Geometric maximal functions, oscillatory

integrals, and non-linear evolution equations

Terence Tao (UCLA)

www.math.ucla.edu/∼tao/preprints
Abstract:
We discuss the interconnections between the study of
geometric combinatorial problems (such as the Kakeya
problem), the study of oscillatory integrals (including
the restriction and Bochner-Riesz conjectures), and the
study of non-linear dispersive equations. These
interconnections have only been systematically exploited
in the last ten years, and much progress on these
problems has already been obtained as a consequence.

1
The Kakeya conjecture

• In 1917 S. Kakeya posed the following problem: what

is the least amount of area necessary to rotate a unit
line segment (a “needle”) continuously by 2π in the
plane? This became known as the Kakeya needle
problem.
• Rotating around the center requires π/4 units of
area, while a more sophisticated “three-point U-turn”
can trim this to π/8.
• In 1925, Besicovitch showed that one can achieve
this rotation using arbitrarily small amounts of area.
There were two basic ideas in the proof:
• Firstly, one can translate a needle for arbitrarily small
area, by sliding very far away, doing a slight rotation,
sliding back, and undoing the rotation.
• Secondly, one can find a collection of rectangles whose
long side is 1, whose orientations cover all possible di-
rections, and whose union has arbitrarily small area
(the Besicovitch set construction).

2
• In general dimension Rn, define a Besicovitch set
to be a set which contains a unit line segment in
every direction. The construction of Besicovitch thus
shows that such sets can have measure zero when
n = 2.
• However, Besicovitch sets in R2 must still have Haus-
dorff and Minkowski dimension exactly 2 (Drury, 1971).
3
 The key observation is that two line segments with
different orientations can intersect in at most one
point.
• This problem can be discretized. For any 0 < δ ≪ 1,
define a δ-Besicovitch set to be a set which contains
a 1 × δ tube in every direction. Besicovitch’s con-
struction gives δ-Besicovitch sets in R2 with area
about log log(1/δ)/ log(1/δ), but Drury’s argument
shows that such sets can never have area less than
about 1/ log(1/δ).
• The Kakeya conjecture states that for any dimen-
sion n, Besicovitch sets must always have Hausdorff
and Minkowski dimension n. On the discretized level,
this means that δ-Besicovitch sets have volume at
least Cεδ ε for any ε > 0.
• In two dimensions, this conjecture is solved, but in
higher dimensions it is still open. Much recent progress
has been made by Bourgain, Wolff, and others. For
instance, when n ≥ 3 the best lower bound on the
Minkowski dimension is
n+2 4n + 3
max( + 10−10, )
2 7
(Katz-Laba-T 1999, Katz-T 1999, Laba-T 2000), while
4
 for the Hausdorff dimension the best bound is
n + 2 6n + 5
max( , )
2 11
(Wolff 1995, Katz-T 1999).
• There are two types of arguments used to prove these
bounds. One is geometric and makes use of inci-
dence geometry facts (e.g. if three lines intersect each
other at different points, they must lie in a plane, and
each plane contains only a one-parameter set of di-
rections). The other is arithmetic and makes use of
work on arithmetic progressions, sum-sets and dif-
ference sets, in particular the work of Heath-Brown,
Szemeredi, Gowers, and Bourgain. Some of the more
recent results have managed to combine the two argu-
ments, but we are still far from settling the problem.
• There are more quantitative versions of the Kakeya
conjecture which involve the Kakeya maximal func-
tion ∫
f ∗(ω) := sup f
l//ω l
where ω ranges over directions, and l ranges over
lines parallel to ω. We will not discuss this here.

5
A link with algebra?

• One can phrase finite field analogues of the Kakeya

problem: in a finite geometry F n, a Besicovitch set
is a set which contains a line in every direction. The
Kakeya conjecture then states that such sets have
cardinality at least cε|F |n−ε for any ε > 0.
• Progress on this problem and the continuous ana-
logue have mostly been equivalent. However, there
are some intriguing features in the finite field case.
For instance, if n = 3 and F contains a subfield G of
cardinality |G| = |F |1/2, then the Heisenberg group
{(x, y, z) ∈ F 3 : Im(z) = Im(xy)},
where y → y is the automorphism fixing G and
Im(z) := (z + z)/2, has cardinality |F |5/2 and is
almost a counterexample to the Kakeya conjecture
in that it contains a large number (|F |2) of distinct
lines (which however are not all parallel).
• It is not known if the analogous situation occurs in
the continuous case. A good test question is the
Erdös ring problem: does R contain a subring of
Hausdorff dimension exactly 1/2? Such a ring would
lead to some interesting near-counterexamples to Kakeya
6
 as well as other geometric measure theory problems
such as the Falconer distance problem and the Fursten-
burg set problem (Katz-T., 2000).
• I think these rings do not exist, and that the reason
has to do with the fact that R is an ordered field.
Unfortunately we know of no way of combining the
order theory of R effectively with the other arith-
metic and geometric tools at our disposal.
• A possible attack route comes from applying inequal-
ities from additive-multiplicative combinatorics, such
as the inequality
max(|A + A|, |A · A|) ≥ c|A|5/4
whenever A is a finite set of integers (Elekes, 1998).
However the arguments used to prove these types of
inequalities do not seem to adapt well to the contin-
uous setting.

7
Kakeya and Fourier summation
• The Kakeya problem is related to several problems in
harmonic analysis and PDE involving oscillatory in-
tegrals. An example arises from the theory of Fourier
summation in higher dimensions.
• If f is a test function on Rn, the Fourier transform
fˆ is defined by
∫
fˆ(ξ) = n
f (x)e−2πix·ξ dx.
R
One has the Fourier inversion formula
∫
f (x) = n
fˆ(ξ)e2πix·ξ dx.
R
• Now suppose that f is a much rougher function, for
instance it is only in the Lebesgue space Lp for some
1 < p < ∞. One can ask to what extent the above
formulae are still valid. A natural guess is that
∫
fˆ(ξ)e2πix·ξ dx → f
|ξ|<R
as R → ∞ in some sense.
• This convergence is true in the weak sense, but one
might ask for stronger convergence, such as Lp con-
vergence or pointwise convergence a.e. (Uniform con-
vergence is easily shown to fail).
8
• The problem can be reduced to the study of the disk
multiplier S 0, defined by
d
S 0 f (ξ) = χ
|ξ|<1 f (ξ).

It is easy to show that one has Lp convergence if and

only if S 0 is bounded on Lp. (a.e. convergence is
related to a maximal version of the disk multiplier
which won’t be discussed here).
• In one dimension, it is a classical result of Riesz that
S 0 is bounded on Lp for all 1 < p < ∞. However,
in 1971, C. Fefferman used Besicovitch sets to prove
the surprising
• Theorem. If n > 2 and p ̸= 2, then S 0 is un-
bounded on Lp.
• Roughly speaking, the idea for p > 2 is as follows.
Using a slight variant of the Besicovitch set construc-
tion, one can find a large number of N × N 2 rectan-
gles Ri with distinct orientations, such that the Ri
are disjoint but their shifts R̃i have large overlap:

9
• On each rectangle Ri, one constructs a “wave packet”
ψi such that S 0ψi is large on R̃i. One then applies the
disk multiplier S 0 to a suitable linear combination of
ψi. The overlap of the R̃i make the Lp norm large.
10
• The p < 2 argument is dual and works in reverse
(swap Ri and R̃i).
• Thus Fourier summation need not converge in Lp.
This can be rectified by introducing a weight in the
Fourier integral:
∫
|ξ|
(1 − )δ fˆ(ξ)e2πix·ξ dx → f
|ξ|<R R
The Bochner-Riesz conjecture asserts that these weighted
Fourier integrals converge in Lp if δ > 0, n| p1 − 21 −
2| − 12 , or if (δ, p) = (0, 2). (These conditions are
known to be necessary).
• If the Kakeya conjecture was false, then one could
exploit the counterexample to give a counterexample
to the Bochner-Riesz conjecture. Thus the Bochner-
Riesz conjecture implies the Kakeya conjecture. (Bour-
gain, unpublished; T., 1998. Heuristically this impli-
cation dates back to the 1970s).
• The converse implication is not known; we do not
know how to convert a complete proof of the Kakeya
conjecture to that of the Bochner-Riesz conjecture.
However, in 1991 Bourgain introduced an argument
which uses partial progress on Kakeya to give partial

11
 progress on Bochner-Riesz, basically by decomposing
an arbitrary function f into wave packets ψ of the
type used above.
• However, progress is still slow. The Bochner-Riesz
conjecture is proven in two dimensions (Carleson,
Sjölin, 1972), but in three dimensions it is only solved
for p > 4 − 8/31 or p′ > 4 − 8/31 (T., Vargas, 1998).
• The Besicovitch set has been used as a counterex-
ample in many other contexts as well. For instance,
in 1993 Bourgain used the Besicovitch set and the
“short sum” construction to disprove the strongest
form of Montgomery’s conjecture on the distribution
of large values of Dirichlet series. (A less strong ver-
sion of Montgomery’s conjecture is still open, and
would imply the Kakeya conjecture just like the Bochner-
Riesz conjecture does).
• We now show some connections between the Kakeya
problem and the linear wave equation.

12
Local smoothing for the wave equation

• Consider the free wave equation in R × Rn with

initial position f and initial velocity zero:
utt − ∆u = 0; u(0, x) = f (x); ut(0, x) = 0

• One can solve for u in terms of f :

√
u(t) = cos(t −∆)f.

• We are interested in how the size and smoothness of

u is determined by that of f and g.
• A basic estimate is the energy estimate
∥u(t)∥L2(Rn) ≤ ∥f ∥L2(Rn)
or more generally
∥∇k u(t)∥L2(Rn) ≤ ∥∇k f ∥L2(Rn)
for any t ∈ R, k ≥ 0.
• This estimate only gives L2 information, and does
not reveal whether a solution u(t) focuses or disperses
as time progresses. To obtain such information we
need to augment the above L2 bounds with Lp esti-
mates.

13
• Such estimates exist, but they lose a lot of regularity.
For instance, one has the decay estimate
∑
s
−(n−1)/2
∥u(t)∥L∞(Rn) ≤ Ct ∥∇sf ∥L1(Rn)
k=0

for all t > 0, provided that s > (n + 1)/2. This esti-

mate has many uses, such as establishing global well-
posedness and scattering for non-linear wave equa-
tions with small data.
• The decay estimate requires so many derivatives be-
cause waves can focus at a point, creating a large L∞
norm even if the initial data has small L1 norm.
• However, one can obtain estimates requiring far fewer
derivatives if one averages in time. Heuristically, this
is because a wave can only focus at a point for very
short periods of time before it disperses again. This
effect is known as local smoothing.
• The first estimates of this type to be discovered were
the Strichartz estimates, and grew out of the work
on such problems as the Bochner-Riesz problem dis-
cussed earlier. A typical estimate is
∥u∥L4(R×R3) ≤ C∥∇1/2f ∥L2(R3).

14
 (Strichartz, 1977) This particular estimate can be
used to establish global well-posedness and scattering
for the non-linear wave equation utt − ∆u = F (u)
for small, low regularity initial data whenever the
non-linearity F is at most cubic.
• Further estimates of Strichartz type were developed
over the decades; the final outstanding endpoint es-
timate was settled recently (Keel, T. 1997).
• However, even Strichartz estimates lose some deriva-
tives, because of travelling wave counterexamples and
the fact that the order of integrability on the left-
hand side is higher than that of the right. In 1991
Sogge conjectured an estimate of this type, namely
√
∥u∥L4([1,2]×R2) ≤ Cε∥(1 + −∆)εf ∥L4(R2)
for all ε > 0. This estimate is only known for
ε > 5/44 (Wolff, 1999, and Vargas, T., 1998). The
argument again requires some estimates related to
the Kakeya conjecture (but with line segments re-
placed by light rays). The idea is to resolve u into a
linear combination of travelling waves, each of which
is concentrated on the neighbourhood of a light ray.
• Unfortunately, one must always lose at least an ep-
silon of derivatives in Lp for p ̸= 2 (Wolff, 1996).
15
This example is again based on the Besicovitch set,
and involves placing a “wave train” as initial data
on each of the rectangles Ri. These wave trains will
interact with large Lp norm for all times 1 ≤ t ≤ 2,
so that averaging in time has no appreciable effect.

16
• Variations of this example show that the local smooth-
ing conjecture (in general dimensions) implies the
Bochner-Riesz and Kakeya conjectures.

17
• Similar situations occur for other dispersive equa-
tions (Schrödinger, KdV, etc.) but will not be dis-
cussed here.

18
Non-linear wave equations

• Estimates for the free wave equation are important

to the analytic study of non-linear wave equations. A
typical example is the cubic non-linear wave equation
utt −∆u = u3; u(0, x) = εf (x); ut(0, x) = εg(x)
where ε > 0 is a small parameter.
• As it turns out, the solution to this equation depends
analytically on the parameter ε, at least for small ε
and initial data f , g which is not too rough. One can
then expand u as a power series
u = εu1 + ε3u3 + ε5u5 + . . .
where u1 is the free solution
(u1)tt−∆u1 = 0; u1(0, x) = f (x); (u1)t(0, x) = g(x)
and u3 is obtained by solving the inhomogeneous
wave equation
(u3)tt−∆u3 = (u1)3; u3(0, x) = 0; (u3)t(0, x) = 0.
More complicated formulas exist for u5, u7, etc.
• Of course, one must establish that this series actually
converges. It is here that the Lp estimates discussed
19
 earlier come into play. These estimates can also be
used to discover further properties of the solution,
such as uniqueness, persistence of regularity, scatter-
ing, decay estimates, etc.
• The above equation is called semi-linear because the
highest order terms are linear, and the non-linear
terms contain no derivatives. Thanks in large part to
Strichartz and local smoothing estimates, the theory
of semi-linear wave equations is now almost complete.
For instance, in almost all cases one knows the min-
imal regularity on the initial data one needs in order
for the above Cauchy problem to be well-posed.
• Our theory is less satisfactory in the semi-linear with
derivatives case, in which the non-linearity is allowed
to contain first-order derivatives. Examples include
the wave map equation
ϕtt − ∆u = −ϕ(|ϕt|2 − |∇ϕ|2)
or the Yang-Mills equation
Dα Dα A = 0,
where Dα = ∂α + [Aα , ·] are covariant derivatives,
and one has chosen an appropriate gauge.

20
• For these equations, the Strichartz estimates are no
longer as useful because they lose too many deriva-
tives. Local smoothing estimates have also been un-
successful.
• However, it was realized in the last decade that one
needs multi-linear estimates rather than linear esti-
mates in order to correctly analyze these equations.
A typical estimate is
∥utvt−∇u·∇v∥L2(R×R3) ≤ C∥∇2f1∥L2(R3)∥∇f2∥L2(R3)

(Klainerman, Machedon, 1993), where

utt − ∆u = 0; u(0, x) = f1(x); ut(0, x) = 0
vtt − ∆v = 0; v(0, x) = f2(x); vt(0, x) = 0.
This estimate can be used to show, for instance, that
the wave map equation is analytically well-posed in
three dimensions if the initial data has two deriva-
tives in L2.
• Bilinear L2 estimates for the homogeneous (Foschi,
Klainerman, 1999) and inhomogeneous (T., 2000)
wave equation have recently been completely clas-
sified. These estimates have led to near-optimal reg-
ularity results for semi-linear wave equations with
21
 derivatives (much work by Klainerman-Machedon and
others; parallel work by Bourgain, Kenig-Ponce-Vega
and others for Schrödinger, KdV, etc.) Interesting
problems still remain, especially in the very diffi-
cult critical regularity case, at which point analyt-
icity of the solution breaks down. (For instance, it
is not known whether singularities can form for two-
dimensional wave maps; the difficulty here is that the
energy class is critical).
• Bilinear L2 estimates are usually proven using Plancherel’s
theorem. Bilinear Lp estimates are more difficult,
with the first non-trivial result being quite recent
(Bourgain, 1995; Vargas, T., 1998). In the pure
space-time norm case, however, this problem has also
been recently solved (Wolff, 1999; T., 1999). These
new estimates may have application to the critical
regularity problem. Again, these estimates rely on
Kakeya type information concerning the combina-
torics of light rays.
• Very recently (Tataru, Bahouri-Chemin, Klainerman,
1999-2000), some of these techniques have begun to
be extended to the quasi-linear case, in which the
spacetime metric is allowed to vary in a non-linear

22
 fashion depending on the data. An extremely diffi-
cult instance of this occurs in the Einstein equations
in general relativity.
• These estimates also have application to controlling
other types of behaviour of PDEs. For instance, one
can control the movement of energy from one fre-
quency range to another over long periods of time,
and thus show that no blowup occurs for such equa-
tions as the KdV equation even when the energy
and L2 norm are infinite (Colliander, Keel, Staffi-
lani, Takaoka, T., 2000). These estimates may also
play a part in solving the soliton conjecture, which
states that dispersive equations of KdV type always
disperse into travelling waves (solitons) even when
the equation is not completely integrable.

23