Physics 116A Winter 2010

Pointwise and Uniform Convergence

A power series,
∞
X
f (x) = an xn ,
n=0

is an example of a sum over a series of functions

∞
X
f (x) = fn (x) , (1)
n=0

where fn (x) = an xn . It is useful to consider the more general case. Let us consider
a sum of the form given in eq. (1) and ask whether the sum is convergent. If we
consider each x separately, then we can determine whether the sum converges at
the point x. Suppose that the sum is determined to converge for all points x ∈ A,
where A is some interval on the real axis. Typical intervals are: the open interval
a < x < b, which we will denote by (a, b) and the closed interval a ≤ x ≤ b, which
we will denote by [a, b]. Of course, we could also consider half-open, half-closed
intervals, such as a < x ≤ b, denoted by (a, b] and a ≤ x < b, denoted by [a, b). In
this notation, the parenthesis indicates that the endpoint is not included in the
interval, whereas the square bracket indicates that the endpoint is included in the
interval.
If f (x) converges for all x ∈ A, we say that the sum given by eq. (1) is
pointwise convergent over the interval x ∈ A. In this case, A is called the interval
of convergence. A classic example is the infinite geometric series,
∞
1 X
= xn , |x| < 1 . (2)
1 − x n=0

The above sum converges pointwise over the open interval −1 < x < 1 to the
function (1 − x)−1 . Using the standard procedures, it is easy to see that the sum
diverges for all |x| ≥ 1.
Let us return to the general case of
∞
X
f (x) = fn (x) , x ∈ A,
n=0

where A is the interval of convergence. Suppose that the fn (x) are continuous
functions. Does this imply that f (x) is continuous? The great mathematician
Augustin Louis Cauchy got the answer wrong. In 1821, he claimed to prove that

1
all infinite sums of continuous functions are continuous. It took over 30 years
before the error was properly corrected.
It is simple to provide a counterexample to Cauchy’s claim. Consider the
series: ∞
X x2
S(x) = . (3)
n=1
(1 + nx2 )[1 + (n − 1)x2 ]
Although this series looks complicated, we can simplify it using partial fractions.
The following is an algebraic identity:

x2 1 1
= − .
(1 + nx2 )[1 + (n − 1)x2 ] 1 + (n − 1)x2 1 + nx2

Thus,
N
x2
   
X 1 1 1
= 1− + − +···
n=1
(1 + nx2 )[1 + (n − 1)x2 ] 1 + x2 1+x2 1 + 2x2
 
1 1
+ −
1 + (N − 1)x2 1 + Nx2

1 Nx2
=1− = .
1 + Nx2 1 + Nx2
Then, (
Nx2 1, for x 6= 0 ,
S(x) = lim =
N →∞ 1 + Nx2 0, for x = 0 .
That is, f (x) is discontinuous at x = 0. Below, I have plotted Nx2 /(1 + Nx2 ) for
N = 2, 4, 6, 8 and 10. As N increases, the curves begin to approach the function
f (x), which is equal to 1 for all x 6= 0. Of course, S(0) = 0, so in the limit of
x → ∞, the functions exhibits a discontinuity.
1.0

0.8

0.6

0.4

0.2

-10 -5 5 10

2
 In this case, the interval of convergence is the entire x-axis, i.e., −∞ < x < ∞.
To formally prove that the sum is pointwise convergent, we define
N
X x2 Nx2
SN ≡ = .
n=1
(1 + nx2 )[1 + (n − 1)x2 ] 1 + Nx2

Then for x 6= 0,
1
|S(x) − SN (x)| = < ǫ, (4)
1 + Nx2
which implies that
1−ǫ
N> . (5)
ǫx2
That is, given an ǫ no matter how small, one can always find an N such that
|S(x) − SN (x)| < ǫ. This implies that for x 6= 0, lim SN (x) = S(x). Note that as
N →∞
x → 0, the value of N required grows arbitrarily large. This means that the rate
of convergence is getting slower and slower the closer x is to the origin. In some
sense, this is the source of the ultimate discontinuity at x = 0.
These observations motivate the following definition.
N
X
Definition: Given SN (x) = fn (x) and lim SN (x) = S(x), where x ∈ A, the
N →∞
n=1
sum is uniformly convergent in the interval x ∈ A if given a positive error bound ǫ,
there always exists a response N such that n ≥ N implies that |S(x)−Sn (x)| < ǫ.
The same N must work for all x ∈ A.

The concept of uniform convergence is illustrated in the above figure, taken

from ref. 2. We see that the approximation fn to f (in our notation, fn is SN (x)
and f is S(x) above) lies completely within an ǫ-band over the entire interval.

3
 By comparing the definition of uniform convergence with eqs. (4) and (5), it
is clear that the sum given in eq. (3) is pointwise convergent but is not uniformly
convergent over any interval that contains the point x = 0. Indeed, uniform con-
vergence is a more stringent requirement than pointwise convergence. However,
the advantage of uniform convergence is that the properties of the functions fn (x)
(such as continuity) are preserved by the infinite sum. We quote a few key theo-
rems without proofs (for details, see e.g., refs. 1 and 2 given at the end of these
notes).
∞
X
Theorem 1: If f (x) = fn (x) converges uniformly over the interval x ∈ A,
n=1
and if the fn (x) are continuous at every point x ∈ A, then f (x) is also continuous
at every point x ∈ A.
Since S(x) defined in eq. (3) is discontinuous at x = 0, it follows that S(x) is
not uniformly convergent over any interval that includes the point x = 0. However,
the converse to Theorem 1 is false. In particular, there are many examples of non-
uniformly convergent sums that are continuous over the interval of convergence.
∞
X
Theorem 2: If f (x) = fn (x) converges uniformly over the open interval
n=1
x ∈ A, and if the fn (x) are differentiable at every point in A, then f (x) is also
∞
X
differentiable at every point x ∈ A, and f (x) =
′
fn′ (x) (where f ′ ≡ df /dx).
n=1

That is, it is legal to interchange the order of summation and differentiation.

While this interchange is clearly valid for finite sums, it requires proof if the
sum is over an infinite number of terms. For non-uniformly convergent sums,
interchanging the order of an infinite summation and differentiation may fail. An
example of this will be given in a separate note.
∞
X
Theorem 3: If f (x) = fn (x) converges uniformly over the closed interval
n=1
x ∈ [a, b], and if the fn (x) are integrable over the interval [a, b], then f (x) is also
integrable over the interval [a, b], and
Z b ∞ Z
X b
f (x)dx = fn (x)dx .
a n=1 a

That is, it is legal to interchange the order of summation and integration.

While this interchange is clearly valid for finite sums, it requires proof if the
sum is over an infinite number of terms. For non-uniformly convergent sums,
interchanging the order of an infinite summation and integration may fail. An
example of this will be given in a separate note.

4
 ∞
X
The Weierstrass M -test: If f (x) = fn (x), and if |fn (x)| ≤ Mn for every
n=1
∞
X
n greater than or equal to some fixed integer N, for all x ∈ A, and if Mn
n−1
∞
X
converges, then fn (x) converges uniformly to f (x) over the interval x ∈ A.
n=1
Moreover, at every point x ∈ A, the convergence is absolute.
This is the analog of the comparison test for numerical series. The converse of
the Weierstrass M-test is false. Namely, it is possible for a uniformly convergent
series to fail the Weierstrass M-test. An example of this is
∞
X (−1)n+1
. (6)
n=1
x2 + n

Clearly, this sum is only conditionally convergent. Thus, it cannot pass the Weier-
strass M-test, since any series that satisfies the Weierstrass M-test must be ab-
solutely convergent. Nevertheless, one can prove that eq. (6) is also uniformly
convergent.
Finally, we can now understand why power series are so nice. As a consequence
of the Weierstrass M-test, we have this important result.
Theorem 4: If f (x) = ∞ n
P
n=0 an x is a power series with a finite radius of conver-
gence R > 0, then for any 0 < r < R, this power series converges uniformly and
absolutely over the closed interval [−r, r]. If the radius of convergence is infinite,
then the power series converges uniformly over the closed interval [−r, r] for any
finite positive value of r.
Proof: For any |x| ≤ r, we have |an xn | ≤ |an |r n . But, ∞ n
P
n=0 |an |r converges
absolutely (as a consequence of the ratio test). Hence the Weierstrass M-test
applies.
As a result, f (x) is continuous over any closed interval [r, r] (for r < R), and
one can differentiate and integrate power series by differentiating or integrating
each term in the series. Returning to the example of the infinite geometric series,
we conclude that eq. (2) is uniformly convergent over any closed interval [−r, r],
where 0 < r < 1. However, it would be incorrect to claim that eq. (2) is uniformly
convergent over the open interval (−1, 1). The problem here is that the rate of
convergence is infinitely slow as x → ±1. This is not surprising, since eq. (2)
diverges at x = ±1. P∞ n
For the general power series f (x) = n=0 an x with radius of convergence
R, the series converges for |x| < R and diverges for |x| > R. However, the
convergence properties at x = ±R must be checked separately. The following
theorem is relevant.

5
 P∞ n
Theorem 5: Given a power series f (x) = n=0 an x with a finite radius of
convergence R > 0 and an interval of convergence A. Then,

1. If A = [−R, R], then the series converges uniformly (but not necessarily
absolutely) over the closed interval [−R, R].

2. If A = (−R, R], then the series converges uniformly (but not necessarily
absolutely) over any closed interval [a, R] for all −R < a < R.

3. If A = [−R, R), then the series converges uniformly (but not necessarily
absolutely) over any closed interval [−R, b] for all −R < b < R.

4. If A = (−R, R), then the series converges uniformly and absolutely over any
closed interval [a, b] for all −R < a < b < R.

The infinite geometric series [eq. (2)] is an example of case 4 of Theorem 5.

An example of case 2 is
∞
X xn
ln(1 + x) = (−1)n+1 , −1 < x ≤ 1 .
n=1
n

This sum is conditionally convergent at x = 1 and divergent at x = −1. Thus,

we conclude that the series converges uniformly on any closed interval a ≤ x ≤ 1
where −1 < a < 1. In particular, the sum is uniformly convergent at x = 1.
Our final example is Euler’s dilogarithm, which is defined by its power series
∞
X xn
Li2 (x) = , |x| ≤ 1 .
n=1
n2

The radius of convergence is 1, and the series converges uniformly and absolutely
at x = ±1. In fact, one can show that:

π2 π2
Li2 (1) = , Li2 (−1) = − .
6 12
This is an example of case 1 of Theorem 5.

References:
1. David, Bressoud, A Radical Approach to Real Analysis (Mathematical Associ-
ation of America, Washington, DC, 1994).
2. Brian S. Thomson, Judith B. Bruckner and Andrew M. Bruckner, Elementary
Real Analysis (Prentice-Hall, Inc., Upper Saddle River, NJ, 2001).