Last time we worked on convolution and showed how to approximate an integrable function by continuous function by convolution. The following code shows the approximation process in problem 4 of problem set 1. It plots Π∗Λ1/n for n=5,10,20.
x = PolynomialRing(QQ,'x').gen()
T = piecewise([[(-1,0),1+x],[(0,1),1-x]])
f = piecewise([[(-1/2,1/2),1]])
P = plot(f)
Q = plot(T)
#N=1
#TN=piecewise([[(-1/N,0),N*(1+x*N)],[(0,1/N),N*(1-x*N)]])
#TNF = f.convolution(TN)
#P=plot(TN)
for N in [5,10,20]:
TN=piecewise([[(-1/N,0),N*(1+x*N)],[(0,1/N),N*(1-x*N)]])
Q = Q + plot(TN,color='red')
TNF=f.convolution(TN)
P=P+plot(TNF)
P.show(title='convolution')
Q.show(title='lambda n')
The idea to do pointwise approximation is, suppose we want to approximate f(x) by a sequence f∗gn. We will have to control ∣f(x)−f∗gn(x)∣=∣∫01f(x−y)g(y)dy−f(x)∣. With assumption ∫01gn(y)dy=1 for every n, we can write the RHS as ∣∫01(f(x−y)−f(x))gn(y)dy∣≤∫01∣f(x−y)−f(x)∣∣gn(y)∣dy.
When in particular gn has very small support (−δ,δ), the integral is taken on y∈(−δ,δ), so that f(x−y) is a small perturbation of f near x. For example, when f is continuous at x, the term ∣f(x−y)−f(x)∣ will be small.
Let (Kn)n=1∞ be a family of kernels. It is called good provided that it satisfies the following properties:
(1) ∫01Kn(x)dx=1 for every n.
(2) For every δ>0, ∫(0,1)∖(−δ,δ)∣Kn(x)∣dx→0.
(3) Either Kn≥0, or ∥Kn∥L1<M for some M>0 and all n.
A good family of kernels is also called an “approximation to the identity”.
The Λn satisfies (1) has support shrinking to {0}, so it is automatically a good kernel. But the Fejer kernel and Dirichlet kernel has no “shrinking support”, as can be seen from appendix of lecture 1.
The notion of “good kernel” is designed so that the following holds:
Theorem. Let (Kn)n=1∞ be a good family of kernels. Let f∈L1(T), then
If f is continuous at x, then limn→∞(f∗Kn)(x)=f(x).
If f is continuous everywhere on T, the limit is uniform.
For ϵ>0 choose δ such that when y∈(−δ,δ), ∣f(x−y)−f(x)∣<ϵ, so ∫−δδ∣f(x−y)−f(y)∣∣Kn(y)∣dy≤ϵ∫−δδ∣Kn(y)∣dy≤ϵM⋅2δ.
The other part of integration is
∫(0,1)∖(−δ,δ))∣f(x−y)−f(x)∣∣Kn(y)∣dy≤2supx∈[0,1]∣f(x)∣⋅∫(0,1)∖(−δ,δ)∣Kn(y)∣dy→0 when n→∞ by assumption (2) of good kernel.
We conclude that when n is large enough (depend on x), ∣f∗Kn(x)−f(x)∣≤∫01∣f(x−y)−f(x)∣∣Kn(y)∣dy<Cϵ for some constatn C>0.
If f is everywhere continuous on T, then it is uniformly continuous on T, so the δ in the proof can be chosen independent of x, so the “when n large enough part” no longer depend on x, this implies uniform convergence. □
Then σn(f)(x)=f∗FN(x). The FN is called Fejer kernel.
Similar to the Dirichlet kernel, the Fejer kernel has also a simpler form.
Lemma.FN(x)=N1sin2(πx)sin2(πNx).
Proof. In problem set 2.
Lemma. The Fejer kernel is a good kernel.
Proof.
Note that since FN≥0, it sufficies to verify (1) and (2) in the definition of good kernel.
To verify (1), ∫−2121FN(x)dx=∫01ND0(x)+⋯+DN−1(x)dx=NN=1.
To verify (2), let Aδ denote the set [−21,21]∖(−δ,δ). Observe that for δ>0, FN→0 uniformly on Aδ. Indeed, choose cδ such that sin2(πx)≥cδ, so FN(x)≤N1⋅cδ1 on Aδ, which goes to 0 when N→∞. By uniform convergence limN→∞∫AδFN=∫AδlimN→∞FN=0.
As a consequence of goodness of FN, we obtain the Fejer’s theorem
Theorem.f∈L1(T). Then
σN(f)(x)→f(x) when f is continuous at x.
If f is continuous on T, then the convergence is uniform.
In particular, we have showed that every continuous function on T can be uniformly approximated by trignomic polynomials, where the trignomic polynomials can be taken as its Cesaro sums. Later, we shall use this result to prove the L2 convergence of Fourier series.
A family of good kernels are also called “an approximated identity”, because it is really an “approximation” to the “identity”. An “identity” element usually means something similar to 1, multiplying with the identity element will return itself. In our case, the identity is the “identity” of convolution. Imagine a “function” I such that f∗I=f, what shall the I be? By the convolution theorem, f^(n)=f∗I(n)=f^(n)I^(n), so one has I^(n)=1 for every n. Then the Fourier series of I should look like ∑n=−∞+∞e2πinx, so I should be something like “limN→∞DN”. With the notion of distributions and measures we can make this discussion precise and the “identity” will be the δ-measure, and an approximate identity is an approximation to the δ-measure in weak sense.
The “goodness” of good kernels is not restricted to continuous category. One can show for example that if f∈Lp(T), then σnf→f in Lp. See Hoffman’s book chapter 2 for more general results in this direction.