Motivations

Distribution Theory

Distributions/Generalized Functions

Prof. Thierry Blu

Office: Barry Lam 514
e-mail: [email protected]
web: www.ee.cuhk.edu.hk/~tblu/
Department of Electrical Engineering
National Taiwan University

February 2024

Prof. Thierry Blu Distributions/Generalized Functions 1/30

 Motivations
Distribution Theory

Outline

1 Motivations

2 Distribution Theory
Duality framework
Calculus in the sense of distributions
Examples
Fourier calculus

Prof. Thierry Blu Distributions/Generalized Functions 2/30

 Motivations
Distribution Theory

Formal Fourier calculations

Since the XIXth century, mathematicians have realized that some
“prohibited” calculations were actually formally valid. For instance, it is
known that the 1D Fourier transform
w +∞
fˆ(ω) = f (x)e−jωx dx
−∞
can be inverted exactly using the formula
1 w +∞ ˆ
f (x) = f (ω)ejωx dω
2π −∞
However, the condition for the integrals to be valid, i.e., f (x), fˆ(ω) ∈ L1 ,
is rather limitating. Yet, a formal proof of the inverse Fourier formula
1 w +∞
w 
+∞
′ jω(x−x′ )
f (x) = f (x ) e dω dx′
2π −∞ −∞
| {z }
???
seems to indicate that “???” behaves like what we now call the Dirac
distribution. At the same time, Physicists started to realize the
advantages of using (formally) point masses in their theories.
Prof. Thierry Blu Distributions/Generalized Functions 3/30
 Motivations
Distribution Theory

Measure theory
The development of measure theory and the extension of Riesz
representation theorem led to a new understanding of the Dirac mass as
a specific (“Radon”) measure: instead of considering the Dirac mass
directly as if it were a function, characterize it by its effect through an
integral
w
def
ϕ(x) dµ = ϕ(x0 )
where dµ = δx0 (x) dx is a measure (different from the Lebesgue
measure dx): µ(A) = 1 iff x0 ∈ A and µ(A) = 0 otherwise.
Riesz representation theorem allows to define a function f (x)
alternatively to the “direct” characterization:
Classical definition: the complete collection of the real numbers
f (x) for x ∈ RN
Riesz
r definition: the complete collection of the real numbers
f (x)ϕ(x) dx for all “nice” functions ϕ(x)—test functions.

Prof. Thierry Blu Distributions/Generalized Functions 4/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Test functions
The alternative definition of functions involves the choice of nice
functions. In the simplest case, these “test” functions are assumed to be
infinitely differentiable and with bounded support. The space of such
functions is denoted by D. Example in 1D:
(
1
exp(− x(1−x) ) if x ∈]0, 1[
ϕ(x) =
0 otherwise
0.02

0.018

0.016

0.014

0.012
φ(x)

0.01

0.008

0.006

0.004

0.002

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x

Prof. Thierry Blu Distributions/Generalized Functions 5/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Definition of distributions
A distribution u(x) is a generalization of integrable functions to the
space of linear operators D ′ acting on D (its dual). More specifically, a
distribution is characterized by
1 its scalar product hu , ϕi with all test functions ϕ(x) ∈ D
2 its order L: for every closed bounded set K ⊂ RN , there exists a
constant CK such that for all ϕ supported within K†
| hu , ϕi | ≤ CK sup sup |∂ α ϕ(x)|
|α|≤L x∈K

The property 2 ensures a form of continuity of the scalar product.

NOTE: hu , ϕi is now essentially a notation for a linear operator

w on ϕ; but
1 N
when u is a true function of L (R ) hu , ϕi coincides with u(x)ϕ(x) dx.

∂ |α|
†α αn ; ∂ α =
P
= (α1 , α2 , . . . , αN ) nonnegative integers; |α| = n α α α .
∂x1 1 ∂x2 2 ···∂xNN
Prof. Thierry Blu Distributions/Generalized Functions 6/30
 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Integrable functions and Dirac

Any function u of L1 (RN ) is a valid distribution of order L = 0 because

the scalar product effectively satifies (triangular inequality)
| hu , ϕi | ≤ kukL1 (RN ) sup |ϕ(x)|
x∈RN
when ϕ ∈ D.

The Dirac distribution δ(x) is completely characterized by

hδ , ϕi = ϕ(0), for all ϕ ∈ D
Hence | hδ , ϕi | ≤ supx∈RN |ϕ(x)| which shows that δ is also of order 0.

NOTE: δ(x) is not a function of L1 (RN ) (why?)!

Prof. Thierry Blu Distributions/Generalized Functions 7/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Support of distributions

Support
The support of a distribution u is the smallest closed set Su of RN such
that, for any ϕ ∈ D whose support does not intersect Su we have
hu , ϕi = 0

Example: the support of δ(x) is Sδ = {0}.

Prof. Thierry Blu Distributions/Generalized Functions 8/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Gradient and duality

Gradient definition
The gradient of a scalar function is defined by a vector that satisfies the
duality formula
du(r) = h∇u(r) , dri
Here, hu , vi = uTv (no integral).

Expressing a gradient in other coordinate systems than the Cartesian

coordinate system essentially amounts to expressing dr in that system.

Prof. Thierry Blu Distributions/Generalized Functions 9/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Example: gradient in spherical coordinates

 
ρ sin θ cos ϕ
∂r ∂r ∂r
r =  ρ sin θ sin ϕ  ⇒ dr = dρ + dθ + dϕ
∂ρ ∂θ ∂ϕ
ρ cos θ |{z} |{z} |{z}
uρ ρ uθ ρ sin θ uϕ

Hence, du(r) = h∇u , uρ i dρ + h∇u , uθ i ρ dθ + h∇u , uϕ i ρ sin θ dϕ

m
∂u 1 ∂u 1 ∂u
h∇u , uρ i = , h∇u , uθ i = and h∇u , uϕ i =
∂ρ ρ ∂θ ρ sin θ ∂ϕ
and so, thanks to the orthonormality of uρ , uθ and uϕ
∂u 1 ∂u 1 ∂u
∇u = uρ + uθ + uϕ
∂ρ ρ ∂θ ρ sin θ ∂ϕ

Prof. Thierry Blu Distributions/Generalized Functions 10/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Vector differentiation and duality

Dual definition of differential operators

The divergence, curl and Laplacian operators are equivalently defined
through duality formulas
h∇ · u , ψi = − hu , ∇ψi
h∇ × u , aψi = − hu , a × ∇ψi
∇2 u , ψ = − h∇u , ∇ψi = u , ∇2 ψ
r
where ψ ∈ D, and a ∈ R3 are arbitrary. Here hu , vi = u(r)Tv(r) d3 r.

Prof. Thierry Blu Distributions/Generalized Functions 11/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Example: computation of ∇2 in spherical coordinates

∇2 u , ψ = − 
h∇u , ∇ψi    
∂u ∂ψ 1 ∂u 1 ∂ψ 1 ∂u 1 ∂ψ
=− , − , − ,
∂ρ ∂ρ ρ ∂θ ρ ∂θ ρ sin θ ∂ϕ ρ sin θ ∂ϕ
w ∂u ∂ψ

1 ∂u ∂ψ 1 ∂u ∂ψ

=− + 2 + 2 2 ρ2 sin θ dρ dθ dϕ
∂ρ ∂ρ ρ ∂θ ∂θ ρ sin θ ∂ϕ ∂ϕ | {z }
Cartesian to spherical
coordinate change
w 
∂ ρ2 ∂u

∂

∂u

1 ∂2u

= sin θ + sin θ + ψ dρ dθ dϕ
∂ρ ∂ρ ∂θ ∂θ sin θ ∂ϕ2
     
1 ∂ ρ2 ∂u 1 ∂ ∂u 1 ∂2u
= + sin θ + , ψ
ρ2 ∂ρ ∂ρ ρ2 sin θ ∂θ ∂θ ρ2 sin2 θ ∂ϕ2

Prof. Thierry Blu Distributions/Generalized Functions 12/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Scaling distributions

Linear change of variables

Consider a distribution u. Then we define the rescaled distribution
uA (x) = u(Ax) by
1
huA , ϕi = u , ϕ(A−1 x)
| det A|

1
For instance, this allows to see that δ(Ax) = δ(x).
| det A|

Prof. Thierry Blu Distributions/Generalized Functions 13/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Differentiation

Derivatives of a distribution
By definition, the partial derivative of a distribution of order L with
respect to xn is a distribution of order L + 1 characterized by
   
∂u ∂ϕ
,ϕ = − u,
∂xk ∂xk
More generally, distributions can be differentiated infinitely
h∂ α u , ϕi = (−1)|α| hu , ∂ α ϕi

NOTE: The “minus” term results from the integration by part for normal
functions
w +∞ h i+∞ w +∞
u′ (x)ϕ(x) dx = u(x)ϕ(x) − u(x)ϕ′ (x) dx
−∞ −∞ −∞
| {z }
= 0 because ϕ has bounded support

Prof. Thierry Blu Distributions/Generalized Functions 14/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Convolutions
Consider the function f (x0 ) = hu(x) , ϕ(x0 − x)i. This function is
infinitely differentiable (because of the order hypothesis). If the support
of u is bounded, then f ∈ D; i.e., it is infinitely differentiable with
bounded support. We denote it by (u ∗ ϕ)(x0 ).
def
It is possible to show that hu ∗ ϕ , ψi = hu , ϕ̌ ∗ ψi where ϕ̌(x) = ϕ(−x),
and ψ ∈ D.

Convolution of distributions
If u and v are two distributions, one of them with bounded support, then
their convolution product is defined by
hu ∗ v , ϕi = hu , v̌ ∗ ϕi, v has bounded support
hu ∗ v , ϕi = hv , ǔ ∗ ϕi, u has bounded support
where f is the distribution defined by fˇ , ϕ = hf , ϕ̌i.
ˇ

Prof. Thierry Blu Distributions/Generalized Functions 15/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Multiplication of distributions

It is usually not possible to multiply two distributions. However, it is

possible to multiply a distribution with a function that is infinitely
differentiable.
Multiplication
If f is infinitely differentiable and u is a distribution, then their product is
defined by
hf u , ϕi = hu , f ϕi

Counterexample: consider H(x) = 1 if x ≥ 0,and 0 otherwise (Heaviside

d
function). We have H(x)2 = H(x)3 ; yet, dx H(x)2 = 2H(x)H ′ (x)
 
d
and dx H(x) = 3H(x) H (x) = 3H(x)H (x) = 32 dx
3 2 ′ ′ d
H(x)2 .

Prof. Thierry Blu Distributions/Generalized Functions 16/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Limits

Limit of a sequence of distributions

By definition, a sequence of distributions un ∈ D ′ converges to another
distribution u ∈ D ′ iff, for every ϕ ∈ D
lim hun , ϕi = hu , ϕi
n→∞
Notation: un ⇀ u. Important consequence:
n→∞
un ⇀ u automatically implies that ∂ α un ⇀ ∂ α u.
n→∞ n→∞

Example: ejnx ⇀ 0 in the sense of distributions.

n→∞

Prof. Thierry Blu Distributions/Generalized Functions 17/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Regularizing sequence

Consider the sequence of functions (approximation of identity)

ϕk (x) = k N ϕ(kx)
where ϕ(x) ∈ D is of integral 1. Then we have that
ϕk (x) ⇀ δ(x)
k→∞
Proof by changing variables in hϕk , ψi = hϕ(x) , ψ(x/k)i where ψ ∈ D,
and use of Lebesgue dominated convergence theorem.

Using the convolution property, this allows to see that any distribution u
is the limit of a sequence of infinitely differentiable functions
u ∗ ϕk ⇀ u
k→∞

Prof. Thierry Blu Distributions/Generalized Functions 18/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

3D Laplacian of 1/r

We want to compute ∇2 u where u is the distribution defined by

1 1
u(x, y, z) = p =
2
x +y +z2 2 ρ
Method: ∇2 u , ψ = − 
h∇u , ∇ψi
∂u ∂ψ
=− ,
∂ρ ∂ρ
w 1 ∂ψ
= ρ2 sin θ dρ dθ dϕ
ρ2 ∂ρ
w 2π wπ w
∞ ∂ψ

= dϕ sin θ dθ dρ
0 0 0 ∂ρ
= −4πψ(0, 0, 0)

1
Hence, ∇2 = −4πδ(x, y, z).
ρ

Prof. Thierry Blu Distributions/Generalized Functions 19/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Distributions with shortest support

Point distributions
The distributions whose support is restricted to one point {x0 } are all of
the form X
u(x) = ak ∂ α δ(x − x0 )
|α|≤L

where L ∈ N.
Proof : This is a consequence of the Lth order property, and of the use of
Taylor formula to approximate a test function ϕ in the neighborhood of x0 .

Prof. Thierry Blu Distributions/Generalized Functions 20/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Tempered distributions
A function f (x) ∈ RN is said to be fastly decreasing if it decreases faster
than the inverse of any polynomial: for all n ≥ 0
lim kxkn |f (x)| = 0
kxk→∞

Schwartz class
The Schwartz class of functions S is made of the functions for which all
the derivatives have fast decrease. Of course, D ⊂ S .

Examples in 1D: the Gaussian function ϕ(x) = exp(−x2 ) and the

hyperbolic secant sech(x) = 1/ cosh(x).

Tempered distributions
Tempered distributions S ′ are the smaller set of distributions u for which
the scalar product hu , ϕi where ϕ ∈ S , can be defined.

Examples in 1D: polynomials of any degree are tempered, but not ex .

Prof. Thierry Blu Distributions/Generalized Functions 21/30
 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Fourier
The Fourier transform of an L1 (RN ) function f (x) is defined by
w T
fˆ(ω) = f (x)e−iω x dN x
1 w ˆ T
and its inverse is given by f (x) = f (ω)eiω x dN ω.
2π N

Fourier transform of a distribution

By definition, the Fourier transform of a distribution is a distribution
characterized by
hû , ϕi = hu , ϕ̂i, for all ϕ ∈ S

r
Example: if u(x) = 1 then hu , ϕ̂i = ϕ̂(x) dN x = (2π)N ϕ(0), which
shows that û(ω) = (2π)N δ(ω).

NOTE: f ∈ S ⇔ fˆ ∈ S .

Prof. Thierry Blu Distributions/Generalized Functions 22/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Fourier properties
Link with the convolution
Through a Fourier transform, the convolution becomes a simple
multiplication
f[
∗ g(ω) = fˆ(ω)ĝ(ω)

Other properties
Fourier
Duality: fˆ(x) −→ (2π)N f (−ω)
Fourier T
Translation: f (x − x0 ) −→ fˆ(ω)e−ix0 ω
T Fourier
Modulation: f (x)eiω0 x −→ fˆ(ω − ω0 )
Fourier
Differentiation: ∂ α f (x) −→ (iω)α fˆ(ω)
Fourier

Affine transformation: f (Ax) −→ det(A−1 ) fˆ (A−1 )Tω
Fourier
Hermitian symmetry: f ∗ (x) −→ fˆ∗D(−ω) E
Parseval theorem: (2π)N hf , g ∗ i = fˆ , ĝ ∗

Prof. Thierry Blu Distributions/Generalized Functions 23/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Fourier in the wilderness

Physical existence of the Fourier transform
The Fourier transform is naturally involved in
Wave propagation: in the far-field Fraunhofer approximation, the
observed diffraction pattern is the CSFT of the aperture.
Crystallography: imaging of ordered crystals or molecules (e.g., the
DNA molecule).
Color separation by prisms.
Resonance: a physical effect widely used, which involves the Fourier
coefficient at the resonant frequency (e.g, Magnetic Resonance).
square aperture hemoglobin dispersive prism K-space NMR

Prof. Thierry Blu Distributions/Generalized Functions 24/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Fourier transform of cosh−1 (x)

1 w +∞ 2e−iωx dx w ∞ 2z −iω dz
If u(x) = then û(ω) = =
cosh(x) −∞ ex + e−x 0 1 + z2
by making the change of variables x = ln z. Im

x +i
For ω ≤ 0, consider the contour
Re
−R R
w R 2z −iω dz w 2z −iω dz w 0 2z −iω dz 2i−iω
+ + = 2iπ ×
0 1 + z2 half circle 1 + z 2 −R 1 + z 2
{z 2i }
| {z } | {z } | {z } |
−→ û(ω) −→ 0 −→ eπω û(ω) =2πeπω/2
R→∞ R→∞ R→∞

π
Hence û(ω) =
cosh(πω/2)

Prof. Thierry Blu Distributions/Generalized Functions 25/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Poisson

For a function f (x) ∈ L1 (RN ) that decreases fastly enough, the Fourier
series decomposition formula becomes the Poisson summation formula
X X T
f (x + n) = fˆ(2πn)e2iπn x
n∈ZN n∈ZN

Poisson formula
X T X
e2iπn x = δ(x − n)
n∈ZN n∈ZN

Prof. Thierry Blu Distributions/Generalized Functions 26/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Fourier transform of H(x)

Consider the Heaviside function H(x) = 1 if x ≥ 0, and H(x) = 0

otherwise. Then, H(x) = limn→∞ un (x) where un (x) = H(x)e−x/n .
The Fourier transform of un (x) is given by
n
ûn (ω) = −→ ?
1 + inω n→∞
w
ϕ(ω) w (1 − inω)ϕ(ω) w nϕ(ω) w n2 ωϕ(ω)
hûn , ϕi = n dω = n 2 2
dω = 2 2
dω − i dω
1 + inω 1+n ω 1+n ω 1 + n2 ω 2
w ∞ n2 ω ϕ(ω) − ϕ(−ω)

= π hnv(nω) , ϕ(ω)i −i dω
| {z } 0 1 + n2 ω 2
−→ ϕ(0)
| {z }
n→∞ −→ hpv( ω
1
) ,ϕi
n→∞
w ϕ(ω) − ϕ(−ω)
1 1


where v(ω) = π(1+ω 2 ) and pv ω , ϕ = dω.
2ω


Hence, Ĥ(ω) = πδ(ω)
− ipv ω1 .
The distribution pv x1 is Cauchy’s “principal value” of 1/x.

Prof. Thierry Blu Distributions/Generalized Functions 27/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Some Fourier transforms

Useful 2D Fourier pairs

Fourier
δ(x) −→ 1
Fourier
1 −→ (2π)N δ(ω)

Fourier p

exp −xTA−1 x/2 −→ (2π)N det A exp −ωTAω/2
P Fourier P
k∈ZN δ(x − k) −→ k∈ZN δ(ω/(2π) − k)

NOTE: A is positive symmetric.

Prof. Thierry Blu Distributions/Generalized Functions 28/30

 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Green’s Theorem
Consider a volume defined by V = {r ∈ R3 s.t. u3 (r) ≤ 0} and a vector
test function ϕ. We want to compute the volume integral †
w

∇· ϕ d3 r = H −u3 (r) , ∇· ϕ
V
w

= δ u3 (r) ∇u3 , ϕ = δ u3 (r) ϕT ∇u3 d3 r

Assume that we can find an invertible change of variables from

r = (x, y, z) to u = (u1 (r), u2 (r), u3 (r)) such that the Jacobian matrix
[∇u1 , ∇u2 , ∇u3 ] has a positive determinant det ∇u ≥ 0, then the
volume integral becomes
w w du1 du2
∇· ϕ d3 r = ϕT ∇u3
V det ∇u u3 =0

i.e., a surface integral, over the boundary of V.

† H(x) is Heaviside’s step function.

Prof. Thierry Blu Distributions/Generalized Functions 29/30
 Duality framework
Motivations Calculus in the sense of distributions
Distribution Theory Examples
Fourier calculus

Green/Gauss/Ostrogradsky’s Theorem
More precisely, considering the reciprocal relation
r = r(u1 , u2 , u3 ) of the
change of variables, the identity u r(u1 , u2 , u3 ) = (u1 , u2 , u3 ) implies
∇u∇r = Id which shows that
∂r ∂r ∂r
∇u3 ⊥ , , and ∇uT3 =1
∂u1 ∂u2 ∂u3
1  ∂r ∂r 
hence that ∇u3 = × so that, finally
det ∇r ∂u1 ∂u2
Divergence formula
Assume that r = r(u1 , u2 , u3 ) defines a 3D change of variables such that
det ∇r > 0 and that in these new variables the volume V is characterized
by u3 ≤ 0; then
w w
T ∂r ∂r
∇· ϕ(r) d3 r = ϕ r(u) × du1 du2
V ∂u1 ∂u2
| {z } u3 =0
dS

Prof. Thierry Blu Distributions/Generalized Functions 30/30