Introduction to Distribution Theory

James ash
Talk,
April 2025
Vibrating String

In [1747] J. L. d’Alembert (1717–1783) showed that the displacement of the

points on a vibrating string fixed at the endpoints can be described by the
expression
fpx, tq “ ψpx ` tq ` φpx ´ tq, (1)

Figure 1

where the arbitrary functions φ and ψ can be determined from the initial
state of the string.
1
d’Alembert vs Euler

• He did not explicitly set down the wave equation

B2 f B2 f
“ (2)
Bx2 Bt2
which does governs the movement of a string, but that was done by
Leonard Euler (1707–1783) in his second paper [1753] on the subject.
• In his first paper [1748] on the vibrating string, which he wrote
immediately after the publication of d’Alembert’s article, Euler derived
expression (1) along the same lines as d’Alembert, But his
interpretation of the arbitrary functions φ and ψ differed decisively
from d’Alembert’s. D’Alembert had explicitly stated that φ and ψ must
be analytic expressions whereas Euler thought that they could be
completely arbitrary.
• Such a controversy would raise a critical question. How REGULAR
should we assume our solutions are?
2
Enter new fighter: Lagrange

While studying the wave equation

B2 z B2 z
2
“c 2 (3)
Bx Bt

Lagrange proposed a idea, he multiplied the equation by a function Mpxq,

integrated over r0, as, then did integration by parts in x. He arrived at the
expression
ża 2 ˙ˇa ża 2
d z dz dM ˇˇ d M
ˆ
2
M dx “ c M´z ˇ `c z 2 dx. (4)
0 dt dx dx ˇ 0 dx
0

Lagrange then assumed that Mp0q “ Mpaq “ 0 and assumed z to be zero at

the end point, hence arriving at
ża 2 ża 2
d z d M
2
M dx “ c z 2 dx. (5)
0 dt 0 dx

3
Lagrange discovery

• Lagrange has officially made a new equation where we made no

assumptions of differentiability on z in the x variable, However while
this attempt wasn’t fruitful since Lagrange would later discover that his
solution is equally discontinuous in t as it is in x, Lagrange had
accidentally discovered a idea at the heart of distribution theory
• But a few questions already present themselves:
1. How did lagrange pick his function M?
2. Does this make sense physically?
3. How will we get rid of the integral?
To answer the first question and looking at the equation we are indeed
already assuming twice differentiability of M in x, as well as forcing M
to vanish on the boundaries, but is that enough?
• Distributions will answer all these questions.

4
Heaviside step function

In 1848, heaviside introduced the following simple step function

#
0, xă0
Hpxq “
1, xě0
One might ask, how should we understand the derivative of such a
function? Well, lets naivly say that
#
1 0 x‰0
H pxq “
8 x“0
żM
One might argue that H1 pxq “ HpMq ´ Hp´Mq “ 1.
´M
Is this possible? Can such a derivative exist?

Hpxq

x 5
Singularities

• The last slide shows that there is some sort of singularity in the
derivative of the heaviside function, it seems as if the derivative of
such a function must satisfy two properties.
#
0 x‰0
1. H pxq “
1
8 x“0
ż
2. H1 pxq “ 1
R
• But does such a function exist?
• Ofcourse not! What we described above is typically called the dirac
delta δ, who, to say the least, had a rough childhood, mathematicians
refused working with it, and never understood it well untill much later.

6
The unloved child isn’t a function

• Lets argue that such a function, cannot exist. Consider a differentiable

function φpxq defined to be zero outside of r´1, 1s, such that φp0q “ 1
and φ “ 0 outside of r´1, 1s.
• Then notice that using IBP we have
ż ż ż8
1
δpxqφpxq “ ´ Hpxqφ pxq “ ´ φ1 pxq “ φp0q “ 1
R R 0
But! we have that δpxqφpxq “ 0 everywhere except when x “ 0, so that
ż
δpxqφpxq “ 0 ‰ 1
R
This is a contradiction! So what shall we do? Has mathematics failed
us?
φpxq
1

x 7
´1 1
The problem with functions

• Imagine that you are trying to study the temperature in a room at a

given time t, you could say that such a thing can be described by a
function fpx, tq “heat at position x and time t.
• But, what if x is to represent the spatial variable, what does something
like x “ p1, 1, 1q mean? are we measuring the temperature at the
atomic position p1, 1, 1q ? In reality we are measuring the temperature
based on the average of all the temperature around it!
• This means in reality, we need a better way to describe the
temperature, and other physical phenomenon. Thats where
distribution theory comes into play.

8
What should we do to measure the average?

• Lets forget about time for a moment, How do we mathematically

describe measuring the temperature at a location a? If we are to use a
thermometer it would measure the average temperature close enough
to where its positioned at a point a.
• Okay one could say lets take a ball Bpa, εq around a of small enough
radius ε then integrate over that, in other words :
ż
tempreture » fpxqdx
Bpx,εq

But this is not exactly accurate, since we expect the value of fpxq near
the point a to have more of a influence on the temprature. The effect
of fpxq should slowly fade away to 0 as we move away from a.
• If we are to imagine the effect of fpxq to be in full effect near a and
slowly decreasing to 0, perhaps we could multiply fpxq by a function
φpxq to help with that process. Lets show how that might work.

9
Bump functions (also known as bumpies)

φpxq

1
2
1.5
1
0.51

Lets consider a function that takes the value 1 when we are near a, and
then slowly decays to zero as we move away from a. We consider
continuous functions so our measurement is accurate. Such functions that
are continous and vanish outside a compact set are said to belong to
Cc pR3 q “ CpR3 q X tf : R3 Ñ R : tx : fpxq ‰ 0u is compact u
10
Almost Generalized Functions

• So it seems we have our solution, if we consider the function φpxqfpxq

this will accurately take into account the effect of fpxq near a and then
the effect fades away as we move away from a.
• Ofcourse we would need to make multiple measurements with a
thermometer, so to take into account for all the ways we can measure
the effect of fpxq near a point, we need to look at φpxqfpxq for all
φ P Cc pR3 q.
• In other words, we arent consider f as a function that takes x P R3 to
fpxq, we are considering f as a Generalized function Tf that takes
ż
3
φ P Cc pR q Ñ Tf pφq “ fpxqφpxqdx
R3

This would almost be enough, but one problem remains.

11
Back to the wave equation

• Remember how the strategy proposed by Lagrange to solve the wave

equation
B2 z B2 z
“ c
Bx2 Bt2
relied on our ability to transfer the burden of differentiability to a new
function Mpxq , yet the equation Lagrange arrived at still assumed
some regularity in time
ża 2 ża 2
d z d M
2
M dx “ c z 2 dx.
0 dt 0 dx
• But what if we instead multiplied by a function φpx, tq, that is also
differentiable in t and applied integration by parts on BOTH sides, what
we’ll arrive at is the equation
B 2 φpx, tq B 2 φpx, tq
ż ż
zpx, tq “ c zpx, tq
R2 B2 t R2 B2 x
ż ˆ 2
B φpx, tq B 2 φpx, tq
˙
ùñ ´ c zpx, tq “ 0
R2 B2 t B2 x
12
The Burden of differentiability

• Notice how the last expression makes sense even when zpx, tq is not
even differentiable, but merely that the integral makes sense. This
motivates a new notion of differentiability where we are
”differentiating the Generalized form of zpx, tq”
• Instead of differentiating the function zpxq in the classical sense,
ż we
can perhaps differentiate the generalized version Tz pφq “ φpxqzpxq,
yet if we are to do integration by parts twice as before, we must
replace our initial assumption of φ P Cc pR3 q with the stronger
assumption of φ P Cc2 pR3 q
• This would allow us to move two derivatives from zpxq onto our Test
function φpxq, giving it the burden of differentiability. It now makes
sense to talk about the derivative of zpxq by doing integration by parts
on its Generalized form
ż
B 2 Tz pφq “ zpxqB 2 φpxq.

Notice how the left side makes sense even when zpxq is not C 2 pR3 q 13
Functional Distributions

• Inspired by our previous problem, if we are to allow differentiability up

to any order, we must assume our Test functions φpxq to be Cc8 pR3 q
Definition (Functions as distributions)
For every continuous function f : R3 Ñ R, we define the distribution

Tf : CC8 ÝÑ R (6)
ż
φ ÝÑ φpxqfpxqdx (7)
R3

And its corresponding derivative as

ż
B α Tfpφq “ p´1qα fpxqB α φpxq
R3

• In reality we can ask for a bit more than just continuous functions, to
those who know measure theory,ż all we need is
f P L1loc pR3 q “ tf : R3 Ñ R : |fpxq| ă 8, for all K compact.u 14
K
Is this notion well defined?

• Now that we have a natural generalizations of functions, and if Tf is to

play the ”representative” role of our function f, we ask a natural
question
Can we say that Tf “ Tg ùñ f “ g ?
• In fact this is more of a pressing issue than whats written above,
indeed recall that in our wave equation we arrived at the equation
ż ˆ 2
B φpx, tq B 2 φpx, tq
˙
3
2t
´ c 2x
zpx, tq “ 0, @φ P C8
c pR ˆ r0, 8sq
R 2 B B
But what if we solved this equation and found out that zpx, tq P C 2 pR3 q
• Reversing the integration by parts, we are back at
żˆ 2
B zpx, tq B 2 zpx, tq
˙
3
´ c φpx, tq “ 0, @φpx, tq P C8
c pR ˆ r0, 8sq
B2 t B2 x
We would like to deduce from this that the wave equation
B 2 zpx, tq B 2 zpx, tq
´ c “0
B2 t B2 x
15
is true!
Fundamental Lemma of the calculus of variations.

Luckily for us, there is a lemma that makes sure both of these problems
have a satisfactory answer.
Lemma
If a continuous function f : R3 Ñ R satisfied
ż
3
fpxqφpxq “ 0, @φ P C8
c pR q

then f ” 0.

Notice how this lemma immedietly solves our problems, indeed if Tf “ Tg

then
ż ż ż
3
fpxq ´ gpxq φpxq “ 0, @φpxq P C8
“ ‰
fpxqφpxq “ gpxqφpxq ùñ c pR q

which by the lemma provides that fpxq “ gpxq.

You should convince yourself that the wave equation problem is also
solved by our lemma!
16
The unloved child problem.

• It seems we found a way to represent every continuous function f as a

”infinitely differentiable” generalized function Tf . But what about the
heaviside function? and its infamous derivative δ?
• Well obviously there is no problems with the expression
ż ż8
Hpxqφpxq “ φpxq ă 8
0
So we canż still consider Hpxq as a generalized function
TH pφq “ Hpxqφpxqdx.
• But what about δ ? Well if we are to consider the derivative of Hpxq in
our new generalized framework, we would arrive as before at
ż
BTH pφq “ ´ HpxqBφpxq “ φp0q

żSo it make sense to say that δpφq “ φp0q. But...we already know that
δpxqφpxq doesnt work, so whats the problem?

17
Dirac Delta Definition

• The problem
ż is that its not enough to consider distributions of the
form fpxqφpxq!

• Indeed, if we simply consider the dirac delta as the Linear function

δ : Cc8 Ñ R
φpxq Ñ φp0q

We realize what we were missing in our definition.

• Indeed the function Tf that we considered earlier, is merely a Linear
functional on C8 c pR q. If we allow all linear functionals on Cc that are
3 8

sufficiently well behaved to be included in our notion of generalized

functions, we can finally make sense of what δ actually is!

18
Fundamental Solution

• A critical application of the dirac delta function is the concept of a

fundamental solution, it is motivated by the admitted fact that f ˚ δ “ f
where ˚ denotes the convolution, which we wont discuss, but it can be
defined using tensor products in distribution theory.
• Lets suppose for a moment that we solve the differential equation
ż
∆u “ δ , to be understood as upxq∆φ “ φp0q

and arrive at some solution u0

• Then notice that if we want to solve ∆u “ f for some f with certain
conditions. Lets simply consider the function u0 ˚ f, notice that
∆pu0 ˚ fq “ ∆u0 ˚ f “ δ ˚ f “ f
and it seems we found our solution!
• Whats important is that we have a powerful tool to solve linear PDEs,
indeed if we can solve Lu “ δ for some linear differential operator L
(for example L “ ∆q, we can deduce a solution for all Lu “ f !
19
Topological Considerations

The topological space C8

c
Consider the topological space C8
c pR q, then we can turn this into a
3

topologcal space by declaring convergence as

φj Ñ φ in C8
c pΩq as j Ñ 8 ðñ
$
&DK Ă Ω compact, φj P tφ P C8
c pΩq : φ vanishes outside Ku for all j,
%and ∥B α φj Ñ B α φ∥8,K Ñ 0, @α P NN
jÑ0

You should convince yourself that this is the natural topology for such a
space, indeed if we simply ask for uniform convergence of all derivatives,
simply consider a smooth test function that keeps ”Expanding”, you can
make sure uniform convergence of all derivatives happen while still losing
the function vanishing outside a compact set.

20
The Space of distributions

We make it more explicit what we mean by ”well behaved linear functions”.

Definition D˚ pΩq
We define the space of distributions D˚ pΩq to be the topological dual of
C8c pΩq when equipped with the topology described above. In other
words, the space of distributions consists of the continuous linear
functionals on C8
C pΩq.

We can also equip our space D˚ pΩq with a natural topology by declaring
convergence as

Tn Ñ T ðñ Tn pφq Ñ Tpφq, @φ P C8
c pΩq

This is sometimes called the weak˚ -topology in functional analysis, which

is simply the topology of pointwise convergence!

21
Shockwaves and Weak Solutions (1/4)

• Consider the inviscid Burgers’ equation:

ˆ 2˙
u
Bt u ` Bx “ 0, px, tq P R ˆ r0, 8q
2
with initial condition:
#
1, xď0
upx, 0q “ gpxq “
0, xą0

• Problem: gpxq is discontinuous at x “ 0, this means we already suspect

our classical derivatives wont apply.
• Indeed instead we are most likely to solve the equation weakly by
distribution theory, as before we multiply by φpxq and integrate by
parts to arrive at
ż8ż8 ˆ 2˙ ż8
u
u φt ` φx dx dt ` gpx, 0q φpx, 0q dx “ 0.
0 ´8 2 ´8

22
Shockwaves and Weak Solutions (2/4)

• One method to solve the equation is through the ”method of

characteristics”, ofcourse we wont be going into all the mathematicals
details but intuitively speaking, it tells us our solutions upx, tq will be
constant along certain curves xptq that start at any point on x0 P R.
• You can think of it as if the solution will get its information
”transfered” from our function gpxq:
t

upxq is constant along these lines

xptq “ gpx0 qt ` x0 Characteristic curves

Initial data 23
x
Shockwaves and Weak Solutions (3/4)

• This means that in our example, the curves will have slope 1 for x ă 0
and 0 for x ą 0. But this suggests that at some point, the curves will
collide! so where do we take our information from at the collision
points? If we take it from the first curve, we get upx, tq “ upx0 , tq “ 1
and if we take it from the second curve we get upx, tq “ upx0 , tq “ 0.
• The problem is that the curves will form a shockwave when they
collide, so as long as we can describe this shockwave curve sptq. We
have our solution! #
1 x ă sptq
upx, tq “
0 x ą sptq
• The speed s of the shock needs to satisfy something called the
Rankine–Hugoniot condition:
Fpuleft q ´ Fpuright q u2
s1 ptq “ , where fpuq “
uleft ´ uright 2
fp1q ´ fp0q 1{2 ´ 0 1
s1 ptq “ “ “
1´0 1 2 24
Shockwaves and Weak Solutions (4/4)
t
shockwave

x
• u“1 u“0
• So the solution is a shock moving at speed 1{2:
#
1, x ă t{2
upx, tq “
0, x ą t{2
• This is a weak solution, it doesn’t exist classically.
25
Conclusion and Questions

• Today distribution theory plays a central role in PDE research, it is

universally accepted that its the correct way to think about
differentiability when solving differential equations and is ubiquitous
in the field.
• Ofcourse that doesnt mean classical solutions dont exist, and in some
cases you can recover classical solutions from weak ones using for
example, Sobolev embeddings, but it seems that reality isnt as well
behaved as we expect it to be and distribution theory captures it much
more accurately than differentiable functions.
• If this talk motivates you to study distribution theory, dont hesitate to
pickup a book after building a solid measure theory and functional
analytic background!
• Do you have any remaining questions or Remark?

26