9.

Green functions

We have already establish a possible solution to a general linear ODE such as

∞
y 00 + 2y = f (x), y(0) = y(π) = 0 where we claim that
X
ck sin kx
k=1

is a solution by just choosing constant ck with

2 ˆ 2Zπ
ck (1 − 2k ) = f (k) = f (x) sin kxdx.
π 0
Indeed, similar computation can be done for periodic boundary condition

y(0) = y(2π), y 0 (0) = y 0 (2π).

However, for example if f is only assumed to be piecewise continuous, the infinite sum
∞
X
ck sin kx may not be twice differentiable or differentiable. Our question is whether it
k=1
will make any sense. The answer is of course yes.
First, let us note that if the infinite sum happens to be twice continuously differentiable,
then for any φ ∈ S2π we have
Z π Z π Z π
00 0 0
y φ + 2yφdx = (−y φ + 2yφ)dx = (yφ00 + 2yφ)dx.
−π −π −π

We say y is weak solution if

Z π Z π
0 0
(−y φ + 2yφ)dx = f φdx for all φ ∈ S2π
−π −π

and we say y is a distribution solution if

Z π Z π
(yφ00 + 2yφ)dx = f φdx for all φ ∈ S2π .
−π −π

Note that for weak solution we need y 0 to ‘integrable’ while for distribution solution, we
only need y to be ‘integrable’.

86
 Mar 29, 23 87

In particular, we can indeed handle a very bad function f , for example, a Delta function
δ0 , (it is indeed not a function, it is a measure with density/pointmass at 0). We can
solve it ‘formally’. Indeed, we see that
1 Z 2π −ikx
e dδ0 = e0 = 1.
2π 0
eikx . Indeed, as long as we interpret the
X
Hence, ‘formally’, the Fourier series of δ0 is
k∈Z
integral Z 2π X X Z 2π
( eikx )φdx as eikx φ(x)dx,
0 k∈Z k∈Z 0
then it is the same as what we expect as
Z 2π
f (x)dδ0 = f (0).
0

The advantage of this expression of Fourier series is that we can then see that
X eikx
y= 2
k∈Z 2 − k

appears to be a solution of y is periodic:

y 00 + 2y = δ0 .

However, as mentioned earlier, this solution may not be in the ’classical sense’, it may
only be either in weak sense or in distribution sense. But the equation with δ function is
important and practical.
88 Mar 29, 23

Let us now recall that if f and g are both periodic (of period 2π), piecewise continuous
functions on R, we have defined
1 Z 2π
f ∗ g(x) = f (x − y)g(y)dy = g ∗ f (x).
2π 0
If one of them ( for example, f ) is piecewise smooth, then

(f ∗ g)0 (x) = f 0 ∗ g(x)

and we may then consider its higher derivative if f has higher derivative. Indeed,

(f ∗ g)00 = f 0 ∗ g 0 .

On the other hand, we could also try to define the convolution of a continuous function
with a (Radon) measure µ
1 Z 2π
f ∗ µ(x) = f (x − y)dµ.
2π 0
In particular, we could look at

f ∗ δ0 (x) = f (x)/2π

where δ0 is the discrete measure that concentrate at 0. We may interpret it as a probability

measure that has measure 1 at the point 0 and has zero measure elsewhere.
 Mar 29, 23 89

We are now looking for solution of the equation L[y] = f where L is a second order
(self-adjoint) differential operator such as

L[y] = (py 0 )0 + qy.

Instead of solving L[y] = f (y is periodic), we will try to solve L[y] = δ0 . Such a solution
will be call a Green function. It has the following wonderful property:
if f is continuous, then we know δ0 ∗ f = f /(2π) and hence

L[G ∗ f ] = L[G] ∗ f = δ0 ∗ f = f /(2π)

and thus (2π)G ∗ f gives us a solution of L[y] = f .

90 Mar 29, 23

10. Fourier transform on R

Recall that if f is periodic with period L, then

ˆ 1ZL
f (n) = f (x)e−2πinx/L dx.
L 0
In particular, when L = 1, we have
Z 1
fˆ(n) = f (x)e−2πinx dx.
0

We would like to define

Z ∞ Z N
fˆ(ξ) = f (x)e−2πiξx dx = lim f (x)e−2πiξx dx.
−∞ N →∞ −N

However, the above limit needs not exists even if f is a bounded uniformly continuous
function on R. Let us consider a special set M(R) when the improper integral converges.

M(R) = {f is piecewise continuous on any bounded interval in R

: there exists a constant M such that |f (x)| ≤ M/(1 + x2 )}.
For easy reference, we will call the set as collection of locally piecewise continuous functions
of moderate decrease.
Remark 10.1
(1) In chapter 5 of ”Fourier Analysis”, the notation denotes the set of continuous
functions of moderate decrease .
(2) There are various definition for the Fourier transform (depending on which text-
book we are using), just to name a few here:

1 Z Z
√ f (x)e−ixξ dx, f (x)eixξ .
2π
However, since we are using the present book, we will stick to
Z ∞ Z N
fˆ(ξ) = f (x)e −2πiξx
dx = lim f (x)e−2πiξx dx.
−∞ N →∞ −N
 Mar 29, 23 91

(3) Indeed, the above is not the standard definition of the improper integral. The
standard definition should be
Z 0 Z ∞ Z 0 Z N
−2πiξx −2πiξx −2πiξx
f (x)e dx+ f (x)e dx = lim f (x)e dx+ lim f (x)e−2πiξx dx
−∞ 0 N →∞ −N N →∞ 0

so that both limits have to converge before we can say the improper integral
Z ∞
f (x)e−2πiξx dx converges. However, for our purpose, we will only need the
−∞
Z N
limit lim f (x)e−2πiξx dx converges.
N →∞ −N

If f ∈ M(R), note that if K > N > 0, then

Z
π
|f (x)|dx ≤ 2M (tan−1 K − tan−1 N ) ≤ 2M ( − tan−1 N ).
N ≤|x|<K 2
Z N
It is now easy to see that the limit lim f (x)dx exists by the Cauchy criterion.
N →∞ −N
Also, note that if f ∈ M(R), then
Z ∞
|fˆ(ξ)| ≤ |f (x)| ≤ M π, if |f (x)| ≤ M/(1 + x2 ).
−∞

10.2 Basic properties of the improper integral: (Proposition 1.1, p.132)

Let f, g ∈ M(R). Then
Z ∞ Z ∞ Z ∞
(i) (linearity) (αf + g) = α f+ g.
−∞ Z ∞ −∞ −∞ Z ∞
(ii) (translation invariance) f (x − h)dx = f (x)dx for all h ∈ R.
Z ∞ −∞ −∞
Z ∞
(iii) (scaling) k f (kx)dx = f (x)dx for all k > 0.
−∞
Z ∞ −∞

(iv) (continuity) |f (x) − f (x − h)|dx → 0 as h → 0.

−∞
92 Mar 29, 23

Some examples: 10.3

For convenience, let us denote F(f ) to be the Fourier transform of the function f .
(1) e−a|x| , a > 0.
Z ∞ Z 0
−a|x| −ax−2πixξ
F(e ) = e dx + eax−2πixξ dx
0 −∞
1 1
= +
a + 2πiξ a − 2πiξ
2a
= 2 .
a + 4π 2 ξ 2
Next find the Fourier transform of the following two functions before proceeding.
(2) χ[−1,1] the characteristic function of the set [−1, 1], that is the function taking the
value 1 on the set but equals to zero otherwise.

(3) e−x (sin x)χ[0,∞) (x).

 Mar 29, 23 93

Schwartz space (p.134)

We will now restrict ourselves to a even smaller subset of M(R), namely, the Schwartz
space S: that is the space of infinitely differentiable functions f such that

sup |x|k |f (l) (x)| < ∞

x∈R

for all k, l ∈ N ∪ {0} where f (l) is the lth derivative of f .

Here are some typical examples:
(1) C0∞ (R) which is the set of infinitely differentiable functions f such that there exists
a constant R > 0 with f (x) = 0 for |x| > R.
2
(2) Gaussian : e−ax , a > 0.
(3) Bump functions g(x) = h(x)h(1 − x) where
 2

 e−1/x when x > 0, 
h(x) =
 0 if x ≤ 0 

Note that g(x) = 0 if x ≤ 0 or x ≥ 1 and g ∈ C0∞ (R).

( See also (p.162).)

Proposition 10.4 (p.136).

Let f ∈ §. Then
(i) τd ˆ
h f (ξ) = f (ξ)e
2πihξ
where τh f (x) = f (x + h);
(ii) F(f (x)e −2πixh
)(ξ) = fˆ(ξ + h) if g(x) = f (x)e−2πixh ;
df (ξ) = fˆ(ξ/δ)/δ where M f (x) = f (δx);
(iii) Mδ δ

(iv) Ff (ξ) = 2πiξ fˆ(ξ);

0

d
(v) F(−2πixf (x))(ξ) = fˆ(ξ) where φ(x) = −2πixf (x).
dξ

Theorem 10.5 (p.137) If f ∈ S, then fˆ ∈ S.

Some details have been given on p.137.
94 Mar 29, 23

A standard computation regarding Gaussians.

Z ∞ Z ∞ Z ∞
−ax2 2 2
( e dx)2
= e−ax e−ay dxdy
−∞ −∞ −∞
Z 2π Z ∞
2
= e−ar rdrdθ using polar coordinates
0 0
h i∞
−ar2
= 2πe /(−2a) = π/a.
0

Of course the above computation works only for a > 0.

Fourier transform of Gaussians (see Theorem 1.4 (p.138) when a = π).

π −π2 ξ2 /a
r
−ax2
F(e )(ξ) = e provided a > 0.
a
For convenience, let
2
F (ξ) = F(e−ax )(ξ).
Then
Z ∞
0 2
F (ξ) = (−2πix)e−ax e−2πixξ dx
−∞
πi Z ∞ 2
= (−2ax)e−ax e−2πixξ dx
a −∞
πi
= (2πiξ)F (ξ)
a
d −ax2 2
(by Proposition 1.2 (iv) since dx
e = −2axe−ax )
−2π 2
= ξF (ξ).
a
This is a linear first order ODE. Solving the equation, we have
d π2 ξ2 /a
(e F (ξ)) = 0.
dξ
Hence
2 2 ξ 2 /a
F(e−ax )(ξ) = F (ξ) = ce−π .
R∞ 2
q
However, F (0) = −∞ e−ax = π/a. Hence
π −π2 ξ2 /a
r
−ax2
F(e )(ξ) = e .
a
In particular,
2 2
F(e−πx )(ξ) = e−πξ .
 Mar 29, 23 95

Moreover, we also have

Corollary 10.6.
2 /δ √ 2
F(e−πx / δ)(ξ) = e−δπξ .
Let us define
2 /δ √
Kδ (x) = e−πx / δ.
Note Zthat
∞ Z ∞
(i) Kδ (x)dx = |Kδ (x)|dx = 1
−∞ −∞
(iii) For every η > 0, Z
lim+ |Kδ (x)|dx = 0.
δ→0 |x|>η

Similar to finite interval, we could define the convolution of two functions in M(R).

Z ∞
f ∗ g(x) = f (x − t)g(t)dt.
−∞

Corollary 10.7 If f ∈ S, then

f ∗ Kδ converges uniformly to f on R as δ → 0+ .
96 Mar 29, 23

Indeed, if f ∈ M(R) and f is continuous at x0 , then

f ∗ Kδ (x0 ) → f (x0 ).

Moreover, f ∗ Kδ ∈ S if f is piecewise continuous and vanishes outside [−M, M ] for some

M > 0.

[Exercise] If fi = 0 outside [−Mi , Mi ], Mi > 0 for i = 1, 2 and f1 , f2 are both piecewise

continuous, show that f1 ∗ f2 vanishes outside [−M1 − M2 , M1 + M2 ]. Is the function
always continuous?
 Mar 29, 23 97

Proposition 10.8. If f, g ∈ S, then

Z ∞ Z ∞
f (x)ĝ(x)dx = fˆ(x)g(x)dx.
−∞ −∞

Indeed, the above is always true as long as Fubini’s theorem can be applied to the function
f (x)g(y)e−2πixy .
98 Mar 29, 23

Theorem 10.9. Fourier inversion If f ∈ §, then

Z ∞
f (x) = fˆ(ξ)e2πixξ dξ.
−∞
2
Proof. Note that if Gδ (x) = e−πδx , then F(Gδ )(ξ) = Kδ (ξ).
Recall from Proposition 10.8 that
Z ∞ Z ∞ Z ∞
2
f (x)Kδ (x)dx = fˆ(x)Gδ (x)dx = fˆ(x)e−δπx dx.
−∞ −∞ −∞
Z ∞ Z ∞
2
Claim: lim+ fˆ(x)e−δπx dx = fˆ(x)dx and hence by Corollary 10.7, we have
δ→0 −∞ −∞
Z ∞
f (0) = lim+ f ∗ Kδ (0) = fˆ(x)dx.
δ→0 −∞

The theorem now follows from Prop 1.2. As the claim involves ‘dominated convergence
theorem’, we will not prove it.

Corollary 10.10. The Fourier transform is a bijective mapping (1-1 and onto) on S.
Z Z ∞ ∞
Fubini’s theorem. If |f (x, y)|dxdy < ∞, that is the improper integral
−∞ −∞
converges absolutely, then
Z ∞ Z ∞ Z ∞ Z ∞
f (x, y)dxdy = f (x, y)dydx.
−∞ −∞ −∞ −∞

Properties of convolution. (This include Proposition 1.11) Let f, g, h ∈ M(R).

Then
(1) f ∗ g = g ∗ f ;
(2) f ∗ (g ∗ h) = (f ∗ g) ∗ h;
(3) (αf1 + f2 ) ∗ g = α(f1 ∗ g) + f2 ∗ g;
(4) F(f ∗ g) = fˆĝ.
 Mar 29, 23 99

Theorem 1.12 Plancherel theorem.

Z ∞ Z ∞
|fˆ|2 = |f |2 .
−∞ −∞

This is indeed true for all f ∈ L2 . However, as that space is beyond the scope of the
course, we will only prove it for f ∈ M(R). First, we will show that it is true if f ∈ S
which is indeed obvious by Proposition 1.8 by choosing g = fˆ, since then ĝ will be equal
to f by the Fourier inversion formula.

Unfortunately, for general f ∈ M(R), we do not know whether the Fourier inversion
formula holds. You cannot do it here.
100 Mar 29, 23

Some applications:
1. Computation of some classical improper integrals:

Z Z
1 Z
sin x/xdx; dx, a, b 6
= 0; (sin x/x)2 dx.
(x2 + a2 )(x2 + b2 )
 Mar 29, 23 101

2. Solving partial differential equations:

Heat equation: ut = uxx , u(x, 0) = f (x).

Theorem 2.1 (p.146) Let Ht (x) = Kδ (x) with δ = 4πt and f ∈ §. Then u(x, t) =
f ∗ Ht (x) is a solution of the heat equation ut = uxx for t > 0. Moreover, u(x, t) → f (x)
uniformly as t → 0+ . Furthermore
Z ∞
|u(x, t) − f (x)|2 dx → 0 as t → 0+ .
−∞