A Periodic Delta-“Function”

as a

Sum of Complex Exponentials

An object under consideration is

N
X
lim e2πi kt.
N→∞
k=−N

This is refered to as an “object” because the limit does not exist as a function, but
is rather a “distribution.” That is, a proposed equivalent form is

N
X ∞
X
2πi kt
lim e = δ (n − t) ,
N→∞
k=−N n=−∞

and like all δ-“functions,” this object must be judged by what it does to other
functions, as opposed to any explicit functional form.
For the purposes of examining the behavior as N → ∞, consider the partial
sum
XN
FN (t) = e2πi kt .
k=−N

Several properties of FN (t) are immediately clear:

• FN (t) is perodic with period 1,
• FN (t) is an even function of t, and
• FN (0) = 2N + 1.
As a consequence of the first property, the remainder of these notes will consider
only the range − 12 ≤ t ≤ 12 .

It should be clear that FN (t) is a geometric series, and so may be found in

closed form;

1
 N
X
FN (t) = e2πi kt
k=−N
2N
−2πi Nt
X
m
=e e2πi t
m=0
2πi(2N+1) t
e −1
= e−2πi Nt
e2πi t
−1
πi(2N+1) t
 πi(2N+1) t −πi(2N+1) t

e e − e
= e−2πi Nt
eπi t [eπi t − e−πi t ]
sin ((2N + 1)π t)
=
sin (π t)
for t non-integral. For t integral, the periodicity of FN (t) gives FN (t) = 2N + 1.
The nature of FN (t) may be examined by simple plots; a MAPLE worksheet may
be downloaded from the 8.03-ESG web page.
To see how FN (t) might be considered a δ-distribution in the limit N → ∞,
consider Z 1/2 Z 1/2
sin ((2N + 1)π t)
FN (t) dt = 2 dt.
−1/2 0 sin (π t)
Make the substitution
du
u = (2N + 1)π t, dt = ,
(2N + 1)π
so that
1/2 (N+1/2)π
sin u
Z Z
FN (t) dt = 2 du.
−1/2 0 (2N + 1)π sin (u/(2N + 1))
In the limit as N → ∞, the upper limit of the integral goes to infinity, giving an
improper definite integral, while the denominator of the integrand approaches π u,
as may be checked using l’Hôpital’s Rule. The result is that
Z 1/2
2 ∞ sin u
Z
FN (t) dt −→ du.
−1/2 π 0 u
This integral is well-known to many, but may be derived by use of a Tauberian
Parameter. What follows is from the 18.023 texts, Calculus: an Introduction to
Applied Mathematics by Greenspan, Benney and Turner, Page 533, or Calculus: an
Introduction to Applied Mathematics by Greenspan and Benney, Page 514. Specif-
ically, consider the improper definite integral
Z ∞
e−a u sin u du, a > 0.
0

2
 1
This integral may be done by many means, both foul and fair, to obtain 2 .
a +1
Then, Z ∞ −a u Z ∞ 
e sin u
Z
−a u
du = − da e sin u du + C
0 u 0
= C − tan−1 (a).
The constant of integration is found by taking a → ∞, in which case the integral
π
with respect to u on the left above must be 0, and so C = 2, and taking a → 0,
∞
sin u π
Z
du = .
0 u 2

Thus, we have
Z 1/2
lim FN (t) dt = 1,
N→∞ −1/2

a necessary candidate for any representation of the δ-distribution.

Now, the fact is that the above representation does not have the property that
the limit converges to 0 uniformly as N → 0 for nonintegral t. That is, consider
2
the first maximum of |FN (t)| (or, equivalently, (Fn (t)) ) that occurs at a value of
t > 0. It can be shown by basic calculus that this value of t, call it tmax , is the least
positive solution of the transcendental (or is it?) equation

1
tan (π t) = tan ((2N + 1)π t) .
2N + 1

This solution may be found numerically for a given N , or graphically. The advantage
to the graphical solution is that it is seen that as N → ∞, the least positive value
of t that is an intersection of the graphs of the two functions is at (2N + 1)π t ∼ 3π
2
,
3
giving tmax ∼ 4N .
Two such graphs, for N = 5, are given in plots linked from the page that gave
these notes. In the first, with the expanded scale, the intersection at t ∼ 0.13 can
be seen (MAPLE gives this point as 0.1304, to larger precision if desired), similar
to the crudely predicted value of 3/22 ∼ 0.1364. The second plot shows that the
tan (11 π t)
successive extrema occur essentially at the places where the graph of
11
diverges. (MAPLE’s plotting algorithm results in having these asymptotes more or
less drawn in by default.)
For N large, then,
 
4N  FN (0)  3π
FN (tmax ) ∼ − ,  FN (tmax )  ∼ 2 .
 
3π

3
We can interpret this as having the “peaks” in FN (t) be roughly proportional to
N , but with alternating signs, and coming closer together as N becomes large.
(This property is also seen from the cited MAPLE worksheet.) Thus, any integral
of FN (t), mulitplied by any well-behaved fuction, will have any contributions away
from the principle peaks at integral t “wash out,” leaving only the area in the region
of the principle peaks, shown above to be 1, times the value of the function at these
values. This is indeed the desired property of the δ-“function.”
To be slightly more rigorous, consider the prior derivation of the integral over
one period, but multiplied by a well-behaved (for our purposes, continuous) func-
tion f (t);

1/2 1/2
sin ((2N + 1)π t)
Z Z
f (t) FN (t) dt = 2 f (t) dt
−1/2 0 sin (π t)
(N+1/2)π
sin u
Z
=2 f (u/(2N + 1)π) du
0 (2N + 1)π sin (u/(2N + 1))
∞
2 sin u
Z
→ f (0) du
π 0 u
2 ∞ sin u
Z
= f (0) du
π 0 u
= f (0),

where the same substitution, u = (2N + 1)π t, has been made. This shows that
lim FN (t) has the proper integral property, even though FN does not converge
N→∞
uniformly to 0 for nonintegral t.