DISTRIBUTIONS

CHRISTIAN REMLING

1. Basic properties
The infamous Dirac δ-function is an object much beloved by theoret-
ical physicists. It cannot be given a mathematically sound definition
(that is, as a function), but rather is usually imagined as having the
values (
∞ x=0
δ(x) = ,
0 x 6= 0
and thenRone assumes that the singularity at x = 0 can be fine tuned
∞
to make −∞ δ(x) dx = 1.
Despite these inauspicious beginnings, the δ-function often seems a
convenient tool at least in formal calculations, and so it is tempting to
try to build a rigorous theory around the idea.
To do this, we take seriously the frequently heard excuse that the
“definition” is not to be taken literally, but that everything will make
sense when put under an integral sign. We can then in fact hope,
slightly more ambitiously, that
Z ∞
f (x)δ(x) dx = f (0)
−∞

(since the integration seems to take place at x = 0 exclusively).

So now the δ-function really does something on other functions, and
this process, whose exact workings are left unexplained, outputs a num-
ber. We now build a precise definition around this idea. For reason
that will become clear later (perhaps), we want our test functions to
be very nice functions.
Definition 1.1. Let U ⊆ Rd be a non-empty open set. A distribution
u ∈ D0 (U ) is a bounded linear functional on D = C0∞ (U ).
More specifically, u : D → C is a linear map, and if K ⊆ U is any
compact set, then there are C = C(K) ≥ 0 and N = N (K) ≥ 0 such

Date: February 28, 2020.

1
2 CHRISTIAN REMLING

that
∂α
(1.1) |u(ϕ)| ≤ C sup ϕ(x)
x∈U,|α|≤N ∂xα
for all ϕ ∈ D with supp ϕ ⊆ K.
We will be mostly interested in the case U = Rd , in fact most of the
time with d = 1, but sometimes the extra flexibility in the definition
will be useful.
I have used multi-index notation for partial derivatives, so α =
(α1 , . . . , αd ), and
∂ αf ∂ α1 ∂ αd
= · · · αd f (x1 , . . . , xd ).
∂xα ∂xα1 1 ∂xd
Recall that for well behaved functions such as f ∈ D, it does not matter
in which order these derivatives are taken, so this is well defined. We
also set |α| = α1 + . . . + αd .
A functional on a vector space X over C is, by definition, a map u :
X → C. The notation X 0 or X ∗ for the space of linear functionals (often
with extra continuity conditions imposed) is common. The action of
u ∈ D0 on a test function ϕ ∈ D is often written as (u, ϕ) instead of
u(ϕ). This suggests the viewpoint of a pairing between D and its dual
space D0 .
The supremum (really: maximum) on the right-hand side of (1.1)
is also denoted by kϕkN . Condition (1.1) can reformulated as follows,
and this indeed has the feel of a continuity condition. (In fact, it is
possible to put a topology T on D in such a way that (1.1) becomes
equivalent to continuity with respect to T and the standard topology
on C, but this T is rather inconvenient to work with, and it is not
useful at all for our purposes.)
Exercise 1.1. Let u : D → C be a linear functional. Show that the
following are equivalent: (a) u satisfies the condition from Definition
1.1; (b) If ϕn ∈ D, supp ϕn ⊆ K for some fixed compact set K ⊆ Rd ,
and kϕn kN → 0 for all N ≥ 0, then u(ϕn ) → 0.
Example 1.1. The motivating example, δ(ϕ) = ϕ(0), is a distribution
in this sense. Clearly, the map ϕ 7→ ϕ(0) is linear, and |ϕ(0)| ≤
sup |ϕ(x)| = kϕk0 , as required.
Example 1.2. The functional u(ϕ) = ϕ0 (0) (and here d = 1) also defines
a distribution, since u is obviously linear and |u(ϕ)| ≤ sup |ϕ0 (x)| ≤
kϕk1 .
 DISTRIBUTIONS 3
P
Example 1.3. u(ϕ) = n∈Z ϕ(n) defines a distribution: Note, first of
all, that there are no convergence issues because ϕ(n) 6= 0 for only
finitely many n for any given ϕ ∈ D. Moreover, if supp ϕ ⊆ [−N, N ],
then |u(ϕ)| ≤ (2N + 1)kϕk0 .
Note that in the last example, the constant in estimate (1.1) does de-
pend on the support of the test function, as anticipated as a possibility
in the definition.
Exercise 1.2. Find an example of a distribution u (perhaps similar to
the one from Example 1.3) for which N from (1.1) also necessarily
depends on the support of test function.
If f ∈ L1loc (U ), then
Z
u(ϕ) = f (x)ϕ(x) dx
U
R
defines a distribution since if supp ϕ ⊆ K, then |u(ϕ)| ≤ kϕk0 K |f |.
Moreover, f can be recovered from the distribution it generates, up
to a change of its values on a null set. So we can identify f with
the distribution u = uf and think of a locally integrable function as a
distribution when this is convenient.
Theorem 1.2. Let f, g ∈ L1loc (U ) and suppose that uf = ug . Then
f = g a.e.
Exercise 1.3. Prove Theorem 1.2. If desired, you can focus on the
case U = Rd exclusively, though the general case isn’t much differ-
ent. Suggestion: It suffices to show that if uf = 0, then f = 0 a.e.
Take convolutions of f with suitable functions and interpret these as
applications of uf to test functions.
Exercise 1.4. Show that Lp ⊆ L1loc for all p ≥ 1.
Given a general u ∈ D0 , it is sometimes interesting to ask if u = uf
for some f ∈ L1loc or, as this is often put, if u ∈ D0 is a function.
Exercise 1.5. Show that δ ∈ D0 is not a (locally integrable) function in
this sense.
More generally, if µ is a Borel measure on Rd (and this, as usual,
includes the requirement
R that µ(K) < ∞ for all compact sets K ⊆ Rd ),
then u(ϕ) = f dµ is another distribution. Again, µ can be recovered
from u, and thus we can think of measures as distributions if desired.
Example 1.4. Define
  Z
1 ϕ(x)
(1.2) PV , ϕ = lim dx.
x h→0+ |x|>h x
4 CHRISTIAN REMLING

Here, PV stands for principal value, and what is meant by this is the
regularization of the integral by omitting the neighborhood (−h, h) of
the singularity at x = 0 and then sending h → 0. Some such device is
needed since ϕ(x)/x is not integrable if ϕ(0) 6= 0.
Before we can use (1.2) as a definition, we must in fact establish that
the limit exists. By the mean value theorem, ϕ(x) = ϕ(0) + ϕ0 (ξ)x,
for some ξ = ξ(x) between 0 and x. If we also fix an L > 0 with
supp ϕ ⊆ [−L, L], then we can write the integral as
Z Z
dx
ϕ(0) + ϕ0 (ξ) dx.
h<|x|<L x h<|x|<L

TheR first integral equals zero for all h > 0, and the second one converges
L
to −L ϕ0 (ξ) dx, by DC. Thus PV(1/x) is well defined, and we also
conclude that |(PV(1/x), ϕ)| ≤ 2Lkϕk1 . So we have indeed defined a
distribution.
Exercise 1.6. Show that PV(1/x) is not a function. Suggestion: If it
were, what would this function have to be equal to away from x = 0?

2. Operations on distributions
If f ∈ C 1 (R), then its derivative f 0 , being continuous, is a locally
integrableR function, so may be viewed as a distribution, which acts as
∞
(f 0 , ϕ) = −∞ f 0 (x)ϕ(x) dx. By an integration by parts, we can rewrite
this as Z ∞
0
(f , ϕ) = − f (x)ϕ0 (x) dx.
−∞

Exercise 2.1. Prove this in more detail. Why are there no boundary
terms?
This motivates:
Definition 2.1. Let u ∈ D0 (R). The distributional derivative u0 ∈ D0
is defined as the distribution (u0 , ϕ) = −(u, ϕ0 ).
We must in fact show that this indeed defines a new distribution,
but this is easy: if K ⊆ R is compact and |(u, ϕ)| ≤ CkϕkN for ϕ ∈ D
with supp ϕ ⊆ K, then |(u0 , ϕ)| ≤ CkϕkN +1 .
In the same way, we can more generally define partial derivatives of
distributions u ∈ D0 (U ), U ⊆ Rd as (∂u/∂xj , ϕ) = −(u, ∂ϕ/∂xj ).
Note that any distribution has (distributional) derivatives, in fact of
any order, since the derivatives are themselves distributions to which
the definition can be applied. For example, (δ 0 , ϕ) = −(δ, ϕ0 ) = −ϕ0 (0)
and, more generally, δ (n) (ϕ) = (−1)n ϕ(n) (0).
 DISTRIBUTIONS 5
2 2
Exercise 2.2. Show that u ∈ D0 (U ) satisfies ∂x∂j ∂x
u
k
= ∂x∂k ∂x
u
j
(that is,
higher order partial distributional derivatives can be taken in any or-
der).

In particular, this holds for arbitrary locally integrable functions. We

must keep in mind, though, that the distributional derivative could in
principle differ from the classical derivative in cases where the latter
also exists (in some sense); we will see examples of this later. However,
if f ∈ C 1 , then the distributional derivative of f agrees with its classical
derivative, and in fact this was the case that motivated the definition.
This option of being able to take arbitrary derivatives is one of the
main attractions of the theory of distributions. It is often useful even
when in the end it turns out no distributions were involved. A typical
situation would be that of a function f (x) that we defined in a com-
plicated way, and we would now like to differentiate it except that we
don’t know at this point if our function is actually differentiable. We
can then always take the distributional derivative without having to
worry about how to justify steps we would like to take.
Let’s now enjoy this new freedom by looking at a few (but relatively
harmless still) functions that are not differentiable everywhere in the
classical sense.

Example 2.1. The standard calculus example for a function that is not
differentiable at a point is f (x) = |x|. Clearly, f ∈ L1loc , so f does
have a distributional derivative, which we’ll denote simply by f 0 (but
perhaps more circumspect notation would have been u0f ). We compute
Z ∞
0 0
(f , ϕ) = −(f, ϕ ) = − |x|ϕ0 (x) dx
−∞
Z 0 Z L
0
= xϕ (x) dx − xϕ0 (x) dx
−L 0
Z 0 Z L
0 L
= xϕ(x) − −L
ϕ(x) dx − xϕ(x) 0
+ ϕ(x) dx
−L 0
Z ∞
= sgn(x)ϕ(x) dx,
−∞

with sgn(x) = 1 for x > 0 and sgn(x) = −1 for x < 0. We also took
L > 0 large enough so that supp ϕ ⊆ [−L, L] in this calculation. We
have found, unsurprisingly, that f 0 ∈ D0 is in fact a function, and it is
equal almost everywhere to the pointwise derivative f 0 (x) = sgn(x) of
f.
6 CHRISTIAN REMLING

Now let’s continue and compute f 00 = sgn0 ∈ D0 . From the definition

of the distributional derivative,
Z ∞ Z 0 Z ∞
0 0 0
(sgn , ϕ) = − sgn(x)ϕ (x) dx = ϕ (x) dx − ϕ0 (x) dx
−∞ −∞ 0
= 2ϕ(0) = 2(δ, ϕ).
In other words, f 00 = 2δ. This is not a function (compare Exercise 1.5).
Example 2.2. Now let’s look at f (x) = log |x|. This is a locally inte-
grable function. What is its distributional derivative f 0 ∈ D0 ? Away
from x = 0, f is smooth and has the classical derivative f 0 (x) = 1/x,
but this cannot be the answer to our question because this function is
not locally integrable, so does not generate a distribution in an obvious
way. Let’s look at this more closely. By DC,
Z ∞ Z
0 0
(f , ϕ) = − log |x|ϕ (x) dx = − lim log |x|ϕ0 (x) dx.
−∞ h→0+ |x|>h

We now again split this last integral into the two parts x < −h and
x > h and then integrate by parts in those. This produces
Z 
0 ϕ(x)
(f , ϕ) = lim dx + log h(ϕ(h) − ϕ(−h)) ,
h→0+ |x|>h x
and here the second term goes to zero because ϕ(±h) = ϕ(0) + O(h)
and h log h → 0.
Let’s summarize:
Theorem 2.2. In D0 , |x|0 = sgn(x), |x|00 = 2δ, (log |x|)0 = PV(1/x).
As our next example, let’s take an increasing function F : R → R.
Such an F is bounded on any compact set, so is clearly locally integrable
and thus generates a distribution. What is its distributional derivative?
Before we answer this, recall that a monotone F is differentiable, in
the classical sense, at almost all x ∈ R. Moreover, F generates a Borel
measure µ = mF , by setting µ((a, b]) = F (b) − F (a), if we also make
F right-continuous here. This changes F at at most countably many
points, so will not affect the distribution F ∈ D0 , and we may thus
indeed insist on this normalization at no cost. It will also be convenient
to make F (0) = 0, by adding a constant to the original function. It
seems intuitively clear that this should not affect F 0 , but let’s establish
this carefully before we proceed.
Theorem 2.3. Let u, v ∈ D0 , c ∈ C. Then (u+v)0 = u0 +v 0 , (cu)0 = cu0 ,
and c0 = 0.
 DISTRIBUTIONS 7

We of course add distributions and multiply them by constants in

the obvious way: for example (u + v)(ϕ) := u(ϕ) + v(ϕ).
Exercise 2.3. Prove Theorem 2.3.
So indeed (F + c)0 = F 0 , as anticipated.
With this further normalization, F (0) = 0, we have F (x) = µ((0, x])
for x ≥ 0 and F (x) = −µ((x, 0]) for x < 0. We are now ready to
compute
Z ∞
0
(F , ϕ) = − F (x)ϕ0 (x) dx
−∞
Z 0 Z ∞
0
(2.1) = µ((x, 0])ϕ (x) dx − µ((0, x])ϕ0 (x) dx.
−∞ 0

Let
R me look at the last integral in more detail: we can write µ((0, x]) =
(0,x]
dµ(t). We obtain an iterated (double) integral, to which we apply
Fubini-Tonelli. Note that the integration is over the set 0 < t ≤ x, so
this produces
Z Z ∞ Z
0
− dµ(t) dx ϕ (x) = ϕ(t) dµ(t).
(0,∞) t (0,∞)

Of course, the Rfirst integral from (2.1) can be given a similar treatment:
it is equal to (−∞,0] ϕ(t) dµ(t). Putting things together, we thus see
that (F 0 , ϕ) = R ϕ(x) dµ(x). In other words, F 0 = µ, the distribution
R
generated by the measure that F induces. Let’s state this formally.
Theorem 2.4. If F : R → R is increasing, then F 0 = µ, with
µ((a, b]) = F (b+) − F (a+).
In particular, note that the distributional derivative need not equal
the pointwise derivative. This holds only if µ is absolutely continuous.
If µ has a singular part, then F 0 is not a function. (You may wonder
how a measure can ever be a function, but in fact it happens, with our
use of terminology, since the measure f dx and the function f are the
same distribution, so in that sense the measure f dx is a function also.)
We in fact already saw this principle in action in the simple example
F (x) = sgn(x), which gave F 0 = 2δ (and this of course is a measure,
though we didn’t emphasize this originally.)
This condition, of having a positive measure as its distributional de-
rivative, characterizes the increasing functions. This criterion is not
particularly useful (usually there will be easier ways to check that a
given function is increasing than computing its distributional deriva-
tive), but it is elegant and satisfying.
8 CHRISTIAN REMLING

Theorem 2.5. Let F : R → R be a locally integrable function. Then

F is increasing (more precisely, F is almost everywhere equal to an
increasing function) if and only if F 0 ∈ D0 is a Borel measure on R.
The qualification spelled out in parentheses is unavoidable since
changing F on a null set will not affect the distribution F ∈ D0 , so
a criterion involving the distributional derivative is clearly insensitive
to such a change.
Proof. We already showed, and stated as Theorem 2.4, that an increas-
ing function has a measure as its derivative. Conversely, suppose now
that F ∈ L1loc has distributional derivative F 0 = µ, for some Borel
measure µ. Let G be an increasing function that corresponds to µ. In
more concrete style, we can set
(
µ((0, x]) x≥0
G(x) = .
−µ((x, 0]) x < 0
By Theorem 2.4, G0 = µ in D0 . I now refer to the general fact, stated
as Theorem 2.6 below, that a distribution can be recovered from its
derivative, up to a constant, to conclude that F (x) = G(x) + c a.e.,
for some c ∈ R; of course, I also use Theorem 1.2 in this step. So F is
almost everywhere equal to an increasing function, as claimed. 
Theorem 2.6. Let u ∈ D0 (R) and suppose that u0 = 0. Then u = c.
More explicitly, the claim Ris that u is the constant function c, for
some c ∈ C, that is, (u, ϕ) = cϕ.
∞ 0 ∞
Lemma 2.7. R ∞Let ϕ ∈ C0 (R). Then ϕ = ψ for some ψ ∈ C0 (R) if
and only if −∞ ϕ(x) dx = 0.
R 0
Proof. It is of course clear
Rx that ψ = 0. Conversely, suppose that
ϕ = 0. Let ψ(x) = −∞ ϕ(t) dt. It is then clear that ψ ∈ C ∞ , ψ 0 = ϕ,
R

and ψ(x) = 0 for all small x. Our assumption on ϕ makes sure that
ψ(x) = 0 also for all large x. So ψ has compact support. 
Proof of Theorem 2.6. Fix once and for all a ϕ0 ∈ C0∞ withR ϕ0 = 1.
R
∞
R arbitrary ϕ ∈ C0 , write ϕ = aϕ
For 0 + (ϕ − aϕ0 ), with a = ϕ. Then
(ϕ − aϕ0 ) = 0, so ϕ − aϕ0 = ψ 0 for some ψ ∈ C0∞ , by the lemma.
However, (u, ψ 0 ) = −(u0 , ψ) = 0, and thus
Z ∞
(u, ϕ) = a(u, ϕ0 ) = (u, ϕ0 ) ϕ(x) dx.
−∞

This says that u = c, as claimed, with c = (u, ϕ0 ). 

 DISTRIBUTIONS 9

Monotonicity can of course also be characterized in terms of the

derivative in a more classical setting, and in fact this is very easy:
if F : R → R is differentiable, then F is increasing if and only if
F 0 (x) ≥ 0. Theorem 2.5 is actually a distributional analog of this
statement, though this is not immediately apparent. To make this
connection explicit, we need one more definition and a major result.
Definition 2.8. A distribution u ∈ D0 is called positive (notation:
u ≥ 0) if u(ϕ) ≥ 0 for all test functions ϕ(x) ≥ 0.
Theorem 2.9. A distribution u ∈ D0 (U ) is positive if and only if it is
a (positive) Borel measure µ on U .
R
One direction is obvious: if ϕ ≥ 0, then also U ϕ(x) dµ(x) ≥ 0, so a
measure defines a positive contribution. The converse is one version of
the Riesz representation theorem. Its proof is long and technical, and
I don’t want to discuss it here. By combining the Riesz representation
theorem with Theorem 2.5, we obtain, as promised, the following.
Corollary 2.10. A locally integrable function F : R → R is increasing
if and only if F 0 ≥ 0 in D0 .
These ideas produce even more satisfying results when applied to
other classes of functions. We call f : R → C (locally) absolutely
continuous and write f ∈ AC if
Z x
(2.2) f (x) = f (0) + g(t) dt
0
for some g ∈ L1loc . In much the same way as above, we can then prove:
Theorem 2.11. Let f ∈ L1loc (R). Then f ∈ AC if and only if the dis-
tributional derivative of f is a function. In this case, the distributional
derivative agrees with the pointwise derivative, which will exist almost
everywhere.
Proof. If f ∈ AC, so (2.2) holds, then we can compute f 0 ∈ D0 exactly
as above as
Z ∞
0
(f , ϕ) = − f (x)ϕ0 (x) dx
−∞
Z 0 Z 0 Z ∞ Z x
0 0
= dx ϕ (x) dt g(t) − dx ϕ (x) dt g(t).
−∞ x 0 0

The term f (0) from (2.2) doesn’t contribute because ϕ0 = 0. We

R
change the order of integration, so for example the first integral be-
comes Z 0 Z t Z 0
dt g(t) dx ϕ0 (x) = g(t)ϕ(t) dt,
−∞ −∞ −∞
10 CHRISTIAN REMLING

and after putting everything back together we see that f 0 = g ∈ L1loc ,

as claimed. From general facts about absolutely continuous functions
we also know that g(x) = f 0 (x), the pointwise derivative, almost ev-
erywhere.
0 1 0
R xConversely, if f = g ∈ Lloc in D , then we introduce h(x) 0
=
0
g(t) dt. By what we just showed, we then know that also h = g
0
in D , so f = h + c by Theorem 2.6. 
As a preparation for our discussion of the analogous result for BV
functions, we make one more general definition.
Definition 2.12. Let u ∈ D0 (U ) and let V ⊆ U be open. Then the
restriction v = u|V ∈ D0 (V ) of u to V is defined as (v, ϕ) = (u, ϕ0 ),
with ϕ0 (x) = ϕ(x) for x ∈ V and ϕ0 (x) = 0 for x ∈ U \ V .
Or, to say this in somewhat less formal style, we restrict a distribu-
tion to V by only applying it to test functions ϕ with supp ϕ ⊆ V . Of
course, a distribution cannot be restricted in literally the same way as
a function since it cannot be evaluated at individual points x ∈ U .
Example 2.3. δ|R\{0} = 0, and this follows at once from the definition:
if 0 ∈
/ supp ϕ, then ϕ(0) = (δ, ϕ) = 0, so δ indeed acts as the zero
distribution on the smaller test function space C0∞ (R \ {0}).
As one would have hoped, distributions can be recovered from their
local restrictions. This is an interesting general fact, and we will also
need it soon when we discuss BV functions.
Theorem 2.13. Let u ∈ D0 (U ), and assume that every x ∈ U has a
neigborhood V = Vx such that u|V = 0. Then u = 0.
Lemma 2.14. Let Wα ⊆ U be an openP cover of U . Then there are ϕn ∈
∞
C0 (U ), 0 ≤ ϕn ≤ 1, such that: (1) n≥1 ϕn (x) = 1; (2) supp ϕn ⊆
Wαn for some αn , for each n ≥ 1; (3) if K ⊆ U is compact, then
PN
n=1 ϕn (x) = 1 for x ∈ K for some N = N (K) ≥ 1.

Such a collection of functions is often called a partition of unity,

subordinate to {Wα }.
Sketch of proof. We can find countably manySopen balls Bn such that
each of them is contained in some Wα , and (1/2)Bn = U . Here cB
denotes the ball with the same center as B and c times the radius.
Pick functions ψn ∈ C0∞ (U ) with 0 ≤ ψn ≤ 1, ψn = 1 on (1/2)Bn and
ψn = 0 on ((2/3)Bn )c . Then define
ϕ1 = ψ1 , ϕ2 = (1 − ψ1 )ψ2 , . . . , ϕn = (1 − ψ1 ) · · · (1 − ψn−1 )ψn .
 DISTRIBUTIONS 11

Property (2) is then clear: supp ϕn ⊆ Bn ⊆ Wαn . Next, I claim that

ϕ1 + . . . ϕN = 1 − (1 − ψ1 ) · · · (1 − ψN ),
and it is in fact straightforward to establish
SN this by an induction on N .
So in particular ϕ1 + . . . + ϕN = 1 on n=1 (1/2)Bn . This implies (1)
and also (3) since any compact K ⊆ U is contained in such a union for
some N . 

Proof of Theorem 2.13. Apply the lemma to {Wα } = {Vx : x ∈ U }.

Let ϕP∈ C0∞ (U ) be an arbitrary test function. We can then write
ϕ = ϕϕn . In fact, property (3), applied to K = supp ϕ, shows
that ϕ = N
P
n=1 ϕϕn for some N = N (ϕ). The products ϕϕn are test
functions themselves, and each of them has its support contained in
some Vx , by (2). Thus
N
X
(u, ϕ) = (u, ϕϕn ) = 0,
n=1

as claimed. 

We say that a function f : R → R is (locally) of bounded variation if

f ∈ BV [a, b] for all [a, b] ⊆ R, and we write f ∈ BV in this case. Recall
also that f ∈ BV if and only if f is the difference of two increasing
functions. In particular, BV functions are locally bounded and thus
also locally integrable.
On bounded intervals, BV functions correspond to signed measures
in the same way increasing functions correspond to positive measures.
On the real line, this is not quite true as examples such as f (x) = |x|
demonstrate: This function is monotone on both half lines, and the
associated positive measures are µ1 = χ(−∞,0) dx, µ2 = χ(0,∞) dx, but
µ2 − µ1 is not a signed measure, as both the positive and the negative
parts are infinite.
Theorem 2.15. Let f : R → R be locally integrable. Then f ∈ BV
(more precisely: f = g a.e. for some g ∈ BV ) if and only if the
restriction of f 0 ∈ D0 (R) to (−L, L) is a finite signed measure for
every L > 0.
Proof. If f ∈ BV , then we can recognize its derivative f 0 (locally) as
the signed measure µ((a, b]) = f (b+) − f (a+) by the same argument
that we already used a number of times, for example in the proof of
Theorem 2.4. Or we could write f = f1 − f2 , with fj increasing, and
then refer to this theorem directly (not its proof).
12 CHRISTIAN REMLING

The converse is also proved by the same arguments as above: If

f 0 = µ on (−L, L), then we define a BV function as
(
µ((0, x]) x≥0
g(x) = ;
−µ((x, 0]) x < 0
note that we can in fact consistently define g on all of R in this way
since the µ = µL will be compatible in the sense that if L2 > L1 and
A ⊆ (−L1 , L1 ), then µL1 (A) = µL2 (A), and this is true because these
measures were obtained as restrictions of one and the same distribution
f 0 ∈ D0 (R).
Exercise 2.4. Strictly speaking, I am using the following (obvious look-
ing) fact here: If W ⊆ V ⊆ U are open sets and u ∈ D0 (U ), then
u|W = (u|V ) |W . Prove this.
Now g has the same distributional derivative as f , after restricting
to (−L, L). By Theorem 2.13, f 0 − g 0 = 0. By Theorem 2.6, this means
that f = g + c ∈ BV . 
The philosophy that led to Definition 2.1 lets us define several other
operations on distributions in a natural way. The guiding principle is:
move the operation over to the test function, in such a way that it gives
the right answer for distributions that are nice functions.
We write (τa f )(x) = f (x − a) for the translation of a function f :
R →RC by a ∈ Rd . For a well
d
R behaved f and a test function ϕ we then
have f (x − a)ϕ(x) dx = f (x)ϕ(x + a) dx. This suggests to define
τa u, for u ∈ D0 (Rd ) and a ∈ Rd as (τa u, ϕ) = (u, τ−a ϕ).
Exercise 2.5. Check that indeed τa u ∈ D0 . Then give an equally natural
definition of the dilation δa u of a distribution that is based on the action
(δa f )(x) = f (ax) (a > 0) on functions.
Definition 2.16. Let u ∈ D0 (R), ϕ ∈ C0∞ (R). Then we define the
convolution of u and ϕ as (u ∗ ϕ)(x) = (u, τx Rϕ), with (Rϕ)(x) =
ϕ(−x).
R
If u ∈ L1loc , then this becomes (u∗ϕ)(x) = u(t)ϕ(x−t) dt, so reduces
to the usual definition of the convolution. Note that the convolution
of a distribution and a test function is a function.
This convolution has the same basic properties as the convolution of
two functions:
Theorem 2.17. Let u ∈ D0 , ϕ ∈ C0∞ . Then u ∗ ϕ ∈ C ∞ , and
(u ∗ ϕ)0 = u0 ∗ ϕ = u ∗ ϕ0 .
 DISTRIBUTIONS 13

Proof. We have
 
1 ϕ(x + h − t) − ϕ(x − t)
((u ∗ ϕ)(x + h) − (u ∗ ϕ)(x)) = u, ,
h h
and here we apply u to the difference quotient as a function of t, for
fixed x, h. If we now send h → 0, then this expression will converge
to ϕ0 (x − t). This is clear in pointwise sense, but this isn’t quite good
enough here. Rather, what we need is that the functions
ϕ(x + h − t) − ϕ(x − t)
gh (t) = − ϕ0 (x − t) ∈ C0∞
h
together with all their derivatives converge to zero uniformly in t and
their supports are contained in one fixed compact set. This is condition
(b) from Exercise 1.1, and by the result of that Exercise, we can then
conclude that (u, gh ) → 0 as h → 0.
Exercise 2.6. Prove this claim about gh in more detail.
It follows that u ∗ ϕ is indeed differentiable, and (u ∗ ϕ)0 = u ∗ ϕ0 .
Since this is still of the same general form (convolution of a distribution
with a test function), we can repeat this argument, and we have in fact
shown that u ∗ ϕ ∈ C ∞ .
Finally,
u0 ∗ ϕ = (u0 , τx Rϕ) = −(u, (τx Rϕ)0 ) = (u, τx Rϕ0 ) = u ∗ ϕ0 .

Definition 2.18. Let g ∈ C ∞ (U ), u ∈ D0 (U ). Then gu ∈ D0 is defined
as the distribution (gu, ϕ) = (u, gϕ).
R
R This definition is of course motivated by the trivial formula (gu)ϕ =
u(gϕ), which is valid for nice functions u. Note that we have defined
the product of a distribution and a (smooth) function; there is no natu-
ral notion of a product of two distributions. This is not surprising. For
example, it is indeed hard to come up with a reasonable interpretation
of what δ · δ might be.
Exercise 2.7. Show that gu indeed is a distribution.
Exercise 2.8. Show that the product rule is valid in the (limited) con-
text of Definition 2.18: if g ∈ C ∞ , u ∈ D0 , then (gu)0 = g 0 u + gu0 .
Example 2.4. What is gδ? We compute (gδ, ϕ) = (δ, gϕ) = g(0)ϕ(0).
This is what the (constant) multiple g(0)δ would have done on ϕ, so
gδ = g(0)δ.
Exercise 2.9. Compute similarly gδ 0 , for g ∈ C ∞ (R).
14 CHRISTIAN REMLING

3. Convergence of distributions
Definition 3.1. Let un , u ∈ D0 . We say that un → u (in D0 , or in the
sense of distributions) if (un , ϕ) → (u, ϕ) for all ϕ ∈ D.
This is a rather weak requirement. For example, einx → 0 in D0 , if
we view these functions as distributions.
Exercise 3.1. Prove this. Suggestion: Riemann-Lebesgue lemma
Can you in fact also show that nN einx → 0 in D0 for any N ≥ 1?
Similarly, for example nδn → 0, where (δn , ϕ) = ϕ(n). To confirm
this, simply observe that (nδn , ϕ) = nϕ(n) will be zero once n is outside
the support of ϕ.
Example 3.1. Let fn (x) = nχ(−1/2n,1/2n) (x). Then fn → δ. To prove
this, we use the mean value theorem to write ϕ(x) = ϕ(0) + ϕ0 (ξ)x =
ϕ(0) + O(x) and compute
Z 1/2n Z 1/2n
(fn , ϕ) = n ϕ(x) dx = ϕ(0)+n O(1/n) dx = ϕ(0)+O(1/n),
−1/2n −1/2n

so this converges to ϕ(0) as n → ∞, as claimed.

1
R 3.2. Prove, more ambitiously, that for any f ∈ L (R), we have
Exercise
fn → ( f )δ, with fn (x) = nf (nx).
Theorem 3.2. If un → u in D0 , then also u0n → u0 .
So we can say that the derivative is a continuous operation on D0 .
Despite the slightly surprising character of this statement, its proof is
extremely simple: (u0n , ϕ) = −(un , ϕ0 ) → −(u, ϕ0 ) = (u0 , ϕ) 
(Differentiation is also continuous on D, if this test function space is
endowed with the topology T mentioned before Exercise 1.1, so Theo-
rem 3.2 is not so very surprising after all.)
Example 3.2. Let’s return to the functions fn → δ from the previous
Example. We have fn0 = n(δ−1/2n − δ1/2n ), and this can be established
by a straightforward calculation, similar to the one that showed that
sgn0 = 2δ. Or simply notice that fn ∈ BV and recall that the distribu-
tional derivative of such a function is the associated (signed) measure.
Theorem 3.2 now shows that
n(δ−1/2n + δ1/2n ) → δ 0 .
This can easily be checked directly: n(ϕ(−1/2n) − ϕ(1/2n)) → −ϕ0 (0)
since ϕ(x) = ϕ(0) + ϕ0 (0)x + O(x2 ), by Taylor’s theorem.
 DISTRIBUTIONS 15

After these easy introductory examples, we are now ready for a more
substantial convergence statement.
Theorem 3.3 (Sokhotski-Plemelj formula). In D0 (R),
1 1
lim = PV + iπδ.
h→0+ x − ih x
Proof. We write
1 x h
= 2 2
+i 2 ,
x − ih x +h x + h2
and we’ll treat the real and imaginary parts separately. The imaginary
part is actually covered by (the continuous version of) Exercise 3.2
because, in the notation used there,
R h/(x2 + h2 ) = f1/h (x), with f (x) =
1/(1 + x2 ), and (by calculus) f = π. But a direct argument is also
easy:
Z ∞ Z ∞ Z ∞
h 1 dt
2 2
ϕ(x) dx = 2
ϕ(ht) dt → ϕ(0) 2
= πϕ(0)
−∞ x + h −∞ 1 + t −∞ 1 + t

The convergence follows from DC, and in this step we use that ϕ is
continuous and bounded.
Next, we look at the real part. Since we anticipate the limit being
the principal value distribution, we first establish that we can remove
a small interval about x = 0. More precisely,
Z h Z h Z h
x x x
2 2
ϕ(x) dx = ϕ(0) 2 2
dx + 2 2
O(x) dx,
−h x + h −h x + h −h x + h

and this goes to zero because the first integral on the right-hand side
has this value, and the second one is O(h).
So it now suffices to show that
Z  
x 1
lim − ϕ(x) dx = 0.
h→0+ |x|>h x2 + h2 x
To do this, we write
x 1 −h2
− =
x2 + h2 x (x2 + h2 )x
and again use the Taylor expansion ϕ(x) = ϕ(0) + O(x) of the test
function. The term with ϕ(0) doesn’t contribute
R since the integrand is
odd, and what is then left is of order O(h2 ) |x|>h dx/(x2 + h2 ). Since
R
R
h/(x2 + h2 ) dx = π, this goes to zero. 
Exercise 3.3. Show that log(x2 + h2 ) → 2 log |x| in D0 .
16 CHRISTIAN REMLING

A slightly slicker proof of the second part can be based on this Ex-
ercise and Theorem 3.2: Take derivatives to obtain
(log(x2 + h2 ))0 → 2(log |x|)0 .
Now recall that this latter distribution equals 2(PV(1/x)); see Theorem
2.2. Moreover, log(x2 + h2 ) is a smooth function for fixed h > 0, so its
distributional derivative is its classical derivative, and thus
2x
(log(x2 + h2 ))0 = 2 .
x + h2
We earlier defined convolutions, and one of their main uses (in the
classical setting) is the approximation of general functions by nice ones.
The convolution from Definition 2.16 can be put to similar use in the
realm of distributions.
Theorem 3.4. Let u ∈ D0 (R), and fix a ϕ ∈ C0∞ (R) with ϕ = 1.
R
Put ϕn (x) = nϕ(nx). Then u ∗ ϕn → u in D0 .
We originally defined u ∗ ϕ as a function (not a distribution), but
we then saw that this is a smooth function, so is in particular locally
integrable and can be viewed as a distribution after all. This is of
course the interpretation of u ∗ ϕn that we need to adopt here, for the
statement to make sense.
So we see that any distribution can be approximated by smooth
functions. We can do even better:
Exercise 3.4. (a) Fix a ψ ∈ C0∞ (R) with ψ(x) = 1 in a neighborhood
of x = 0, and let ψn (x) = ψ(x/n). Show that ψn u → u for any u ∈ D0 .
(b) Combine this with Theorem 3.4 to conclude that for any u ∈
D (R), there are ϕn ∈ C0∞ (R), ϕn → u in D0 .
0

Sketch of proof. Let ψ be a test function. We want to show that (u ∗

ϕn , ψ) → (u, ψ) or, if we write this out,
Z ∞
n (u, ϕ(n(x − t))ψ(x) dx → (u, ψ).
−∞

Here it is again understood that u is applied to ϕ(n(x−t)) as a function

of t (not x). I would now like to rewrite the left-hand side as
 Z ∞ 
(3.1) u, n ϕ(n(x − t))ψ(x) dx ;
−∞
this looks plausible, based on the linearity of u. To justify this step,
we would have to write the integral as a limit of Riemann sums, and
the limit must take place in D, in the sense of Exercise 1.1, so that we
can then refer to the continuity of u. I’ll leave the matter at that.
 DISTRIBUTIONS 17
R
The final step then is to note that n ϕ(n(x − t))ψ(x) dx → ψ(x).
Again, this looks exceedingly plausible; for example, it is clear that
we will have pointwise (in fact: uniform) convergence, but even that
isn’t quite good enough here because we want to refer to Exercise 1.1
one more time and so need to establish convergence in the sense of
condition (b) from that Exercise. I will again skip all the details.
If all this is granted, then it indeed follows that (3.1) converges to
(u, ψ), and we are done.
(Note that even if this seemed unconvincing, the general plan of the
proof was simple and straightforward: move the operation of convolving
with the approximate identity ϕn over to the test function, and use that
convergence on that level is clear and that u is continuous.) 

Let’s try this out for u = δ: We obtain approximations un = δ ∗ ϕn ,

and what are these equal to? In fact, what is δ ∗ ψ? By the definition
of the convolution, (δ ∗ ψ)(x) = (δ, ψ(x − t)) = ψ(x). So we simply
recover the approximations ϕn → δ that we already discussed above,
in Exercise 3.2.
We have also inadvertently proved the interesting fact that δ ∗ϕ = ϕ,
that is, δ acts as the identity element for the convolution product.
Exercise 3.5. Recall the definition of the translation τh u of a distribu-
tion from Section 2. Then show that for any u ∈ D0 (R),
1
(τ−h u − u) → u0
h
as h → 0 in D0 . (So distributional derivatives can also be computed as
limits of difference quotients.)

4. Tempered distributions
The Fourier transform of an f ∈ L1 (Rd ) is defined as
Z
(4.1) fb(t) = f (x)e−2πit·x dx.
Rd

The Fourier inversion theorem says that if f, fb ∈ L1 , then

Z
f (x) = fb(t)e2πit·x dt.
Rd

We would now like to define the Fourier transform u b also for distribu-
tions u ∈ D0 . We follow the usual strategy of moving the operation
over to the test function, taking the case of a nice function u as our
18 CHRISTIAN REMLING

guideline. In this situation,

Z Z Z
u
b(x)ϕ(x) dx = dx ϕ(x) dt u(t)e−2πit·x
Rd d d
ZR Z R
= dt u(t) dx ϕ(x)e−2πit·x = (u, ϕ),
b
Rd Rd

so the definition we want to make is (b u, ϕ) = (u, ϕ).

b Unfortunately,
this doesn’t work: ϕb is not guaranteed to be a test function. It will be
smooth, but in fact ϕ b will never have compact support, unless ϕ ≡ 0.
In other words, the Fourier transform is not an operation on D, and
this suggests the path we are going to take: we need a new test function
space, and the appropriate one for our current purposes is the Schwartz
space S(Rd ). It is defined as follows:
S(Rd ) = f : Rd → C : f ∈ C ∞ , kf kj,n < ∞ for all j, n ≥ 0 ,


kf kj,n = sup (1 + |x|j )|∂ α f (x)|

x∈Rd ,|α|≤n

So S is the space of all smooth functions that together with all their
derivatives decay faster than any power. Clearly, C0∞ ⊆ S, but S also
contains (many) not compactly supported functions such as f (x) =
2
e−x .
The norms k · kj,n can be combined into a metric d, for example as
follows:
X kf − gkj,n
d(f, g) = 2−j−n
j,n≥0
1 + kf − gkj,n

Exercise 4.1. Show that d(fk , f ) → 0 if and only if kfk − f kj,n → 0 for
all j, n ≥ 0.
That we have made the right choice is confirmed by
Theorem 4.1. The Fourier transform is a continuous bijection on S.
I don’t want to prove this in detail here, so will just make a few
general remarks. The proof is straightforward, but somewhat tedious
to write down. Essentially, we have to show that the Fourier transform
maps S continuously back into itself; the rest will then follow quickly
from Fourier inversion. The two defining properties, rapid decay and
smoothness, are dual to each other in the sense that one will become
the other after taking Fourier transforms, so the combination of both
is preserved. If this is done carefully, we will then also obtain control
on kfbkj,n in terms of the kf kj 0 ,n0 , and this gives the asserted continuity
of the Fourier transform.
 DISTRIBUTIONS 19

Definition 4.2. A tempered distribution u ∈ S 0 is a continuous linear

functional u : S → C. Here, continuity means that there are C ≥ 0
and j, n ≥ 0 such that |u(ϕ)| ≤ Ckϕkj,n .
Exercise 4.2. Let u : S → C be a linear functional on S. Show that
the following are equivalent: (a) u ∈ S 0 ; (b) If ϕn ∈ S and kϕkj,N → 0
for all j, N ≥ 0, then u(ϕn ) → 0; (c) u is continuous with respect to d.
Of course, the reason for doing all this was to have the following
definition available.
Definition 4.3. Let u ∈ S 0 . Then the Fourier transform u
b ∈ S 0 is
defined as (b
u, ϕ) = (u, ϕ).
b
Note that this does define a tempered distribution: the map ϕ 7→
u(ϕ)
b is continuous, being the composition of the two continuous maps
ϕ 7→ ϕb and u.
As before, a function f ∈ L1 generates a tempered R distribution
(which we’ll frequently identify with f ) R(f, ϕ) = f ϕ. RIndeed, this
map is obviously linear, and |(f, ϕ)| ≤ ( |f |) sup |ϕ| ≤ ( |f |)kϕk0,0 .
We repeat the motivating calculation we already did at the beginning of
this section to confirm that the distributional Fourier transform fb ∈ S 0
satisfies
Z Z Z
b dt = dt f (t) dx ϕ(x)e−2πit·x
(f , ϕ) = f (t)ϕ(t)
b
Z Z
= dx ϕ(x) dt f (t)e−2πit·x .

In other words, fb is a function, and this function is still given by the

original definition (4.1).
Next, let’s look at u = δ. First of all, observe that |δ(ϕ)| = |ϕ(0)| ≤
kϕk0,0 , so δ ∈ S 0 . (Strictly speaking, we have two δ’s at this point:
the original distribution δ ∈ D0 , and its extension to S, which is the
tempered distribution δ ∈ S 0 . This confusing situation will be rectified
in a moment; let’s not get distracted by it right now.) From the defi-
nition of the Fourier transform, we have δ(ϕ)b = δ(ϕ)
b = ϕ(0),
b but this
R
is ϕ(x) dx, by (4.1). In other words, δ = 1 (the function f (x) ≡ 1).
b
R
Similarly, b
1(ϕ) = ϕ b = ϕ(0), by Fourier inversion. We summarize:
1 = δ in S 0 (Rd ).
Theorem 4.4. We have δb = 1, b
We could also have tried to argue based on (4.1) that
Z
b = δ(x)e−2πitx dx = e−2πitx
δ(t) = 1,
x=0
20 CHRISTIAN REMLING

and this gives the correct answer, but is not a formally correct ma-
nipulation because we tried to apply δ to e−2πitx , which is not a test
function.
Now let’s take a closer look at the relation between S 0 and D0 . We
already observed that D ( S. Because of this, any tempered distri-
bution u : S → C can be restricted to D (and don’t confuse this type
of restriction with the one from Definition 2.12). Then u|D will lie in
D0 : By assumption, we have |u(ϕ)| ≤ Ckϕkj,n for some j, n. If now
also ϕ ∈ D, supp ϕ ⊆ B(0, R), then kϕkj,n ≤ (1 + Rj )kϕkn , and this
latter norm is the one from Definition 1.1. This verifies that u ∈ D0 ,
as claimed; more precisely, it is the restriction u|D that is in D0 , but
we will often not make this distinction explicit in the notation.
Moreover, the original tempered distribution u ∈ S 0 can be recovered
from its restriction u|D ∈ D0 . This follows from the fact that D ⊆ S is
dense in S with respect to d.
Exercise 4.3. Prove this. More explicitly, suppose that ϕ ∈ S, and
then show that there are ϕk ∈ C0∞ such that kϕk − ϕkj,n → 0 for each
j, n ≥ 0. Suggestion: Fix a ψ ∈ C0∞ with ψ = 1 near x = 0 and then
try ϕk (x) = ψ(x/k)ϕ(x).
However, if we conversely start out with a u ∈ D0 , then this may
or may not have an extension to a tempered distribution u0 ∈ S 0 ; if
it does, then u0 is unique, as we just discussed. As an example for
such a u ∈ D0 with no continuousRlinear extension, we can consider
the function u(x) = ex . So u(ϕ) = ex ϕ(x) dx, and it is already clear
intuitively what will go wrong here: there seems to be only one natural
way to attempt an extension to S, namely, try to use this formula for
general ϕ ∈ S also, but then such a function is not guaranteed to have
enough decay to make the integral convergent.
To make a proof out ofR this intuition, we need to proceed differently.
Fix a ψ ∈ C0∞ (R) with ψ 6= 0, and let ϕn (x) = ψ(x − n)e−x ∈ C0∞ ⊆
S. Then kϕn kj,N → 0.
Exercise 4.4. Prove this in detail.
R R
However, u(ϕn ) = ψ(x − n) dx = ψ(x) dx does not converge to
zero. So u has no continuous extension to S.
What we actually showed is that u is already discontinuous on D,
at ϕ = 0, when we use the metric d on D. So the general message is
that the continuity requirement on a tempered distribution u ∈ S 0 is
stronger than the one on a u ∈ D0 (when only test functions ϕ ∈ D are
considered). Let’s summarize:
 DISTRIBUTIONS 21

Theorem 4.5. The restriction of a tempered distribution to D is a dis-

tribution. Conversely, a distribution may or may not have a continuous
linear extension to S; if it does, then the extension is unique.
It is best to think of this as saying that every tempered distribu-
tion is a distribution, but, conversely, only some distributions are also
tempered (in this version, we don’t distinguish between a distribution
and its restriction or extension, which is justified by the uniqueness of
these). Symbolically, we can write S 0 ( D0 , which is a good way to
commit this whole discussion to memory, as long as you don’t take this
inclusion literally.
The obstacle illustrated by the example u(x) = ex ∈ D0 , u ∈ / S0
is the only one that can prevent a distribution from being tempered:
too rapid growth near infinity. For example, if f ∈ L1loc has at most
polynomial growth, |f (x)| ≤ C(1 + |x|N ) for some N ≥ 0, then f ∈ S 0 .
Exercise 4.5. Prove this.
Exercise 4.6. Let f ∈ Lp (Rd ), for some 1 ≤ p ≤ ∞. Show that f ∈ S 0 .
Another statement along these lines can be established with the help
of the notion of the support of a distribution.
Definition 4.6. Let u ∈ D0 (U ). The support of u is defined as
[ c
supp u = V ,
where the union is over all open V ⊆ U with u|V = 0.
Exercise 4.7. Show that if f ∈ C(Rd ), then this agrees with the usual
notion of the support of a function as the closure of the set {x : f (x) 6=
0}.
By Theorem 2.13, if W = (supp u)c , then u|W = 0. Conversely, if
u|W = 0 for an open set W , then W ∩ supp u = ∅, by the definition of
the support. This gives a description of the complement of the support
as the largest open set on which u is zero.
Example 4.1. supp δ = {0}: indeed, δ R\{0} = 0, as we observed earlier.
On the other hand, if U is open and 0 ∈ U , then there are ϕ ∈ C0∞ (U )
with ϕ(0) 6= 0, so 0 ∈ supp δ.
Theorem 4.7. Suppose that u ∈ D0 (Rd ) has compact support. Then
u ∈ S 0.
Here, I have again applied the convenient convention of not distin-
guishing between u ∈ D0 and its unique extension to S 0 . A more
explicit version of the statement would be: u ∈ D0 has a continuous
linear extension to S.
22 CHRISTIAN REMLING

Sketch of proof. Fix a ψ ∈ C0∞ with ψ = 1 on an open set containing

supp u. Then u = ψu since for any ϕ ∈ D, we can write
ϕ = ψϕ + (1 − ψ)ϕ,
and the second function on the right-hand side is annihilated by u
because its support is contained in (supp u)c . Now supp (ψϕ) ⊆ K for
a fixed compact K, for all ϕ ∈ D, and thus |u(ϕ)| ≤ CkψϕkN (the
point here is that we have one C and one N that work for all ϕ ∈ D).
Now ψ and all its derivatives are bounded, so kψϕkN ≤ Dkϕk0,N ,
and then the inequality |u(ϕ)| ≤ Bkϕk0,N allows us to extend u con-
tinuously to S, by approximating a general θ ∈ S by ϕn ∈ D, so
kϕn − θkj,k → 0. We can then define u(θ) = lim u(ϕn ). (It needs to be
checked that this limit exists and is independent of the approximating
sequence, but I don’t want to go into these details here.) 
Theorem 4.8. Suppose that u ∈ S 0 (R), g ∈ C ∞ (R), |g (j) (x)| ≤ C(1 +
|x|)Nj (j ≥ 0). Then also u0 , gu ∈ S 0 .
Strictly speaking, we haven’t even defined these operations for tem-
pered distributions yet, but of course there are two obvious answers to
this complaint. We can just mimic the old definitions (and then The-
orem 4.8 is claiming that everything is well defined in S 0 also), or we
refer to Theorem 4.5 and proceed as follows: consider the restriction
u ∈ D0 of u ∈ S 0 , and take its distributional derivative (or multiply
it by g). Then the claim of Theorem 4.8 is that the new distributions
obtained in this way are also tempered.

Proof. To establish the (important) claim about u0 , we only need to

observe that kϕ0 kj,N ≤ kϕkj,N +1 . So if |(u, ϕ)| ≤ Ckϕkj,N , then also
|(u0 , ϕn )| = |(u, ϕ0n )| ≤ Ckϕkj,N +1 ,
as required.
As for the product gu, we can similarly observe that kgϕkj,n ≤
Dkϕkj+M,n , with M = max0≤k≤n Nk , and then argue as above. 

While distributions that do not grow too fast near infinity will be
tempered, the converse is not true. Consider the function f (x) =
ex cos ex . This does become large (some of the time, at least) for large
x. However, we have f ∈ S 0 anyway, and this follows from Theorem
4.8 because f = g 0 , with g = sin ex ∈ S 0 .
Definition 4.9. Let un , u ∈ S 0 . We say that un → u in S 0 if (un , ϕ) →
(u, ϕ) for all ϕ ∈ S.
 DISTRIBUTIONS 23

Of course, this is the expected analog of Definition 3.1. As a con-

sequence, if un → u in S 0 , then also un → u in D0 because this is the
same condition, except that we now only test on ϕ ∈ D. There is a
pitfall that must be avoided in this context: if un , u ∈ S 0 and un → u
in D0 , then it does not follow that un → u also in S 0 . In other words,
convergence in S 0 is a stronger condition than convergence in D0 , and it
does not simply follow from the fact that all the distributions involved
are tempered.
For a concrete example, consider the tempered distributions un =
2
en δn . As elements of D0 , these converge to (the tempered distribution)
2
u = 0 in D0 (why?), but un 6→ 0 in S 0 as (un , e−x ) = 1.
Let’s now develop the theory of the Fourier transform in S 0 a bit.
We define the convolution of a tempered distribution u ∈ S 0 with a
test function ϕ ∈ S as expected as
(u ∗ ϕ)(x) = (u, τx Rϕ).
This has the same basic properties as before: u ∗ ϕ ∈ C ∞ and (u ∗ ϕ)0 =
u0 ∗ ϕ = u ∗ ϕ0 . Another important observation is that |(u ∗ ϕ)(x)| ≤
C(1 + |x|)N for some C, N , so u ∗ ϕ defines a tempered distribution
itself. To prove this estimate, observe that
(4.2) 1 + |t| = 1 + |t − x + x| ≤ (1 + |t − x|)(1 + |x|)
and thus
|(u ∗ ϕ)(x)| ≤ Ckϕ(x − t)kj,N ≤ C(1 + |x|)j kϕkj,N ,
since the first norm (which must be taken of ϕ(x − t) as a function
of t, for fixed x) involves suprema over t of expressions of the form
(1 + |t|)j |ϕ(n) (x − t)|, with 0 ≤ n ≤ N , and we then use (4.2).
Proposition 4.10. Let ϕ, ψ ∈ S(R), a ∈ R. Then: (a) (ϕ)b b = Rϕ;
0 0
(b) ϕ (t) = 2πitϕ(t);
d b (c) (−2πixϕ(x)) b (t) = ϕb (t); (d) (τa ϕ) b (t) =
−2πiat 2πiax
e ϕ(t);
b (e) (e b (f ) (ϕ ∗ ψ)b= ϕ
ϕ)b= τa ϕ; bψb
Part (a) is Fourier inversion, and the other parts follow from quick
calculations.
Theorem 4.11. Let u ∈ S 0 , ϕ ∈ S. Then u also satisfies (a)-(e), and
(u ∗ ϕ)b= ϕb
bu.
We have not yet formally introduced all the operations on tempered
distributions that are involved here, but it is of course clear how to pro-
ceed. We define (Ru, ϕ) = (u, Rϕ) and, as before, (τa u, ϕ) = (u, τ−a ϕ).
Parts (a)-(e) prove themselves if we just move the operations over to
the test function and then refer to Proposition 4.10. Let me do part
24 CHRISTIAN REMLING

(d) as an illustration:
((τa u)b, ϕ) = (τa u, ϕ)
b = (u, τ−a ϕ)
b
This test function equals (e−2πiax ϕ)b, by Proposition 4.1(e), and un-
packing this, we then obtain
(u, (e−2πiax ϕ)b) = (b
u, e−2πiax ϕ) = (e−2πiax u
b, ϕ).
Since this is true for any ϕ ∈ S, we have (τa u)b= e−2πiax u
b, as claimed.
Only part (f) presents some technical challenges, similar actually to
the ones from the proof of Theorem 3.4. Let ψ ∈ S be an arbitrary
test function. Then
Z
((u ∗ ϕ)b, ψ) = (u ∗ ϕ, ψ) = ψ(x)(u,
b b ϕ(x − t)) dx.
R
I will again rewrite this as (u, ψ(x)ϕ(x−t)
b dx) without fully justifying
R
this step. We can then interpret ψ(x)ϕ(x − t) dx = (ψb ∗ Rϕ)(t), and
b
Fourier inversion says that the inverse operation of b is R b. Thus
ψb ∗ Rϕ = fb, with

f = R(ψb ∗ Rϕ)b= R(ψb · (Rϕ)b) = ψR(Rϕ)b.

b

Putting things together, we deduce that

((u ∗ ϕ)b, ψ) = (b
u, ψR(Rϕ)b) = (R(Rϕ)bu, ψ),
and since R(Rϕ)b= ϕ, b by a calculation, this finally gives (f).
Let’s run a quick check on this result. We already know that δ ∗ ϕ =
ϕ. Taking Fourier transforms, we obtain ϕ bδb = ϕ,b and since δb = 1,
everything is indeed in perfect order.
Theorem 4.12. Suppose that un → u in S 0 . Then also u b in S 0 .
cn → u
Exercise 4.8. Prove Theorem 4.12.
Recall now Theorem 3.3. What do we obtain from the identity
1 1
(4.3) lim = PV + iπδ
h→0+ x − ih x
after taking Fourier transforms?
Exercise 4.9. Show that all distributions are tempered and that (4.3)
also holds in S 0 .
The distribution 1/(x − ih) is a function for fixed h > 0, so it would
be nice if we could compute its Fourier transform from (4.1), but this
fails because 1/(x − ih) ∈/ L1 . We can actually make good use of
 DISTRIBUTIONS 25

Theorem 4.12 right away and get around this difficulty by observing
that
1 1
lim χ(−L,L) (x) =
L→∞ x − ih x − ih
0
in S ; this is a routine application of DC, but do it in more detail
please if it’s not immediately clear to you. Now this cut off function is
integrable, so its Fourier transform is given by
Z L −2πitx
e
(4.4) dx.
−L x − ih

This integral can be evaluated in the limit L → ∞, most conveniently

by a typical application of the residue calculus. Let’s say t < 0. We
then close the contour [−L, L] by a semicircle in the upper half plane,
which can be parametrized as γ(s) = Leis , 0 ≤ s ≤ π. This part of the
integral can thus be estimated by
Z π
L
(4.5) e2πtL sin s ds,
L−h 0
and this goes to zero as L → ∞, by DC (recall that t < 0 currently).
The only singularity of the integrand from (4.4) occurs at x = ih, thus
Z L −2πitx
e
lim dx = 2πie2πht .
L→∞ −L x − ih

A similar calculation is possible for t > 0, and in this case we find that
Z L −2πitx
e
lim dx = 0
L→∞ −L x − ih

(because we now close the contour in the lower half plane, to make the
exponential small, and there will be no singularities inside our region
this time).
So, summing up, we have shown that
Z L −2πitx (
e 2πie2πht t < 0
lim dx = .
L→∞ −L x − ih 0 t>0
On reflection, we have actually shown this as a pointwise limit, but
this is not quite what we need here because what we know is that the
cut off integrals will converge to the Fourier transform of 1/(x − ih)
in S 0 . However, it is easy to go over the argument one more time and
confirm that this follows, too. We have shown that
 b
1
(t) = 2πiχ(−∞,0) (t)e2πht .
x − ih
26 CHRISTIAN REMLING

When h → 0+, this converges (in S 0 !) to

2πiχ(−∞,0) (t) = iπ(1 − sgn(t)).
When we compare this with (4.3), we obtain the following interesting
result.
Theorem 4.13. (PV(1/x))b = −iπ sgn(t)
The operation of convolving with PV(1/x) is called the Hilbert trans-
form of a function, so
f (x − t)
Z Z
f (t)
(Hf )(x) = lim dt = lim dt,
h→0+ |t|>h t h→0+ |t−x|>h x − t

and in our current context, this is defined for f ∈ S. Theorem 4.13

shows that (Hf )b(t) = −iπsgn(t)fb(t).
By Theorem 4.11(c),
1 1
(xPV(1/x))b= (PV(1/x))b0 = sgn0 = δ.
−2πi 2
This can be confirmed directly since
Z Z ∞
xϕ(x)
(xPV(1/x), ϕ) = lim dx = ϕ(x) dx,
h→0+ |x|>h x −∞

so xPV(1/x) = 1, and indeed b

1 = δ, as we saw earlier in Theorem 4.4.
Theorem 4.14. Let u ∈ D0 (R) and suppose that supp u = {a}. Then
(j)
u= N
P
j=0 cj δa .

Proof. For ease of notation, let’s assume that a = 0. We showed earlier,

in the proof of Theorem 4.7, that a compactly supported distribution
satisfies
|u(ϕ)| ≤ CkϕkN
(the point is that a fixed N and C work for all ϕ ∈ D).
I now claim that if ϕ ∈ D, ϕ(0) = ϕ0 (0) = . . . = ϕ(N ) (0) = 0, then
u(ϕ) = 0. This we can prove as follows. Given such a ϕ and  > 0, it
will of course be true for all sufficiently small b > 0 that |ϕ(N ) (x)| < 
on |x| < b. By successively integrating these derivatives, we then find
that also
(4.6) ϕ(j) (x) < bN −j for j = 0, 1, . . . , N.
Then we pick a ψ ∈ C0∞ with 0 ≤ ψ ≤ 1, suppψ ⊆ [−1, 1] and ψ = 1 on
a neighborhood of x = 0, and we consider ϕn (x) = ψ(nx)ϕ(x). These
functions will have their supports inside [−1/n, 1/n]. We now take
 DISTRIBUTIONS 27

b = 1/n and assume that n is large so that (4.6) becomes available.

From the j = 0 case, we conclude that |ϕn | ≤ bN = n−N . Next,
ϕ0n (x) = nψ 0 (nx)ϕ(x) + ψ(nx)ϕ0 (x),
and this gives us a bound of the type
|ϕ0n | ≤ C(nbN + bN −1 ) = (C + 1)n1−N .
We can continue in this style and bound the first N derivatives by
(4.7) ϕ(j)
n ≤ Cj nj−N , j = 0, 1, . . . , N.
Here, it is important that the constants Cj only depend on ψ (more
precisely, on bounds on ψ and its derivatives) and not on . Stated more
succinctly, (4.7) says that kϕn kN ≤ C. Since ϕn = ϕ near x = 0, we
have u(ϕ) = u(ϕn ) and, as we showed, |u(ϕn )| ≤ D. Since  > 0 was
arbitrary, we conclude that u(ϕ) = 0, as claimed.
Exercise 4.10. Explain in more detail why u(ϕ) = u(ϕn ). This is per-
haps best done by establishing the following general fact: If u ∈ D0
has compact support and ϕ = ψ on an open set V ⊇ supp u, then
u(ϕ) = u(ψ).
(n)
Now pick ϕj ∈ C0∞ with ϕj (0) = δjn for 0 ≤ j, n ≤ N . For example,
we could take ϕj (x) = xj /j! near x = 0. For arbitrary ϕ ∈ D, write
N
X n
X N
X
ϕ= aj ϕj + ϕ − aj ϕj ≡ aj ϕj + ψ,
j=0 j=0 j=0

with aj = ϕ(j) (0) (this is similar to a Taylor expansion of ϕ, except

that the compactly supported functions ϕj take over the role of the
polynomials xj /j!). Then ψ(0) = ψ 0 (0) = . . . = ψ (N ) (0) = 0, thus, by
what we just proved, u(ψ) = 0 and hence
N
X
u(ϕ) = u(ϕj )aj .
j=0

Since aj = ϕ(j) (0) = (−1)j (δ (j) , ϕ) this is what we claimed, with cj =

(−1)j (u, ϕj ). 
Theorem 4.15 (Poisson summation formula). u = n∈Z δn ∈ S 0 and
P
u
b = u.
The name Poisson summation formula is usually not given to this
rather highbrow version, but instead to
P P
Corollary 4.16. Let ϕ ∈ S(R). Then n∈Z ϕ(n) = n∈Z ϕ(n).
b
28 CHRISTIAN REMLING

This or also more general versions of the statement can be and usu-
ally are proved directly.
δn ∈ S 0 .
P
Exercise 4.11. Prove that
P
Proof of Theorem 4.15. Clearly, u = δn has the following two invari-
ance properties: τn u = u and e2πinx u = u, for all n ∈ Z. By taking
Fourier transforms, this implies that v = u b has the same two properties.
In particular, (e2πix − 1)v = 0, and this implies that supp v ⊆ Z, by
arguing as follows: Let V ⊆ R be an open set with V ∩ Z = ∅, and let
ϕ ∈ D be arbitrary with supp ϕ ⊆ V . Since supp ϕ is a compact subset
of V , it has positive distance to V c , and thus ψ = ϕ/(e2πix − 1) ∈ D
also. Thus
(v, ϕ) = (v, (e2πix − 1)ψ) = ((e2πix − 1)v, ψ) = 0.
We have shown that V ∩ supp v = ∅, and this holds for any open V
with V ∩ Z = ∅, thus supp v ⊆ Z, as claimed.
Consider now the restriction v0 of v to (−3/4, 3/4), say. Actually,
we want an element of D0 (R), not of D0 (−3/4, 3/4), so we really take
v0 = ψv with a v ∈ C0∞ with supp ψ ⊆ (−1, 1), ψ = 1 on (−3/4, 3/4).
(The purpose is to make Theorem 4.14 applicable, but this is a technical
point and not essential; the argument would also work with the actual
restriction of v, if Theorem 4.14 is slightly adapted instead.)
We have supp v0 ⊆ {0}, so
N
X
(4.8) v0 = cj δ (j) ,
j=0

by Theorem 4.14. We can say more here: in fact, v0 = cδ. To see this,
assume that N ≥ 1 and consider a trigonometric polynomial p(x) =
2πinx
P
|n|≤L an e , and here we want to choose L ≥ 1 and then the an ∈ C
such that p(0) = 1, p(j) (0) = 0 for 1 ≤ j ≤ N − 1, and p(N ) (0) 6= 0.
Exercise 4.12. Show that this is possible.
Finally, we build a test function by multiplying p by a ϕ ∈ C0∞ .
Then
(ϕp)(j) (0) = ϕ(j) (0) (j < N ), (ϕp)(N ) (0) = ϕ(N ) (0) + p(N ) (0)ϕ(0),
and thus, if v0 is as in (4.8) and invariant under multiplication by e2πinx ,
as we know it is, then
(v0 , ϕ) = (pv0 , ϕ) = (v0 , pϕ)
N
X
= cj (−1)j ϕ(j) (0) + cN (−1)N p(N ) (0)ϕ(0).
j=0
 DISTRIBUTIONS 29

On the other hand, if we simply

PN apply vj0 to ϕ directly, then we see that
(j)
this must also be equal to j=0 cj (−1) ϕ (0). It follows that cN = 0.
The whole argument can then be repeated to see that also cN −1 = 0
and so on.
So, reconstructing the distribution Pfrom its restrictions, based on
Theorem 2.13, we conclude that v = c n δn . P
Since τn v = v, the cn
must actually be independent of n, so v = c δn = cu. There are
2
ϕ ∈ S with ϕ b = ϕ, for example ϕ(x) = e−πx , and applying both u and
v=u b to such a ϕ shows that c = 1. 
Theorem 4.17. Let 0 < p < d. Then u = |x|−p ∈ S 0 (Rd ) and u
b=
cp,d |t|p−d .
Formally, this looks easy, as
Z Z ∞ Z
−p −2πit·x
|x| e dx = dr r d−1−p
dω(v) e−2πirv·t ,
Rd 0 v∈S d−1
if the integration is done in spherical coordinates, with dω denoting
the surface measure on the unit sphere. By the spherical symmetry of
this measure, the second integral, which we could denote by ω b , does
not depend on the direction of t, so ω b = ω b (r|t|) and thus, after the
substitution r|t| = s,
Z Z ∞
−p −2πit·x p−d
|x| e dx = |t| b (s)sd−1−p ds,
ω
Rd 0
as claimed (with cp,d apparently given by the integral).
Of course, none of this makes rigorous sense in this form since |x|−p ∈ /
L1 (Rd ), so the very first integral was undefined, and we cannot use (4.1)
here. Also, the last integral is not guaranteed to converge. The Fourier
transform of the surface measure has asymptotics |b ω (s)| ' s(1−d)/2 , so
our answer only makes sense if p > (d + 1)/2. (The decay of ω b in
dimension d > 1, by the way, is an initially rather surprising, but well
studied phenomenon.)
Still, the effort was not completely wasted because one useful general
conclusion can be drawn: the computation of u b seems to depend mainly
on the symmetry of u (more precisely, its spherical symmetry and ho-
mogeneity), and indeed this is exactly how the rigorous argument will
proceed.
Sketch of proof. The function u(x) = |x|−p is locally integrable and
bounded near infinity, so u ∈ S 0 .
In Exercise 2.5, you defined the dilation δa u of a distribution u as
(δa u, ϕ) = a−d (u, δ1/a ϕ), and (δb ϕ)(x) = ϕ(bx). This is done in such
a way that if u is a function, then δa u = u(ax). We now call u ∈ D0
30 CHRISTIAN REMLING

homogeneous of degree r if δa u = ar u for all a > 0. So if u is a function,

then this means that u(ax) = ar u(x), which is the familiar definition
of homogeneity for functions.
In now claim that ub ∈ S 0 is homogeneous of degree p−d. We confirm
this by a calculation, which will actually establish the general fact that
if v ∈ S 0 is homogeneous of degree r, then vb will be homogeneous of
degree −r − d.
We have
b, ϕ) = a−d (b
(δa u u, δ1/a ϕ) = a−d (u, (δ1/a ϕ)b).
It is easy to see, along the lines of what we did in Proposition 4.1, that
a−d (δ1/a ϕ)b= δa ϕ.
b
Exercise 4.13. Prove this in detail.
Thus
b, ϕ) = a−d (δ1/a u, ϕ)
(δa u b = ap−d (u, ϕ)
b = ap−d (b
u, ϕ)
Here, the second equality follows because the function u(x) = |x|−p is
clearly homogeneous of degree −p.
In the remainder of the argument, I will discuss explicitly only the
case d = 1. As our next step, we then show that u b is also invariant
under reflection in the sense that Rb
u=ub, with Rv defined as (Rv, ϕ) =
(v, Rϕ) and, as above, (Rϕ)(x) = ϕ(−x).
Exercise 4.14. Prove that Rb
u=u
b, by a calculation similar to the one
just given.
Now if we already knew that u b = f is a function f ∈ L1loc , then
everything seems clear: homogeneity and reflection invariance say that
f (x) = |x|p−1 f (1), which is our claim, with cp,1 = f (1).
Exercise 4.15. Actually, things are not quite as easy because a distribu-
tion that is a function determines its function only almost everywhere,
so homogeneity for example only says that f (ax) = f (x) off a null set,
which could depend on a > 0. Provide the missing details.
We are in this situation for 1/2 < p < 1 because we can then write
|x|−p = χ{|x|<1} |x|−p + χ{|x|>1} |x|−p as the sum of two functions, one in
L1 and one in L2 . The Fourier transform of an L1 is a function (in C0 ),
and so is the Fourier transform of an L2 function (this is in L2 also).
If, on the other hand, 0 < p < 1/2, then 1 − p is in the range just
discussed, so (|x|1−p )b= c|x|−p and then Fourier inversion or rather the
fact that a tempered distribution is uniquely determined by its Fourier
transform gives the claim in this case also.
 DISTRIBUTIONS 31

This leaves only the case p = 1/2, which we can handle by a limiting
argument, sending p → 1/2, p 6= 1/2.
Finally, this argument also works in higher dimensions if we establish
one additional symmetry before entering the final part of the argument:
u
b is spherically symmetric. 
Exercise 4.16. Give a precise definition of what is meant by this final
claim, and then prove it.
With additional trickery, the constants cp,d can actually be identified.
One finds
Γ((d − p)/2)
(4.9) cp,d = π −d/2+p .
Γ(p/2)
The differential operator
∂2 ∂2
∆= + . . . +
∂x21 ∂x2d
is called the Laplacian. The corresponding partial differential equa-
tion ∆u = 0 is called the Laplace equation, and its solutions are also
called harmonic functions or distributions, if we admit distributional
solutions.
Theorem 4.18. Let u ∈ S 0 , ∆u = 0. Then u is a (harmonic) polyno-
mial.
The assumption that the distribution is tempered is crucial here.
The theorem does not hold for general distributions, and in fact there
are many counterexamples that are smooth functions such as u(x, y) =
ex cos y.
Exercise 4.17. Find all u ∈ D0 (R) with ∆u = 0.
Corollary 4.19 (Liouville). A bounded harmonic function is constant.
Proof. Such a function u can be viewed as a tempered distribution, so
is a polynomial by Theorem 4.18, but the only bounded polynomials
are the constants. 
In the same way we can establish the more general statement that a
harmonic function u of at most polynomial growth, |u(x)| . (1 + |x|)N ,
actually is a polynomial of degree at most N .
Proof of Theorem 4.18 (sketch). Take Fourier transforms in ∆u = 0
and use Theorem 4.11 to conclude that
(4.10) −4π 2 |t|2 u
b = 0.
32 CHRISTIAN REMLING

Note that |t|2 = t21 + . . . + t2d is a smooth function satisfying the as-
sumptions of Theorem 4.8, so |t|2 u b ∈ S 0 is well defined.
Since 1/|t|2 is smooth away from t = 0, (4.10) implies that supp u ⊆
{0}. We now need the multi-dimensional analog of Theorem 4.14.
Assuming this, we obtain
X ∂α
u
b= cα δ,
∂tα
|α|≤N

and
P then we|α| take (inverse) Fourier transforms to deduce that u =
cα (−2πi) xα . 

A Green function is a G ∈ S 0 with ∆G = δ. Usually, one imposes

additional conditions on the asymptotics to make G unique and can
then meaningfully speak of the Green function (assuming for now its
existence, which we will establish in a moment). The G I will construct
below is the usual choice. Without such extra requirements, G is clearly
not unique since we can add an arbitrary harmonic u ∈ S 0 to a given
G.
Such a G is an interesting object. For example, it allows us to
construct fairly explicit solutions to the Poisson equation ∆u = f , at
least for f ∈ S, as u = G ∗ f . Indeed, ∆(G ∗ f ) = (∆G) ∗ f = δ ∗ f = f ,
as desired.
Theorem 4.20. For d ≥ 3, we can take G(x) = −cd |x|2−d .
In particular, G indeed is a function (not just a tempered distribu-
tion). The constants can be identified as
Γ(−1 + d/2)
cd = .
4π d/2
In particular, since Γ(1/2) = π 1/2 , we have c3 = 1/(4π).

Proof. To motivate what the theorem says is true, we can again take
Fourier transforms in ∆G = δ as in the proof of Theorem 4.18, to see
that (in S 0 ) −4π 2 |t|2 G b = −1/(4π 2 |t|2 ).
b = 1. This suggests to try G(t)
Theorem 4.17 then shows that indeed G(x) = −cd |x|2−d , with cd =
1/(4π 2 cd−2,d ). 
Exercise 4.18. The way I wrote it up, the argument doesn’t quite follow
the required logic. We should really define G(x) as in the theorem and
then check that ∆G = δ. Clean up the argument by reorganizing it
along these lines, and also derive the formula for cd from (4.9).
 DISTRIBUTIONS 33

Theorem 4.20 also works for d = 1 and is in fact much easier to

obtain in this case. Since c1 = −1/2, it then says that G(x) = |x|/2,
and we verified long ago that indeed ∆G = G00 = δ, as required.
If we argue as in the proof of Theorem 4.20 for d = 1, then we learn
c satisfies −2π 2 t2 u = 1 and thus
that u = |x|
−1
u = .
R\{0} 2π 2 t2
This is a smooth function away from t = 0 and presents no problems
there, but clearly it cannot be the case that u itself is this function
because 1/t2 is much too singular near t = 0 and fails to be integrable
there. So what is u? To answer this, recall from Theorem 4.13 that
1
iπ sgnb = PV ;
x
to write the result in this form, I have also used that b b = R. So, by
Theorem 4.11,
 0
2 1
2π (t sgn(t))b= PV .
x
Now t sgn(t) is simply |t|, so we have already obtained the interesting
result that  0
1 1
|t| = 2 PV
b .
2π x
Finally, let’s look at this derivative more closely. We have
0 
ϕ0 (x)
 Z
1
PV , ϕ = − lim dx.
x h→0+ |x|>h x
R −h R∞
We integrate by parts in the two integrals −∞ . . . and h . . . and then
put things back together to obtain
 0   Z 
1 ϕ(−h) + ϕ(h) ϕ(x)
PV , ϕ = lim − 2
dx .
x h→0+ h |x|>h x

Since
ϕ(−h) + ϕ(h) − 2ϕ(0)
→0
R h
and |x|>h dx/x2 = 2/h, we can rewrite this as
 0 
ϕ(0) − ϕ(x)
Z
1
PV , ϕ = lim dx.
x h→0+ |x|>h x2
We summarize:
34 CHRISTIAN REMLING

Theorem 4.21.
ϕ(x) − ϕ(0)
Z
(u, ϕ) = lim dx
h→0+ |x|>h x2
defines a (tempered) distribution. We have u = −2π 2 |t|
b and also u =
0
−(PV(1/x)) .