Annals of Mathematics

The Gamma Function in the Integral Calculus

Author(s): T. H. Gronwall
Source: Annals of Mathematics, Second Series, Vol. 20, No. 2 (Dec., 1918), pp. 35-124
Published by: Annals of Mathematics
Stable URL: http://www.jstor.org/stable/1967180 .
Accessed: 24/05/2014 14:24

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].

Annals of Mathematics is collaborating with JSTOR to digitize, preserve and extend access to Annals of
Mathematics.

http://www.jstor.org

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE GAMMA FUNCTION IN THE INTEGRAL CALCULUS.
By T. H. GRONWALL.

TABLE OF CONTENTS.
PAGEE.
Introduction................................................................. 35
CHAPTERI: Definiteintegrals.
? 1. Uniformcontinuity. ........................................................... 36
? 2. Upper and lower integrals of a function...................................... 37
? 3. The definiteintegral. ......................................................... 39
? 4. Double and repeated integrals......... ....................................... 46
? 5. Infiniteintegrals. ........................................................... 50
? 6. Integrals of complex functionsof a real variable ............................... 58
? 7. Fourier series. ................................................................ 60
? 8. Bernoulli functionsand numbers ......... ..................................... 64
? 9. Euler's summation formula........... ........................................ 70

CHAPTER II: Definiteintegralsfor the Gamma functionderivedfromits definitionas an infinite

product.
? 10. Expressionof the Gamma functionby Euler's integral of the second kind. Integrals
forP(s) and Q(s) ....................................................... 75
? 11. The hypergeometric seriesF(a, ,3, y; s) and its Value fors = 1. Euler's integralof the
firstkind .............................................................. 78
? 12. Applicationof Euler's summationformulato log r(s + a) and 4b(s+ a) ........... 83
? 13. Integralsforlog r(s) and V&(s)of Cauchy and Gauss .............................. 89
? 14. Integralsof Raabe and Binet ............ .................................... . 92
? 15. The generalized Schaar integral .......... ..................................... 95
? 16. Sonin's formof the remainderterm in Stirling's formula...... .................. 99
? 17. Kummer's and Lerch's trigonometricseries..................................... 101

CHAPTER III: The Gammafunctiondefinedas a definiteintegral.

? 18. The Euler integralsof the firstand second kind and their elementaryproperties.... 105
? 19. Integralsfor sb(s). The infiniteproduct for r(s) ................................ 110
? 20. Integralsforlog r(s). Raabe's integral........................................ 112
? 21. Binet's integraland the asymptoticexpansion of log r(s + a) and 46(s+ a) .115
? 22. The integralsof Schaar and Landsberg.117
? 23. Kummer's trigonometricseries.121

Introduction. The obj ect of this paper is to give an exposition, as ele-

mentary as possible, of some of those aspects of the theory of the Gamma
functionwhich are not dealt with in Jensen's "An elementary exposition
of the theory of the Gamma function."' Chapter I presents briefly
1 Authorizedtranslation,with additional notes, by T. H. Gronwall. These Annals, ser.
2,
vol. 17 (1916), pp. 124-166. Referencesto paragraphs and footnotesin this paper will be given
thus: J ? 2 and J2.
35

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
36 T. H. GRONWALL.

those definitionsand theoremson definiteintegralswhich are indispensable

in the following. In chapter II, the classical applications of the integral
calculus to the Gamma function are set forth in a form which adheres
quite closely to the point of view of Jensen's paper, while in chapter III
the same body of theorems is derived from the definitionof F(s) as a
definiteintegral.*
While the paper contains little that is new in subject matter, a con-
siderable number of proofshave been remodeled, or replaced by new ones.

CHAPTER I.

DefiniteIntegrals.
1. Uniformcontinuity. Let us consider a function f(x1, x2, **, xn)
defined for all values of xi, x2, *.., Xn belonging to a pointset P which
is bounded (i. e., there exists an A such that I xi I < A, I X21 < A, ** *,
I Kn I < A for all points of P). The functionis continuouson the pointset
P when, for any point xi, x2, ***, x, of P and any E (arbitrarily small),
there exists a 8 = 8(E, x1, x2, *.*, Xn) such that for every point xi', X2',
-**, xn/ of P where

(1) X1I - XI I < 1 X2' -X2 I <8, * X** x/n Xn I <8

we have
(2) f (xI', X2', * Xn) -f (xI, X2, * Xn) < E.

The function is uniformlycontinuous on the pointsetP when, for any C,

a 6 - (E) may be chosen independentlyof xi, x2, ..., Xn such that the
inequality(2) holds forany two pointsxi, x2, *..*, Xnand xl', x2', *..., Xn
of P satisfying(1). The followingtheorem is of fundamental importance
for the definitionof an integral:
THEOREM I. When f(xi, x2, *, Xn) is continuouson a CLOSED2 point-
set P, thisfunctionis also uniformlycontinuouson P.
Suppose the theoremto be false; then there exist an E > 0, an infinite
sequence of pairs of points on P: x1,x2I, *, x., and xi', X2v', ,
(v = 1, 2, 3, v)
* such that
(3) Xilf XI, <6v ln-Xn,
Xn,, I< 5, 5V - 0asv co,
and for which
(4) | f(xIl, X2', **,x') - f(x1v,x2v, **, v) 1E.
Let a limitingpoint of the sequence xi,, x2,, *, xnvbe x10,x20, * *, xnO;
this point belongs to P, since P is closed, and f(xi, x2, *, Xn) being
* See the note at the end ofthis paper.
2 1. e., a pointsetcontainingall its limitingpoints.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 37

continuous on P, there exists a So such that for i - xI1 I < So

I x2 -x20 I < 30, ** I Xn -xnoI < S0 we have If(x1, x2, * X*,xn)-f(xio,
x20, . **, xno) I < E/2. We may now choose v so large that 3, < So/2and
xI - x1i0 I < So/2,..., I Xnv - xnI < So/2; hence, by (3), I/ - x1ioI
c | x~v- x10I + I xiv - xivI < So/2+ *,< o,) IXn' xno I < so. We
have thereforeIf(xi,, x2,, *. *n)x - f(xo, x20, * * , xnOI < E/2, f(xl,',
2v'y..., yXn') - f(X1, x20; , xnO) I < E/2.whence by addition If(xi,',
X2, ***, xnv') - f(xly, x2v,** , xv) I < E, which contradicts(4), and this
contradictionproves our theorem.
2. Upperand lower integralsof a function. We definethe interval (a, b),
where a < b, as the set of all points x such that a c x _ b. This pointset
is evidently closed; a and b are called the end points, and the points x
such that a < x < b the interiorpoints of the interval. The lengthof the
interval is b - a.
A set of values of a real variable t is said to be bounded below,when
there exists a finiteA such that t > A forevery t in the set, and bounded
above, when t < B for every t. It is a well-known theorem that every
set of i's which is bounded below has a lower bound m such that t _ m
for every t but that for any positive E there exists at least one t in the set
such that < m + E; similarly every set which is bounded above has an
upper bound M such that t _ M for every t but t > M - E for at least
one t in the set. It is evident that the lower and upper bounds of -
are -M and - m respectively.
In particular, let t be the set of values assumed by a functionf(x)
when x varies in the interval (a, b); when this set of {'s is bounded both
above and below, f(x) is said to be bounded in the interval (a, b), and m
and M are called the lower and upper bounds off(x) in this interval.
Now consider a functionf(x), bounded in the interval (a, b), and let
us divide this interval into a number of subintervals (x,-i, x,) (of length
X- xvI) by means of the points of division x0, xI, ***, xn, where
a= XO < XI < X2 < *.. < Xn1 < Xn = b. Denote by M, the upper,
and by m, the lower bound off(x) in the interval (xv,-, xv), and write
S - M1(x1 - Xo) + M2(X2 - XI) + * + Mn(Xn - Xn-1) = ,MJv
(vXv-l),

s = m1(x1 - Xo) + m2(x2 - X1) + * + mn(n-Xn- = 2mV(x, P-X,).

Evidently M - M m, m, and hence

(5) M(b -a) A S A s - m(b -a).
If we consider all possible values of S and s correspondingto all possible
modes of subdivision of (a, b), S and s are bounded above and below by
(5); hence there exist an upper bound j of s and a lower bound J of S.
We shall now prove

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
38 T. H. GRONWALL.

THEOREM II. Whenthenumberofsubintervals is increasedindefinitely

in sucha mannerthatthelengthofeach tendstowardzero,thesumsS and s
tendtowardthelimitsJ and j respectively.
Let us change the sum s into s' by introducinga new division point x'
between x,_1and xv. Denoting by m,' and m," the lower bounds of f(x)
in (x,-1, x') and (x', xv), we obviously have
-s = m' (x' - X'_.) + mV"(x, - x') - mV(x -X1)
= (m' - m") (x'- x ) + (m," - m) (x -X)
and since O - m '-ml _M-m O -_<m" -m, M-m, we find
0 s' -s c (M-m)(x,-x xi) < (M-I') if xv-x xi < S. More
generally, if s' is obtained from s by the introduction of k new division
points, we find by introducing them one at a time and successive appli-
cation of the above inequality that,
(6) 0 c s'-s < k(M-m)8,
if every subinterval corresponding to s is less than S. Since - J and
- S stand in the same relation to -f (x) as j and s to f(x), it is sufficient
to show that for any e > 0 there exists a a such that j - s > j - e for
any s in which the length of each subinterval is less than S. Since j is
the upper bound of all s, it is clear that j - s and that there exists an so
such that so > j - E/2. Let the number of subdivisions correspondingto
so be k, and consider any s in which the length of each subdivision is less
than

8 2k(M -m)
Finally let s' correspond to the subdivision having as division points all
those occurringin s and in so; then, by (6),
s > s' m)8 = s' - E/2,
- k(M -

and applying (6) to so and s', we find s' : so, so that

s > so - e/2 > (j - E/2) - E/2= j-E,
which proves our theorem.
The numbers J and j are called the upper and lower integrals of f(x)
between the limits a and b, and are denoted by
rb bh
J = fb(x) j = fi(x)dx.

Since S : s, it follows from theorem II that J j or

rb fb

(7) I f(x)dx I f(x)dx.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 39

Let a < c < b and considerthe sums S and s forwhichc is a division

point,it followsimmediatelyfromtheoremII that
(b

I fxd fbx
f(x)dx = ff(x)dx + ff(x)dx,
(8) b c b

f(x) dx = f(x)dx + f(x)dx,

and from(5) and (7) it is seen that

(9). m(b - a) - f f(x)dx c f f(x)dx M (b - a).

Furthermore,when k is a constant,the sums S and s forkf(x) equal k

timesthoseforf(x), whence
rb b rb rb

fkf(x)dx = k ff(x)dx, fkf(x)dx = kJ f(x)dx for k 0_

(10) b b b

J'kf(x)dx = k ff(x)dx, fkf(x)dx = k f(x)dx for k c 0.

Finally,letfi(x), f2(x) *. ** fn(x) all be boundedin (a, b), and let M, and
mY,Ml, and ml,, *., Mn, and Mn be the upper and lower bounds of
(fi(x) + f2(x) + *..+ f.(x)), fi(x), - fn(x) in (x,-1, x,). Evidently
X M14 + M2v + * + Mnv and m. ml, + m2v+ + *Mnv whence
2;MV(Xy - xV-1) _ Mi~(xv - x1-,) + *. + 2;Mnv(x, - x,-,)
and
;mv(X,- X,1) 2 ml,(xv - X-1) + + 2Mn,(X, -Xv-).

Passing to the limit,we obtain

rb

f(fi(x) + f2(X) + * + fn(x))dx

b rb- b

(11 .< fbf(x)dx+ f2(x)dx+ + fn(x)dx,

s(fi(x)+ + + fn(x))dx
fb
f2(x)

Jq
ff(xdxb
- fi(x)dx + + *** +dx.

3. The definiteintegral. When the function f(x) is bounded in the

interval (a, b) and the upper and lower integralsof f(x) between the
limitsa and b are equal, the functionis said to be integrable
in (a, b), and
the commonvalue of the upper and lowerintegralsis called the integral

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
40 T. H. GRONWALL.

off(x) betweenthe limitsa and b and is denotedby

rb

f (x)dx.

The necessaryand sufficient

conditionforintegrability
is obviouslythat
the difference
(whichis O 0 by (5))
S - s = 2(M - m3)(x, - xvl)

shall tend toward zero for any one particularmode of subdivisionof

(a, b) whenthe numberof subdivisionsis increasedindefinitely in such a
mannerthat the lengthof each approacheszero.
We shall now prove
THEOREM III. A bounded functionf(x) is integrable in the interval
(a, b) when it is possible to find a finite number of subintervals,the total
lengthof which is as small as we please, and such that everypoint of dis-
continuityoff(x) is interiorto one of thesesubintervals.
We may note that the hypothesisof this theoremis fulfilledwhen
f(x) has a finitenumberof pointsof discontinuityin (a, b).
To prove our theorem,it suffices,by the remarkabove, to exhibita
subdivisionsuch that the corresponding sums S and s differby less than
any assigned E. By hypothesis,we can finda finitenumberof intervals
A', of total lengthless than 2(ME-)' such that every point of dis-
continuityof f(x) is interiorto one of them. Denote by A" the finite
numberof intervalsobtainedby removingfrom(a, b) the interiorpoints
of all intervalsA', and considera subdivisionof (a, b) in whichthe end
pointsof all intervalsA' are among the divisionpoints. Then we may
write
S - s = m'('A-m)(x, - x_) + m(MAI-mX)(xX -X l),
where !' and I" referto the intervals belonging to A' and A" respectively.
For the former,we have, since M,- m M - m,
m'M- )(X- x"_1)
_ (M - m)2'(x. - x_1) < (M-7r) ( - m) 2

Regarding A", we observe that A" is a closed pointset on which f(x) is

continuous,and by theoremI we may assign a a so small that

If(x') - f(x") I < 2(b-a)

forany x', x" both belongingto A" and such that Ix' - x" I < &. Now

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 41

make every subdivision in A" so small that xv - x,1 < 3, and let x', x"
be values of x in the interval(xvi, x,) forwhichthe continuousfunction
f(x) takes its maximumand minimumvalues M, and m, respectively.
Then
M; M"< 2(b-a)

forevervsubintervalbelongingto A", and the total lengthof thesesub-

intervalsnot exceedingb - a, we have

E/ .)*(b -a) 2
Consequently
0 S -s <2+2= es

whichprovesour proposition.
THEOREM IV. Whenf(x) is boundedand integrable in (a, b), thesame
is trueof If(x) I
It is evidentthat If(x) I is bounded whenf(x) is bounded. Let Mv
and mv,M.' and m^' be the upper and lower bounds of f(x) and If(x)
in (xl, X.). When M, -mv A, we have M,' = M , mv'=m,; when
0 M, -m^, then M^' = -m M, = - ,M, and when M ? 0
mM' M,
Mv' equals the greaterof M, and - m,, whilem^' = 0, so that in all three
cases 0 _ Mv' - m,' M, - Mn. Since N(M,- m^)(x, - x,-) tends
toward zero, the same is true of 2(M^' -m,-)(x, - ), so that the
upperand lowerintegralsof If(x) I are equal.
From (9) it followsthat whenf(x) is integrablein (a, b)
fb

(12) m(b - a) c f(x)dx M(b - a).

More generally,when three bounded and integrablefunctionsf(x),

m(x) and M(x) satisfythe inequalities m(x) f(x) _ M(x) in (a, b),
we have
(b fb fb

(13) Jm(x)dx c f(x)dx c M(x)dx.

In fact,since m(x) - f(x) c M(x) in (x^_-,xv), the lowerbound of m(x)

in thisintervaldoes not exceed that off(x), and the upper bound off(x)
does not exceed that of M(x); hence the sum s relativeto m(x) does not
exceedthatrelativetof(x), and theS relativetof(x) does not exceedthat
relativeto M(x). Passing to the limit,we obtain (13).
In particular, theorem IV allows us to take m(x) = - f(x) I

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
42 T. H. GRONWALL.

M(x) = if(x) (, and (13) gives

(14) Ijbff(x)dx _ f (f(x) I dx.

When a < c < b, (8) becomesforan integrablef(x)

rb

,ff(x)dx = f c

f(x)dx +
rb

f(x)dx,

and the repeatedapplicationof thisformulagives

rb XI rX2

(15) x~d = f+(x)d f(x)dx

aX
(15)
J
X
b

f(x)dx + Jof(x)
n-l

+ ***+ dx
fora <xl <X2 < *.. < X1<b.

It is sometimesconvenientto considerthe case where the upper

limitin an integralis smallerthan the lowerone; in thiscase we introduce
the definition
ra b

(16) f(x) = - f(x)dx, a < b.

From (16) it followsthat (15) is valid whena, x1, x2, * * xn1, b arein any
orderof magnitude,providedf(x) is integrablein an intervalcontaining
them all. When b < a, the inequalities (12), (13), (14) must however
be reversed.
The inequalities (10) and (11) give when f(x), fi(x), , fn(x) are
integrable
rb obb

(17) kf(x)dx = k a(x)dx,

rb

(fb(x) + f2(x) + + fn(x))dx

(18) a b b b

= fJ
A (x)dx + J7f2(x)dx + + Jfn(x)dx
THEOREM V. When fi(x) and f2(x) are boundedand integrablein
(a, b), thesame is true oftheirproduct
f(x) = f1(x)f2(x).
Let M and m, M' and m', M" and m" be the upperand lowerbounds
of f(x), fi(x) and f2(x)in (a, b); furthermore,let M, and m,,M.' and mr',
MV"'and mr" be the corresponding boundsin (x,1, xv). Evidently I f(x) I
does not exceed the greatestof the quantitiesM'M", m'm", -Mimi/
- m'M", so thatf(x) is bounded in (a, b). To prove the integrability,
let us firstsuppose thatf1(x) and f2(x) (and consequentlyf(x)) are never

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 43

negative in (a, b). Then obviously M, c Mf'Mj'", m, m'mvt",and

AMi
Mv-mv - I'I"
_Mv'Mv'
- MnV - mv'mv"f
n
= f'(Mv'-mv) + mv'(Mv''-mv') _Mf'(Mv' - v')

+ Mf (Mv'' - mv");
hence
f
2:(MV -mv) (xv -XV_1) C M2 (MV' -mv') (Xv- Xv-l)

+ M3'(mv1" -
mvr")(xv-xv-).
fi(x) and f2(x) beingintegrable,both sums to the rightmay be made as
small as we please by taking the subdivisionssufficiently small, and
consequentlythesame is trueof thesum to theleft,i. e.,f(x) is integrable.
Second, iffi(x) or f2(x) assumes negativevalues in (a, b), the productof
the non-negativefunctionsfi(x) - m' and f2(x) - mi" is integrable,and
by (17) and (18), the same is the case for
fl(x)f2(x) = (fi(x) - m')(f2(X) - Mi") + mffi(x) + m'f2(X) -MM".

THEOREM VI. Whenf(x) is integrablein (a, b), and for any x betweena
and b we write
rx
F(x) = J'f(t)dt, x $ a; F(a) = 0,

thenF(x) is continuousin (a, b), and at any point wheref(x) is continuous,

we have
dF(x) = (x)
dx =~)
By (15) we have, whenboth x and x + h are in (a, b),
rx+h .x+ h
rx
F(x + h) - F(x) = (t)fdth - f(t)dt = (t)dt,

and if M and m are the upper and lower bounds off(t) in (x, x + h),
(12) gives
mih F(x + h) -F(x) Mh
(if h is negative,the inequalitiesmust be reversedby (16)), whichproves
the continuityof F(x). When f(x) is continuousat x, m and M tend
towardthe same limitf(x) as h -O 0, and consequently

f(x) = limF(x + h) = dF(x)

- F(x)
THEREV. hef) s h =o dx
THEOREM VII. Whenf(x) satisfiestheconditionsof theorem
III, and

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
44 T. H. GRONWALL.

thereexists a functionF1(x) continuous in (a, b) and the derivativeof which

equals f(x) at everypoint wheref(x) is continuous,then

Fi(x) - Fi(a) = f(t)dt

for any x in (a, b).
Let F(x) be the function defined in theorem VI; then Fj(x) and F(x)
have the same derivative f(x) in any subinterval of (a, b) in which f(x) is
continuous, and by the theorem of the mean in the differentialcalculus,
Fj(x) = F(x) + const. in the subinterval considered. Now let (x1, x2)
and (X3, X4) be two subintervals of continuity forf(x), separated by the
subinterval (X2, X3) in the interior of which f(x) becomes discontinuous.
Then we have Fi(x) = F(x) + C in (x1, x2) and Fi(x) = F(x) + C' in
(X3,X4), and consequently
C' - C = (F1(x3) - Fi(x2)) - (F(x3) - F(x2)).
By the conditions of theorem III, X3 - X2 may be made as small as we
please; since Fj(x) is continuous in (a, b) by hypothesis and F(x) by
theorem VI, the expression to the right may be made as small as we
please by taking X3 - X2 sufficientlysmall, and C' - C, being inde-
pendent of x2 and X3,thereforeequals zero. Consequently
Fi(x) = F(x) + C
throughout the interval (a, b), and since F(a) = 0, we have C = Fi(a)
and the theorem is proved.
THEOREM VIII. When f(x), fi(x) and f2(x) are bounded and integrable
in (a, b), then
Ib
(19) ff(x)dx= lim If(Q,)(x, -x"A)
rb
(20) fi(x)f2(x)dx = lim lf1((,')f2((V") (x, -x,-),
a woo~~~~~8_>
where(V,(v' and (v" are any points in (xvl, xv), and xv - xv, < 8 for all P.
When (, is a point of discontinuityof f(x), we may take as thevalue off(%)
any numberbetweenm and M.
In (19), the sum to the right lies between s and S, and consequently
tends toward the same limit f b
f(x)dx as these. In (20), we have
1f1( v')f2(v"') (xv -Xv-)

h f flt)sm to (tvh 2
(Xvr-Xv-1)
) t +
rfl(x) )
-f2(d as -(
802)) XVt
(Xvy- )

the firstsum to the righttends toward as 5 O- by theorem

f(f2(x)dx

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 45

V and (19), while the absolute value of the second sum does not exceed

2 fi(') I *i f2(v") I (xv -Xv)

-f2(v')
c 2I t) I WV" - M.,") (xv - X.,-,) < ( t M' Z -+ IM
m x
) (M"
- )
-mvi')(xv - xv-)
and f2(x) being integrable, the last sum approaches zero as a -> 0.
THEOREM IX. Integrationby parts. Let u(x) and v(x) be continuous
functionsof x such thatboundedfunctionsu'(x) and v'(x) exist in (a, b) and
are continuous and equal to the derivativesof u(x) and v(x) exceptpossibly
on a pointset which may be enclosed in the interior of a finite number of
subintervalsof arbitrarilysmall total length; then
b b
(21) a
u(x)v'(x)dx = u(b)v(b) - u(a)v(a) - v(x)u'(x)dx.
~~~~~~~~~~~~~~a
Except on the pointset specified,we have

uv' + vu' d(uv)

and the integral f rb

(uv' + vu')dx, which exists by theoremV and (18),
equals u(b)v(b) - u(a)v(a) by theorem VII, whence (21) follows by
applying (18).
THEOREM X. Integrationby substitution. Let f(x) be bounded and
integrablein (a, b), and sp(t) a functionsuch that sp(a) = a, p(f) = b, and
~p'(t)is continuousand different fromzero in (a, f); then
rb r?
(22) j'f(x)dx = ff(ep(t)) p'(t)dt.

Supposing that sp'(t) > 0 in (a, f), there exists a k > 0 such that sp'(t) - k
in (a, fi). Subdivide (a, fi) at the points a = to,ti, *.., t.-1, t, = f, and
write xv = sp(t,). Let M. and m. be the upper and lower bounds of
f(p(t)) in (t,1, tQ)or, which is the same, the upper and lower bounds of
f(x) in (x,,, xv). Then, by the mean value theorem,
Xv- x"- = sp'(tv') (tv -tv-1) > k (tv -tv-1) > O.

where t' is some value in (t.b, Q), so that

O < 2 (MV -MV) (tv -tv-1) < k; (MV -MV) (X,,- X,,),

and f(x) being integrable in (a, b), the sum to the right approaches zero
when all xv - xvi tend toward zero, which is evidently the case when all

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
46 T. H. GRONWALL.

t,- tVi tend toward zero, and the above inequalityshows that f(so(t))
is integrable in (a, fi). The same is true when So'(t) - k in (a, fi),
the above inequalitybeing then reversed. Since So'(t) is integrable, the
product f(Sp(t)) p'(t) is integrable by theorem V, and for t," in (t,-1, tv),
xv' = 0(t~"), we have
Y2f(x,')(xv,- xi~) o(tvl)(tv - t^_).
= YfAs0(t"'T~')

Passingto the limitand usingtheoremVIII, our propositionfollows.

4. Double and repeatedintegrals. Let f(x, y) be bounded in the rec-
tangle a c x _ b, c _ y _ d, and subdivide this rectangleinto the mn
rectangles x,-, c x _ x,,,-Y -,
y yc , where

a =xo < X < .. <xm-i < xm =b

and
C =Yo < Yl < ... < Yn-i < Yn d.
Denote by M and m the upperand lowerboundsoff(x, y) in the rectangle
_
a c x _ b, c c y d, and by MMV the corresponding
and mM,.V boundsin
the rectangle x,_ _ x _ x,,, y -1-cy _ yV,and consider the two sums
in n
S= Ei E
'"=1 V=1
MIV(xM -xIA-1) (y V-Y ))

m n

s = E E MAV(x A- xI,-)(yV YV-1

l= V= 1

Since evidentlyM : M mAv

: m, we have

(23) M(b -a) (d -c) :--S :--s :--m(b -a) (d-c),

so that forall possiblemodes of subdivisionof the rectanglea c x _ b,
c _ y c d, S and s are bounded above and below; hence there existan
upperboundj of s and a lowerbound J of S. In exactlythe same way
as we provedtheoremII, we find
THEOREM XI. When thenumberofsubrectanglesis increasedindefinitely
in such a manner thatthe sides x,. - x,,1 and yv,- yv-' of each tendtoward
zero, the sums S and s tend toward the limits J and j respectively. The
numbersJ and j are called the upperand lowerintegralsoff(x, y) overthe
rectanglea c x = b, c = y = d, and are denotedby
rbr j =
b rd
= fbJ ff(x, y)dxdy, f(x, y)dxdy.

Whenf(x,y) is boundedin therectangle,and the upperand lowerintegrals

off(x, y) overthe rectangleare equal, the functionis said to be integrable
in the rectangle,and the commonvalue of J and j is called the (double)

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE. THEORY OF THE GAMMA FUNCTION. 47

integral off(x, y) and denoted by

rb

fb ff(x y)dxdy.

The necessary and sufficientcondition for integrabilityis obviously that

the difference(which is O 0 by (23))

S -s = E E(M'"V -
mLV)(x,.-1x41)(yV - YV)

shall tend toward zero for some particular mode of subdivision of the
rectangle when the number of subrectangles is increased indefinitelyin
such a manner that the lengths of all their sides approach zero.
The followingtheorem is proved in exactly the same way as theorem
III:
THEOREM XII. A boundedfunction in therectangle
f(x, y) is integrable
a _ x b, c _ y _ d, whenit is possibleto find a finitenumberof sub-
thetotalarea of whichis as small as we please, and such that
rectangles,
everypointofdiscontinuity off(x, y) is interiorto oneofthesesubrectangles.
For any value of x in (a, b), f(x, y) is bounded when c _ y _ d, and
rd
consequently the upper and lower integrals f(x, y)dy and ff(x, y)dy
exist. Denote by
F(x) = J'f(x, y)dy
any number such that
rd rdr
f(x,
f y)dy f y)dy
f(x, J'f(x, y)dy,
d

so that, in particular, F(x) = J f(x, y)dy for every value of x for which
f(x, y) is integrable in respect to y between the limits c and d. Then the
integral (which we suppose for the moment to exist) F(x)dx is called
the repeatedintegralof f(x, y) in respect to y and x and is denoted by
rb r
Jbdxf f(x, y)dy.

Similarly, we may define the repeated integral in respect to x and y,

rd rbr

Jddyf f(x, y)dx = fG(y)dy,

wb
where G(y) = J (x, y)dx. We then have

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
48 T. H. GRONWALL.

THEOREM XIII. Whenf(x, y) is boundedand integrablein therectangle

a x< b, c y c d, thenbothrepeatedintegrals off(x, y) existand are
equal to thedoubleintegral:

f
ob rrd

dx ff(x, y)dy = f rb

dy ff(x, y)dx = f
rb rd

ff(x, y)dxdy.

Since mAVc f(x, y) c M_. for xA- 1 x c x,,, YI-1 c Y Y

y, it follows
from(9) that
mAV(yv -Yv-1) fYJf(x, y)dy _ f(x, y)dy _ MIAV(YP -YV-);
Yv_1 YV_-1

adding these inequalities for v = 1, 2, *.*, n and applying (8), we find

n d d n

v=l
ILV(Yv - YV-1) _ Jf(x, y)dy c jf(x, y)dy Z Mv(yv
V=1
-YV

for any x in (x,-,, x,.), and consequently the lower bound mA and the
rd
upper bound M,. of F(x) = f(x, y)dy for x.1 _ x _ xM,satisfythe
inequalities
n n

EmAV(yV -YV-1) M
mA - MM c ZM (ylv- V-1
v=l V=1

Multiplicationby x,, - ,x,_ and summationin respectto A now gives

m n m
E EmAV(xA - xI-1)(yV- yV-) _ Z mA,(x,,-x/1
MA1V=1l=

Z
,u=1
MIA.(x, ,-x1)
mn n

ZEE MV(xA -xI)(y V-A

But as m and n tend toward infinitywhile all xM xA1 and yv- M-

tend toward zero, the two outside membersof this inequalityapproach
the same limitff f(x, y)dxdy; consequentlythe two innermembers
must also approach this same limit,that is, fF(x)dx existsand equals
rb

the double integral. The same proofobviouslyapplies also to the other

repeatedintegral.
THEOREM XIV. When f(x, y) and f-( - are boundedfor x in (a, b)
and y in (c, d) and
bY t)
daf(X,
Ja
-dxdt

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 49
existsforany y in (c, d),

at
existsforany x in (a, b) and any y in (c, d), andfinally f f(x, c)dx exists,
then
d J'f(x, y)dx = af(x y)dx

foreveryvalueofy forwhichthelast integralis a continuous functionofy.

It may be noted that all the conditionsof the theoremare satisfied
whenf(x, y) and af(x, y)lay are continuousforx in (a, b) and y in (c, d).
For the proofwe observethat by theoremXIII

(24) (Y af
f(x,t) dxdt= b
doaf(x, t) dt y
dt Ub
af(X, t) dx
~ Ja .)c at dt dx at - d at

and by theoremVII

('8 Ofx, t) dt = f(x, y) - f(x, c);

hence, by theoremXIII, frb

(f(x, y) - f(x, c))dx exists,and fromthe

rb

existenceof J'f(x, c)dx and (18) we concludethat

J f(x, y)dx =
rb pb

J
rb
Jf(x, c)dx + (f(x, y) - f(x, c))dx.

Consequently(24) gives

fbf(x2 y)dx - fbf(x2 c)dx = dt f af dx,

and differentiating
in respectto y, theoremVI gives
ajb f(x( y)dx = J af(dyy)

for any value of y for which the integralto the rightis a continuous
functionof y.
The followingpropositionis proved in exactly the same way as
theoremIV:
THEOREM XV. Whenf(x, y) is boundedand integrable in therectangle
a x - b, c c y- d, thesame is trueof If(x, y) I .
Formulas (13) and (14) are also immediatelyextended to double

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
50 T. H. GRONWALL.

integrals:
rb dc
J
rb rdb rd

(25) m(x, p)dxdy ff(x, y)dxdy c J'M(x, y)dxdy

when the three bounded and integrable functions f(x, y), m(x, y) and
M(x, y) satisfythe inequalitiesm(x, y) c f(x, y) c M(x, y) in the rec-
tangleofintegration;and form(x,y) = -I f(x, y) I , M(x, y) = If(x, y) j,

(26) b fdX y)dxdy c f f If(x, y) Idxdy.

Also (17) and (18) extendimmediatelyto double integrals:

(27) ffkf(x,
rb rd

y)dxdy = k f ob Sd

f(x, y) dxdy,
rb rd

FbJd(fi(x y) +f2(x, y) + +fn(x, y))dxdy

~~
f f
sb d

+
(28) = fi(x,y)dxdy J'f2(x, y)dxdy +

+ fbXfn(fx y)dxdy.

In (25)-(28), any double integralmay evidentlybe replaced by either

of the repeatedintegrals. Finally,theoremVII extendsimmediatelyto
double and repeatedintegrals.
integrals.In the precedingparagraphs,we have assumed a
5. Infinite
and b finiteand f(x) boundedin (a, b). When one of these conditionsis
we may define,by a limitingprocess,whatis called an infinite
not fulfilled,
integral(the termis chosenin analogy to "infiniteseries").
Firstassumef(x) to be boundedand integrablein (a, b) forany b > a.
The integralbetweenthe limitsa and oo is thendefinedas

ff(x)dx = lim fb(x)dx,

wheneverthislimitexists. Similarlywe may define
rb rb r0 rb

ff(x)dx = limZJ f(x)dx, f(x)dx = lim' fb(x)dx.

_oo a-4-oo a -o a---oo a

Next let us assume thatf(x) is bounded and integrablein (a, b-

for an arbitrarily small positive e, but tends toward + oo or - oo (or
oscillates between these bounds) as x -) b. The integralbetween the
limitsa and b is thendefinedas
pb
fb-f
Jaf(x)dx = lim f(x)dx
a a
f~~~~C-O

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 51

wheneverthis limitexists. Further,whenf(x) ceases to be bounded as

x -> a, we define
fi(x)dx = lim f(x)dx,
a >Ob a+C

and whenc is a pointbetweena and b in the neighborhoodof whichf(x)

is not bounded

a-f
rb roof rb
ff(x)dx = lim f(x)dx + lim ff(x)dx.
a ~~~eo es*oce
Similarlywe may defineinfiniterepeated integrals; forinstance,when
f(x, y) is bounded and integrablefor a + e c x c b, c + et c y c I
wheree and e' are as small, b and d as large as we please, we have the
definitions

a
dx fa x, y)dy
ceo
= lm
a->Oe
J x [ijjAx,
ep
so ce y)dy]
b->oo d oo

fdyJ f(x, y)dx = lim dy[limib f(x, y)dxl

a e
ego ef e -O a+e
d-> (* b
b->oo

In the following,we shall consider only integrals of the type

f(x)dx wheref(x) may cease to be bounded at x = a, and the corre-
spondingcase forrepeatedintegrals. The theoremswe shall provehold,
however,in the othercases; the necessarymodifications of the proofsare
obvious.
A convenientmeansofestablishingthe existenceof an infiniteintegral
consistsin comparingit to anotherwithpositiveintegrand.
THEOREM XVI. Letf(x) beboundedand integrablewhena + E x b, c
wheree is as smalland b as largeas we please,and let If(x) l c M(x) in this
fb

interval. If M(x)dx exists,and is boundedwhen e -0 and b -> oo,

then f 0+
(x)dx exists,and
f

ff(x)dx q M(x)dx.

Since M(x) O 0, fiM(x)dx increases (or remains constant) when

fb

e
decreasesor b increases,and being bounded,this integralmust therefore
approach a definitelimit as e -> and b -> oo, that is, M(x)dx exists.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
52 T. H. GRONWALL.

Consequently
lim M(x)dx= 0, i'm M(x)dx = 0,
e-*O Sa+e b b
e'-t0 0o

and by (14) and (12)

rai e
ra~~+ rb rb
rJ f
Ja-~e
d
Ja+e
M(x)dx fJ(x)dx
blb
M(x)dx,
so that
limi
e->o0
J ae
f(x)dx = 0, lim ff(x)dx
b9 b-
= 0,

whichare the conditionsforthe existenceof ,ff(x)dx.

When If(x) I < K(x -a)-a, wherea < 1,fora < x < xi,
f
COROLLARY.

and f(x) I < Kx-O, where, > 1,forx > x2, then f(x)dx exists.
In fact, when e and e' are so small that a + E < a + E' < xi, we have
+elf ~ a+i-ef K
fAx)dx K (x - a)j-dx= (E0-'~-*,
Sza+e Sa+ee a
and forb' > b> x2,

x Kxdx K (.
Jf(x)dx =

When f f(x)dx exists,then,by the definitionof an infiniteintegral,

J'f(x)dx and f(x)dx also exist,and (15) becomes

(29) J'f(x)dx
00

a
= Jaf(x)dx +
Xx
f f(x)dx +*
r2

Il
+
rftl

xn-2
rX
f(x)dx + J'f(x) dx
xl
for a < xi < X2 < ... < Xn-1. The equation of definition (16) becomes
na noo
(30) J'f(x)dx = - f(x)dx,

and (17) and (18) also extendimmediatelyto the presentcase:

C* *00
(31) fkf(x)dx = k ff(x)dx,

f(fi(x) + f2(x) + + f.(x))dx

,f f
oo *o
(32) a

= f1(x)dx+ f2(x)dx+ + fn(x)dx.

By a passage to the limitin theoremVII, we find

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 53

THEOREM XVII. Whenf(x) satisfiestheconditionsof theorem III in

(a + e, b) where e is as small and b as large as we please, and thereexists
a functionFi(x) continuousin (a + e, b) and the derivativeof which equals
f(x) at everypoint wheref(x) is continuous,whenfinally
lim F1(a + E) = Fi(a)

f
e-o

and limFi(b) = F1( oo) bothexist,then f(x)dx existsand

b-~~~~~~~~~~~
00~ ~ ~ ~ ~ ~~~~0
F1(oo) - Fi(a) = f (x)dx.

Similarly, it is seen that when the conditions of theorem IX are fulfilled

in (a + e, b), where e is as small and b as large as we please, and three
of the four expressions f u(x)v'(x)dx, J'v(x)u'(x)dx, u (a)v(a) =

lim u(a + E)v(s + e) and u(0o)v(oo) = lim u(b)v(b) exist, then the fourth
e* 0
b0>0
exists and
,00 r
(33) fu(x)v'(x)dx = u(oo)v(oo) - u(a)v(a) - v(x)u'(x)dx.

Finally, it is readily seen how theorem X, on integration by substitution,

may be extended to the present case.
A convergent infiniteseriesu0 + u1 + * + un+ maybe written
as an infiniteintegralby makingf(x) = unforn c x < n + 1. In the
interval (0, b), where n + 1 b < n + 2, f(x) has the only discontinuities
x = 1, 2, ***, n + 1 and is integrable by theorem III, and (15) gives
rb n V+bv+1 7

tf O (x)dx = E
' f(x) dx -'
xd = E U+ (b - n-1)u1n+l
=O?v In+ v=20

The series being convergent,un+1 -* 0 as n -o oo, so that

I (b - n - 1)un+1 I < I Un+1 I -o 0.

In the following,the concept of uniformapproach to a limit (or uniform

convergence)is of fundamental importance. Let f(x, a, w) depend on the
parameters a and w; then f(x, 6, c) tends toward the limitf(x) uniformlyin
(a, b) as a -o 0 and X -* oo when, for any e > 0 as small as we please,
there exist a 60 = 60(e) > 0 and an coo= co(e) > 0 independent of x such
that for every x in (a, b)
If(x, a, c) - f(x) I < e for0 < a < So, co > co.
In particular, let f(x, 6, w) be independent of a and make

f(X, C) =
v=O
UVu(x)

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
54 T. H. GRONWALL.

for n c co< n + 1; we then obtain the familiardefinitionof uniform

convergenceof a series.
THEOREM XVIII. When for a c x _ b, f(x, 6, c) is boundedand
integrablein respectto x and tendsuniformly towardf(x) as 8 -* 0 and
rb
co -* o, then f(x)dx existsand
rb rb

lim fb(x 8, w)dx = fbf(x)dx.

Box0

For any e > 0, thereexist,by hypothesis,a 8o = bo(e)and an coo= c(e)

such thatf(x, 6, co)- e c f(x) f(x, 8, co) + E in (a, b) for0 < a < boy
w > coo. Hence f(x) is bounded in (a, b) so that its upper and lower
integralsexist; the applicationof (11) and (9) to the above inequality
gives
ff(x)dx c j f(x, 6, cw)dx+ E(b - a),
rb rb

Jf(x)dx _ f(x, 6, c)dx - E(b - a).

Since f(x, 6, co) is integrable,its upper and lower integralsare equal,

ob bt
whence 0 c f(x)dx - f(x)dx 2E(b - a); but f(x) being inde-
rb

pendentof E,it followsthat f(x)dx exists,and the precedinginequalities

become
- E(b - a) f(x)dx - f(x, 6, co)dx (b - a)

for0 < 6 < boyw > wo,whichprovesour proposition.

fory in (c, d) when
The integralJ'f(x, y)dx is said to existuniformly
f(x, y) is bounded and integrablein respect to x in (a + E, b) and the
fb rX
fory in (c, d),
integral ff(x, y)dx tendstowardJ'f(x, y)dx uniformly
that is, when
pa+el pb
lim f(x, y)dx = 0, lim f(x, y)dx = 0
e-O a+ e b-- Xb
e
f 1-. by X>

uniformlyfor y in (c, d). A convenienttest for uniformexistenceis

containedin
THEOREM XIX. When f(x, y) is bounded and integrablein respectto x
in (a + (, b), while y is in (c, d), and when, for x > a and y in (c, d),

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 55

f(x,y) I < M(x) where f M(x)dx exists,thenff(x, y)dx existsuni-

fory in (c, d).
formly
In fact, the proof of theorem XVI then shows that both the integrals
fa+ef Jb

f(x, y)dx and f(x, y)dx

a+e

as
approachzero uniformlye, -' 0 and b, b-' oo.
O
We furthermorehave
THEOREM XX. Whenf(x, y) is boundedand doublyintegrable
for
a + E ~ x ~ b,c ~ y d, and ff(x, foryin (c, d),
y)dxexistsuniformly

f
then
pd pco 0o pd

fddyf f(x, y)dx = dxf f(x, y)dy.

By theorem XIII, we have

pd

fdy J
pb

f(x, y)dx = f b

dx
pd

f(x, y)dy,

and since ff(x, y)dx exists uniformlyfor c y d, theorem XVIII

fcdy lim f dy
shows that
af(x, y)dx exists and equals~~~~~ c
~~~~~e--O +e
f(x, Y.)dx.

f
b-4co

The preceding equation then shows that lim dx ff(x, y)dy exists
a c

f
e-* e
b-)> oo
pd c

and equals dyJf (x, y)dx. But the last limit, by the definitionof an
a
c~~~~0 poo pd
J
infiniteintegral,equals J dx f(x, y)dy, whence our theorem.
Making f(x, y) = un(y) forn x < n + 1, where all un(y) are bounded
and integrablefor c c y _ d, the uniformexistenceof J'f(x, y)dy is
00

equivalent to the uniform convergence of X un(y)

0
in (c, d), whence
writing x, a, b for y, c, d, we find the
COROLLARY. WhenXun(x) converges in (a, b), everyUn(X)
uniformly
0

beingboundedand integrable,then

r
b oo oo b

EUn (x) dx = Un(x)dx.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
56 T. H. GRONWALL.

THEOREM XXI. When f(x, y) and Of(x, y)/dy are continuousfor

a + e c x c b, c _ y c d, ff(x, c)dx exists and f (Of(x, y)/Oy)dx

fory in
fory in (c, d), then ff(x, y)dx existsuniformly
existsuniformly
(c, d), and
dy f f(x, y)dx = Of(x, dx.

All conditionsof theoremXIV beingfulfilledforx in (a + E, b) and y in

(c, d), we have

ff(x, y)dx = f(x c)dx + fqdtf t) dx.

a+ a~~e c crv
eaf(s
Since the firstintegralto the rightis independentof y and tends toward
ff(x, c)dx as e -* 0, b -+ oo, and since ff dt existsuni-
(Of(x, t)/Ot)dx

fory in (c, d) by the hypothesison

formly f (Of(x,y)/ly)dxand theorem

XVIII, it followsthat J'f(x, y)dx existsuniformly

for y in (c, d) and

('f(x, y)dx = (x, c)dx + Fdt U f(-, )dx.

From the continuityof of(x, y)/dyfor a + c x b, c y _ d, and
f
E

the uniformexistenceof (Of(x, y)Idy)dx fory in (c, d), it followsat

once that the latterintegralis a continuousfunctionof y in (c, d). Con-
sequently,differentiatingthe last equation in respectto y and applying
theoremVI, we obtain our proposition.
Making f(x, y) = un(y) for n c x < n + 1, and replacing y, c, d by
x, a, b, we have the
COROLLARY. WhenE (dun(x)/dx)converges uniformly in (a, b), every
0
~~~~~~~~~~~~00
E
X
then un(x)
(du.(x)/dx) being continuous,and Xun(a) converges, converges
0 0
uniformly in (a, b), and
d X0
do E = E-d
x 0un'(X)
dun(x)
0 ~dx
X0

Regardinginfiniterepeatedintegrals,we shall now prove

THEOREM XXII. Whenf(x, y) is boundedand integrable fora + E = x
= b, c + Et c y c d, whereE, (' are as smalland b, d as largeas we please,

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 57

and ONE of the infinite repeated integrals f dx If(x, y) I dy and

f dyf If(x,y) Idx exists,thenboththe infiniterepeatedintegralsof

f(x, y) existand are equal:

f dx fJ(x, y)dy = f dy fJ(x, y)dx.

From the existenceof fae cef

rb rd
J'f(x, y)dxdywe infer,by theoremXV,

that of
rb
f If(x, y) Idxdyand henceby theoremXIII that
rd

rb rd rd rb
f dxf f(x,y) dy= f dyJ If(x,y)Idx.
ae clei ce a
When in eitherof theserepeatedintegralsE and E' are decreased,b and d
increased,the integralcannot decreasesince If(x, y) I is never negative.
Consequentlyassumingthe existence.of dx f If(x, y) I dy,we have
rb rd rX r
f dxJ If(x, y) I dy cfdxf If(x, y) I dy,whence

rd rb rX r
J'dyJ If(xy) dxcJ'dxJ If(x,y)I dy,

so that the integralto the leftis bounded forall values of E, E b and d

considered. Since thisintegraldoes not decreaseas E, E' decreaseto zero
approachesa limitnot exceeding
it therefore
and b, d increaseindefinitely,
the right member, that is, f fa I dy f(x, y) I dx exists and
00c
fdyf If(x,y) I dx c fdxf If(x, y) I dy.

Having establishedthe existenceof f dyf f(x, y)I dx,we apply to it

repeatedintegral,
the same argumentas was applied to the otherinfinite
and therebyestablishthe inequality

fdx f f(x, y) I dy cfdyf If(x, y) I dx.

Combiningthe two inequalities,we find

fdx If(x,y) I dy = fdyf If(x, y) I dx.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
58 T. H. GRONWALL.

Furthermore,
the integral
rb

f (If(x,y)I
rd

- f(x, y))dxdyexists,being
of the double integralsof If(x, y) [ and f(x, y), and from
the difference
I
o f(x, y) I-f(x, y) c 2 If(x, y) I we see that
0 Jrb A

dxJ
d

(If(x,y) I-f(x,y))dy 2f dxff(x,y)Idy.

Since the integral to the left cannot decrease as e, E' decrease to zero and
b, d increaseindefinitely,
it followsthat dx f (If(x, y) I - f(x, y))dy
exists; we maynowshowexactlyas beforethatthe otherinfinite
repeated
integralexistsand that

dx (I f(x,y) |-f(x, y))dy = dy (If(x,y)I -f(x, y))dx.

Our theoremnow followsby subtractingthisequation fromthat relative
to If(x,y) I
Replacing f(x, y) by un(y) for n c x < n + 1 and writing x, a for
y, c we have the
COROLLARY. When everyu. (x) is finiteand integrable
in (a + E, b)
J E 0 I Un(x) I dx and ~~~~~~~~0
and eitheroftheexpressions a J Iun(x)Idx exists,
thenboththecorrespondingexpressionsin u.(x) exist,and

f Eu0 (x)dx = ~~~0JFun(x)dx.

6. Integralsof complexfunctionsof a real variable. In the followingtwo

chapters,we shall make extensiveuse of such integrals,whichare defined
in the followingmanner: let f(x) = g1(x) + ig2(x), whereg1(x) and g2(x)
are real functionsofx, boundedand integrablefora c x c b. Then, by
definition,
rb b rb
Jaf(x)dx + iJ 92(x)dx,

and the double and repeatedintegralsoff(x, y) = gi(x, y) + ig2(x,y), as

well as the various kinds of infiniteintegrals,are definedin the same
mannerby decompositioninto theirreal and imaginaryparts.
TheoremsII and XI, as well as formulas(7)-(13), apply to real func-
tionsonly; but all the remainingtheoremsand formulasmay be extended
to the presentcase. In fact,the proofof theoremI subsistsunchanged,
since it makes no use of the realityof f(xi, ..., Xn). Theorem III is

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 59

immediately extended by decomposition of f(x) into its real and complex

parts. In theorem IV, let M,, m,, M,,, mi,, M2., M2., be the upper and
lower bounds of I f(x) 1, I g1(x) I and Ig2(x) I in (x,1, x,), and x', x" any
two points in this interval. We have

f(x') I- i f(x") I - If(x') - f(x") I

= I (X') -91(X") + i(92(x') - 92(X")) I
1 9g(X') - 91(X") I + 192(X') - 92(X") I

(M1b - m1L)+ (M2v-mev),

and since for suitably chosen values of x' and x" i f(x') will differfrom
M., and If(x") I fromm,, as little as we please, it follows that
MV mv c(Mb - Mlv) + (M2v - l2v),
and consequently the S - s relative to I f(x) I may be made as small as
we please, since it is less than the sum of the S - s relative to Igi(x) I
and that relative to 192(x) I , both of which may be made as small as we
please by theorem IV. Theorems V-X, XII-XIV are immediately
extended by decomposition (it being observed that in X so(t) is a real
functionof the real variable t), XV is extended in the same way as IV, and
the proof of XVI remains unchanged, since (14) subsists in the complex
case. In fact, with the notations of theorem VIII, we have, since
XV- xvl > 0, forf(x) complex
l If (U (xv - owA Iz I AUiv I (x v-xv-);
as we have seen that theorem VIII extends to the complex case, it follows
that for 6 .- 0, the left side approaches fb(x)dx , and the right,

f
zb

f(x) I dx, whence we obtain (14). In exactly the same way, we extend
(26), while (15), (16), (17), where k may now be a complex constant,
(18), (27), (28), (32), (33) are extended immediately by decomposition.
Theorems XVII and XVIII extend by decomposition, the proof of XIX
remains unchanged, XX and XXI (with corollaries) are again extended
by decomposition, while in XXII (and corollary) we firstconclude from
Igl(x, y) I x, y) and I g2(x, y) I If(x, y) I that the same infinite
repeated integrals of g1(x, y) I and '192(x, y) I exist as that of I f(x, y) i
included in the hypothesis; the extensionis then finishedby decomposition.
We shall finallyprove the following
THEOREM XXIII. When,foran e > 0 as smalland a b as largeas we
f(s, t) oftwocomplexvariabless and t is holomorphic
please,thefunction for

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
60 T. H. GRONWALL.

a + e t b and I s-so l r, where r is independentof E and b, and

whenfinally J'f(s, t)dt convergesuniformlyfor I s -so I < r, thenthisin-
tegraldefinesa functionof s whichis holomorphicfor Is - sI < r.
Write s - so = x + iy and f(s, t) = g1(x, y, t) + ig2(X, y, t), and
denote by P the pointset defined by a + e ? t c b, x2 + y2 r2; since
f(s, t) is holomorphic on P, g9and 92 are continuous on P and this pointset
being closed, gq and 92 are uniformlycontinuous on P (theorem I). Since
g9and g2 are continuous, we may take the whole interval (a + e, b) as A"
in theorem III, and on account of the uniformcontinuity on P, we may
choose, for any e' > 0, a 6 independentof x and y such that S and s differ
by less than e' when all subdivisions of (a + E, b) are less than 6. Conse-
quently, as 6 0, the sum
-

If(s,
(8) (t) -t-1) = .g1(x, y,r,) (t - tv1) + i292(X, y, Tv)(t -tv-1))
fb
formed according to theorem VIII, tends toward f(s, t)dt uniformly
a+e
for I s - so r. But each f(s, r,) is expansible in a power series in
s - so convergent for s - 5o I c r, and thereforealso If(s, r,) (t -tv-,)
which contains a finitenumber of such terms. Let 61,62, * be a sequence
of positive numbers such that &n -* 0 as n -* 0, and let
(D.(s; E, b) = If(s, -r)(t -tv-,)
(s; E,b)
fora certain mode of subdivision where all tv - tV, < 6.; since (J,,,
r and
f
is a power series in s - So convergent for j s o- j since 5
rb

4),(s; e b) tends uniformlytoward f(s, t)dt,we conclude froinWeier-

b
strass' theorem on sequences of analytic functions that f(s, t)dt is
aLe

expansible in a power series in s - so convergent for j s o- j r. 5

Now choose sequences El, -2, * and bl, b2, * such that c71 -* 0 and

b, * oo as n - oo; since J'f(s, t)dtconverges uniformlyfor I s - so r,

a

I
rb,

the expressions f(s, t)dt,which are power series in s -so convergent

f
Sa+ e,

for Is - so j r, converge uniformlytoward the limit f(s, t)dt which

is therefore,by a second application of Weierstrass' theorem, a power
series convergent for I s - so I < r.
7. Fourierseries. Let f(x) be bounded and integrable forE c x = 1-,
where E is as small as we please, and suppose that f If(x) I dx exists.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 61

Since If(x) sin 2nwxxI I f(x) I and I f(x) cos 2nrx j If(x) i, it follows
fromtheoremXVI that the integrals
(34) a. = 2 ff(x) cos 2nirxdx, bn = 2 ff(x) sin 2nirxdx

both exist. Then the Fourierseriesoff(x) is definedas

a0
(35) 2? + Z1 (ancos 2nirx+ bn sin 2nrx).

In regardto the convergenceof thisseries,we shall need in the following

only this very special theorem:
When f(x) is holomorphicfor 0 < x < 1 (but not necessarily at the end
I
points of the interval) and If(x) I dx exists, the Fourier series of f(x)
to thesumf(x) uniformly
converges for e x 1 - e, wheree > 0 butas
small as we please.
Using the integration variable t instead of x in (34), we find
an cos 2n7rx+ bnsin 2n7rx
= 2 ff(t) (cos 2n7rtcos 2n7rx+ sin 2n7rtsin 2n7rx)dt
1
= 2 ff(t) cos 2nwr(t- x)dt,
and sincewe have'
1 + 2Zcos2 (t ) sin (2n + 1)r(t - x)
r(t - x)
-
V=1 ~~~~~sin
it followsthat
ao
+ x:(a, cos 2v7rx+
v=1
b, sin 2v-xx)

sin (2n ?w(t

1>r - x)
r(tdt)
=_ Ji f~t)
W sin - x)
3 Proof: in the finitegeometricseries
n e2(n+)i -
.e(2n+I)$s e-s
VO e 1is -e-i

e(a+(2nf+s)i _ e(a-I)i
2i sin ,
comparethe real and imaginaryparts on both sides:

cos(a+2)= sin ( + (2n + 1) sin (a

v=o 2 sin 0
> sin (ax+ 2vjB)forl
=cos
= (a + ) - cos (a + (2n + 1)3)
F, sin (at+2he

From the firstof these,we obtain the formulain the text by making a = 0, ~ r(t -x).

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
62 T. H. GRONWALL.

In the special case f(x) = 1, (34) gives ao = 2, a, = b, = 0 (v > 0), and

(36) becomes
("sin (2n + 1)ir(t-x)
sin r(t - x)
thisbyf(x) and subtractingfrom(36), it is seen that
multiplying

-2 + E (a, cos 2v~rx+ b, sin 2vrx) - f(x)

v=1

(37) = i( (f (0 - f W) sin~~~~~~~sin
(f(t)-f(x)) (2n +7r(t dt
- x) x) d
1)ir(t -

= J1 + J2 + J3,
whereJi, J2and J3are the integralscorresponding to the decomposition
oftheinterval(0, 1) intothethreeintervals(0, 6), (6, 1 - 6) and (1 - 6, 1).
Our theoremwill then be proved if we show that after assigningan
arbitrarilysmall e, a 6 may be chosensufficiently small and an no suffi-
cientlylarge,both independentof x, such that IJ1I + I J21 + I J31 will
be as small as we please forn > no and everyx in (E, 1 - E). Assume
0<6<E/2; for E x l-e and 0_ t 6 or 1- 6 t=1, we
then have E/2< I t -x < 1 - (E/2) and consequently I sin ir(t - x) I
> sin (ire/2); since If(t) - f(x) I If(t) I + If(x) I, the expressionunder
theintegralsignin J, and J3willbe less than ( If(t) I + If(x) I )/sin r(E/2),
whence
I + I 31 < sin (ire/2)[tt(If(t) I +I f(x) I )dt
+ f (I f(t) I + I f(x) I )dt].

As f(x) is holomorphicin (E, 1 - thereexistsa constantM such that

E),

If(x) I < M forall x in (E, 1 - E), whence

1l 1lI + IJ31 < sin (rE/2)(26M + jf(t) I dt + j f(t) I dt),

Since f If(t) I dt exists by hypothesis,the last two integralsmay be

made as small as we please by taking 6 sufficiently small, and the same
being true of 26M, it followsthat IJ1I + I J31 may be made as small
as we please by taking6 sufficiently small and independent of x. In J2,
thefunction
f(t) -f(x) (X,
sin r (t - x) 1- ox t)
is evidentlyholomorphicfor E x 1 - E, 6 _t <1 - 6, so that con-

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 63
stantsM1 and M2 existsuch that I p(x, t) I < M1 and

a(x, t) < M2
forall thesevalues of x and t. Integratingby parts,

J2= j (x, t) sin (2n + 1)w(t -x)dt

-[(2n 4- 1)w (x, t) cos (2n + 1)ir(t - x)1

1 ~ - 06a(X,t)
+ ( 1)cJ8 at C05 (2n + 1) (t- x)dt,
and since sp(x,t) cos (2n + 1)7r(t- x) [ p t) I < M1,
So(x,

a0 tt' t) cos (2n + 1)7r(t -x) < M2,

we have
2M1 + M2
[ 21<(2n -+ 1) 7r

Consequently,afterfixinga 8 so small that IJ +I? I < ?7,where 7 is

as small as we please, M1 and M2, whichdepend on 8 but not on x, are
fixed,and thelast inequalitynowpermitsus to determinean no(dependent
on a but independentof x) such that IJ2 I < - forn i 0noand all x in
(e, 1 - E). The expression to the left in (37) is then less than 2-qin abso-
lute value, whichprovesour theorem.
We shall now derive a formulawhich is useful in estimatingthe
remainderin some seriesoccurringin the next chapters: Let
Cn C ..
Cn+*p-1 __Cnp -0 0

k a constantand E > 0 but as small as we please, then

(38) Ec, sin (2v + k)7x sin 'Xc -e.

Usingthe identity
2 sin rx sin (2 v + k)7rx= cos (2v - 1 + k)7rx- cos (2v + 1 + k)rx,
we have
n-p
2 sin 7rxE c sin (2v + k)7rx
v=n
n+p n+p

= - C Cos (2v - -
1 + k)wrx cos (2v + 1 + k)rx
v=n v=n

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
64 T. H. GRONWALL.

n+p
-n COS (2n - 1 + k)7rx- i, (cl- cV) cOs (2v -1 + k)7rx
v=n+l1

-cn+, cOs(2n + 2p + 1 + k)rx,

and taking the sum of the absolute values of all terms to the right and
remembering that c, - cv ? 0,
n+p |nip
2 sin 7rx E cVsin (2v + k)7rx cn+ Z (Cv-1 -cv) + Cn+, = 2Cny
v=n+l I v=n+l
whence (38) follows at once.
8. Bernoulli functions and numbers. We now apply the theorem of the
preceding paragraph to f(x) = - x; (34) gives, upon integration by
parts, ao = 0, an = 0, bn 1/nr (n _ 1), whence
=
X=
sin 2n7xx
2 E 0 < x < 1,

the series being uniformly convergent for e _ x 1 - e. The series to

the right converges also for x = 0 and x = 1, the sum being zero in either
case, while the left side becomes i 2. Let [x] be the greatest integer
contained in x, so that 0 x - [x] < 1, then 2 + [x] - x remains un-
changed when an integer is added to x, and equals 2-x for 0 < x < 1.
Since the Fourier series to the right also has the period unity, we see that
if we write
(39) Pi(x) + [X]-X

then for any x which is not an integer

.0 sin 2nwxx
(40) Pi(x) = E
while for an integer m, Pi(m + 6) -* 2 and Pi(m - -)
- as a -
0
through positive values. Now write
0
Ecos 2nrxx
P2(X) = E 2nw2 '

the series being uniformly convergent for all values of x, since its general
term is less than 1/n2 in absolute value. Then, since (40) converges uni-
formly for E c x 1 - E, we have
dP2(x)1
-Pi(x) = x -2
dx -
(theorem XXI cor.) and therefore
x
P2 (X) = - 2+
XP2 )
c

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 65

we observe
in thelatterinterval; to determinethe constantofintegration,
that,on account of the uniformconvergenceof the series,
O'P cos2n7rX 0' cos 2n-x7rx
fP2(x) dx 2=2dx2 x 0

and therefore

fl(xZ<Zx?)dox = or

Now the polynomial2X2 - 2X + takes the same value (= -1'2) for

x = 0 and x = 1; consequentlythe expressionI (x - [X])2 (x - [x])
+ -112admitsthe periodunityand is continuousfor all values of x, so that
the relation

P2(x) = (x - [X])2 (X - [X]) + 1 =- E 2r,-

whichwe proved for 0 < x < 1, subsistsfor all values of x since both
membersare continuouseverywhereand admit the period unity. We
observe that the integral-valuesof x formno exceptionas in the case
ofPi(x).
Continuingthis process, we definefunctionsPk(X), k = 2, 3, *..,
by the relations

Pk (X) =(-l)
=(- k ddPk+l(X)
Pkr))'ldx

(41) fPk+l(x)dx = 0,

Pk+l(X + 1) = Pk+l(X),

and show by completeinductionthat Pk(X) is a polynomialof degreek

that Pk(X) is continuousforall values
in x - [x] withrationalcoefficients,
of x when k > 1, since Pk+1(l)- Pk+1(O) = (-1)k Pk(x)dx = 0, and
that
X cos 2nwrx
P2k(X) = 22k-1 2k 2k

(42) 00 sin 2nrxx

P2k+l(X) = 22kn2k+1 2k+1l

convergentforall values ofx.

theseseriesbeingabsolutelyand uniformly

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
66 T. H. GRONWALL.

From (42) it followsthat,whenm is an integer

P2k+l(m) = P2k+1(m + D) = 0,

(43) P2 k(m) = 22k-17r2k En 2kX

P2k(m + 2) = - - _ 2k(M),

the last relationresultingfromthe identity

00 1 n 00 1 001 1' 1\001
Z
E 2k- 2Z (2)k
2k i jn n2k
(2)=21 n2k 22k-1

Furthermore(42) gives
(44) Pk(l - x) = (- 1)kPk(X) = Pk(- X).

We shall now show thatfor k odd and > 1, theequation Pk(x) = 0 has the
roots0, 2, 1 and thatPk(x) > Ofor0 <x < 1, PJ(X) <0 for 2 <x < 1,
whilefor k even,thereis one and only one rootinteriorto each of theintervals
(0, 2) and (2, 1). From (43) it is seen that Pk(0) = Pk(2) = Pk(1) = 0
fork odd, but that fork even none of these expressionsvanishes. For
k = 2, our theoremis true, since P2(X) = -X2-x + 12 in (0, 1); we
may thereforeassume the theoremproved up to the index k - 1 and
deduceits validityforthe indexk in the followingmanner. If, fork odd,
Pk(X) = 0 has a rootin (0, 1) distinctfrom0, I, 1, we mayassume,by (44),
that it is interiorto (0, 2) But since

Pk-l(X) - dk(X)

by (41), Pk-l(X) = 0 would then,accordingto Rolle's theorem,have two

rootsinteriorto the interval(0, 1), contraryto the provenfactsfork - 1
even. Thus Pk(x) does not change its sign for0 < x < ', and since for
x = 0,
dPk(X) - P (O)
dx Pk10

is positiveby (43), we findthat Pk(X) > 0 for0 < x < 2, and (44) then
gives Pk(X) < 0 for 2 < x < 1. If, for k even? Pk(x) = 0 has two roots
interiorto (0, 1), Pk-l(x) = 0 would have a root interiorto the same

interval,contraryto theprovenfactsfork - 1 odd. The same argument

applies to the interval (2, 1).
The expressions P2k(m) = P2k(0) of (43) also occur in the expansion
of the cotangentin a power series. From the decompositionof the co-

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 67

tangentintopartialfractionswe obtain
t t 0 2t2
-cot
2 21 = 1 2E 412 2-t2

convergentfor I t
the seriesbeinguniformly 2w - E. For such values
of t,we have
2t2 t2k

47r2 2 - t2 =1 22k-l1r 2k 2k

and by Weierstrass'stheoremon seriesof analyticfunctions,

we obtain
ct t 0 1 _

-2cot-2= 1 - ktE
=___'0 tk
22k-1 2k 2k

We now definethe kthBernoullinumberBk by4

(2k)! 1
(45) Bk 22kl172k 2k
and obtain5
t t 00Bkt2k
(46) ~~~
~~2 2 ~ k=l (2k)
In
t t tieti'2 + e-ti2 ti ti
2 -c
o 2
cot = 2 eti2
ti 2 ti__ e- /2 =
2 ++ eti _i _ _

we now replace t by - ti; (46) thenbecomes

(47) 1 - 1= 2+ E (2k) |t
< 27r.

Multiplyingboth sides by
tk
Z
o

-t_1 =

we find
00 tk 00 0
{t ??0(
)
tk 1)kBkt2k

Ek~k'rk!-= k!Jk 2 (2k)T

of t2k+1 on both sides
and comparingcoefficients
1 1 _ B1 B2
(2k + 1)! 2(2k)! (2k - 1)!2! ? (2k -3)!4!
B3 (_ B)kBBk
(2k -5)!6!?+ +
1!(2k);
4These numbersappear forthe firsttime in Jacob Bernoulli's Ars conjectandi,Basel, 1713,
p. 97, in connection with the problem of expressingthe sum 1k + 2k + 3k + . . . + (n - 1)k
as a polynomialin n.
6 Euler, L., Introductioin analysis infinitorum,
Lausanne, 1748, p. 161.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
68 T. H. GRONWALL.

this recurrentformulaallows us to calculate successivelyB1, B2,

whichare clearlyall rationalnumbers. The firstten Bernoullinumbers
are thus foundto be
B- 1 B2= w, B3=A, , B5=w, B6= 69
B4=.
B7- 67 B8 =- 356Qo-7, B9 4T9X7 Blo= 134 3%o1

The comparison of (43) and (45) gives, since Pk(m) = Pk(O) for k > 1,
P2k+1(0) = P2k+l(2) 0,

(48) P2k(0) = (2k)! Bk P2k(2) = - (2k)!(1 - 2k-1 )Bk

The left side of (47) is the special case x = 0 of 1 - [(tetx)/(et- 1)],

and the latterexpressionmay evidentlybe expandedin a powerseriesin t,
beingpolynomialsin x, the seriesconverging
the coefficients uniformly for
] tI 2 r- E and I x I < a, where E is as small and a as large as we
please. We writethis expansion

(49) 1- e - E (- l)k(k1) I2pk(x)tk.

et k= 1
Since
1- _
1 + etx(1 et 1
we obtain from(47) and (49)
Z(- _1\)k(k-1) /2 (tk = 1 + ( E (- )kt2k

oflike powersof t,
and comparisonof coefficients
and upon multiplication
X2k 2k-1 BIX2k-2 B2X2k-4
()' (2k-4)!4!
-~1)k<2k~x) 2-(2k)! 2(2k-1)!(2k-2)!2!
B3X2k-6 + * k + (-lk B k-1X2
(2k -6)!6!?2(k )
1 )k Bk
+(~ (2k)!
2k+1 X2k Bix2k-1 B2x2k-3
(~ 1) <2k1(x) (2k + 1)! + 2(2k)! (2k 1)!2! + (2k - 3) !4!

B3X 2k5 + B k-1X3

(2k -3-5)-!6 3!2k - 2)!

+ (- ) k -BkX

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 69

for k > 0, while spi(x) = - X, SO that for 0 < x < 1, spi(x) = Pi(x).
We shall now provethat fork > 1 and 0 c x c 1
SPk(X) = Pk(X).
Since for0 c x c 1, the Pk(x) are completelydeterminedby the firsttwo
relations(41), it is evidentlysufficientto show that the sPk(x) satisfythe
same relations. Differentiating (49) in respectto x, we obtain,since the
differentiationoftherighthand seriestermby termis legitimateaccording
to Weierstrass'stheoremon seriesof analyticfunctions,
p
Eii(- )~k(k-1) /2
dpk(x) tk = te = t( - 1)k(k-1) /2 k(X)tk -
1 ~~~dx el -1 \

and comparingcoefficients
of tk+1
d(
(_ )k d+ (X
dx= = ~(X)*

Integrating (49) fromx = 0 to x = 1, we find

EZ( _1l)k(k-1) /2tk X Pk(X)dX = [X et ] II

whence
Xfk(x)dx = 0.

Outside of (0, 1), 'Pk(x) and Pk(x) do not coincide,since Pk(x) has the
periodunity,but f k (x)has not. To findthedifference Pk(X + 1) - (Pk(X),
changex into x + 1 in (49) and subtract:
?? ?o Ok-itk
_)k(k1)/2(Pk(X ? 1) - Pk(X))tk = - tea = - E-1)

whence
xk-1
(Pk(X +1) - k(X) = - ( k) (k-1) /2(k1_

Changingk into k + 1 and adding these equations forx = 0, 1, 2,

n- 1, we find
lk + 2k + 3k + * + (n - 1)k = (_ 1)k(k 1)j2k!(0k+l(n) - Pk+1(0)),

whichis essentially
6 the formulaestablishedby Bernoulli.
6 The Bernoulli polynomials (as distinct from the periodic Bernoulli functions Pk(x)) are
commonlydefinedby
et -1 s ( h

and by comparison to (49), we readily find the relation -pk(X) =(_ l)k(k-l)12(<k(O) - 'k(x))
between this definitionand that of the tbxt; the latter has been chosen in view of its application
to Euler's summation formula in the next paragraph. For furtherinformationon Bernoulli
numbers and polynomials, see Saalschiitz, L., Vorlesungen uber die Bernoullischen Zahlen,
Berlin, 1893.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
70 T. H. GRONWALL.

9. Euler's summationformula. Let F(s) be a real or complex function

of the real variable s, having continuous derivatives up to the order 2m
inclusive, or we may suppose s complex and F(s) an analytic function;
furthermorelet 0 c a < 1 and b > 0. In the identity

F(s + nb) - F(s + ab + vb) = bJ F'(s + tb)dt

make v = 0, 1, 2, *.* ,n - 1 and add, obtaining

n-1 n1 n
nF(s + nb) - F(s + ab + vb) = bEf F'(s + tb)dt.

Decomposing each integral by the formula

F(s
iinaF'(8 + tb)dt=J ('v+1
F(s ++l tb)dt+ /L~v1 /+1
F'(s + tb)dt
va va uvlX

we obtain
n-1

nF(s + nb) - F(s + ab + vb)

v=O

_ J
n-1 (V+vu n-1 n-1 +

bE >JvF'(s + tb)dt+ b F'(s + tb)dt;

v=O Va v=O ,u vf X

in the double sum, each integral between limits ,uand ,u+ 1 occurs ,utimes,
namely for v = O, 1, 2, *-- , u- 1, so that
n-I n-1 [ffi+l n-1 + 1

J F'(s + tb)dt = Z J F'(s + tb)dt,

v=o ,u=v+l lu ,=l

or adding to the last sum the zero term correspondingto u = 0, and writing
v instead of ,u
n-1

nF(s + nb) _ZF(s + ab + vb)

v=O

vJ
n-1 (+l (v~l
= bZ( J F'(s + tb)dt+ F'(s + tb)dt
v=o vav
In the integrals with lower limit v + a, we may reduce this limit to v by
observingthat forv _ t < v + a, [t - a] = - 1, or [t - a] -v + 1 = 0,
while for v+a t< v+1, [t-a]= v, or [t-a]-v+1=1;
consequently
(v+1 v+1
([t -a]- v + 1)F'(s + tb)dt= J F'(s + tb)dt,
v vf~~~~~~~~~~~~~~a
and substituting in the previous equation
n-1 n-1 pvrl

nF(s + nb) -ZF(s + ab + vb) = b E ([t - a] + 1)F'(s + tb)dt

v=O v= v
n
= b |([t-a] + l)F'(s + tb)dt.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 71

But from (39), [t - a] + 1 = t - a + 2 + PF(t - a), and integratingby

parts,we find

bf(t - a + 1)F'(s + tb)dt

n
= (n -a + )F(s + nb) -(-a)F(s) -JF(s + tb)dt;
all thisin the precedingformula,we obtain
introducing
n-1 Pn

ZF(s + ab + vb) = rF(s + tb)dt+ (2-a)(F(s) - F(s + nb))

(50) V=n
- b f'P(t - a)F'(s + tb)dt.

Integratingthe last integralby parts, using the firstand thirdof (41)

and the secondpart of (44), we see at once by completeinductionthat
"-1
,F(s + ab + vb)
v=O

= J'F(s + tb)dt+ (2-a)(F(s) - F(s + nb))

(51) + E (- 1)[k(k+l) 12]+1bkPk+l(a)(F(k)(s + nb) -

F(k) (s))
k=l

+ (- 1)mb2mf P2m(t - a)F(2m)(s + tb)dt,

0 a < 1, b > 0.
In the special case a = 0, the values of Pk+l(a) are given by (48), and
adding F(s + nb) to both sides in (51), we obtain Euler's summation
formula7
n rn
,F +vb)= JF(s + tb)dt+ I(F(s + nb) + F(s))
v=OO

(52) + Z
1) _ kb2k-l(F(2k-1)(s + nb) - F( -)(S)
k=1 (2k)

+ (- 1)mb2mJ' P2m(t)F(2m)(S + tb)dt.

7 Euler, L., Methodus generalis summandi progressiones. Comment. Acad. Petrop., vol. 6
(1732-33, published 1738), pp. 68-97, and independently,Maclaurin, C., A treatise on fluxions.
Edinburgh, 1742, art. 352 and 828-30, 847. Neither author gives the remainderterm. The
proofin the text was given,in the special case a = 0, by Wirtinger,W., Einige Anwendungender
Euler-MaclaurinschenSummenformel,insbesondere auf eine Aufgabe von Abel. Acta Math.,
vol. 26 (1902), pp. 255-271.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
72 T. H. GRONWALL.

In view of the application of (51) to the approximate evaluation of

the sum to the left,it will be necessary to investigate the remainder term,
or the last term to the right. The term corresponding to k = 2m - 1
in the sum to the right in (51) is
nb) -
f
(- 1)m-lb2m-lP2m(a) (F(2m-l) (s + F(2m-l) (s))

- (- l)m-1b2mP2m(a) F(2m)(s + tb)dt,

so that (51) may be writtenin the form
n-1 vn
ZF(s + ab + vb) = fF(s + tb)dt + (2 - a) (F(s) - F(s + nb))
(53) 22m?
+ X (- 1)[k(k+l)12+1lbkPk+l(a)(F(k)(s + nb) - F(k) (S))+ R2m,
k=1

(54) R2m = (- )mb2mf

(P2m(t - a) - P2m(a))F(2m)(S+ tb)dt.

Now suppose s and F(s) real, and assume that the sign of F(2m)(s + tb) does
not change for 0 c t _ n, so that, writing a = 1 or - 1, aF2m(s + tb) :O
in the interval of integration in (54). Since P2m(t) takes its maximum
value for t = 0, we have
| P2m(t - a) - P2m(a) | - P2m(t - a) | + I P2m(a) I < 2P2m(O),

and by (14) and (13),

f (P2m(t - a) - P2m(a)) -aF(2m)(s + tb)dt

vn
J 2P2m(0) -aF(2m)(s + tb)dt

b )
2P2m(- (F(2m-l)(S + nb) -F(2m-1)(S)),
so that 8

(55) | R2mI < 2b2m-lP2m(O)I F(2M-1)(s + nb) - F(2m-l)(S) I

In the two most important special cases a = 0 and a = we shall now
1,
obtain more accurate approximate values of R2mby introducing stronger
assumptions regarding F(2m)(s).9 Let a = ?+- 1, and assume first, that
8 For a = 0, a slightlymore accurate formulawas given by Poisson, S. D., " Sur le calcul

numeriquedes Int6gralesdefinies,"Mem. Ac. Sc. Paris, vol. 6 (1827), pp. 571-602. Other inves-
tigationsof the remainderterm (in the special case a = 0) by Jacobi, C. G. J., "De uso legitimo
formulae summatoriaeMaclauriniana," Journ.f. Math., vol. 12 (1834), pp. 263-272, and Malmsten,
C. J., "Sur la formule,etc.," Acta Math., vol. 5 (1884), pp. 1-46 (correctedreprintfromJourn.f
Math., vol. 35 (1847), pp. 55-82).
9 For a = 0, this methodis due to Sonin, N., "Sur les termescomplementairesde la formule
sommatoired'Euler et de celle de Stirling,"Ann. de l'Ec. Norm.,ser. 3, vol. 6 (1889), pp. 257-262.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 73

aF (2m)(s + tb) is positiveand decreasingas t increasesfrom0 to n + 2, and

second, that
n-1

aE (F(2m)(s + vb + tb) + F(2m)(s + vb + b - tb))

v=0

is positiveand decreasingas t increasesfrom 0 to 2. Then,for a = 0,

(56) (_)m-lR2m = b2m-lP2m(0)(F(2m-l)(5 + nb + hb) -F(2m-l)(S + hb)),

0 < h< 12
Since a = 0, we may write (54) in the form

( l)M-1bR-2maR2
= frn
(P2m(0) - P2m(t))aF(2m)(S + tb)dt,

and since P2m(t)has the period 1,

I1 n-1
(- l)M-1b-2mR2m = (P2m(0) - P2m(t))a ZEF(2m)(S + vb + tb)dt.
v=O

Decomposing the integral in two with limits 0, 1 and 1, 1 respectively,and

introducing 1 - t as integration variable in the latter, we find by the
aid of (44) that
)m-lb-2maR2m= j (P2m(0) - P2m(t))a (F (2m)(s + vb + tb)
(57) 0v=0
+ F(2m)(s + vb + b - tb)dt.
To obtain an upper bound for this expression, we shall use a theorem due
to Tchebychef: When u(t) and v(t) are continuousfor a ! t c b, and both
thesefunctionsincrease (or bothdecrease) as t increases, then

(b -a) fob
u(t)v(t)dt > jbb
u(t)dt f ob
v(t)dt,
but
ob b ob
(b -a)f u(t)v(t)dt < fu(t)dtf' v(t)dt

when one of the functions increases and the other decreases.'0 Writing
on
10 The followingsimpleproofis due to Franklin,F., "Proof of a Theorem of Tchebycheff's
Definite Integrals," Am. Journ.,vol. 7, pp. 377-379: Consider the double integral

(u(t) - u(x)) (v(t) - v(x))dtdx= dt [u(t)v(t) -u(t)v(x) -u(x)v(t) + u(x)v(x)]dx

a a ~~~~~~a
a

= (b -a) u(t)v(t)dt u(t)dt .f v(x)dx-f0 v(t)dt.f u(x)dx + (b-a) u(x)v(x)dx,

or replacingx by t in the singleintegralsto the right

(u(t) - u(x)) (v(t) - v(x))dtdx = (b -a) u(t)v(t)dt- Ju(t)dt * f v(t)dt.

When u(t) and v(t) vary in the same sense, u(t) -u(x) and v(t) - v(x) have the same sign, so

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
74 T. H. GRONWALL.

u(t) = P2m(O) - P2m(t), we have, by (41), u'(t) = P2m-l(t) > 0 for

0 < t < 2, so that u(t) is positive and increases with t in (0, 1), and by
(41) and (43),

f'u(t)dt = [P2m(0) t - P2l(t)]t_= = 12_

_ (O)

The second factor in (57),

n-1
v(t) = aE (F(2m)(s + vb + tb) + F(2m)(s + vb+ b -tb)),
v=O

is positive and decreasing by hypothesis, and

r
f'v(t)dt = f r~~~~n
aF(2n) (s + tb)dt;

consequently, by Tchebychef's theorem,

(58) P2m(O) f rn

aF(2m)(S + tb)dt > (- 1)m-lb-2maR2.

To obtain a lower bound for (57), we observe that u(t) is positive and
v(t) decreasing in the interval of integration,whence
i ~~~~~~~U-
1

fu(t)v(t)dt > v(-) J (t)dt = P2m1(0) aFS(2)(S + vb + 1b).

v=O

(s + vb + lb + tb) is positive and decreasing when t increases

But acF(2m)
from0 to 1, and consequently

aF(2m)(S + vb + 1b) > J cxF(2m)(s+ vb + 1b + tb)dt,

so that, by the preceding inequality,

rn
(59) (-1)mlb-2mnaR2m> P2m(O0)faF(2m)(s + lb + tb)dt.

Since by the firsthypothesis on F(2m)(s), the integral

fn

J'aF(2m)(s + hb + tb)dt

is a continuous and decreasing function of h as h increases from 0 to i,

the comparison of (58) and (59) shows that there exists an h interior to
that theirproductis positive,and thereforealso the double integral,whichproves the firstpart of
the theorem. The second part followsfromthe firstupon replacingu(t) by - u(t).

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 75

(0, 2) such that

( 1)M-1b-2maR2m
= P2m(O) aF (2m)(s + hb + tb)dt,

and evaluating the integral, we obtain (56).

Now let a = ? 1, and assume first,that aF(2m) (s + tb) is positiveand
decreasingas t increasesfrom0 to n + 1, and second, that
n-1
aE (F(2m)(s + lb + vb+ tb) + F(2m)(s+ 2b+ vb - tb))
v=O

is positiveand increasing as t increasesfrom 0 to 2. Then, for a =

(-1) m-lR2m = b2m-lP2m( 2)(F(2m-l) (S + nb + hb) -F(2m-l) (s+ hb)),
(60) 0<h
We obtain from (54), in the case a = 2

3
rn-i

( 1)mb-2mR2m
= (P2m(t) - P2m(2)) *aF(2m)(s + 2b + tb)dt

= Jq(P2m(t) - + 2b + vb+ tb)dt

-O ~~~~~v=O
P2m(())aZF(2m)(S
and decomposing the integral in two with limits 0 and 0, 2 respec-
tively, and introducing - t as integration variable in the first,(44) gives
i* n-1

( =
1)mbb-2mR2m (P2m(t) - P2m(2)) *a E (F(2m)(s + 2b + vb + tb)
v=O

+ F(2m) (S + 2b + vb - tb))dt.

Here the firstfactor under the integral sign is seen at once to be positive
and decreasing, while the second factor is positive and increasing by
hypothesis, and (60) is now readily proved in the same way as before.

CHAPTER II.
Definite integrals for the Gamma function derived from its definitionas an
infiniteproduct.

10. Expression of the Gamma functionby Euler's integral of the second kind.
IntegralsforP (s) and Q (s). Assume T (s) > 0 (when s = x + yi, T (s) = x
is the real part of s), let t be a real variable, and definets as e logt where
log t denotes the real value of the logarithm of t; since

ts-1 =d s

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
76 T. H. GRONWALL.

it follows that
ft t~[1]
dt =hlm -
8
e-> 0 gt=e

Replacing s by s + 1 and subtracting the new integral fromthe previous

one, we find
1~ ~ ~
Jt81(1 - t)dt +1

and, generally, from

C' ~~~~~(n-i)
- -t)n-'dt = s(s + 1).*.(s + n-1)
it follows that
()l
~~~~~(n-i) (n-i)
jt~ t5- lU'at1nd =s(s +1) (s + n-1)(s + 1)(s + 2) ...(s + n)
n!
s(s +l (s +n)'
or finally,replacing t by tin,
, t8-l 1 _ tdt =fl fs

n, s(s + 1).(s + n)
For n - co, the expression to the right approaches the limit F(s), as is
seen from (J4") upon replacing n by n + 1 and a by unity, and since
(1 - t/n)n e-t, the above equation suggests that

(61) r(s) = e-tt8-ldt, TY(s) > 0.11

To prove this, we must first establish the convergence of the infinite

integral, which is known as Euler's integral of the second kind. Writing
s = x + yi, and assuming A - x - e > 0, where A is as large and e as
small as we please, we have Ie-It-1' = e-ttx-1, and for 0 < t c 1,
e- ttX-l< te-1,while for t sufficientlylarge, et > tA+1and
e ttxl1 < t-(A+l)tA-1= t-2.

Hence by theorem XVI corollary and theorem XIX, the integral in (61)
exists uniformlyfor A : TJ(s) : e > 0. In the second place, it must be
shown that the left and right members of (61) are equal, or that
rnt /o n
limJ ts-1'1- dt = J e-tt-ldt, T(s) > 0.
11For s real and positive,this expressionis due to Euler (J10).

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 77

We have

e- tta-ldt t 1- ) dt

= f( (et-(i -_) t
)t8-1dt +f e-tt-ldt,

and thesecondintegralto the rightapproacheszeroas n - oo, on account

of the convergenceof the integralin (61). Using the inequality12

0 _e- t - t) 2e-te~t for O t n,

we find

fy(e- ( 1 ) )t8 dt fJ(e- (1 ) t dt

_ me-
ttx-ldt,
< - we- ttx+ldt 0O,
n
no n n
the last integralbeingconvergentand independentof n. Thus the proof
of (61) is complete.
Replacingt by at, wherea > 0, we obtainfrom(61)

(62) (,) = C-a tt8-ldt, a > 0, T (s) > 0.

The integral
e- tt8-1 log tdt

convergesuniformlyfor A i x e c > 0, since for t sufficiently

small,
e-tt8-1log t I < tf- | log t I < t(e12)-l, and fort sufficiently
large,
e-tt8-1log t I < t-2 log t < t-312;
12 Proof: forb a _O. we have
0 b- - an = (b - a)(bn- + bn-2a + ... + an-)c nbn-1(b - a).
When b = e-tln, where0 t n, and a = 1-(t/n), we have
3
= 1 (t
2
b -a _ 1 (t +...

so that, the seriesbeing alternatingwith numericallydecreasingterms,

0 0~~~i
b- -a _ ~
2
Consequently
0 Ce-t t- 1t(iqn
_ . -2(-
(1_t)cne-('-I/n)t )= 2-*
t
e-t-
e
-2e-t-
t2
t2e n
2e-t

The inequality and its use in proving (61) are due to Watson, G. N., "An inequality associated
withthe Gamma function,"Messengerof Math., vol. 45 (1915), pp. 28-30.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
78 T. H. GRONWALL.

(61) in respectto
consequentlyby theoremXXI, we may differentiate
x underthe integralsign,and since
ar(s) drP(s)
,Ox ds
we find
(63) dF(s) =fi tts-'log tdt, 9?(s) > 0,
underthe integralsign gives
and the repeateddifferentiation

(64) dnF(s) = t)ndt

e-1t8-1(log T (s) > O.

Integralexpressionsforthe functionsP(s) and Q(s) may be obtained in

the followingmanner: the expansion

e- tts-l - E (- 1): t8-l+v

V=O

by theoremXXI
for0 < e_ t _ 1, and therefore,
convergesuniformly
corollary
'e-tt8-ldt = E (_ 1)" 1 - E8+V

1
For T (s) = x > 0, the integral e- tt8-ldtconverges,and since

C-'8+1' EX 1s+v P Ix+ P+ yt _x+ Px,

we have

v=o P .S + v|-x '= !4?a O

so that, by the definitionof P(s) (J33),

(65) P(s) = Z
( +)L P) = e-t8ldt ST(s) > 0.
V=OP!s+. o

Since Q(s) = r(s) - P(s), (61) and (65) give

(66) Q(s) = f tt8-ldt.

As the integranddoes not become infiniteat the lower limit for any
finitevalue ofs, thelast integralconvergesuniformly for- A T 9(s) _ A,
whereA is as large as we please, and consequentlythe entirefunction
Q(s) equals the integral(66) forall finitevalues of s, by theoremXXIII.
The integrals(65) and (66) are due to Legendreand Prym(J34 32).
11. The hypergeometric series F(a, i, y; s) and its value for s = 1.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 79

Euler's integralof the firstkind. The name hypergeometricwas applied by

Wallis 13to the power series in s

+ a-0 + a(a + 1) *0( + 1) S2

.'y8 1.2.y(,y + 1)
a(a + 1)(a + 2) f3(0 + 1)(G + 2) 3 +
+
1*2*3.'y(,y + 1)(y + 2)
where a, f and y may have any real or complex values, except that oy
must not equal zero or a negative integer (for when y = - n, the n + 1st
and followingcoefficientsbecome infinite). Using the notations
( (a+ 1)(a +2) (a +n-1),
(67)
(a, n) = (n - l)!na/(a)n, (a, n) r
P(a) as n - oo,
already introduced in J ? 14 and J ? 2 respectively,we may write this;
series
(68) F(a, f, y; s) E (a) v

Many of the elementary functions are special cases of this series; for
instance
(1+ s)n = F(-n,f , 0; s), log (1 + s) = sF(l, 1, 2; -s),
ea= lim F(1, 3, 1; s/l);

the essential properties of the hypergeometric series, and its intimate con-
nection with the Gamma function, were discovered by Gauss (J6). By
(67), the absolute value of the general term in (68) may be written

(1, v)(y, v) SV = (1, )y, v) sIS

|(amp V) (13 ) V+ ) | I(a, V) (13 V) | v+(--B

and since the first factor approaches

P(,y)
I(a) I(0)
as v - o, and z l/pl+k converges for k > 0, but diverges for k c 0, it
is evident that the series (68) converge absolutely for I s I < 1, but is
divergent for I s I > 1, and that for I s I = 1, the series is absolutely
convergent when TJ(Qy- a - () > 0.14 To find the value of F(a, (, oy; 1),
which is a convergent series when TJ(Qy- a - () > 0, we begin by
Wallis, J., Arithmeticainfinitorum,
13 Oxford,1656.
By more elaborate methods,it may be shown that for Is I = 1, the series is (conditionally)
14

convergentfor0 _ R(-y - a - #) > - 1, divergentforR(-y - a - __ - 1, except fors = 1,

wherethe series divergesforR(-y - a - . 0

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
80 T. H. GRONWALL.

establishing the following relation between the three hypergeometric

series F(a, fi,'y; s), F(a, f, zy-1; s) and F(a, ,B,oy+ 1; s) for !I < 1:
(69) 'y['y -1- (2y - a - - 1)s]F(a, 3, y; s)

+ (y - a)Q(y - f)sF(a, f3,y + 1; s)

-
y(y - 1)(1 - s)F(a, f, - 1; s) = 0.
In fact, the constant term on the left equals y(-y- 1) -y(y - 1) = O0
and the coefficientof Sn, where n > 0, is seen by (68) to be

- 1) (1)n(1)n 2 a (a) n-1i(f3)n-1

'y(0y
t(7 1) 7( 8
-
(1)n(-f)n (1)n-l (a)n-1

+ ( - a)
a)Qy -
1) (l)n-1(0+)n-1

(a) n (13) n (a)n-1(13)n-1

-(' -1) - ln -
V(l)n(7 (1)n-l(Qy l)n-1
or denoting by
Ufl - (1) 1

of Sn in F(a, fi, y-1;

the coefficient s),

Sy(7y- 1 n(1+ + -2y)(-y + n-1)

+ n-1- + (a + n-1)( + n-1)
+ n
n - a)(-y-
(a + (-1- 1)(j3 + n )+ )-(r+n-1 )

+ n(-y+ n -2) (-y+ n-i)-

(a + n -l) + n -1)
_n-_) _1 (y-Oa)(y- )
+n-1
+y - (a + n -1)( + n-)
+ ( +n -1\ 1 - +n0
+ (a + n-1)( + yn - 1) )

In (69) we now let s tend toward unity through positive and increasing
values; since F(a, f, 'y; 1) and F(a, f, y + 1; 1) are convergent series on
account of T y-o- a ) > 0, Abel's theorem on power series (J31)
shows that
lim F(a, 3, 'y; s) = F(a, 3,'y; 1)
s-*1
and
lim F(a, 'y + 1; s)
-y, F(a, f, 'y + 1; 1).
=

+ 00
Furthermore(1 -s)F(a, fi,y - 1; s) =1 >U1( - u.1)s" for IsjI < 1

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 81

and since

1 +1+00(U
(u-t u =V)lmu
) (>8 n),y
lim 1bn=~ IiM_(1, 7 ,- 1,n)
,n 1 = 0

on account of TJ(yy
- a ) > 0, the theorem in question gives
lim (1 -s)F(a, (3, y -1; s) = 0.
8s-1

Thus we finallyobtain from (69)

F(a, (, y; 1) = (Y( )( ))F(a, ,y+1; 1)

and replacing y by y + 1, y + 2, .., y+ n -1 and multiplying the

correspondingequations

F(a, (, 'y; 1)- (7) e a)-na nF((, -y+ n; 1)

(y, n)( y -a (, n) F(a,-

y+ n; 1).
,y - a, n)Q(y-3n)
The first factor to the right approaches the limit r(-y)r(y- a -()1
r(,y- a)r(y - () as n -> c, and for n > I ^y I, we have
(1)(Qy + n)~
F(a,, ,y + n; 1) 1I Z
V=i

v=I (l)v(y + n), v=l(1),(n - )v

I?eI 00 a + 1)v(l 0 I + l)v .

<n - y! (1),(n+ 1 - II)v
the last series converges forn > -y + a + - 1 and is evidently a
of
decreasing function n, and since 1/(n - r
y ) - 0 as n - oo, we have
lim F(a, (, y + n; 1) = 1. We thereforeobtain the final result
n-> oo
I
(70) F(a, 3, y; 1) = (,y-a)(-y 3)'aey-a- > 0

From this formula we shall now derive the relation

(71) JIta1
t-( - t),s-ldt = F (a~f)'
T~)()(()0,0)~j)0w(a) > O. 13 > O.1

where the integral is known as Euler's integral of the firstkind15

Multiplying the binomial expansion, convergent for I t I < 1,
15 J10 (fora and , real and positive).

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
82 T. H. GRONWALL.

(1 - t),- = (_ ltv = 0ft(a)v(l- tv

by ta-1 and integratingfromE to s (O < < s < 1), we obtain,since the

seriesconvergesuniformly

JE tal(1 - t)8-ldt = E (a)V(l - fi (s )

v=O (1),,(a)X(a + v')

-F(a, 1 -, a + 1; s) --F(a, 1-, a + 1; e).

a ~~~~a
Now let e -) 0 and s -- 1; since J(a) > 0, J(3) > 0, the integral tends
towardthe leftmemberof (71), em-> 0 and on account of
T1(a + 1 - a - (1 - 3)) = T (3) > 0,

F(a, 1-,fi a + 1; 1) is convergent,so that Abel's theoremon power

series gives F(a, 1 - f, a + 1; s) ->F(a, 1-,f, a + 1; 1). Applying
obtain
(70), we therefore

,gt'- - t)-ldt 1 1- , +1; a 1) = 1 r(a + 1)()

ar (1)Pr(a + j3)
-
a

_r(a)r (o)
r(a + f)
We may now express the hypergeometric series as a definiteintegral.
Let s have a fixedvalue such that I s < 1, then stt sI forO - t _ 1,
and the binomialexpansion
(1 - st)- = v0 (1)T

convergesuniformlyin t for 0 t_1. AssumingT (a) >0, T(Ty-a) >0,

multiplyingby tal(- t)V-0-l and integratingfrom 0 to 1, we may
integratetermby termon the rightside and obtain
therefore

jt a-4-
l( 0- (ai)
t)va1(1-st)-dt = E00 ()
f
ta+V-1(-_ t) ~aldt
-Y- dt
o v~~~~~~~~~~~~;=O
v O

+ P)r(
(a r y a) 1-0 r)v
r(iX1y + )
?(
by (71), or since r(a + v) = (ac)r(a) by (J2),
1 - a)
r~~~~~~(a)r(,y
(72) J0
tQ-(1 t)Y--1(1
- - st)-dt = rT(o) F(a, f, y; s),

I sRI < 1. T(a) > O.2 w(,y-a) > O.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 83

ofEuler'ssummation
12. Application formulato log r(s + a) and 4I(s+ a).
To secure the single-valuednessof the logarithmswith which we shall
deal in the following,we introducea cut in the complexplane along the
negativereal axis (from0 to - oo) and subject the complexvariable s
that it shall nevercrossthiscut. Then everybranchof
to the restriction
the functionlog s is single-valued,and we shall definelog s as that
particularbranchwhichis real forreal and positivevalues of s. In par-
ticular,we have log (s)n = log s + log (s + 1) + *** + log (s + n-1).
We shall now apply (50) and (51) to the case F(s) = log s, b = 1 and
obtain firsta proofthat (s, n) approaches a limit as n -) x for any s
not on the negativereal axis (that is, a new proofof the convergenceof
Euler's infiniteproduct for F(s), J3 and 3', for these values of s) and
second,Stirling'sasymptoticseriesforlog F(s).
Making F(s) = log s, b = 1, 0 c a < 1 in (50), we find, since
tn, 1
Jlog (s + t)dt= (s +- n) (log (s + n) -1)- s(log s -1) and F'(s) =-
n-1

log (s + a), = E log (s + a + v) = (n + s + a - -) log (s + n)

v=O

- n (s + a l) gs fP(t a) dt;

in particular,we findfors = 1, a = 0 and replacingn by n - 1

n-2 pn-i
log (n-1)! = Elog (1 + v) = (n - 1) logn - n + 1- P1(t) dt.

Adding (s + a) log n to the last equation, and subtractingthe first,we

obtain
log (s + a, n) = log (n - 1)! + (s + a) logn - log (s + a)n

(73 -(n + s + a-2) log( 1 + ) + (s + a-2) log s

+ 1 - f n4 1 () dt +f P(1t- a) dt.

Since P1(t) = - dP2(t) by

dt (41), an integrationby parts gives

fn p,(t -a) dt- P2(a) _P() n2 t )d

a P a t- P b s thn t( gr t-t)2 d
"OP2 a) dt con-
and as I P2 (t- a) cP2(0) by (42), the integral (t-

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
84 T. H. GRONWALL.

verges for every s not on the negative real axis (theor. XVI cor.). Con-
sequently

(74) rimf sP(t

- ) dt = j Pi(t -
a) dt = P2(a) _ P2_(t-a)
__ dt

.L1?t~
-I
im
andimilrly
and similarly lim P1 -dt
1 +(t)
dt =+ pi(t)
exists; furthermore
,1*o I;
1
Is
lim (n
n-00o
+ s+ a )log (1 +-)n = lim
n->oo
(n + s + a-2)-= n s and
substitut ng all this in (73) we find

lim log (s+a, n) = (s+a-2 ) log s-s+ 1 - j' lP-t dt+ f Pl( -ta) dt.

From the existence of the limit to the left we infer at once that
rlim(s + a, n) = F(s + a) exist foreverys not on the cut, and that for
such values of s

(75) log r(s + a) = (s + a- log 6 s + k' +f P +t a) dt,

where we now only have to evaluate the constant k' = 1 - f l+9tdt.

To this purpose we observe that f 1-dt = if P(t)t dt 0 as n - oo

P
on account of the convergenceof 1+ ( dt; makinga = 0, s = nand
s = 2n in (75), we obtain

2 log r(n) = (2n -1) log n -2n + 2k' + 2 f4 < (t)tdt,

rWP1(t)
log r(2n) (2n- )log n-2n + k' + J2+t dt,
and subtracting

k' +2 + Pi
Pd dt - ; t)dt =logPr(nfl2f2Th
Sn + t 2n + t g r(2n)
2Dr(n)n'
'
2(2 + )***(2+ n -1'
so that, as n - o
k' = log (2'r(2)) = log f2ir. This is essentially the
method given in J ? 4; another is the following: making s = + yi,

y > 0 and a = 0, so that for t 0, , I s + t > y and Is + t > + t,

I t
+ 1 2 > y( + t)2-,it follows from (74) that

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 85

fP;
________2

t)+yi + t
(0

y
P2(0) 00

+
2 t)
dt__

Now, by (J6)

r(A + yi)r(A- yi) = sinw(A+ yi) = ewi(Yi)-

2_7r 27re-ry

e e
+ e'Y= 1+
and adding the two equations (75) for a = 0,= A + yi and s= 2-yi,

log 2r- ry -log (1 + e-2'y)

= yi log (2 + yi) - (2 + yi) - yi log (A - yi)- ( -yi)

+ 2k' + P1 (t) Pi (t) dt.

k+ fA dt + dL .i+t.
But log (2 + yi) = log y + log i + log (1 +2yi) = log y + 2 +2yi +
71i 1
and similarlylog (A - yi) = logy- -2 + **; introducing

this in the preceding equation and letting y - co, we find 2k' = log 2r.
Thus (75) finallybecomes16
log r(s + a) = (s + a -) log s - s + log 42w + c(s, a),

(76) co(s,a) = fPt- + )dt.

When a = 0,

.Wo(S'0) = f ds ( dt - Lfv -l(t) dt

v=o
0i> Pit) td
Ss + v + td VO S +
2
v + tdt
and performing the integrationsin the last expression,we obtain Guder-
mann'sformulaforc(s, 0), (J 13).
In (76), we now integratethe expressionfor c(s, a) by parts; the
resultmay be read offimmediatelyfroma comparisonof (50) and (51),
observing that Fk(s) = (- 1) k-l(k - 1) !/Sk. Thus we obtain
16 For a = 0 and s real, this formulais due to Gilbert,Ph. (J22),and fora = 0 and s complex
to Stieltjes, T. J., Sur le developpementde log r(a), Journ.de Math., ser. 4, vol. 5 (1889), pp.
425-444. For 0 c a < 1 and s complex, see Lindeldf,E., Le calcul des residus, Paris, 1905,
pp. 95-97.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
86 T. H. GRONWALL.

((k
2m-1 -
E
1)!Pk~l(a)
(s, a) (-_ 1)[(k-1)1/2 k

+ (- 1)m(2m- 1)! J( (s+ t)2m)dt.

To form an estimate of the remainder, let s = j s I e0, - 7r < 0 < 7r;

then
s + t 12= (I s j cos 0 + t)2 + I s 12sin2 6

= (js I + t)2 cos2 ?+ (I s -t)2 sin2 2 (I sj + t)2 cos22

and since I P2m(t- a) I c P2m(0), we obtain

( 1)m(2m - 1)! f(P2m+ t)dt < (2m

<) !S2m(O)J S + t)2m

(2m - 2) !P2m(0)
COS2m6/2. s| 2m-1

Joining the term with k = 2m - 1 in (77) to the remainder, we may

thereforewrite
2m-2
E (k - '1 !Pkil(a) + (2m
( - h
- 2) ! P2,m(0)
(S, a) = (-_ )[k(k-1)1/2 kl
mm6/2m- I

jhI <2.
In the particularcases wherea = 0 or a = 2, Pk+l(a) = 0 for k + 1 odd,
and therefore,replacing k by 2k - 1, (78) becomes
n-1 (2k -_2) !P2k(a) (2m - 2) !P2m(O) h
ws, a)
(

(79) = ( 2 k+1
82+ +
+ cO 2m0~/2< | '3 2m-1
1

1hI <2, a = 0or 2.

When s is real and positive, we may use the much more accurate remainder
terms (56) and (60). In fact, for a = 0 and ao = - 1,
(2m - 1)!
aF 2m)(S + t) = (S + t)2m

is positive and decreases as t increases from 0 to oo, and

a(F(2m)(s + v + t) + F(2m)(s + v+ 1 - t))
(2m - 1)! (2m - 1)!
(s + v + t)2m +(S + v + 1 - t) 2m
is positive and decreases as t increases from 0 to 1 (since the derivative

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 87

in respect to t is negative). Similarly we see that for a = 2 and a = -1,

the conditions for the applicability of (60) are satisfied. Taking the
values of P2k(0) and P2k(2) from (48), we thereforeobtain17
(,0- ( 1) B,. 1 (- 1)m-B 1
(80) - k=I (2k - 1) .2k S2k-1+ (2m -1) .2m (s + h)2m-1I
0 < h < 2, S > 0.
co(s, 12~h<, >O

k=i 1 22k-1) (2k- 1) -2k S2k-1

(81) + ~ ~ ( )~~ ~ <,sO
X
+(122m-1) (2m -
1< 2
) l2m (s +
>.
h)2m
The infiniteseries
E (-_ 1 ) [k(k-1)] /2 ( ~ a

k=1 8

the first2m -2 terms of which occur in (78), does not converge for any
value of s. First suppose a * 0 and a 2, then the ratio of the terms *

for which k = 2m + 2 and k = 2m may be written,using the expressions

(42) for P2m+3(a) and P2m+l(a):
0
sin 2nra
_ m(2m + 1) I- nm+2
27r2s2 00sin 2nra'
~ 2m
00 sin 2nria
and since Z, converges uniformly in respect to m for m _ 1,
sin 2nra
we have Z i2m -sin 2ra as m o
0, so that the second factor in
our ratio approaches unity as m - 00. The first factor, however, ap-
proaches infinity,and consequently the series cannot converge. When
a = 0 or a = 2, a similar argument applies to the ratio of the terms with
k = 2m+3andk = 2m+ 1.
Proceeding to the logarithmic derivative of the Gamma function,
we find,making F(s) = sb = 1 0 ca < 1 in (50), that

+
ZE 1 = log (s+n) -logs+('-a) 2 s +n, \ j (Pit)
&,=s + a + v (s +t) 2 ,

and since the last integral converges as n -< 00, and ,6(s + a) =
\
ES
ni-1 1
1 -og Z a by (J 19'), it is seen that
nm- jogn V-O + a+ vJ
17 (80) is due to Stirling (J21) and (81) to Gauss (J"), both without remainderterms. The

formof the remainderterm given in (80) is due to Sonin.9 For a = 0, the remaindertermin
(79) was given by StieltjesW(who in this particularcase obtains I h I 1).

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
88 T. H. GRONWALL.

VI(s + a) = log s - w*(s, a)

(82) P1
1a) = a) + Pi-(t- a)dt

Treatingthe integralin the same manneras the corresponding

one for
w(s,a), we readilyobtain
1 2m
w*(s, a) = (1- a) - + E (- 1) [(k-I Xkl-2)] /2 (k - 1) !Pk(a)
S kk2 S
(83) - a) dt
m0Pn(t
+ (- 1)m(2m)!j (S + t)2m+1

co*(s,a) =
C))*(,
a)a)-+
- 1 z
2m-1
[(l2)/(
_1[k]k212( - 1)!PAk(a)
(84)
(2m - 1)!P2m(O)h
+ cos2mea/2.|IsI2m j|h i < 2,
and fors real and positive

0) 1~~~~(- 1) klBk 1 (-1) mlBm 1

(85) 2s+ k 2k h)2m+ 2m (s +
0 < h < 22
cow(o 2) En(1 - 2 )k BA; 1

(86)
1 0 -lMBm <h<
+t-22- 2m (s + h)2
furthermore, the same argumentas beforeshows the divergenceof the
infiniteseriesof whichthe initialtermsoccurin (84).
A divergentseries
ao + a,+a2+ + a,+ '
S S2 Sf

in whichthe sum ofthe n + 1 firsttermsis Sn(S), is said to be an asymp-

totic expansionof a functionf(s) fora givenrangeof values of 0 = arc s,
if the expansion
Rn(s) = Sn(f(s) - Sn(s))

satisfies the condition lim Rn(s) for n fixed, even though = 0

1!-4 00

lim I Rn(s) = X when s is fixed. In this sense, (78) and (79) are
n7-wco
asymptoticexpansionsof (s, a), and (84) ofc* (s, a), for- + e c 6 _ -
wheree is as small as we please, i. e., fors insideof any infinitesectornot
containingthe negativereal axis.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 89

It is clear fromthe definitionthat a functionf(s) cannot have two

differentasymptotic expansions ao + a,+ *. and bo + -i + *. both
S ~~~~~S
valid in the same sector(or even fora singlecommonvalue of 0 = arc s),
sincethenwe shouldhave

im en (ao - bo + - + + a-n )

whenceao = bo, a, = bi, *- . This remarkwill be used in the following

to identifyvariousexpressionsforthe coefficientsin (78) and (84).
13. Integralsfor log r(s) and 41t(s)of Cauchy and Gauss. We begin
by showingthat
r~e-at -t _e-bt b
(87) dt = log- forT (a) >0,O J(b) > 0,
t a

where log aa is the principallogarithm. By directintegrationwe find,

forx > 0,
ie.(X+io) tdt 1 (1 - e-(x+jy)T),
Jo ~~x+ iy
e(x+iy)T exT T
and since x = e -e
E > O as T co, it followsthat
+iy

(88) *je- (+iy)tdt = X + i'

the integral being uniformlyconvergentfor x _ e. Let 9Z(a) el

9Z(b) ie and write x + iy = a(l -u) + bu where 0 c u c 1, then
x El and fromthe uniformconvergenceof our integralit follows,by
theor.XX, that
a: (b-a)du_ l
adu (b - a)e-(a+(b-a)u)tdt
a + (b -a)u
J dtJ (b - a)e (a+(b-a) u) tdu

SXe-at - e-bt

~~t dt. =

But the firstsingleintegralis foundby directintegrationto equal one of

the values of log , and since the real part of the denominatorin the
integrandis not less than e forT(a) :, E1 R(b) 1 E, it followsthat the

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
90 T. H. GRONWALL.

integralis continuousforsuch values of a and b. Since the integralis

a
obviouslyreal forreal values of a and b, it followsthat the value of log
in questionmustbe theprincipalone,and (87) is proved. In theequation
(J 19"), making a = 1,
n1\
1(s) = im (log (n + 1) - E +
we now assume T(s) > 0; then we have, by (88) and (87), forv = 0, 1,
1 bo+o ) _et
=
t e-(n+l)t

+ e-d tt log (n + 1) = d10

t dt,
and consequently
rX
-t - e (n+l) t n-1\

(S) = lim-J e -e - e-(+)t) dt.

nea:) t v=O

Upon summationof the finitegeometricseriesto the right,this becomes

(s) = lim j( e t -
e-t
(1e- nt)dt.

Now let S(s) - e and I s I < A, where e < 1 is as small and A as large
as we please; usingthe powerseriesfore-t and e-st,we see that the first
factorundertheintegralsignis expansiblein a powerseriesin t,convergent
forI t ] < r, wherer is independentof s, and consequently
e-t e-st
Te - 1 -et < M1 = Ml(E, A)

for0 = t < r. On the otherhand, we have fort - r, since e < 1

e-tet e~~~~~~t
e-st 1e-t et e-et {1 + 1 het -e
t 1-et = r 1 e-
1e- r e
so that finally,fort _ 0 and SR(s) _ E, I s j = A
et e-(t

t 1- e-t < M(,E, A).-e-ei.

Consequentlythe integral

(t 1 -e-t~d
in respectto s forSR(s)
convergesuniformly E, I s j = A, and

J;
(et t
-1 r-t) e-ntdt < M(e) A)
1 _-eJ-tE,~ee~~d
-e-(,+n) tdt = M(e
e A)) 0

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 91

as n - oo, so that finally

(89) ,6r(s)= f (Se - 1 eet)dt, T (S) > 0.

We may transformthis integral by writing t = log (1 + u , A = es-1,

00e-t 1C e-t 00 du
e-t
iI'(s) = lim t dt = lim t t u(l+u)8)

and since
e- 1
O
0 <Jt
e-t
dt <J-dt log
0lg 6 >as 6 -0,

thismay be written
,6(s) = lim (J; dt - fu(l + )
or finally,replacingu by t in the last integral,

(90) 4I(s) = (6-t - (1 + t)-8) t T(s) > 0.

Again,replacinge-t by t in (89), we obtain

(91) 4I(s) Jbl(- t-1 S(s)>0.

For s = 1, we have ,6(s) = -C, whence by (89), (90), (91)

C= t~UIe 't1)e-tdt

- (1--t+ e-
logt~dt,
and consequently dt
p000-t.._...l-8t
s + c=dt
t)-(1 + >0
(1+ 41+Ct
~ J 1- dt t)l)o t R()

-rl 1e- te-

Since (89) convergesuniformlyforS(s) , E, s l A, the corresponding

integralforlog F(s) may be obtained by integrationaccordingto theor.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
92 T. H. GRONWALL.

XX, but it is equally simpleto startfrom(J 3") fora = 1, or ratherthe

equivalentformula
log r(s) = lim [(s -1) log (n + 1) -E log + ]
n--> y= 1 V

For S(s) > 0, (87) gives

log
S+
v-
=-1 W e-t - e-(B+-')
t
t
dt, log (n + 1) J We-t-e-(n+l)
-
t
dt,
So t

lim FF(s
log F(s) = n->oo - 1) (e-t - e-(n+l) t) (e-vt e-(8+vl) t) t
o v=l

= lim J(s - 1)e-t + eet e-] (1-ent) dt

WhenST(s) i e (e < 1), I s i A, it is seen exactlyas beforethat

- 1e-t +?;_je-i7] t< M(e, A)e-et

fort 0, whereM dependsonlyon E and A but not on s, and that

log
F~s~ = e-s ~e-tdt
(94) log r(S) = ([(s - 1)e-t + 1 - eFtJt ST(s) > 0

the integralbeing uniformlyconvergentforTY(s): e, s A. In the

same manneras (90) was derivedfrom(89), we also find

j,r00
(1 + t)8 -(1 + t)-1 dt
(95) log F(s) = [(s- 1)et + log (1 + t) ]t -

14. IntegralsofRaabe and Binet. The integralof Raabe'9

1
(96) flog F(s + a)da = s log s - s + log -'I7, (s not on cut),

resultsmostreadilyfrom(76), whichgivesupon in egration

flog F(s + a)da = s log s-s + log 2 + ,da -(s + t)

Here the double integralvanishes,beingthe limitas n oco of

18 Eq. (89) and (91) are due to Gauss (Jil), (90) to Dirichlet, P. G. L., Sur les integrales

Euleriennes,Journ.f. Math., vol. 15 (1836), pp. 258-263 = Werke,vol. 1, pp. 271-278. The last
of (93) was given by Legendre,A. M., Exercices de calcul int6gral,vol. 2 (1817), p. 45, and (94)
and (95) by Cauchy, A. L., Exercices d'analyse et de physique mathematique, vol. 2 (Paris,
1841), p. 380.
19Raabe, J. L., AngendherteBestimmungder Factorenfolge,etc., Journ.f. Math., vol. 25
(1843), pp. 146-159.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 93

fdaf P1(t a)dt + tfPI (t -a)da,

and by (41), fPi(t - a)da = P2(t - 1) - P2(t) = 0. Another proof

may be derivedby means of (J 3'):

0~~~~ 0~~~1
[lgn-itm')(~
J=ourn
log P(s+a)da = lim J
o) [lo logr(n)+(s+a)
17(n)+ (s + log
log n n-Efloglog(s(s+a+)
+ a~da]
da
-

= lim logJ(n) + (s + 2) log n - log (s + a)da

=lim[logF(n) + (s + log n- (s +n)log (s +n)

2)
nf-- 0

+ s log s + n]
= s log s - s + lim [log r(n)- (n - 2) log n + n],
n-- 00

and the last limitis log -.2;7,as is seen fromStirling'sformula. We shall

now derivean expressionforcw(s,a) similarto the integralforlog r(s) in
(94). From (76) we obtain,since Pi(x) has the periodunity,

w(s, a) = lim fi7i(u- du = lim n P(uu

S
da)u
V
w O
n----) + U n-w v=O u +
and from(88), forTJ(s)> 0,
1 0
= e-(+u+v)tdt
+ u +1
so that
n-1I1 aoo
w(s, a) = lim
n-->
Ef du
v=OO?
Pi(u - a)e-(8+u+v)tdt

By theoremXX, the orderofintegrationin the repeatedintegralsmay be

inverted,and sincePi(x) = - -X for-1 < x < O but Pi(x) = - x
for0 < x < 1, we have
r1
fPi(u - a)e-utdu =
ra
)(a - u)e-tudu + J(a +
I
1 -u)e-utdu

t[a 2- -t)e-t) +eat]

as is readilyseen upon integrationby parts,So that

no-I
r0oor/d
w(s, a) = limZJ [(a -2-)(1 -e-t) + e-at]e-(s+v) t
00[ 1 e-a l t dt
= lim ~a- T? ee]1 -
ent)e- t
n0*o.41 e0

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
94 T. H. GRONWALL.

Exactly as in the proofof (89), we see that fort _ 0,

[a-i~ ~
eat
2 t+1-et] tt
and that consequently,for R(s) > 0, the limitingprocess may be per-
formedunderthe integralsign,so that finally20

(97)
Writing
co(s, a) a (a -et
T
9(s) > 0.

(98) f(t, a) = (a - ' 1

this functionis holomorphicin the neighborhoodof t = 0, and conse-

aIf (t, a) is bounded in this neighborhood. Using Leibniz's
quently k
at
formulaforobtainingthis derivative,each termwill contain a negative
power of t as factor,and consequently-fd(t,a) - ? as t - oo. We
have
therefore
a kf(tja)<Mfr O
< Mk for t _ O
atk'

where Mk is independent of t. Now integratingby parts in (97), we obtain

e-8t +t=
n
2 (t, a)
a= ak-1f(t, a) l a2m-2f
(s, a) -1 +L 1 at 8kJ 62m-2J at2m-2 e

and since 9T(s) = s I cos 0 > 0 and the partial derivativesof f(t, a) are
bounded,each termin the sum vanishesat the upperlimitand
--to (t, a)
a92m-2f _________

.10 at2m-2--e-8tdt < M2n-2e-I8IcoSe tdt = lcoso

we see that
E [dk-lf(t, a) ] 1 h |h
atk~-1_ j < 1
M2m-2
(99-a)
(99) co(s,a)=E II i1m-lCoOS

Consequently(99) is an asymptoticexpansionof co(s,a) for9(s) > 0, and

mustthereforecoincidewith (78), whence

1)Ek(k)Ii3(k - 1)!Pk+l(a), k = 1, 2,
Latk-i
[ak-lf(t, a) ]
=
- (-

Using this in the expansionoff(t, a) by Taylor's theorem,we obtain a

formulawhichis essentiallyidenticalwith (49).
20 For a = 0, this expressionis due to Binet (J24).

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 95
It shouldbe notedthat thismethodof derivingtheasymptoticexpan-
sion of co(s,a) only shows the expansion to be valid in the half-plane
TJ(s)> 0, whereasthemethodof ? 12 is valid whens is noton the negative
real axis.
15. The generalizedSchaarintegral. To obtain this integral,we shall
replaceeach termin the expression(J 24)
d_(s) _co 1
ds v=0 (s + p)2

by an integralover a rationalfunction(instead of the exponentialused

in ?13). Lets = x+yiandx > O;then
rX.sdt rXis 1 1 X
Jot2+
s2 Jo2 at- y+ xi t + y- xi~
i fe t-y-X% ( t+y)+x) x dt
2 O - dt
(t y) 2 + X2 (t + y) 2 +2
-
y) 2
xY
(t arctan - i arctan +x Y ]
= 2 log ++X2 -i

7r

=2
Now let au = + ji and t > 0, then

-r( _ S = ,fG(
( 2 ,-2 +-2
whence
1 2 rn t~dt ~ ?s ,~~)>0
> 0
8? + a= i;J (t2 + s2)(t2 + u2) ' T() 2 (

and thisis also trueforo- = 0, since the integralthenreducesto the pre-

cedingone. Differentiating in respectto x, we obtain
1 2C M2stdt
(100) s> >
___

(100) R(s) T 0>0>

(s + u) 2=7J (t2 + S2)2(t2 ? u2) 0, (o) 0,

and this is also trueforo- = 0. In factwe have 21

(101) | t2+ s21 |(t2 +I|s12)cos 0 fort O. s = Is eo, -2 < 2<

and fors=s a Ieli, -2 + e

I e, a=
0, -2- , A _js
I 6,
j jI _ 6, the integrandin (100) is consequentlyless than
21 Since I t2 + S2 12 = I t2 + I s 12e2' 12 = (t2 + I s 12 cos 20)2 + (I s 12 sin 20)2 = (t2 + Is 12)2 cos2 0
+ (t2 - I s 12)2 sinf2 0.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
96 T. H. GRONWALL.

2At2
sin3 E(t2 + 62)3

so that the integral (100) is uniformlyconvergent in respect to s in the

region considered, and theorem XXI is applicable; a similar proof applies
to the case o- = 0. In order to obtain a formula applying to all values of
s except those on the negative real axis, let us assume - r+2e=0 ir-2e;
7r
then we may determine an angle a such that- + e =- a -=
=2
- E and

+e-
-2 - a2-e. In (100) now replace s by se-i and o- by
where v is a positive integer or zero, and formthe sum from v
Pe -fi, = 0
to v = n - 1; we obtain
_1 1 2 f n 2se3ait2 n-1d
v0 (s + v)2 T Jo (t2e2i + s2)2 '= t2e2ai +42 d?
But by (101), t2e2f? ,2 _ (t2 +
v 2) cos a v2 sin E, and for s
It2e2ai + S2 =t2 + S2e-2ai | (t2 + I S 2) COS (0O)

12(t2 + 62)sinE
consequently
fW 2se3ait2 E 1 2--s I m1 . r t2dt
|J (t2e24-ais2)2 t2e2ai+ ,2 = sin3 E v v2 J (t2 + 62)2

1
as n o, since the last integral converges and I is a convergent series;
therefore
, 1 2' 2se ait2 00 1
1t
=
(8) = d(S + )2 = J(t2
2ai
+ s2)2 1:t2e2ai + p2dt

But fromthe partial fractionexpansion of the cotangent it follows (see ? 8)

that
0 2t 2r 1
V= t2 + p2 = e22re-1 + + W,

and changing t into tea, we find

- 2 0 seaidt2 se2aitdt + se2ait dt
41)1(s) -
J
I
(t2e2ai + 2)
~~~~+
21
(t2e2ai + s2) 2
+ 4 Jo (t2e
2ai
+ S2)2 e2-teoai_ 1

The firstof these integrals is found, by (100) for o- = 0, to equal 1/2S2,

the value 1/sof the second is obtained by direct integration,and the third
converges uniformlyfor A | s 6
8, as is seen from (100) and the fact
that

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 97

-.
-Efor 0 < t< 1
(103) <
e2,,teati 1 1 frt> 12
t2 sin3 efor t >

Hence the application of theorem XXI to the third integral gives

1 1 la
__ e2___i tdt
s + - xJ t2e2ai + 1 '
S' e27ra

and integration in respect to x,

41 (S) = log s 1 + c(y)

- -+cy)-2 - 2a t2 + tai
d -1
eai S2

where we have denoted the integration constant by c(y). Now in.

s = x + yi let x -* oo; then 0 = arctan y/x ->0, so that for any a where
a 7r/2- c and a has the same sign as y, we have Io - a I 7/2- e
for x sufficientlylarge. Then, by (102) and (103)
00 e2ai tdt 1 (l dt _dt
IJ
t2-2ai + 82 e-2rtea <
S 12 sinE J0sin +J t2sinr3 )J

as x -* oo, and since also 41(s) - log s -> 0 by (84), it follows that c(y) = 0
for all values of y. Consequently
1 aW e2i
e ~ tdt
41)(s) = log s- -i -
2s 2f t2
e~ai+ S2 e27rtea1i
td_

(104)

0 < t7, 0 < 0 - <a K

and upon integration by parts

1 1 C0
~eai(S2 - t2e2ai)a
(105) 4t(s) = log s - ?-+ 1 ( e2ai -+
2)2
log (1 - e-2,rti)dt.

Since I a I < w/2, we have I e-27rIea = e-27rtcosa < 1 for t > 0, and con-
sequently

I
00 x
e-27rte"
e e-27rt C08a

log (1 - e27teat) = V e o
= =
(106)
= - log (1 - e-27r
tcos
a);

the last two expressions show that log (1 - e-27rteai) becomes infinite as
log t for t-* 0 and vanishes as e-2rtcosa when t -) oo. Together with
Since Iertea- 1 _ 1 e2ftcoa 1 _e2fftsin e-1 _ either of 27rtsin e or
(27rtsin E)3
3!

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
98 T. H. GRONWALL.

(102), thisshowsthat the integralin (105) convergesuniformly in respect

to s in the regionpreviouslyconsidered,and we may therefore integrate
(105) in respectto x, obtaining
1 C0 se
J
a

log r(s) = (s - ) log s -s + c(y) - t2e2ai + s2 log (1,-ae-27')dt

From (102) and the precedingremarkson the logarithmit followsthat

the last integral approaches zero as x - co, and from (76) and (78)
log r(s) - (s - 1) + s -: log -12ir,so that c(y) equals the latter constant
and we finallyobtain
log r(s) = (s - 2) log s - s + log -27r
(107) 1 00 seai
-
t2e2e-
1 + 2 log (1 te)dt,
e-27m

< -T) a
Ile < a2 I Il <2

where,by (76), the last integralequals W(s, 0).23

Makings = 1 and a = 0 in (104) and (105), we obtaintwoexpressions24
forEuler's constant
C
=
+ 2Jt2 +1 tdt__
(108) 1 2
- 1 _
,lr -+ 12 log (1 --2rt)dt.
2 0(t2+

To obtainthe asymptoticexpansionof w(s,0) from(107), we expand the

firstfactorunderthe integralsignin negativepowersof s:
seai m-i (- 1) k-le(4k-3) ait2 k-2 (- 1 e(4m-3)ait2m-2

t2e2ai + 2 k s2 k-i S2m-3 (t2e2al + S2)

and obtain
! log (1 -e
W(s,
7t(.1
0) )k
=
-
2 k1
.
J ec4k-3)ait2k-2
(+Rm (1 ta )dt + RmY

(-1 )m 1fo e(4m-3)ait2m-2

Rm= 2m3 Jr t2e2ai + S2 log (1 - )dt

From (102) and (106) it followsthat

1 1 00 t2m-2 2ra

iS
3 T S 12 log(1e2tcosa)dt
23 For a = 0, T(s) > 0, (107) is due to Schaar, Memoire sur les integraleseul'riennes et sur

la convergenced'une certaine classe de series,Mem. Ac. Belgique, vol. 22 (1848), pp. 3-25.
24 Poisson, S. D., Memoire sur les int6gralesddfinies,Journ. de l'cole Polytechn., cahier
18 (1813), p. 305.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 99

or introducing t cos a as integration variable

1C
R -=-:=1
s I 12m-1Cos2m-la cos (O - a) * J -tn1lg(1-e')t
log (1
- t2m-2 - e2)d
Hence I s1s -2 I Rm I -* 0 as I s I < so and the expansionabove is asymp-
totic, so that it must coincide with (79) for a = 0. Expressing P2k(0)
in terms of Bk and identifyingcoefficients,we thereforeobtain

(109) Bk
2k- 1) .2k = -
! f e(4
k-3)att2k-2 log (1 - e2wte )dt;

making k = m, a = 0 and introducing in the last expression for I Rm |

we find
(110) R.mI Bm 1
I(2rn-
l).2m Cos2m-1a cos (0a-a) I8 12m1
By choosing a for a given 0 so as to make the second denominator a
maximum, this form of the remainder will frequently be much smaller
than the one in (79).25
16. Sonin's form of the remainder in Stirling's formula. In the re-
mainder term in w(s, 0) obtained at the end of the preceding paragraph,
we now assume s real and positive and make a = 0, whence
1 - t2M-2
__ 1
82m-3 7J t2 + 82 lg1 e-2t
or writingt = su,
- 1 0_ _ _ du
(_ 1)M-R, = u2m-2.og1
-leoru *1 + u2

For s > 0, u > 0, 0 < e-2ru < 1, and

1 X
lo1 e2u = E e-_2nsu

X0 e-2 n 7 (s+X) _u
________
E . e2n rku;

assuming X > 0, e2nnAu is positive and increases with n, whence

1 00 e-2n r(s+A) u

log 1-e-2 ; > e2rku . E

= e2vAulog 1
1 e-2 r(s+A)

Now assume X so large that for every u O 0

25For furtherdetails on this point, compare Lindelof,Calcul des rdsidus,pp. 98-102.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
100 T. H. GRONWALL.

(111) e u 1 + U2
then the last expression for Rmgives
1L 2m2l _ > (- 1)m-1R,

1 C 2m-21 _ _ _ _ _
> -
u og1 - -2(r(,+A)udu

and introducingthe new integration variables su in the firstand (s + X)u

in the second of these integrals

S-2-l*
1
1 1,(t2m- log l 11
-e_7--dt > (- 1)m- Rm (
-2

> (s + X,)2m-l J
7r,( t2m-2log dt.
_e27rt
Making k = m and a = 0 in (109), it follows that
_Bm_ 1 M- Bm 1
(2m-1)-2m * s)m-1 (2m - 1).2m (s + X)2m-1
To secure as close a limitation of Rm as poss ble, we have to find the
smallest value of X which will satisfy (111). It is clear that for this
smallest value of X, the equality sign in (111) must obtain for one or more
values of u besides u = 0, and consequently the smallest admissible
value of X equals the maximum of

X = - *_u log (1 + u2)

2wr
This maximum must satisfy the equation

d(27 =-)
du U2log (1 + U2) + 1+u2
+ u2 = ?

let
f(u2)-= 2d(2wX) 1 2+u2
1 + U 2-log (1 + u2),
du -

so that
1t(U2) =1_

and consequently f(u2) increases for u < 1 and decreases for u > 1. But
for u small, f(u2) = u2 + *-- > 0, so that finallyf(u2) = 0 can have
only one root, for which u > 1. The numerical calculation of this root
gives 1 + u2 = 4.92154 ... and the correspondingvalue of X is 0.12808.**,
so that we may now replace (80) by the following:

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 101

'r-' (
- 1 ) k1Bk 1 (-1) mlBm, 1
(112) - 1) .2k s2k1+(2m - 1) .2m (s + h)
wo(sO?)= Z(2k
0 < h < 0.12808--, s > 0.

series. The function r(1

17. Kummer's and Lerch's trigonometric - s)
is holomorphic for 0 < s < 1, and becomes infiniteas log 1/s when s -> 0
F(s)
and as og (1-s) as s - 1, so that log r(1 - s) ds exists. Conse-
quently, this functionis expansible in a Fourier series (35) uniformlycon-
vergent for e c s _ 1 - E, where the coefficientsare given by (34):

an = 2 1oglo10gF(1 -r() s) - cos 2nrsds, = 2 - 2nFsds,sin

= . bn 1og
rjagP(
- s)si2nd,

and replacing s by 1 - s in the expression for any it is seen that

1
s) s) cos 2nsds=
a-
=2jg log -an

or an = 0. To calculate bn,we observethat,by (J 3'),

r(s) -
lim m2s-l (1- ) (2 * s)* (m -
s)
F (1 -s) m-*oo ~ +S ..(M +S
so that
bn = lim log m*F (2s - 1) sin 2n7rsds
+ log (v - s) sin2nwsds
m1 lo
+ E
v=Oo
log (v + s) sin 2nrsds
Integratingby parts,

J(2s - 1) sin2nwsds=
and taking - s and v + s respectivelyas new integration
variables

E log (v - s) sin 2nwsds = - f IIog s sin

2nwsds

- -j flog s sin 2n7rsds,

E
v=O o
log (v + s) sin2nrsds
log E log slog s sin 2n7rsds
M-1
fv-F

= log s sin 2nrsds,

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
102 T. H. GRONWALL.

whence
1b= lim [ _ log m -
2 log s sin 2nlrsdsl,

and integrating
the last integralby parts

b= lim [Js 1-cos 2nirsds - log ml

1 r1/21 cos 4mn7rs
= lim [Jo s ds-log mn + logn
so that finally,replacingmnby m in the limit,

2bn-=n1fir (log n + K),

where
K = limIl
-r1/21 cos 4mirs1
ds -log
-

m.
L
m-> O S

To calculate K, we shall replace the integralby one whichmay be eval-

uated in finiteterms. Observingthat
1 7r
S sin irs

is holomorphicfor0 _ s c 2, we may write

'1/2 1- cos 4m7rs r1/2 1 - cos 4mr
ds

f
S J sin 7rs
11/2 {1 2
= fJ/2 (p(s)ds p(s) cos 4mrsds.
Now
1/2 F ~ s=1/2
, /p(s)d='1-_-- = log + log2 = log 27r -3 log 2,
L tan s ]8?0

and integrating
by parts
J1/2 s 4 =11i2
(s) cos4mrsds =
=

------
-p(s) sin 4mrs
- 1 1for/2 9'(s) sin 4mrsds

-40 as m -> oo,

sinceso(s)and sp'(s)are holomorphic,

and therefore
bounded,in theinterval
of integration. Consequently

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 103

K =lim 'Ir f 1 -cos 4m~d

rs -log m
]+ log 27 g

and sincewe have

Cos 4mirs= 2
11-cosm1r- I
m sin (2v -1)7rs,
sin 7rs V=1

the integralis readilyevaluated, giving

K = im E 2v - 1-log m + log 27r- 3 log 2.

of Euler's constant,
By the definition
4m 2 2m1

E--2 log 4m -4 2C, - log 2m C

as m -> and by subtraction,

oo,
2m 2
~ _1-log m -3 log22-
_2v C,
so that finally
K = C + log 2r,
and consequently
r(s) log n + C + log 2s
-

P (1 -s) 1nirns O s1
But by (40) we have, for0 < s < 1

+nlg 2sin 2nrs = (C + log 27r)( s),

and furthermore,
by (J 6),

lglog(j s) = 2lglog(s)r(1 - s) = log r(S) + 2

2 logsin 7 71S

whencewe obtain Kummer'sseries26

log r(s) + I logsin + (C + log 27r)(s- = flog sin 2nws,

0 <S < 1,
convergentfore
the seriesbeinguniformly - s -- 1 - e.
termby termof the series (113) gives a divergent
The differentiation
26Kummer, E. E., Beitrag zur Theorie der Function r(x), Journ.f. Math., vol. 35 (1847),
pp. 1-4.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
104 T. H. GRONWALL.

expansion of 4'(s) is howeverfurnishedby the

series; a trigonometric
followingtheorem:27
Let
00
f(x) = 2nwx
sin 2r
,n

=n
be convergentfor0 < x < 1; thenthe derivativeoff(x) is givenin this
intervalby
sin irX 0
f(x) -- 1rn=O = (cn - cn+1) sin (2n + 1)irx,
__

where co = 0, provided that the last series convergesuniformlyfor

e c x c 1- E, whereE is as small as we please.
Writing
Z (cn -
00

g(x) =
n0O
cn+1)sin (2n + 1)7rx
N
= Z (cn -
n=O
c,,1) sin (2n + 1)rx + RN,
N 1
NV+

=E cn sin (2n + 1)rx - ECn sin (2n - 1)rx + RN

n=1 n=1

N
= Cn * 2 sin rxcos 2nrx - CN+1 sin (2N + 1) rx + RN,
n=I

where,on account of the uniformconvergence,I RN I may be made as

small as we please for Ec x 1 - E by taking N sufficiently
large,
we have
N
1r sin (2N + 1) rx rRN
s-1n g(x) = ECn 2r cos 2n7rx- 7rCN+1 sin WX sin
i x
irx'
sin wx n1snr
betweenthe limitsE and x
and integrating
oX N

. g(x)dx = Z (sin 2n7rx- sin 2nwre)

e sin Wx
n
n=1 n
( r sin (2N + 1)drx + rRN
1
Cv~lJe
~CN,+ + d)x +
JesinRNdx.
rx
sin rx
Integratingby parts,
r sin (2N + 1) rx N+1 cos (2N + 1)Er-
CN+1I
sinwrdx 2N + 1 sin wr-
cos (2N + 1)7x _ rXi- cos (2N + 1)l7rxcos lrX 1
sin wx sin2i rx
Lerch, M., Sur la diff6rentiation
27 Ann. Ec. Norm.,
d'une classe de series trigonometriques,
ser. 3, vol. 12 (1895), pp. 351-361.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 105

and here the expressionin brackets is evidently bounded forec x c 1- e

and all N. so that our integralvanishes as N -> ox,since then -CN+1 -4028.
2N+ 1
Furthermore,

IrR--d| < RN I dx ->O as N -4 oo

5111 rx sin irE

on account of the uniformconvergence of the series g(x), and we therefore

have, letting N -4 oo,

F(sin g(x)dx
Sin1flrXn
= Z n (sin 2n7rx- sin 2n7rE)= f(x) -f()

whence our proposition follows by differentiation.

In the special case cn = log n/-r,

g(x) =-ZE log n+ * sin (2n + 1)urx,

7 n=1 n+
and since log n+-is positive and decreases as n increases, we have
n
by (38)
EZlog
log + sin (2v + 1)7rx| 1- lrE *logfl+lfl
sin
so that the series g(x) converges uniformlyfor c x 1 - e. We may
thereforeapply Lerch's theorem to (113) and obtain Lerch's series (1. c.)

,p(s) sin irs + 2

2
cos 7rs+ (C + log 2r) sin irs
(114)
=Z *sin (2n + 1) rs,
log 0 < s < 1,
n+1 n=1

the series being uniformlyconvergent for e c s _ 1 -e. It should be

noted that this series is not a Fourier series according to the definition
in ? 7.
CHAPTER III.

The Gamma functiondefinedas a definiteintegral.

18. The Euler integralsof the firstand second kind and theirelementary
properties. In the preceding chapter, the various integrals connected
28This is an immediateconsequence of the convergenceof the series forf(x) in (0, 1) and the

followingtheorem,due to Cantor: When Y, (an cos nx t bn sin nx) converges in an interval

0
(a, b), then an -4 0, b -4 0 as n -4 oo. We omit the proof,since we shall apply Lerch's theorem
only to the case c, = log n/ir,where it is evident that cn+1/(2n + 1) -40 as n-> oo.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
106 T. H. GRONWALL.

with the Gamma functionwere derived fromthe definitionof r(s) as

Euler's infiniteproduct. As we shall now proceed to show, the same
resultsmay also be obtainedby definingr(s) by meansofEuler's integral
of the second kind (61)

(115) r(s) e-jtt8-ldt, 9R(s) > 0.

As shown in ? 10, the integralconvergesuniformlyfor e c 9R(s)_ A,

wheree is as small and A as largeas we please, but does not convergefor
SR(s) < 0. Since therefores mustbe restricted to thehalf-plane9R(s)> 0
throughout thepresentchapter,certainformulas(forinstance,the asymp-
totic expansionsin ? 12, whichwere there proved for all s not on the
negativereal axis) will have a largerregionof validitythan that shown
by the proofsto be givenin the following. Integratingby parts,we find
00 t=00 we r 0
e-tt'dt = e-tt8 + s e-tt-ldt = s fe-t-ldt,
_t=Oe
or
(116) r(s + 1) = sr(s),
whichis one ofthe fundamentalpropertiesof the Gamma function(J4a).
By direct integration,(61) gives r(1) = 1, whence from (116), for a
positiveintegraln, r(n) = (n - 1)! The formulas(62) and (63)

(117) (s) = e-atta-ldt a > 0,

drP(s)
(118) ds e-tt81- log tdt

are derivedfrom(115) exactlyas in ? 10.

Euler's integralof the firstkind is definedas

ta-1( - t)P-'dt, 9Z(a) > 0, 9(3) > O;

.~~~~~~~~~
o
introducingthe new integrationvariable u = 1 whence t = 1+us
we find
(119) j t-1(1 t)-l = J (1 + u)a+ du.

To express this integral in terms of the Gamma functionwe write

a = 1 + u, s = a + j in (117):

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 107

4 e-(l+u)tta+P-ldt =
r(a + i3)
(1 + U)a+9

multiplyby u'-1du and integratefrom0 to oc, whichgives

raw ra4 ua-1

du e-(1")tualta+-ldt
e-(+ )ta-ta9
= r(a + fl)f (1 + u)a+

If we replacetheintegrandto theleftby its absolutevalue (whichamounts

to replacinga and 13by theirreal parts), the new repeatedintegralcon-
verges,since it equals the righthand memberwith9(a) and 9(,B) sub-
stitutedfor a and 1; by theoremXXII, we may thereforereversethe
orderof integrationin the repeatedintegraland obtain

Udt Ue-(l+u)tUa-lta+-1du = r(a + A) UO ua-+ du.

J0 J0 (1 + U)
.~~~~~~~
But by (117)
J:e-utu
a-1du2
~~feutud~d P(a)
(o
and consequently
r(a)f
ox
e-tt-ldt = r(a+ I+
rx~~~~~
ua-1
du)
(1 +u)a+U

or finally,using (115)

(120) + 3) T (1) >

(120) Ax _du = 1(a
J(1+(1 +uu)a?+, P(a)P(3) 9(a) > 0, 0,
whichby (119), is equivalentto (71).
To provethat
(121) r()r( - s) = sin s' <9 (s) < 1,

(whichis J 6 with a restrictionon SR(s)imposed by the methodof the

present chapter), we make a = s, = 1 - s in (117), so that

r(a + 1) = r(m) = 1,
and showthat
(122) 1 Udu = sins, 0 < 9(s) < 1.

The integralto the left convergesuniformlyfor e c (s) _ 1 - by

theoremXIX, since u8-1/(1
u + u) U u -1 for0 < u < 1 and Iu-31/(1 + u)
u-1 for u > 1, and the integrals

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
108 T. H. GRONWALL.

fuC-1du, C-d
both converge. By theoremXXIII, our integral thereforedefinesa
functionof s whichis holomorphicfor0 < SR(s) < 1, and the expression
to the rightin (121) is also holomorphicin the same region. To show
thatbothsidesin (121) are equal throughout thisregion,it is consequently
to prove theirequality foran infinityof values of s having at
sufficient
least one limitingvalue interiorto the region,for instance the values
$ = (2p + 1)/2q,wherep and q are any positiveintegers such that p < q.
Making u = t2q in (119), it is therefore to showthat
sufficient

Re0tip 00 tip
Jol + t2qdt =J2 1+t2qdt 2p+1
q sin 2q

The roots of 1 + t = 0 are t2 = e+l)ti/2

(2 and tn e-(2n+l)tiI2s(n - = 0,
1, *, q - 1), and decomposing into partial fraction,, we have

t2P _q n
+ tt-4
_An An

1 +
1 t~t = =o Vt - tn - tn
where

An A-
1- tn21 n 2p+1.
t2Pq
2qtn2q-1= - qtn2P+l A
A -
2q
Now tntn= 1,
~2n+ 1
tn + tn = 2 cos
-

2q
is real, and
t-tn
si 2n +
1 0
2i -sn 2q
hence

An + An (An + An)t- (Antn + Antn)

t - tn t - tn t2 - (tn + tn)t + 1

1 (An + An)(2t -tn - tn)

2 t2 - (tn+ tn)t+ 1

1
+ (An-An)?(tn -tn)
2 2
2an
dn tn+ tn t

and

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 109

Jty
(t-tn + -n )dt = 2(An+ An) limlog t2 -(tn + t )t + 1
_ _ ~~~1
+ 2(An -An)(tn -tn)
tn -tn

2i
tn + tn t=00
t- 2
X arctan j

L ~ 2i t=-00

= iri(An - n) = 27r9R(iAn)

= 7r (- ie(2n+l) (2p+1) fri/2)

so that finally

4-t2p dt E ie(2n+1)(2p+l)wii2q\

T / e2q(2p+l)7iI2q-1
= - _ je(2p+l)rii2q
2(2p+l)wi/2q
q e~~~~~~

= q
7r~~~2~l7i~
-ie(2p+-) * i/2q e-(2p+1)i/2q)

7r 1
q . 2p+l1
s5n 2q

which proves our proposition.29

7r8. For decomposing
29 From (121) we may obtain the expansionin partial fractionsof /srSin

the integralin two with limits 0, 1 and 1, oo, and changingu into 1/u in the second of these
integrals,we obtain
= O < 9(s) < 1
sin7 JO 1 +U du,
and since
-
1 n- 1)vuv+ 1)n

1 +U v(=+O 1 +

it is seen that
7r n-I
1)V + J~~~~n
I U8+7-l
+ un-sd'
8in r8 V=O s ? + V=0 }' +1 $ + 1) +.u
For e 9?s) 1 - e, we have

o
I un-
+ un- du <J (ue+n-l+un-(l*))ds= 2 -40 as n- oc
,

and consequently

sinirs 8 + ,+ 0 +1-& .-0, +v+

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
110 T. H. GRONWALL.

In (120), we now make a = = s, obtaining

Go u16 1 u
r(s)2 CIUs81 U-1
r(2s) = u( + U)28du+ (1 + U)(sd = 2 1+ U)2sdu
and writingu = (1 - tl/2)/(l + t1'2) in the last integral,

- tOa-ldt= 1 -(
r7(2s) 222s--1 ot-'(l
P(2s)2= 2 2s' r~)2(s)
+ 1)
by (119) and (120). Since (121) gives r(4) = '17r,we obtain Legendre's
formula(J13)
(123) r(s)r(s + 2) = 2 r1P(2s), T(S) > 0.

19. Integralsfor y6(s). The infiniteproductfor r(s). In (118), we

replacelog t by the expression
w e-u _ '-tu
iog t J u du
obtainedfrom(87), and find

P'(s) =
Go r4
fdt e-tt-le
eu - e-tudu = f dt
r
f(t, u)du;

since e-u e-tuas t : 1, we have

If(t, u) -= 4 e-tt-(8)-1 e as t 1),

so that
r0
J'dtf
,s0
f(t,u) Idu = fdtf I roO
If(t,u) I du
rX
dt
rX
If(t,u) I du

= - dt f e-tt(8) e' -e du
eu
rX rX ~~~e-U_e-eu
+ dt e-tt(8)10-e du

and performingthe integrations in respect to u

dtJ If(t,u) I du
Go 00

=-J
1

e-tt(8)-l log t dt + J e-tt(8)1-log t dt.

G

for 0 < S(s) < 1, but since the infiniteseries converges uniformlyin any finiteregion of the
s-plane to which the points s = 0, - 1, :1: 2, *. are exterior,and consequentlythe series is a
functionof s holomorphicin this region,the above equation is valid for all values of s distinct
from0, 1 1, :1:2, *-* . Conversely,we may deduce this partial fractionexpansionfromthat of
the cotangent by means of the identity 2/sin7rs = cot 7rs/2+ cot [7r(1- s)]/2, and then the
above expansionof the integralgives us a new proofof (121).

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 111

Bothofthelast integralsconvergesince9?(s) > 0; hencefdtif (t,u) Idu

convergesand by theoremXXII, we may reversethe orderofintegration
in the repeatedintegralforP'(s), whence

du fta-i
e--t _
'(s) Gof e-(l+u)tct
= Jdu
(uOO[ Jt-
Fr(s) du
[er(s) -(1+u)8

by (115) and (117), or finally,writingt insteadof u in the integral,

(124) APt(S)
= r(s) =f (e-' - (1 + t)-8) t 9?() >0)

which is identical with (90). We now defineEuler's constant by the

relation
(125) - C = (1) = '(1) = (e-t - (1 + t-1) dt
whenceby subtraction
I/'(s) + C = f ((i + t)- - (1 + 0-8) t

and replacing1 + t by et
-
(126) IP(s) + C = e-t dt.

underthe integralsign,
Differentiating

(127) +'e(s) = 1 te dt 1
1~ ~ ~ 1-e
whereevidently 1- e < M fort O. so that forSR(s) E

f0te-t
e-t dt
00M
Mte-dt
1 =f2

Consequentlythe integralin (127) convergesuniformlyfor S?(s)(

so that the differentiation
underthe integralsignis legitimateby theorem
XXI. We have
teaSt 00te-t t teN(8) -t
1es- =-- t 2: te-
= Ze+t, X8V Ie
1e- fol= 1--t
s - et =
-I
E ta- (91(8)+^?
o

and sincef 1 e-t dt convergesfor W(s) > 0, theoremXXII cor.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
112 T. H. GRONWALL.

gives
e
+t1s)= v=O te-(+,) tdt,

or evaluatingthe integralsby means of (117)

= Z0(8?v)2dso(1+v s+v)

convergentforS(s) _ (, s$ l c A, since
wherethe last seriesis uniformly
1 - 1 _
Is-11 A+1 A+1
1+ v s + v (1 +v) Is+ v I (1 + v)(E+v)- V2

and consequently, rememberingthat t,'(1)= -C,

A(s) + C E 4+ 4+)
This may also be written
d log r(s + 1) d 1 4( 8_log 8 + V)

the last seriesbeinguniformly convergentas before,as is readilyseen by

expandingthelogarithm, and sincefors = 0, r(s + 1) = 1, it followsthat

log r(s + 1) + Cs = ( -log(1 ?-)).

For s = 1, we obtain,since r(2) = 1,
00

C = E - -log(1 +-))

=lim 11+ +n-1 - g )

which is the definitionof Euler's constant given in J ? 1, and passing

from logarithms to numbers, the preceding equation gives, since
r(s + 1) = sr(s),
1 LS~a~ %
= e s1s 11 (l
CST0

F(s) + -e-8V) > 0.

Except forthe condition R(s) > 0, thisis the productformula(J 5).

20. Integralsfor log r(s). Raabe's integral. In the same way as in
the case of (89), we may showthat the integral(126) convergesuniformly
forSR(s)i E, I s I c A, and consequently

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 113
r0 e-t e- t U
(8-1)t
(s - 1)6t'(1 + u(s - 1)) + C(s-1) = f(s-1) e dt

convergesuniformlyfor St(s) e, | s | A, 0 c u c 1. By theorem

XX, the integrationin respectto u betweenthe limits0 and 1 may there-
forebe performedunder the integralsign to the rightand gives, since
r (1) = 1,
-
log r(s) + C(s -1) =
(s-1)te-t + e-8t e-t _

For s = 2 we obtain, since r(2) = 1,

Can mu
ltilye-itnb

and multiplyingby (s - 1) and subtracting fromthe preceding equation,

(128) log F(s) = J(s - 1)e-t + 1 e Jt] 9(s) > 0

whichis identicalwith (94) and is seen,as in the case of (89), to converge

uniformlyfor 9?(s) e, | s l A. We shall now evaluate Raabe's

I
integral
log r(s + a)da, 9?(s) > 0,

where,on accountoftheuniformconvergenceof (128), log r(s + a) is con-

tinuousfor9?(s)> 0 and O a_ 1. Fromlog I'(a) = log 12(1+ a) - log a
and the existenceof 3 log ada it thereforefollowsthat 3 log I'(a)da
01 pl

exists. The integral

I = flog rJ(us+ a)da,

where9?(s) , I sl A, 8 u c 1 (8 being arbitrarilysmall) may be

underthe integralsign,since the resultingintegral
differentiated

s{l(us + a)da
convergesuniformly forthe values ofs and u considered. But integrating
in respectto a and applying(116), we find
aI = s(log r(us + 1) - log ip(us)) = s log us,

in respectto u betweenthe limits0 and 1,

whenceintegrating

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
114 T. H. GRONWALL.

log r(s + a)da log r(a)da = f slog usdu = s log s - s.

Now, replacing a by 1 -a,

J log r(a)da = J log r1( - a)da

and consequently, applying (121)

J log r(a)da = 2 J (log r(a) + log r(1 - a))da

= 2 I (log r - log sin ra)da = log ir- 2j~ log sin rada.
o
To the last integral, we apply the formula sin 7a = 2 sin 2
sin ira = 2 ira r(1+a)
whence

log sin = log 2 + log sin 2 da + logsin )da

irada (1+

= log 2 + 2 log sin 7rada+ 2 log sin + 7ra) da

r Il
= log 2 + 2 log sin irada + 2 log sin rada

= log 2 + 2 f log sin irada,

so that
J'log sin rada =-log 2
and
rlog r (a)da = log - + I log2 = log 12r,
and finally
*1
(129) jlog r(s + a)da = s log s - s + log 2-,, 1(s) > 0.

We may obtain another expression for Raabe's integral by replacing s

by s + ain (128):
a
logr(s + a) =I (s + a -1 )eet +
- e-
]t
t
(130) J 1- e

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 115

and integratingin respectto a from0 to 1; the integralto the rightbeing

uniformly convergentin respectto a from0 _ a c 1, we may perform
the integrationunderthe integralsign and obtain

(131) flog r(s + a)da =f;[(s -)e- + -t

- 1- e-t]'t 9?(s) > O.

andtheasymptotic
21. Binet'sintegral expansionoflog r(s+a) andV/(s+a).
From (130) we subtract(131) and findby the aid of (129):
log rI(s + a) - (s log s - s + log n27r)
0
- D)e-t - eS + e-s-at
=
,[(a i~ t+1 -e-t ]dt
T;
from(87) we obtain
coe-t -e-8t
og Js edt,
t
and multiplyingby a -2 and subtractingfromthe precedingequation,
log r(s + a) = (s + a - 2) log s - s + log is1-J7
1'( 1 em'at e-Stdt

Rewritingequations (97) and (98),

(132) co(s, a) i(a 2 t + l -e ') t'

(133) 2 t+ 1 e t

we findas in ? 14, upon successiveintegrationby parts in (132),

(134) c(s, a) = e-Et ]tk- k ] + 2 - I t2m-2

It is now proposed to show that this formulayields the asymptotic

expansion (78) for 9?(s) > 0, and the proofwill be conducted without
any comparisonof (134) to (77) (since we have not establishedthe latter
formulaby the methodsof the presentchapter). To this purpose,we
note thatf(t, a) is a functionof a holomorphicfor0 c a c 1, and may
thereforebe expanded in a Fourier series uniformlyconvergentin this
interval
f(t,a) = + x:(an cos 2n7ra+ bnsin 2n7ra)

= + +E (A e2nria + A-ne-2nfia),

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
116 T. H. GRONWALL.

where, by (34),

An=2(an-ibn)=fJf(t,a)e-2nfriada,An= 2(an+ ibn)=J'f(ta)e2nwiada,

From (133) we find
ao=2 f2(t, a)da = 0

and
An= (a- - + e_)e2nfriada

1r 1 ae2nffiaa=1 + 2nr -t e(2ri )a]a=

t L2nri Ja=o t L2n7ri -t 1 -e-t Ja=o

t (2n~ri 2n7ri - 2n7ri(2n7ri -t)

Anis obtained fromthis by changing the sign of i, so that

Go e2nrixt e-2n ria
(135) f(t,a) = -E 1(2n7ri(2n7ri+ t) + 2n7ri(2n7ri
-t)

Differentiatingterm by term in respect to t, which is legitimate since

the series obtained are uniformlyconvergent for t : 0, it is seen that

a f(t,a) (( 1)c k-le2n7ria e-2n7ria

(136) atk-1 = -(k-1)! - El *(2i + k+ t) 2n~ri(2n~ri-t) k/c
a
and fromthe uniformconvergence for t _ 0 it follows that -1f(t, a) 40
as t- oo. For t = 0, we obtain
a) ( 1)k-1e2n7ria+ e-2n7ria
a k-1f(t,
at JeI =(=O (k 1)! E
~~~n=1 (2n7ri) k+1

and comparing the cases k odd and k even separately to (42), we find

[a-f(t,
a)k ] - 1)k(k1)'2(k - 1)!Pk+l(a).

Since I 2n7ri? t I 2n7r,it follows from (136) for k = 2m - 1 that

a)
a2m!.-2f(ty
(2m 2)! E
"O 2 (2m 2) !P2m(O),
f
82m- a)|< - (2n)2m
= -

and for s = s e' (- 7r/2 < 0 < 7r/2since 9?(s) > 0),

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 117
"Oa~-2f (t, a) 0
e-stdt < (2m - 2) !P2m(O)fe181 cotdt

=(2m - 2)! P2m(O)

js ICos 0

Introducingall thisin (134), we finallyobtain

(us, a) = E- 1)k(k-)I2(k - 1)!Pk+l(a) (2m - 2)!P2m(O) h

(137)
|h I< 1, 0 ca < 1, 12 6 < 72r

which,except for the slightlydifferent

formof the remainderterm,is
identicalto (78).
We have
W*(s,a) = log s - V(s + a) = (2-a)
2
1 (s, a)
s O9s
and by theoremXXI, it is legitimateto differentiate
in respectto s under
the integralsignin (132), whence

(138) C)*(s, a) = (2 2- a) - +so i tf(t,a)e-stdt, T (s) > 0.

Integratingby parts, and using (135) in the same manneras above, we

find
c*(s, a) = (1 - a) - + E (- 1) (k-1)(k-2)2 (k 1)!Pk(a)
S k=2 S

(139) + (2m - 1)!P2m(O)*h

hIh < 1, 0 Pa < 1, - < a < '

whichdiffers from(84) onlyin the formof the remainderterm.

22. The integralsof Schaar and Landsberg. Let us firstassume s real
and positiveand replacef(t, a) in (132) by the series(135), so that

co(s a)
= - -'
co 0 ( ___eniaeni
e2n~ria e-2 n7ria \ e..tdt.
an= 2n-iri(2n-iri+ t) + 2n-7ri(2n-7ri
-
t))

Since I 2nri ? I _ 2n-r,the absolute value of the gene-al termin the

seriesdoes not exceed ((2n)2 + (2n )2) e-8t, and the integral

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
118 T. H. GRONWALL.

JI 27r1n2e-Lt = 27r2SE 12

beingconvergent,the same is the case with

'-' go ( e2nfia e-2n7wia e tdt.
J L1 2nri(2n~ri + t) 2n7ri(2n7ri-t)
+
By theoremXXII cor.,we may therefore reversethe orderof integration
and summationin the expressionforw(s,a), and write
(' (
wo(s,a) =~l - e) V
e2nfria
2n'ri(2n ri + t) + 2n7ri(2n7ri-
e-2nfria
e))
\
dt

in each integral(this is a real substitutionsince s

Replacing t by 2nirt/s
is real), we obtain
c) 20 e1 2nffia e2nf ia e2nfft

cos}a)= lJo2nr ss+ ti + s-ti) -8dt

since Is i ti I s, the absolutevalue of each integranddoes not exceed
and the series
(1/n7rs)e-2nw',
00 coe-2nff
t 1 W1
EJ 7rsdt 27rsEn2

beingconvergent,the same is the case withthe series

E
-n=l (s + +
fli-s
2n7r ti + s-ti +2I )e-2nlrsdt.
)e2wt|d

By theoremXXII cor., we may thereforeagain reverse the order of

summationand integrationin wcs,a), whichgives

co(s, a) = Jo 2fl7 A +ti + s-ti )dt

and summingthe infiniteseries

(4 (s, a) 7( ti log 1 -2
7r(t1 a

(140) 1
+a -ti log 1 - e-2r(t-ai) dt.
Passing to the case of s complex,we may now writeforST(s) > 0

co(s, a) = log 1

f
s - e-2f (t+ai)dt

(141)
(141) +1 2 + 1 1 dt -~log 1
~~~~~~~~~~~~~~~+ ~~a)dt.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 119

In fact, 1/(s + ti) and 1/(s - ti) are holomorphicin s for R(s) -,
s j c A and t : 0; since j s i ti l e and

log 1l -wta)=og j
eg e-27r(t=1ai) Cg _e-2,w t

(whichis seen by expansionin series,as in (106)), and furthermore

001 1 d
f log -e-2dt

converges(the integrandbecominginfiniteas log t fort -* 0 and vanish-

ing as e-27rtfor t -+ co), both integralsin (141) convergeuniformlyfor
T(s) E) I s I c A. The applicationof theoremXXIII now showsthat
the rightside in (141) is a functionof s holomorphicforT(s) > 0; the
same beingthe case withcw(s,a), and both sides in (141) beingequal fors
real and positive by (140), it followsthat the equality subsists in the
entirehalf-plane ?(s) > 0.
Writing1 - e-2fr(tIai) = rekei,we have r2 = 1 e-2ert cos 27ra+ e47t
sin 27ra
0 = arctan e2fft- cos 2ra
and making
a1t(t,a) = 4-log (1 - e-2rt cos 27ra+ e-4frt),
(142) 1 sin 27ra
x(t, a) = - aretan e2ff-cos 2-a'

we findby a simplealgebraictransformation
of (141)

(143) co(s, a) = C2tx (t,28+'ta)

a) dt -, dt 9 (s) > 0.
J2o t Js2 +t2

(141) in respectto x = 9S(s), we find

Differentiating

(2 -a) -co*(s, a) =-(s, )

(144) -
8 + i) 2 log 1-e-2 (t+ai) dt

- X1(s - ti)2 log 1 -e-27r( t-ai) dt;

the differentiationunderthe integralsign is legitimateby theoremXXI,
since the integralsobtained convergeuniformlyfor9?(s) e, I s I A,
as is seen in exactlythe same way as forthe corresponding integralsin

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
120 T. H. GRONWALL.

(141). Integratingby parts in the formulajust obtained, we findfor

0 < a < 1

271J7
- (s + ti)'2 g 1 - e-2(t+ai) dt = [2ri s+ tilog 1 e-2f ( t+ai) IL
- 1 e-2ff(t+ai)
JOs+ ti 1-2 e27r(t+ai)

and the firsttermto the rightequals

1 log (1 2iras (/ -2ei
2-rislo 1-e-2,ff
-

ai) =2-i log ~2e ff.(1I2a)i 2i /

2is [w(4 - a)i + log (2 sin ra)].

The second integralin (144) is foundby changingthe sign of i, and we
finallysee that
1 (t+ai) d
e-2ff
,a) = ?0 s + ti 1 - e-''tai)
(145) 00 1 e-2r(t-ai)
Jo
s0 ti * 1_- e2r(t-ai) dt, 9?(s) > 0, O < a < 1,
or writing
cos 2wa - e
alt'1(t,a) =e2fft - 2 cos 2ra + e-2fft

(146) sin 2wa

Xi(t, a) = e2fft- 2 cos 2ra + e-2wrt'

(145) algebraically
and transforming
2sx, (t, a) d a) dt,
+ Jo 2t4f1,(t, () > 0.
=
(147)
(14) * s, a))
cwo*(s, = s 2dt s82 +td ~(S)9
J 82 + t2 + t2

(147) is also seen to be true fora = 0 by makinga = 0 in (143), differ-

entiatingin respectto x and integratingby parts. In the special cases
a = 0 and a = 2 these expressionsare considerablysimplified,and we
obtain
c(8, 0) = s2+ 2log 1 e irdtX
('
-
2t dt
sO*(S'?) J8s2 + t2e2'ft.
(148) 1
W(S, 2) = 7J S2 + t2log 1 + e-2wrdt'

@*(S, 2) = I
2t e dt
+2
82+ t e2irt
+1

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 121

and (143) and (147) to Landsberg.30

The integrals(148) are due to Schaar,23
Usingthe identity

1 (- 1) k-lt2k-2 (_ l)m-lt2m-2
2 + t2k 82k + 82m-2(82 + t2)

in (143) and (147), and proceedingin the same manneras at the end of
? 15, thereresultasymptoticexpansionswhichmust coincidewith (137)
and (139). Comparingcoefficients, we readilyobtain

P2k(a) (2k - 2)!f a1(t, a)t2 2dt

p00
2

= (2k -1)
1' (t, a)t2k-ldt,
(149)2 X

P2k+l(a) = - (2k-1j-,J'X(tja)t2k-ldt

= (2k) xi (t, a)t2kdt.

!J2
series. Let s be real and
23. Kummer'strigonometric e s 1 -e
we thenobtain from(128)

e-st - e(-Btd
1 rF(s)s)
l00~ + (2s -)e-t
log t
og -- s 1-et
_

and subtracting(125) multipliedby 2s - 1

(150) log 1(1- s) + C(2s -1) = f+[e t=e-t-)+ 21 Jt]

We now have the Fourierexpansion

e-st ~e-(l8) t2s -1 ao00
1-_ etz + 1 + t = 2?+ Z(an cos 2nrs + bn sin 2nrs)

uniformly convergentfor0 _ s c 1 and a fixedvalue of t, since the left

side is a holomorphicfunctionof s. It is seen from(34) that ao = 0
and that forn > 0
30Landsberg,G., Sur un nouveau developpementde la fonctiongamma, M6m. Ac. Belgique,
vol. 55 (1897).

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
122 T. H. GRONWALL.

an + ibn 2,(+ _ 1+ t e2nriads

2 r
_(2nri-t) e-t+(2nvi+t)s-1 2
re(2s-l)e2nsi8=1
1-e-tL2nwri t t 2= + 1 + t( 2nwri 8=

2 2 2i
2nri-t 2nwri+ t n7r(1 + t)

8nri 2i
4n27r2 + t2 n(1 + t)'
so that
[e-st e-(l8)t 2s -1] 1 ,0 sin 2nws
(151) [ 1e-t + + t t = (t) n7r
0 S- _1 t 0
where
1 / 27_2 2 2 2t
(152) Cn(t) =t4 2 + t2-1+t 1+t4n22 + t2

increaseswithn forany fixedpositivevalue oft. We shallnowprovethat

it is permissibleto substitute(151) in (150) and integratetermby term,
or in otherwords,that
00 00
J7tEcv(t)
(
sin 2vwrS
r) dt - 0 as n -* oo

uniformlyfor e c s 1 - e. Writing

? sin 2yrS

so that
sin 2v7s
= SV- SV+1
v7r

it follows from (38) that for e _ s 1 -e

1SI -v7r sin re'

and fromthe identityof summationby parts

n+p n+p

I: Cv'(Sv' Sv'+,) = I: (Cv' - CV-l)SY + CnSn+l -Cn+p+lSn+p+l

v=ZS+l v=Z++l

we conclude,since c,(t) - c,+1(t) > 0 fort > 0, that

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
 THE THEORY OF THE GAMMA FUNCTION. 123
1
|n~p
P tsin 2vrs c .~ (c^(t) -

c sinr 1
(13 i1+ V(t) ,+

(153) L v~~1 + n(t) | + IcCn+p+l(t)|

n+ 1 n+ p+1
From (152) we obtain

(154) ('cn(t)dt = [o4n2 + t

12 2+ t2 I t=o = 2 log 2n7r,

and since cn(t) ' 0 as t 4n

fr0 o
Cn(t) idt = J
* o
e/4n2,,2

Cn(t)dt
00
Cn(t)dt
~~~~~~~~~~~~4n2ir2

=2 log (1 +4 + 2 log 2nr <4 log2n,

so that,by (153)

ZCv(t) s2 dt< w sin l +

+2 log 2n~r+2 log 2(n + p + 1)7r

+ n+ 1 + n+p +1

Letting p -* , and observing that log (1 + _1) < and

00

v - = -' we finallysee that

00 00 sin2vwrs 2 1 2log2nw
f((v=+t) ) dt rsin7rE( n+ n+1 ) *O as n o
uniformlyfor e c s1 - e. Substituting (151) in (150) and integrating
termby termwiththe aid of (154), we findfor0 < s < 1

log r(s)
-
+ C(2s -1) = 2l nl sin 2nwrs,
roa( s) +r
the series being uniformly convergent for 1 - e. Using (40) e 5 s
and (121) (the latter being identical with J 6), we now immediately
obtain Kummer'sseries (113).
Additionalnote. Applicationof complexintegration to the theoryof
the Gamma function.We observed at the beginningof chapter III
that whenthe Gamma functionis definedby means of Euler's integralof
the second kind,we are therebyrestrictedto the regionof convergence
is overcome
of thisintegral,viz.,the halfplane 9?(s) > 0. This difficulty
by the use of complexintegrationand othermoreadvanced parts of the

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions
124 T. H. GRONWALL.

theory of functions of a complex variable. The author's original inten-

tion to devote a fourth chapter to an exposition of these methods had
to be abandoned through lack of space, and it was found feasible only to
give the followingbriefreferencesto the literature.
A paper by Birkhoff3'contains what is probably the shortest way of
arriving at the main properties of the Gamma functionfromthe function
theoretic point of view. Starting fromthe definition

F(s)== lim (o(s +n +1)

n-;s) ,
8m(s + 1) (s + n)'

p(s) = S8-1/2e-8 2w, arc s < or,

which is essentially equivalent to (J 15), Birkhofffirstproves the formula

(J 14) log r(s) = (s - 2) log s - s + log >I2w+ W(S)

with the Gudermann expansion of c)(s) (J 13) and the fundamental
property of this function, viz. Ic)(s) I < K/p, where K is a constant and
p the distance of s fromthe nearest point on the negative real axis. Next,
the formula r(s) r (1 - s) = 7r/sin7rs (J 6), the infinite products for
r(s) (J 3 and 5), and Euler's integral of the firstkind are established,
and after a digression on i/i(s) = log s -co*(s), ic*(s) i < K/p2, Birk-
hoff'spaper closes with the expression of Euler's integral of the firstkind
in terms of the Gamma function (71 in the present paper).
By the use of Cauchy's integral, a systematic study may be made of
a whole class of summation formulas, containing that of Euler, and as an
application, the formulas (76) and (82), Binet's integral (97) and those of
Schaar (143, 147) may be derived froma common point of view, together
with the various asymptotic expansions to which they lead. An excellent
exposition of this aspect of the theory is given by Lindelof.32
Attention is finallycalled to an important paper by Mellin,33in which
is set forththe intimate connection of the Gamma function with a large
class of other transcendental functions (for instance, the Riemann Zeta
function). The main purpose of Mellin's paper is however to apply the
Gamma function to the systematic solution of a class of linear difference
equations (the simplest case of which is solved in J ? 11) as well as to a
corresponding class of linear differentialequations, the coefficientsof
which are polynomials of the firstdegree.
31 Birkhoff,G. D., Note on the Gamma function,Bulletin Am. Math. Soc., vol. 20 (1913-14),
pp. 1-10.
32 Lindel6f,E., Le calcul des residus,Paris, Gauthier-Villars,1905 (Collection Borel).
33 Mellin, H., Abriss einer einheitlichenTheorie der Gamma- und der hypergeometrischen
Funktionen,Math. Annalen, vol. 68 (1910), pp. 305-337.

This content downloaded from 91.229.248.39 on Sat, 24 May 2014 14:24:38 PM

All use subject to JSTOR Terms and Conditions