MATH 220: DISTRIBUTIONS AND WEAK DERIVATIVES

ANDRAS VASY

Suppose V is a vector space over F = R or F = C. The algebraic dual of V is the

vector space L(V, F) consisting of linear functionals from V to F. That is elements
of f ∈ L(V, F) are linear maps f : V → F satisfying
f (v + w) = f (v) + f (w), f (cv) = cf (v), v, w ∈ V, c ∈ F.
When V is infinite dimensional, we need additional information, namely con-
tinuity. So if V is a topological space with the compatible with the vector space
structure, i.e. if V is a topological vector space, we define the dual space V ∗ as the
space of continuous linear maps f : V → F.
For us, V is the class of ‘very nice objects’, and V ∗ will be the class of ‘bad
objects’. Of course, normally there is no way of comparing elements of V with
those of V ∗ , so we will also need an injection
ι:V →V∗
so that elements of V can be regarded as elements of V ∗ (by identifying v ∈ V
with ι(V ). As we want to differentiate functions, as much as we desire, V will
consist of infinitely differentiable functions. As we need to control behavior at
infinity to integrate, the elements of V will be compactly supported. So we define
V = Cc∞ (Rn ) to be the space of infinitely continuously differentiable functions of
compact support. Here recall that the support supp φ of continuous function φ is
the closure of the set where φ 6= 0; so φ ∈ Cc∞ (Rn ) means that there is a compact
(i.e. closed and bounded) subset K of Rn such that φ ≡ 0 outside K.
As Cc∞ (Rn ) is infinite dimensional, we also need to put a topology on this. Tech-
nically this means that we should define what open sets are. Rather than doing
this (to avoid complexity) we define what convergence of a sequence φj of functions
in Cc∞ (Rn ) means.
Definition 1. Suppose {φj }j=1,2,... is a sequence in Cc∞ (Rn ), and φ ∈ Cc∞ (Rn ).
We say that limj→∞ φj = φ if
(i) there is a compact set K such that φj ≡ 0 outside K for all j,
(ii) and all derivatives of φj converge uniformly to φ, i.e. for all multiindices α,
supRn |Dα (φj − φ)| → 0 as j → ∞. Explicitly, for all α ∈ Nn and ǫ > 0
exists N such that for j ≥ N , supRn |Dα (φj − φ)| < ǫ.
Lemma 0.1. For all x0 ∈ Rn and ǫ > 0 there is a function φ ∈ Cc∞ (Rn ) such that
φ(x0 ) > 0, φ ≥ 0 and supp φ ⊂ {x : |x − x0 | < ǫ}.
Proof. First one checks that the function χ defined by
χ(t) = e−1/t , t > 0; χ(t) = 0, t ≤ 0,
is in C ∞ (R). Then we let
ǫ2
 
φ(x) = χ − |x − x0 |2 .
2
This has all the desired properties. 
An immediate corollary is that Cc∞ (Rn ) is indeed infinite dimensional. Note also
that a smooth function is not determined by its Taylor series at any point.
We can now define distributions:
1
2 ANDRAS VASY

Definition 2. A distribution u ∈ D′ (Rn ) is a continuous linear functional on

Cc∞ (Rn ). That is, a distribution u is a map u : Cc∞ (Rn ) → F such that
(i) u is linear: u(φ + ψ) = u(φ) + u(ψ), u(cφ) = cu(φ) for φ, ψ ∈ Cc∞ (Rn ) and
c ∈ F,
(ii) and u is continuous, so if φj is any sequence such that φj → φ in Cc∞ (Rn )
then u(φj ) → u(φ) in F.
It is straightforward to check that continuity of u is equivalent to the following:
for all K compact there exists m and C > 0 such that for all φ ∈ Cc∞ (Rn ) with
supp φ ⊂ K,
X
|u(φ)| ≤ C sup |∂ α φ|.
Rn
|α|≤m

The simplest distributions is the delta-distribution at some point a ∈ Rn ; it is

given by
δa (φ) = φ(a).
One can also generate many similar examples, e.g. the map u : Cc∞ (Rn ) → C given
by u(φ) = (∂1 φ)(a) − (∂22 φ)(b), where a, b ∈ Rn , n ≥ 2, is also a distribution.
A large class of distributions is obtained the following way. Suppose that f ∈
L1loc (Rn ) (i.e. is locally Lebesgue integrable) – for instance, (piecewise) continuous
f would satisfy this. Then f defines a distribution u = ιf as follows:
Z
u(φ) = ιf (φ) = f φ dx.
Rn

Note that the integral converges since φ has compact support.

Lemma 0.2. The map ι : L1loc (Rn ) → D′ (Rn ) is injective.
Indeed, if ιf = 0 then f φ dx = 0 for all φ ∈ Cc∞ (Rn ), which is easily seen
R
to imply f = 0. Note that restricted to C(Rn ) the injectivity argument does not
involve anything remotely sophisticated: if f ∈ C(Rn ), and f (x0 ) 6= 0 for some x0 ∈
Rn , by the continuity of f , for ǫ > 0 sufficiently small, |f (x) − f (x0 )| < |f (x0 )|/2
for |x − x0 | < ǫ. Now let φ be as in Lemma 0.1, so |f (x) − f (x0 )| < |f (x0 )|/2 on
supp φ. So
|f (x0 )|
Z Z Z Z
| f φ dx − f (x0 ) φ dx| ≤ |f (x) − f (x0 )| φ(x) dx ≤ φ dx,
2
R R
so f φ dx 6= 0 as φ dx > 0.
Because of this lemma, we can consider L1loc (Rn ) as a subset of D′ (Rn ), via the
identification ι.
One usually equips D′ (Rn ) with the so-called weak-* topology, so one says that
a sequence uj ∈ D′ (Rn ) converges to u ∈ D′ (Rn ) if u(φ) = limj→∞ uj (φ) for all
φ ∈ Cc∞ (Rn ).
As an example, fix ǫ = 1 and let φ be as in Lemma 0.1 with x0 = 0. Let δj be a
sequence of positive constants with lim δj = 0. Let uj beR the distribution given by
δj−n φ(./δj ), i.e. uj (ψ) = δj−n φ(x/δj )ψ(x) dx. Let c = φ dx. Then lim uj = cδ0 .
R

Indeed, by the continuity of ψ, given ǫ′ > 0, there is δ ′ > 0 such that |x| < δ ′
implies |ψ(x) − ψ(x0 )| < ǫ′ . Then
Z Z
−n
|uj (ψ) − cδ0 (ψ)| = | δj φ(x/δj )ψ(x) dx − ψ(0) φ(x) dx|
Z Z
= | δj−n φ(x/δj )ψ(x) dx − δj−n φ(x/δj )ψ(0) dx|
Z
≤ δj−n φ(x/δj )|ψ(x) − ψ(x0 )| dx.
 3

For j sufficiently large, δj < δ ′ , and φ(x/δj ) = 0 if |x|/δj ≥ 1, i.e. if x ≥ δj , so

certainly if x ≥ δ ′ . Correspondingly, in the integral, one can restrict the integration
to |x| ≤ δj , where |ψ(x) − ψ(x0 )| < ǫ′ to deduce that
Z
|uj (ψ) − cδ0 (ψ)| ≤ ǫ′ δj−n φ(x/δj ) dx = cǫ′ .

This proves our claim. This is a rather typical example, and it is not hard to show
that one can approximate any u ∈ D′ (Rn ) in the weak-* topology by uj which are
given by Cc∞ (Rn ) functions, i.e. Cc∞ (Rn ) is dense in D′ (Rn ).
We can now consider differentiation. The idea is that we already know what the
derivative of a C 1 function is, so we should express it as a distribution in such a
way that it obviously extends to the class of all distributions. Now, for f ∈ C 1 (Rn ),
the distribution associated to the function ∂j f satisfies
Z Z
ι∂j f (φ) = ∂j f φ dx = − f ∂j φ dx = −ιf (φ)

for all φ ∈ Cc∞ (Rn ). Motivated by this, we make the definition:

Definition 3. The partial derivatives of u ∈ D′ (Rn ) are defined by
∂j u(φ) = −u(∂j φ).
Note that for φ ∈ Cc∞ (Rn ),
∂j φ ∈ Cc∞ (Rn ), so this definition makes sense.
It is straightforward to check that ∂j u is a distribution, in particular is continuous
as a map Cc∞ (Rn ) → C. Note also that this is the only reasonable notion of a
derivative as the map u 7→ ∂j u is continuous, i.e. uk → u implies ∂j uk → ∂j u, and
Cc∞ (Rn ), on which we already know ∂j , is dense in D′ (Rn ).
Some examples: on R,
δa′ (φ) = −δa (φ′ ) = −φ′ (a), φ ∈ Cc∞ (R).
Also, if H is the Heaviside step function, so H(x) = 1 for x ≥ 0, H(x) = 0 for
x < 0, then
Z ∞
H ′ (φ) = (ιH )′ (φ) = −ιH (φ′ ) = − φ′ (x) dx = φ(0)
0
for all φ ∈Cc∞ (R),where we used the fundamental theorem of calculus. Therefore
H ′ = δ0 .
Note also that distributions u may be multiplied by C ∞ functions g. Indeed, we
proceed again by analogy with ιf where f ∈ C(Rn ). In that case
Z Z
ιf g (φ) = (f g)φ dx = f (gφ) dx = ιf (gφ),

so for arbitrary u ∈ D′ (Rn ) we define gu ∈ D′ (Rn ) by

gu(φ) = u(gφ).
Note that for φ ∈ Cc∞ (Rn ),
gφ ∈ Cc∞ (Rn ) since g ∈ C ∞ (Rn ), so the definition makes
sense.
As an example of calculation with distributions, consider the following:
Lemma 0.3. Suppose that u ∈ D′ (R) and xu = 0. Then there is a constant c ∈ F
such that u = cδ0 .
Proof. First note that if u = cδ0 , then xu = 0 indeed:
xu(φ) = u(xφ) = cδ0 (xφ) = c(xφ)(0) = 0
for all φ ∈Cc∞ (R)since x(0) = 0, so (xφ)(0) = 0.
Suppose now that xu = 0, i.e. u(xφ) = 0 for all φ ∈ Cc∞ (R). If ψ is a test function
such that ψ(0) = 0 then Taylor’s theorem allows one to write ψ = xφ with φ ∈
Cc∞ (R) (namely φ = x−1 ψ for x 6= 0 extends to be C ∞ at 0), so u(ψ) = u(xφ) = 0.
4 ANDRAS VASY

So now suppose that ψ ∈ Cc∞ (R). Let φ0 ∈ Cc∞ (R) be such that φ0 (0) 6= 0. We
choose α ∈ F such that ψ − αφ0 vanishes at 0, i.e. let α = φψ(0)
0 (0)
. Then by the
argument of the previous paragraph, u(ψ − αφ0 ) = 0. Thus,
u(ψ) = u((ψ − αφ0 ) + αφ0 ) = u(ψ − αφ0 ) + αu(φ0 )
u(φ0 ) u(φ0 )
= αu(φ0 ) = ψ(0) = cδ0 (ψ), c = .
φ0 (0) φ0 (0)
This finishes the proof. 
If L is a linear partial differential operator, so L is of the form
X
L= aα D α ,
|α|≤m

and aα are in C (R ), then for all u ∈ D′ (Rn ), Lu makes sense as an element of

∞ n

D′ (Rn ). In particular, we make the following definition:

Definition 4. Suppose f ∈ L1loc (Rn ), u ∈ L1loc (Rn ), L is linear with C ∞ (Rn )
coefficients. We say that u is a weak solution of Lu = f if Lu = f in the sense of
distributions.
All these considerations go through unaffected if Rn is replaced by an open set
Ω ⊂ Rn . In this case, we consider only test functions φ ∈ Cc∞ (Rn ) with supp φ ⊂ Ω.
Even if L is not linear, it may make sense to talk about weak solutions. For
instance, for conservation laws, L(u) = ut + (f (u))x , u ∈ L∞ 2
loc (R ) (so e.g. u ∈
2
C(R ) or u has jump discontinuities) and f ∈ C(R) we say that u is a weak
solution of L(u) = 0 if Lu = 0 in the sense of distributions. Note that in this case,
f (u) ∈ L∞ 2 1 2 ′ 2
loc (R ) ⊂ Lloc (R ), so f (u) can be considered as an element of D (R ), so
the derivative makes sense.
One often wants solutions only for t ≥ 0, in which case it is helpful to refor-
mulate the definition of a weak solution by rewriting the distributional derivatives
explicitly. If u ∈ C 1 (Rx × [0, ∞)t ), and φ ∈ Cc∞ (Rx × [0, ∞)t ), then
Z
(ut + f (u)x )φ dx dt
Rx ×[0,∞)t
Z Z
=− (u φt + f (u)φx ) dx dt − u(x, 0) φ(x, 0) dx.
Rx ×[0,∞)t Rx

So we say that u ∈ L∞
loc (Rx × [0, ∞)t ) is a weak solution of
ut + f (u)x = 0, u(x, 0) = v(x),
if for all φ ∈ Cc∞ (Rx × [0, ∞)t ),
Z Z
(u φt + f (u)φx ) dx dt + v(x) φ(x, 0) dx = 0.
Rx ×[0,∞)t Rx

In particular, if φ is supported in t > 0, so φ(x, 0) = 0 for all x, then

Z
(u φt + f (u)φx ) dx dt = 0,
Rx ×[0,∞)t

so ut + f (u)x = 0 in the sense of distributions on Rx × (0, ∞)t .