CAAM/Stat 31460 Applied Fourier Analysis, Winter 2024

Homework 3

Instructions: Same as for earlier assignments. Reminder: In all cases where there is collaboration or
external resources are consulted or used, proper citation must be given.
Due date: Friday, Feb. 23, 3pm, by online Gradescope submission.
Reading Assignment:
• Sections 3.4 and 7.1–7.3 of Folland, together with the discussion of the discrete Fourier transform in
Section 7.6.
• Sections 3.1–3.3 of Mallet.
• Optional supplemental reading:
– Chapter 2 of Mallet.
– You may also find it interesting to take look at the survey article by R. Osserman, The isoperi-
metric inequality (Bull. Amer. Math. Soc. 84 (1978), no 6, 1182–1238) posted on Canvas
(under Files Reference Material). A brief overview of the proof we gave in class is discussed
in Section 1, with a few references. While there have also been more recent developments, the
article of Osserman gives a nice overview of many important classical and contemporary ideas
in the subject.
– Another survey article relevant to some of the applications we’ve discussed is G. Folland and A.
Sitaram, The uncertainty principle: A mathematical survey (J. Fourier Anal. Appl. 3 (1997),
no. 3, 207–238), also posted to Canvas (in the same Files section).

Written Assignment:

(1) We say that a function f : R → R is Hölder continuous of order α ∈ (0, 1) if there exists a constant
C > 0 such that
|f (x) − f (y)| ≤ C|x − y|α
for all x, y ∈ R. In this exercise, we will see how the Hölder continuity of a 2π-periodic function f
implies a decay property for the Fourier coefficients.
Suppose that f : R → R is 2π-periodic and integrable on [−π, π].

(a) Let (cn ) denote the Fourier coefficients of f . Show that

Z π
1
cn = − f (x + (π/n))e−inx dx,
2π −π
and thus that Z π 
1 
cn = f (x) − f (x + (π/n)) e−inx dx.
4π −π
(b) Show that if f in addition satisfies a Hölder condition of order α ∈ (0, 1), then there exists
C > 0 such that
|cn | ≤ C/|n|α
for all n ∈ Z.
(c) Show that the decay rate in (b) cannot be improved, by showing that for fixed α ∈ (0, 1), the
function f : R → R defined by
∞
X n
f (x) = 2−nα ei2 x ,
n=0

for x ∈ R satisfies a Hölder condition of order α, while its Fourier coefficients (cn ) satisfy
cm = m−α for m = 2−n , n ≥ 0.
(Hint: To show the Hölder condition, write
X X
f (x) − f (y) = ··· + ··· .
{n:2n ≤|x−y|−1 } {n:2n >|x−y|−1 }
 To estimate the first sum, use |1 − eiθ | ≤ C|θ| for |θ| small (if you use this, explain why it
holds!); to estimate the second sum, use |eix − eiy | ≤ |eix | + |eiy | = 2.)

(2) Even with the example of 1(c) above in mind, a bit more can be said when α > 1/2. Suppose that
f : R → R is a 2π-periodic function which is Hölder continuous of order α for some α > 1/2, and
let (cn ) denote the Fourier coefficients of f . A classical resultP of Bernstein asserts that under these
inx
P
hypotheses the Fourier series of f converges absolutely, i.e. n |cn e | = n |cn | < ∞ (and so,
in particular, the Fourier series of f converges uniformly to f on R in this case). Follow the steps
below to prove this result:
(a) Show that there exists C > 0 such that for all h > 0
∞
!1/2
X
inh 2 2
|e − 1| |cn | ≤ Chα .
n=−∞

(Hint: First observe that for all h > 0, gh : R → R defined by gh (x) = f (x + h) − f (x) for
x ∈ R is a 2π-periodic continuous function, and thus in particular belongs to L2 (−π, π). Now,
compute the Fourier coefficients of gh and use Parseval’s equation.)
(b) Show that for all k ≥ 0, if n ∈ Z satisfies 2k ≤ |n| < 2k+1 , then one has
k
|e2inπ/(3·2 )
− 1| ≥ 1.
(c) Use (a) and (b) to show that there exists C > 0 (potentially depending on α) such that for all
k ≥ 0,
X C
|cn |2 ≤ 2kα .
k k+1
2
2 ≤|n|≤2

(Hint: For each k ≥ 0, apply (a) with h = 2π/(3 · 2k ) and invoke (b).)
(d) Show that
 1/2
∞
X ∞
X X
|cn | ≤ |c0 | + 2(k+1)/2  |cn |2  .
n=−∞ k=0 2k ≤|n|<2k+1
P∞ P∞ P
(Hint: Write n=−∞ |cn | = |c0 | + k=0 2k ≤|n|<2k+1 |cn | and use Cauchy-Schwarz.)
P∞
(e) Combine (c) and (d) to show that α > 1/2 implies n=−∞ |cn | < ∞.1

(3) Folland, pg. 80, exercises 6, 9, and 11.

(4) Some practice with the space L2 :

(a) Folland, pg. 86, exercises 6 and 7.
(b) Folland, pg. 213, exercises 2 and 8.

(5) Folland, pg. 224–225, exercises 4, 7 and 13(a).

(6) Fix f ∈ L2 , and let (fn ) and (gn ) be sequences of functions such that fn ∈ L1 ∩ L2 for all n ≥ 1,
gn ∈ L1 ∩ L2 for all n ≥ 1, and
kfn − f kL2 → 0 and kgn − f kL2 → 0

1
PTo recover f as the limit, one way to argue
P is as follows: note that, as we discussed in class, it follows from the convergence
inx converges uniformly to a continuous limit g. On the other hand,
of |cn | and the Weierstrass M test that |n|≤N cn e
since f belongs to L2 (−π, π), |n|≤N cn einx converges to f in the L2 norm – it now follows from the result stated as Theorem
P

3.2 in Folland that the partial sums converge to g in the L2 norm as well, so the uniqueness of limits in L2 implies that f = g
almost everywhere. Since f and g are both continuous, we obtain that f (x) = g(x) for all x ∈ R, and thus the partial sums
converge uniformly to f .
 as n → ∞. Show that if
kfˆn − F kL2 → 0 and kĝn − GkL2 → 0
as n → ∞ then F = G a.e. on R.
(Hint: Use Parseval’s identity to note that kfˆn −ĝn kL2 = kfn −gn kL2 ; now try to estimate kF −GkL2 .)

(7) Folland, pg. 234–236, exercises 1, 2, 6, 8, 9, 10.

(Note: There is a misprint in the statement of 9(b) in some editions – the correct text is “Show that
F̂ (ξ) = eia(ξ+α) fˆ(ξ + α) and thense that ∆α fˆ = ∆0 F̂ .” Similarly, some editions have a misprint in
line 2 of exercise 10, where the ODE should read f 0 (x) + cxf (x) = 0. For a full list of misprints in
various editions, see https://sites.math.washington.edu/~folland/Homepage/oldfourier.pdf
and https://sites.math.washington.edu/~folland/Homepage/fourier.pdf.)
Suggested Additional Problems (not collected or graded):

(1) Folland, pp. 212–213, exercises 1, 7.

(2) Working with convolutions: Folland, pg. 213, exercises 3, 4, and 5.
(3) Folland, pg. 224–225, exercises 1, 3, 5, 6, 14.
(4) Consider u solving the wave equation utt −uxx = 0 for u = u(t, x) : R×R → R, with initial conditions
u(x, 0) = f (x) and ut (x, 0) = g(x).
(a) Assuming that all the Fourier transforms that appear exist (you may also assume this in the
remainder of the problem), show that
sin(tξ)
û(ξ, t) = fˆ(ξ) cos(tξ) + ĝ(ξ) ,
ξ
where û(·, t) = Fx [u(x, t)], i.e. the “hat” denotes the Fourier transform in the spatial variable
only.
(b) Show that the energy
Z
1
E(t) := |ut (x, t)|2 + |ux (x, t)|2 dx
2 R
satisfies Z
1
E(t) = |ξ|2 |fˆ(ξ)|2 + |ĝ(ξ)|2 dξ
4π
for all t ∈ R, and is thus independent of t. (Hint: Use Plancherel.)
(c) Suppose E(0) < ∞. Show that
Z
1 1
|ut (x, t)|2 dx → E(0)
2 2
as t → ∞. (Hint: One approach is to combine a calculation using the Plancherel identity as in
(b) with the formula from (a). Now use the Riemann-Lebesgue lemma.)
(d) Invert the Fourier transform from (a) to obtain the usual d’Alembert formula for u,
1 x+t
Z
1
u(t, x) = (f (x − t) + f (x + t)) + g(y)dy.
2 2 x−t
(5) Folland, pg. 235, exercises 3, 7.
(6) (DFT/FFT) Fix an integer N ≥ 1. Recall that for x = (x0 , · · · , xN −1 ) ∈ CN , we defined the discrete
Fourier transform x̂ ∈ CN via
x̂ = (x̂0 , · · · , x̂N −1 ),
N
X −1
x̂m = e−2πimn/N xn , 0 ≤ m ≤ N − 1.
n=0
(i) Suppose that N is even. Write, for 0 ≤ m ≤ N − 1,
(N/2)−1 N −1
X X
x̂m = xn e−2πimn/N + xn e−2πimn/N
n=0 n=N/2
(N/2)−1 (N/2)−1
X X
= xn e−2πimn/N + x(N/2)+n e−2πim((N/2)+n)/N
n=0 n=0
(N/2)−1
X
= (xn + x(N/2)+n e−πim )e−2πimn/N .
n=0

Use this to show that for each x = (x0 , · · · , xN −1 ) ∈ CN and for every 0 ≤ k < N/2,
(N/2)−1
X
x̂2k = (xn + xn+(N/2) )e−4πikn/N ,
n=0
 and
(N/2)−1
X
x̂2k+1 = (xn − xn+(N/2) )e−4πikn/N e−2πin/N .
n=0
(Note that this latter identity corrects a +/− sign misprint in the Lecture 11 notes.)
(ii) Show that if x ∈ CN is as above, setting
y := (x0 + xN/2 , x1 + x1+(N/2) , · · · , x(N/2)−1 + xN −1 ) ∈ CN/2
and
z := (z0 , · · · , z(N/2)−1 ) ∈ CN/2 , zn = e−2πin/N (xn + xn+(N/2) ),
leads to
x̂ = (ŷ0 , ẑ0 , ŷ1 , ẑ1 , · · · , ŷ(N/2)−1 , ẑ(N/2)−1 ),
so that x̂ can be computed by computing two discrete Fourier transforms of size N/2.
(iii) The following Matlab/Octave code implements a FFT algorithm based on the above identity,
when N is a power of 2 (for simplicity). This numerical scheme is sometimes referred to as
“decimation in frequency.” The implementation here uses recursion, but this can be avoided in
favor of a purely iterative algorithm.

function xhat=fft_altfreq(x)
global xtemp;
N=size(x)(1);
xtemp=x;
fft_altfreq_recurse(0,N);
xhat=xtemp;
end

function fft_altfreq_recurse(start,N)
global xtemp;
start_index=start+1;
start2_index=start_index+N/2;
if (N>1)
temp=xtemp(start_index:start_index+N/2-1);
xtemp(start_index:start_index+N/2-1)= ...
xtemp(start_index:start_index+N/2-1) ...
+xtemp(start2_index:start2_index+N/2-1);
xtemp(start2_index:start2_index+N/2-1)= ...
exp(-2*pi*I*linspace(0,N/2-1,N/2)’/N) ...
.*(temp-xtemp(start2_index:start2_index+N/2-1));
fft_altfreq_recurse(start,N/2);
fft_altfreq_recurse(start+N/2,N/2);
xhat=xtemp;
for k=[0:2:N-1]
xhat(start_index+k)=xtemp(start_index+k/2);
xhat(start_index+k+1)=xtemp(start2_index+k/2);
end
xtemp=xhat;
end
end

For timing comparisons, a naive Matlab/Octave implementation of the discrete Fourier trans-
form (as defiend above) is given by
function xhat=naive_dft(x)
N=size(x)(1);
xhat=exp(-2*pi*I*kron(linspace(0,N-1,N)’,linspace(0,N-1,N))/N)*x;
 end

Use the Matlab/Octave command x=rand(N,1) (for various values of N = 2k , k ≥ 1) to

generate a random vector of N real numbers and use the tic and tok commands to compare
these routines. Compare also with the built-in fft command. Verify that each method computes
the same FFT with a command like norm(fft(x)-naive dft(x)).

(iv) The above algorithm (which is the one we discussed in lecture) is closely related to the FFT
algorithm described in the Folland text. We now discuss a special case of this (sometimes called
“decimation in time”). We again work with even N . The approach is based on decomposing
the summation that defines x̂m into the contributions of even-indexed and odd-indexed terms,
(N/2)−1 (N/2)−1
X X
−4πimk/N
x̂m = x2k e + x2k+1 e−2πim/N e−4πimk/N
k=0 k=0
(N/2)−1
X
= (x2k + e−2πim/N x2k+1 )e−4πimk/N .
k=0
This leads to the following observation: if we set
y = (x0 , x2 , x4 , · · · , xN −4 , xN −2 ) ∈ CN/2
and
z = (x1 , x3 , x5 , · · · , xN −3 , xN −1 ) ∈ CN/2
then
x̂ = (x̂0 , · · · , x̂N −1 )
satisfies
x̂m = ŷm + e−2πim/N ẑm
for 0 ≤ m < N/2, and
x̂(N/2)+m = ŷm + e2−πi((N/2)+m)/N ẑm
for 0 ≤ m < N/2, where the latter identity results from writing
(N/2)−1
X
x̂m = (x2k + e−2πi((N/2)+m)/N x2k+1 )e−4πi((N/2)+m)k/N
k=0
(N/2)−1
X
= (x2k − e−2πim/N x2k+1 )e−4πimk/N , 0 ≤ m < N/2.
k=0
You may find it interesting to implement the FFT algorithm that results from iteratively ap-
plying this reduction in Matlab/Octave, or your programming language of choice.