Tips and Tricks in Real Analysis

Nate Eldredge
August 3, 2008
This is a list of tricks and standard approaches that are often helpful when
solving qual-type problems in real analysis.
Approximate. There are a lot of results that say that a function f can be approx-
imated by a sequence of nicer functions f
n
, so that f
n
f in some ap-
propriate sense. So if you have some property thats continuous in that its
preserved by taking the right kind of limit, you can show that an arbitrary
function has this property by showing that the nicer ones do. Here are
some examples.
Measurable functions can be approximated pointwise by simple func-
tions. If the limiting function is nonnegative (consider f
+
and f

), the
simple functions can be chosen monotonically increasing, so the mono-
tone convergence theorem can be used.
Continuous functions on a compact interval can be approximated uni-
formly by step functions (linear combinations of characteristic func-
tions of intervals), piecewise linear functions, or smooth functions.
Also: polynomials, trigonometric polynomials, or anything else that
satises the hypotheses of the Stone-Weierstrass theorem. You can
also use rationals for most of these things, which is useful for show-
ing something is separable. (E.g. if polynomials work, then so do
polynomials with rational coecients.)
Convolution is a useful tool for approximating functions by smooth
(C

) functions. Let be your favorite smooth function which inte-

grates to 1. (I imagine something Gaussian-ish.) Let
n
(x) = n(nx),
so that
n
gets a sharper and sharper bump at 0 while still integrating to
1. As a general principle,
n
f is a smooth function, and
n
f f
in an appropriate sense. Specically:
If f is uniformly continuous then
n
f f uniformly. In par-
ticular this works if f is continuous with compact support.
1
If f C
k
c
then
n
f f in C
k
c
, i.e. the rst k derivatives of

n
f converge uniformly to the corresponding derivatives of f .
If f L
p
then
n
f f in the L
p
norm.
If f is a distribution then
n
f f in the usual sense for distri-
butions.
In a function space (or indeed any space), look for dense subsets. In
a Hilbert space, orthonormal bases are often useful, just as in R
n
; any
element is a limit of (nite!) linear combinations of the basis elements
e
n
.
If you prefer, you can rephrase any of these statements as something about a
subset being dense. E.g. piecewise linear functions are dense in C([0, 1]),
smooth functions are dense in L
p
.
Deal with measurable sets in terms of generators. If a -algebra Jis generated
by a collection of sets c, and you are trying to prove something about all
measurable sets in J, it is good to start by proving it holds for sets in c.
But you should not then try to write an arbitrary measurable set in terms
of sets from c by taking unions and complements; it cant be done in any
simple way.
1
Instead, you can use the denition: if J = (c), then J is
the smallest -algebra containing c, and so any other -algebra containing
c will contain J as well. Thus, consider the collection of all sets that have
the property you want. If you can show that it contains c and is (or contains)
a -algebra, youll be done.
Pass to subsequences. If a sequence x
n
converges to x, then so do all of its subse-
quences x
n
k
. Conversely, if every subsequence x
n
k
has a further subsequence
x
n
k
j
which converges to x, then x
n
itself also converges to x. (Intuitively,
if x
n
did not converge to x, you could nd a subsequence that converged to
something else y, and all of its subsequences would converge to y as well.)
This fact is commonly useful in connection with the theorem (2.30 in Fol-
land) that if f
n
f in measure, there is a subsequence f
n
k
such that f
n
k
f
a.e.; you can often take a theorem that requires a.e. convergence and weaken
the hypothesis to convergence in measure (thus strengthening the theorem).
Example 0.1. Show that the dominated convergence theorem holds for con-
vergence in measure; i.e. if f
n
f in measure and f
n
g L
1
for all n,
then f L
1
and

f
n

f .
1
Technically, there is a way to do this, but you might have to take unions of complements of
unions of complements, etc, and repeat the process up to an uncountable number of times. I dont
recommend it unless you are very used to dealing with innite ordinals and transnite induction.
2
Proof. Let x
n
=

f
n
, and let x
n
k
be an arbitrary subsequence. The subse-
quence f
n
k
also converges to f in measure (check this if you like). By the
theorem above it has a subsequence of its own, f
n
k
j
which converges to f
a.e. By dominated convergence f L
1
and

f
n
k
j

f . So every subse-
quence x
n
k
of x
n
has a further subsequence x
n
k
j
which converges to

f , and
therefore x
n

f as desired.
Use the triangle inequality, in all its forms, generously. Any time you see an ab-
solute value or norm on the outside of a sum, innite sum, or integral, con-
sider crashing through with it. More often than not its the right thing to
do, and if its not, you can usually tell right away. Dierences usually arent
as good an idea, since you are often subtracting two things that are close to-
gether, but consider it anyway. Another good trick for estimating dierences
is to add and subtract the same thing and then use the triangle inequality.
Example 0.2. Let X, Y be normed spaces, x
n
x a convergent sequence
in X, and T
n
T a norm-convergent sequence of bounded operators in
L(X, Y). Show that T
n
x
n
T x in Y.
Proof. Well, wed better show that jT
n
x
n
T xj
Y
0. We have two things
changing at once, so lets try to get them one at a time by adding and sub-
tracting T x
n
.
jT
n
x
n
T xj
Y
= jT
n
x
n
T x
n
+ T x
n
T xj
Y
jT
n
x
n
T x
n
j
Y
+ jT x
n
T xj
Y
jT
n
Tj
L(X,Y)
jx
n
j
X
+ jTj
L(X,Y)
jx
n
xj
X
Now the second term is going to zero because we know x
n
x. For the
rst term, note that since x
n
is a convergent sequence, it must be bounded
(a very useful fact), so there is an M such that jx
n
j
X
M for all M. Since
jT
n
Tj
L(X,Y)
0 by assumption, the rst term is also going to zero, so we
are done.
Watch for the Big Three Baire theorems for Banach spaces. These are the Prin-
ciple of Uniform Boundedness (PUB), the Open Mapping Theorem, and the
Closed Graph Theorem. See section 5.3 of Folland. It can be easy to over-
look an opportunity to use one of these theorems, which is unfortunate be-
cause they can be very powerful. I think its because they are so very false
in general normed spaces, and so they dont come so naturally to mind when
thinking in normed-space terms. You may need to explicitly remind yourself
when you have Banach spaces that these theorems are in play.
3
Shoot rst and ask questions later. Real analysis has a lot of theorems of the
form: Under [complicated set of conditions], [simple conclusion] holds.
Perhaps the best examples are the integral convergence theorems and the
Fubini/Tonelli theorems, which tell you when you can interchange various
combinations of limits, integrals, and sums. When you want to make such an
interchange, its helpful to rst try it and see if it gets you what you want.
2
If it does, then you can check the hypothesis of the theorem that justies it.
Work informally with distributions. Although the technical denitions for dis-
tributions are a bit complicated, we should try to remember that they are
supposed to be generalized functions: they are designed to behave a lot
like functions. So in the spirit of the previous item, it can be helpful in early
stages of a problem to pretend that they are functions. Add them, dieren-
tiate them, integrate by parts (a biggie), convolve, take Fourier transforms.
This will help you decide what the answer actually is, and you can then go
back and verify it at a more technical level.
Example 0.3. Compute the Fourier transforms of the Dirac delta distribution

0
and its derivative

0
.
Proof. Informally, the Fourier transform of
0
is supposed to be
/[
0
]() =

e
ix

0
(x) dx. (1)

0
is supposed to be a point mass at 0, so this should be the value of e
ix
when x = 0; we expect /[
0
]() = 1.
For

0
, we integrate by parts:
/[
0
]() =

e
ix

0
(x) dx
=

d
dx
e
ix

0
(x) dx
=

ie
ix

0
(x) dx = i.
Nowthat we knowwhat the answers are, well check themmore carefully.
0
is denitely a tempered distribution, so by denition /[
0
] is the distribution
2
It almost always does. When I took 240 from Professor Bruce Driver, he propounded what
we students came to call Drivers Theorem: Whenever you see two integrals, you should try to
change their order.
4
such that (/[
0
], ) = (
0
, /[]) for all S, the Schwartz class. But
(
0
, /[]) = /[](0) =

e
i0x
(x) dx =

(x) dx = (1, ). (2)

Next, (/[

0
], ) = (

0
, /[]) = (
0
, /[]

) = /[]

(0). But

d
d

=0

e
ix
(x) dx =

d
d
e
ix
(x) dx
=0
=

(ix)e
ix
(x) dx
=0
=

ix(x) dx = ( f , ) where f (x) = ix.

The interchange of integral and derivative in the rst line is an example of
dierentiating under the integral sign and should be justied; see Folland
2.27. The theorem looks scary; its perhaps easier just to remember the
proof: its just the denition of the derivative, the DCT and the mean value
theorem.
Well, perhaps that wasnt a great example; knowing the answer in advance
didnt really help. But it did make for a good check. Maybe there is a better
example.
5