where the estimate was concluded by the following estimate X
θ
kPθfkpLp
!1/p
.p kfkLp,
which holds for 2 ≤p < ∞. This can be inferred from Plancherel’s theorem for p= 2 and from Young’s inequality for p =∞. The remaining cases follow from interpolation.
7.4 L
p-smoothing estimates for elliptic phase
Proposition 7.4.3. Let p >2 +n+14 andφbe a normalized elliptic phase function giving rise to the H¨ormander operatorTλ. Then, for anyε >0, we find the following estimate to hold:
kTλfˆkLp(Rn+1).ελn(1/2−1/p)+εkfkLp(Rn). (7.52) For a phase functionφsatisfyingH1) letq(t, x;ξ) =∂tφ(t, x; [∂xφ(t, x;·)]−1(ξ)) denote the local parametrization of the surface associated to the adjoint Fourier re-striction operator. To prove theLp-smoothing estimate for variable-coefficient phase functions, crucial use is made of the following bilinear estimate using transversality, which is a generalization of the theorem from [Tao03]:
Theorem 7.4.4 ([Lee06b, Theorem 1.1, p. 58]). For i = 1,2, let φi be smooth functions satisfyingH1). Suppose that the Hessian∂ξξ2qi satisfies
det∂ξξ2 qi(t, x;∂xφi(t, x;ξi))6= 0 on the support ofai and for(t, x;ξi)∈supp(ai)
|h∂xξ2 φi(t, x;ξi)δ(t, x;ξ1, ξ2),
[∂xξ2 φi(t, x;ξi)]−1[∂2ξξqi(t, x;ui)]−1δ(t, x;ξ1, ξ2)i| ≥c >0 fori= 1,2, whereui=∂xφi(t, x;ξi)and
δ(t, x;ξ1, ξ2) =∂ξq1(t, x;u1)−∂ξq2(t, x;u2).
Then, for anyε >0, there is a constantC=C(ε)such that forp≥(n+ 3)/(n+
1),
kT1λf T2λgkLp(R×Rn)≤Cελεkfk2kgk2,
whereTiλ denotes the H¨ormander-operator associated to the data(φi, ai).
For normalized elliptic phase functions this yields the following corollary:
Corollary 7.4.5. Let φ be a normalized elliptic phase function and suppose for f1, f2 ∈L2(Rn) with supp(fi)⊆Q(2). Then, for any ε >0 there is Cε such that forp≥(n+ 3)/(n+ 1)we find the following estimate to hold:
kTλf1Tλf2kLp(Rn×R)≤Cελεkf1k2kf2k2.
Proof. We check that the assumptions of the more general Theorem 7.4.4 are sat-isfied: Indeed, from normalization we infer ∂xξ2 φ ∼ In, δ(x, t;ξ1, ξ2) ∼ ξ1 −ξ2,
∂2ξξqi∼In and consequently,
|h∂2xξφ(x, t;ξi)δ(x, t;ξ1, ξ2),[∂xξ2 φ(x, t;ξi)]−1[∂ξξ2qi(x, t;ui)]−1δ(x, t;ξ1, ξ2)i|
∼ |ξ1−ξ2|2≥c2. This yields the claim.
Due to a localization property of the kernel, this yields the followingLp -estima-tes.
Lemma 7.4.6. Let p > 2(n+3)n+1 , B1, B2 ⊆ Q(1) balls with dist(B1, B2) ≥ c > 0 andφbe a normalized elliptic phase function. Then, forf1, f2 with supp( ˆfi)⊆Bi, i= 1,2, we find the following estimate to hold:
kTλfˆ1Tλfˆ2kLp/2(R×Rn).p,ε λn(1−2/p)+εkf1kLp(Rn)kf2kLp(Rn). Proof. First, we note the localization property for
K(t, x;y) =aλ0(t, x) Z
Rn
ei(φλ(t,x;ξ)−hy,ξi)β(ξ)dξ.
In fact,
∂ξ[φλ(t, x;ξ)− hy, ξi] =∂ξφλ(0, x;ξ) + Z t/λ
0
∂t∂ξφ(t0, x/λ;ξ)dt0−y
=x−y+λ Z t/λ
0
dt0 Z 1
0
ds∂t∂ξξ2φ(t0, x/λ;sξ)·ξ
=x−y+O(λ).
Consequently, integration by parts yields that there is a constantCwhich depends only on the constants in the definition of normalized elliptic phase functions so that for |y| ≥ Cλ we find |K(t, x;y)| ≤ CNλ−N|x−y|−N from integration by parts.
LetχQ be a smooth function which is essentially supported onQ(Cλ) with Fourier support inB(0, c/10) and decomposefi=χQfi+fi2=fi1+fi2.
From trivial kernel estimates we find
kTλfˆi1kLp(R×Rn)≤λC0kfikLp(Rn),
whereC0=C(n, p) and from the kernel estimate and Young’s inequality kTλfˆi2kLp(R×Rn)=
Z
K(t, x;y)fi2(y)dy Lp(
R×Rn)
≤CNλ−Nkfi2kLp(Rn)≤CNλ−NkfikLp(Rn). We split
kTλfˆ1Tλfˆ2kLp/2(R×Rn)≤ kTλfˆ11Tλfˆ21kLp/2(R×Rn)+kTλfˆ12Tλfˆ21kLp/2(R×Rn)
+kTλfˆ11Tλfˆ22kLp/2(R×Rn)+kTλfˆ12Tλfˆ22kLp/2(R×Rn)
≤I+II+III+IV.
II,III,IV are estimated by the above elementary estimates. For instance, II≤ kTλfˆ12kLpkTλfˆ21kLp≤CNλ−Nkf1kLpλC0kf2kLp.
Only for I we have to invoke the bilinear estimate from above. Due to the support properties ofχQ we can still use Corollary 7.4.5
kTλfˆ11Tλfˆ21kLp/2(Rn×R).ελεkχ[Qf1kL2kχ[Qf2kL2
≤CελεkχQf1kL2kχQf2kL2
≤Cελελn(1−2/p)kf1kLp(Rn)kf2kLp(Rn), and the proof is complete.
Lemma 7.4.6 is the most important building block for further arguments. To utilize it efficiently, we carry out a Whitney decomposition of Tλfˆ1Tλfˆ2 in terms of frequency support and separation.
Precisely, let
B(1)×B(1)\D= [
0≤j≤j0
[
dist(Qjk
1,Qjk
2)∼2−j
Qjk
1×Qjk
2, where for 0≤j < j0= 12log2λ1/2the cubesQjk
1, Qjk
2have side length comparable to 2−j and are separated with a distance comparable to 2−j. Forj=j0the separation is less or equal to 2−j.
Adapted to the Whitney-decomposition, we write (Tλfˆ)2= X
0≤j≤j0
X
k1,k2
Tλ(βkj
1
fˆ)Tλ(βkj
2
fˆ)
=: X
0≤j≤j0
B˜j[f, f],
where theβjk denote a smooth partition of unity adapted to the Whitney-decom-position.
Forj < j0 we prove the following estimate by scaling:
Lemma 7.4.7. Let φ be a normalized elliptic phase function andfi have Fourier support in cubes of side length2−j which are also separated with a distance of2−j. Then, we find the following estimate to hold provided that1≤22j ≤λ:
kTλfˆ1Tλfˆ2kLp/2.ελε24jp(λ/22j)n(1−2/p)kf1kLp(Rn)kf2kLp(Rn). (7.53) Proof. We use parabolic rescaling to write
kTλfˆ1Tλfˆ2kLp/2(Rn+1)
= 22jnp 2−2jn24jpk Z
eiφλ(22jt0,2jy,ξ0+2−jξ0)aλ(22jt0,2jy;ξ0+ 2−jξ0) ˆf1(ξ0+ 2−jξ0)dξ0 Z
eiφλ(22jt,2jy;ξ0+2−jξ0)aλ(22jt0,2jy;ξ0+ 2−jξ0) ˆf2(ξ0+ 2−jξ0)dξ0kLp/2(Rn+1)
= 2−2jn22jnp 24jpk Z
eiφ˜λ/22
j(t0,y;ξ0)˜aλ/22j(t0, y;ξ0)ˆg1(ξ0)dξ0 Z
eiφ˜λ/22
j(t0,y;ξ0)˜aλ/22j(t0, y;ξ0)ˆg2(ξ0)dξ0kLp/2(Rn×R).
(7.54) Due to ˆgi(ξ) = ˆfi(ξ0+ 2−jξ), we havekgikLp= 2nj2−njpkfikLp. We argue that the latter expression is amenable to Lemma 7.4.6: Note that
supp(˜aλ/22j(t,·;ξ))⊆Q(λ/2j).
However, we want to apply Lemma 7.4.6 with parameterλ/22j. For this purpose we use again the localization property of the kernel. Note that this holds for ˜φλ/22j independently of x and in particular, there is no renormalization with an affine change of coordinates necessary for this property to hold.
CoverQ(λ/2j) byλ/22j cubes Ql and letg=P
lχQlg, whereχQl are smooth functions essentially supported onQl withP
χQl= 1 and
supp(χdQl)⊆B(0, c) (7.55)
for somec1. This is possible due to the Poisson summation formula andλ/22j≥ 1.
Effectively, we work with the decomposition g=
N
XχQlg+χQ∗g, χQ∗g= X
Ql∩Q(λ1+ε/2j)=∅
χQlg.
Like in the proof of Lemma 7.4.6, we find that the terms containingχQ∗gcan safely be neglected.
In fact, we find the main contribution of (7.54) raised top/2 to be given by X
Qk,l,m:
Ql∩λεQk6=∅, Qm∩λεQk6=∅
T˜λ/22j(χ\Qlg1) ˜Tλ/22j(χ\Qmg2)
p/2
Lp/2([−λ/22j,λ/22j]×Qk) (7.56)
The conditions onQk, l, min the above display are denoted byQk, l, m: (∗) in the following.
From (7.55) the supports of χ\Qlg1and χ\Qmg2 are still separated of unit order, and after the change of variablesx→x0+x, we find
kT˜λ/22j(χ\Qlg1) ˜Tλ/22j(χ\Qmg2)kLp/2([−λ/22j,λ/22j]×Qk)
=kT˜kλ/22j(eiφ˜χ\Qlg1) ˜Tkλ/22j(eiφ˜χ\Qmg2)kLp/2([−λ/22j,λ/22j]×Q0),
(7.57)
whereφk= ˜φ(t, x+x0;ξ)−φ(0, x˜ 0;ξ).
Note that theφkare normalized elliptic phase functions after an additional affine transformation, which is close to the identity mapping. Hence, (7.57) is amenable to Lemma 7.4.6 with parameter λ/22j, and we conclude the bound by means of Cauchy-Schwarz:
X
Qk,l,m:(∗)
kT˜λ/22j(χ\Qlg1) ˜Tλ/22j(χ\Qmg2)kp/2Lp/2
([−λ/22j,λ/22j]×Qk)
.ε X
Qk,l,m:(∗)
(λ/22j)[n(1−2/p)+ε](p/2)kχQlg1kp/2LpkχQmg2kp/2Lp
.ε(λ/22j)[n(1−2/p)+2ε](p/2) X
l
kχQlg1kpLp
!1/2 X
m
kχQmg2kpLp
!1/2
.ε(λ/22j)[n(1−2/p)+2ε](p/2)kg1kp/2Lpkg2kp/2Lp, and the proof is complete.
For possibly vanishing separation for cubes with side lengthλ−1/2 we have the following lemma:
Lemma 7.4.8. Let supp( ˆf)⊆B, whereB is a ball of radiusλ−1/2. Then, we find the following estimate to hold:
kTλfˆkLp(R×Rn).ελελ1/pkfkLp(Rn).
Proof. We use parabolic rescaling to write like in the proof of Lemma 7.4.7 kTλfkˆ Lp(R×Rn).λ−n/2λn/2pλ1/pkT˜1gkˆ Lp(R×Rn),
wherekgkLp=λn/2λ−n/2pkfkLp. It is enough to prove kT˜1ˆgkLp(R×Rn).ελεkgkLp(Rn).
Again, we use the localization property of the kernel to consider cubes of unit size and adapted functions with localized Fourier transformχQ. Decompose like in the proof of Lemma 7.4.7
g= X
l:Ql∩Q(λ1/2+ε)6=∅
χQlg+χQ∗g.
Like in the proof of Lemma 7.4.7 one finds kT˜1ˆgkpLp([−1,1]×Q).X
Q,l
kT˜1χ[QlgkpLp([−1,1]×Q)+CNλ−NkgkLp. The proof is concluded following along the lines of the proof of Lemma 7.4.7.
To separate the contribution of the differentki, we use a natural orthogonality of the oscillatory integrals, which was used in a similar context in [Lee06b].
Lemma 7.4.9([Lee06b, Lemma 3.3, p. 82]). Letφλ be a normalized elliptic phase function, 1 ≤ p/2 ≤ 2 and j ≥ 12log2λ. Then, we find the following estimate to hold:
X
dist(Qjk
1,Qjk
2)∼2−j
Tλ(βkj
1
fˆ)Tλ(βjk
2
f)ˆ p/2
≤Cελε
X
k1,k2
kTλ(βjk
1
fˆ)Tλ(βkj
2
fˆ)kp/2p/2
2/p
+CNλ−Nkfk2p.
For the error term recall that we can always suppose thatf in position space is localized to a cube of lengthλfrom the kernel estimate. Hence, using the estimate from [Lee06b, Lemma 3.3] withr= 2 yields the claim after applications of H¨older’s inequality and Plancherel’s theorem.
We prove the following bound for ˜Bj.
Proposition 7.4.10. Let 0 ≤j ≤ 12log2(λ) and B˜ like above. Then, we find the following estimate to hold:
kB˜j[f, f]kLp/2 .ε
( λε24j/p(λ/22j)n(1−2/p)kfk2Lp, 2 +n+4n+1 ≤p≤4,
λε24j/p(λ/22j)n(1−2/p)kfk2Lp, 4< p <∞. (7.58)
Proof. We have to distinguish between 1 ≤ p/2 ≤ 2 and p > 4. In fact, for 1 ≤p/2 ≤2 we can apply Lemma 7.4.9, but for p >4 we use instead the trivial observation
X
dist(Qjk
1,Qjk
2)∼2−j
Tλ(βjk
1
f)Tˆ λ(βkj
2
fˆ) L∞
.2jn sup
k1,k2
kTλ(βkj
1
fˆ)Tλ(βkj
2
fˆ)kL∞
and interpolation yields kX
Tλ(βjk
1
fˆ)Tλ(βkj
2
fˆ)kLp/2 ≤Cελε2jn(1−4/p)
X
k1,k2
kTλ(βkj
1
fˆ)Tλ(βkj
2
fˆ)kp/2Lp/2
2/p
.
(7.59)
Consequently, for 1≤p/2≤2:
k X
dist(Qjk
1,Qjk
2)∼2−j
Tλ(βkj
1
fˆ)Tλ(βkj
2
fˆ)kLp/2
.ελε X
dist∼2−j
kTλ(βkj
1
fˆ)Tλ(βjk
2
f)kˆ p/2Lp/2
!2/p
.ελ2ε24j/p(λ/22j)n(1−2/p)
X
k1,k2
kPkj
1fkp/2LpkPkj
2fkp/2Lp
2/p
.ελ2ε24j/p(λ/22j)n(1−2/p) X
k1
kPkj
1fkpLp
!1/p
X
k2
kPkj
2fkpLp
!1/p
≤Cελ2ε24j/p(λ/22j)n(1−2/p)kfkLpkfkLp
and forp >4 by the above means, but appealing to (7.59) than Lemma 7.4.9,
X
dist(Qjk
1,Qjk
2)∼2−j
Tλ(βkj
1
fˆ)Tλ(βjk
2
fˆ) Lp/2
.ελ2ε2jn(1−4/p)
24j/p(λ/22j)(1−2/p)kfk2Lp.
Proof of Proposition 7.4.3. We conclude that for small p the main contribution comes fromj= 0 and summing a geometric series yields the bound
kTλfkˆ Lp=k(Tλfˆ)2k1/2Lp/2 ≤
X
0≤j≤j0
kB˜j[f, f]kLp/2
1/2
.ε
X
0≤j≤j0
λελn(1−2/p)24j/p2−2j(n(1−2/p))kfk2Lp
1/2
.λελn(1/2−1/p)kfkLp.
The same argument works forp→ ∞and interpolation between the bounds for smallpand largepproves the claim.
Conclusively, we argue how Proposition 7.4.3 implies Theorem 7.1.3.
Proposition 7.4.3⇒Theorem 7.1.3. Let ˆf be supported in a ρ−1-ball centered at ρ0. It follows from parabolic rescaling (cf. Section 7.2)
eiλφ(Φξ0(ρ2t/λ,ρx/λ);ξ0)Tλfˆ◦Φλξ
0◦Dρ= ˜Tλ/ρ2ˆg, where
T˜λ/ρ2g(t, x) =ˆ Z
Rn
eiφ˜λ/ρ
2(t,x;ξ)a˜λ/ρ2(t, x;ξ)ˆg(ξ)dξ, and the phase ˜φis given by
hx, ξi+ Z 1
0
(1−r)h∂ξξ2 φ(Φξ0(t, Dρ0−1x), ξ0+rξ/ρ)ξ, ξidξ.
A change of spatial variables gives
kTλfˆkLp(R×Rn).φρn+2p ρ−nkT˜λ/ρ2ˆgkLp((Φλξ
0◦Dρ)−1(R×Rn), where like abovekgkLp =ρnρ−n/pkfkLp.
The support of ˜Tλ/ρ2ˆgis essentially aλ/ρ×. . .×λ/ρ×λ/ρ2-ellipse. As argued in Section 7.2, ˜φis up to an affine transformation a normalized elliptic phase function.
Note that in the current context, we chooseρ depending on φ so that the phase function, we arrive at after rescaling and an additional affine change of coordinates, is actually normalized. Moreover, the magnitude of the Jacobian of Φξ0 andL(see Section 7.2 for notation) depend on φ. Thus, the implicit constant depends on φcontrary to the applications of parabolic rescaling for normalized elliptic phase functions. To utilize Proposition 7.4.3, we perform an additional decomposition of the ellipse intoλ/ρ2-cubes and use again the orthogonality property which follows from the localization property of the kernel. The argument is concluded like in the proofs of Lemma 7.4.7 and 7.4.8. In order to avoid repetition, the details are omitted.
In the case of constant coefficients it was shown in [RS10] how a localization property of the kernel and an interpolation argument yield globalization.
We point out that the bilinear approach depicted above for variable coefficients also applies to the hyperbolic Schr¨odinger equation (cf. [Lee06a, Var05]) in two dimensions:
i∂tu(t, x) + (∂xx−∂yy)u(t, x) = 0, (t, x)∈R×R2,
u(0, x) =u0(x)∈Lpβ(R2), (7.60) and using the localization and interpolation argument from [RS10, Section 4] which applies mutatis mutandis yields the following result:
Theorem 7.4.11. Forp >10/3we find the following estimate to hold:
keit(∂xx−∂yy)u0kLp([0,1]×R2).ku0kLp
β(R2), β≥4(1 2 −1
p)−2
p (7.61)
This covers the endpoint Sobolev regularity, which is not covered in [Rog08].