —ml.
Therefore,
/ = Ce~mz.
To find C, set z = 0 in the integral for / as well as in the above equation and
equate to get
By (1.5.3), C = (27r) (m " 1) / 2 m- 1 / 2 and / = (2nfm-l)!2m-x/2e-mz. Substitution

in (1.5.4) gives
f-OO
mz ma l
Y(a)T(a + 1/m) • • • T{a + (m - \)/m) = / e- z ~
Jo
which is Gauss's formula.

Remark 1.5.2 We pointed out earlier that 1 / F (x) is a half of sin nx. In this sense
the duplication formula is the analog of the double angle formula
sin 2nx = 2 sin itx sin n [ x -\—

2
V
This is usually written as sin27rx = 2 sin ixx cos ixx and so is thought of as a
special case of the addition for sin(jc -h y). The gamma function does not have an
addition formula.
1.6 Integral Representations for Log T(x) and i/>(x)
In (1.2.13), we obtained
x- 1
from the product for 1/ F(x). We start this section by rederiving it from the beta
integral. Note that, for x > 1,
Jo
by term-by-term integration, which is valid because of uniform convergence in
[0, 1—e]. Now let e —> 0. By Abel's continuity theorem for power series,
rl °° (Y _
i \ # . J C — 2 i _ _ / i ^\ i, \ ^ vA
We can introduce log(l — t) in the beta integral JQ tx 2

(l — t)ydt by taking the
derivative with respect to y. In that case,
dy Jo ^y F(JC + y)
or
(x - l ) / g
r(x)r(y + i)r'(* + y)
y)2
The case v = 0 gives the necessary result. The differentiation is justified since
the integrands involved are continuous. Some care should also be taken of the fact
that the integrals are improper. The details are easy and left to the reader. The next
theorem gives the integral representations of \l/(x) due to Dirichlet and Gauss.
Theorem 1.6.1 For Re x > 0,
(i) ir(x) = [ - (e~z - l

\dz (Dirichlet),
x
Jo z\ (l+z) J
z
f°° (e~ e~xz \
(ii) = )dz (Gauss).
Jo Vz 1-e-zJ
1.6 Integral Representations for Log T(x) and x//(x) 27
Proof, (i) Evaluate the integral /0°° /* e~tzdtdz in two different ways by changing
the order of integration to get the formula
e-z _ e-sz
/ = logs. (1.6.1)
Jo
Similarly, the double integral
dsdz
Jo Jo z
when first integrated with respect to z yields (by (1.6.1))
d
f°° -s x-U a f°° s x-lj
/ e s logs ds = — / e s ds =
Jo dx Jo
If we integrate the double integral with respect to s we get
Equate the last two expressions to get Dirichlet's formula.

(ii) Gauss's formula is obtained from Dirichlet's by a change of variables:
a oo -z
—dz-
re
/ z(l+zY
dz
= lim { /
a 00 z
e~z
Js00
—dz - /
f
—dz +
e~'x
-dt
\
)dt}
since
rlog(l+8) e-z f8 1 8
/ —dz < -dz = log -— ^ 0 as
JS Z J\og(l+8) Z l0g(l + 8)
This proves (ii). •
The integrated form of the last theorem is given in the next result.
iogr(;c) =
Jo
and
f ( e e \ dt
Proof. The integrals in Theorem 1.6.1 are uniformly convergent for Re x > 8 >
0, so we can integrate from 1 to x under the sign of integration. The integrals in
Theorem 1.6.2 are the corresponding integrated forms.
A change of variables u = e~* in (ii) gives
) ^ (1.6.2)
There are two other integrals for log F(x) due to Binet that are of interest. These
are given in the next theorem. A proof of one of them is sketched and the other is
left as an exercise. See Exercise 43. •
/ 1\ 1 r°°/l 1 1 \e~tx
(i) l 0 g r ( * ) = [X-- l0gX-X + -l0g27T+ / - - T + -7—7 ) — d t
and
(IV
1\ , 1, fr*rcton(t/x)
(ii) log F(x) = JC - - logx - x + - Iog2jr + 2 /
\ 1J I Jo e— — i
Proof Gauss's formula in Theorem 1.6.1 together with Equation (1.6.1) give
f(x + l) = 4~ logr(* + 1) = -!- + logx

b - / (\ - - + —!—)e-txdt.
dx 2x Jo \2 t e'-lj
Integrate from 1 to x, changing the order of integration to get
Use log F(x + 1) = log F(x) + log JC to rewrite the above formula as
1 1 1
where
/"(' I •)!! d.6.3.

Jo \2 t e* - \) t
Stirling's formula applied above gives 7 = 1 — (1/2) log In.
The second Binet formula can be used to derive the asymptotic expansion for
log T(x) contained in Corollary 1.4.5.
1.7 Kummer's Fourier Expansion of Log F (x) 29
Expand l/(e2nt — 1) by the geometric series and integrate term by term to see
that
JO
The last equality comes from Theorem 1.2.4. Now,
1 +z2 1 +z2
gives, after integration,
f n3 n5 (-1)"" 1 tln~l {-\)n r* z2ndz

arctan(f/*) = - - - — + - — • • . +
JC 3 r 5 r 2w — 1
Substitute this in Binet's formula (ii) and use (1.6.4) to arrive at
n
1\ 1 B
(
x--J logx - x + -\og2n + Y, 2 . (2 • _'
- i ) nn /*°°00 / r*
- l )l Z" /^ /"' z 2n
2 (2n \ dt
~ Jo \Jo
For |arg JC| < | — 6, 6 > 0, it can be seen that 13^21 ^ c s c ^ f° r a ^ z > 0.
This implies that the last term involving the integral is O{, }n+\)- ô we have
\x 1
the asymptotic series but only for | arg JC | < | — € instead of | arg x | < n — €.
Whittaker and Watson [1940, §13.6] show how to extend the range of validity.
It is also possible to derive an asymptotic formula for log V(x) from Binet's first
formula. See Wang and Guo [1989, §3.12]. For references to the works of Gauss,
Dirichlet, Binet, and others, see Whittaker and Watson [1940, pp. 235-259]. •
1.7 Kummer's Fourier Expansion of Log T(x)

Kummer [1847] discovered the following theorem:
Theorem 1.7.1 For 0 < x < 1

F(x) 1 1 1 ^ : :
log —= = — Iog(2sin7rjc) + - ( y +log2jr)(l - 2x) + - Y j ^ sin27r^x,
V 2TT 2 2 7T f~^ A:
where y is Euler's constant.
Proof. Start with the identity
-log(l - e2nix) = e2nix + — + — + •••, 0 < x < 1.
The real and imaginary parts are
-\og(2sm7tx) = Yl 1 ( L7 - ! )
and
I(l-2,) = f ; ^ [ f . (1.7.2)
Since log T(x) is differentiable in 0 < x < 1, it has a Fourier expansion

OO 00
log r(jc) = Co + 2 ^2 ck c o s 2k7tx + 2 ^ D* sin 2knx,

k=\ k=\
where
Ck= Iogr(;c)cos2)brjtdjc and Dk = logT(jc) sm2knxdx. (1.7.3)

Jo Jo
We use Kummer's method to compute Ck and Dk. The Ck are easy to find. Take
the logarithm of Euler's reflection formula (1.2.1):
logF(jc)+logr(l -x) =
= log 2ix + cos 2nx -\— cos 4nx + • • • .
The Fourier series of log T(x) gives

logT(jc) + l o g T ( l - x) = 2C0 + 4Ci cos 27TJC -f 4C2 cos4nx -\
Equating the last two relations gives
Co = - Iog27r and Ck — — foxk> 1.

2 4k
Now use integral (1.6.2) for log T{x) in (1.7.3) so that
i i i sin2knxdudx
logw
But
sin 2knxdx = 0, / x sin 2knxdx = —

/
Jo
and
rl
ux sin 2knxdx =
The first two integrals are easy to solve and the third is the imaginary part of
1 fl 1 M-1
u Jo u logu + 2kni'
1.7 Kummer's Fourier Expansion of Log F (x) 31
Therefore,
2knt
-i K
—2kn
u((\ogu) 2
+4k2n2)
1 \ du
2kn ) logw'
or, with u = e~ ,
-Iknt
Take £ = 1 and we have
Moreover, x = 1 in Dirichlet's formula (Theorem 1.6.1) gives
2TT 2n Jo V l+ tjt
where y is Euler's constant. Therefore,
~ Y 1
By (1.6.1), thefirstintegral is log 2n and a change of variables from t to 1/f shows

that the second integral is 0. Thus
To find Dk, observe that

1 n~2nt
f°° e~2nt _ e-2knt
— e~ 1
kDk-Dx = — dt = — log/:,
In Jo 1t "" 2ll
In
where the integral is once again evaluated by (1.6.1). Thus
Dk = (y + log2A:7r), k = 1, 2, 3 , . . . .
The Fourier expansion is then
logF(x) = -log27T + ^ C ° S o 7 7 r X
+-(y+ Iog27r)"
k=\ LK n
}2
k=\ k=\
, 1 v^log/: .
Apply (1.7.1) and (1.7.2) to get the result. •
Kummer's expansion for log (F (X)/V2TT) and Theorem 1.3.4 have applications
in number theory. Usually they give different ways of deriving the same result.
This suggests that the Hurwitz zeta function itself has a Fourier expansion from
which Kummer's result can be obtained. Such a result exists and is simply the
functional equation for the Hurwitz function:
2F(1— s) ) . 1 ^coslmnx 1 ^ sinlmixx |

r(x,s) = — < sm -us > hcos -ns > >. (1.7.4)
I m=l m=l J
The functional equation for the Riemann zeta function is a particular case of this
when x = 1. See Exercises 24 and 25 for a proof of (1.7.4) and another derivation
of Kummer's formula.
1.8 Integrals of Dirichlet and Volumes of Ellipsoids

Dirichlet found a multidimensional extension of the beta integral which is use-
ful in computing volumes. We follow Liouville's exposition of Dirichlet's work.
Liouville's [1839] presentation was inspired by the double integral evaluation of
the beta function by Jacobi and Poisson.
Theorem 1.8.1 If V is a region defined by xt > 0, / = 1, 2 , . . . ,n,and^Xi < 1,

then for Re, at > 0,
Proof The proof is by induction. The formula is clearly true for n = 1. Assume
it is true for n = k. Then for a (k + 1)-dimensional V
f'. • dxk+l
l-xi-Jt2 xk
= f f~xx••• f' * \rlxrl---xakk+T1dxk+i---dXl
Jo Jo Jo
1 r\ pl-Xy—Xk-x
i-1 Ofjfc-1
= — / • • • / x\
(*k+\ Jo Jo
• ( 1 - JCI xk)ak+ldxkdxk-x' • • dxx .

Now set xk = (1 — xx — • • • — xk-x)t to get
&k+\ Jo Jo Jo Jo
ff-\..f-'-"-
Jo Jo Jo
Xakk_T\\ -xx xk.x)ak+ak+ldxk.x • • • dxx
1.8 Integrals of Dirichlet and Volumes of Ellipsoids 33
Compare this with the integral to which the change of variables was applied and
use induction to get
r(ak)T(aM + 1) («* +a*+i)(n?=i !>,-))I"(a*
This reduces to the expression in the theorem. •

Corollary 1.8.2 IfV is the region enclosed by xt > 0 and ]T(JC;/<2;)A < 1, then
f
r(i+ £(«,/„» '
Proof. Apply the change of variables, y/ = (xi/ai)Pi, / = 1 , . . . , n. Then
9X; 1 JC/
and the Jacobian is
The integral becomes
P\Pl'"Pn J
where V is defined by y; > 0 and J^ yt < 1. The corollary now follows from the
theorem. •
Corollary 1.8.3 The volume enclosed byXO/MK' < 1. î >0is

In particular the volume of the n-dimensional ellipsoid J2(xi/at)2 < 1 is
r(l+n/2)
Proof For the first part of the corollary take ax•. = 1. For the particular case take
Pi = 2 and use the fact that r ( | ) = \ V^r. •
Corollary 1.8.4 IfV is given by xt > 0 and ^ ( ^ - ) A < A. in Dirichlet's integral,
then its value is
Liouville also gave the following extension of Dirichlet's result, which can be
proven in the same way.
Theorem 1.8.5 If V consists of xt > 0 , t\ <Y^(xi/ai)p' < h and f is a continuous

function on {t\, t-i), then
I••• I Jt? 1 " 1 • • •xan"-lf{(xl/a1y< +••• + (xn/an)p"}dxi •••dxn
A related integral is given next.
Theorem 1.8.6 IfV is the setxt > 0, J2"=\ xi = h then

n
[••• J x ? - l
This is a surface integral rather than a volume integral, but it can be evaluated
directly by induction or from Corollary 1.8.2. It is also a special case of Theorem
1.8.5 when f(u) is taken to be the delta function at u = 1. This function is not
continuous, but it can be approximated by continuous functions.
1.9 The Bohr-Mollerup Theorem

The problem posed by Euler was to find a continuous function of x > 0 that equaled
n\ at x = n, an integer. Clearly, the gamma function is not the unique solution to
this problem. The condition of convexity (defined below) is not enough, but the
fact that the gamma function occurs so frequently gives some indication that it
must be unique in some sense. The correct conditions for uniqueness were found
by Bohr and Mollerup [1922]. In fact, the notion of logarithmic convexity was
extracted from their work by Artin [1964] (the original German edition appeared
in 1931) whose treatment we follow here.
Definition 1.9.1 A real valued function f on (a, b) is convex if
f(kx + (1 - X)y) < Xf(x) + (1 - X)f(y)
for x, y e (a, b) andO < X < 1.
Definition 1.9.2 A positive function f on(a, b) is logarithmically convex if log /

is convex on (a,b).
It is easy to verify that if / is convex in (a, b) and a < x < y < z < b, then
f(y) ~ fM ^ f{z) - fix) ^ f{z) - f(y)
y-x ~ z-x ~ z-y
With these definitions we can state the Bohr-Mollerup theorem:
1.9 The Bohr-Mollerup Theorem 35
Theorem 1.9.3 If f is a positive function on x > 0 and (i) / ( I ) = 1, (ii)

f(x + 1) = xf(x), and (Hi) f is logarithmically convex, then f(x) = T(x)for
x > 0.
Proof Suppose n is a positive integer and 0 < x < 1. By conditions (i) and (ii)
it is sufficient to prove the theorem for such x. Consider the intervals [n, n + 1],
[n + 1, n + 1 + JC], and [n + 1, n + 2]. Apply (1.9.1) to see that the difference
quotient of log /(JC) on these intervals is increasing. Thus
/(n + 1) 1 /(* + !+*) ^ / ( / i + 2)
log < - log < log .
/(") * fin + 1) /(w + 1)
Simplify this by conditions (i) and (ii) to get
[fr+W)(jc+wl) */(*)[ ^
x logrc < log < x log(rc + 1).
L n\ J
Rearrange the inequalities as follows:
Therefore,
n\i
/(JC) = lim
and the theorem is proved. •

This theorem can be made the basis for the development of the theory of the
gamma and beta functions. As examples, we show how to derive the formulas
poo
POO
fx l
T(JC) = / e-e-t
t - dt, J C > 0,
Jo
and
I o L \x -t y)
c > 0 and y > 0. (1.9.2)
We require Holder's inequality, a proof of which is sketched in Exercise 6. We

state the inequality here for the reader's convenience. If / and g are measurable
nonnegative functions on (a, b), so that the integrals on the right in (1.9.3) are
finite, and p and q positive real numbers such that \/p + \/q = 1, then
< I / fpdx 1 / gqdx I . (1.9.3)

\Ja ) \Ja )
It is clear that we need to check only condition (iii) for log F(JC). This condition
can be written as
>0 and a + )8 = l. (1.9.4)
Now observe that

/»O
= /
Jo
and apply Holder's inequality with a = \/p and ft = l/q to get (1.9.4).
To prove (1.9.2) consider the function
r(x + y)B(x,y)
T{y)
Once again we require the functional relation (1.1.14) for B(x, y). This is needed
to prove that / (x +1) = xf (x). It is evident that / (1) = 1 and we need only check
the convexity of log fix). The proof again uses Holder's inequality in exactly the
same way as for the gamma function.
We state another uniqueness theorem, the proof of which is left to the reader.
Theorem 1.9.4 If f(x) is defined for x > 0 and satisfies (i) / ( I ) = 1,

(ii) f(x + 1) = xfix), and (iii'jlinwoo fix + n)/[nxf(n)] = 1, then fix) =
For other uniqueness theorems the reader may consult Artin [1964] or
Anastassiadis [1964]. See Exercises 26-30 at the end of the chapter. Finally, we
note that Ahern and Rudin [1996] have shown that log |F(JC + iy)\ is a convex
function of x in Rex > 1/2. See Exercise 55.
1.10 Gauss and Jacobi Sums

The integral representation of the gamma function is
C
=Jor e-,-d±
h t
Here dt/t should be regarded as the invariant measure on the multiplicative group
(0, oo), since
diet) _ dt
ct ~~ t '
To find the finite field analog one should, therefore, look at the integrand e~cttx.
The functions e~ct and tx can be viewed as solutions of certain functional relations.
This point of view suggests the following analogs.
Theorem 1.10.1 Suppose f is a homomorphism from the additive group of real

numbers R to the multiplicative group of nonzero complex numbers C*, that is,
/ : / ? - > C*
1.10 Gauss and Jacobi Sums 37
and
f(x + y) = f(x)f(y). (1.10.1)

cx
/ / / is differentiable with /'(0) = c # 0, then f(x) = e .
Remark 1.10.1 We have assumed that f(x) ^ 0 for any x but, in fact, the relation
g(x + y) = g(x)g(y), where g : R —> C, implies that if g is zero at one point it
vanishes everywhere.
Proof. First observe that / ( 0 + 0) = / ( 0 ) 2 by (1.10.1). So /(0) = 1, since /(0)
cannot be 0. Now, by the definition of the derivative,
J w
r-+o t t-+0 t
= f (JC) lim
tô t
= cf(x).
Sof(x)=ecx. m
Remark 1.10.2 In the above theorem it is enough to assume that / is continuous
or just integrable. To see this, choose a y e / ? such that f* f(t)dt / 0. Then
fix) jy f(t)dt = jj f(X + odt = jxx+y f{t)dt. so
This equation implies that if / is integrable, then it must be continuous and hence
differentiable.
Corollary 1.10.2 Suppose g is a homomorphism from the multiplicative group
ofpositive reals R+ to C*, that is,
g(xy) = g(x)g(y). (1.10.2)

c
Then g(x) = x for some c.
Proof Consider the map f = g o exp : R —• C*, where exp(jc) = ex. Then /
satisfies (1.10.1) and g(ex) = ecx. This implies the result. •
A finite field has pn elements, where p is prime and n is a positive integer. For
simplicity we take n = 1, so the field is isomorphic to Z(/?), the integers modulo
p. The analog of / in (1.10.1) is a homomorphism
Since Z(/?) is a cyclic group of order p generated by 1 we need only specify

Also, tff(l)p = if(0) = 1 and we can choose any of the pth roots of unity as the
value of V^(l). We therefore have p different homomorphisms
2nijx/p
^.(JC) = e , j = 0 , 1 , . . . , / ? - 1. (1.10.3)
These are called the additive characters of the field. In a similar way the multi-
plicative characters are the p — 1 characters defined by the homomorphisms from
Z(p)* to C*. Here Z(p)* = Z(p) - {0}. Since Z(p)* is a cyclic group of order
p — 1, we have an isomorphism Z(/?)* = Z(p — 1). The p — 1 characters on
Z(/?)* can be defined by means of this isomorphism and (1.10.3). We denote a
multiplicative character by either x or 77, unless otherwise stated.
It is now clear how to define the "gamma" function for a finite field.
Definition 1.10.3 For an additive character ^ and multiplicative character Xi
we define the Gauss sums gj(Xi), j = 0 , 1,...,/? — 1 by the formula
P-\
x=0
where we extend the domain of Xi by setting X/(0) = 0.
It is sufficient to consider g(x) •= gi(x)> f° r when j ^ 0
gj(x) = X)xW^;W
= XU)8(X). d-10.5)
This formula corresponds to /0°° e~jxxs~ldx = T(s)/js, where j is a nonzero
complex number with positive real part. When j = 0 in (1.10.4) the sum is
)i which can be shown to be zero when x (x) ^ 1 for at least one value of x.
Theorem 1.10.4 For a character x>
, , : d)
V p-l ifX=id.
Remark 1.10.3 The identity character is the one that takes the value 1 at each
point in Z(/?)*.
The result is obvious for x = id. If x ¥" id, there is a y e Z(p)* such
l.Then
which implies the theorem. There is a dual to (1.10.6) given by the following
theorem: •
Theorem 1.10.5 For the sum over all characters we have
Proof. It is sufficient to observe that if x / 1, then there is a character x such

that x (x) 7^ 1. The theorem may now be proved as before. •
We now define the analog of the beta function.
an
Definition 1.10.6 For two multiplicative characters x d r\ the Jacobi sum is
defined by
Y, (1.10.8)
x+y=l
The following theorem gives some elementary properties of the Jacobi sum. We
denote the trivial or identity character by e. The reader should notice that the last
result is the analog of the formula B{x, y) = r(x)T(y)/[r(x + y)].
an
Theorem 1.10.7 For nontrivial characters x d ??, the following properties
hold:
J(e,X) = 0. (1.10.9)
J(e,e) = p-2. (1.10.10)
1
(l.io.ii)
then J ( x , ly) = * ( x ) * ( > y ) . (1.10.12)
g(x*i)
Remark 1.10.4 From the definition of characters it is clear that the product of
two characters is itself a character and so the set of characters forms a group.
The additive characters form a cyclic group of order p and the multiplicative
characters a cyclic group of order p — 1. Also, x~l(x) — x(x~!) — VxOO and
since IxOOl = 1 it follows that x - 1 ( ^ )
Proof The first part of the theorem is a restatement of Theorem 1.10.3 and the
second part is obvious. To prove (1.10.11), begin with the definition
x JC/0,1
Now note that as x runs through 2 , . . . , / ? — 1, then x(l — x) runs through 1 , . . . ,

p — 2. The value y = p — \ = — 1 (mod p) is not assumed because x = y ( l + y ) ~ 1 .
Therefore,
J(x,x~l)= J]
by Theorem 1.10.4. This proves the third part. The proof of the fourth part is
very similar to Poisson's or Jacobi's proofs of the analogous formula for the beta
function. Here one multiplies two Gauss sums and by a change of variables arrives
at a product of a Jacobi sum and a Gauss sum. Thus, for x V / e,
liziylp
X(x)r](t-x)e27tit/p.
x+y=0
The first sum is J2X X(*)*7(—*) = 0 since XV ¥" id. The

second sum with x = st is
s)
"
This proves the fourth part of the theorem. •
We were able to evaluate F (s) in a nice form for positive integer values and half-
integer values of s. Evaluations of special cases of Gauss sums are also possible and
important, but in any case the magnitude of the Gauss sum can always be found.
Theorem 1.10.8 For nontrivial multiplicative and additive characters x and \jf,
Proof. By (1.10.5) it is enough to prove that |gi(x)| 2 = \g(x)\2 = P
\g(x)\2 =
Setx = ty. Then
tyÔ
E
Ô or 1
The first sum is p — 1 and the inner sum in the second term is — 1. Thus
and the result is proved. •
Corollary 1.10.9 If x, r\ and x *7 are nontrivial characters, then
\J{X,ri)\ = Jp. (1.10.13)
Proof. This follows from Theorems 1.10.7 and 1.10.8. •
As an interesting consequence we have:
Corollary 1.10.10 If p = An + 1 is a prime, then there exist integers a and b

such that p = a2 + b2.
Proof The group Z(/?)* is of order p — 1 = An, which is also isomorphic to

the group of multiplicative characters on Z(/?)*. Since the latter group is cyclic
there exists a character x of order 4 that takes the value ± 1 , =b". It follows that
J(X, X) = a + bi for integers a and b. Since x 2 7^ id, apply Corollary 1.10.9 to
obtain the desired result. •
Corollary 1.10.10 is a theorem of Fermat, though Euler was the first to publish a
proof. See Weil [1983, pp. 66-69]. Later we shall prove a more refined result that
gives the number of representations of a positive integer as a sum of two squares.
This will come from a formula that involves yet another analog of the beta integral.
We have seen that characters can be defined for cyclic groups. Since any abelian
group is a direct product of cyclic groups, it is not difficult to find all the characters
of an abelian group and their structure. The following observation may be sufficient
here:
If xi is a character of a abelian group G\, and xi °jGi> then we can define a
character x : G\ x G2 -» C* by / ( * , y) = X\MX2(y)-
We thus obtain n additive characters of Z(n) and </>(n) multiplicative characters
of Z(n)*. The Gauss and Jacobi sums for these more general characters can be
defined in the same way as before. Gauss [1808] found one derivation of the law
of quadratic reciprocity by evaluating the Gauss sum arising from the quadratic
character. (A character x 7^ id is a quadratic character when x 2 = id.) Details of
this connection are in Exercise 37 at the end of the chapter. One problem that arises
here, and which Gauss dealt with, is evaluating the sum G = Ylx=o el7Tlx2/N. As
in Theorem 1.10.8 one can show that G2 — ±N depending on whether N = I
(A) or 3 (4). The problem is to determine the appropriate square root for obtaining
G. According to Gauss, it took him four years to settle this question. Dirichlet's
evaluation of Ylx=o e2nix2/N by means of Fourier series is given in Exercise 32.
Jacob! and Eisenstein also considered the more general Jacobi sum
J(Xi,X2, • • •, Xt) = Yl Xi(h)X2(t2) • • • xiift). (1.10.14)
This is the analog of the general beta integral in Theorem 1.8.6. Eisenstein's result,
corresponding to the formula in Theorem 1.8.6, follows.
Theorem 1.10.11 If xi, X2> • • • > Xi are nontrivial characters and X\Xi" ' Xt ^
nontrivialy then
J(Xu X2, • • •, Xi) = ; :—• (1.10.15)

g(XiX2'-Xi)
The proof of this is similar to that of Theorem 1.10.7, and the reader shouldfillin
the details.
In Section 1.8 the volume of n-dimensional objects of the form a\xs^ + a^x^ +
hfljfcjCfc* < b was determined by means of the gamma function. In the same way,
for finite fields, the number of points satisfying a\x\x +a2Xs22 H \-akXskk = b can
be found in terms of Gauss sums. Gauss himself first found the number of points
on such (but simpler) hypersurfaces and used this to evaluate some specific Gauss
sums. Weil [1949] observed that it is easier to reverse the process and obtain the
number of points in terms of Gauss sums. For an account of this the reader should
see Weil [1974]. It may be mentioned that Weil's famous conjectures concerning
the zeta function of algebraic varieties over finite fields are contained in his 1949
paper. It also contains the references to Gauss's works. One may also consult
Ireland and Rosen [1991] for more on Jacobi and Gauss sums and for references
to the papers of Jacobi and Eisenstein.
The form of the Gauss sums also suggests that they are connected with Fourier
transforms. Let T denote the vector space of all complex valued functions on
Z(N), the integers modulo N. Let F be the Fourier transform on T defined by
N-l
(Ff)(n) = -y=Y, fWe2ninx/N- (1.10.16)

**N x=0
It can be shown that the trace of this Fourier transform with respect to the basis
{So, S\, ...,SN], where
fO, x^y,
&x(y)
11, x = y,
is the quadratic Gauss sum Ylx=o e2ltlx2/N. Schur [1921] gave another evaluation
of this sum from this fact. The details are given in Exercise 47. One first proves
that the fourth power of F is the identity so that the eigenvalues are ± 1 , ± / and
the essential problem is to find the multiplicity of these eigenvalues.
1.11 A Probabilistic Evaluation of the Beta Function 43
Discrete or finite Fourier analysis was not applied extensively before 1965 be-
cause of the difficulty of numerical computation. This changed when Cooley and
Tukey [1965] introduced an algorithm they called the Fast Fourier Transform
(FFT) to reduce the computation by several orders of magnitude. The reader may
wish to consult the paper of Auslander and Tolimieri [1979] for an introduction
to FFT, which emphasizes the connection with group theory. Some of the earlier
instances of an FFT algorithm are mentioned here. Computational aspects are also
interesting. See de Boor [1980] and Van Loan [1992, §1.3].
1.11 A Probabilistic Evaluation of the Beta Function

When a and ft are positive integers,
l
dx=
/ (a
It seems that it should be possible to arrive at this result by a combinatorial argu-

ment. But working with only a finite number of objects could not give an integral.
Here is a combinatorial-cum-probabilistic argument that evaluates the integral.
Choose points at random from the unit interval [0, 1]. Assume that the probability
that a point lies in a subinterval (a, b) is b — a. Fix an integer n and let P(xk < t)
denote the probability that, of n points chosen at random, exactly k of them have
values less than t. The probability density function for P(xk < t) is
P(xk<t + At)-P(xk<t)
pit) = lim .
^ v } A?Ô At
Now
P(xk <t + At)- P(xk <t)

= the probability that one point lies in (t, t + At),
k — 1 points less than t and n — k points greater than t + At,
+ the probability that two points lie in (t, t + At),
k — 2 points less than t and n — k points greater than t + At,
+ •••.
Since there are n points, the number of ways that one point is in (t, t + At), k — \
points are less than t, and n—k, are greater than t + Af is
n-\\(n-k\ (n-\
- 1 \n-k (
The probability of each such event is Attk~l(l —t — At)n~k, and since the events
are mutally exclusive, we get
P(xk<t + At)-P(xk<t)
l k n n kk
= n\ ? -\lL(l - t -At)
i " 1\ )r- At)~ ~ At + O((AtY).
Therefore,
Since
we obtain
f
Jo
Jo
p(t)dt = 1,
(k - l)!(/i - k)\ r(k)T(n - k
We made use of probability theory here to indicate its relationship with the beta
function. Though we do not use it elsewhere, probability theory can be used to
derive formulas involving some extensions of the beta function.
1.12 Thep-adic Gamma Function

In number theory there are completions of the rationals other than the reals that
are of great importance. These are the p-adic completions of the rationals. There
is an analog of the gamma function defined on the p-adic numbers that is useful.
The following is a very brief account of the /?-adic gamma function. The interested
reader should consult the references given later.
Suppose a is an integer and p a prime. Define ovdpa to be the highest power of
p that divides a. Let Q be the set of rational numbers. For x = a/b e Q, where a
and b are integers, define ord^jc = ordpa — ordpb. The /7-adic norm | \p on Q is
defined by
\X\B =
x=0.
1.12 The /7-adic Gamma Function 45
Thus in the /?-adic norm, pn gets small as n gets large. In contrast, for negative
values of n, pn becomes big. So it is reasonable to write numbers in powers of p.
An integer would have an expansion of the form
a0 + a\p -f a2p2 H h anpn,
where at e {0, 1, 2 , . . . , p — 1}. For rational numbers negative powers of p will

also be involved. The /?-adic norm is non-Archimedean, that is,
\\x + y\\p<max(\\x\\p,\\y\\p),
and so the triangle inequality holds. This gives a metric on Q.

We can then obtain a completion of Q with this metric in the same way as we
get the real numbers by taking the ordinary metric on Q. This involves taking
the Cauchy sequence of rationals. The p-adic completion is denoted by Qp. The
p-adic numbers can be represented by the series
b2pz + • • •.
pm pm~[ p
The subset of Qp which contains all numbers with nonnegative powers of p forms
a ring denoted by Zp. This is the ring of /?-adic integers. The positive integers
Z + form a dense subset of Zp. This makes sense because a member of Zp can be
represented as an infinite series
ao + a\p + a2p2 -\ , at e {0, 1 , . . . p — 1},
and the partial sums are integers that converge to the p-adic number. So, if there
is a function / defined on the positive integers and the values of / at two integers
that are p-adically close are close to each other, then / has a unique continuous
extension to Zp.
Define a function / on the positive integers n by the formula
k=\
pKk
It is not difficult to show that f(n + pml) = f(n)modpm, where n,m, and
n
I are positive integers. Now n + p l and n are p-adically close to each other
and the values of / at these points are also p-adically close. Consequently, /
has an extension to Zp. This extension gives the /?-adic gamma function due to
Morita [1975]. The /7-adic gamma function is defined by Tp(x) — —f(x — 1).
This function also has a functional relation and other useful properties. There is
a formula of Gross and Koblitz [1979] that gives the Gauss sum as a product of
values of the /?-adic gamma function.
A good treatment of the /?-adic numbers and functions is given in Koblitz
[1977]. An account of the /?-adic gamma function and the Gross-Koblitz formula
is available in Lang [1980], including a reference to a paper by "Boyarsky" [ 1980].
In fact, p-adic extensions of the beta function, and more generally, the Mellin
transform, are also available.
Exercises
1. Use the change of variables s = ut to show that
poo poo
r(jc)r(;y)= / / tx-lsy-le"(s+t)dtds
Jo Jo
is T(JC + y)B(x, v). (Poisson)
2. Let / = /0°° e~xldx. Observe that I2 = /0°° /0°° e~(x2+y2)dxdy. Evaluate this
double integral by converting to polar coordinates and show that / = ^/TT/2.
3. A proof of Wallis's formula is sketched below:
(a) Show that
Jo x2+a
dx
f n
(b) Take the derivative of bothJosides n times with respect to the parameter a
to conclude that
r
°° dx 1 • 3 • 5 • • • In - 1 n 1
/o (x2 + a)n+1 2'4-6-"2n 2
(c) Set x = y/ojn, a = 1, let n - • oo, and use Exercise 2 to obtain Wallis's
formula.
4. Evaluate / ^ ( l — t2)x~ldt in two different ways to prove the duplication
formula given in Theorem 1.5.1. To get another proof evaluate
rn/2
in two ways. Jo
I" Jo
sin 2 *" 1 20 dO
5. Suppose that / is twice differentiate. Show that f" > 0 is equivalent to

f(ax + fiy) < af{x) + Pf(y) for a and /3 nonnegative and a + fi = 1.
6. Convexity can be used to prove some important inequalities, for example,
Holder's inequality:
b f >b
rb >i \/p
>j
fgdx
-{Lm'dx\
/
where / and g are integrable functions and - + - = 1. We sketch a proof
here.
Exercises 47
x
(a) Note that e is a convex function; use this and the result of Exercise 5 to
show that if u and v are nonegative real numbers then
vq UP
1 . uv <
p q
Equality holds if and only if up = vq.
(b) Deduce Holder's inequality from (a).
It might be appropriate to call this the Rogers-Holder inequality since Rogers
[1888] had the result before Holder [1889]. Other important results of
L. J. Rogers are discussed later in the book.
7. Here is another proof of the functional relation
Write
= / dt
and perform an integration by parts to show that
x +y
8. Show that
B(x, y) _ f°° tx'ldt
cy " Jo {c + ty+y'
Take the derivative with respect to c and derive the functional equation
y
Give a similar argument using
ft x ~\c-t) y dt.
9. Write Gauss's formula as
Show that the right side satisfies all the conditions of the Bohr-Mollerup
theorem. This proves the formula.
10. Give a proof of Gauss's formula by using the definition of F(JC).
11. Prove Gauss's formula by the method given in the remark after Theorem 1.5.2.
12. It is clear from Y{x + 1) = xY(x) that f*+l logV (t)dt = x logx - x + C.
Show that C = \ Iog27r. Stirling's formula will work, but there is a more
elegant argument using Gauss's multiplication formula first.
13. There is another beta integral due to Cauchy defined by

dt TT22-*
C(x, y) =
(a) To prove this, show:
(i) Integration by parts gives C(x,y + 1) — -C(x + l,y).
(ii) Write
This together with (i) gives
C(x,y) =
x+ y-1
(iii) Iteration gives
_ x 22n(x)n(y)n
(x + y - \)ln
dt
Set t -> r/ V^ in the second integral and let n -* oo.

(b) The substitution r = tan ^ leads to an important integral. Find it.
14. Use the method for obtaining Stirling's formula to show that
1 1 1 n 1 / 1
/T + /? + + A 72V^
v
+ 2C +/ ^ +\ O w3/2 / '
where
C = -(1 + y/2) (1 - -^ + ^ - 4 T + '

V V2 V3 V4
Sum
n n +
Cn = ^ f o — ck-\) With Cn = ^ — -
and use some algebra to change cn — cn-\ to an expression that goes to zero
like n~3/2 to show that
See Ramanujan [1927, papers 9 and 13] for further results of this type.
Exercises 49
15. Here is an outline of a real variable proof of (1.2.1). Let
N
1
g(x) = lim \ ^
tan;!* N>OC *-^ n+ x
—N
(a) g'(x) = -7T2/ sin2 nx + ^ \/(n + x)2.
(b) g'{x) is continuous for 0 < x < 1 if g'(0) = g'(l) = 0.

(c) $'(*/2) + g'((* + l)/2) = 4g'(*).
(d) Let M = maxo<x<i |g'(jc)|. Then M < M/2 so M = 0.
oo
(e) g(x/2) - g((x + l)/2) = 27r/(sinjrx) - 2 ^ (-!)"/(« + x).
—oo
(f) g(x + 1) = g(x).

(g) g(x) = constant.
(h) / /*-7(l + r)dr = 2^(-l)7(«

OO
= 5^ (-i)7(n + *).
-00
so (1.2.1) holds. This proof, due to Herglotz, was published by Caratheodory

[1954, pp. 269-270]. Bochner's [1979] review of the collected works of
Herglotz also includes this proof.
16. The following is Dedekind's [1853] proof of r(x)T(l - x) = n / sin7rjc. Set
_ r ° tx~l
~ Jo l+t
(a) Show that
f°° tx~x
/ ——rdt = 4>{x)
Jo st + l
and
TOO ,*-l
/ dt =:<t>{x)SX-\
Jo t+s
Jo t-\-s
(b) Deduce that
(c) Use the second formula in (a) to get

/•oo i / poo t
[Hx)f= 4 T ( / 7+ s
Jo s + 1 V Jo t
then change the order of integration to obtain
poo +x—x l
2 _ r°°t - \ogt
~ Jo t-\
(d) Deduce
y [°° ty-\ _ t-y
[(/)(x)]2dx = / t
—dt.
/ \-y-y Jo — A
(e) Integrate (b) with respect to s over (0, oo) and use (d) to derive
000 / [4>(t)]2dt = 2 [ l
\ lOgtdt = 2(j)Xx).
Ji-x Jo 1 +1
(f) Show that 0(JC) = 0(1 — x) implies (j)f{\) = O and
x px
/
[c/)(t)]2dt = 2 [c/)(t)]2dt.
-x Jl/2
(g) Deduce that
Jl/2
(h) Show that 0 satisfies the differential equation 00 r/ — (0') 2 = 0 4 -
(i) Solve the differential equation with initial condition 0 ( | ) = TT and
0 r (l) z= Otoget0(;c) = n csc7rjc.
17. Show that
-, Rex > 0,Rey > 0,« > 0.

Jo Vat +
18. Show that
19. Prove that for a > 0,

sinaut
and
/•OO
f°° cos ax 1 . ! stc(nb/2)
/ :—JJC = -na , 0 < ReZ? < 1.
Jo *b 2 T(b)
20. For A > 0, JC > 0 and — n/2 < a < TT/2, prove that
Jo
Exercises 51
and
/•OO
/ fX-ig-ktcosa s m ( ^ sina)dt = X~xT(x) sin ax.

Jo
21. Prove that n-s/2r(s/2)$(s) = 7r- ( 1 - 5 ) / 2 r((l - s)/2)S(l - s) as follows:
(a) Observe that
00
sin(2/2 + l);c
E
n—\
2n + l
= (—1) 7T/4 for run < x < (ra + l)7T,ra= 0, 1,
(b) Multiply the equation by xs~l(0 < s < 1) and integrate over (0, oo).
Show that the left side is F(s) sin(>7r/2)(l - 2" 5 " 1 )f (s + 1) and that
the right represents an analytic function for Res < 1 and is equal to
(c) Deduce the functional equation for the zeta function. (Hardy)
22. Let C be a contour that starts at infinity on the negative real axis, encircles
the origin once in the positive direction, and returns to negative infinity. Prove
that
This formula holds for all complex s.

(a) Note that the integral represents an analytic function of s.
(b) C may be taken to be a line from — oo to — <5, then a circle of radius 8 in
the positive direction, and finally a line from —8 to —oo. Show that
/ elt~sdt = 2i sinns / e~uu~sdu

Jc J8
where / is the integral on the circle |f| = 8.
This representation of the gamma function is due to Hankel; see Whittaker
and Watson [1940, p. 244].
23. Prove that
e-i7VSF(l-s) x
where C starts at infinity on the positive real axis, encircles the origin once
in the positive direction, excluding the points ±2rmi, n > 1 an integer, and
returns to positive infinity.
Hint: First prove that
and then apply the ideas of the previous exercise. Note also that £ 0 , s) is now
defined as a meromorphic function by the contour integral with a simple pole
at ^ = 1.
24. Prove the functional equation
2F(1 — s) f ^ coslmnx ^
K
?(*, s) = (2ixy~ s;; ^sin(7r5/2)
^sin(7r5/2) V
V + C o s ( s 7 r / 2 ) V r yl s }.
*-^ m[~s +Cos(s7r/2) *-^ m~
K m=\ m=\ )
Hint: Let Cn denote the line along the positive real axis from oo to (In + l)n,
then a square with corners (2n + 1)TT (± 1 d= /), and then the line from (2n + 1)
to oo. Show that
r ts-i^-xt
-dt
— the sum of the residues at d= Imni, m = 1 , . . . , n,

where C is the curve in the previous exercise.
Note that the sum of the residues at ±2mni is
-2(2m7t)s-leins sin(2m7rx + its/2).
Now let n —> oo and show that \n —>• 0.
25. Show that the functional equation for f (JC, 5) easily implies
(a) the functional equation for f (5),
(b) Kummer's Fourier expansion for log T{x)/\/2jt.
The next five problems are taken from Artin [1964].
26. For 0 < x < 00, let 0(JC) be positive and continuously twice differentiable
satisfying (a) 0(JC + 1) = 0(JC), (b) 0(f)0(*±±) = d0(jc), where J is a
constant. Prove that 0 is a constant.
Hint:Letg(x) = £2 log 0(JC). Observe that g(jc + l) = g(jc)and|(g(f) +
^(^)) = gW.
27. Showthat0(jc) = F(X)F(1—JC) sin 7TJC satisfies the conditions of the previous
problem. Deduce Euler's reflection formula.
28. Prove that a twice continuously differentiable function / that is positive in
0 < x < 00 and satisfies (a) f(x +1) = xf(x) and (b) 22x~xf{x)f{x + \) =
^/nf(2x) is identical to F(JC).
29. It is enough to assume that / is continuously differentiable in the previous
problem. This is implied by the following: If g is continuously differentiable,
g(x + 1) = g(x), and £(f) + g(*±!) = g(jc), then g = 0.
Hint: Observe that
A:=0
The left side tends to J* g'{x)dx = g(l) - g(0) = 0 as n -> 00.
Exercises 53
30. Prove that the example g(x) = Y1T=\ h sin(2n7rx) shows that just continuity
is insufficient in the previous problem.
31. Suppose/ and g are differentiable functions such that fix+y) = f(x)f(y) —
g(x)g(y)mdg(x+y) = f (x)g (y)+g(x)f(y). Prove that /(*) = eaxcosbx
and g(x) = eax sin bx, unless f(x) = g(x) = 0.
32. Prove that ^=o e2nixl/N = l-^VN, where i = </=l.
(a) Set f(t) = J2x=oe27Ti(x+t)2/N>° < * < I. Note that /(0) = / ( l ) and
extend fit) as a periodic function to the whole real line.
(b) Note that f(t) = E-oo ane2**"', where an = f* fit)e-2*intdt.
Conclude that /(0) = E t " o e27Tix2/N = E-oo *»•
(c) Show that an = e~2«iNn2/4 ]%n2/2) e
(d) Show that
( r
"odd
(e) Use Exercise 19 to evaulate the integral. Another way is to take N = I in

id). (Dirichlet)
33. If p is an odd prime, then there is exactly one character x2 that maps Zip)*
onto {±1}. Recall that Zip)* is the integers modulo p without 0. Prove that
X2ia) = 1 if and only if x2 = a mod/7 is solvable, that is, a is a square in
Zip)*. Usually one writes x2 (a) = (~)»which is called the Legendre symbol.
34. Prove that if a is a positive integer prime to p, then ap~1/2 = ( £ ) (mod p).
Here p is an odd prime. (Use the fact that Zip)* is a cyclic group.)
35. For pan odd prime, use the previous problem to prove that (—) = (— l)^" 1 )/ 2
and ip = (-1)^ 22- 1} / 8 . (Use 2^ 2 = ieni/4 +
(mod/?). Consider the two cases /? = ±1 (mod8) and p = ±3 (mod8)
separately.)
36. Prove the law of quadratic reciprocity: For odd primes p and q, ( - ) ( - ) =
(a) For S = J2x=l(pe2nix/p^ s h o w t h a t s2

= (-^P- ( T h e P r o o f i s similar
to that of Theorem 1.10.8.)
(b) Use (a) and Exercise 34 to prove that Sq~l = ( - l ) V ^ (£) (modg).
(c) Show that ^ = E ^ i ( f V 2 7 r ^ / / ? = ( f ) ^ (mod^).
(d) Deduce the reciprocity theorem from (b) and (c). (Gauss)
37. For integers a and N with N > 0, define G(a, A0 = E^To ^27r/flx2/iV.
(a) For p prime, show that G(l, /?) = E ? = i ( f )^ I > J : / p -
(b) For p prime show that G(<a, /?) = (|)G(1, p).
(c) Prove that G(q, p)G(p, q) = G(l, pq) when p and q are odd primes.
(d) Now use the result of Exercise 32 to deduce the reciprocity law. (Gauss)
For a discussion of Exercises 32-37 and for references, see Scharlau and
Opolka [1985, Chapters 6 and 8].
38. Prove Theorems 1.8.5 and 1.8.6.
39. Prove Theorem 1.9.4.
40. Weierstrass's approximation theorem: Suppose / is a continuous function
on a closed and bounded interval, which we can choose to be [0, 1] without
any loss of generality. The following exercise shows that / can be uniformly
approximated by polynomials on [0,1].
(a) Show that it is enough to prove the result for / ( 0 ) = / ( I ) = 0. Now
extend / continuously to the whole real line by taking / = 0 on x < 0
and x > 1.
(b) Observe that
is a polynomial such that
Qn(f)dt = l.
Show that Pn(x) = / ^ f(x + t)Qn(t)dt is a polynomial in x for x e

[0, 1].
(c) Use Stirling's formula to show that for 8 > 0and<5 < \t\ < 1, Qn(t) —• 0
uniformly as n - • oo.
(d) Note that for 0 < x < 1, Pn(x) - f(x) = f^[f(x + t) - f(x)]Qn(t)dt.
To show that Pn(x) —• f(x) uniformly on [0,1], break up the integral
into three parts, J~^ + J^_s + / 5 \ and use (c).
41. Prove Plana's formula (see Whittaker and Watson [1940, p. 145] for references
to Plana): For positive integers m and n
+ / (/)(x)dx - i
<p(n + iy) - 0(m + iy) - 0(n - iy) + 0(m - iy)

x Jy
where 0(x + /y) is a bounded analytic function in m < x < n.

Hint:
(a) Consider the integral §c(j){z)/(e~27liz — \)dz where C is a suitable in-
dented rectangle with vertices k,k + \,k + 1 + L/, and k + Li. Then let
L —> oo.
(b) Now replace / with —/ in the contour C and repeat the process in (a).
(c) Add the results in (a) and (b) and sum over k.
Exercises 55
42. (i) In Plana's formula let m = 0, n —>> oo, and suppose that 4>(n) —> 0,
(p(n ± /y) —• 0, to get
d>(iy) - (b(-i
Jo
(ii) Deduce Hermite's formula (for reference, see Whittaker and Watson,
[1940, p. 269])
x~s xx~s f00 (x2 + t2)~s/2 sin(s arctan^/jc)
f (*, j ) = — - + + 2 / —— dt.
^ -
1 1 f™ Axt dt
(iii) Conclude that f (JC, 2) = ^-^
2JC2 H x h Jo/ (x2 + t2)2(e2*t -1)'
Jo
43. (a) For x//(x) = T\x)/ T(JC), note that f'(x) = $(x, 2).
(b) Deduce that
1 f00 ltdt
(x2 + t2)(e27Tt - 1)'
(Use part (iii) of the previous exercise.)
(c) Deduce Binet's second formula
in roc) = ([xx - I )) hue

in, --J Cx++ -ln(27r) + 2 /
1
V 2J 2 Jo
where x is complex and Re x > 0.
(d) Use Hermite's formula in the previous problem to obtain Lerch's formula
(1.3.7) for ( £ ? ( * , s)) 5 = 0 .
44. Prove the following properties of Bernoulli polynomials:
U ~p 1
n=M
(c) Bn(x) = ^
k=0
(d) Bn{\ — x) = (— \)nBn(x).
(e) Bn(lx) =
45. Prove that

\dj £>2q — 1 v-^ L-^J/ — ^ v
and
00
B2q(x-[x]) = 2(-ir
^ (Inn)2*
(b) Deduce
and
n=0
46. Prove that
B2n =
P
(p-l)\2n
where G^n is some integer and p is a prime such that p — 1 divides In.
(Clausen-von Staudt)
Hint: Define Y^=o ft*" = Y^T=o fa" ( m o d k)if ^ divides an - Z?n for all
n > 0. Show that
/z3 z 5 z7 \
(a) (ez - I) 3 = 2f - + - + - + • • • J (mod4).
(b) For prime /?,
(c) For composite m > 4

(^ _ i)^-i = o (modm).
Deduce the result on Bernoulli numbers. (See Polya and Szego [1972, Vol. II,
p. 339].
47. Let C(Z(n)), where Z(n) is the integers modulo n, be the set of all complex
functions on Z(w), where n is an odd positive integer. Define F: C(Z(n)) ->
C(Z(/i))by
n-l
= — V f{k)elnikx'n for x G Z(/i).
(a) Show that Trace F = ^ ^ ^ I Q e2nik^n.

Hint: Use the functions 8x,x e Z(n), where <5x(y) = 0, x ^ y, and
<$*(.*) = 1, as a basis for Z(n).
Exercises 57
(b) Prove that (F2f)(x) = f(-x). Conclude that F4 = id and hence that
± 1 , db/ are the eigenvalues of F. Let m\, m2, m$, ra4 be the multiplicities
of 1, /, — 1, and —/ respectively. Thus m\ + m2 + m^ + ra4 = 1.
(c) Show that Trace F2 = 1 and conclude that m\ — rri2 + mi — m4 — 1.
(d) Show that \jj^YXX e2lzik2'n\2 = 1. Use (a) to get {mx - mi)1 + (m3 -
(e) Prove that

(1
det F _ .( {{ ~")/2, " = K4),
~ I (m2 - m 4 ) / ( 1 - w ) / 2 , >z = 3(4),
and also
detF = det ( ^-e27tixy/n\ = Ki(l~n)/2,

V J 0<x,y<n-l
where K is a positive number.
(f) Show that m\ = a + 1, andra2 = m^ = m4 = a when n = 4a + 1 and
m\ = m2 = m^ = a andra 4 = a — 1 when n = 4a — I.
(g) Obtain the value of -^ X^=o e27n^ ^n for w odd. (Schur)
Let m be a positive integer and let x be a character on the group Z(m)*.
The function x can then be defined on all the integers by setting x (k) = 0
when gcd(A:, m) > 1. Clearly x has period m. We call x primitive if it
does not have a smaller period. Also, x is even if x (—1) = 1 and odd if
x(-D = - i .
Also, define
m-\
gk(x) = Y^x(n)e2lllknlm and gi(x) = g(x)-
n=0
48. For a e Z(p), let N(xn = a) denote the number of solutions of the equation
xn = a. If n | p — 1, then prove that
Ar(jc" = a) = 1 4
where the sum is over all nontrivial characters of order dividing n.

Let a be a nonzero integer. Consider the elliptic curve E defined by
XQX\ — x\ — axl = 0, which in affine coordinates is y2 = x3 + a. Suppose
p / 2 or 3 is a prime that does not divide a. Then y2 = x3 + a is an elliptic
curve over Z(p) with a point at infinity. If Np denotes the number of Z(p)
points on the curve, then Np = 1 + N(y2 = x3 + a).
(a) Show that if p = 2 (mod 3), then Np = p + 1.
(b) Let p = 1 (mod 3) and let X3 and xi denote the cubic and quadratic
characters of ZQ?)*. Note that N(y2 = x 3 + a) = Y,u+v=a N(y2 = w)
N(x3 = -v). Deduce that

Np = p + 1 + X2X3(a)J(X2, X3> + X2X3O3) J(X2, X3).
(c) Show that if Np = p + 1 - ap then \ap \ < /

49. By the method used in the previous problem, show that
\N(X3 +y3 = 1) - P + 2\ <2jp.
For Exercises 48 and 49, see Ireland and Rosen [1991, Chapters 8 and 18].
50. Prove that if x is primitive, then
when
( } = f XWSW 8cd(*> m ) !'
8kKX)
\0 when gcd(ik,m) > 1.
Define the Dirichlet L-function by
The series converges for Re s > 0, when x is a nontrivial character, that is,
X (n) T^ 1 for at least one w e Z(m)*.
51. (a) Prove that when x is nontrivial
^ m—1
L(x,D = —
(b) Show that if x is primitive
V xW logg sin— in— +

m
++
m *—J V m m )
(c) Prove that when x is even, ^ x (^)^ = 0> a n d also, when x is odd,
£x(*)log
(d) Prove that
{ ^^ £
^ E m ,
( ) X (k) log sin ^ ,
when x is odd.
when x is even,
52. Prove that

(a) 1 — | + i — ^H = 5- (Madhava-Leibniz)
(b) 1 + 1 - 1 - 1 + 1 + ^ - . . . = ^ (Newton)
(c) 1 - 1 + 1 - 1 + 1 - 1 + . . . = ^ (Euler)
Exercises 59
(d) 1 + 1 - 1 + 1 - 1 - 1 + 1 + . . . = ^ . (Euler)
(e) i _ I _ I +I+I _ I _ i +I+± _ . . . = 2 j 1^5
The series for n/4, usually called Leibniz's formula, was known to Madhava
in the fourteenth century. See Roy [1990]. Newton [1960, p. 156] produced
his series in response to Leibniz's formula by evaluating the integral
l+x2
0 1+*4
Adx
I '
in two different ways. Series (c) and (d) are attributed to Euler by Scharlau
and Opolka [1985, pp. 30 and 83].
Define the generalized Bernoulli numbers by the formula
00 „
V^Xfa Xn
z—4 emx __ Y
a=\ n=0
53. (a) Prove the following functional equation for L(x, s), X primitive:
, 8(x)f2nY L(x,l-s)
)
~ 2i* \m) r ( S ) c o s ^ '
where 8 = 0 or 1 according as x is even or odd.
Hint: Consider the integral
I
Jc emi - 1
where C is as in problems 23 and 24. Follow the procedure given in those
-dt,
problems.
(b) For any integer n > 1, show that
L(XA-n)
= - ^ ^ .
n
(c) For n > 1 and n = 8 (mod 2) (8 as defined in (a)), prove that
54. Let P be any point between 0 and 1. Show that
r
The notation implies that the integration is over a contour that starts at P,
encircles the point 1 in the positive (counterclockwise) direction, returns to
P, then encircles the origin in the positive direction, and returns to P. The
1—, 0— indicates that now the path of integration is in the clockwise direction,
first around 1 and then 0. See Whittaker and Watson [1940, pp. 256-257].
55. Let G(z) = log F(z). Show that
(a) If x > 1/2, then Re G"(x + iy) > 0 for all real y.
(b) If x < 1/2, then Re G"(x + iy) < 0 for all sufficiently large y.
(c) If 1/2 < a < fe, then
is an increasing function of y on (—oo, oo).

(d) The conclusion in (c) also holds if 0 < a < 1/2 and b > I — a.
(Ahern and Rudin)
56. Show that
This problem was given without the value by Amend [1996]. FOXTROT
© 1996 Bill Amend. Reprinted with permission of Universal Press Syndicate.
All rights reserved.

The Gamma and Beta Functions

Uploaded by

Copyright:

Available Formats

The Gamma and Beta Functions

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Gamma and Beta Functions

Uploaded by

Copyright:

Available Formats

Cambridge Books Online

George E. Andrews, Richard Askey, Ranjan Roy

Book DOI: http://dx.doi.org/10.1017/CBO9781107325937

Online ISBN: 9781107325937

Hardback ISBN: 9780521623216

Paperback ISBN: 9780521789882

1 - The Gamma and Beta Functions pp. 1-60

Chapter DOI: http://dx.doi.org/10.1017/CBO9781107325937.002

Cambridge University Press

1.1 The Gamma and Beta Integrals and Functions

(fl) B =a(a + l)---(a + i i - l ) forn > 0, (a) 0 = 1, (1.1.2)

and a is any real or complex number. Rewrite (1.1.1) as

Therefore, the infinite product

I I ( J C ) = lim ' (1.1.4)

Definition 1.1.1 For all complex numbers x ^= 0, — 1, — 2 , . . . , the gamma func-

F(JC + 1) =xT(x). (1.1.6)

The infinite product in (1.1.9) exists because

B(x,y)= I tx~\\-t)y~ldt. (1.1.12)

B{x, y) = ^-^B(x, y + 1). (1.1.14)

B(x,y + 1)= \-tx(l-ty] +- / tx(l-ty~ldt

Set y = 1 in (1.1.12) and (1.1.17) to get

I = f f-Ut = B(x, 1) = -P^- (

An important consequence of the proof is the following corollary:

Corollary 1.1.5 For Re x > 0

r(jc) = / tx-xe-fdt. (1.1.18)

T / 2 sin 2 - 1 0 cos2^"1 Odd = r ( * ) r ° ° . (1.1.21)

/ e~x2dx = 2 e~x2dx = / r ^ V d f = T(l/2) = Jn. (1.1.23)

The special case a = — l,b = 1 is worth noting as it is often used:

oo poo /»oo ppoo y l s

1.2 The Euler Reflection Formula

Theorem 1.2.1 Euler's reflection formula:

when zx~l has its principal value. Thus

it sec 7i x = lim (-1)* (1.2.8)

Formula (1.2.4) follows from the product formula

proved in the previous section and from Theorem 1.2.1 intheformr(jc)r(l—x) =

and (1.2.5) gives the expansion

Corollary 1.2.6 Log F(x) is a convex function of x for x > 0.

Proof The right side of (1.2.14) is obviously positive. •

Proof The first formula is the logarithmic derivative of

T(x + n) = (x + n - 1)(% + n - 2) • • • xT{x).

where w = e2ni^k is a primitive &th root of unity. This is a consequence of

= lim V ( — - — ) t p + n q = : lim s(t)

s(t) = -tp~q log(l - tq) + Yl^^ lo (1

= -tp~q log ]j^- - (tp~q - 1) log(l - 0

Replace p by q — p and add the two expressions to obtain

\Kp/q) - \/f (I - (p/q)) = -n cotnp/q. (1.2.18)

Add this identity to (1.2.17) to get

, f P\ n np ^ 1^-y lixnp ( 2nn\

1.3 The Hurwitz and Riemann Zeta Functions

= lim ( V - - l o g n ) =y. (1.3.2)

Theorem 1.3.1 For all complex s,

Proof From the functional equation and the fact that

Now (1.3.2) implies that (s - l)f(s) = 1 + y (s - 1) + A(s - I) 2 H .So take

?(*, s) - (Us) ~ sx$(s + D) = x~ + ]TV 5 [(1 + x/ny5 - (1 - sx/n)].

1.4 Stirling's Asymptotic Formula

T{x) ~ \/7jzxx~xl1e~x as Rex —> oo.

Proof. Denote the right side of the equation

cn = (n + x) \og(n +x)-(n+x) + dn.

Substitute this in the previous equation to obtain

log |r(a + ib)\ = ( a - ^ log || - | | | + i log2^ + O (±- ).

' 1 — ^3 • ^4 I * I in • Jv\ I <J / v^3 fit \/*

r(ak)T(aM + 1) («* +a+i)(n?=i !>,-))I"(a