Divergent Series Sum
Divergent Series Sum
Divergent Series Sum
Bryden Cais
“Divergent series are the invention of the devil, and it is shameful to base on them any demonstration
whatsoever.”—N. H. Abel
1. Introduction
P∞
The notion of convergence of a series is a simple one: we say that the series n=0 an converges if
N
X
lim an
N →∞
n=0
exists and is finite. So for example the series
1 1 1 1
1+ + + + + ···
2 4 8 16
1 1 1 1
1 − + − + − ···
2 3 4 5
P
both converge (to 2 and log 2, respectively). If a series an does not converge, it is said to diverge. Two prototypical
examples of divergent series are
1 + 2 + 3 + 4 + 5 + ···
1 − 1 + 1 − 1 + 1 − ··· ,
where the first series diverges because the partial sums tend to +∞ and the second series diverges because the
partial sums sN do not tend to any limit (even though lim s2N = 0 and lim s2N −1 = 1).
One might think that not much can be said for divergent series. The goal of this these notes is to show that this is
not the case, and that divergent series are in fact an interesting object of study. Historically, divergent series occur
in the work of Euler, Poisson, Fourier, and Ramanjuan (among many others), and although it was not until Cauchy
that the definitions of convergence were formally stated, these masters knew well enough when a series converged
and when it did not. Part of the reason that divergent series were so abhorred by mathematicians after Cauchy
is because no one formally defined what the sum of a divergent series should be. Rather, mathematicians sought
for some intrinsic meaning and rapidly found themselves in rather difficult terrain. We will begin by highlighting
some of the problems that arise when one tries to make sense of a divergent series without clear definitions, and
this will lead naturally to two generalizations of the notion of “sum,” both of which can be used to assign meaning
to divergent series.
Problems arise, however, as soon as one tries to rearrange any terms. For example, we might try to argue that
s = 1 − 1 + 1 − 1 + · · · = (1 − 1) + (1 − 1) + (1 − 1) + · · · = 0 + 0 + 0 + · · · = 0,
but such arguments can be dismissed on the grounds that even for convergent series rearranging terms can spell
disaster. Thus, for example, one can rearrange terms of the series
1 1 1 1
1 − + − + − ···
2 3 4 5
to obtain any real number as the sum! As for the second argument, one might object (as Callet did) that for m < n
we have
1 − xm
= 1 − xm + xn − xm+n + x2n − · · · ,
1 − xn
so that by l’hopital’s rule
1 − xm m
1 − 1 + 1 − 1 + · · · = lim = .
x→1− 1 − xn n
But here again, it is not so difficult to see what is happening: the limit as x → 1− of 1 − xm + xn − xm+n + · · · is
not the series
1 − 1 + 1 − 1 + ··· ,
but rather the series
1 + 0 + 0 + · · · + 0 −1 + 0 + 0 + · · · + 0 +1 + · · · ,
| {z } | {z }
m−1 zeroes n−m−1 zeroes
and there is no reason why these two series should have the same value.
For a different kind of “problem,” consider the series
s = 1 + 2 + 3 + 4 + 5 + ··· .
What should be the value of s? Laying aside for the moment our reservations, we have
−3s = (1 − 4)s = (1 + 2 + 3 + 4 + 5 + · · · ) − 2(2 + 4 + 6 + · · · )
= 1 − 2 + 3 − 4 + 5 − ···
= 1 − (2 − 3 + 4 − 5 + · · · )
= 1 − (1 − 2 + 3 − 4 + 5 − · · · ) − (1 − 1 + 1 − 1 + · · · )
= 1 + 3s − 1/2,
we conclude that −6s = 1/2, that is
1
1 + 2 + 3 + 4 + 5 + ··· = − .
12
Such a statement obviously presents philosophical difficulties. Namely, one is forced to ask how the “sum” of a
divergent series of entirely positive terms can be negative. Yet the manipulations involved in our determination of
s are no more outlandish than those used in determining 1 − 1 + 1 − 1 + · · · = 1/2. We will see later that in a very
precise sense, −1/12 is the correct value of 1 + 2 + 3 + 4 + · · · .
P
(1) (an + bn ) = s + t. P
(2) For any real number α, we require αan = αs.
These notions can be formulated in a precise way. Let V denote the infinite-dimensional real vector space consisting
of sequences of real numbers (an ), and let W ⊆ V be the subspace of convergent sequences. Then we can think of
X
: W −→ R
P P∞
as the linear operator defined by ((an )) := P n=1 an . In seeking to generalize the idea of “sum” to divergent
series, we wish to P extend the linear operator to subspaces V 0 ⊆ V strictly containing W . We might hope to
be able to extend to all of V , but as we shall see, this is a somewhat ambitious goal. Our generalizations are
classical, and motivated by the calculations of the preceding section.
3.1. Cesaro summability. Cesaro summation formalizes our claim that 1 − 1 + 1 − 1 + · · · should be 1/2 as
this is the average value of the partial sums. Concretely, let (an ) be any sequence of real numbers and put
sN = a1 + a2 + · · · + aN . Let Vc,1 be the subset of V for which
s1 + s2 + · · · + sN
lim
N →∞ N
exists and define the map C1 : Vc,1 −→ R by
s1 + s2 + · · · + sN
C1 ((an )) = lim .
N →∞ N
It is easy to see that Vc,1 is actually a subspace of V , and not just a subset. Moreover, suppose that a1 + a2 + · · ·
converges, say to L, so that for any > 0 there exists N > 0 such that |L − sM | < for all M > N . Then we have
s1 + s2 + · · · + sN sN +1 + · · · + sN +k
C1 ((an )) = lim +
k→∞ N +k N +k
sN +1 + · · · + sN +k
= lim .
k→∞ N +k
But since |sM − L| < for all M > N , we have the bounds
k(L − ) sN +1 + · · · + sN +k k(L + )
L − = lim ≤ lim ≤ lim = L + ,
k→∞ N +k k→∞ N +k k→∞ N + k
P
and since > 0 was arbitrary, this shows that C1 ((an )) = n≥1 an when this sum converges. Thus, Vc,1 ⊇ W and
P
C 1 W = .
It is clear that C1 is a linear map. We can say more, though. Indeed, suppose that C1 ((a1 , a2 , a3 , . . .)) = s.
Then we claim that for any finite j > 0 we have
j−1
X
(1) C1 ((aj , aj+1 , . . .)) = s − an .
n=1
Indeed, let sN denote the N th partial sum of the sequence (a1 , a2 , . . .) and tN the N th partial sum of (aj , aj+1 , . . .),
Pj−1
and put S := n=1 an . Then tN = sj+N −1 − S and
t1 + t2 + · · · + tN sj + sj+1 + · · · + sj+N −1 − N S
lim = lim
N →∞ N N →∞
N
N + j − 1 s1 + s2 + · · · + sj+N −1 s1 + s2 + . . . sj
= lim · − −S
N →∞ N j+N −1 N
s1 + s2 + · · · + sN
= lim − S,
N →∞ N
which proves our claim.
We see almost immediately that the containment Vc,1 ⊇ W is strict, for the sequence (1, −1, 1, −1, . . .) has partial
sums (
1 if N is odd
sN =
0 if N is even
4
3.2. Abel summability. Now we touch briefly on another important method of defining the sumPof a divergent
series. Let Va ⊆ V be the subset of all sequences (an ) such that the corresponding power series n≥1 an xn has
radius of convergence 1, and represents a function f (x) such that limx→1− f (x) exists. We observe right away that
we have the containment W ⊆ Va . We defineP a map A : Va −→ R by A((anP )) = limx→1− f (x). It is clear that this
is a linear map. Moreover, suppose that an is convergent.
P Then f (x) := an xn has radius of convergence 1 and
by a Theorem P of Abel, we have limx→1− f (x) = an , so that the linear operator A, when restricted to W ⊆ Va
agrees with .
It is clear from our earlier examples that the containment W ⊆ Va is strict. For we saw that
1 1
A((1, −1, 1, −1, . . .)) = lim = .
x→1− 1+x 2
Observe that our definition makes it impossible to claim that the sum of 1 − 1 + 1 − 1 + · · · could be m/n for
any m < n on grounds of the identity
1 − xm
= 1 − xm + xn − xm+n + x2n − · · · ,
1 − xn
5
because as we have pointed out, this identity does not correspond to the sequence (1, −1, 1, −1, . . .), but to the
sequence (1, 0, 0, . . . , 0, −1, 0, 0, . . . , 0 , 1, . . .)
| {z } | {z }
m−1 zeroes n−m−1 zeroes
In the next section we will reconstruct Euler’s derivation of the functional equation for the Riemann zeta function
using the linear operator A.
with the product being over all primes, and used this to give an analytic proof of the infinitude of primes. This
viewpoint ultimately enabled Dirichlet to prove the infinitude of primes in arithmetic progressions.
One of the most important properties of the zeta function is that it satisfies a “functional equation” that relates
ζ(s) to ζ(1 − s):
sπ
(2) ζ(1 − s) = 2(2π)−s cos Γ(s)ζ(s),
2
where Γ(s) is a function of the complex variable s satisfying Γ(s) = (s − 1)! for integers s ≥ 1. This functional
equation—which exhibits the symmetry of the ζ-function about the line s = 1/2—played a large role in Riemann’s
formulation of his infamous “hypothesis.” What is fascinating is that Euler conjectured (2) based on calculations
with divergent series over 100 years before Riemann wrote his influential paper on the zeta function. We now
reconstruct Euler’s arguments.
Euler works exclusively with the linear operator A. He begins with the power series
1
e−y − e−2y + e−3y − e−4y + · · · = ,
ey +1
which converges for all y > 0. By differentiating n times, he obtains
dn
1
1n e−y − 2n e−2y + 3n e−3y − 4n e−4y + · · · = (−1)n ,
dy n ey + 1
which again converges for any y > 0.
Now the function 1/(ey + 1) can be expanded as a Taylor series about y = 0. Specifically, we have
∞
1 X
(3) = ak y k
ey + 1
k=0
We observe that since tan z is an odd function of z we necessarily have a0 = 1/2 and a2k = 0 for all k > 0.
Combining (5) and (6) and setting z = 0, we get, for odd values of n
∞
X 1
(7) 2 n+1
= an (−πi)n+1 .
m=0
(2m − 1)
But ∞
1 1 X 1
2(1 − 2−n−1 ) 1 + + + ··· =2 = an (−πi)n+1 ,
2n+1 3n+1 m=0
(2m − 1)n+1
so that
an (−πi)n+1
ζ(1 + n) = ,
2(1 − 2−n−1 )
which since n is odd we may write as
an (−π)n+1
n+1
(8) ζ(1 + n) = cos π .
2(1 − 2−n−1 ) 2
On the other hand, we computed
(1 − 2 · 2n )ζ(−n) = A((1n , −2n , 3n , −4n , . . .))
= (−1)n n!an ,
whence
(−1)n n!an
(9) ζ(−n) = ,
1 − 2n+1
from which we can immediatly conclude that ζ(−2k) = 0 for k > 0 since we saw above that a2k = 0 for k > 0.
Thus, combining (8) and (9) we get
n+1 n+1
2n!ζ(n + 1) = (2π) cos π ζ(−n),
2
or by replacing n with n − 1 and rearranging,
nπ
(10) ζ(1 − n) = 2(2π)−n cos (n − 1)!ζ(n).
2
7
Observe that even though this a priori holds only for even n, we can see that it is also true for odd n ≥ 3 as both
sides are zero in these cases. Now Euler reasons as follows: if we replace (n − 1)! with Γ(n) then both sides of (10)
make sense for non-integral values of n. He thus conjectures that for all complex s for which both sides make sense,
one has
sπ
(11) ζ(1 − s) = 2(2π)−s cos Γ(s)ζ(s).
2
√
He then proceeds to verify this equation for s = 1/2 by using the fact that Γ(1/2) = π.
References
[1] Hardy, G.H. Divergent Series. Clarendon Press, Oxford. 1949.