LECTURE NOTES 2 FOR 247A

TERENCE TAO

1. Complex interpolation

It is instructive to compare the real interpolation method of the preceding section

with the complex interpolation method. Now we need to take T to be linear rather
than sublinear; we now make the mild assumption that the form
Z
T f gd (1)
X
makes sense for all simple functions f, g of finite measure support.
Theorem 1.1 (Riesz-Thorin complex interpolation). Let T be a linear operator
such that the form (1) is well-defined. Let 0 < p0 , p1 , 1 q0 , q1 and
A0 , A1 > 0 be such that we have the strong-type bounds
kT f kLqi (Y ) Ai kf kLpi (X)
for all simple functions f of finite measure support, and i = 0, 1. Then we have
kT f kLq (Y ) A kf kLp (X)
for all simple functions f of finite measure support and 0 1, where p , q ,
A is as in equation (23) of last weeks notes.

Proof We may assume that (p0 , q0 ) 6= (p1 , q1 ) otherwise the claim is trivial. We
may then normalise A0 = A1 = 1 as in the previous weeks notes. By duality and
homogeneity it suffices to show that
Z
| T f g d| 1 whenever kf kLp (X) = kg kLq (Y ) = 1
X
for all simple functions f , g of finite measure support and all 0 1. Note
that the claim is already true by hypothesis when = 0, 1.

The idea is to use the three lines lemma from last weeks notes. The problem is
that the inequality is not complex analytic in as stated. However we can fix this
as follows. Observe that if f is a simple function with Lp (X) norm 1, then we
can factorise
f = F01 F1 a
where F0 , F1 are non-negative simple functions with Lp0 (X) and Lp1 (X) norms
respectively equal to 1, and a is a simple function of magnitude at most 1. Indeed
we can set a = sgn(f ) and Fi = |f |p /pi . (Some minor changes need to be made
1
2 TERENCE TAO

for the limiting case when one or both of the pi are equal to infinity; we leave this
to the reader.) Similarly we can factorise

g = G1
0 G1 b

where G0 , G1 are non-negative simple functions with Lq0 (Y ) and Lq1 (Y ) norms
respectively equal to 1, and b is a simple function of magnitude at most 1.

Now consider the complex-valued function

Z
H(z) := T (F01z F1z a)G1z
0 Gz1 b d.
X

Because T is linear and all functions are simple, it is not difficult to see that H is
an entire function of z of at most exponential growth. It is bounded by 1 on both
sides of the strip 0 Re(z) 1, hence bounded also within the strip. Setting z =
we obtain the result.

Comparing the Riesz-Thorin theorem with the Marcinkiewicz theorem, we observe

some key differences. On the plus side, we do not lose an unspecified constant in
the estimates, and we do not have the restriction q p. On the other hand, the
hypotheses require strong-type control on T , not restricted weak-type control.

The Riesz-Thorin theorem can also be rephrased as follows. Let kT kLp(X)Lq (Y )

be the operator norm of T from Lp (X) to Lq (Y ); then this quantity is log-convex
as a function of (1/p, 1/q). Because of this, the above theorem is sometimes known
as the Riesz-Thorin convexity theorem.

An important observation of Elias Stein is that the above proof can be generalised
to the case where the operator T itself varies analytically:

Theorem 1.2 (Stein interpolation). Let Tz be a family of linear operators on the

strip 0 Re(z) 1 such that the for each f DX , g DY , the form
Z
(Tz f )g dX
Y

is absolutely convergent and a complex analytic function of z on the strip which

grows slower than double-exponentially in z. Let 0 < p0 , p1 , 1 q0 , q1
and A0 , A1 > 0 be such that we have the strong-type bounds
kTz f kLqi (Y ) Ai kf kLpi (X)
for all simple functions f of finite measure support, i = 0, 1, and Re(z) = i. Then
we have
kT f kLq (Y ) A kf kLp (X)
for all simple functions f of finite measure support and 0 1.

Note that the Riesz-Thorin theorem corresponds to the case where Tz = T is

independent of z.
 LECTURE NOTES 2 3

Proof We repeat the previous argument; the only observation to make is that the
function Z
H(z) := Tz (F01z F1z a)G1z
0 Gz1 b d
X
continues to be complex-analytic; this is easiest seen by decomposing all the simple
functions into indicator functions.

One particularly common special case of this theorem is

Corollary 1.3 (Stein-Weiss interpolation theorem). Let T be a linear operator
such that (1) is well-defined, and let w0 , w1 : X R+ and v0 , v1 : Y R+
be non-negative weights which are integrable on every finite measure set. Let 0 <
p0 , p1 , 1 q0 , q1 and A0 , A1 > 0 be such that we have the strong-type
bounds
kT f kLqj (Y,vj dY ) Aj kf kLpj (X,wj dX ) (2)
for all simple functions f of finite measure support, and j = 0, 1. Then we have
kT f kLq (Y,v dY ) A kf kLp (X,w dX )

for all simple functions f of finite measure support and 0 1, where v :=

v01 v1 and w := w01 w1 .

Proof We begin with a number of qualitative reductions. A monotone convergence

argument (replacing the weights vj , wj by various truncated versions) lets one
reduce to the case where vj are bounded from above, and wj are bounded from
below. Another monotone convergence argument (this time replacing Y by various
truncated versions) lets us reduce to the case where v is bounded from below, which
implies that the vj are also bounded from below. One can also similarly restrict Y
to have finite measure, and another truncation argument (to the support of f ) also
reduces X to have finite measure. Then another limiting argument (restricting f
to the region where w is bounded by some N , and then letting N go to infinity,
using the hypotheses (2) and Fatous lemma to get enough convergence) lets one
reduce to the case where w is bounded from above, which implies that wi are also
bounded from above. Now that all the vj and wj are bounded both above and
below one can approximate by simple functions and use a limiting argument to
reduce to the case where vj , wj are simple and never vanishing.

We can now legimately define the analytic family of operators

(1z)/q0 z/q1 (1z)/p0 z/p1
Tz := v0 v1 T w0 w1
for any complex z, which is well defined on simple functions. The hypotheses (2)
then ensure that kTz kLpj (X,X )Lqj (Y,Y ) Aj whenever j = 0, 1 and Re(z) =
j, the point being that multiplication by imaginary exponentials such as v0it are
contractions on every Lp space; one also easily checks that the other hypotheses of
the Stein interpolation theorem are satisfied. We conclude that
kT kLp (X,X )Lq (Y,Y ) A
from which the claim follows.
4 TERENCE TAO

The real and complex interpolation methods provide a surprisingly powerful way
to prove estimates involving Lp norms: in order to prove a range of such estimates,
it suffices to do so for the extreme cases (possibly weakening strong type to weak
type or restricted weak type in the process). We shall shortly give several disparate
examples of this interpolation approach.

2. Duality

Both real and complex interpolation give a means to deduce new bounds on an
operator T from old ones. Another important technique is that of duality, which
basically replaces the input function f with an output function T g, and the output
function T f with an input function g. Generally speaking, the situation is as
follows. One has a domain X, and some class of test functions DX on X (such
as simple functions, or perhaps the Schwartz class), and one also has a domain Y
with a class DY . We shall require DX and DY to be vector spaces, though we
will not require any topological structure on these spaces (in particular, they do
not need to be complete with respect to any particular topology). We can then
define the dual space DY of linear functionals g 7 hg, hiY on DY ; we do not
require these functionals to be continuous in any topology, and so DY is in fact a
very large space. In particular, it typically contains very general objects such as
locally integrable functions, measures, or distributions, where we (formally at least)
identify a function (or distribution, etc.) h(y) with the linear functional
Z
g 7 hg, hiY := g(y)h(y) dY (y).
Y

We also adopt the convention hh, giY := hg, hiY . One can of course have similar

conventions for DX and DX .

We assume that T is a linear operator from DX to DY , thus given any f DX

and g DY the sesquilinear form
(f, g) 7 hT f, giY
makes sense. We then automatically have an adjoint T : DY DX

defined by
hf, T giX := hT f, giY .
At present, T g only takes values in the abstract dual space DX

. However in many

cases T g can be identified more concretely with a function, measure, or at least a
distribution. For instance, if X, Y are topological measure spaces, DX and DY are
the compactly supported continuous functions on X and Y , and T is given by an
integral operator
Z
T f (y) := K(x, y)f (x) dX (x)
X
for some locally integrable kernel K : X Y C, then the sesquilinear form
Z Z
hT f, giY = K(x, y)f (x)g(y) dX (x)dY (y)
Y X
 LECTURE NOTES 2 5

makes sense for all f DX , g DY , and by the Fubini-Tonelli theorem we see

that the adjoint T is given by
Z
T g(x) := K(x, y)g(y) dY (y).
Y

The point of introducing the adjoint T is that many bounds on the original oper-
ator T are equivalent to those for T . We illustrate this point with the Lp spaces,
although from the proof one sees that the same would hold true for any spaces
whose norm can be characterised by a duality relationship.
Theorem 2.1 (Duality). Let 1 p, q and A > 0. Let DX and DY be linear

spaces of functions on X and Y which are dense in Lp (X) and Lq (Y ) respectively.
Let T : DX DY be a linear transformation, and let T : DY DX
be its adjoint.
Suppose we make the (qualitative) assumptions that T maps DX to Lq (Y ) and T

maps DY to Lp (X). Then the following (quantitative) assertions are equivalent.

(i) For all f DX we have kT f kLq (Y ) Akf kLp(X) .

(ii) For all g DY we have kT gkLp (X) AkgkLq (Y ) .
(iii) For all f DX and g DY we have |hT f, giY | Akf kLp (X) kgkLq (Y ) .
(iv) For all f DX and g DY we have |hf, T giX | Akf kLp(X) kgkLq (Y ) .

Proof The equivalence of (i) with (iii) follows from the dual characterisation of the

Lq (Y ) norm from last weeks notes (and the density of DX in Lq (Y )). Similarly
for the equivalence of (ii) with (iv). The equivalence of (iii) with (iv) follows from
the definition of T .

Remark 2.2. This theorem differs very slightly from the usual duality theorem,
which asserts that if T is continuous from a normed vector space V to another W ,
then its adjoint T is well defined from W to V and has the same operator norm.
That duality theorem is fine for most applications, but runs into a slight difficulty
when V is an L type space, as the dual can then get excessively large. (Having
the domain space W too large is of course not a problem.) In particular when
dealing with non-reflexive spaces such as L1 or L , one loses the equivalence of the
two statements. By passing to dense classes and making the a priori qualitative

hypotheses that T maps DX to Lq (Y ) and T maps DY to Lp (X) (which are often
easily verified in practice, since we can usually1 take the dense classes DX and DY
to be very nice functions such as the Schwartz class).

3. Conditional expectation

Let us now start using our interpolation theorems. The first application is to the
useful operation of conditional expectation. This operation was first developed in
probability theory but has since proven to be useful in ergodic theory and har-
monic analysis; furthermore, several further operations in harmonic analysis (such
1One caveat comes when L spaces are involved, because classes which decay at infinity, such
as the Schwartz class, are usually not dense in L .
6 TERENCE TAO

as Littlewood-Paley projections) behave sufficiently like conditional expectation

that it is worth studying conditional expectation simply to gain intuition on these
other operators.

Let (X, Bmax , ) be a measure space (which need not necessarily be a probability
space, though this is certainly an important special case), and let B be2 a -finite
sub--algebra of Bmax . Then L2 (B) = L2 (X, B, ) is a closed subspace of the
Hilbert space L2 (Bmax ) = L2 (X, Bmax , ), and thus has an orthogonal projection,
which we shall denote by the map f 7 E(f |B). Thus E(f |B) is B-measurable, and
is equal to f if and only if f is also B-measurable.
Problem 3.1. Let X = R with Lebesgue measure d = dx, and let Bmax be the
Borel or Lebesgue -algebra. Let B be the sub--algebra generated by the intervals
[n, n + 1) for n Z. Show that for f L2 (Bmax ), E(f |B) is given by the formula
Z x+1
E(f |B)(x) := f (y) dy
x

for all x R, where x is the greatest integer less than x. This model example is
worth keeping in mind for the rest of this discussion.

We have the orthogonal projection identities

Z Z Z
f E(g|B) d = E(f |B)g d = E(f |B)E(g|B) d
X X X
for all f, g L2 (Bmax ). In particular we have
Z Z
E(f |B)g d = f g d
X X
whenever f L2 (Bmax ) and g L2 (B). By the self-duality of the Hilbert space
L2 (B), this is a full description of the function E(f |B) L2 (B). From this descrip-
tion we quickly infer that

If f is real-valued, then so is E(f |B);

If f is real and non-negative, then so is E(f |B);
If h L (B), then
E(f h|B) = hE(f |B). (3)
2
In other words, conditional expectation is linear when L (Bmax ) is viewed
as a module over L (B), not just over C.

By linearity, we thus see that E(f |B) = E(f |B) and E(Ref |B) = ReE(f |B). Also,
if f , g are real-valued and f g pointwise, then E(f |B) E(g|B) pointwise. For
real-valued functions this already gives the triangle inequality |E(f |B)| E(|f ||B).
For complex-valued functions, a naive splitting into real and imaginary parts will
lose a factor of two: |E(f |B)| 2E(|f ||B). This factor can eventually be recovered
by a tensor power trick (try it!) but this is overkill. Instead, we exploit phase
2A small subtlety here: the hypothesis that B is -finite does not automatically follow from
the hypothesis that B is a sub--algebra of Bmax . Of course this is not an issue when X has finite
measure, which is for instance the case in probability theory.
 LECTURE NOTES 2 7

invariance to deduce the complex triangle inequality from the real one. Observe
that the complex triangle inequality |E(f |B)| E(|f ||B) is invariant if we multiply
f by a B-measurable phase h (thus |h| = 1). Using this (and the B-measurability
of E(f |B), and the module property) we can reduce to the case where E(f |B) is
real and non-negative. But in that case
|E(f |B)| = ReE(f |B) E(Ref |B) E(|f ||B)
as desired.

By definition of orthogonal projection we know that the map f 7 E(f |B) is a

contraction on L2 (Bmax ):
kE(f |B)kL2(Bmax ) kf kL2(Bmax ) .
We can then ask whether the same is true for other Lp :
Proposition 3.2 (Conditional expectation contracts Lp ). If 1 p and
f L2 (Bmax ) Lp (Bmax ), then
kE(f |B)kLp(Bmax ) kf kLp(Bmax ) .

Proof By complex interpolation3 it suffices to check the endpoint cases p = 1, .

We start with the p = case. For any set E in B of finite measure and f
L2 (Bmax ) L (Bmax ), we have the pointwise bound
|E(f |B)|1E = |E(f 1E |B)| E(|f |1E |B) kf kL (Bmax ) E(1E |B) = kf kL (Bmax ) 1E
and thus (since X is the countable union of sets of finite measure in B) the desired
bound follows.

Now take p = 1. If f L1 (Bmax ) L2 (Bmax ), and E B has finite measure then

Z Z
|E(f |B)|1E d E(|f ||B)1E d
X
ZX
= |f |1E d
X
kf kL1 (Bmax )
and hence by monotone convergence (and the -finiteness)
kE(f |B)kL1(B) kf kL1 (Bmax ) .
(One can also deduce the p = 1 case from the p = case using the self-adjointness
of conditional expectation and a duality argument; we omit the details.)

For 1 p < we know that L2 (Bmax ) Lp (Bmax ) (which contains all simple func-
tions of finite measure support) is dense in Lp , and thus there is a unique continuous
extension of conditional expectation to Lp (Bmax ), which is also a contraction. The
same argument works when p = and X has finite measure. When X has infi-
nite measure, the Hahn-Banach theorem gives an infinity of possible extensions to
3Real interpolation would also work, but then one needs the tensor product trick to eliminate
the constant. This in turn requires some knowledge of product measures, and how conditional
expectation reacts to tensor products.
8 TERENCE TAO

L (Bmax ) which are contractions, however there is a unique such extension which
still retains the module property (3); we leave the verification of this to the reader.
Remark 3.3. When p = 2, E(f |B) is the closest B-measurable function to f in Lp
norm; however we caution that this claim is not true for other values of p.

Let us give an application of this contractibility. Suppose we have an increasing

family
B1 B2 . . . Bn
W
of -finite sub--algebras of Bmax
S. We can then form their limit n=1 Bn , defined
as the -algebra generated by n=1 Bn .
Example 3.4. If X = R, and Bn is W the -algebra generated by the dyadic intervals
[k/2n , (k + 1)/2n ) for k Z, then
n=1 Bn is the Borel -algebra (why?).

Proposition 3.5 (Norm convergence of nested expectations). W If 1 p < and

f Lp (Bmax ), then E(f |Bn ) converge in Lp norm to E(f | n=1 Bn ).

Proof From the nesting property we have

_
E(f |Bn ) = E(E(f | Bm )|Bn )
m=1
W
and so we see that (by replacing f with E(f | m=1 Bm ) if necessary)
W to prove the
proposition it suffices to do so in the case where f is already n=1 Bn -measurable.

Observe that the space of sets in Bmax which can be approximated inWmeasure to
S
arbitrary accuracy by a Wset in n=1 Bn is a -algebra and thus contains n=1 Bn . In

particular, everyWset in n=1 Bn can be so approximated. This implies that simple

functions in Lp ( n=1 Bn ) can be approximated in Lp norm by simple functions in
p
L (BSn ) for some n; since simple functions
W are themselves dense in Lp , this implies
p p
that n=1 L (Bn ) is dense in L ( n=1 Bn ).
S
Now for f n=1 Lp (Bn ) it is clear that E(f |Bn ) converges to f in Lp norm as
n . We also know that the conditional expectation operators f 7 E(f |Bn ) are
contractions on Lp , and in particular are uniformly bounded on Lp . The claim now
follows from the previously mentioned density and a standard limiting argument.

Remarks 3.6. The claim is false at p = ; consider for instance the indicator
function of [0, 1/3) in Example 3.4. We shall also address the issue of pointwise
almost everywhere convergence in next weeks notes. The trick of using uniform
bounds on operators, together with norm convergence on a dense class of functions,
to deduce norm convergence on all functions in the space, is a common4 one and
will be seen several times in these notes.

4Indeed, the uniform boundedness principle implies, in some sense, that this is the only way
to establish a norm convergence result.
 LECTURE NOTES 2 9

4. Norm interchange

Consider a function f (x, y) of two variables. More precisely, consider two measure
spaces X = (X, BX , X ) and Y = (Y, BY , Y ), and consider a function f on the
product space X Y = (X Y, BX BY , X Y ). We can consider the mixed
norms Lpx Lqy (X Y ) of f for 1 p, q by
kf kLpxLqy (XY ) := kkf (x, )kLqy (Y ) kLpx (X)
thus for instance when p, q are finite we have
Z Z
kf kLxLy (XY ) = ( ( |f (x, y)|q dY (y))p/q dX (x))1/p .
p q

X Y
We can similarly define the interchanged norm Lqy Lpx (X Y ). By iterating the
triangle inequality we can see that these are indeed norms. The relationship between
the norms is as follows.
Proposition 4.1 (Interchange of norms). Let the notation and assumptions be as
above.

(i) When p = q, then kf kLpxLqy = kf kLqy Lpx = kf kLpx,y (XY ) .

(ii) When p q, then kf kLpxLqy kf kLqy Lpx .
(iii) When p q, then kf kLpxLqy kf kLqy Lpx .

Proof Claim (i) follows from the Fubini-Tonelli theorem (with the case p = q =
treated separately). For claim (ii), observe that the claim is already true for q = p,
and is also true for q = 1 by Minkowski. For the intermediate values of p, we can
modify the proof of the Riesz-Thorin interpolation theorem (adapted now to mixed
norms) to conclude the argument; we leave the details to the reader as an exercise.
Claim (iii) of course follows from claim (ii) by reversing p and q.

A more direct proof of (ii) is as follows. We can take p, q finite since the cases
p = and q = are easy. Raising things to the q th power, it suffices to prove
k|f |q kLp/q L1 k|f |q kL1 Lp/q ,
x y y x

p/q
but this follows from Minkowskis inequality (since Lx is a Banach space). This
proof is shorter, but perhaps a little harder to remember as it relies on an ad hoc
trick; in contrast, the interpolation proof, while more complicated, is at least relies
on a standard method and is thus easier to remember.
Problem 4.2. Let 1 q < p < , and suppose that kf kLpxLqy = kf kLqy Lpx < .
Show that |f | is a tensor product, i.e. there exist functions fx Lpx and fy Lqy
such that |f (x, y)| = fx (x)fy (y). Thus we only expect interchanging norms to be
a good idea when we believe the worst case example of f in our problem to be
roughly tensor product-like in nature.

Let us note two special (and easy) cases of norm interchange. Firstly, that maximal
functions control individual functions:
sup kfn kp k sup |fn |kp .
n n
10 TERENCE TAO

Secondly, we have the triangle inequality

X X
k fn kp kfn kp .
n n

The moral here is that maximal functions are harder to upper-bound, but are
conversely more powerful for dominating other expressions; conversely, it is easier to
bound norms of sums than sums of norms, though the latter are good for dominating
other things.
Problem 4.3 (Duality for mixed norm spaces). Show that if 1 p, q and
f Lpx Lqy , then
Z Z
kf kLpxLqy = sup{| f (x, y)g(x, y) dX (x)dY (y)| : kgkLp Lq 1}.
x y
X Y

Note that Proposition 4.1 is in some sense self-dual with respect to with this
duality relationship.

5. Schurs test

We now consider a general class of operators, the integral operators T = TK , defined

(formally, at least) by
Z
T f (y) = TK f (y) = K(x, y)f (x) dX (x) (4)
X

where K : X Y C is some measurable function, called the integral kernel or

simply the kernel 5 of the operator T . Of course, one needs some sort of integrability
condition6 on K in order for this operator to make sense, even for simple functions
f . For instance, if K is bounded, then T f makes sense (as a bounded function) for
any simple function f of finite measure support; or if X, Y are topological spaces
(with X , Y Radon measures) and K is locally integrable, then T f makes sense
(as a locally integrable function) for all continuous, compactly supported functions,
basically thanks to the Fubini-Tonelli theorem.
Remark 5.1. If X, Y are finite sets with counting measure, then K is just the
matrix associated to the linear transformation TK .

A fundamental problem in harmonic analysis is to determine conditions on the

kernel K which will ensure that T is of strong-type (p, q) (or maybe weak-type or
restricted weak-type), and if so, to estimate the constant. In general, these problems
are very difficult. However, if one is mapping from L1 into a Banach space, or
mapping from a Banach space into L , the problem becomes much simpler. We
illustrate this with the Lp spaces, 1 p :

5Unfortunately, the word kernel is also used for the null space of T ; in cases where this could
lead to confusion the full terminology integral kernel is of course recommended instead.
6Later on we shall study singular integrals, in which K is a distribution which typically does
not obey any absolute integrability condition, but nevertheless for which the integral operator can
make sense due to cancellation inherent in the kernel.
 LECTURE NOTES 2 11

Proposition 5.2 (Mapping from L1 ). Let 1 q and K : X Y C.

Suppose also that kK(x, )kLq (Y ) is uniformly bounded. Then the operator (4) is of
strong-type (1, q) (and absolutely convergent for all L1 (X)), and furthermore
kTK kL1 (X)Lq (Y ) = sup kK(x, )kLq (Y ) (5)
xX

(as always, supremum is understood to be essential supremum).

Proof For any f L1 (X), Minkowskis inequality7

Z Z
k |K(x, y)||f (x)| dX (x)kLqy (Y ) k|K(x, y)||f (x)| kLqy (Y ) dX (x)
X X
Z
|f (x)| dX (x) sup k|K(x, y)| kLqy (Y )
X yY
<
and so in particular we see that T f (y) is absolutely convergent for almost every y,
and from the triangle inequality we also conclude that
kTK kL1 (X)Lq (Y ) sup kK(x, )kLq (Y ) .
xX
To get the converse inequality, we observe from limiting arguments that we may
take K to be a simple function on X Y ; indeed we may go further and take
K to be a product simple function, i.e. a linear combination of tensor products
of indicator functions on X and Y separately. In such a case, we can partition
X into finitely many sets of positive measure, on each of which K is independent
of x (excluding sets of measure zero of course). On one of these sets, call it A,
the supremum kK(x, )kLq (Y ) is attained. If we then test TK against 1A for some
A A of finite measure, we obtain the claim.

Remark 5.3. If kK(x, )kLq (Y ) is unbounded, then (5) strongly suggests that T is
not of strong type (1, q). However proving this rigorously is remarkably difficult,
because the lack of even qualitative control on K (other than measurability) defeats
the use of most limiting arguments. Nevertheless, for heuristic purposes at least, the
above proposition gives a necessary and sufficient condition to map from L1 to Lq
for any q 1. A similar argument works with Lq replaced by other Banach spaces,
but it unfortunately does not work for quasi-normed spaces such as L1, ; indeed,
weak-type (1, 1) estimates can be remarkably delicate to establish (or disprove).

There is a dual statement:

Proposition 5.4 (Mapping into L ). Let 1 p and K : X Y C.
Suppose also that kK(, y)kLp (X) is uniformly bounded. Then the operator TK
defined in (4) is of strong-type (p, ) (and absolutely convergent for all f Lp (X)),
and furthermore
kTK kLp (X)L (Y ) = sup kK(x, )kLp (X) (6)
xX

7This can be deduced from the triangle inequality by a monotone convergence argument,
approximating |K| and |f | from below by simple functions. The corresponding claim for K and f
then follows from a similar dominated convergence argument.
12 TERENCE TAO

(as always, supremum is understood to be essential supremum).

We leave the proof as an exercise to the reader; it is similar to the previous propo-
sition but relies on H
olders inequality (and its converse) rather than Minkowskis
inequality.

The above propositions give useful necessary and sufficient criteria for strong-type
(p, q) bounds when p = 1 or q = . For other cases, it does not seem possible to
find such a simple criterion for strong-type (p, q), although when the kernel K is
non-negative one can at least handle the p = or q = 1 cases easily:
Problem 5.5. Let K : X Y R+R be non-negative and 1 q . Then TK is
of strong type (, q) if and only if X K(x, y) dX (x) lies in Lq (Y ), and in fact in
this case we have
Z
kTK kL (X)Lq (Y ) = k K(x, y) dX (x)kLqy (Y ) .
X
R
Dually, if 1 p , then TK is of strong type (p, 1) if and only if Y K(x, y) dX (y)

lies in Lp (X), and in fact in this case we have
Z
kTK kLp (X)L1 (Y ) = k K(x, y) dY (y)kLp (Y ) .
y
Y

By combining Propositions 5.2, 5.4 with the Riesz-Thorin theorem we can obtain
a useful test for Lp boundedness:
Theorem 5.6 (Schurs test). Suppose that K : X Y C obeys the bounds
Z
|K(x, y)| dX (x) A for almost every y Y
X
and Z
|K(x, y)| dY (y) B for almost every x X
Y
for some 0 < A, B < . Then for every 1 p , the integral operator TK in
(4) is well-defined (in the sense that the integral is absolutely integrable for almost
every y) for all f Lp (X), with

kTK f kLp (Y ) A1/p B 1/p kf kLp(X) .

Proof The statement of the theorem is invariant8 under multiplying either X or

Y by a positive constant; hence we may normalise A = B = 1. By the triangle
inequality (replacing K and f with |K| and |f | respectively) it suffices to show the
claim when K and f are non-negative. By monotone convergence we may then take
f to be a simple function with finite measure support. More monotone convergence
then allows us to reduce to the case where X and Y have finite measure. In
particular from Fubinis theorem we now know that K is absolutely integrable, and

8Note how we are exploiting the non-assumption that the domain (X, ) and range (Y, )
X Y
are not required to be equal. Thus we see that generalising a proposition may in fact make it
easier to prove, as it can introduce more symmetries or other structural features.
 LECTURE NOTES 2 13

R
so the form Y
TK f g dY makes sense and is absolutely integrable for all simple f ,
g.

From the preceding two propositions we know that the claim is true for p = 1 and
p = , and the general case follows from the Riesz-Thorin theorem. (One can also
use real interpolation coupled with the tensor power trick.)

We can give an alternate proof of Schurs test, which is more elementary (using
real convexity rather than complex analyticity) but relies on a non-obvious trick,
as follows. Again we normalise A = B = 1. As the cases p = 1, are trivial (or
can be obtained by limiting arguments) we assume 1 < p < . By duality and
monotone convergence it suffices to show that
Z Z
|K(x, y)||f (x)||g(y)| dX (x)dY (y) kf kLp(X) kgkLp (Y )
X Y
for all simple functions f, g. We can normalise kf kLp(X) = kgkLp (Y ) = 1. We now
use the weighted arithmetic mean-geometric mean inequality
1 1
|f (x)||g(y)| |f (x)|p + |g(y)|p
p p
which if we rewrite it as
1
log |f (x)|p + p1 log |g(y)|p

1 log |f (x)|p 1 p
ep e + elog |g(y)|
p p
is seen to just be the convexity of the exponential function in disguise. We conclude
that
Z Z Z Z Z Z
1 1
|K(x, y)||f (x)||g(y)| dxdy |K(x, y)||f (x)|p dX (x)dY (y)+ |K(x, y)||g(y)|p dX (x)dY
X Y p X Y p X Y
Computing the first term by integrating in y first, and the second term by integrat-
ing in x first, we obtain the claim.
Remark 5.7. A variant on this elementary approach is to factor |K(x, y)||f (x)||g(y)|

as |K(x, y)|1/p |f (x)| and |K(x, y)|1/p |g(y)| and use H
olders inequality, estimating

these two factors in Lp (X Y ) and Lp (X Y ) respectively. (We thank Kenley
Jung for pointing out this simple argument.)
Problem 5.8. By observing what happens when one multiplies K, X , or Y by
positive constants, show that the hypotheses of Schurs test cannot be used to
deduce a strong type (p, q) (or even a restricted weak-type (p, q)) bound on T when
p 6= q.
Remark 5.9. An illustrative case of Schurs test is as follows: let K be an n n
matrix whose entries are non-negative, and whose row and column sums are all
bounded by A. Then the operator norm of K is also bounded by A.

Schurs test can be sharp. Suppose that X, Y have finite measure, that K is non-
negative, and that the hypotheses of Schurs test are satisfied exactly, in the sense
that Z
|K(x, y)| dX (x) = A
X
14 TERENCE TAO

for all y Y and Z

|K(x, y)| dY (y) = B
Y
for all x X. Then Fubinis theorem gives the identity
AY (Y ) = BX (X).
Also, we see that
TK 1X = A1Y
and so if we set f = 1X then we easily compute that

kTK f kLp(Y ) = A1/p B 1/p kf kLp(X)
for all 1 p , so Schurs test is completely sharp in this case. From this
analysis we see that we expect Schurs test to be sharp when (a) there is no os-
cillation in the kernel K, and (b) K Ris distributed homogeneously
R in x and y, in
the sense that the marginal integrals X |K(x, y)| dX (x) and Y |K(x, y)| dY (y)
behave like constant functions. When the latter condition is not true, then some-
times a weighted version of Schurs test is more effective (see Q1); alternatively, a
decomposition of the operator into components, on each of which Schurs test can
be applied, and which can be summed in some efficient manner, is often a better
approach.

When K oscillates, one usually does not expect Schurs test to be efficient. For
instance, consider the Fourier transform operator f 7 f on Rd , which is an integral
operator with kernel K(x, y) = e2ixy . The values of A and B for this kernel are
infinite, yet Plancherels theorem (which we shall review below) shows that this
operator is bounded in L2 (Rd ). Handling oscillatory integrals is in fact a major
challenge in harmonic analysis, requiring tools such as almost orthogonality and
various wave packet decompositions; we shall return to these issues in later notes.
Remark 5.10. There is in fact a sense in which the weighted version of Schurs test
is always sharp for non-negative kernels K 0, provided one chooses the weights
optimally. Indeed, suppose that kTK kLp (X)Lp (Y ) = A, and furthermore that this
bound is attained by some non-zero f , thus kTK f kLp(Y ) = Akf kLp(X) . Since K is
non-negative, we can
R assume without loss of generality that f is also non-negative;
we can normalise X f p dX = 1, thus
Z
(TK f )p dY = Ap .
Y
Elementary calculus of variations using Lagrange multipliers then reveals that

TK ((TK f )p1 ) = f p1
for some 0; integrating this against f reveals that = Ap . If we write w = f
and v = (TK f )p1 we thus have
Z
K(x, y)w(x) dX (x) = v(y)1/(p1)
X
and Z
K(x, y)v(y) dY (y) = Ap w(x)p1 .
Y
 LECTURE NOTES 2 15

It is then not difficult to formulate a weighted version of Schurs test which gives
kTK kLp (X)Lp (Y ) A. Thus we see that by choosing the weights to be associated
to the extremal functions for T , Schurs test does not lose any constants whatsoever.

6. Youngs inequality

We have the following variant of Schurs test.

Proposition 6.1 (Generalised fractional integration). Suppose that K : XY C
obeys the bounds
kK(, y)kLr,(X) . A for almost every y Y
and
kK(x, )kLr, (Y ) . A for almost every x X
for some 1 < r < and A > 0. Then the operator TK defined in (4) is of weak
type (1, r) with norm Or (A) and of restricted type (r , ) with norm Or (A). For
any 1 < p < q < and 1 s with p1 + r1 = 1q + 1, TK maps Lp,s (X) to
Lq,s (Y ) with norm Or,p,q,s (A). In particular, TK is of strong-type (p, q) with norm
Or,p,q (A).

Proof By splitting K into real and imaginary parts, and then into positive and
negative parts, we may assume that K is real and non-negative. We may also
restrict the functions f we are testing to also be real and non-negative, for similar
reasons. By multiplying K by a constant we may normalise A = 1.

From the dual characterisation of Lr, (X) (Problem 6.9 of last weeks notes) we
know that the Lr, (X) quasi-norm is equivalent to a norm. One can then mod-
ify the proof of Proposition 5.2 to obtain the weak type (1, r). A similar argument
(modifying Proposition 5.4) gives us the restricted type (r , ). The Lorentz claims
then follow from Marcinkiewicz interpolation, and the final strong-type claim fol-
lows by setting s = p and recalling that Lq,p (Y ) embeds into Lq .

A virtually identical argument (using complex interpolation instead of real) gives

Proposition 6.2 (Generalised Youngs inequality). Let 1 p, q, r be such
that p1 + 1r = q1 + 1, and suppose that K : X Y C obeys the bounds
kK(, y)kLr (X) A for almost every y Y
and
kK(x, )kLr (Y ) A for almost every x X
for som A > 0. Then the operator T defined in (4) is of strong-type (p, q) with
norm at most A.

We leave the proof to the reader as an exercise.

A particular case of Proposition 6.1 is the following inequality, which is of funda-

mental importance in PDE.
16 TERENCE TAO

Corollary 6.3 (Hardy-Littlewood-Sobolev fractional integration inequality). Let

d 1, 0 < s < d, and 1 < p < q < be such that
d d
= s.
q p
Then we have
Z
f (x)
k dxkLqy (Rd ) .d,s,p,q kf kLpx(Rd )
R d |x y|ds
for all f Lpx (Rd ).

1 d
Indeed, one simply sets K(x, y) := |xy| ds and r := ds and verifies the hypotheses

of Proposition 6.1. The reason for the terminology fractional integration will be
explained later in this course.
R f (x)
The quantity Rd |xy| ds dx is of course a convolution of f with the convlution

kernel |x|1ds . Convolutions can in fact be studied on more general groups than
Rd . Suppose we have a domain X which is both a measure space, X = (X, B, ),
and a multiplicative9 group. We assume that the measure structure and group
structure are compatible in the following senses. First, we assume that the group
operations (x, y) 7 xy and x 7 x1 are measurable, and that the translations
x 7 xy, x 7 yx and reflection x 7 x1 are all measure preserving10. We can then
define (formally, at least), the convolution f g of two functions f, g : X C by
Z Z
f g(x) := f (y)g(y 1 x) d(y) = f (xy 1 )g(y) d(y).
X X

This operation is bilinear and (formally) associative (the association being justified
from the Fubini-Tonelli theorem when f, g are either both non-negative or both
absolutely integrable), but is only commutative when the underlying group X is
also. The convolution algebra is a continuous version of the group algebra CX of
X, which corresponds to the case when is counting measure and all functions are
restricted to have finite support.

From Proposition 6.2 and the hypotheses on we quickly obtain

Theorem 6.4 (Youngs inequality). Let X be as above, and let 1 p, q, r be
such that p1 + r1 = q1 + 1. Then for any f Lp (X) and g Lr (X), the convolution
f g is absolutely convergent for almost every x and
kf gkLq (X) kf kLp(X) kgkLr (X) .
Remark 6.5. This looks very similar to Holders inequality, but note the +1 in
the scaling condition p1 + r1 = 1q + 1 (which, by the way, can be easily seen to

9We use multiplication whenever we do not wish to assume that the group is commutative. Of
course, many important examples, such as Euclidean space Rd , are additive (and thus commuta-
tive) groups.
10This is for instance the case if X is a Lie group, with a bi-invariant Haar measure. Such
measures exist whenever X is unimodular, which in particular occurs when X is compact or
abelian. Another example is if is counting measure and X is at most countable (for -finiteness).
 LECTURE NOTES 2 17

be necessary by the symmetry of multiplying the measure by a constant). This

shows that convolutions raise the integrability exponent, in contrast to products
which lower it. (This will be consistent with the Hausdorff-Young inequality in the
next section.)
Remark 6.6. For some groups, Youngs inequality is sharp. For instance, if X
contains a measurable subgroup H of positive finite measure, then 1H 1H =
(H)1H , and one verifies that Youngs inequality is sharp for all p, q, r with this
example. However, for other groups such as Rd , which do not contain subgroups of
positive finite measure, one can lower the constant below 1. We shall establish this
in later notes. On the other hand, when f, g are non-negative, Youngs inequality
at L1 is an equality:
kf gk1 = kf k1 kgk1 .
This follows easily from the Fubini-Tonelli theorem.

Youngs inequality does not directly imply the Hardy-Littlewood-Sobolev inequal-

ity (the problem being that |x|1ds lies in weak Ld/(ds) but not strong Ld/(ds) ).
However it does imply the following variant.
Corollary 6.7 (Non-endpoint Hardy-Littlewood-Sobolev). Let d 1, s < d, and
1 p q be such that
d d
< s.
q p
Then we have
Z
f (x)
k ds
dxkLqy (Rd ) .d,s,p,q kf kLpx(Rd )
Rd hx yi

for all f Lpx (Rd ). Here hxi := (1 + |x|2 )1/2 is the Japanese bracket of x.

One corollary of Youngs inequality is the norm convergence of approximations to

the identity.
Theorem 6.8 (Norm R convergence of approximations to the identity). Let
L1 (Rd ) be such that Rd (x) dx = 1, and for each > 0 let (x) := d ( x ).
Then for any 1 < p and f Lp (Rd ), the functions f converge to f in the
Lp (Rd ) topology.

Proof The key observation here is the quantitative estimate

kf kLp (Rd ) kf kLp(Rd ) k kL1 (Rd ) = kf kLp(Rd ) kkL1 (Rd ) .
This quantitative bound lets us make a number of qualitative reductions. First we
may restrict to a dense subclass of L1 (Rd ), such as the test functions C0 (Rd ).
Indeed the general case then follows from a limiting argument taking advantage of
the linearity of f in , together with the above bound. A similar argument lets
us also restrict f to be in C0 (Rd ). But then one easily verifies from the identity
Z Z
f (x) = f (x y)(y) dy = f (x) + (f (x y) f (x))(y) dy
Rd Rd
18 TERENCE TAO

that f converges uniformly to f . Since f also vanishes outside a fixed

compact set for sufficiently small, f also converges to f in Lp (Rd ) norm, and
we are done.

Problem 6.9. Show that the theorem fails for p = , but is true when L (Rd ) is
replaced by the closed subspace C 0 (Rd ) of bounded continuous functions.

Problem 6.10. If 1 p , f Lp (Rd ), and g Lp (Rd ), show that f g is
continuous and decays to zero at infinity.

7. Hausdorff-Young

We will now very quickly introduce the Fourier transform on Rd in order to demon-
strate a classical application of interpolation theory, namely the Hausdorff-Young
inequality. In this section we shall present this Fourier transform in a rather un-
motivated way; much later in this course we shall systematically study the Fourier
transform on various groups in a more unified context.

For any f L1x (Rd ), we define the Fourier transform F f = f : Rd C by

Z

F f () = f () := f (x)e2ix dx.
Rd

Note that while f is a function of position x Rd , it is useful to think of f as a

function of frequency Rd , which should be kept separate from position11 It
is clear that the Fourier transform F is a bounded linear operator from L1x (Rd ) to
L d
x (R ). The Fourier transform enjoys the easily verified symmetries

F Transx0 = Modx0 F
F Mod0 = Trans0 F

F Dilp = Dilp1 F
where Transx0 for x0 Rd is the spatial translation operator
Transx0 f (x) := f (x x0 ),
d
Mod0 for 0 R is the frequency modulation operator
Modx0 f (x) := e2ix0 f (x),
and Dilp for > 0 and 1 p is the Lp -normalised dilation operator
1 x
Dilp f (x) := f ( ).
d/p
11To be truly finicky, the frequency variable should live in the dual space (Rd ) of Rd ; this
is of course canonically identifiable with Rd once one imposes the Euclidean inner product x
on Rd , but one could also work on a more abstract finite-dimensional vector space without a
preferred inner product, in which case no canonical identification is available. It is occasionally
useful to work in the latter abstract setting, in order to more easily exploit the GL(Rd ) symmetry,
but in many applications the Euclidean structure is already being exploited (e.g. through the use
of balls, or of the magnitude function x 7 |x|), and so there is little point in trying to remove
that structure from the Fourier transform.
 LECTURE NOTES 2 19

Thus the Fourier transform interchanges translation with modulation, and reverses
dilation. There is in fact a more general symmetry

F DilpU = Dilp(U )1 F
for any invertible linear transformation U : Rd Rd (i.e. U GL(Rd )), where
1
DilpU f (x) := f (U 1 x).
| det U |1/p
Thus, for instance, the Fourier transform commutes with orthogonal transforma-
tions such as rotations and reflections. Finally, we observe the commutation rela-
tions

F = 2ij F
xj

F 2ixj = F
j
which can be viewed as infinitesimal versions of the translation and modulation
symmetries. One consequence of these symmetries is the observation that F maps
Schwartz functions to Schwartz functions, continuously in the Schwartz topology.

The adjoint Fourier transform F is clearly given by

Z

F g(x) = g(x) := g()e2ix dx
Rd
and is well defined for g L1 (Rd ) at least. Being the conjugate of the Fourier
transform as well as the adjoint, it obeys essentially the same symmetries:
F Transx0 = Modx0 F
F Mod0 = Trans0 F

F Dilp = Dilp1 F

F = 2ixj F
j

F 2ij = F .
xj
In particular the compositions F F and F F , which maps Schwartz functions with
Schwartz functions continuously in the Schwartz topology, commute with transla-
tions, modulations, dilations, multiplication by linear functions, and derivatives.
2
One also easily computes that the standard Gaussian e|x| is left unaffected by
F and F , and hence by F F and F F . One then easily deduces that F F and
2
F F leaves invariant all polynomial multiples of e|x| , which happen to be dense
in the Schwartz space (essentially by the Weierstrass approximation theorem). Thus
we obtain the Fourier inversion formula
F F f = f ; F F g = g
for all Schwartz f and g. In other words, F is formally unitary. As a corollary we
obtain the Parseval identity 12
hf, gi = hF F f, gi = hF f, F gi
12Also known as the Plancherel identity.
20 TERENCE TAO

for all Schwartz f, g. Specialising to the diagonal case f = g we obtain the

Plancherel identity
kF f kL2(Rd ) = kf kL2x(Rd ) ,
which allows us by density to extend the Fourier transform uniquely to a continuous
linear unitary transformation on L2x (Rd ). The transform F now maps L1x (Rd ) to
13
L d 2 d 2 d
x (R ) and Lx (R ) to L (R ), both with norm 1. Riesz-Thorin interpolation
then yields the Hausdorff-Young inequality
kF f kLp (Rd ) kf kLpx (Rd ) (7)

for all 1 < p < 2 and simple functions f with compact support, and (thus by
density) we may also extend F uniquely and continuously to Lp (Rd ). (For p > 2,
the Fourier transform of an Lp (Rd ) function can still be defined as a distribution,
but it need not correspond to a locally integrable function.)
Problem 7.1. Show that up to constant multiplication, the Fourier transform F is
the only continuous map from Schwartz space to itself which obeys the translation
and modulation symmetries.
Problem 7.2. Obtain the Hausdorff-Young inequality using real interpolation and
the tensor power trick.

The Hausdorff-Young inequality (7) inequality gives an upper bound of 1 for the
operator norm kF kLp (Rd )Lp (Rd ) . In the converse direction, we have:
x

Problem 7.3. Show that if f is a modulated gaussian function, thus f (x) :=

ce2ix0 eA(xx0 )(xx0 ) for some positive definite operator A : Rd Rd , some
c > 0, and x0 , 0 Rd , then for all 1 p we have
p1/2p
kF f kLp (Rd ) = kf kLpx(Rd ) .
(p )1/2p
(Hint: use symmetries to reduce to the case of the standard gaussian f (x) =
2
e|x| .)

1/2p
This gives a lower bound of (pp )1/2p for the operator norm. This bound turns out
to be sharp (a famous result of Beckner) but we will not demonstrate it here.

One can ask whether the Fourier transform has any other Lp mapping properties,
i.e. to determine all the (p, q) for which F is strong type (or restricted weak
type, etc.). The scale invariance of the Fourier transform shows that the duality
condition q = p is necessary for any of these type properties to hold. The Hausdorff-
Young inequality also shows that this necessary condition is sufficient (for any of
the types) for p 2. Unfortunately the inequality fails for p > 2. In fact the
failure is quite severe and can be demonstrated in a number of ways. One is by
Littlewoods principle, which asserts that on non-compact domains (such as Rd ),
it is not possible for an operator with any sort of translation symmetry to map
high-exponent Lebesgue spaces to low-exponent spaces. (The higher exponents
are always on the left.) We give one rigorous formulation of this principle in Q12.
13This is one reason why we normalise the Fourier transform by placing the 2 in the exponent.
 LECTURE NOTES 2 21

That principle does not directly apply here, but a variant of it will suffice here. Let
N be a large integer, let v Rd have large magnitude, and consider the function
N
X
f (x) := e2ixnv g(x nv)
n=1
2
where g is the standard Gaussian g(x) := e|x| . Then the symmetries of the
Fourier transform show that
N
X
f() := e2inv g( nv) = f ().
n=1

For any 1 p , it is not hard to show that

kf kLp(Rd ) p N 1/p
if N and |v| are sufficiently large depending on p. (Hint: establish the upper bound
at p = 1 and p = , which gives the upper bound for all intermediate p also by
log-convexity. The lower bound can be established by considering the value of f
near integer multiples of v, or else by computing the L2 norm by hand and using
log-convexity and the upper bounds to then obtain lower bounds elsewhere.) In
particular
1
1
kF f kLp (Rd ) p N p p kf kLp(Rd ) .
If p > 2, then the exponent of N here is positive; since N can be made arbitrarily
large, we see that F cannot be of strong type (p, p ) when p > 2.
Problem 7.4. Show also that F cannot be of restricted weak type (p, p ) for p > 2.
Remark 7.5. The Hausdorff-Young inequality maps spaces at L2 or below to spaces
at L2 or above. It is very difficult to map back to the original space (in part due to
the Littlewood principle) so in practice the Hausdorff-Young inequality is not very
useful for the type (p, p) of operators connected to the Fourier transform (such as
Fourier multipliers), except of course in the important case p = 2. The Hausdorff-
Young inequality is also a prototype of a Fourier restriction estimate (also known
as a restriction theorem), of which more will be said later.

8. Kernel truncation

We continue our study of integral operators (4). Intuitively, the smaller one makes
the kernel K, the easier it should be to bound it. One basic result in this direction
is:
Theorem 8.1 (Positive domination). Let K : X Y C, K : X Y R+ be
kernels, and let TK , TK be the corresponding integral operators. Suppose that we
have the pointwise bound |K| K . Then if TK is of strong type (p, q), then TK
is also (and the integral is a.e. absolutely convergent); in fact we have the bound
kTK kLp (X)Lq (Y ) kTK kLp (X)Lq (Y ) .

Proof From the triangle inequality we have the pointwise bound |TK (f )|
TK (|f |). The claim follows.
22 TERENCE TAO

Note that similar results also hold for weak type, restricted type, etc. because of
the monotonicity property of the underlying function spaces.

In particular, when K is non-negative, and is any subset of X Y , we see that

TK1 has a smaller (Lp , Lq ) operator norm than TK ; thus truncation always reduces
the operator norm. However, this property is not quite true for signed kernels, the
problem being that the boundedness of T might arise from delicate cancellations
within the kernel which might be destroyed by truncation.
Problem 8.2. Let X = Y = R, let K be the Fourier kernel K(x, y) := e2ixy ,
and let R R be those pairs (x, y) for which e2ixy has positive real part.
Show that even though TK is of strong type (2, 2), the operator TK1 is not. (Hint:
compute a lower bound for the real part of hTK1 1I , 1I i for a large interval I.)

On the other hand, certain truncations are acceptable for signed kernels. For in-
stance, if p = 1 or q = then all truncations contract the operator norm, thanks
to Propositions 5.2, 5.4. Secondly, it is clear that if is a product set = A B,
then from the14 identity
TK1 (f ) = 1B TK (1A f ) (8)
we see that if TK is of strong-type (p, q) then TK1 is also, with the expected
comparison principle kTK1 kLp(X)Lq (Y ) kTK kLp (X)Lq (Y ) . Also note in this
case that TK1 can also be viewed as an operator from Lp (A) to Lp (B), and
kTK1 kLp (X)Lq (Y ) = kTK1 kLp (A)Lq (Y ) = kTK1 kLp (A)Lq (B) .

Let us write the identity (8) in operator form as

TK1 = 1B TK 1A
where we now think of 1A , 1B not as functions, but rather as the multiplier operators
f 7 1A f , f 7 1B f . Motivated by this identity, let us now study other ways in
which an operator T can be multiplied on the left and right by indicator functions.
Problem 8.3 (Block diagonal operators). Suppose Q is at most countable, and we
have disjoint subsets (Xn )nQ , (Yn )nQ of X and Y respectively. For each n Q,
let Tn : Lp (Xn ) Lq (Yn ) be a bounded linear operator for some 1 p, q ,
and define the block diagonal15
X
T := 1Yn Tn 1Xn .
nQ

Here we identify functions on Xn (resp. Yn ) with functions on X (resp. Y ) by

extending by zero.

14Here we are implicitly assuming that there is enough regularity in K and f that the expres-
sions are well-defined; we gloss over this subtlety here.
15Strictly speaking, block diagonal refers to the case when X = Y and X = Y . A more
n n
appropriate term here would be block permutation, but we prefer block diagonal as it is more
familiar (and covers the most important cases).
 LECTURE NOTES 2 23

If q p, show that
kT kLp(X)Lq (Y ) = sup kTn kLp (Xn )Lq (Yn ) .
n

If q < p, show that

kT kLp(X)Lq (Y ) = kkTnkLp (Xn )Lq (Yn ) klrn
1 1
where r is such that p + r = 1q .

Thus the behaviour of the block-diagonal operator T is determined entirely by the

behaviour of the components Tn . In particular, if q p, T is of strong type (p, q)
if and only if the Tn are uniformly of strong type (p, q).
Corollary 8.4 (Truncation to a block diagonal). Let Q, (Xn )nQ , (Yn )nQ be as
above, and let T : Lp (X) Lq (Y ) be a bounded linear operator for some 1 p
q . Let T be the block-diagonal restriction of T ,
X
T := 1Yn T 1Xn .
nQ

Then T is also of type (p, q) and kTkLp (X)Lq (Y ) kT kLp(X)Lq (Y ) .

Proof Apply the above problem with Tn := 1Yn T 1Xn .

Thus we see that we may freely

S restrict the kernel of an integral operator TK to
a block-diagonal set such as n Xn Yn without increasing the operator norm, as
long as we respect Littlewoods principle (keeping the higher exponent on the left).

A similar argument works for almost block diagonal operators16:

Problem 8.5 (Almost block diagonal operators). Let (Xn )nQ , (Ym )mQ be disjoint
subsets of X and Y respectively, where Q, Q are at most countable. Let be a
relation between Q and Q which has bounded degree in the sense that for each
n Q there are at most O(1) values of m Q for which n m, and conversely
for each m Q there are at most O(1) values of n Q for which n m. Let
1 p q , and for each (n, m) with n m, let Tn,m : Lp (Xn ) Lq (Ym ) be
a bounded linear operator, and let T be the sum
X
T := 1Ym Tn,m 1Xn .
(n,m)QQ :nm

(Note that the bounded degree hypothesis ensures that this sum is well-defined.)
Then for any A > 0 the following are equivalent up to changes in the implied
constants.

kT kLp(X)Lq (Y ) .p,q A.
kTn,m kLp (Xn )Lq (Ym ) .p,q A for all n, m with n m.

16Again, a better term here would be almost block permutation operators.

24 TERENCE TAO

Remark 8.6. The point about block-diagonal or almost block-diagonal operators

are that they are local ; the influence of the subdomain Xn on the operator T is
limited to a bounded number of subranges Ym . Thus we see that for local operators,
global boundedness and local boundedness are essentially equivalent. One can now
obtain an analogue of Corollary
S 8.4, namely that one can restrict a kernel to an
almost block diagonal region nm Xn Ym without increasing the (Lp (X), Lq (Y ))
norm by more than a multiplicative constant; we leave the details to the reader.

Having considered truncations to diagonal regions, let us now consider trunca-

tions to upper diagonal regions. Here it turns out we need to strengthen the
Littlewood condition p q to a strict inequality p < q to obtain good results
(otherwise we lose a logarithm, see Q14).
Theorem 8.7 (Christ-Kiselev lemma, finite version). Let N 1, and suppose
SN SN
we have partitions X = n=1 Xn , Y = n=1 Yn . Let 1 p < q , and let
T : Lp (X) Lq (Y ) be a bounded linear operator. Let T be the modified operator
X
T := 1Ym T 1Xn .
1nmN

Then T is also bounded from L (X) to Lq (Y ), and we have

p

kT kLp (X)Lq (Y ) .p,q kT kLp(X)Lq (Y ) .

Proof We will prove this by induction on N ; however, one must be careful because
of the use of the Vinogradov notation .p,q in the conclusion. To do things rigor-
ously, let Ap,q be a large constant depending only on p and q to be chosen later.
We claim that for every N , we have the bound
kT kLp (X)Lq (Y ) Ap,q kT kLp(X)Lq (Y ) .
The claim is trivial for N = 1 (if Ap,q is large enough) so suppose inductively that
N > 1 and the claim has already been proven for smaller values of N . We may
normalise kT kLp(X)Lq (Y ) = 1. Let us also choose f Lp (X), normalised so that
kf kLp(X) = 1. Our task is to show that
kT f kLq (Y ) Ap,q .
The idea is to divide-and-conquer f in an intelligent fashion, adapted to both f and
the blocks X1 , . . . , Xn . Consider the sequence of numbers kf 1X1 ...Xn kpLp (X) for
n = 0, 1, . . . , N . This sequence increases from 0 to 1, so we can find 0 n0 < N
such that
1
kf 1X1 ...Xn0 kpLp (X) < kf 1X1 ...Xn0 +1 kpLp (X)
2
and hence
1
kf 1X1 ...Xn0 kLp (X) , kf 1Xn0 +2 ...XN kLp (X) 1/p .
2
Applying the induction hypothesis, we conclude that
1
kT (f 1X1 ...Xn0 )kLq (Y1 ...Yn0 ) , kT (f 1Xn0 +2 ...XN )kLq (Yn0 +2 ...YN ) 1/p Ap,q .
2
On the other hand, we have
kT (f 1X1 ...Xn0 )kLq (Yn0 +1 ...YN ) = kT (f 1X1 ...Xn0 )kLq (Yn0 +1 ...YN ) = O(1)
 LECTURE NOTES 2 25

and
kT (f 1Xn0 +1 )kLq (Y ) kT (f 1Xn0+1 )kLq (Y ) = O(1)
by the normalisations on T and f and the definition of T . Also, T (f 1Xn0 +2 ...XN )
vanishes outside of Yn0 +2 . . . YN . From the triangle inequality we conclude that
1
kT f kLq (Y1 ...Yn0 ) , kT f kLq (Yn0 +1 ...YN ) 1/p Ap,q + O(1)
2
and thus
21/q
kT f kLq (Y ) 1/p Ap,q + O(1).
2
If we choose Ap,q large enough depending on p, q, then we can make the right-hand
side less than Ap,q (here we use the hypothesis q > p). The claim follows.

In terms of kernels,
S the above lemma allows one to truncate the kernel to upper-
diagonal regions nm Xn Ym without losing too much in the operator norm, so
long as q > p. The claim is unfortunately false when p = q (except when p = q = 1
or p = q = , where the claim is easy to verify), although producing an example
will have to wait for our analysis of the Hilbert transform in next weeks notes.
S S
The fact that the partitions X = n Xn , Y = n Yn are arbitrary makes the
Christ-Kiselev lemma quite powerful. Indeed, we can conclude
Corollary 8.8 (Christ-Kiselev lemma, maximal version). Let Q be a countable
ordered set, and for each Q let E be a set in X with the nesting property
E E whenever < . Let 1 p < q , and let T : Lp (X) Lq (Y ) be a
bounded linear operator. Define the maximal operator
T f (y) := sup |T (f 1E )(y)|. (9)
Q

Then the sublinear operator T is also bounded from Lp (X) to Lq (Y ), and in fact
kT kLp (X)Lq (Y ) .p,q kT kLp(X)Lq (Y ) .

Proof We may normalise kT kLp(X)Lq (Y ) = kf kLp(X) = 1. By monotone con-

vergence we may assume that Q is finite, and by relabeling we can then take
Q = {1, . . . , N }. By adding dummy indices if necessary we may assume EN = X,
and we also adopt the convention E0 = . Then for each y Y there is an
1 (y) N which attains the supremum in (9):
T f (y) = |T (f 1E(y) )(y)|,
and so we need to show that
kT (f 1E(y) )kLqy (Y ) .p,q 1.
(This trick of replacing a maximal operator T by a linear operator f 7 T (f 1E(y) ),
albeit an operator which depends on an arbitrary function , is known as linearisa-
tion.) Now for each 1 n N , let Xn := En \En1 and Yn := {y Y : (y) = n}.
Then we observe the telescoping identity
X
T (f 1E(y) ) = 1Ym T (1Xn f ).
1nmN

The claim now follows from Lemma 8.7.

26 TERENCE TAO

Let us give a sample application of this lemma. Given any locally integrable function
f : R C, define the maximal Fourier transform F f by
Z
F f () := sup | f (x)e2ix dx| = sup fd1I ()
I I I
where I ranges over compact intervals in R.
Corollary 8.9 (Menshov-Paley-Zygmund theorem, quantitative version). For any
1 p < 2 we have
kF kLp (R)Lp (R) .p 1.

Proof By the triangle inequality we may restrict I to be an interval with one

endpoint 0; by symmetry we may take I to have the form [0, y] for some positive
y. By monotone convergence we may take y to be rational. Thus we wish to show
that
k sup f\1[0,y] kLp (R) .p kf kLp(R) .
yQ+
But this follows from the Hausdorff-Young inequality and Corollary 8.8.

Corollary 8.10 (Menshov-Paley-Zygmund theorem, qualitative version). R If 1

p < 2 and f Lp (R), then for almost every R the Fourier integral R f (x)e2ix dx
is conditionally convergent to f(), thus
Z N+
lim f (x)e2ix dx = f()
N+ ,N + N
for almost every .

Proof We can rewrite the claim as

Z N+
k lim sup | f (x)e2ix dx f()|kLp (R) = 0
N+ ,N +
N

The claim is trivial if f lies in L1 (R)Lp (R), which is a dense subset of Lp (R). The
claim then follows from a limiting argument and Corollary 8.9 (and the Hausdorff-
Young inequality).

Remark 8.11. The above results are also true for p = 2, but is much more difficult
to prove, and is known as Carlesons theorem.

One can pass from the finite version of the Christ-Kiselev lemma to an infinite
version if one has enough regularity on the kernel. For instance, we have
Problem 8.12 (Christ-Kiselev lemma, upper diagonal version). Let K : R R C
be a locally integrable kernel, and suppose that the operator TK (which is defined
and locally integrable for compactly supported continuous functions f Cc0 (R) at
least) is bounded from Lp (R) to Lq (R) for some 1 p < q (i.e. it is bounded
on the dense subclass Cc0 (R) and thus has a unique continuous extension). Let K>
be the restriction of K to the upper diagonal {(s, t) R R : t > s}. Then the
operator TK> (which is also defined on Cc0 (R)) is also of strong type (p, q), with
kTK> kLp (R)Lq (R) .p,q kTK kLp (R)Lq (R) .
 LECTURE NOTES 2 27

Remark 8.13. A vector-valued version of this lemma, in which the input and
output functions are not complex values, but instead take values in some infinite-
dimensional vector space, is very useful in the study of nonlinear dispersive PDE,
and in particular in the theory of Strichartz estimates, but we will not discuss that
topic further here.

In the above results we have only discussed rough truncations, in which a ker-
nel K was multiplied by an indicator function 1 . One can do somewhat better
with smooth truncations, but this requires significantly more structure on the
underlying domain.
Definition 8.14 (Adapted bump function). Let Rd be a bounded open set,
and let L : Rd Rd be an invertible affine transformation (the composition of
a general linear transformation and a translation), thus L() is another bounded
open set in Rd . We say that a function : Rd C is a bump function adapted to
L() if is smooth, supported in L(), and we have the bounds
k ( L)(x) = Ok, (1)
for all k 0 and x Rd (actually one just needs x ).
Remark 8.15. This definition only becomes meaningful if we hold fixed and let L
vary. Typically is a standard set, such as the unit ball, unit cube, unit cylinder,
or unit annulus, so that L() describes balls, cubes, boxes, tubes, or annuli of
various radius, eccentricity, and location; the point is then that the bump functions
adapted to these sets are adapted uniformly in these additional parameters.

Proposition 8.16 (Smooth cutoffs). Let K : Rd Rd C be a locally integrable

function, and suppose that TK is bounded from Lp (Rd ) to Lq (Rd ) for some 1

p, q . Let L() Rd and L ( ) Rd be affine images of bounded open

sets Rd , Rd , and let : Rd Rd C be a bump function adapted to

L() L ( ). Then TK is bounded from Lp (Rd ) to Lq (Rd ), and
kTK kLp (Rd )Lq (Rd ) ., ,d,d kTK kLp (Rd )Lq (Rd )
where we define TK first for test functions Cc0 (Rd ) and then extend by density.

Proof The idea is to exploit the fact that while bump functions adapted to product
sets such as L() L ( ) are not necessarily tensor products of bump functions
adapted to L() and L ( ) separately, they can be efficiently decomposed as a
convergent sum of such.

By rescaling by L and L we may assume that these transformations are the iden-
tity. By rescaling a little more to shrink and we may assume that these sets

are supported in (say) the cubes [1/4, 1/4]d and [1/4, 1/4]d respectively. In par-

ticular, is now a smooth bump function supported on [1/4, 1/4]d [1/4, 1/4]d .

We can identify this set with a subset of the torus (R/Z)d (R/Z)d in the usual
manner, thus creating a function : (R/Z)d (R/Z)d C whose k th derivatives

are Ok,, (1). We now appeal to the theory of Fourier series to write
X X b
y) =
(x, m)e2inx e2imy
(n, (10)
nZd mZd
28 TERENCE TAO

b
where (n, m) is the Fourier coefficient
Z
b m) =
(n, y)e2inx e2imy dxdy.
(x,
(R/Z)d (R/Z)d

combined with a standard integration by parts argument,

The smoothness of ,
b
m) decays rapidly, thus for any k we have
shows that (n,
b m) = O k
(n, k,, ((1 + |n| + |m|) ).
In particular (taking k larger than d + d ) we obtain that these coefficients are
summable: X X b
m)| .d,d ,, 1.
|(n,
nZd mZd
Thus the summation in (10) is in fact absolutely convergent. This leads to the
kernel identity
X X
(x, y)K(x, y) = m)e2inx 1[1/4,1/4]d (x)e2imy 1
(n, [1/4,1/4]d (y)K(x, y)
nZd mZd

which in turn leads to the operator identity

X X
TK = m)1
(n, [1/4,1/4]d em TK 1[1/4,1/4]d en
nZd mZd

(at least when applied to test functions), where em is the multiplier operator
f (y) 7 f (y)e2imy , and similarly for en . Since the operators 1[1/4,1/4]d , en are
contractions on Lp (Rd ), and 1[1/4,1/4]d , em are contractions on Lq (Rd ), the claim
now follows from the triangle inequality

One of the morals of the above proposition is that smooth bump functions, when
applied to an integral kernel, have essentially a negligible impact on the behaviour
of the integral operator, other than to localise that operator to the support of that
bump function. In particular the precise choice of bump function used is almost
always irrelevant. The boundedness of an operator is instead controlled by those
features unaffected by multiplying by smooth bump functions, such as singularity
or rapid oscillation.

9. Exercises

Q1 (Weighted Schurs test) Suppose that K : X Y C obeys the bounds

Z
|K(x, y)|wX (x) dX (x) AwY (y) for almost every y Y
X
and
Z
|K(x, y)|wY (x) dY (y) BwX (x) for almost every x X
Y
for some 0 < A, B < and some strictly positive weight functions wX :
X R+ , wY R+ . Show that the integral operator (4) is bounded from
L2 (X) to L2 (Y ) with an operator norm bound kT kL2(X)L2 (Y ) AB.
 LECTURE NOTES 2 29

Q2 (Hilbert-Schmidt norm) Suppose that K : X Y C is square-

integrable. Show that the integral operator (4) is bounded from L2 (X)
to L2 (Y ) with operator norm bound
kT kL2(X)L2 (Y ) kKkL2(XY ) .
Show that equality occurs if and only if T is rank one, thus K(x, y) =
f (x)g(y) for some f L2 (X), g = L2 (Y ). The quantity kKkL2(XY ) is
known as the Hilbert-Schmidt norm or Frobenius norm of T . From the
above analysis of equality, we only expect the above estimate to be efficient
when T is approximately low rank (i.e. it has some good compactness
properties; more informally, this estimate is only efficient when T is only
sensitive to a few degrees of freedom of the input function).
Q3. Let X be a measure space which is non-atomic in the sense that
given any set E X of positive measure, there exists a subset E with
0 < (E ) < (E). (The real line R is a typical example.)
(a) Show that if E has positive measure and 0 < < (E), then there
exists a subset E of E with (E ) = . (Feel free to use Zorns lemma,
though it is possible to establish this result using only the axiom of count-
able choice instead. You may wish to first establish the weaker result that
there exists a subset E with 0 < (E ) .)
(b) Suppose that T is a linear operator mapping simple functions on X to
functions on Y which is of restricted weak-type (p, q) for some 1 q <
and 0 < p < 1. Show that T = 0. Indeed, it is very rare to see any
non-trivial boundedness statement originating from below the L1 norm,
especially given that such functions are usually not even locally integrable.
Q4. Show by means of examples (or by some sort of dimensional analysis
or scaling argument) that Corollary 6.3 fails when p = 1, when q = , or
when the condition pd = dq s is violated.
Q5. (Hardy inequality, I.) Let R+ be given Lebesgue measure dx. Establish
the one-dimensional Hardy inequality
Z
1 x
k f (t) dtkLpx (R+ ) .p kf kLp(R+ ) .
x 0
for all 1 < p and f Lp (R+ ), and show that the inequality fails at
p = 1. (Hint: first establish restricted weak type for 1 < p < .)
Q6. (Hardy inequality, II.) Show that
Z
1 x
k f (t) dtkL2x (R+ ) 2kf kL2(R+ )
x 0
for all f L2x (R+ ), and that the constant 2 cannot be replaced by any
smaller constant. (Hint: there are many ways to prove this; one is by Q1
with some perturbation of the weight |x|1/2 .)
Q7. (Hardy inequality, III.) Let 1 < p < . Show that
Z
1 x p
k f (t) dtkLpx (R+ ) kf kLp(R+ )
x 0 p1
p
for all f Lpx (R+ ), and that the constant p1 cannot be replaced by any
smaller constant. (You may need to develop an Lp version of Q1.)
30 TERENCE TAO

Q8. Let d 1, 1 p, q be such that p1 + q1 1, and let f Lp (Rd )

and g Lq (Rd ). Let , R obey the scaling condition
d d
+ =
.
p q
If > pn (thus < qn ), show that
Z Z
|x| |y| |f (x)||g(y)| dxdy .,,p,q kf kLp(Rd ) kgkLq (Rd ) .
|x|.|y|

If < pn (thus > qn ), show that

Z Z
|x| |y| |f (x)||g(y)| dxdy .,,p,q kf kLp(Rd ) kgkLq (Rd ) .
|y|.|x|

Q9. (Weighted one-dimensional Hardy-Littlewood-Sobolev) Let 1 p, q

, 0 < s < 1, and , R obey the conditions
1
>
p
1
>
q
1 1
1 + 1+s
p q
and the scaling condition
n n
+ 1+s = ,
p q
with at most one of the equalities
1 1
p = 1, p = , q = 1, q = , + = 1 + s
p q
holding. Let f Lp (R) and g Lq (R). Show that
Z Z
|x| |y|
1s
|f (x)||g(y)| dxdy .,,p,q kf kLp(R) kgkLq (R) .
R R |x y|
Q10. Let X be a measure space with (X) = 1. Let T be a linear operator
mapping simple functions on X to some normed vector space W . Suppose
we have the bound
1
kT 1E kW . A|E|(1 + log )
|E|
for all sets E X and some A > 0. Conclude that
kT f kW . Akf kL log L
for all simple f . (Hint: first do this for quasi-step simple functions, and
then perform a decomposition similar to that in the previous weeks notes.)
Thus in some sense L log L and restricted L log L are equivalent.
Q11 (Yano extrapolation theorem). Let X be a measure space with (X) =
1. Let T be a linear operator defined on simple functions which is of strong
type (p, p) for all 1 < p 2. Furthermore, assume that
1
kT f kLp(X) . kf kLp(X)
p1
 LECTURE NOTES 2 31

for all 1 < p 2 and all simple functions f . Show that we also have
kT f kL1(X) . kf kL log L(X)
for all simple functions f . (Hint: use Q10.)
Q12 (Littlewoods principle). Let 1 q < p , and suppose that
T : Lp (Rd ) Lq (Rd ) is a bounded linear operator which commutes with
translations, thus T Transx0 = Transx0 T for all x0 R. Show that T is
identically zero. (Hint: if T is not identically zero, then there is a non-
zero f Lp (Rd ) such that T f is also not identically zero. Use monotone
convergence to make f and T f small in a suitable sense outside of a large
PN
ball. Now investigate the Lp (Rd ) norm of n=1 Transxn f , where the xn
are widely separated points in space, as well as the Lq norm of the image of
that function under T .) Give a counterexample to show that Littlewoods
principle fails if Rd is replaced with a compact domain such as the torus
Rd /Zd .

Q13. Let BX be a -finite sub--algebra of BX , and similarly let BY be
a -finite sub--algebra of BY . Show that if K : X Y C is bounded

and measurable with respect to BX BY , and TK extends to a continuous
p q
linear operator from L (BX ) to L (BY ), then we have the
kTK kLp (BX )Lq (BY ) = kTK kLp (BX
)Lq (B ) = kTK kLp (B )Lq (B )
Y X Y

for all 1 p, q . Thus to establish strong type (p, q) bounds on TK

on the original -algebra BX , it suffices to do so on the sub-algebra BX .
The qualitative hypotheses that K is bounded and TK is continuous can
be relaxed, but we shall not do so here.
Q14. Suppose that T is a continuous linear operator from Lp (X) to Lp (Y )
for some 1 p . Let E1 . . . EN be a nested sequence of sets in X
for some N > 1, and let T be the maximal operator
T f (y) := sup |T (f 1En )|.
1nN

Show that
kT kLp (X)Lp (Y ) . log N kTK kLp (X)Lp (Y ) .
Q15 (Radamacher-Menshov inequality). Let f1 , . . . , fN be an orthonormal
set of functions in L2 (X) for some N > 1. Show that
Xn
k sup | fm |kL2 (X) . N 1/2 log N.
1nN m=1

(Hint: use Q14 somehow.)

Pn Comparing this with what Pythagorass theorem
gives for sup1nN k m=1 fm kL2 (X) , we see that putting the sup inside
only costs at most log N .

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

E-mail address: [email protected]