LECTURE NOTES 9 FOR 247B

TERENCE TAO

1. Fourier analysis on finite abelian groups

We have been using Fourier analysis on the Euclidean group Rd (and to a lesser
extent, on the toral group Td ) for some time now. It turns out that Fourier analysis
can in fact be formalised on any locally compact Hausdorff abelian group. We will
come to this generalisation later in this set of notes, but to keep things simple
for now we will work with the special setting of finite abelian groups, in which
qualitative issues such as integrability, measurability, continuity, etc. are irrelevant,
and one can do things like induct on dimension without having to invoke things
like Zorn’s lemma.

Our viewpoint shall be representation-theoretic. Other approaches to the subject

include the Gelfand-theoretic approach (viewing the convolution algebra as a B ∗ -
algebra), spectral-theoretic (focusing on trying to diagonalise various translation-
invariant operators), or algebraic (using explicit classifications of the group and
its characters). Each of these approaches is worthwhile (and they are of course
closely related to each other), but we select the representation-theoretic approach
as it extends relatively easily to the non-abelian case, whereas the other approaches
have more difficulty.

In contrast to previous notes, the material here will be largely algebraic in nature;
what analysis there is here is largely of a qualitative nature (e.g. dealing with issues
of continuity, integrability, etc.) rather than quantitative (estimates, etc.).

We now turn to the details. Fix a finite additive group (G, +), where we use the
addition symbol for the group operation to emphasise the abelian nature of the
group; in particular we use 0 for the group identity and −x for the group inverse
of x. It will not be terribly relevant for this finitary analysis, but we give G the
discrete topology and the discrete σ-algebra. Let #G denote the cardinality of G.
We define the normalised Haar measure dx on this group to be the normalised
counting measure
Z
1 X
f (x) dx := f (x).
G #G
x∈G

1
The normalising factor #G will almost never be seen again, being concealed within
the measure dx at all times.
1
2 TERENCE TAO

We have the (finite-dimensional) Hilbert space L2 (G) of functions f : G → C, with

inner product Z
hf, giL2 (G) := f (x)g(x) dx.
G
Note that as G is finite, all norms are equivalent, thus for instance L2 (G) = L1 (G) =
L∞ (G). Every group element y ∈ G gives rise to a translation operator Transy :
L2 (G) → L2 (G), defined by
Transy f (x) := f (x − y);
observe each such operator is a unitary operator on L2 (G), and furthermore the
map y 7→ Transy is a homomorphism:
Transy Transz = Transy+z . (1)
In other words, the map y 7→ Transy is a unitary representation of the group G
acting on the Hilbert space L2 (G); this particular representation is known as the
regular representation. One consequence of (1) is that all the translations commute
with each other.

Given two functions f, g ∈ L2 (G), we can form their convolution f ∗ g ∈ L2 (G) by

the formula Z
f ∗ g(x) := f (y)g(x − y) dy.
G
One easily verifies that convolution is bilinear, associative, and commutative (the
latter of course relies on the abelian nature of G). There is also an identity element
δ ∈ L2 (G), defined by δ(x) := (#G)1{0} , thus f ∗ δ = δ ∗ f = 1.

Given any g ∈ L2 (G), we can define the associated convolution operator Tg :

L2 (G) → L2 (G) by Tg f := g ∗ f . Equivalently, we have
Z
Tg = g(y)Transy dy,
G
thus convolution operators are nothing more than linear combinations of trans-
lation operators. The convolution operators commute with each other, and are
translation-invariant; conversely, it is not hard to show that every translation-
invariant operator is a convolution operator. (One can view the map g 7→ Tg as a
representation of the convolution algebra L2 (G) acting on itself.)

Define a translation-invariant subspace V of L2 (G) to be any subspace of V which

is invariant under all of the translations Transy , thus Transy V = V for all y; one
can also think of such a space as a component of the regular representation of
G. Equivalently, V is preserved by all of the operators Tg . Besides the trivial
examples of {0} and L2 (G), two other simple examples of such spaces
R are the space
{c : c ∈ G} of constant functions, and the space {f ∈ L2 (G) : G f = 0} of mean-
zero functions. Another important pair of examples: given any convolution operator
Tg , the kernel {f ∈ L2 (G) : Tg f = 0} and the range {Tg f : f ∈ L2 (G)} are also
translation-invariant (this is basically because convolution operators commute with
translations). The vector sum or intersection of two translation-invariant spaces is
translation-invariant; since all the translation operators are unitary, we also see that
the orthogonal complement of a translation-invariant space is translation-invariant.
 LECTURE NOTES 9 3

Example 1.1. In the case of the cyclic group G = Z/N Z, given any ξ ∈ Z/N Z
we can form the one-dimensional translation-invariant space Vξ generated by the
character eξ : x 7→ e2πixξ/N . If we secretly allow ourselves the use of the finite
Fourier transform, it is not hard to see that a space is translation-invariant iff it is
the direct sum of some of these Vξ , or equivalently if it is of the form {f ∈ L2 (G) :
fˆ|E = 0} for some fixed set of frequencies E ⊂ Z/N Z. Readers familiar with the
finite Fourier transform may find it instructive to keep this example in mind for
the rest of this section.

Since Tg δ = g for all g ∈ L2 (G), we see that any translation-invariant space

which contains the convolution identity δ, must be all of L2 (G). Equivalently,
a translation-invariant space is proper if and only if it avoids δ.

From these facts it is clear that (a) every proper translation-invariant subspace is
contained in a maximal translation-invariant subspace; and (b) every translation-
invariant subspace can be expressed (possibly non-uniquely) as the direct sum of
irreducible translation-invariant subspaces, i.e. non-zero subspaces which cannot be
nontrivially split as the sum of two smaller translation-invariant subspaces. It is not
hard to show that a translation-invariant subspace is irreducible iff its orthogonal
complement is maximal.
Remark 1.2. The orthogonal projection to a translation-invariant space is a translation-
invariant operator, and thus given by convolution with some function µ, in particu-
lar µ∗µ = µ. Such functions are known as idempotent measures and were influential
in the classical development of harmonic analysis. We have however structured our
presentation here so that these measures are not needed.

We now come to the first really non-trivial result about these spaces. There are two
equivalent forms of this result. The first is phrased in terms of maximal translation-
invariant subspaces:
Proposition 1.3 (Gelfand-Mazur theorem, special case). All maximal translation-
invariant subspaces are hyperplanes (i.e. their codimension is one).

Proof Suppose for contradiction that we have a maximal translation-invariant

subspace V of codimension two or greater. Observe that V is an ideal of the
Banach algebra L2 (G), and hence the quotient L2 (G)/V is also a Banach algebra,
with dimension at least two. In particular, it contains an element g which is not a
multiple of the identity. From Liouville’s theorem we know that (g − z)−1 cannot
exist for every complex z, thus there is a z ∈ C for which g−z is non-invertible. Thus
the kernel of g − z is a proper non-zero ideal in L2 (G)/V , which induces a proper
translation-invariant subspace of L2 (G) which strictly contains V , contradicting
maximality.

The second is phrased in terms of the orthogonal complement, irreducible translation-

invariant subspaces:
Proposition 1.4. All irreducible translation-invariant subspaces have dimension
one.
4 TERENCE TAO

Proof Let V be an irreducible translation-invariant subspace, and suppose for

contradiction that V had dimension at least two. Then V will contain a non-zero
function f which vanishes at at least one point. Then there exists y ∈ G such that
Transy f is not a constant multiple of f (simply choose y so that Transy shifts a
zero of f to a non-zero of f ). In particular, Transy is not a constant multiple of
the identity on V . On the other hand, it is a unitary transformation on V . Thus
by the spectral theorem, it has an eigenvalue, and hence a non-trivial eigenspace
W in V . Since all the translations commute with each other, we see that W is
also translation-invariant, as is the orthogonal complement of W in V . But this
contradicts the irreducibility of V .

Of course, the above two propositions are equivalent to each other; interestingly,
they both secretly use complex analysis in the proof (the latter via the spectral
theorem for unitary operators). In the special case of finite abelian groups, an-
other way to establish these propositions is to use the classification of finite abelian
groups and then explicitly construct a Fourier transform with which to analyse
these translation-invariant spaces. This is the most direct proof, but relies heavily
on the classification, which is not easily available for more general abelian groups.

Now let V be an irreducible translation-invariant subspace, thus one-dimensional.

This means that each of the translation operators Transy acts by multiplication by
a complex constant χV (y) on V ; since Transy is unitary, we must have |χV (y)| = 1.
Also, from (1) we see that χV is a homomorphism: χV (y + z) = χV (y)χV (z). In
other words, χV : G → S 1 is a multiplicative character of G; conversely, given any
multiplicative character χ : G → S 1 , the one-dimensional space generated by χ is
clearly an irreducible translation-invariant space. Thus the irreducible translation-
invariant spaces are in one-to-one correspondence with multiplicative characters.
Introducing the exponential function e : R/Z → S 1 by e(x) := e2πix , we can write
any multiplicative character χ as χ = e(ξ), where ξ : G → R/Z is an additive
character, i.e. an additive homomorphism from G to R/Z. Thus the irreducible
translation-invariant spaces are also in one-to-one correspondence with additive
characters.
Remark 1.5. As a corollary, one also concludes that every maximal translation-
invariant subspace is the orthogonal complement of a multiplicative character χ.

Define the Pontryagin dual Ĝ of G to be the space of all additive characters of G;

this is clearly an additive group. Given x ∈ G and ξ ∈ Ĝ, we shall also write ξ · x
for ξ(x) ∈ R/Z. We refer to elements ξ of Ĝ as frequencies. Given a frequency ξ,
the associated irreducible translation-invariant space Vξ is described as the linear
span of the multiplicative character eξ : x 7→ e(ξ · x).
Lemma 1.6 (Orthogonality). If ξ, η are two distinct frequencies, then the associ-
ated translation-invariant subspaces Vξ and Vη are orthogonal to each other.

Proof It suffices to show that eξ and eη are orthogonal, or in other words that the
expression Z
I := e(ξ · x)e(−η · x) dx
G
 LECTURE NOTES 9 5

is zero. Shifting x by y we see that

I = e(ξ · y)e(−η · y)I
for all y ∈ G. But since ξ, η are distinct, then there exists y such that ξ · y 6= η · y,
and so I = 0 as claimed.

Splitting the regular representation into irreducible (and now also orthogonal) com-
ponents, we conclude
L
Corollary 1.7 (Peter-Weyl theorem, finite abelian case). L2 (G) = ξ∈Ĝ Vξ . In
particular (by dimension count) #G = #Ĝ. Equivalently, the space {eξ : ξ ∈ Ĝ} of
multiplicative characters is an orthonormal basis for L2 (G).

Observe that if f ∈ L2 (G) and ξ ∈ Ĝ, the orthogonal projection of f to Vξ is given

R
by fˆ(ξ)eξ , where fˆ(ξ) = hf, eξ iL2 (G) = G f (x)e(−ξ · x) dx. We thus conclude

Corollary 1.8 (Fourier inversion formula). For any f ∈ L2 (G), we have f =

P ˆ
ξ∈Ĝ f (ξ)eξ .

Other quick consequences are the Plancherel identity

kf kL2 (G) = kfˆkl2 (Ĝ)
and more generally Parseval identity
hf, giL2 (G) = hfˆ, ĝil2 (Ĝ) .
It is also easy to check the convolution identity

f[
∗ g = fˆĝ
and dually that
fcg = fˆ ∗ ĝ
where the ∗ on the right now refers to discrete convolution (using counting measure
on Ĝ rather than normalised counting measure on G).

Observe that every x ∈ G can be viewed as a character x 7→ ξ · x on Ĝ, thus

ˆ
providing a canonical map from G to Ĝ. This map is injective. To see this, suppose
that x was in the kernel of this map, then ξ · x = 0 for all ξ ∈ Ĝ, or equivalently
Transx fixes each of the characters eξ . By the Fourier inversion formula this implies
that Transx fixes all functions, which is observe. From the Peter-Weyl theorem we
ˆ
also know that G and Ĝ have the same cardinality. Thus the map is bijective. In
other words, the Pontryagin dual of Ĝ is canonically identifiable with G itself.
Example 1.9. When G = Z/N Z, each ξ ∈ Z/N Z generates a character, defined by
ξ · x := ξx/N . These are N distinct characters, and so by the Peter-Weyl theorem
there are no further characters. Thus the abstract Fourier transform given here
corresponds to the usual finite Fourier transform on Z/N Z.
6 TERENCE TAO

We have the usual duality relationship between translation and modulation: for
every f ∈ L2 (G) and y ∈ G, ξ ∈ Ĝ we have
\y f = Mod−y fˆ;
Trans \
Mod ˆ
ξ f = Transξ f

where Mod−y F (ξ) := e(−ξ · y)F (ξ) and Modξ f (x) := e(ξ · x)f (x).

2. Locally compact abelian groups

Many (though not all) of the theory developed above continues to apply in the
more general category of locally compact abelian (LCA) groups - topological groups
which are both abelian and locally compact. (A topological group is a group with a
topology under which the group operations are continuous. To avoid minor techni-
calities, we also assume that the topology is Hausdorff.) We will omit some details,
and refer the reader to Rudin’s “ Fourier analysis on groups” for a full treatment.

Fix a LCA group G. The first task is to determine the analogue of the measure
dx used earlier. This is Haar measure - a non-negative translation-invariant Radon
measure which is not identically zero. The fact that such a measure exists at all is
non-trivial - akin to the construction of Lebesgue measure, which is Haar measure
for Rd . Roughly speaking, the way the measure is constructed is to use covering
by translates of a small neighbourhood of the origin to create some sort of outer
measure, and then send that neighbourhood to zero (normalising the measure so
that some fixed compact set has measure 1, say). We will not give the details here,
as it is a little tricky. However, it is relatively easy to show that Haar measure
is unique up to constants. For if dx and dx ˜ are two Haar measures, then for any
f, g ∈ Cc (G) one easily verifies using Fubini’s theorem that
Z Z Z Z Z Z
˜ ˜
g(y)f (x + y) dxdy = ( f (x) dx)( g(y)dy) = ( f (x) dx)( g(−y) dy) ˜
G G G G G G
and so by choosing g so that the two g-integrals are non-zero, we see that dx and
˜ differ by a constant multiple. From this we also conclude that Haar measure is
dx
always reflection-symmetric, since the reflection of Haar measure is also translation-
invariant and assigns the same weight to symmetric sets as the original measure.

We can now define convolution operators as before, but on L1 (G) rather than
L2 (G). This turns L1 (G) into a B ∗ -algebra (a Banach algebra with a conjugation,
namely f˜(x) := f (−x)). The analogue of “maximal translation-invariant subspace”
is now “maximal convolution ideal” of L1 (G). The Gelfand-Mazur theorem argu-
ment works in this setting and shows that such maximal convolution ideals have
codimension one, and are therefore identifiable with non-trivial B ∗ -algebra homo-
morphisms λ from L1 (G) to C. By 1
R the duality of L (G) and L ∞
∞
(G), such homo-
morphisms take the form λ : f 7→ G f (x)χ(x) dx for some χ ∈ L (G). Using the
homomorphism property and approximations to the identity we can show that
λ ◦ Transy = χ(y)λ;
since translations are continuous in L1 (G) we conclude that χ is continuous. Also,
the identity also reveals that χ is multiplicative: χ(x + y) = χ(x)χ(y). Since χ is
also in L∞ (G) and non-trivial, we conclude that |χ| = 1. Thus χ is a continuous
 LECTURE NOTES 9 7

multiplicative map from G to S 1 , otherwise known as a multiplicative character.

Again, we can write χ(x) = e(ξ · x), where ξ : G → R/Z is a continuous homomor-
phism from G to R/Z, otherwise known as a additive character or frequency.

Let Ĝ be the collection of additive characters; this is an abelian group. We can

define the Fourier transform fˆ : Ĝ → C for any f ∈ L1 (G) by
Z
fˆ(ξ) := f (x)e(−ξ · x) dx.
G

Thus each frequency ξ generates a linear functional on L1 (G), i.e. we have an

embedding of Ĝ in L1 (G)∗ . We give Ĝ the topology induced by the weak-* topology
on L1 (G); this is easily verified to make Ĝ a locally compact abelian group. By the
preceding discussion, fˆ is essentially the Gelfand transform of f in the B ∗ -algebra
L1 (G).

We let A(Ĝ) := {fˆ : f ∈ L1 (G)} (this is known as the Wiener algebra of Ĝ.)
One easily verifies that f[ ∗ g = fˆĝ, and so A(Ĝ) becomes a B ∗ -algebra with point-
wise product (rather than convolution) as the multiplication operation. It is also
translation-invariant. There is an analogue of the Riemann-Lebesgue lemma, which
asserts that A(Ĝ) ⊂ C0 (Ĝ), where C0 (Ĝ) is the space of continuous functions which
go to zero at infinity. As A(Ĝ) is also closed under conjugation and separates points,
the Stone-Weierstrass theorem then shows that A(Ĝ) is in fact dense in C0 (Ĝ) in
the uniform topology. This makes A(Ĝ) a useful class of “test functions” to verify
a variety of Fourier-analytic identities.

Now we move towards the all-important Fourier inversion formula. Let M (Ĝ) be
the space of finite Radon measures on Ĝ. Given such a measure ν, one can define
its inverse Fourier transform F ∗ ν ∈ C(G) by
Z
F ∗ ν(x) := e(ξ · x) dν(ξ).
Ĝ
We let B(G) denote the space of all such inverse Fourier transforms. One easily
verifies that this is a translation-invariant sub-algebra of C(G). From Fubini’s
theorem we have the duality relationship
Z Z
f F ∗ ν(x) dx = fˆ(ξ)dν(ξ)
G Ĝ
1
for all f ∈ L (G) and ν ∈ M (Ĝ).

So far we have not shown that Ĝ is large in any sense. One convenient way to
obtain this largeness is to show that B(G) is large. A key tool for achieving this is
Theorem 2.1 (Bochner’s theorem). Let f ∈ C(G) be positive semi-definite, in
the sense that we have hf ∗ µ, µi ≥ 0 for any finitely supported complex measure µ.
Then f ∈ B(G), and furthermore we have f = F ∗ ν for a non-negative measure ν.

We remark that the converse is easily established: if ν is a non-negative finite Radon

measure on Ĝ, then F ∗ f is positive semi-definite.
8 TERENCE TAO

Proof Let f be positive semi-definite. It is easy to see that f (x) = f (−x) and
|f (x)| ≤ f (0) for all x; we can then normalise f (0) = 1.

It is not hard to use the hypothesis and limiting arguments to show that hf ∗g, gi ≥ 0
for all simple functions g of finite measure support, and thence also for all g ∈ L1 (G).
This implies that the inner product hg, hif := hf ∗ g, hi is positive semi-definite,
and thus by the Cauchy-Schwarz inequality
|hg, hif | ≤ |hg, gif |1/2 |hh, hif |1/2 .
Letting h be an approximation to the identity, and taking limits, using the fact
that f is continuous and equals 1 near the origin, one soon concludes that
|f ∗ g(0)| ≤ |f ∗ g ∗ g̃(0)|1/2
where g̃(x) := g(−x). We can iterate this repeatedly and obtain
|f ∗ g(0)| ≤ |(Tg Tg∗ )2 f (0)|1/2
n n+1

1/2n+1
≤ k(Tg Tg∗ )2 kL∞ (G)→L∞ (G) ,
n

1/2n+1
= k(Tg Tg∗ )2 kL1 (G)→L1 (G) ,
n

where Tg is the operation of convolution by g. Taking limits as n → ∞ we can thus

bound |f ∗ g(0)| by the spectral radius ρ(Tg ). But Gelfand theory tells us that
ρ(Tg ) ≤ kĝkL∞ (Ĝ) .
(Quick sketch of proof: observe that g−z is invertible (hence holomorphic) in L1 (G)
if |z| > kĝkL∞ (Ĝ) (otherwise the range of g − z is a proper ideal, hence contained
in a maximal ideal, contradiction) and Cauchy’s integral formula outside of a circle
then quickly shows that
k(f f ∗ )n kL1 (G) /kĝk2n
L∞ (Ĝ)
→ 0 as n → ∞
and the claim follows. The reverse inequality is also true.)

To summarise so far, we have shown that

|f ∗ g(0)| ≤ kĝkL∞ (Ĝ)
for all g ∈ L1 (G). Thus the map ĝ 7→ f ∗ g(0) is a bounded linear functional with
norm at most 1 on A(Ĝ), and hence extends by density to C(Ĝ). Applying the
Riesz representation theorem, we obtain a finite Radon measure ν on Ĝ with total
mass one Z
f ∗ g(0) = ĝ dν
Ĝ
for all g ∈ L1 (G). The right-hand side is also equal to hg, F ∗ νi, and so by duality
we see that f is (up to some harmless conjugations) the inverse Fourier transform
of a measure of total mass 1. Since f (0) = 1, we also see that this measure has to
be non-negative as required.

One easily verifies that for any f ∈ L2 (G), the function f ∗ f˜ is positive definite,
indeed hf ∗ f˜ ∗ µ, µi = kf ∗ µk2L2 (G) ≥ 0. By Bochner’s theorem we conclude that
f ∗ f˜ ∈ B(G). From Young’s inequality we thus also see that if f ∈ L1 (G) ∩ L2 (G)
 LECTURE NOTES 9 9

then f ∗ f˜ ∈ L1 (G) ∩ B(G). From this and translation invariance, we obtain a large
class of functions in L1 (G) ∩ B(G). The significance of this is that we can establish
a preliminary inversion formula in this class:
Proposition 2.2 (Preliminary inversion theorem). There exists a Haar measure
dξ on Γ̂ such that whenever f ∈ L1 (G) ∩ B(G), then fˆ ∈ L1 (Ĝ) and
Z
f (x) = fˆ(ξ)e(ξ · x) dξ.
Ĝ

Proof Since f ∈ B(G), there exists νf ∈ M (Ĝ) such that f = F ∗ νf . Observe that
if f, g ∈ L1 (G) ∩ B(G) and h ∈ L1 (G), then
Z Z
fˆĥ dνg = f[
∗ h dνg = f ∗ h ∗ F νg (0) = f ∗ g ∗ h(0).
Ĝ Ĝ
By symmetry of convolution we thus have
Z Z
fˆĥ dνg = ĝ ĥ dνf
Ĝ Ĝ

for all ĥ ∈ A(Ĝ). Since A(Ĝ) is dense in C(Ĝ), we conclude from the Riesz repre-
sentation theorem that
fˆdνg = ĝdνf .
To put this another way, the quantity dνf /fˆ is independent of f on any region on
which the quotient is well-defined (i.e. fˆ is non-zero). Given any ξ ∈ Ĝ, one can
engineer an f in L1 (G) ∩ B(G) whose Fourier transform does not vanish near ξ by
taking an expression of the form f = h ∗ h̃ for some approximation to the identity
h, by the previous discussion; we can then glue all the dνf /fˆ measures together to
create a Radon measure dξ such that dνf = fˆ dξ for all f ∈ L1 (G) ∩ B(G). By
translating νf (which modulates f , which translates fˆ) we see that dξ is translation-
invariant and is thus a Haar measure. Since f = F ∗ νf , the claim now follows.

One consequence of the inversion theorem is that

Z
f (0) = fˆ(ξ) dξ
Ĝ

whenever f ∈ L1 (G) ∩ B(G), which implies also that

Z
f ∗ g(0) = fˆ(ξ)ĝ(ξ) dξ
Ĝ
1
for f, g ∈ L (G) ∩ B(G). This easily implies the Plancherel and Parseval identities
in this class. Using approximations to the identity, one can show that L1 (G)∩B(G)
is dense in L2 (G), and so we obtain Plancherel’s theorem, namely that there is a
unique unitary extension of the Fourier transform to L2 (G).

Given that Haar measure is defined up to a multiplicative constant, the only re-
maining question is to determine what this constant is. But by testing against a
single explicit non-trivial function, one can recover this constant. Some standard
examples:
10 TERENCE TAO

Example 2.3. Let G = Z. Then it is easy to see that the only additive characters
x 7→ ξ · x from Z to R/Z can be parameterised by a frequency ξ ∈ R/Z, given by
the formula ξ · x = xξ. Thus Ĝ ≡ R/Z. Testing against f (x) := 1x=0 (so fˆ(ξ) = 1
we see that the Fourier measure dξ on Ĝ is just Lebesgue measure.
Example 2.4. Let G = Z/N Z. Then the additive characters x 7→ ξ · x from Z/N Z
to R/Z lift to one from Z to R/Z so are parameterised by ξ ∈ R/Z as before, but
also we must have N ξ = 0 in order to be able to descend back to Z/N Z. Thus Ĝ
is equivalent to the N th roots of R/Z, which can also be identified with Z/N Z (if
we change the pairing (x, ξ) 7→ ξ · x to ξ · x = ξx/N ). Thus Ĝ ≡ Z/N Z. Testing
against f (x) := 1x=0 (so fˆ(ξ) = 1 we see that the Fourier measure dξ on Ĝ is just
counting measure.
Example 2.5. Let G = R. Then it is easy to see that the only additive characters
x 7→ ξ · x from R to R/Z can be parameterised by a frequency ξ ∈ R, and given
by the formula ξ · x = ξx mod 1; this can be seen by first using the continuity to
observe that for x close enough to zero, ξ · x lies in (say) a 0.1 neighbourhood of
the origin in R/Z, and then one can easily show using the homomorphism property
and continuity that ξ · x is linear near the origin, then one can extend to all of
R using the homomorphism property again. Thus Ĝ ≡ R. Testing against the
2
function f (x) := e−πx (say) one sees that the required Fourier measure dξ on Ĝ
is just Lebesgue measure.
Example 2.6. Let G = R/Z. The additive characters x 7→ ξ · x from R/Z to
R/Z descend from characters from R to R/Z, but in order for the descent to work
properly the frequency ξ must be an integer. Thus Ĝ = Z. Testing against f (x) := 1
(so fˆ(ξ) = 1ξ=0 ) we see that dξ is just counting measure.

Observe that each x ∈ G induces an additive character on Ĝ by the map x : ξ 7→ ξ·x.

ˆ
This allows one to canonically embed G in Ĝ; this map is easily seen to be an
injective homomorphism and homeomorphism. We have two Plancherel theorems
ˆ
on Ĝ, one with frequencies in Ĝ and one with frequencies in G. The only way these
ˆ ˆ
can be compatible is if G has full measure in Ĝ and the Fourier measure of Ĝ is
ˆ
dx. In fact one can show that G is equal to all of Ĝ, by observing that (by the full
measure property) G is dense, while also being locally compact. This equivalence
ˆ
G ≡ Ĝ is sometimes referred to as Pontryagin duality.

3. The Walsh ring

For most applications in real-variable harmonic analysis, the general theory of

Fourier analysis on LCA groups can be specialised to the classical cases of the
real line, torus, integers, cyclic groups, and products thereof. There are however a
number of other interesting groups (or rings) on which this theory is useful. One
is Fourier analysis on the ring AQ of adeles, which is of significant importance
in algebraic number theory but which will not be discussed here. Another is to
multiplicative abelian groups such as R+ (which leads to the Mellin transform) or
 LECTURE NOTES 9 11

(Z/N Z)∗ (which leads to the theory of characters in number theory). The other
is the Walsh ring or Cantor-Walsh ring 1 Wp , which is a dyadic variant of the real
number ring R and which serves as a useful model for harmonic analysis on the real
line (among other things, it manages to avoid the perennial presence of Schwartz
tails which plague the analysis on R.

To define the Walsh ring, we need an prime p; in many applications we set p = 2

(giving the dyadic Walsh ring) or p = 3 (giving the ternary Walsh ring). We let Fp
be the field of p elements (identified with Z/pZ and with {0, . . . , p − 1} in the usual
manner), and let Fp (t) be the ring of Laurent polynomials in one indeterminate t,
thus a typical non-zero element of Fp (t) is of the form
B
X
x= a n tn
n=A
for some finite integers A ≤ B and coefficients an ∈ Fp with aB 6= 0. We can also
represent this element more suggestively in “base p notation” as
x = aB aB−1 . . . a0 .a−1 . . . aA .
PB
Indeed, if we identify an element n=A an tn of Fp (t) with the base p terminating
PB
decimal n=A an p−n (recall that we identify Fp with {0, . . . , p − 1}), we obtain
a ring which resembles the ring of base p terminating decimals in the reals, but
where addition and multiplication are performed without the “carry” operation (or
in more modern jargon, the cocycle is trivial).

Just as the reals R are the metric completion of the rationals Q, the Walsh ring
Wp are the metric completion of the Laurent ring Fp (t), using the metric d(x, y) :=
kx − ykp := p−B , where B is the degree of x − y (i.e. the largest power of t which
appears; we use the convention that the degree of 0 is −∞). One easily observes
that the ring operations are locally uniformly continuous on Fp (t) and thus extend
in a unique continuous manner to Wp . The completion can also be identified with
the one-sided infinite series
XB
x= a n tn
n=−∞
over Fp ; one can also almost identify this ring with the half-line R+ by identifying
PB
x with n=−∞ an p−n , but of course the infamous 0.999 . . . = 1 problem shows
that there is a problem with doing so at every element of Fp (t). Strictly speaking,
one thus has to adjoin an additional element infinitesimally to the left of every
non-zero terminating base p decimal in R+ before one has a fully accurate repre-
sentation of the Walsh ring Wp , but we will gloss over this annoying technicality;
in particular, we shall abuse notation and describe elements of Wp using the for-
PB PB
mula n=−∞ an p−n rather than the more accurate n=−∞ an tn . (An alternate
PB
approach is to identify x with n=−∞ an q −n for some q > p, which identifies Wp
bijectively with a Cantor set, but then the analogy between Wp and R+ is dimin-
ished.) We can pull back Lebesgue measure on R+ to create a measure dx on Wp ,
which is easily seen to be a Haar measure.
1This notation is not standard.
12 TERENCE TAO

The metric d extends of course to the completion Wp ; it is somewhat related to

the usual Euclidean metric on R+ by observing that d(x, y) ≥ 12 |x − y|, but it is
possible for d(x, y) to much larger than |x − y| (this occurs for instance if x and y
are close to but on opposite sides of a terminating decimal). The metric d is in fact
better than a metric, it is an ultrametric (or non-archimedean metric), thus

d(x, y) ≤ max(d(x, z), d(z, y)).

This implies that the metric balls are nested. Indeed, the metric balls here are
nothing more than the p-dyadic intervals [ pjn , j+1
pn ) for n ∈ Z, j ≥ 0 (note that we
Pn−1
include the element pjn + m=−∞ p−1
pn , which is infinitesimally to the left of j+1
pn ,
in this ball).

The space Wp is homeomorphic to an unbounded Cantor set, as the identification

PB
x ↔ n=−∞ an q −n shows; in particular, Wp is locally compact as well as abelian.
Now let’s compute the Pontryagin dual of Wp . Let ξ : x 7→ ξ · x be an additive
character. Observe that for each integer n, ξ · pn must be a pth root of unity in
R/Z, since adding pn to itself p times in Wp gives zero (note this is not the same
as p × pn in Wp !). Thus we can write ξ · pn = a−n−1 /p for some a−n−1 ∈ Fp . By
continuity, we see that ξ · pn → 0 as n → +∞, and hence an = 0 for all sufficiently
large n, say n > N . If we then write
N
X
ξ˜ := an p−n ∈ Wp
n=−∞

and define the dot product

N
X M
X N
X
( an p−n , bm p−m ) := an b−1−m /p ∈ R/Z
n=−∞ m=−∞ n=−M

then we see that

ξ · x = ξ˜ · x

for all x ∈ Fp (t), and hence by continuity for all x ∈ Wp also. Thus Ŵp is identifiable
with Wp , using the dot product · : Wp × Wp → R/Z described above.

If one lets f := 1[0,1) in Wp , one easily checks that fˆ = f , which by Plancherel’s

theorem tells us that the Fourier measure dξ is the same as the original measure dx.
Note that in sharp contrast to the Fourier theory on R, in Wp it is perfectly possible
to have a function of compact support whose Fourier transform is also of compact
support. This leads to a very pleasant harmonic analysis in Wp , which can be
regarded as a simplified model of harmonic analysis in R: many of the foundational
theorems of the subject are somewhat simpler in this setting while still capturing
the main essence of the real-variable theory (a good example arises by comparing
the ordinary Hardy-Littlewood maximal function with its Walsh variant, namely
the dyadic (or p-adic) maximal function). We will however not pursue this theme
in detail here.
 LECTURE NOTES 9 13

4. Exercises

• Q1. (Poisson summation formula for finite abelian groups) Let G be a

finite abelian group, and let H be a subgroup of G. Define the orthogonal
complement H ⊥ ⊂ Ĝ of H as H ⊥ := {ξ ∈ Ĝ : ξ·x = 0 for all x ∈ H}. Show
that H ⊥ is a subgroup of Ĝ, and that Ĝ/H ⊥ is canonically identifiable with
Ĥ. In particular we see that
#G = (#H)(#H ⊥ ).
Show that
#H
1̂H (ξ) = 1 ⊥ (ξ)
#G H
and conclude the Poisson summation formula
1 X X
f (x) = fˆ(ξ)
#H
x∈H ⊥ ξ∈H
2
for all f ∈ L (G). More generally, show that
#G d X
f 1H (ξ) = fˆ(ξ + η)
#H ⊥
η∈H

and
#G \
f ∗ 1H (ξ) = 1H ⊥ (ξ)fˆ(ξ)
#H
for all f ∈ L2 (G) and ξ ∈ Ĝ (thus “the Fourier transform of restriction is
projection”, and vice versa).
• Q2. (Uncertainty principle for finite abelian groups) Let G be a finite
abelian group. Show that for any non-zero f ∈ L2 (G) we have
#(supp(f ))#(supp(fˆ)) ≥ #G.
(Hint: use the Plancherel and Hölder inequalities, together with the trivial
bound kfˆkl∞ (Ĝ) ≤ kf kL1 (G) .
• Q3. (Entropy uncertainty principle for finite abelian groups) Let G be a
finite abelian group. Establish the Hausdorff-Young inequality
kfˆklp′ (Ĝ) ≤ kf kLp(G)
for all 1 ≤ p ≤ 2 and f ∈ Lp (G); differentiate this at p = 2 to conclude the
entropy uncertainty principle
Z X
1 1
|f (x)|2 log + |fˆ(ξ)|2 log ≥0
x∈G |f (x)| 2 ˆ
|f (ξ)|2
ξ∈Ĝ

for all f ∈ L2 (G) with kf kL2 (G) = 1, using the convention 0 log 01 = 0.
Using Jensen’s inequality, give a second proof of the uncertainty principle
from Q2.
• Q4. (Adjoints in finite abelian groups) Let G be a finite abelian group,
and let φ : G → G be an automorphism. Define the adjoint operator
φ∗ : Ĝ → Ĝ by
φ∗ (ξ) · x := ξ · φ(x).
14 TERENCE TAO

Show that φ∗ is also an automorphism, and that for any f ∈ L2 (G) we have
f[
◦ φ = fˆ ◦ (φ∗ )−1 .
• Q5. Show that the Pontryagin dual of a compact abelian group is a discrete
abelian group, and vice versa.
• Q6. Let Wp be a Walsh group. Define the operators Dilqpn for 1 ≤ q ≤ ∞
and n ∈ Z and f ∈ Lq (Wp ) by
Dilqpn f (x) := pn/q f (p−n x).
(Note that Walsh multiplication of x by p−n corresponds with classical
multiplication.) Show that Dilqpn is an isometry on Lq (Wp ), and that
\
Dil q q′ ˆ
pn f = Dilp−n f .

• Q7 (Walsh uncertainty principle). Let Wp be a Walsh group, f ∈ L2 (Wp ),

and n ∈ Z. Show that fˆ is supported in the interval [0, pn ) if and only if f
is constant on every ball of radius p−n . We remark that the Walsh-Fourier
projection to the interval [0, pn ) is highly analogous to the Littlewood-Paley
projection ψ≤n , and enjoys essentially the same theory (in fact, the Walsh
theory is cleaner and simpler, and is also closely related to martingale
theory).
• Q8 (Tiles). Define a tile to be a set of the form P = I × w ⊂ Wp × Wp ,
where I, w are balls with |I||w| = 1. Show that the space VP := {f ∈
L2 (I) : fˆ ∈ L2 (w)} is one-dimensional, and that VP and VQ are orthogonal
whenever P, Q are disjoint tiles. If a set Ω ⊂ Wp × Wp can be partitioned
Ω = P1 ∪ . . . ∪ Pn into finitely many disjoint tiles, show that the vector
space VΩ := VP1 + . . . + VPn is independent of the choice of partition.

Department of Mathematics, UCLA, Los Angeles CA 90095-1555

E-mail address: [email protected]