Princeton Companion to Mathematics Proof 1

Distributions smooth ones. Then, given an operation that is nat-

urally defined for smooth functions, there is a good
By Terence Tao chance that there will be exactly one natural way
to extend it to an operation on rough functions:
A function is normally defined to be an object one takes a sequence of better and better smooth
f : X → Y which assigns to each point x in a set approximations to the rough functions, performs
X, known as the domain, a point f (x) in another the operation on them, and passes to the limit.
set Y , known as the range (see Section ?? in The Distributions, or generalized functions, belong at
Language and Grammar of Mathematics). the rough end of the spectrum, but before we say
Thus, the definition of functions is set-theoretic what they are, it will be helpful to begin by consid-
and the fundamental operation that one can per- ering some smoother classes of functions, partly for
form on a function is evaluation: given an element comparison and partly because one obtains rough
x of X, one evaluates f at x to obtain the element classes of functions from smooth ones by a process
f (x) of Y . known as duality: a linear functional defined on a
However, there are some fields of mathematics space E of functions is simply a linear map φ from
where this may not be the best way of describing E to the scalars R or C. Typically, E is a normed
functions. In geometry, for instance, the fundamen- space, or at least comes with a topology, and the
tal property of a function is not necessarily how it dual space is the space of continuous linear func-
acts on points, but rather how it pushes forward or tionals.
pulls back objects that are more complicated than
points (e.g. other functions, bundles and sections, The class C ω [−1, 1] of analytic functions.
sheaves and schemes, etc.). Similarly, in analy- These are in many ways the “nicest” functions of
sis, a function need not necessarily be defined by all, and include many familiar functions such as
what it does to points, but may instead be defined exp(x), sin(x), polynomials, and so on. However,
by what it does to objects of different kinds, such we shall not discuss them further, because for many
as sets or other functions. The former concept leads purposes they form too rigid a class to be useful.
to the notion of a measure; the latter to that of a (For example, if an analytic function is zero every-
distribution. where on an interval, then it is forced to be zero
Of course, all these notions of function and everywhere.)
function-like objects are related. In analysis, it is The class Cc∞ [−1, 1] of test functions. These
helpful to think of the various notions of a function are the smooth (that is, infinitely differentiable)
as forming a spectrum, with very “smooth” classes functions f , defined on the interval [−1, 1], that
of functions at one end and very “rough” ones at vanish on neighbourhoods of 1 and −1. (That is,
the other. The smooth classes of functions are very one can find δ > 0 such that f (x) = 0 whenever
restrictive in their membership: this means that x > 1 − δ or x < −1 + δ.) They are more numer-
they have good properties, and there are many ous than analytic functions and therefore more
operations that one can perform on them (such tractable for analysis. For instance, it is often use-
as, for example, differentiation), but it also means ful to construct smooth “cutoff functions,” which
that one cannot necessarily ensure that the func- are functions that vanish outside some small set
tions one is working with belong to this category. but do not vanish inside it. Also, all the operations
Conversely, the rough classes of functions are very from calculus (differentiation, integration, compo-
general and inclusive: it is easy to ensure that one sition, convolution, evaluation, etc.) are available
is working with them, but the price one pays is that for these functions.
the number of operations one can perform on these
functions is often sharply reduced (see function The class C 0 [−1, 1] of continuous functions.
spaces). These functions are regular enough for the notion
Nevertheless, the various classes of functions can of evaluation, x → f (x), to make sense for every
often be treated in a unified manner, because it x ∈ [−1, 1], and one can integrate such functions
is often possible to approximate rough functions and perform algebraic operations such as multipli-
arbitrarily well (in an appropriate topology) by cation and composition, but they are not regular
2 Princeton Companion to Mathematics Proof

enough that operations such as differentiation can function g ∈ Cc∞ ([−1, 1]), producing a number
be performed on them. Still, they are usually con- µ, g. A famous example is the Dirac distribution
sidered among the smoother examples of functions δ0 , defined as the functional which when paired
in analysis. with any test function g returns the evaluation
g(0) of g at zero: δ0 , g := g(0). Similarly, we have
The class L2 [−1, 1] of square-integrable the derivative Dirac distribution −δ0 , which when
functions. These are measurable functions f : paired with any test function g returns the deriva-

1 1] →2 R for which the Lebesgue integral
[−1, tive g  (0) of g at zero: −δ0 , g := g  (0). (The rea-
−1
|f (x)| dx is finite. Usually one regards two son for the minus sign will be given later.) Since
such functions f and g as equal if the set of x such test functions have so many operations available
that f (x) = g(x) has measure zero. (Thus, from to them, there are many ways to define continu-
the set-theoretic point of view, the object in ques- ous linear functionals, so the class of distributions
tion is really an equivalence class of functions (see is quite large. Despite this, and despite the indi-
Section ?? in The Language and Grammar of rect, virtual nature of distributions, one can still
Mathematics).) Since a singleton {x} has mea- define many operations on them; we shall discuss
sure zero, we can change the value of f (x) without this later.
changing the function. Thus, the notion of evalua-
tion does not make sense for a square-integrable The class C ω ([−1, 1])∗ of hyperfunctions.
function f (x) at any specific point x. However, There are classes of functions more general still
two functions that differ on a set of measure zero than distributions. For instance, there are hyper-
have the same Lebesgue integral (see measures), functions, which roughly speaking one can think of
so integration does make sense. as linear functionals that can only be tested against
A key point about this class is that it is self-dual analytic functions g ∈ C ω ([−1, 1]) rather than test
in the following sense. Any two functions in this functions g ∈ C ∞ ([−1, 1]). However, as the class of
analytic functions is so sparse, hyperfunctions tend
class can
be paired together by the inner product
1 not to be as useful as distributions in analysis.
f, g := −1 f (x)g(x) dx. Therefore, given a func-
tion g ∈ L2 [−1, 1], the map f → f, g defines At first glance, the concept of a distribution has
a linear functional on L2 [−1, 1], which turns out limited utility, as all a distribution µ is empow-
to be continuous. Moreover, given any continuous ered to do is to be tested against test functions g
linear functional φ on L2 [−1, 1], there is a unique to produce inner products µ, g. However, using
function g ∈ L2 [−1, 1] such that φ(f ) = f, g for this inner product, one can often take operations
every f . This is a special case of one of the Riesz that are initially defined only on test functions, and
representation theorems. extend them to distributions by duality. A typical
example is differentiation. Suppose one wants to
The class C 0 [−1, 1]∗ of finite Borel mea- know how to define the derivative µ of a distri-
sures. Any finite Borel measure µ gives rise to bution, or in other words how to define µ , g for
0
a continuous linear

1 functional on C [−1, 1] defined any test function g and distribution µ. If µ is itself
by f → µ, f  = −1 f (x) dµ. Another of the Riesz a test function µ = f , then we can evaluate this
representation theorems says that every continuous using integration by parts (recalling that test func-
linear functional on C 0 [−1, 1] arises in this way, so tions vanish at −1 and 1). We have
one could in principle define a finite Borel measure  1
to be a continuous linear functional on C 0 [−1, 1]. f  , g = f  (x)g(x) dx
−1
The class C ∞ ([−1, 1])∗ of distributions.  1
Just as measures can be viewed as continuous lin- =− f (x)g  (x) dx = −f, g  .
ear functionals on C 0 ([−1, 1]), a distribution µ is a −1

continuous linear functional on Cc∞ ([−1, 1]) (with Note that if g is a test function, then so is g  . We
an appropriate topology). Thus, a distribution can can therefore generalize this formula to arbitrary
be viewed as a “virtual function”: it cannot itself distributions by defining µ , g := −µ, g  . This is
be directly evaluated, or even integrated over an the justification for the differentiation of the Dirac
open set, but it can still be paired with any test distribution: δ0 , g = −δ0 , g   = −g  (0).
Princeton Companion to Mathematics Proof 3

More formally, what we have done here is to “disappear” in the limit. One advantage of work-
compute the adjoint of the differentiation opera- ing with distributions instead of smoother func-
tion (as defined on the dense space of test func- tions is that one often has some compactness in
tions). Then we have taken adjoints again to define the space of distributions under weak limits (e.g.
the differentiation operation for general distribu- by the Banach–Alaoglu theorem). Thus distribu-
tions. This procedure is well-defined and also works tions can be thought of as asymptotic extremes of
for many other concepts, thus one can add two behavior of smoother functions, just as real num-
distributions, multiply a distribution by a smooth bers can be thought of as limits of rational num-
function, convolve two distributions, and compose bers.
distributions on both left and right with suitably
smooth functions. One can even take Fourier trans-
forms of distributions. For instance, the Fourier
transform of the Dirac delta δ0 is the constant
function 1, and vice versa (this is essentially the
Fourier
 inversion formula), while the distribution
n∈Z δ0 (x − n) is its own Fourier transform (this
is essentially the Poisson summation formula).
Thus the space of distributions is quite a good
space to work in, in that it contains a large class
of functions (e.g. all measures and integrable func-
tions), and is also closed under a large number of
common operations in analysis. Because the test
functions are dense in the space of distributions,
the operations as defined on distributions are usu-
ally compatible with those on test functions. For
instance, if f and g are test functions and f  = g in
the sense of distributions, then f  = g will also be
true in the classical sense. This often allows one to
manipulate distributions as if they were test func-
tions without fear of confusion or inaccuracy. The
main operations one has to be careful about are
evaluation and pointwise multiplication of distri-
butions, both of which are usually not well-defined
(e.g. the square of the Dirac delta distribution is
not well-defined as a distribution).
Another way to view distributions is as the weak
limit of test functions. A sequence of functions
fn is said to converge weakly to a distribution µ
if fn , g → µ, g for all test functions g. For
instance,

1 if ϕ is a test function with total inte-
gral −1 ϕ = 1, then the test functions fn (x) :=
nϕ(nx) can be shown to converge weakly to the
Dirac delta distribution δ0 , while the functions
fn = n2 ϕ (nx) converge weakly to the derivative
δ0 of the Dirac delta. On the other hand, the func-
tions gn (x) := cos(nx)ϕ(x) converge weakly to
zero (this is a variant of the Riemann–Lebesgue
lemma). Thus weak convergence has some unusual
features not present in stronger notions of conver-
gence, in that severe oscillations can sometimes