Null Form Estimates For The Cone: Tao/preprints

Null form estimates for the cone
Terence Tao (UCLA)
www.math.ucla.edu/∼tao/preprints
Abstract
Linear and bilinear Lp estimates for the wave equa-

tion have applications to non-linear wave equations, but
the full range of estimates which hold is not completely
known. However, Wolff has recently obtained a nearly
sharp Lp estimate in the bilinear setting, and the author
has obtained the endpoint of this estimate.
1
Introduction
• Consider waves
∂ 2u
− ∆u = 0
∂t2
in Rn with initial data in a Sobolev space H s ×H s−1.
• Q1. Can we control u in some spacetime Lebesgue
norm Lpx,t?
• Q2. Can we control bilinear expressions of u, e.g.
|ut|2 − |∇u|2 in Lpx,t?
• Question 1 is useful in controlling the non-linear ef-
fects of the semi-linear wave equation 2u = |u|p,
while Question 2 addresses the non-linear effects of
equations such as the wave map equation
2u = −uQ(u, u)
where
Q(u, v) = utvt − ∇u · ∇v.
• Q is known as a null form, because Q(u, v) vanishes
when u = v is a travelling wave.
• The estimates in Q1 are known as generalized Strichartz
estimates, and have been completely classified (around
2
the early 90s, with the final endpoint result obtained
last year).
• The estimates in Q2 are known as null form esti-
mates. Although we do not yet have a complete clas-
sification, there has been much recent progress, and
many near-sharp results have been obtained, at least
in the pure spacetime-norm case. (We do not con-
sider mixed space-time norm estimates in this talk).
3
Fourier transform
• Waves can be studied using the space-time Fourier
transform. With the notation of the previous slide,
we have
u=[ F dσ
where dσ is induced surface measure on the double
light cone Γ+ ∪ Γ−, with
Γ± = {(ξ, τ ) ∈ Rn+1 : τ = ±|ξ|}
and F is a function on Γ+ ∪Γ− satisfying the weighted
L2 estimate
∫
n
|F (ξ, ±|ξ|)|2|ξ|2s dξ . 1
R
• By dyadic decomposition we can reduce to when F
is supported on
Γ0 = {(ξ, τ ) ∈ Rn+1 : τ = |ξ|, 1 ≤ |ξ| ≤ 2}.
In this case all spacetime derivatives are essentially
irrelevant, and Q1 reduces to
∥[
F dσ∥Lpx,t . ∥F ∥2
• This type of estimate is known as an adjoint restric-

tion estimate. By adapting the T T ∗ arguments of
4
Tomas and Stein, Strichartz proved the above esti-
mate for all p ≥ 2(n+1)
n−1 . The key ingredient is the
curvature of the surface Γ0, as reflected in the decay
c
of dσ:
c t)| . (1 + |t|)− n−1
|dσ(x, 2 .
• Strichartz’ estimate is sharp. The example for this is

the Knapp example, in which F is a bump function
adapted to the sector
{(ξ, τ ) ∈ Γ0 : ∠ξ, e1 . 1/r}
for some r ≫ 1:
5
6
Bilinear estimates
• We now consider the question of obtaining control on
the bilinear expression ∥Q(u, v)∥Lp(x,t) for waves u,
v and suitable Sobolev control on the initial data.
• We can write u = [
F dσ and v = [
Gdσ.
• As a model case, we that F and G are supported on
Γ0 [This is not the most typical case]. The estimate
then simplifies to
∥Q([ F dσ, [
Gdσ)∥p . ∥F ∥2∥G∥2.
• To understand Q, suppose first that u, v are travel-
ling waves
u = cei(x·ξ+t|ξ|), v = c′ei(x·η+t|η|)
with |ξ|, |η| ∼ 1. Since
ut = |ξ|u, vt = |η|v, ∇u = ξu, ∇v = ηv
we have
Q(u, v) = (|ξ||η| − ξ · η)uv ∼ ∠(ξ, η)2uv.
• Thus Q is only large when ∠(ξ, η) ∼ 1, in which
case Q(u, v) is like uv. We are thus led to study the
estimate
∥F[ [2∥p . ∥F ∥2∥G∥2
dσ1Gdσ (∗)
7
where dσ1 and dσ2 are surface measure on disjoint
sectors in Γ0 of angular width ∼ 1 and separation
∼ 1; if one can establish this estimate, one has a
good chance of proving null form estimates for this
value of p.
• From Strichartz and Hölder’s inequality, (*) holds for
p ≥ (n + 1)/(n − 1). However, one does not expect
this to be sharp, because the Knapp examples on dσ1
and dσ2 are so different:
8
• Klainerman and Machedon observed that (*) also
holds for p = 2. This is because we can use Plancherel
to rewrite (*) as
∥(F dσ1) ∗ (Gdσ2)∥2 . ∥F ∥2∥G∥2
and use a Cauchy-Schwarz argument, exploiting the
9
transversality between dσ1 and dσ2 via the estimate
∥dσ1 ∗ dσ2∥∞ . 1
• This L2 estimate (and the null form estimates it gen-

erates) have been crucial in the low-regularity local
existence theory of non-linear wave equations.
• Using some bilinear Knapp variants, Machedon noted
10
that the condition p ≥ (n + 3)/(n + 1) is necessary
for (∗), and conjectured that this was also sufficient.
• In 1995 Bourgain showed that when n = 2, that (*)
could be pushed down to p > 2 − 13/2408. This was
improved by Vargas and the author to p > 2 − 8/121
last year.
• The full conjecture was mostly proved by Wolff early
in this year; he obtained the estimate for all p >
(n + 3)/(n + 1). The endpoint p = (n + 3)/(n + 1)
has since been obtained by the author.
11
Red and blue waves
• We now undo the Fourier transform and rewrite our

simplified conjecture back in terms of waves.
• We say that a wave u is red if û is supported on the
sector
{(ξ, |ξ|) : 1 ≤ |ξ| ≤ 2, ∠(ξ, e1) < 1/10}
and blue if û is instead supported in
{(ξ, |ξ|) : 1 ≤ |ξ| ≤ 2, ∠(ξ, −e1) < 1/10}.
• Red and blue waves propagate along different sets

of characteristics; red waves propagate along null di-
rections close to (−e1, 1), and blue waves along null
directions close to (e1, 1).
• For any red wave u, define the energy E(u) by
∫
E(u) = |u(x, t)|2 dx
for any t ∈ R; this definition is independent of

t. (Because frequency is localized, one can dispense
with the derivatives when defining energy). Simi-
larly define E(v). The previous results can then be
rewritten as
12
• Theorem. (Wolff, T.) For all p ≥ (n + 3)/(n + 1),
all red waves u and all blue waves v we have
∥uv∥p . E(u)1/2E(v)1/2.
• We now indicate some of the ideas used in the proof.
13
Spatial localization
• Following standard techniques, we prove the global
spacetime estimate by first proving local spacetime
estimates.
• Let p0 = (n + 3)/(n + 1), and for all R ≫ 1 let A(R)
be the best constant such that
∥uv∥Lp0 (Q) ≤ A(R)E(u)1/2E(v)1/2
for all red and blue waves u and v, and all spacetime
cubes Q of side-length R.
• Wolff showed that A(R) = O(Rε) for all ε > 0;
a technical argument of Bourgain then shows that
(∗) holds for all p > p0. Wolff’s estimate can be
tightened to A(R) = O(1), which yields the endpoint
p = p0 .
• One new innovation in Wolff’s argument is that bounds
on A(R) are obtained inductively, by first proving
A(R) . (log R)C A(R1−ε) + RCε,
The endpoint argument is similar but is based on the
more precise inequality
A(R) ≤ (1 − ε)A(R/2) + C.
14
Frequency localization
• Fix R, and let Q be a cube of radius R.

• Let us first bound ∥uv∥Lp(Q) under a very strong
assumption on u, namely that û is√supported in a
frequency sector of angular width 1/ R. This means
that u has very little angular dispersion.
• In this case, we have the strong L2 estimate
∥uv∥2 . R−(n−1)/4E(u)1/2E(v)1/2;
without the low angular dispersion hypothesis we
could not obtain the R−(n−1)/4 factor. This estimate
goes back at least to Mockenhaupt. The gain comes
from the observation that ∥dσ1 ∗ dσ2∥∞ is not just
bounded by 1 as before, but is now much smaller,
about R−(n−1)/2, when dσ1 is restricted to a sector
of width R−1/2.
15
• On the other hand, from the definition of energy we
have
∥u∥L2(Q) ≤ R1/2E(u)1/2
so by Hölder we have
∥uv∥L1(Q) ≤ RE(u)1/2E(v)1/2.
16
Interpolating between the L1 and L2 estimate we get
∥uv∥Lp0 (Q) ≤ E(u)1/2E(v)1/2;
note that 1 < p0 < 2 for all dimensions n.
17
General frequencies
• To handle the general case, we split u and v into low

frequency pieces, and try to argue that each piece of
u interacts with at most one piece of v.
• We can partition the red portion of √Γ0 into about
R(n−1)/2 sectors of angular width 1/ R, which in-
duces a decomposition of u into many smaller red
waves, each of which has low frequency dispersion.
• Suppose one of these pieces has frequency support in
a sector centered about the null vector (ω, 1) for some
|ω| = 1. Then one can exploit the uncertainty princi-
ple and (essentially) partition this piece further
√ into
pieces which are physically concentrated on R × R
tubes oriented in the dual direction (−ω, 1).
• Combining these two decompositions, we obtain a
partition
∑ ∑
u= uT , v = vT
T T
where uT has low frequency

√ dispersion and is spa-
tially concentrated in a R × R tube T oriented in
a red null direction, and similarly for v.
18
• Wolff’s argument then proceeds by analyzing the ge-
ometry of the tubes T , locating those slightly smaller
cubes (of side-length R1−ε) where the tubes T are
concentrating in a “bush”:
• It turns out that one can isolate a quite small number

(only (log R)C ) of high density cubes this way.
• Outside these cubes we exploit the low frequency-
19
dispersion estimate. The point is that every tube
vT can only interact with at most O(RCε) tubes uT ′
coming out of each bush:
• The contribution inside the bad cubes is O((log R)C A(R1−εa)),

so we have
A(R) . (log R)C A(R1−ε) + RCε.
20
The endpoint case - an incomplete sketch
• This is an oversimplified, and slightly inaccurate ac-

count of how the endpoint is obtained.
• By going through Wolff’s argument carefully, one can
separate the cube Q into N R/2 cubes which are
“bad” (1 ≤ N ≤ 2n+1) and tighten the inequality to
A(R) ≤ N −εA(R/2) + C.
• If N > 1 then N −ε < 1, and we can use this to

obtain uniform bounds on A(R).
• However, there is still the case in which Q has only
one “bad” R/2 sub-cube. In this case we take the
R/2 cube and subdivide it further into R/4 cubes.
If we again have one bad cube, we divide it further.
Eventually we end up with a bad r-cube Qr for some
r < R which does split up into more than one bad
r/2-cube, so that
∥uv∥Lp(Qr ) ≤ (1 − ε)A(r/2) + C.
• To deal with the remaining portion Q\Qr we use

the fact that u and v must be highly concentrated
near Qr , and that u and v propagate in different
21
directions, to obtain
∥uv∥Lp(Q\Qr ) ≤ εA(R) + C
• If one chooses ε small enough, then the two estimates

together give
∥uv∥Lp(Q) ≤ (1 − ε/2)A(R) + C,
which is enough to obtain uniform bounds on A(R).
22
Remarks
• Similar bilinear estimates have been conjectured for

other equations (Klein-Gordon, Schrödinger, Helmholtz
equations), but only partial results are known. The
techniques used for the cone fail completely, because
of the breakdown in the strong Huygens principle. In
particular, the crucial fact that a blue tube can only
intersect one red tube in a bush is false for any other
equation.
• For instance, in R2+1 the analogous bilinear estimate
for the Schrödinger equation is expected to hold for
p ≥ 5/3, but the best result is p > 15/8 (Vargas,
T.) by techniques somewhat different from the ones
here.
• Our model bilinear estimate can be generalized to
the more difficult (but more useful) null form es-
timates that are within an epsilon of being sharp.
(The epsilon loss is caused by various dyadic decom-
positions). Hopefully, these estimates should have
application to non-linear wave equations.
23

Null Form Estimates For The Cone: Tao/preprints

Uploaded by

Copyright:

Available Formats

Null Form Estimates For The Cone: Tao/preprints

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Null Form Estimates For The Cone: Tao/preprints

Uploaded by

Copyright:

Available Formats

Null form estimates for the cone

Terence Tao (UCLA)

Linear and bilinear Lp estimates for the wave equa-

• This type of estimate is known as an adjoint restric-

• Strichartz’ estimate is sharp. The example for this is

• This L2 estimate (and the null form estimates it gen-

• We now undo the Fourier transform and rewrite our

• Red and blue waves propagate along diﬀerent sets

for any t ∈ R; this definition is independent of

• We now indicate some of the ideas used in the proof.

• Fix R, and let Q be a cube of radius R.

• To handle the general case, we split u and v into low

where uT has low frequency

• It turns out that one can isolate a quite small number

• The contribution inside the bad cubes is O((log R)C A(R1−εa)),

• This is an oversimplified, and slightly inaccurate ac-

• If N > 1 then N −ε < 1, and we can use this to

• To deal with the remaining portion Q\Qr we use

• If one chooses ε small enough, then the two estimates

• Similar bilinear estimates have been conjectured for

You might also like