In Week 7, we saw that even if fn→f uniformly, f is not guaranteed to be differentiability at even a single point. This tells us that uniform convergence of fn is independent from the differentiability of f, so our conditions will need to look relatively different:
Theorem
Assume that fn,g:[a,b]→R are functions such that:
fn′ exists for all n∈N
fn′→g uniformly on [a,b]
there exists x0∈[a,b] such that {fn(x0)}n converges
Then fn converges uniformly to a function f on [a,b] and f is differentiable with f′=g on (a,b).
Remark.
The theorem tells us that under the conditions above, we can swap limits in the following:
The proof of the theorem is quite hard, which tells you that swapping limits is hard in general, so always take a second to think before you try swapping things!
This theorem is very useful when dealing with power series. For now, I'll illustrate its usefulness with a concrete example:
Example 1.
Let f(x)=∑n=0∞xn. Show that f is differentiable on (−1,1) and that
f′(x)=n=1∑∞nxn−1.
Solution.
Notice that the formula for f′ says
dxdn=0∑∞xn=n=0∑∞dxdxn.
To take a derivative, we need to take a limit, and to calculate an infinite sum, we need to take another limit. This tells us that even though it looks obvious, this fact is not trivial since we need to interchange two limits.
To show that f is differentiable on (−1,1), we can just show that f is differentiable on any strict subset (−r,r), where 0≤r<1. We want to restrict ourselves to a smaller set for two reasons:
Differentiability is a local property (i.e., f′(x) only depends on values of f really close to x), meaning we don't have to use the whole set to prove things about it.
Bad things happen close to the endpoints −1 and 1, so we want to avoid them.
Let fN(x)=∑n=0Nxn. We'll try to apply the theorem above to this sequence.
Because fN is a finite sum, we already know
fN′(x)=n=1∑Nnxn−1,
so (i) holds. To show (ii), we need to show that fN′ converges uniformly, i.e., we need to show that an infinite series of functions converges uniformly. This tells us that we'll want to use the Weierstrass M-test, so first calculate
x∈[−r,r]sup∣∣nxn−1∣∣=nrn−1.
To show that ∑n=1∞nrn−1 converges, we'll use the root test. Recall that limn→∞nn1=1, so by our limit laws,
(We also used the fact that rx is continuous here.) This step is also where working on the compact set [−r,r] pays off, since the root test would be inconclusive if we tried to use it on the whole interval. The root test tells us
n=1∑∞x∈[−r,r]sup∣∣nxn−1∣∣=n=1∑∞nrn−1<∞,
and so by the Weierstrass M-test, fN′ converges uniformly to g(x)=∑n=1∞nxn−1, so (ii) holds.
Finally, a power series always converges at its center, i.e., fN(0)=1, which converges as N→∞, so (iii) holds. By the theorem, f is differentiable on (−r,r) and
f′(x)=n=1∑∞nxn−1.
for any x∈(−r,r).
Given any x∈(−1,1), we can find 0≤r<1 (e.g., r=2∣x∣+1) such that x∈(−r,r), so we conclude that f is differentiable on all of (−1,1) with the same formula above.
Exercise 1.
With the same f as above, show that F(x)=∑n=0∞n+1xn+1 is an antiderivative for f.
Hint: Show that F′=f (as opposed to trying to show that ∫f(x)dx=F(x)+C). The proof should look mostly the same as the example.
Stone-Weierstrass
Theorem (Stone-Weierstrass)
Let (X,d) be a compact metric space and let A⊆C(X,R) be a subset such that:
1∈A
(A is an R-algebra) for any f,g∈A and λ∈A, we have λf+g,fg∈A
(A separates points) for any x=y in X, there exists f∈A such that f(x)=g(x)
Then A is dense in C(X,R).
This theorem is important because continuous functions are determined uniquely by their values on dense sets:
Proposition
Let f∈C(X,R) and assume that A is a dense subset of X (i.e., that A=X). If f(a)=0 for all a∈A, then f(x)=0 for all x∈X.
Proof.
Let x∈X. By density of A, there exists a sequence {an}n⊆A such that an→x. Thus, by continuity of f,
f(x)=f(n→∞liman)=n→∞limf(an)=n→∞lim0=0.
□
Corollary
Let f,g∈C(X,R) and assume that A is dense a dense subset of X. If f(a)=g(a) for all a∈A, then f(x)=g(x) for all x∈X.
Proof.
h=f−g∈C(X,R) and is 0 on a dense set, so h=0 on all of X, i.e., f(x)=g(x) for all x∈X.
□
Example 2.
Show that there is exactly one linear function F∈C(X,R), where X=C([0,1],R), such that F(x2n)=1 for all n≥0.
Solution.
To prove something like this, we need to show two things:
Existence: there is at least one F that works.
Uniqueness: there is at most one F that works.
To show existence, we can just give an example of a function. Define F:X→R via F(f)=f(1) (so F sends a continuous function f to a number by evaluating it at x=1). We need to show that F is linear and continuous: let f,g∈X and let λ∈R. Then
i.e., F is Lipschitz, hence continuous (by a old homework problem). This proves existence.
To show uniqueness, let G be another such function, i.e., G is a continuous linear function such that G(x2n)=1 for all n≥0. We need to show that F=G on all of X. Notice that
F(n=0∑Nanx2n)=n=0∑NanF(x2n)=n=0∑NanG(x2n)=G(n=0∑Nanx2n),(F is linear)(F(x2n)=1=G(x2n))(G is linear)
so F=G on the set
A={n=0∑Nanx2n∣∣N∈N,an∈R}⊆X.
If we can show that A is dense in X, then by the corollary above, we get F=G on all of X. To show that, we can use Stone-Weierstrass:
(i) holds since if n=0, then 1=x2n∈A. To show (ii), let f(x)=∑n=0Nanx2n, g(x)=∑m=0Mbmx2m, and λ∈R. Without loss of generality, we may assume that N≤M (otherwise, we can just swap the roles of f and g), so we get
so A is an R-algebra. Finally, for (iii), notice that x2∈A is injective, so if x=y in [0,1], then x2=y2. Thus, A separates points.
By Stone-Weierstrass, A is dense in X, so F=G on all of X, which proves uniqueness.
True or False
Example 3.
True or false: If fn→f uniformly on compact sets of R, then fn→f uniformly on all of R.
Solution.
This is false. A simple example is a moving step function, i.e.,
fn(x)={10if x∈[n,n+1],otherwise.
Any compact set K⊆R is bounded, so there exists N∈N such that supK<N. Thus, fn(x)=0 on K for all n≥N, so fn→f uniformly on K, hence on any compact set. However,
x∈Rsup∣fn(x)−0∣=1,
so fn does not converge uniformly to 0 on all of R.