Aequationes Mathematicae 26 (1983) 208-220 0001-9054/83/002208-1351.50 + 0.

20/0
University of Waterloo t~ 1983 Birkh/iuser Verlag, Basel

Series like Taylor's series

JOHN BORIS MILLER

Abstract. A theorem and some examples are given concerning the convergence, in a space of generalized
functions, of power series whose terms contain successive derivatives of a given function. One example is
the Euler-Maclaurin sum formula, where there are some novelties.

1. Introduction

The Taylor series for a function x belongs to the class of series of the general type

~oX(t) + a~2x'(t) + ~222x"(t) + ~323x'"(t) + .... (1)

an in this case being 1/n!. We state a theorem concerning the convergence and sum
of such series in a space of generalized functions, with somewhat arbitrary coefficient
sequences (an), and we give some examples. Although the convergence in each case
is in a generalized sense, but norm convergence nevertheless, the derivatives can
always be interpreted as ordinary derivatives for suitable x. In many cases the sum
of the series is a convolution.
Undoubtedly the two most celebrated examples of such series are Taylor's series
itself, and the Euler-M aclaurin series ~ ( - l)n - 1B~22~- 1xt2n - l>(t)/(2n)! which appears
in the sum formula of that name. There are other fairly simple series in the class,
such as (19), (23), (26), (27) below, which can be proved easily enough for ordinary
functions under suitable conditions using successive integrations by parts, and which
may be of some interest. The formulation of the general case given here clearly shows

AMS (1980) subject classification: Primary 40A30, 44A40. Secondary 65B15, 301305, 44A10, 44A35.

Manuscript received February 7, 1983.

208
Vol. 26, 1983 Series like Taylor'sseries 209

a natural common origin for these examples: the coefficients ~n are the coefficients
in the power series expansion of the Laplace transform of a generalized function a,
and the series (1) has for its sum the generalized convolution of x with a. We are not
restricted to power series expansions which converge everywhere, as this description
might suggest, because of the way the space of generalized functions is defined.

2. The space In of generalized functions

The space to be used is that described by the author in [3], there called ~ . and
here written I n . It is particularly suitable for dealing with series of the type (1), and
is defined as follows. Let L t denote the Banach algebra of (classes of) Lebesgue-
integrable complex-valued functions on • + = [0, ~ ) with convolution product,
so that for t~R ÷,

x" y(t) =
fo x(u)y(t - u)du, l[x I[ =
;; Ix(u) ldu. (2)

For x ~ L 1 the Laplace transform

(L~'x)(s) = ~¢(s) = f ; e-Stx(t)dt (3)

is holomorphic on the half-plane Ilo = {s~C: Re(s) > 0}, and

(x. y)^(s) = ~(s).'~(s). (4~

Let fl be any nonfinite compact subset of this open half-plane, and write

Ixln = maxl~(s)l. (5)

$¢fl

Then [. [n is an algebra on L ~. The space t n is defined to be the completion of

L ~ with respect to this norm; it is a commutative complex Banach algebra and its
elements are called generalized functions. The Laplace operator La extends naturally
to a bounded linear operator ~ : In---, C(f~).
The justification for regarding the elements of I n as generalized functions is
explained in [3]; I n contains a unit, namely the "delta function" 6, for which oe(s) = 1
on f~, and it contains a generalized derivative
210 ,o.r~ BoR,s M~LLER AEO MATH

x [11, with (x°l)A(s) = sx(s) (sef~) (6)

for each of its members x, and a generalized integral

K x , with (Kx)A(S) = s-l~c(S) (s~f~). (7)

Differentiation and integration are bounded operators on ~n and each is the inverse
of the other. The space ~[Bn, of all (classes of) locally integrable functions w on R ÷
whose Laplace transform integral ~v(s)converges on some open half-plane containing
f~, can be embedded in a natural way in 3En. So too can LP(R +) for 1 ~<p< oo.
For functions x in these spaces,

(Kx)(t) =
fo x(u)du. (8)

Any locally integrable function with compact support in R ÷ belongs to ~[Bn, for
every fL
Moreover, by judicious use of the translation operator Th, where

( T , x ) ( t ) = 0 (0 < t < h), x(t - h) (t >t h), (9)

( Thx)^(s) = e-h'~c(s), (10)

all functions in L I ( - oo, oo) or even in ~ n ( - oo, oo) whose support is bounded below
can be embedded in a;n also, in a consistent way; and the generalized derivative x f~l,
defined initially as the limit in norm l'ln of h - l ( T h x - x ) as h ~ 0, can be viewed as
a two-sided derivative by taking (10) to define T~ also when h < 0.
The connection between the nth order generalized derivative x [nJ and the ordinary
derivative x (n) can be formulated as follows ([3], p. 417).

L E M M A 1. L e t n e N and let x, x (~). . . . . x (~) exist and belong to ~2Bn, with x ~ - ~

absolutely continuous on R +. Then

x t~ = x (~) + x (~- "(0)~ + x ~ - 2~(0),~['] + . . . + x(O)6 t~- ~3, (11)

that is, x t~t and x (n) differ by a linear combination o f b and its first n - I generalized
derivatives, the coefficients being the first n - 1 righthand derivatives of x at O. If
OSsupport(x) then x tnl = x (~).

We shall abbreviate the notation for the ordinary derivatives x (~, x ~2), ... to
X', X", . . . .
Vol. 26, 1983 Series like Taylor's series 211

Much of the above theory depends upon the fact (5) that ~Se is an isometric
isomorphism from ~n onto its range ran(Le) in C(f~), where the latter has the supremum
norm Ilfll = maxlf(s)l. The nature of ran(ZP), that is, of the Banach algebra of
sell
continuous complex-valued functions on the compact set f~ which arise as Laplace
transforms of generalized functions, is discussed in [3]. In summary we have ([3],
p. 412):

L E M M A 2. If f~ does not separate the plane then ran(.LP) is the algebra of all
functions which are continuous on f] and holomorphic componentwise on its interior.

According to the nature of~, always assumed nonfinite and compact and contained
in Ho, one obtains different types of spaces 3~n. For example if f~ is at our disposal
then given x e L 1, x # O, we can always find a space 3Ea containing an inverse x -1,
by suitable choice of f~ ([3], p. 419). These matters are explored in [3], pp. 4t 1413.
Let ,,,~o~
~'~ 0 , ~ denote the algebra of all infinitely differentiable functions x on R + such
that x(k)(o) = 0 and x~k)e~tBn for k = 0, 1, 2. . . . .

T H E O R E M . Let

f(s) = ~ or.s" (12)

n=O

be any power series whose open disc of convergence contains fL Then f[ f~ is the Laplace
transform of a generalized function a, and for all x e En,

x'a = Otox + ctlx111+ 0C2X[2] q- . . . . (13)

lf aeffl3ta and xeC(o~ then

f l x(u)a(t - u)dt = CtoX(t) + cqx'(t ) + ot2x"(t) + .... (14)

that is, the partial sums of the series on the right form a sequence of functions converging
in norm ['In to the function on the left. If, in addition, the Laplace transform integral
(La a)(s) converges for all Re(s) > 0, then for 0 < 2 < 1 we have

fo x ( u )1- ~ a (t- -u- f - ) d u = e o X ( t ) + ¢q2x'(t)+ 0~222X"(t) + •. • (15)

in the same sense.

212 JOHN BORIS MILLER AEQ. MATH.

The proof is an immediate consequence of the preceding theory.

We mention in the next section some examples wheref(s) is expanded instead as
a power series in s-~, giving rise to series whose terms contain successive integrals
rather than derivatives.
The product in (13) has other interpretations than convolution of functions in
particular cases; for ezample it may be a translate or a derivative (as in (10), (6)
respectively). The integral (8) is an important convolution.
In [3] two other species of spaces of generalized functions were described: firstly
the projective limit 3 ~ of spaces 3Eoas f~ runs through a directed family of sets covering
II,~ = {s: Re(s) > o9}; secondly the inductive limit ~ of these projective limits 3£~ with
respect to o9 > 0. The Laplace transforms in the first case constitute all functions
holomorphic on II~,; in the second case they constitute all functions hotomorphic on
right half-planes of their own choosing. Clearly the theorem is extendible to these
spaces only when (12) converges for all s.

3. Examples

3.1 Taylor series. Take

~, = ( - h ) " / n ! , f(s) = e -~. (16)

The generalized function a in this case is not an element of ~Bta (being the translate
6(t - h)); for x~C~o~d equation (13) becomes

v
x(t -- h) = Thx(t) = x(t) -- x'(t) + - ~ x " ( t ) - - . . . . (17)

When h < 0, ThX is the generalized function identifiable with the function x translated
to the left, with support not necessarily in R +.

3.2 Take f(s) = (p + s)- 1, with p > 0, so that

a(t) = e -p' (t >i 0), a ~ L 1. (18)

If f~ lies in the open disc of radius p and centre 0 (and of course in Flo) then (14)
gives

e-°t l' eP,x(u)du

do
x(t)
= p-
x'(t)
p2
x"(t)
Jr" p3 ....
(19)

For example, when x is the Heaviside function h we have as a case of (13)

Vol. 26, 1983 Series like Taylor's series 213

1-e -p' = h(t)---+f(t) ¢Stll(t) 6t2J(t)

+ ....
/9 /92 p3

On the other hand, if 11 lies wholly outside the circle of radius/9, then we get instead

e -p'
;o eP"x(u)du = K x ( t ) - pKAx(t) + pAKSx(t)-..., (20)

where K is the integration operator of (7) and (8). Of course equations (19) and (20)
are easily verified for pointwise convergence under suitable extra conditions on x
and p, by integrating by parts appropriately in each case.
If p < 0, then there is no choice of set fl giving rise to (19); but (20) is still possible.
On the other hand, we can get the series in (19) when p < - 1, but a different sum,
as follows. Write p = - 1/2, 0 < A < 1. Thenf(s) = - 2 ( I - As)-~ can be expanded as
-A(1 + As + XAs2 + ...) (if Ist < A-x) and also as

2 _ i (if I s ( s - 1)-11 < ( 1 - 2)-~)" (21)

These domains intersect in the degenerate annulus A bounded by two circles touching
at s = 2 -1, and when -,=,-,o.nv~c'l®}and D. c A c~{s:Re(s)> 1} we obtain, as the inverse
transform o f f (s)~c(s),

I ~ 1{ 1ytt x(t) x'(t) x"(t)

-P,~o~..~ 1 +-P/doe"U"X'")(t- u)du = p p2 F p3 "'" (P < - 1),

(22)

3.3 Take f(s) = p(p2 + s2)-1 with p > 0, so that

a(t) = sinpt (t t> 0), a ~ n .

If f~ lies in the open disc of radius p, then (14) becomes

fo x(u)sin(p(t-u))du = x(t)
P x"(t)
~ + x'~(t)
~ .... (23)

whereas if f~ lies outside the circle, then

f l x(u)sin(p(t -- u))du = pK 2x(t) -- paK4x(t) + pSK6x(t) - . . . . (24)

214 JOHN BORIS MILLER AEQ. MATH.

3.4 Let ~ > fl > 0, and take

e-at _ e-#~
f ( s ) = log(S~+.~fl~, so that a ( t ) = (25)
\s+~/ t

aeL 1. If t2 lies in the open disc of radius fl, then (14) becomes

fo' e-au --u e-~U.x(t - u)du

= log ( f l ) x(t) + 0c~flflx,(t ) ~2_f1220¢2fl

2 x"(t) + ~33~_x f13 ,,,( t ) - . . . (26)

and there are variants of this when f2 lies in the annulus or outside the larger circle.
If instead • > fl = 0, then aq~L1, and instead of (26) there is, for f~ in Isl < ~,

o[U- l(e -au -- 1)x(t -- u) + (C + logu)x'(t - u) ]du

= -logctx(t)- x'(t) "t x"(t) x"'(t) + (27)

ct 2ct2 30C3 "'"

(C = Euler's constant 0.57721566 ...).

3.5 The Euler-Maclaurin sum formula. Let co > 0, and take

1 /cos'X 1 BICO B2CO3S2 B3coSs 4

f(s) = ~scoth[\-~-/] = ~ s 2 4 2! 4 . ~ -I 6! "'" (28)

for 0 < I s [ <27rco -1, where Bz = 1/6, B2 = 1/30 . . . . are the Bernoulli numbers.
Suppose that f~ lies in Hoc~{s:)sl <27rco-1}. Let yeC~o~d and apply the theorem
to x = y', so that ;(s) = s~(s). We have f(s) = 8(s) for Re(s) > 0, where

a(t) = n - ½ for (n - 1)co < t < nco (n = 1, 2 . . . . ) (29)

([1], p. 256, (8)), ae~/13n for all D, and

Blco A (30)
"d(S)X(S) = sf(s)'y{s) = I s - ly(s) + -~. sy(s) - - . . . .
(I)
Vol. 26, 1983 Series like Taylor'sseries 215

The lefthand side is the Laplace transform of

(n -- ½) y'(t -- u)du + ([t/09] - ½) y'(t - u)du

n= i n- 1)~
[t/wlco

[t/~ol
=½y(t) + 2 y(t -- r09),
r=l

so we get

[t/toly(t - r09) = If y(u)du + ½y(t) + ~ ( - 1)k-1Bk092k-1 y(2k -

r=o --, ~= 1 (2k)? l~(t)" (31)
0

This is the Euler-Maclaurin sum formula for 3En; for the classical theory see [2], Ch.
13, or [8], p. 128.
Let

tt/col
(S~,y)(t) = ~ y(t - r09) = T~y(t) (32)
r=O r=O

define the standard summation operator S,~. On algebras of functions with pointwise
product S,o satisfies the Baxter identity with parameter value 1,

S x . S y = S(Sx" y + x . S y - xy); (33)

and (31) is the application to y of the operator equation

S~ = 1 K + ½I + ~ ( - 1)k-1Bk092k- 1D 2 k - 1 (34)
o9 k~ 1 (2k)?

where D is differentiation. This equation is discussed in some detail by the author in

[4], I-5] and [6] (though the cases treated there do not include the case (31)). In the
present circumstances where the product is convolution, D x = x' determines not a
derivation D but a multiplier:

D ( x . y ) = D x ' y = x ' D y = x ' y ' ~ tll (all x, y ~ ) . (35)

Since K is also a multiplier, being convolution with h, So, is itself a multiplier, in fact
216 JOHN BORIS MILLER AEQ. MATH.

So,y = (a tl; + ½6)'y = ~ (T,o,6)'y (all y~3En), (36)

r=O

and does not satisfy (33). For the Euler-Maclaurin sum formula in its most general
form in ~n we take (36) to be the definition of So and the formula is then

1 Bkco2k - 1
So,y = co- IKy + ½y + ~ ( - 1) k- yi2k- 11 (37)
k= i (2k)! '

the series converging in ~n for all yEXn when

0 < co < 2n(maxlsl ~-1. (38)

\ s~q /

We have

(S,oy)"(s) = e'°'(e o,s- 1)-l~(s). (39)

The Laplace transform lends itself to a nice description of the partial sums and
remainder term in the Euler-Maclaurin formula. Introduce the operators

1)k- 1Bkco2k- 1
SO,,N = co- IK + ½1 + N-
)-"l ( -
k=l (2k)! O2k- t, RO,,N = So, - SO,.N, (40)

so that the righthand side of (31) is lim SO,,Ny, in 3En. Write ~.(t) for the nth
N~<x)
Bernoulli polynomial,

1,,
qb.(t) = t"-½nt "-1 + ~. ( - 1 ) k-1 Bkt "-2k (41)
k=l

Since t p has Laplace transform (p + 1)!s-P- 1, direct calculation shows that s" +1q~.(s)/n!
is the partial sum, up to the term in s "-2 or : - 1 , of the power series for

s(e' -- 1)-I = ½s(coth(½s) - 1).

This leads to the elegant formulae

(cos) 2N ~
(S,o,NX)" (S) = ( 1 + ~4a2N(O~S) )X(S), (42)
Vol. 26, 1983 Series like Taylor's series 217

valid for all x ~ 3Ea, and

co2N-1 ~t ~t-u\
(s°,N,,)(,) -- ,, ) + Jo T (4,)

valid for all t >~ 0 when x~C~o2,~), N >t 1. The conditions on x can be relaxed since
(43) can be proved directly by integration by parts provided x, x', .... Xt2N-l) all
vanish at 0.
Let $.(t) denote the periodic extension outside [0, 1) of the Bernoulli polynomial,

$.(t) = $.(t)(0 <~ t < 1), $.(t + 1) = $.(t) for all t s R +. (44)

To obtain a formula for Ro,m, we use the formula

$.(s) = ~.(s) + ~-(e" - 1) -1 (45)

connecting the Laplace transforms of ok. and its periodic extension, proved briefly
as follows:

~[k+, e-s'~a.(t )dt

= ~'z. e-~S
k=O
f, e-~"4~(u + k)du
dO

= y ' e -I'~ e -~" ¢ . ( u ) + n 2 ( u + j ) ' - i du

k=o do j=o

= e-~('+k)~h.(u+k) au+n E E e-('~+l)" | e-"(u+J)'-Idu

j=Om=j dO

=
k=Odk
e-'tdj.(t)dt + e-V~_lj
of,
oo n F(n)
= e-"~J.(t)dt + e , _ 1 s"
do

Using (45) we have, for any x~Kn,

( l + ( -("~s)~"
~ . $ 2 s ( o ) s ) ) x ( s~) " " +~)x(~)
= ~2stogs,[(e~s)2N'~ e~" ^
218 JOHN BORIS MILLER AEQ, MATH.

which by (39) and (42) gives

O.12N
Sco,NX = _(2N)!
_,t,~
eeN . x [ Z N ] + S,ox (46)

( w h e r e f f ' ( t ) = co- if(co- it), (ff')^(s) = f(o)s)), so we have, when Y == c,(2N)

"-~o , n ,

(z)2N- 1 ~t /t-uS
(47)

This is, in another guise, the form of remainder which appears in some treatments
of the sum formula (see for example [2], (13.6.5)). F o r m u l a (43) on the other hand,
for a partial s u m rather than a remainder, is, if known, certainly not well-known.
Another novel formula worth mentioning is

On = SlOn, (48)

where 0, is the truncated Bernoulli function O,(t) = ~b,(t) (O ~< t < 1), 0 (t ~> 1). This
follows from the formula

~,(s) = : ( e ~ - 1)-1
fo' e - ~"d~,(u)du = e~(e~ - 1)-10,(s)

and (39). N o t e that the terms of SjO, are not all functions.
F r o m (31) we deduce that

lim ¢oS,oy = Ky (49)

~o~,0

in Xn, when y=,~o.n.

,,~:(®) T h a t is, the integral is the limit of its Riemann sums with
equipartitions. Another formula of a similar kind is: for ~ > 0 and ~ = r ( * )

(0 < t < ~),

N~ .=o f, x(u)du (~ < t).

(5o)
Vol. 26, 1983 Series like Taylor'sseries 219

3.6 Poisson's summation formula for the function e -'°' and its Fourier cosine
transform x/(2/rr)co(u 2 + co2)- 1 gives

1 e -"~'s 1 2 ~ s
(51)
2 + .= £, = -~ss + - o . ~ , s 2 + (2nrt/o)

This holds for Re(s) > 0, o > 0. So for xe~13n, any fL we have

½X(t ) + £
n=l
T.o,x(t ) = ~o- '~ox(u)du + 2o9-1 £r'
n=lJO
cos 2n~(t -
O)
u)x(u)du

(52)

which can be written

Sox(t) = o ) - ' K x ( t ) + ½x(t)+ 209 -1 oo~ |'cos-'~"'(t

f" 9mr - u)x(u)du; (53)
n=ld o 09

compare Poisson's proof of the Euler-Maclaurin formula, [2] p. 330, particularly

(13.8.3). Using instead the sine transform x/(2/n)u(u 2 + 092)- x gives

• ,k [" • ( 2 k + 1)ft,
~.~ (-- l)kr(2k+ , ) o X ( t ) = 0) - 1 Z ( - - l) | sin -a-- (t -- u)x(u)du. (54)
k=O k=O JO ZfD

(See [7], pp. 60, 64.)

Again, putting in the Poisson summation formula the function te -½'2 and its
Fourier sine transform u e - ½"~ gives an identity which, suitably interpreted in a variable
s, leads to the equation

(409)3/2 £ (-- 1)k(2k + 1)XU1(t -- (2k + I)2o9)

/~ k=O

t" 1 f ( 2 k + 1)rc./u'~
(55)
:
d
3.7 Suppose that a in the theorem is C1[0, oo). Then s~s~(s ) is the Laplace

transform of - d ( t a ( t ) ) . Differentiation of (12) with respect to s leads to the formula

dt

-
fo ua'(u)x(t - u)du = ¢toX(t) + 2cqx'(t) + 3ct2x"(t) + .... (56)
220 JOHN BORIS MILLER AEQ. MATH.

This equation holds in 3En,for every choice of ~ for which (14) holds. More generally
if aeC(k)[o, ~) then

t
( - 1)k
doI uka(k)(u)x(t -- u)du = ~oX(t) + 2 k g l x ' ( t ) + 3kct2X"(t) +... (57)

Further formulae for sums of translations can also be found by differentiating

with respect to s. For example, by differentiating

l+~e-.O~ = (l_e-,~s)-t = - -1+ 1 + ~, (-1) k-lnkco2k-ls2k-1,

.=1 cos 2 k=l (2k)!

we are led to the formula

~, n x ( t - nco) = (t - u)x(u)au - x(t )

n=O -~

~ (--1)k+lBk+lCO 2k((2k
+k=1 - ~ .+ 1)*x(2k)(t)" (58)

REFERENCES

[1] ERDELYI,A., MAGNUS,W., OBERHETTINGER,F., and TRICOMI,F. G., Tables of integral transforms,
Vol. 1. Bateman Manuscript Project, McGraw Hill, New York, 1954.
[2] HARDY,G. H., Divergent series. Oxford University Press, 1963.
[3] MILLER,J. B., Generalized-function calculi for the Laplace transformation. Arch. Rational Mech. Anal.
12 (1963), 409-419.
[4] MILLER,J. B., The Euler-Maclaurin sum formula for an inner derivation. Aequationes Math. 25 (1982),
42-51.
[5] MILLER,J. B., The Euler-Maclaurin sum formula for a closed derivation. J. Austral. Math. Sot., Series
A, to appear.
[6] MILLER, J. B., The Euler-Maclaurin formula generated by a summation operator. Proc. Royal Soc.
Edinburgh, to appear.
[7] TITCHMARSH,E. C., Introduction to the theory of Fourier integrals. Oxford, Clarendon, 1950.
[8] WHITrAgER,E. T. and WATSON,G. N., A course of modern analysis. Cambridge University Press,
Cambridge, 1946.

Department of Mathematics
Monash University
Clayton, Victoria, 3168
Australia