Quantitative Bounds For Critically Bounded Solutions To The Navier-Stokes Equations

Quantitative bounds for critically bounded solutions to the
arXiv:1908.04958v2 [math.AP] 10 Jul 2020
Navier-Stokes equations
Terence Tao
Abstract. We revisit the regularity theory of Escauriaza, Seregin, and Šverák

for solutions to the three-dimensional Navier-Stokes equations which are uni-
formly bounded in the critical L3x (R3 ) norm. By replacing all invocations of
compactness methods in these arguments with quantitative substitutes, and
similarly replacing unique continuation and backwards uniqueness estimates
by their corresponding Carleman inequalities, we obtain quantitative bounds
for higher regularity norms of these solutions in terms of the critical L3x bound
(with a dependence that is triple exponential in nature). In particular, we
show that as one approaches a finite blowup time T∗ , the critical L3x norm
must blow up at a rate (log log log T 1−t )c or faster for an infinite sequence of
∗
times approaching T∗ and some absolute constant c > 0.
1. Introduction
This paper is concerned with quantitative bounds for solutions u : [0, T ]×R3 →
R , p : [0, T ] × R3 → R to the Navier-Stokes equations
3
∂t u + (u · ∇)u = ∆u − ∇p
(1.1)
∇ · u = 0.
Here we have normalised the viscosity to equal one for simplicity. To avoid tech-
nicalities, we shall restrict attention to classical solutions, by which we mean
solutions that are smooth and such that all derivatives of u, p lie in the space
L∞ 2 3
t Lx ([0, T ] × R ). As our bounds are quantitative and do not depend on any
smooth norms of the solution, it is possible to extend the results here to weaker no-
tions of solution, such as mild solutions of Kato [K], the weak Leray-Hopf solutions
studied in [ESS2], or the suitable weak solutions from [CKN], by using the regular-
ity theory of such solutions; we leave the details to the interested reader. As is well
known, such solutions have a maximal Cauchy development u : [0, T∗ ) × R3 → R3 ,
p : [0, T∗ ) × R3 → R for some 0 < T∗ ≤ ∞, with the restriction to [0, T ] × R3 a
classical solution for all T < T∗ , but for which no smooth extension to time T∗ is
possible if T∗ < ∞. We refer to T∗ as the maximal time of existence of such a
classical solution.
1991 Mathematics Subject Classification. Primary 35Q35, 37N10, 76B99.
Key words and phrases. Navier-Stokes, blowup criterion.
The author is supported by NSF grant DMS-1266164 and by a Simons Investigator Award.
We also thank Stan Palasek and Jiayan Wu for corrections.
1
2 TERENCE TAO
The Navier-Stokes system enjoys the scaling symmetry (u, p, T ) 7→ (uλ , pλ , λ2 T )

for any λ > 0, where
uλ (t, x) := λu(λ2 t, λx)
and
pλ (t, x) := λ2 p(λ2 t, λx),
Among other things, this means that the norm
kukL∞ 3 3
t Lx ([0,T ]×R )
is scale-invariant (or critical ) for this equation. In [ESS2] it was shown that as
long as this norm stays bounded, solutions to Navier-Stokes remain regular. In
particular, they showed an endpoint of the classical Prodi-Serrin-Ladyshenskaya
blowup criterion [Pr], [S2], [La] or the Leray blowup criterion [Le]:
Theorem 1.1 (Qualitative blowup criterion). [ESS2] Suppose (u, p) is a clas-
sical solution to Navier-Stokes whose maximal time of existence T∗ is finite. Then
lim sup ku(t)kL3x (R3 ) = +∞.
t→T∗+
There are now many proofs, variants and generalisations [ESS2], [KK], [GKP],
[GKP2], [S1], [Ph], [GIP] [DD], [A], [BS], [SS], [WZ] of this theorem, including
extensions to higher dimensions or other domains than Euclidean spaces, replacing
L3 with another critical Besov or Lorenz space, or replacing the limit superior by a
limit. However, in contrast to the more quantitative arguments of Leray, Prodi, Ser-
rin and Ladyshenskaya, the proofs in the above references all rely at some point on
a compactness argument to extract a limiting profile solution to which qualitative
results such as unique continuation and backwards uniqueness for heat equations
(as established in particular in [ESS]) can be applied. As such, the above proofs
do not easily give any quantitative rate of blowup for the L3 norm.
On the other hand, the proofs of unique continuation and backwards uniqueness
rely on explicit Carleman inequalities which are fully quantitative in nature. Thus,
one would expect it to be possible, at least in principle, to remove the reliance on
compactness methods and obtain a quantitative version of Theorem 1.1. This is
the purpose of the current paper. More precisely, in Section 6 we will establish the
following two results:
Theorem 1.2 (Quantitative regularity for critically bounded solutions). Let
u : [0, T ] × R3 → R3 , p : [0, T ] × R3 → R be a classical solution to the Navier-Stokes
equations with
(1.2) kukL∞ 3 3 ≤ A
t Lx ([0,T ]×R )
for some A ≥ 2. Then we have the derivative bounds

j+1
|∇jx u(t, x)| ≤ exp exp exp(AO(1) )t− 2
and
j+2
|∇jx ω(t, x)| ≤ exp exp exp(AO(1) )t− 2
whenever 0 < t ≤ T , x ∈ R3 , and j = 0, 1, where ω := ∇ × u is the vorticity field.

(See Section 2 for the asymptotic notation used in this paper.)
QUANTITATIVE BOUNDS FOR NAVIER-STOKES 3
Remark 1.3. It is not difficult to iterate using Schauder estimates in Hölder

spaces and extend the above regularity bounds to higher values of j than j = 0, 1
(allowing the implied constants in the O() notation to depend on j), and also control
time derivatives (conceding a factor of t−1 for each time derivative); we leave this
extension of Theorem 1.2 to the interested reader.
Theorem 1.4 (Quantitative blowup criterion). Let u : [0, T∗ ) × R3 → R3 ,
p : [0, T∗ ) × R3 → R be a classical solution to the Navier-Stokes equations which
blows up at a finite time 0 < T∗ < ∞. Then
ku(t)kL3x (R3 )
lim sup 1 c
= +∞
t→T∗− (log log log T∗ −t )
for an absolute constant c > 0.

We now discuss the method of proof of these theorems, which uses many of the
same key inputs as in previous arguments (most notably the Carleman estimates
used to prove backwards uniqueness and unique continuation), but also introduces
some other ingredients in order to avoid having to make some rather delicate results
from the qualitative theory (such as profile decompositions) quantitative, as doing
so would almost certainly lead to much poorer bounds than the ones given here.
The main estimate focuses on bounding the scale-invariant quantity
(1.3) N0−1 |PN0 u(t0 , x0 )|
for various points (t0 , x0 ) in spacetime, and various frequencies N0 , where PN0 is
a Littlewood-Paley projection operator to frequencies ∼ N0 (see Section 2 for a
precise definition). Using (1.2) and the Bernstein inequality, one can bound this
quantity by O(A). It is well known that if one could improve this bound somewhat
for sufficiently large N0 , for instance to O(A−C0 ) for a large constant C0 , then
(assuming A is large enough) the L3 norm becomes sufficiently “dispersed” in space
and frequency that one could adapt the local well-posedness theory for the Navier-
Stokes equation (or the local regularity theory from [CKN]) to obtain good bounds.
Hence we will focus on establishing such a bound for (1.3) for N0 large1 enough
(see Theorem 5.1 for a precise statement).
The first step in doing so is to observe (basically from the Duhamel formula
and some standard Littlewood-Paley theory) that if the quantity (1.3) is large for
some N0 , t0 , x0 with t0 not too close to the initial time 0, then the quantity
(1.4) N1−1 |PN1 u(t1 , x1 )|
is also large (with exactly the same lower bound) for some (t1 , x1 ) a little bit to
the past of (t0 , x0 ) (but more or less within the “parabolic domain of dependence”,
in the sense that x1 = x0 + O((t0 − t1 )1/2 )) and with N1 comparable to N0 ; see
Proposition 3.1(iv) for a precise statement. If one takes care to have exactly the
same lower bounds for both (1.3) and (1.4), then this claim can be iterated, creating
a chain of “bubbles of concentration” at various points (tn , xn ) and frequencies Nn ,
propagating backwards in time, and for which
Nn−1 |PNn u(tn , xn )|
1Strictly speaking, it is the scale-invariant quantity N 2 T that needs to be large, rather than
0
N0 itself, where T is the amount of time to the past of x0 for which the solution exists and obeys
the bounds (1.2).
4 TERENCE TAO
is bounded from below uniformly in n. Furthermore, by using a “bounded total

speed” property first observed in [T], one can ensure that (tn , xn ) stays in the
“parabolic domain of dependence” in the sense that xn = x0 + O((t0 − tn )1/2 ). Due
to the well known fact (dating back to the classical work of Leray [Le]) that solutions
to Navier-Stokes enjoy large “epochs of regularity” in which one has control of high
regularity norms of the solution in large time intervals outside of a small dimensional
singular set of times (see Proposition 3.1(iii) for a precise quantification of this
statement), one can show that there are a large number of points (tn , xn ) for which
the frequency Nn is basically as small as possible, in the sense that
Nn ∼ |t0 − tn |−1/2 .
The (Littlewood-Paley component PNn u of) the solution u is large near (tn , xn ),
and it is not difficult to then obtain analogous lower bounds on the vorticity
ω := ∇ × u
near (tn , xn ). The importance of working with the vorticity comes from the fact
that it obeys the vorticity equation
(1.5) ∂t ω = ∆ω − (u · ∇)ω + (ω · ∇)u
which can be viewed as a variable coefficient heat equation (in which the lower
order coefficients u, ∇u depend on the velocity field) for which the non-local effects
of the pressure p do not explicitly appear. Using a quantitative version of unique
continuation for backwards parabolic equations (see Proposition 4.3 for a precise
statement) that can be established using Carleman inequalities, one can then ob-
tain exponentially small, but still non-trivial, lower bounds2 for enstrophy-type
quantities such as Z Z
|ω(t, x)|2 dxdt
In 0
Rn ≤|x−xn |≤Rn
for various cylindrical annuli In × {x : Rn ≤ |x − xn | ≤ Rn0 } surrounding (tn , xn ),
with Rn0 a large multiple of Rn . Crucially, one can set Rn to be as large as one
pleases (although the lower bound exhibits Gaussian decay in Rn ). In order to apply
the Carleman inequalities, it is important that the time interval I lies within one
of the “epochs of regularity” in which one has good L∞ estimates for u, ∇u, ω, ∇ω,
but this can be accomplished without much difficulty (mainly thanks to the energy
dissipation term in the energy inequality).
For many choices of scale Rn (a bit larger than |t0 − tn |1/2 ), one can use an
“energy pigeonholing argument” (as used for instance by Bourgain [B]) to make
the energy (or more precisely, a certain component of the enstrophy) small in an
annular region {x : Rn ≤ |x − xn | ≤ Rn0 } at some time t0n a little bit to the past
of tn ; by modifying the somewhat delicate analysis of local enstrophies from [T]
that again takes advantage of the “bounded total speed” property, one can then
propagate this smallness forward in time (at the cost of shrinking the annular region
{Rn ≤ |x − xn | ≤ Rn0 } slightly), and in particular back up to time t0 , and parabolic
regularity theory can then be used to obtain good L∞ estimates for u, ∇u, ω, ∇ω
in these regions. This allows us to again use Carleman inequalities. Specifically,
2One can think of this as applying (a quantitative version) of unique continuation “in the
contrapositive”. Similarly for the invocation of backwards uniqueness below. Actually in practice
the Carleman inequalities also require an additional term such as |∇ω(t, x)|2 in the integrand, but
we ignore this term for sake of discussion.
Figure 1. A schematic depiction of the main argument. Start-

ing with a concentration of critical norm at a point (t0 , x0 ) in
spacetime, one propagates this concentration backwards in time
to generate concentrations at further points (tn , xn ) in spacetime.
Restricting attention to an epoch of regularity In × R3 (depicted
here in purple), Carleman estimates are then used to establish
lower bounds on the vorticity at other locations in space, and in
particular where the epoch intersects an “annulus of regularity”
(depicted in green) arising from an energy (or enstrophy) pigeon-
holing argument. A further application of Carleman estimates are
then used to establish a lower bound on the vorticity (or veloc-
ity) in the annular region at time t = t0 , thus demonstrating a
lack of compactness of the solution at this time which can be used
to obtain a contradiction when N0 (or more precisely the scale-
invariant quantity N02 T , where T is the lifespan of the solution) is
large enough, by letting n vary.
by using the Carleman inequalities used to prove the backwards uniqueness result
in [ESS2] (see Section 4 for precise statements), one can then propagate the lower
bounds on In × {x : Rn ≤ |x − xn | ≤ Rn0 } forward in time until one returns to the
original time t0 of interest, eventually obtaining a small but nontrivial lower bound
for quantities such as
Z
|ω(t0 , x)|2 dx
0
Rn ≤|x−xn |≤Rn
(ignoring for this discussion some slight adjustments to the scales Rn , Rn0 that occur
during this argument), which after some routine manipulations (and using the fact
that (tn , xn ) lies in the parabolic domain of dependence of (t0 , x0 )) also gives a
lower bound on quantities like
Z
|u(t0 , x)|3 dx.
0
Rn ≤|x−x0 |≤Rn
Crucially, this lower bound is uniform in n. If one now lets n vary, the annuli
{Rn ≤ |x − x0 | ≤ Rn0 } end up becoming disjoint for widely separated n, and one
can eventually contradict (1.2) at time t = t0 if N0 is large enough.
6 TERENCE TAO
Remark 1.5. The triply exponential nature of the bounds in Theorem 1.2
(which is of course closely tied to the triply logarithmic improvement to Theorem
1.1 in Theorem 1.4) can be explained as follows. One exponential factor comes
from the Bourgain energy pigeonholing argument to locate a good spatial scale
R. A second exponential factor arises from the Carleman inequalities. The third
exponential arises from locating enough disjoint spatial scales Rn to contradict
(1.2). It seems that substantially new ideas would be needed in order to improve
significantly upon this triple exponential bound.
Remark 1.6. Of course, by Sobolev embedding, the L3x (R3 ) norm in the above
1/2
theorems can be replaced by the critical homogeneous Sobolev norm Ḣx (R3 ). It is
likely that the arguments here can also be adapted to handle other critial Besov or
Lorentz spaces (as long as the secondary exponent of such spaces is finite, so that the
critical norm cannot simultaneously have a substantial presence at an unbounded
number of scales), but we will not pursue this question here; based on Theorem 1.4,
it is also reasonable to conjecture that the Orlicz norm ku(t)kL3 (log log log L)−c (R3 ) of
u also must blow up as t → T∗− for some absolute constant c > 0. On the other
hand, our argument relies heavily in many places on the fact that we are working
in three dimensions. It may be possible to obtain a higher-dimensional analogue
of our results by finding quantitative versions of the argument in [DD], but we do
not pursue this question here. Similarly, our arguments do not directly allow us to
replace the limit superior in Theorem 1.1 with a limit, as is done in [S1] (see also
[A]); again, it may be possible to also find quantitative analogues of these results,
but we do not pursue this matter here.
2. Notation
We use the notation X = O(Y ), X . Y , or Y & X to denote the bound
|X| ≤ CY for some absolute constant C > 0. If we need the implied constant C
to depend on parameters we shall indicate this by subscripts, for instance X .j Y
denotes the bound |X| ≤ Cj Y where Cj depends only on j.
Throughout this paper we will need a sufficiently large absolute constant C0 ,
which will remain fixed throughout the paper. For instance C0 = 105 would suffice
throughout our paper, if one worked out all the implied constants in the exponents
carefully.
If I ⊂ R is a time interval, we use |I| to denote its length. If x0 ∈ R3 and R > 0,
we use B(x0 , R) to denote the ball {x ∈ R3 : |x − x0 | ≤ R}, and if B = B(x0 , R) is
such a ball, we use kB = B(x0 , kR) to denote its dilates for any k > 0.
We use the mixed Lebesgue norms
Z 1/q
q
kukLqt Lrx (I×R3 ) := ku(t)kLr (R3 ) dt
x
I
where
Z 1/r
ku(t)kLrx (R3 ) := |u(t, x)|r dx
R3
with the usual modifications when q = ∞ or r = ∞. For any measurable subset
Ω ⊂ I × R3 , we write kukLqt Lrx (Ω) for ku1Ω kLqt Lrx (I×R3 ) , where 1Ω is the indicator
function of Ω.
Given a Schwartz function f : R3 → R, we define the Fourier transform

Z
fˆ(ξ) := f (x)e−2πiξ·x dx
R3
and then for any N > 0 we define the Littlewood-Paley projection P≤N by the
formula
ˆ
≤N f (ξ) := ϕ(ξ/N )f (ξ)
P\
where ϕ : R3 → R is a fixed bump function supported on B(0, 1) that equals 1 on
B(0, 1/2). We also define the companion Littlewood-Paley projections
PN := PN − PN/2
P>N := 1 − P≤N
P̃N := P2N − PN/4
P∞
where 1 denotes
P∞ the identity operator; thus for instance P≤N f = k=0 P2−k N f and
P>N f = k=1 P2k N f for Schwartz f (with the convergence in a locally uniform
sense). Also we have PN = PN P̃N . These operators can also be applied to vector-
valued Schwartz functions by working component by component. These operators
commute with other Fourier multipliers such as the Laplacian ∆ and its inverse
∆−1 , partial derivatives ∂i , heat propagators et∆ , and the Leray projection P :=
−∇ × ∆−1 ∇× to divergence-free vector fields. To estimate such multipliers, we use
the following general estimate:
Lemma 2.1 (Multiplier theorem). Let N > 0, and let m : R3 → C be a smooth

function supported on B(0, N ) that obeys the bounds
|∇j m(ξ)| ≤ M N −j
for all 0 ≤ j ≤ 100 and some M > 0. Let Tm denote the associated Fourier
multiplier, thus
m f (ξ) := m(ξ)f (ξ).
Td
Then one has
3 3
(2.1) kTm f kLq (R3 ) . M N p − q kf kLp (R3 )
whenever 1 ≤ p ≤ q ≤ ∞ and f : R3 → R is a Schwartz function. More generally, if

Ω ⊂ R3 is an open subset of R3 , A ≥ 1, and ΩA/N := {x ∈ R3 : dist(x, Ω) < A/N }
denotes the A/N -neighbourhood of Ω, then we have a local version
(2.2)
3 3 1 1 3 3
kTm f kLq1 (Ω) . M N p1 − q1 kf kLp1 (ΩA/N ) + A−50 M |Ω| q1 − q2 N p2 − q2 kf kLp2 (R3 )
of the above estimate, whenever 1 ≤ p1 ≤ q1 ≤ ∞ and 1 ≤ p2 ≤ q2 ≤ ∞ are such

that q2 ≥ q1 , and |Ω| denotes the volume of Ω.
By the usual limiting arguments, one can replace the hypothesis that f is
Schwartz with the requirement that f lie in Lp . Also one can extend this theorem
to vector-valued f : R3 → R3 by working component by component. In practice,
the A−50 factor will ensure that the second term on the right-hand side of (2.2) is
negligible compared to the first, and can be ignored on a first reading.
8 TERENCE TAO
Proof. By homogeneity we can normalise M = 1; by scaling (or dimensional

analysis) we may also normalise N = 1. We can write Tm f as a convolution
Tm f = f ∗ K of f with the kernel
Z
K(x) := m(ξ)e2πiξ·x dξ.
R3
By repeated integration by parts we obtain the bounds K(x) . (1 + |x|)−90 (say),
so in particular kKkLr (R3 ) . 1 for all 1 ≤ r ≤ ∞. From Young’s convolution
inequality we then conclude that
kTm f kLq (R3 ) . kf kLp (R3 )
giving (2.1). To prove (2.2), we see that the claim already follows from (2.1)
when f is supported in Lp (ΩA ), so by the triangle inequality we may assume that
f is supported on R3 \ΩA . In this case we may replace the convolution kernel
K by its restriction to the complement of B(0, A), which allows us to improve
the bound on the Lr norm of the kernel to (say) O(A−50 ). The claim follows
from Young’s convolution inequality, after first using Hölder’s inequality to bound
1 1
kTm f kLq1 (Ω) ≤ |Ω| q1 − q2 kTm f kLq2 (Ω) .
Thus for instance, we have the Bernstein inequalities
3 3
(2.3) k∇j f kLq (R3 ) .j N j+ p − q kf kLp (R3 )
whenever 1 ≤ p ≤ q ≤ ∞, j ≥ 0, and f is a Schwartz function whose Fourier
transform is supported on B(0, N ), as can be seen by writing f = P≤2N f and
applying Lemma 2.1. In a similar spirit, one has
3 3
(2.4) kPN et∆ ∇j f kLq (R3 ) .j exp(−N 2 t/20)N j+ p − q kf kLp (R3 )
for any t > 0 and any Schwartz f . Summing this, we obtain the standard heat
kernel bounds
j 3 3
(2.5) ket∆ ∇j f kLq (R3 ) .j t− 2 − 2p + 2q kf kLp (R3 ) .
3. Basic estimates
The purpose of this section is to establish the following initial bounds for L∞ 3
t Lx -
bounded solutions to the Navier-Stokes equations.
Proposition 3.1 (Initial estimates). Let u : [t0 − T, t0 ] × R3 → R3 , p : [t0 −
T, t0 ] × R3 → R be a classical solution to Navier-Stokes that obeys the bound
(3.1) kukL∞ 3 3 ≤ A.
t Lx ([t0 −T,t0 ]×R )
for some A ≥ C0 . We adopt the notation

j
Aj := AC0
for all j, thus A0 = A and Aj+1 = AC
j .
0
(i) (Pointwise derivative estimates) For any (t, x) ∈ [t0 − T /2, t0 ] × R3 and
N > 0, we have
(3.2) PN u(t, x) = O(AN ); ∇PN u(t, x) = O(AN 2 ); ∂t PN u(t, x) = O(A2 N 3 );
similarly, the vorticity ω := ∇ × u obeys the bounds
(3.3)
PN ω(t, x) = O(AN 2 ); ∇PN ω(t, x) = O(AN 3 ); ∂t PN ω(t, x) = O(A2 N 4 ).
(ii) (Bounded total speed) For any interval I in [t0 − T /2, t0 ], one has
4 1/2
(3.4) kukL1t L∞ 3 . A |I|
x (I×R )
.
(iii) (Epochs of regularity) For any interval I in [t0 − T /2, t0 ], there is a subin-
terval I 0 ⊂ I with |I 0 | & A−8 |I| such that
k∇j ukL∞ ∞ 0 3 . A
t Lx (I ×R )
O(1)
|I|−(j+1)/2
and
k∇j ωkL∞ ∞ 0 3 . A
t Lx (I ×R )
O(1)
|I|−(j+2)/2
for j = 0, 1.
(iv) (Back propagation) Let (t1 , x1 ) ∈ [t0 − T /2, t0 ] × R3 and N1 ≥ A3 T −1/2
be such that
(3.5) |PN1 u(t1 , x1 )| ≥ A−1
1 N1 .
Then there exists (t2 , x2 ) ∈ [t0 − T, t1 ] × R3 and N2 ∈ [A−1

2 N1 , A2 N1 ] such
that
A−1
3 N1
−2
≤ t1 − t2 ≤ A3 N1−2
and
|x2 − x1 | ≤ A4 N1−1
and
(3.6) |PN2 u(t2 , x2 )| ≥ A−1
1 N2 .
(v) (Iterated back propagation) Let x0 ∈ R3 and N0 > 0 be such that

|PN0 u(t0 , x0 )| ≥ A−1
1 N0 .
Then for every A4 N0−2 ≤ T1 ≤ A−1

4 T , there exists
(t1 , x1 ) ∈ [t0 − T1 , t0 − A−1

3 T1 ] × R
3
and
O(1) −1/2
N1 = A3 T1
such that
O(1) 1/2
x1 = x0 + O(A4 T1 )
and
|PN1 u(t1 , x1 )| ≥ A−1
1 N1 .
(vi) (Annuli of regularity) If 0 < T 0 < T /2, x0 ∈ R3 , and R0 ≥ (T 0 )1/2 , then
there exists a scale
O(1)
R0 ≤ R ≤ exp(A6 )R0
such that on the region
Ω := {(t, x) ∈ [t0 − T 0 , t0 ] × R3 : R ≤ |x − x0 | ≤ A6 R}
we have
k∇j ukL∞ ∞
t Lx (Ω)
. A−2 0 −(j+1)/2
6 (T )
and
k∇j ωkL∞ ∞
t Lx (Ω)
. A−2 0 −(j+2)/2
6 (T )
for j = 0, 1.
10 TERENCE TAO
As C0 is assumed large, any polynomial combination of A = A0 , A1 , . . . , Aj−1

will be dominated by Aj for any j ≥ 1; we take advantage of this fact without
comment in the sequel to simplify the estimates. The various numerical powers of
A (or Aj ) that appear in the above proposition are not of much significance, except
that it is important for iterative purposes that the negative power A−1 1 appearing
in (3.5) is exactly the same as the one appearing in (3.6).
In the remainder of this section t0 , T, A, u, p are as in Proposition 3.1. Our
objective is now to establish the claims (i)-(vi).
We begin with the proof of (i). It suffices to establish (3.2), as (3.3) then follows
from the Bernstein inequalities (2.3). The first two claims of (3.2) are immediate
from (3.1) and (2.3). For the final claim, we first apply the Leray projection P to
(1.1) to obtain the familiar equation
(3.7) ∂t u = ∆u − P∇ · (u ⊗ u)
where the divergence ∇ · (u ⊗ u) of the symmetric tensor u ⊗ u is expressed in
coordinates as
(∇ · (u ⊗ u))i = ∂j (ui uj )
with the usual summation conventions. We apply PN to both sides of (3.7). From
(3.1) and (2.3) we have
3
kPN ∆u(t)kL∞ 3 . N A.
x (R )
From (3.1) and Hölder we have ku ⊗ u(t)kL3/2 (R3 ) . A2 , hence by Lemma 2.1 we
x
have
3 2
kPN P∇ · (u ⊗ u)(t)kL∞ 3 . N A ,
x (R )
and the final claim of (3.2) follows from the triangle inequality.
Now we prove (ii), (iii). It is not difficult to see that these estimates are
invariant with respect to time translation (shifting I, t0 , u accordingly) and also
rescaling (adjusting T, t0 , I, u accordingly). Hence we may assume without loss of
generality that I = [0, 1] ⊂ [t0 − T /2, t0 ], which implies that [−1, 1] ⊂ [t0 − T, t0 ].
It will be convenient to remove3 a linear component from u, as it is not well
controlled in L2x type spaces. Namely, on [−1, 1] × R3 we split u = ulin + unlin ,
where ulin is the linear solution
(3.8) ulin (t) := e(t+1)∆ u(−1)
and unlin := u − ulin is the nonlinear component. From (3.1) we have
(3.9) kulin kL∞ 3 3 , ku
t Lx ([−1,1]×R )
nlin
kL∞ 3 3 . A.
t Lx ([−1,1]×R )
From (3.7) and Duhamel’s formula one has

Z t
0
nlin
u (t) = − e(t−t )∆ P∇ · (u ⊗ u)(t0 ) dt0 .
−1
3/2
From (3.1), u ⊗ u has an Lx (R3 )
norm of O(A2 ). From (2.5), the operator
0 3/2
e(t−t )∆ P∇· maps Lx to L2x with an operator norm of (t−t0 )−3/4 . From Minkowski’s
inequality we conclude an energy bound for the nonlinear component:
(3.10) kunlin kL∞ 2
2
3 . A .
t Lx ([−1,1]×R )
3See also [C] for a similar technique to apply energy methods to Navier-Stokes solutions that
lie in a function space other than L2x .
We now restrict attention to the slab [−1/2, 1]×R3 . Here t+1 lies between 1/2 and
2, and we can use (3.1), (3.8), and (2.5) to obtain very good bounds on ulin (but
only in spaces with an integrability exponent greater than or equal to 3). More
precisely, we have
(3.11) k∇j ulin kL∞ p 3 .j A
t Lx ([−1/2,1]×R )
for any 3 ≤ p ≤ ∞ and j ≥ 0.

To exploit the bound (3.10), we use the energy method. Since ulin solves the
heat equation ∂t ulin = ∆ulin , we can subtract this from (1.1) to conclude that
(3.12) ∂t unlin = ∆unlin − ∇ · (u ⊗ u) − ∇p.
Taking inner products with unlin , which is divergence-free, and integrating by parts,
we conclude that
Z Z Z
1
∂t |unlin |2 dx = − |∇unlin |2 dx + (∇unlin ) · (u ⊗ u) dx
2 R3 R3 R3
where the quantity (∇unlin ) · (u ⊗ u) is defined in coordinates as

(∇unlin ) · (u ⊗ u) := (∂i unlin
j )ui uj .
From the divergence-free nature of unlin and integration by parts we have
Z
(∇unlin ) · (unlin ⊗ unlin ) dx = 0
R3
and hence
Z Z Z
1
∂t |unlin |2 dx = − |∇unlin |2 dx + (∇unlin ) · (u ⊗ u − unlin ⊗ unlin ) dx.
2 R3 R3 R3
Integrating this on [−1/2, 1] using (3.10) we conclude that
Z 1 Z Z 1 Z
nlin 2 2
|∇u | dxdt . A + |∇unlin ||u ⊗ u − unlin ⊗ unlin | dxdt,
−1/2 R3 −1/2 R3
and hence by Young’s inequality

Z 1 Z Z 1 Z
nlin 2 2
|∇u | dxdt . A + |u ⊗ u − unlin ⊗ unlin |2 dxdt.
−1/2 R3 −1/2 R3
Splitting u ⊗ u − unlin ⊗ unlin = ul ⊗ u + unlin ⊗ ul and using (3.1), (3.9), (3.11)

(with p = 6, j = 0) and Hölder’s inequality, one has
Z 1 Z
|u ⊗ u − unlin ⊗ unlin |2 dxdt . A4
−1/2 R3
and thus
Z 1 Z
(3.13) |∇unlin |2 dxdt . A4 .
−1/2 R3
By Plancherel’s theorem this implies in particular that

X
(3.14) N 2 kPN unlin k2L2 L2 ([−1/2,1]×R3 ) . A4
t x
N
where N ranges over powers of two. Also, from Sobolev embedding one has
(3.15) kunlin kL2t L6x ([−1/2,1]×R3 ) . A2 .
12 TERENCE TAO
We are now ready to establish the bounded total speed property (ii), which is a
variant of [T, Proposition 9.1]. If t ∈ [0, 1] and N ≥ 1 is a power of two, we see
from (3.7) and Duhamel’s formula that
Z t
(t+ 21 )∆ 1 0
nlin
PN u (t) = e PN unlin
− − PN e(t−t )∆ P∇ · P̃N (u ⊗ u)(t0 ) dt0 .
2 −1/2
0
From (2.4) the operator PN e(t−t )∆ P∇· has an operator norm of O(N exp(−N 2 (t −
(t+ 12 )∆
t0 )/20)) on L∞
x , while from (3.9), (2.4) we see that e PN v(− 21 ) has an L∞
x
norm of O(AN exp(−N 2 /20)). Thus by Young’s inequality
−1
kPN unlin kL1t L∞ 2
3 . AN exp(−N /20) + N
x ([0,1]×R )
kP̃N (u ⊗ u)kL1t L∞ 3 .
x ([−1/2,1]×R )
We split u ⊗ u = ulin ⊗ ulin + ulin ⊗ unlin + unlin ⊗ ulin + unlin ⊗ unlin . From (3.11)
one has
kP̃N (ulin ⊗ ulin )kL1t L∞ 2
3 . A .
x ([−1/2,1]×R )
From (2.3), Hölder’s inequality, and (3.11), (3.15) one has

kP̃N (ulin ⊗ unlin )kL1t L∞ 3 . N
x ([−1/2,1]×R )
1/2
kulin ⊗ unlin kL1t L6x ([−1/2,1]×R3 )
. A3 N 1/2 .
Similarly with ulin ⊗ unlin replaced by unlin ⊗ ulin . We then split unlin ⊗ unlin =
P≤N unlin ⊗ P≤N unlin + P≤N unlin ⊗ P>N unlin + P>N unlin ⊗ P≤N unlin + P>N unlin ⊗
P>N unlin . We have from Hölder that
kP≤N unlin ⊗ P≤N unlin kL1t L∞ 3 . kP≤N u
x ([−1/2,1]×R )
nlin 2
kL2 L∞ ([−1/2,1]×R3 )
t x
kP≤N unlin ⊗ P>N unlin kL1t L2x ([−1/2,1]×R3 ) ,

kP>N unlin ⊗ P≤N unlin kL1t L2x ([−1/2,1]×R3 ) . kP≤N unlin kL2t L∞ 3
x ([−1/2,1]×R )
× kP>N unlin kL2t L2x ([−1/2,1]×R3 )

kP>N unlin ⊗ P>N unlin kL1t L1x ([−1/2,1]×R3 ) . kP>N unlin k2L2 L2 ([−1/2,1]×R3 )
t x
and hence by (2.3), the triangle inequality, and Young’s inequality

nlin 2
kP̃N (u⊗u)kL1t L∞ 3 . kP≤N u
x ([−1/2,1]×R )
kL2 L∞ ([−1/2,1]×R3 ) +N 3 kP>N unlin k2L2 L2 ([−1/2,1]×R3 ) .
t x t x
Putting all this together, we conclude that

3 −1/2
kPN unlin kL1t L∞ 3 . A N
x ([0,1]×R )
+ N −1 kP≤N unlin k2L2 L∞ ([−1/2,1]×R3 )

t x
+ N 2 kP>N unlin k2L2 L2 ([−1/2,1]×R3 ) .

t x
By (2.3) and Cauchy-Schwarz we have

 2
X
kP≤N unlin k2L2 L∞ ([−1/2,1]×R3 ) .  (N 0 )3/2 kPN 0 unlin kL2t L2x ([−1/2,1]×R3 ) 
t x
N 0 ≤N
X
. N 1/2 (N 0 )2 kPN 0 unlin k2L2 L2 ([−1/2,1]×R3 )
t x
N 0 ≤N
where N 0 ranges over powers of two, while from Plancherel’s theorem one has
X
kP>N unlin k2L2 L2 ([−1/2,1]×R3 ) . kPN 0 unlin k2L2 L2 ([−1/2,1]×R3 ) .
t x t x
N 0 >N
Summing in N , and using the triangle inequality followed by (3.14), we conclude

that
X
kP≥1 unlin kL1t L∞ 3
3 . A +
x ([0,1]×R )
(N 0 )2 kPN 0 unlin k2L2 L2 ([−1/2,1]×R3 ) . A4 .
t x
N0
From (3.9) and (2.3) we also have

kulin kL1t L∞ 3 , kP<1 u
x ([0,1]×R )
nlin
kL1t L∞ 3 . A
x ([0,1]×R )
and we conclude
4
kukL1t L∞ 3 . A
x ([0,1]×R )
which gives (ii).

Now we establish (iii). For t ∈ [0, 1] we define the enstrophy-type quantity
Z
1
E(t) := |∇unlin (t, x)|2 dx,
2 R3
Taking the gradient of (3.12) and then taking the inner product with ∇unlin , we
see upon integration by parts that
Z Z
∂t E(t) = − |∇2 unlin |2 dx + ∆unlin · (∇ · (u ⊗ u)) dx
R3 R3
1
∂t E(t) ≤ − k∇2 unlin k2L2x (R3 ) + O k∇ · (u ⊗ u)k2L2x (R3 ) .
2
By the Leibniz rule and Hölder’s inequality, one has
k∇ · (u ⊗ u)kL2x (R3 ) . kukL6x (R3 ) k∇ukL3x (R3 ) .
From (3.11) and the triangle inequality one has
kukL6x (R3 ) . A + kunlin kL6x (R3 )
and
k∇ukL3x (R3 ) . A + k∇unlin kL3x (R3 )
while from Sobolev embedding and Hölder one has
kunlin kL6x (R3 ) . k∇unlin kL2x (R3 ) . E(t)1/2
and
1/2 1/2 1/2
k∇unlin kL3x (R3 ) . k∇unlin kL2 (R3 ) k∇2 unlin kL2 (R3 ) . E(t)1/4 k∇2 unlin kL2 (R3 ) .
x x x
We conclude that
1
∂t E(t) ≤ − k∇2 unlin k2L2x (R3 ) + O (A2 + E(t))(A2 + E(t)1/2 k∇2 unlin kL2x (R3 ) )
2
1
(3.16) ∂t E(t) ≤ − k∇2 unlin k2L2x (R3 ) + O((A2 + E(t))A2 + (A2 + E(t))2 E(t)).
4
In particular we have
(3.17) ∂t E(t) ≤ O(A4 + A4 E(t) + E(t)3 ).
From (3.13) we have
Z 1
E(t) dt . A4 ,
0
14 TERENCE TAO
and hence by the pigeonhole principle, we can find a time t1 ∈ [0, 1/2] such that
E(t1 ) . A4 .
A standard continuity argument using (3.17) then gives E(t) . A4 for t ∈ [t1 , t1 +
cA−8 ] = [τ (0), τ (1)], where τ (s) := t1 + scA−8 and c > 0 is a small absolute
constant. Inserting this back into (3.16) one has
1
∂t E(t) ≤ − k∇2 unlin k2L2x (R3 ) + O(A12 )
4
and hence by the fundamental theorem of calculus
Z τ (1) Z
(3.18) |∇2 unlin |2 dxdt . A4 .
τ (0) R3
Thus we have
(3.19) k∇unlin kL∞ 2
2 nlin
3 + k∇ u
t Lx ([τ (0),τ (1)]×R )
kL2t L2x ([τ (0),τ (1)]×R3 ) . A2 .
From the Gagliardo-Nirenberg inequality
1/2 1/2
(3.20) kunlin kL∞
x
. k∇unlin kL2 k∇2 unlin kL2
x x
and Hölder’s inequality, one concludes in particular that

kunlin kL4t L∞ 3 . A
x ([τ (0),τ (1)]×R )
2
and hence by (3.11)

2
(3.21) kukL4t L∞ 3 . A ;
x ([τ (0),τ (1)]×R )
also from Sobolev embedding and (3.19) one has

k∇unlin kL2t L6x ([τ (0),τ (1)]×R3 ) . A2
and hence by (3.11)
(3.22) k∇ukL2t L6x ([τ (0),τ (1)]×R3 ) . A2 .
These are subcritical regularity estimates and can now be iterated to obtain even
higher regularity. For t ∈ [τ (0.1), τ (1)], we see from (3.7) that
Z t
0
(3.23) u(t) = e(t−τ (0))∆ u(τ (0)) − e(t−t )∆ P∇ · (u ⊗ u)(t0 ) dt0 .
τ (0)
0
From (2.5) the operator e(t−t )∆ P∇· has norm O((t−t0 )−1/2 ) on L∞
x , while e
(t−τ (0))∆
3 ∞ −1/2 O(1)
maps Lx to Lx with norm O((t − τ (0)) ) = O(A ). We conclude from (3.1)
that Z t
ku(t)kL∞ 3 . A
x (R )
O(1)
+ (t − t0 )−1/2 ku(t)k2L∞ 0
3 dt .
x (R )
τ (0)
From (3.21) and Young’s convolution inequality, we conclude that
O(1)
kukL8t L∞ 3 . A
x ([τ (0.1),τ (1)]×R )
Repeating the above argument, we now also see for t ∈ [τ (0.2), τ (1)] that
Z t
ku(t)kL∞
x (R 3) . A
O(1)
+ (t − t0 )−1/2 ku(t)k2L∞ 3 dt
x (R )
0
τ (0.1)
so from Hölder’s inequality we conclude that

O(1)
(3.24) kukL∞ ∞ 3 . A
t Lx ([τ (0.2),τ (1)]×R )
.
Now we differentiate (3.7) to conclude that

Z t
0
∇u(t) = ∇e(t−τ (0.2))∆ u(τ (0.2)) − ∇e(t−t )∆ P∇ · (u ⊗ u)(t0 ) dt0
τ (0.2)
for t ∈ [τ (0.3), τ (1)]. From (3.1), the first term ∇e(t−τ (0.2))∆ u(τ (0.2)) has an L∞ x
0
norm of O(AO(1) ). From (2.5), the operator ∇e(t−t )∆ P maps L6x to L∞ x with norm
O((t − t0 )−3/4 ), thus
Z t
k∇u(t)kL∞ 3 . A
x (R )
O(1)
+ (t − t0 )−3/4 k∇ · (u ⊗ u)(t0 )kL6x (R3 ) dt0 .
τ (0.2)
From (3.22), (3.24), Leibniz and Hölder one has

k∇ · (u ⊗ u)kL2t L6x ([τ (0.2),τ (1)]×R3 ) . AO(1)
and hence by fractional integration
O(1)
k∇ukL4t L∞ 3 . A
x ([τ (0.3),τ (1)]×R )
.
From this, (3.24), Leibniz, and Hölder one has
O(1)
k∇ · (u ⊗ u)kL4t L∞ 3 . A
x ([τ (0.3),τ (1)]×R )
.
0
By (2.5), ∇e(t−t )∆ P has an operator norm of O((t − t0 )−1/2 ) on L∞ x , thus
Z t
k∇u(t)kL∞ 3 . A
x (R )
O(1)
+ (t − t0 )−1/2 k∇ · (u ⊗ u)(t0 )kL∞ 3 dt
x (R )
0
τ (0.3)
for t ∈ [τ (0.4), τ (1)], and hence by Hölder’s inequality

O(1)
k∇ukL∞ ∞ 3 . A
t Lx ([τ (0.4),τ (1)]×R )
.
From the vorticity equation (1.5), we now have
∂t ω = ∆ω + O(AO(1) (|ω| + |∇ω|))
on [τ (0.4), τ (1)] × R3 , and also ω = O(AO(1) ) on this slab. Standard parabolic
regularity estimates (see e.g., [LSU]) then give
O(1)
k∇ωkL∞ ∞ 3 . A
t Lx ([τ (0.5),τ (1)]×R )
.
Setting I 0 := [τ (0.5), τ (1)], we obtain the claim (iii). We remark that it is also
possible to control higher derivatives ∇j u, ∇j ω with j > 1, for instance by using
parabolic Schauder estimates in Hölder spaces, but we will not need to do so here.
Now we establish (iv). Let t1 , x1 , N1 be as in that part of the proposition. By
rescaling we may normalise N1 = 1, and by translation invariance we may normalise
A2
(t1 , x1 ) = (0, 0), so that t0 −T ≤ − T2 ≤ − 23 , so in particular [−2A3 , 0] ⊂ [t0 −T, t0 ].
From (3.5) we have
(3.25) |P1 u(0, 0)| ≥ A−1
1 .
Assume for contradiction that the claim fails, then we have

kPN ukL∞ L∞ ([−A3 ,−A−1 ]×B(0,A4 )) ≤ A−1
1 N
t x 3
for all A−1

2 ≤ N ≤ A2 . From (3.2) and the fundamental theorem of calculus in
time, we can enlarge the time interval to reach t = 0, so that
kPN ukL∞ ∞
t Lx ([−A3 ,0]×B(0,A4 ))
. A−1
1 N.
16 TERENCE TAO
Suppose now that N ≥ A−1 2 . For t ∈ [−A3 , 0], we can use Duhamel’s formula,
(3.7), and the triangle inequality to write
kPN u(t)kL3/2 (B(0,A4 )) ≤ ke(t+2A3 )∆ PN u(−2A3 )kL3/2 (B(0,A4 ))
x x
Z t
0
+ ke(t−t )∆ PN ∇ · (u(t0 ) ⊗ u(t0 ))kL3/2 (R3 ) dt0 .
x
−2A3
From (2.4), e(t+2A3 )∆ PN has an operator norm of O(exp(−N 2 A3 /20)) on L3x , and
0 3/2
e(t−t )∆ PN ∇· similarly has an operator norm of O(N exp(−N 2 (t − t0 )/20)) on Lx .
Applying (3.1) and Hölder’s inequality, we conclude that
kPN u(t)kL3/2 (B(0,A4 )) . AA4 exp(−N 2 A3 /20) + A2 N −1
x
and hence in the range N ≥ A−1

2 we have
(3.26) kPN ukL∞ L3/2 ([−A3 ,0]×B(0,A4 )) . A2 N −1 .

t x
−1/2
Now suppose that N ≥ For t ∈ [−A3 /2, 0], we again use Duhamel’s
A2 .
formula, (3.7) and the triangle inequality to write
kPN u(t)kL1x (B(0,A4 /2)) ≤ ke(t+A3 )∆ PN u(−A3 )kL1x (B(0,A4 /2))
Z t
0
+ ke(t−t )∆ PN ∇ · P̃N (u(t0 ) ⊗ u(t0 ))kL1x (B(0,A4 /2)) dt0 .
−A3
From (2.4), (3.1), and Hölder as before we have

ke(t+A3 )∆ PN u(−A3 )kL1x (B(0,A4 /2)) . AA24 exp(−N 2 A3 /40).
From (2.2) one has
0
ke(t−t )∆ PN ∇ · P̃N (u(t0 ) ⊗ u(t0 ))kL1x (B(0,A4 /2)) . N exp(−N 2 (t − t0 )/20)

1/2
× kP̃N (u(t0 ) ⊗ u(t0 ))kL1x (B(0,3A4 /4) + A−504 A 4 kP̃ N (u(t 0
) ⊗ u(t 0
))kL
3/2
(R3 )
x
and hence by (3.1)

kPN ukL∞ 1
t Lx ([−A3 /2,0]×B(0,A4 /2))
. A−40
4 +N −1 kP̃N (u(t0 )⊗u(t0 ))kL∞ 1
t Lx ([−A3 ,0]×B(0,3A4 /4))
.
Since P̃N (P≤N/100 u(t0 ) ⊗ P≤N/100 u(t0 )) vanishes, we can write

(3.27)
P̃N (u(t0 ) ⊗ u(t0 )) = P̃N (P>N/100 u(t0 ) ⊗ u(t0 )) + P̃N (P≤N/100 u(t0 ) ⊗ P>N/100 u(t0 )).
From (2.2), (3.1) we have
kP̃N (P>N/100 u(t0 )⊗u(t0 ))kL∞ 1
t Lx ([−A3 ,0]×B(0,3A4 /4))
. kP>N/100 u(t0 )⊗u(t0 )kL∞ 1
t Lx ([−A3 ,0]×B(0,A4 ))
+A−40
4 .
From (3.26) (and the triangle inequality) as well as (3.1) and Hölder’s inequality,
we thus have
kP̃N (P>N/100 u(t0 ) ⊗ u(t0 ))kL∞ 1
t Lx ([−A3 ,0]×B(0,3A4 /4))
. A3 N −1 .
Similarly for the other component of (3.27). We conclude that
(3.28) kPN ukL∞ 1
t Lx ([−A3 /2,0]×B(0,A4 /2))
. A3 N −2
−1/2
for all N ≥ A2 .
−1/3 1/3
Now suppose that A2 ≤ N ≤ A2 . For t ∈ [−A3 /3, 0], we again use
Duhamel’s formula, (3.7), and the triangle inequality as before to write
kPN u(t)kL2x (B(0,A4 /4)) ≤ ke(t+A3 /2)∆ PN u(−A3 /2)kL2x (B(0,A4 /4))
Z t
0
+ ke(t−t )∆ PN ∇ · P̃N (u(t0 ) ⊗ u(t0 ))kL2x (B(0,A4 /4)) dt0 .
−A3 /2
Arguing as before we have

1/2
ke(t+A3 /2)∆ PN u(−A3 /2)kL2x (B(0,A4 /4)) . AA4 exp(−N 2 A3 /120)
and
0
ke(t−t )∆ PN ∇ · P̃N (u(t0 ) ⊗ u(t0 ))kL2x (B(0,A4 /4)) . N 5/2 exp(−N 2 (t − t0 )/20)

kP̃N (u(t0 ) ⊗ u(t0 ))kL1x (B(0,A4 /3)) + A−50
4 N −1 kP̃N (u(t0 ) ⊗ u(t0 ))kL3/2 (R3 )
x
and thus
(3.29)
kPN ukL∞ 2
t Lx ([−A3 /4,0]×B(0,A4 /4))
. A−40
4 +N 1/2 kP̃N (u(t0 )⊗u(t0 ))kL∞ 1
t Lx ([−A3 /2,0]×B(0,A4 /3))
.
We can split P̃N (u(t0 )⊗u(t0 )) into O(1) paraproduct terms of the form P̃N (PN 0 u(t0 )⊗
P≤N/100 u(t0 )) where N 0 ∼ N , O(1) terms of the form P̃N (P≤N/100 u(t0 ) ⊗ PN 0 u(t0 )),
and a sum of the form N1 ∼N2 &N P̃N (PN1 u(t0 ) ⊗ PN2 u(t0 )). For the “high-low”
P
term P̃N (PN 0 u(t0 ) ⊗ P≤N/100 u(t0 )), we observe from (3.26), (3.2) and the triangle
inequality that
kP≤N/100 ukL∞ L3/2 ([−A3 ,0]×B(0,A4 )) . A2 N −1 .
t x
Using this, (2.2), (3.28) (for the high frequency factor PN 0 u(t0 )), and Hölder’s in-
equality, we conclude that the contribution of this term to (3.29) is O(A3 A−1 1 N
−1/2
).
0 0
Similarly for the “low-high” term P̃N (P≤N/100 u(t ) ⊗ PN 0 u(t )). Finally, to control
the “high-high” term N1 ∼N2 &N P̃N (PN1 u(t0 ) ⊗ PN2 u(t0 )), we use (2.2), the trian-
P
gle inequality, Hölder, and (3.28) to control this contribution by

X
. A−40
4 + N 1/2 A3 N1−2 kPN2 ukL∞ L3/2 ([−A3 ,0]×B(0,A4 )) .
t x
N1 ∼N2 &N
Using (3.26) when N2 ≤ A2 and (3.2) otherwise, we see that this term also con-
tributes O(A3 A−1
1 N
−1/2
). We have thus shown that
(3.30) kPN ukL∞ 2
t Lx ([−A3 /4,0]×B(0,A4 /4))
. A3 A−1
1 N
−1/2
−1/3 1/3
for A2 ≤ N ≤ A2 .
We now return once again to Duhamel’s formula to estimate
Z 0
0
|P1 u(0, 0)| ≤ |e A3 ∆/4
P1 u(−A3 /4)|(0)+ |e(t−t )∆ P1 ∇· P̃1 (u(t0 )⊗u(t0 ))|(0) dt0 .
−A3 /4
From (2.4), (3.1), the first term is O(A3 exp(−A23 /320)), thus from (3.25) we have
Z 0
0
|e(t−t )∆ P1 ∇ · P̃1 (u(t0 ) ⊗ u(t0 ))|(0) dt0 & A−1
1 .
−A3 /4
From (2.2), (3.1) one has

0
|e(t−t )∆ P1 ∇·P̃1 (u(t0 )⊗u(t0 ))|(0) . exp(−(t−t0 )/20)(kkP̃1 (u(t0 )⊗u(t0 ))kL1x (B(0,A1 )) +A−50
1 )
18 TERENCE TAO
and hence by the pigeonhole principle we have

kP̃1 (u(t0 ) ⊗ u(t0 ))kL1x (B(0,A1 )) & A−1
1 .
for some −A3 /4 ≤ t0 ≤ 0.

Fix this t0 . As before, we can split P̃1 (u(t0 ) ⊗ u(t0 )) into the sum of O(1) “low-
high” terms P̃1 (PN 0 u(t0 ) ⊗ P≤1/100 u(t0 )) and “high-low” terms P̃1 (P≤1/100 u(t0 ) ⊗
PN 0 u(t0 )) with N 0 ∼ 1, plus a “high-high” term N1 ∼N2 &1 P̃1 (PN1 u(t0 )⊗PN2 u(t0 )).
P
−1/3
For the first two types of terms, we use (2.2) (for frequencies larger than A2 ),
(3.1), and Hölder to conclude that
kP≤100 u(t0 )kL2x (B(0,2A1 )) . A3 A−1
1
and then from (3.30), (2.2) (and (3.1) to control the global contribution of (2.2))
we see that the contribution of those two types of terms is O(A6 A−2
1 ). For the high-
1/3
high terms with N1 , N2 ≤ A2 , we again use (3.30), (2.2), (3.1) to again obtain
1/3
a bound of O(A6 A−21 ). For the cases when N1 , N2 & A2 , we use (3.26), (3.1) to
−1/3
obtain a much better bound O(A3 A2 ). Putting all this together we obtain
A−1 6 −2
1 . A A1
giving the required contradiction. This establishes (iv).

Now we prove (v). We may assume that A4 N0−2 ≤ A−1 4 T , since the claim is
trivial otherwise. Thus we have N0 ≥ A4 T −1/2 .
By iteratively applying (iv), we may find a sequence (t0 , x0 ), (t1 , x1 ), . . . , (tn , xn ) ∈
[t0 − T, t0 ] and N0 , N1 , . . . , Nn > 0 for some n ≥ 1, with the properties
(3.31) |PNi u(ti , xi )| ≥ A−1
1 Ni
(3.32) A−1
2 Ni−1 ≤ Ni ≤ A2 Ni−1
(3.33) A−1 −2 −2
3 Ni−1 ≤ ti−1 − ti ≤ A3 Ni−1
−1
(3.34) |xi − xi−1 | ≤ A4 Ni−1
for all i = 1, . . . , n, with ti ∈ [t0 − T /2, t0 ] and Ni ≥ A3 T −1/2 for i = 0, . . . , n − 1

and either tn ∈ [t0 − T, t0 − T /2] or Nn < A3 T −1/2 . To see that this process
terminates at a finite n, observe from the classical nature of u that the PNi u(ti , xi )
are uniformly bounded in i, which by (3.31) implies that the Ni are uniformly
bounded above, and hence by (3.33) ti−1 − ti are uniformly bounded below; since ti
must stay above t0 − T , we obtain the required finite time termination. By (3.33),
the first time t1 after t0 lies in the interval
t1 ∈ [t0 − A2 N0−2 , t0 − A−1 −2
2 N0 ].
If Nn < A3 T −1/2 , then by (3.33), (3.32)

tn−1 − tn ≥ A−1 −2 −2 −2 −4 −4
3 Nn−1 ≥ A3 Nn ≥ A3 (tn−1 − tn ) ≤ A3 T
so in particular tn ≤ t0 − A−4 3 T . Of course this inequality also holds if tn ∈

[t0 − T, t0 − T /2]. In either case, we see from the hypothesis A4 N0−2 ≤ T1 ≤ A−1
4 T
that
tn < t − T1 ≤ t1 .
Let m be the largest index for which tm ≥ t − T1 , thus 1 ≤ m ≤ n − 1 and

tm+1 > t − T1 . By telescoping (3.33), we conclude that
m
X m+1
X
(3.35) A3 Ni−2 = −2
A3 Ni−1 ≥ t − tm+1 ≥ T1 .
i=0 i=1
On the other hand, from (3.31) and (3.2) we have

|PNi u(t, xi )| & A−1
1 Ni
for t ∈ [ti − A−2 −2

1 Ni , ti ]; as PNi is bounded on L
∞
by (2.3), this implies that
−1
kPNi u(t)kL∞ 3 & A
x (R ) 1 Ni
for such t. From (3.33) we see that the time intervals [ti − A−2 −2
1 Ni , ti ] are disjoint
and lie in [t − T1 , t] for i = 0, . . . , m − 1. Applying (3.4), we conclude that
m−1
1/2
X
A−1 −2 −2
1 Ni × A1 Ni . A4 T1
i=0
and thus
m−1
1/2
X
Ni−1 . A41 T1 .
i=0
Using (3.32) to extend this sum to the final index m, we conclude that
m
1/2
X
(3.36) Ni−1 . A22 T1 .
i=0
Comparing this with (3.35), we conclude that there exists i = 0, . . . , m such that
1/2
Ni−1 & A−2
3 T1 .
Since A4 N0−2 ≤ A−1

4 T , i cannot be zero, thus 1 ≤ i ≤ m. From (3.33), (3.32) we
have
t0 − ti ≥ ti−1 − ti
≥ A−1 −2
3 Ni−1
≥ A−2 −2
3 Ni
& A−6
3 T1 .
Since t0 −ti is also bounded by T1 , we also have from (3.33) that A−1
3 Ni
−2
≤ T1 , thus
−1/2 −1/2
Ni ≥ A3 T1 . Finally, from telescoping (3.34) and using (3.36), we conclude
that
1/2
|xi − x0 | . A24 T1 ,
and the claim follows.
Finally, we prove (vi), which is the most difficult estimate. The claim is in-
variant with respect to time translation and rescaling, so we may assume that
[t0 − T 0 , t0 ] = [0, 1]. In particular [−1, 1] ⊂ [t0 − T, t0 ], so we may decompose
u = ulin + unlin as before with the estimates (3.10), (3.11), (3.13).
From (3.13) we can find a time t1 ∈ [−1/2, 0] such that
Z
|∇unlin (t1 , x)|2 dx . A4 .
R3
20 TERENCE TAO
Fix this time t1 . From (3.11) we thus have

Z 4
X
|∇unlin (t1 , x)|2 + |∇j ulin (t1 , x)|3 dx . A4 .
R3 j=0
By the pigeonhole principle, we can thus find a scale

O(1)
(3.37) A100
6 R0 ≤ R ≤ exp(A6 )R0
such that
Z 4
X
(3.38) |∇unlin (t1 , x)|2 + |∇j ulin (t1 , x)|3 dx . A−10
6 .
A−10
6 R≤|x|≤A10
6 R j=0
Fix this R. We now propagate this estimate forward in time to [t1 , 1]. We first
achieve this for the linear component ulin , which is straightforward. From Sobolev
embedding we have
sup |∇j ulin (t1 , x)| . A−3
6
A−9 9
6 R≤|x|≤A6 R
for j = 0, 1, 2. Since ∇j ulin solves the linear heat equation, we conclude from this,
(2.2), and (3.11) that
(3.39) sup sup |∇j ulin (t, x)| . A−3
6
t1 ≤t≤1 A−8 R≤|x|≤A8 R
6 6
for j = 0, 1, 2. This estimate (when combined with (3.11)) will suffice to control
all the terms involving the linear component ulin of the velocity (or the analogous
component ω lin := ∇ × ulin of the vorticity).
The vorticity ω := ∇ × u obeys the vorticity equation (1.5). On [t1 , 1] × R3 , we
decompose ω = ω lin + ω nlin , where ω lin := ∇ × ulin is the linear component of the
vorticity and ω nlin := ∇ × unlin is the nonlinear component. As ω lin solves the heat
equation, we have
(3.40) ∂t ω nlin = ∆ω nlin − (u · ∇)ω + (ω · ∇)u.
As in [T, §10], we apply the energy method to this equation with a carefully chosen
time-dependent cutoff function. Namely, let
(3.41) R− ∈ [A−8 −8
6 R, 2A6 R]; R+ ∈ [A86 R/2, A86 R]
be scales to be chosen later, and define the time-dependent radii
Z t
R− (t) := R− + C0 (A6 + ku(t)kL∞ 3 ) dt
x (R )
t1
Z t
R+ (t) := R+ − C0 (A6 + ku(t)kL∞ 3 ) dt
x (R )
t1
that start at R− , R+ respectively, and contract inwards at a rate faster than the
velocity field u. From the bounded total speed property (3.4), (3.37), and the
hypothesis R0 ≥ 1, we conclude that
R− (t) ∈ [A−8 −8
6 R, 3A6 R]; R+ (t) ∈ [A86 R/3, A86 R]
for all t ∈ [t1 , 1].
For t ∈ [t1 , 1], we define the local enstrophy

Z
1
E(t) := |ω nlin (t, x)|2 η(t, x) dx
2 R3
where η is the time-varying cutoff
η(t, x) := max(min(A6 , |x| − R− (t), R+ (t) − |x|), 0),
thus η is supported in the annulus {R− (t) ≤ |x| ≤ R+ (t)}, is Lipschitz with norm
1, and equals A6 in the smaller annulus {R− (t) + A6 ≤ |x| ≤ R+ (t) − A6 }. From
(3.38) we have the initial bound
(3.42) E(t1 ) . A−9
6 .
Now we control the time derivative ∂t E(t) for t ∈ [t1 , 1]. From (3.40) and integration
by parts we have
∂t E(t) = −Y1 (t) − Y2 (t) + Y3 (t) + Y4 (t) + Y5 (t) + Y6 (t) + Y7 (t) + Y8 (t) + Y9 (t)
where Y1 is the dissipation term
Z
Y1 (t) := |∇ω nlin (t, x)|2 dx,
R3
Y2 (t) is the recession term

Z
1
Y2 (t) := − |ω nlin (t, x)|2 ∂t η(t, x) dx,
2 R3
Y3 (t) is the heat flux term

Z
1
Y3 (t) := |ω nlin (t, x)|2 ∆η(t, x) dx,
2 R3
Y4 (t) is the transport term

Z
1
Y4 (t) := |ω nlin (t, x)|2 u(t, x) · ∇η(t, x) dx,
2 R3
Y5 (t) is a correction to the transport term arising from ω lin ,

Z
Y5 (t) := − ω nlin (t, x) · (u(t, x) · ∇)ω lin (t, x) η(t, x) dx,
R3
Y6 (t) is the main nonlinear term

Z
Y6 (t) := ω nlin (t, x) · (ω nlin (t, x) · ∇)unlin (t, x) η(t, x) dx
R3
and Y7 (t), Y8 (t), Y9 (t) are corrections to the transport term arising from the ulin
and ω lin ,
Z
Y7 (t) := ω nlin (t, x) · (ω nlin (t, x) · ∇)ulin (t, x) η(t, x) dx
3
ZR
Y8 (t) := ω nlin (t, x) · (ω lin (t, x) · ∇)unlin (t, x) η(t, x) dx
R3
Z
Y9 (t) := ω nlin (t, x) · (ω lin (t, x) · ∇)ulin (t, x) η(t, x) dx.
R3
22 TERENCE TAO
Here all derivatives of the Lipschitz function η are interpreted in a distributional

sense. We now aim to control Y3 (t), . . . , Y9 (t) in terms of Y1 (t), Y2 (t), E(t), and
some other quantities that are well controlled. From definition of η we see that
−∂t η(t, x) = C0 (A6 + ku(t)kL∞ 3 )|∇η(t, x)|
x (R )
so in particular we have that Y2 (t) is non-negative and

Y4 (t) ≤ C0−1 Y2 (t).
A direct computation of ∆η in polar coordinates yields the bound
|ω nlin (t, x)|2
Z
Y3 (t) . dx
|x|∈[R− (t),R− (t)+A6 ]∪[R+ (t)−A6 ,R+ (t)] |x|
X Z
2
+ r |ω nlin (t, rθ)| dθ
r=R− (t),R− (t)+A6 ,R+ (t)−A6 ,R+ (t) S2
where dθ is surface measure on the sphere (in fact the r = R− (t) + A6 , R+ (t) − A6
terms are non-positive and could be discarded if desired). This expression is difficult
to estimate for fixed choices of R− , R+ . However, if selects R− , R+ uniformly at
random from the range (3.41), we see from Fubini’s theorem that the expected
value E|Y3 | of |Y3 | can be estimated by
|ω nlin (t, x)|2
Z
E|Y3 (t)| . A6 dx
|x|∈[A−8 −8 8 8
6 R,3A6 R]∪[A6 R/3,A6 R]
|x|2
and hence by (3.13), (3.37)
Z 1
E |Y3 (t)| dt . A−10
6
t1
(say). Thus we can select R− , R+ so that

Z 1
(3.43) |Y3 (t)| dt . A−10
6
t1
and we shall now do so.

To treat Y5 (t), we use Young’s inequality to bound
Z
Y5 (t) . E(t) + |(u · ∇)ω lin |2 η dx.
R3
Using (3.1), (3.11), (3.39), Hölder’s inequality, we then have
Y5 (t) . E(t) + A−2
6
(say).
In a similar vein, from (3.39) and Hölder’s inequality one has
Y7 (t) . E(t)
(with plenty of room to spare) and from Young’s inequality one has
Z
Y9 (t) . E(t) + |(ω lin · ∇)ulin |2 η(t, x) dx
R3
and hence by (3.1), (3.11), (3.39), and Hölder
Y9 (t) . E(t) + A−2
6 .
For Y8 , we again use Young’s inequality to bound

Z
Y8 (t) . E(t) + |(ω lin · ∇)unlin |2 η(t, x) dx
R3
and hence by (3.39)

Y8 (t) . E(t) + Y10 (t)
where Z
Y10 (t) := A−3
6 |∇unlin (t, x)|2 dx.
R3
Observe from (3.13) that
Z 1
(3.44) |Y10 (t)| dt . A−2
6 .
t1
We are left with estimation of the most difficult term Y6 (t). Following [T], we
cover the annulus {R− (t) ≤ |x| ≤ R+ (t)} by a boundedly overlapping Whitney
decomposition of balls B = B(xB , rB ), where the radius rB of the ball is given
1
as rB := 100 η(t, rB ). In particular, we have η(t, x) ∼ rB on the dilate 10B =
B(xB , rB ) of the ball. We can then write
X Z
Y6 (t) ∼ rB |ω nlin |2 |∇unlin | dx
B B
where we suppress the explicit dependence on t, x for brevity. Similarly one has
X Z
(3.45) E(t) ∼ rB |ω nlin |2 dx
B 10B
and
X Z
(3.46) Y1 (t) ∼ rB |∇ω nlin |2 dx
B 10B
To control Y6 (t), we need to control ∇unlin . The Biot-Savart law suggests that this
function has comparable size to ω nlin , but we need to localise this intuition to the
ball B and thus must address the slightly non-local nature of the Biot-Savart law.
Fortunately this can be handled using standard cutoff functions. Namely, we have
∆unlin = −∇ × ω nlin , hence if we let ψB be a smooth cutoff adapted to 3B that
equals 1 on 2B, then
unlin = −∆−1 (∇ × (ω nlin ψB )) + v
where v is harmonic on 2B. From Sobolev embedding and Hölder one has
3/2
kvkL2x (2B) . kω nlin ψB kL6/5 (R3 ) + kunlin kL2x (2B) . rB kω nlin kL3x (3B) + kunlin kL2x (2B)
x
and hence by elliptic regularity for harmonic functions

−5/2 −1 −5/2
k∇vkL∞
x (B)
. rB kvkL2x (2B) . rB kω nlin kL3x (3B) + rB kunlin kL2x (2B) .
We conclude the pointwise estimate
(3.47)
−1 −5/2
∇unlin = −∇∆−1 (∇ × (ω nlin ψB )) + O(rB kω nlin kL3x (3B) ) + O(rB kunlin kL2x (2B) )
24 TERENCE TAO
on B. By elliptic regularity, ∇∆−1 (∇×(ω nlin ψB )) has an L3x (B) norm of O(kω nlin kL3x (3B) ).
From Hölder’s inequality we thus have
Z
−5/2
|ω nlin |2 |∇unlin | dx . kω nlin k3L3x (3B) + rB kω nlin k2L2x (3B) kunlin kL2x (3B)
B
and hence Y6 (t) . Y6,1 (t) + Y6,2 (t), where
X
Y6,1 (t) := rB kω nlin k3L3x (3B)
B
and
−3/2
X
Y6,2 (t) := rB kω nlin k2L2x (3B) kunlin kL2x (3B) .
B
For Y6,2 (t), we first consider the contribution of the large balls in which rB ≥ A10 .
−3/2
Here we simply use (3.10) to bound kunlin kL2x (3B) . A2 . Since rB A2 . rB for
large balls B, the contribution of this case is O(E(t)) thanks to (3.45). Now we
look at the small balls in which rB < A10 . Here we use Hölder to bound
3/2 3/2
kunlin kL2x (3B) . rB kunlin kL∞ 3 . r
x (R )
2
B (A + kukL∞ 3 )
x (R )
so the contribution of this case is bounded by

X
kω nlin k2L2x (3B) (A2 + kukL∞ 3 ).
x (R )
B:rB <A10
For small balls B, 3B is completely contained inside the region in which ∂t η &
−1
C0 (A6 +kukL∞ 3 ), so the contribution of this case can be bounded by O(C
x (R ) 0 Y2 (t)).
Thus
Y6,2 (t) . E(t) + C0−1 Y2 (t).
Now we control Y6,1 (t). For each ball B, define the mean vorticity ωB by
nlin
R
3 ω ψB dx
ωB := R R .
ψ dx
R3 B
From the Poincaré inequality, Sobolev embedding, and the triangle inequality we
have
kω nlin − ωB kL6x (3B) . k∇ω nlin kL2x (10B)
and similarly
−1/2
(3.48) |ωB − ωB 0 | . rB k∇ω nlin kL2x (10B)
when B, B 0 are overlapping Whitney balls. We can now use Hölder’s inequality to
bound
X X
4 3
Y6,1 (t) . rB ωB + rB kω nlin − ωB k3L3x (3B)
B B
3/2 3/2
X X
4 3
. rB ωB + rB kω nlin − ωB kL2 (3B) k∇ω nlin kL2 (10B) .
x x
B B
By Young’s inequality and (3.46), we then have

1 X
4 3
X
Y6,1 (t) ≤ Y1 (t) + O( rB ωB + rB kω nlin − ωB k6L2x (3B) ).
2
B B
From the triangle inequality and Cauchy-Schwarz, and (3.45), one has
−1/2
kω nlin − ωB kL2x (3B) . kω nlin kL2x (3B) . rB E(t)1/2
and hence from Hölder, (3.46)

X X
rB kω nlin − ωB k6L2x (3B) . rB kω nlin − ωB k2L6x (3B) E(t)2 . E(t)2 Y1 (t).
B B
4 3
P
Now we estimate B rB ωB .
We can arrange the Whitney decomposition so that
all the radii rB are powers of 1.001, and that every ball B of radius less than (say)
A6 /100 has a “parent” ball p(B) that overlaps B and has radius 1.001rB . From
the triangle inequality we have
k−1
X
|ωB | ≤ |ωpk (B) | + |ωpi (B) − ωpi+1 (B) |
i=0
for any Whitney ball B, where k = kB is the first natural number for which the
1/2
iterated parent pk (B) has radius larger than A6 . By Hölder we then have
k−1
X
3 3
|ωB | . |ωpk (B) | + (1 + i)10 |ωpi (B) − ωpi+1 (B) |3 .
i=0
From a volume packing argment we see that for a given i, a Whitney ball B 0 is
of the form pi (B) for at most O((1.001)2i ) choices of B. One can then sum the
geometric series (exactly as in [T, §10]) and conclude that
X X X
4 3 4 3 4
rB ωB . rB ωB + rB |ωB − ωp(B) |3 .
B B:rB ≥A6 /100 B:rB <A6
1/2
For the small balls in which rB < A6 /100, we observe from (3.46) and Cauchy-
Schwarz that
−2
ωB , ωp(B) . rB E(t)1/2
and thus from (3.48), (3.46)
X
4
rB |ωB − ωp(B) |3 . E(t)1/2 Y1 (t).
B:rB <A6 /100
For the large balls in which rB ≥ A6 /100, we write ω nlin = ∇ × unlin and integrate
by parts using Cauchy-Schwarz to find that
−5/2
ωB . rB kunlin kL2x (3B)
and hence using (3.10) and the bounded overlap of the Whitney balls
−7/2 −7/2
X X
4 3
rB ωB . A2 rB kunlin k2L2x (3B) . A4 A6 . A−2
6 .
B:rB ≥A6 /100 B:rB ≥A6 /100
Thus we have
1
Y6,1 ≤ Y1 (t) + O(E(t)1/2 Y1 (t) + A−2 2
6 + E(t) Y1 (t)).
2
Putting all this together, we see that
1
∂t E(t) ≤ − Y1 (t)+O E(t) + |Y3 (t)| + |Y10 (t)| + A−2 6 + E(t)1/2
Y1 (t) + E(t)2
Y1 (t) .
2
A standard continuity argument using (3.42), (3.43), (3.44) then gives
(3.49) E(t) . A−2
6
26 TERENCE TAO
for all t1 ≤ t ≤ 1, and also

Z 1
(3.50) Y1 (t) dt . A−2
6 .
t1
These are subcritical regularity estimates and can now be iterated4 as in the proof
of (iii) to obtain higher regularity. First we move from control of the vorticity back
to control of the velocity. From (3.47) and elliptic regularity one has
−2
k∇unlin k2L2x (B) . kω nlin k2L2x (3B) + rB kunlin k2L2x (2B)
for any ball B; summing this on balls of radius A10 6 (say) using (3.49), (3.10), we
conclude that
Z
(3.51) |∇unlin (t, x)|2 dx . A−2
6
A−7 7
6 R≤|x|≤A6 R
for all t1 ≤ t ≤ 1. Similarly we have

−2
k∇2 unlin k2L2x (B) . k∇ω nlin k2L2x (3B) + rB k∇unlin k2L2x (2B)
and using (3.51), (3.50) in place of (3.10), (3.49) we conclude that
Z 1Z
|∇2 unlin (t, x)|2 dxdt . A−2
6 .
t1 A−6 6
6 R≤|x|≤A6 R
Using the Gagliardo-Nirenberg inequality (3.20) as before we see that

kunlin kL4 L∞ ([t1 ,1]×{A−5 R≤|x|≤A5 R}) . A−2
6
t x 6 6
which when combined with (3.39) gives

kukL4 L∞ ([t1 ,1]×{A−5 R≤|x|≤A5 R}) . A−2
6 .
t x 6 6
By repeating the arguments in (iii) (using (2.2) in place of (2.1) to handle the
long-range components of the heat kernel, which can be controlled with extremely
good bounds using (3.1)), one can then show iteratively that
kukL8 L∞ ([t1 ,1]×{A−4 R≤|x|≤A4 R}) . A−2
6 ,
t x 6 6
then
kukL∞ L∞ ([t1 ,1]×{A−3 R≤|x|≤A3 R}) . A−2
6 ,
t x 6 6
then
k∇ukL4 L∞ ([t1 ,1]×{A−2 R≤|x|≤A2 R}) . A−2
6 ,
t x 6 6
then finally
k∇ukL∞ L∞ ([t1 ,1]×{A−1 R≤|x|≤2A6 R}) . A−2
6
t x 6
and
k∇ωkL∞ ∞
t Lx ([t1 ,1]×{R≤|x|≤A6 R})
. A−2
6
giving (vi).
4It is likely that one can also proceed at this point using the local regularity theory from
[CKN].
4. Carleman inequalities for backwards heat equations

We will need some Carleman inequalities for backwards heat equations which
are essentially contained in previous literature (most notably [ESS2], [ESS]), but
made slightly more quantitative for our application (also it will be convenient to not
demand that the functions involved vanish at the starting and final time). Following
[ESS2], we shall reverse the direction of time and work here with backwards heat
equations rather than forward ones.
Our main tool is the following general inequality (cf. [ESS, Lemma 2]):
Lemma 4.1 (General Carleman inequality). Let [t1 , t2 ] be a time interval, and
let u ∈ Cc∞ ([t1 , t2 ] × Rd → Rm ) be a (vector-valued) test function solving the back-
wards heat equation
Lu = f
with L the backwards heat operator
(4.1) L := ∂t + ∆,
and let g : [t1 , t2 ] × R → R be smooth. Let F : [t1 , t2 ] × Rd → R denote the function
d
F := ∂t g − ∆g − |∇g|2 .
Then we have the inequality
Z Z
2 1 2 g 1 2 2 1 2
∂t |∇u| + F |u| e dx ≥ (LF )|u| + 2D g(∇u, ∇u) − |Lu| eg dx
Rd 2 Rd 2 2
for all t ∈ I, where D2 g is the bilinear form expressed in coordinates as
D2 g(v, w) := (∂i ∂j g)vi · wj
with the usual summation conventions. In particular, from the fundamental theorem
of calculus one has
Z t2 Z
1
(LF )|u|2 + 2D2 g(∇u, ∇u) eg dxdt
t1 Rd 2
Z t2 Z Z
1 1
≤ 2 g
|Lu| e dxdt + 2
|∇u| + F |u| 2
eg dx|t=t
t=t1 .
2
2 t1 R d Rd 2
The above inequality is valid in all dimensions, but in this paper we will only
need this lemma in the case d = m = 3.
Proof. By breaking u into components, we may assume without loss of gen-
erality that we are in the scalar case m = 1.
We use the usual commutator method. Introducing the weighted (and time-
dependent) inner product Z
hu, vi := uv eg dx
Rn
for test functions u, v : I × Rn → R, we compute after differentiating under the
integral sign and integrating by parts
Z
hLu, vi + hu, Lvi = (∂t (uv) + ∆(uv) − 2∇u · ∇v) eg dx
Rn
Z
−(∂t g)uv + (∆g + |∇g|2 )uv − 2∇u · ∇v eg dx

= ∂t hu, vi +
Rn
= ∂t hu, vi − hF u, vi − 2h∂i u, ∂i vi
28 TERENCE TAO
with the usual summation conventions. We can write

1
(4.2) − h∂i u, ∂i vi − hF u, vi = hSu, vi
2
where S is the differential operator
1
Su := ∆u + ∇g · ∇u − F u
2
which is then formally self-adjoint with respect to the inner product h, i; one can
view S as the self-adjoint component of L. We can then rewrite the above identity
as
∂t hu, vi = hLu, vi + hu, Lvi − 2hSu, vi.
In particular, by the self-adjointness of S we have for any test functions u, v that
∂t hSu, vi = hLSu, vi + hSu, Lvi − 2hSu, Svi
= h[L, S]u, vi + hSLu, vi + hSu, Lvi − 2hSu, Svi
= h[L, S]u, vi + hLu, Svi + hSu, Lvi − 2hSu, Svi
1 1
= h[L, S]u, vi + hLu, Lvi − h(L − 2S)u, (L − 2S)vi.
2 2
Among other things, this shows that the differential operator [L, S] (which does
not involve any time derivatives) is formally self-adjoint with respect to the inner
product h, i. Specialising to the case u = v, we conclude in particular the inequality
1
(4.3) ∂t hSu, ui ≤ h[L, S]u, ui + hLu, Lui.
2
Now we compute [L, S]. As previously noted, [L, S] is a formally self-adjoint differ-
ential operator that does not involve any time derivatives. Since the second order
operator L commutes with the second order component ∆ of S, we see that [L, S]
is a second-order operator. The highest order terms can be easily computed in
coordinates as
[L, S]u = 2(∂i ∂j g)∂i ∂j u + l.o.t.
and hence after integrating by parts the symmetric quadratic form h[L, S]u, vi must
take the form
Z
h[L, S]u, vi = (−2D2 g(∇u, ∇v) + Huv) eg dx
Rd
for some function H; setting u = 1, we see that H must equal
1
H = [L, S]1 = LS1 = − LF.
2
We conclude that
Z
1
h[L, S]u, ui = (−2D2 g(∇u, ∇u) − (LF )|u|2 ) eg dx.
R d 2
Inserting this identity back into (4.3) and using (4.2), we obtain the claim.
The inequality below is a quantitative variant of [ESS, Lemma 4].
Proposition 4.2 (First Carleman inequality). Let T > 0, 0 < r− < r+ , and
let A denote the cylindrical annulus
A := {(t, x) ∈ R × R3 : t ∈ [0, T ]; r− ≤ |x| ≤ r+ }.
Let u : A → R3 be a smooth function obeying the differential inequality

−1/2
(4.4) |Lu| ≤ C0−1 T −1 |u| + C0 T −1/2 |∇u|
on A. Assume the inequality
2
(4.5) r− ≥ 4C0 T.
Then one has
Z T /4 Z r− r+
2
T −1 |u|2 + |∇u|2 dxdt . C02 e− 4C0 T (X + e2r+ /C0 T Y )

0 10r− ≤|x|≤r+ /2
where Z Z
2
X := e2|x| /C0 T
(T −1 |u(t, x)|2 + |∇u(t, x)|2 ) dxdt
A
and Z
Y := |u(0, x)|2 dx.
r− ≤|x|≤r+
r− r+
The key feature here is the gain of e− 4C0 T , which can be compared against the
2 2
trivial bound of e−2r− /C0 T X that follows by lower bounding the factor e2|x| /C0 T
2
appearing in X by e2r− /C0 T . Thus, this lemma becomes powerful when the ratio
r+ /r− is large. Informally, Proposition 4.2 asserts that if u solves (4.4) on A,
has some mild Gaussian decay as |x| → ∞, and is extremely small at t = 0,
then it is also very small in the interior of A near t = 0. The various numerical
constants such as 1/4 or 10 appearing in the above proposition can be modified
(and optimised) if desired, but we fix a specific choice of constants for sake of
2
concreteness. The weight e2|x| /C0 T in X is inconvenient, but it is negligible when
2
compared against the “natural” decay rate of e−|x| /4t arising from the fundamental
solution of the heat equation, and it can be managed in our application by using
the second Carleman inequality given below. Specialising Proposition 4.2 the case
u(0, x) = 0 (so that Y = 0) and sending r+ to infinity, one recovers a variant of the
backwards uniqueness result in [ESS, Lemma 4].
Proof. We may assume that

(4.6) r+ ≥ 20r−
since the claim is vacuous otherwise. By the pigeonhole principle, one can find a
time T0 ∈ [T /2, T ] such that
Z
2
(4.7) e2|x| /C0 T (T −1 |u(T0 , x)|2 + |∇u(T0 , x)|2 ) dx . T −1 X.
r− ≤|x|≤r+
Fix this time T0 . In the discussion below we implicitly restrict (t, x) to the region
A ∩ ([0, T0 ] × R3 ). We set
r+
(4.8) α :=
2C0 T 2
and observe from (4.5), (4.6) that
40
(4.9) α≥ .
r− T
30 TERENCE TAO
Following [ESS2], we apply Lemma 4.1 on the interval [0, T0 ] with the weight
1
g := α(T0 − t)|x| + |x|2
C0 T
and u replaced by ψu, where ψ(x) is a smooth cutoff supported on the region
r1 ≤ |x| ≤ r2 that equals 1 on 2r1 ≤ |x| ≤ r2 /2 and obeys the estimates |∇j ψ(x)| =
O(1/|x|j ) for j = 0, 1, 2. Since α(T0 − t)|x| is convex in x, we have
1
D2 g(∇(ψu), ∇(ψu)) ≥ 2 |∇(ψu)|2 .
C0 T
The function F defined in Lemma 4.1 can be computed on A as
2
2α(T0 − t) 6 2|x|
F = −α|x| − − − α(T0 − t) +
|x| C0 T C0 T
2α(T0 − t) 6 T0 − t 4|x|2
= −α|x| − − − α2 (T0 − t)2 − 4α |x| − 2 2 .
|x| C0 T C0 T C0 T
In particular F is negative. We also calculate
4α|x| 8α(T0 − t) 24
LF = 2α2 (T0 − t) + − − 2 2.
C0 T C0 T |x| C0 T
T0 −t
We see from (4.5) that |x| ≤ 41 |x|, so that
2α|x| 24
LF ≥ − 2 2.
C0 T C0 T
56
By (4.9) we thus have LF ≥ C02 T 2
. Applying Lemma 4.1 and discarding some
terms, we conclude that
Z T0 Z
28C0−2 T −2 |u|2 + 4C0−1 T −1 |∇u|2 eg dxdt

0 2r− ≤|x|≤r+ /2
Z T0 Z
1
≤ |L(ψu)|2 eg dxdt
2 0 R3
Z Z
+ |∇(ψu)(T0 , x)|2 eg(T0 ,x) dx + |(ψu)(0, x)|2 |F (0, x)| eg(0,x) dx.
R3 R3
In the region 2r− ≤ |x| ≤ r+ /2 we have from (4.4) that
|L(ψu)|2 = |Lu|2 ≤ 2C0−2 T −2 |u|2 + 2C0−1 T −1 |∇u|2 .
In the regions r− ≤ |x| ≤ 2r− or r+ /2 ≤ |x| ≤ r+ , we have
|L(ψu)|2 . |Lu|2 + |x|−2 |∇u|2 + |x|−4 |u|2 . C0−2 T −2 |u|2 + C0−1 T −1 |∇u|2
thanks to (4.4), (4.5). For all other x, L(ψu) vanishes. A similar calculation gives
|∇(ψu)|2 . |∇u|2 + C0−1 T −1 |u|2 .

We therefore have
Z T0 Z
C0−2 T −2 |u|2 + C0−1 T −1 |∇u|2 eg dxdt

0 2r− ≤|x|≤r+ /2
Z T0 Z
. (C0−2 T −2 |u|2 + C0−1 T −1 |∇u|2 ) eg dxdt
0 |x|∈[r− ,2r− ]∪[r+ /2,r+ ]
Z Z
−1 −1 2 2 g(T0 ,x)
+ (C0 T |u| + |∇u| )(T0 , x) e dx + |u(0, x)|2 |F (0, x)| eg(0,x) dx.
r− ≤|x|≤r+ r− ≤|x|≤r+
From (4.7) one has

Z
(C0−1 T −1 |u|2 + |∇u|2 )(T0 , x) eg(T0 ,x) dx . T −1 X.
r− ≤|x|≤r+
When t ∈ [0, T0 ] and |x| ∈ [r+ /2, r+ ], one has

|x|2 2 2
eg ≤ eαT |x|− C0 T e2|x| /C0 T
≤ e2|x| /C0 T
by (4.8). When instead t ∈ [0, T0 ] and |x| ∈ [r− , 2r− ], one has
|x|2 2 2
eg ≤ eαT |x|− C0 T e2|x| /C0 T
≤ e2αT r− e2|x| /C0 T
.
We conclude that
Z T0 Z
C0−2 T −2 |u|2 + C0−1 T −1 |∇u|2 eg dxdt

0 2r− ≤|x|≤r+ /2
Z
. e2αT r− T −1 X + |u(0, x)|2 |F (0, x)| eg(0,x) dx.
r− ≤|x|≤r+
In the region t ∈ [0, T /4], 10r− ≤ |x| ≤ r+ /2, one has

αT |x| |x|2 5
+C T
eg ≥ e 4 0 ≥ e 2 αT r−
and hence
Z T /4 Z
C0−2 T −2 |u|2 + C0−1 T −1 |∇u|2

dxdt
0 10r− ≤|x|≤r+ /2
Z !
−αT r− /2 −1 2 g(0,x)
.e T X+ |u(0, x)| |F (0, x)| e dx .
r− ≤|x|≤r+
From (4.8) we have

r− r+
e−αT r− /2 = e− 4C0 T .
Finally, for t = 0 and r− ≤ |x| ≤ r+ one has

2
|x|2 3r+
eg ≤ eαT |x|+ C0 T ≤ e 2C0 T

32 TERENCE TAO
and
|F | . αr+ + αT /r− + C0−1 T −1 + α2 T 2 + C0−1 αr+ + C0−2 T −2 r+
2
2 2 2 2
r+ r+ 1 r+ r+ r+
. + + + + +
T2 r− T T T2 T2 T2
2
r 1
. + (T −1 + )
T r+ r−
r2
. + T −1
T
2
3r+ 2
2r+
r2
since r+1r− ≤ r12 . T1 by (4.6), (4.5). Bounding T+ e 2T . e T and multiplying by
−
T , we conclude that
Z T /4 Z
C0−2 T −1 |u|2 + C0−1 |∇u|2 dxdt

0 10r− ≤|x|≤r+ /2
Z !
−r r+
− 4C 2
2r+ /T 2
.e 0T X +e |u(0, x)| dx
r− ≤|x|≤r+
giving the claim.

Our second application of Lemma 4.1 is the following quantitative version of
standard parabolic unique continuation results.
Proposition 4.3 (Second Carleman inequality). Let T, r > 0, and let C denote
the cylindrical region
C := {(t, x) ∈ R × R3 : t ∈ [0, T ]; |x| ≤ r}.
Let u : C → R3 be a smooth function obeying the differential inequality (4.4) on C.
Assume the inequality
(4.10) r2 ≥ 4000T.
Then for any
T
(4.11) 0 < t1 ≤ t0 <
1000
one has
Z 2t0 Z
2 r2 2
T −1 |u|2 + |∇u|2 e−|x| /4t dxdt . e− 500t0 X + t0 (et0 /t1 )O(r /t0 ) Y
3/2
t0 |x|≤r/2
where Z T Z
X := (T −1 |u|2 + |∇u|2 ) dxdt
0 |x|≤r
and Z
−3/2 −|x|2 /4t1
Y := |u(0, x)|2 t1 e dx.
|x|≤r
As with the previous inequality, the numerical constants here such as 1000, 500
can be optimised if desired, but this explicit choice of constants suffices for our ap-
r2
plication. The key feature here is the gain of e− 500t0 . Specialising to the case where
r2
u vanishes to infinite order at (0, 0), sending t1 → 0 (which sends (et0 /t1 )O( t0 ) Y
to zero thanks to the infinite order vanishing), and then sending t0 → 0, we ob-
tain a variant of a standard unique continuation theorem for backwards parabolic
equations (see e.g., [ESS2, Theorem 4.1]).
Proof. By the pigeonhole principle, we can select a time
T T
(4.12) ≤ T0 ≤
200 100
such that
Z
(4.13) (T −1 |u|2 + |∇u|2 ) dx . T −1 X.
|x|≤r
We define
r2
(4.14) α=
400t0
so from (4.10), (4.11) we have
(4.15) α ≥ 10.
We apply Lemma 4.1 on [0, T0 ] × R3 with the weight
|x|2 3 t + t1 t + t1
g := − − log(t + t1 ) − α log +α
4(t + t1 ) 2 T0 + t1 T0 + t1
2
1
(which is a modification of the logarithm of the fundamental solution t3/2 e−|x| /4t of
the heat equation) and u replaced by ψu, where ψ(x) is a smooth cutoff supported
on the region |x| ≤ r that equals 1 on |x| ≤ r/2 and obeys the estimates
(4.16) |∇j ψ(x)| = O(r−j )
for r/2 ≤ |x| ≤ r and j = 0, 1, 2. Clearly
1
D2 g(∇(ψu), ∇(ψu)) = − |∇(ψu)|2 .
2(t + t1 )
We can calculate
2
|x|2

3 α α 3 x
F = 2
− − + + +
4(t + t1 ) 2(t + t1 ) t + t1 T0 + t1 2(t + t1 ) 2(t + t1 )
α α
= −
T0 + t1 t + t1
and hence
α
LF = .
(t + t1 )2
From Lemma 4.1 we thus have
Z
α α
∂t |∇(ψu)|2 − |ψu|2 + |ψu|2 eg dx
R3 2(t + t1 ) 2(T0 + t1 )
Z
α 2 1 2 1 2
≥ 2
|ψu| − |∇(ψu)| − |Lu| eg dx.
R 3 2(t + t 1 ) t + t 1 2
To exploit this differential inequality we use the method of integrating factors. If
we introduce the energy
Z
2 α 2 α 2
E(t) := |∇(ψu)| − |ψu| + |ψu| eg dx
R3 2(t + t1 ) 2(T0 + t1 )
34 TERENCE TAO
then we conclude from the product rule that

(t + t1 )2

∂t (t + t1 ) + E(t)
10T0

t + t1
≥ 1+ E(t)
5(T0 + t1 )
(t + t1 )2
Z
α 2 1 2 1 2
+ (t + t1 + ) |ψu| + |∇(ψu)| − |L(ψu)| eg dx
10T0 R3 2(t + t1 )2 t + t1 2
(t + t1 )2
Z
t + t1 2 5(T0 + t1 ) − (t + t1 ) 2 1 2
= |∇(ψu)| + α |ψu| − (t + t1 + )|L(ψu)| eg dx
R3 10(T0 + t1 ) 10(T0 + t1 )2 2 10(T0 + t1 )
Z
t + t1 2 α 2 2
≥ |∇(ψu)| + |ψu| − (t + t1 )|L(ψu)| eg dx
R3 10(T0 + t1 ) 10(T0 + t1 )
and hence by the fundamental theorem of calculus
Z T0 Z
t + t1 2 α 2
|∇(ψu)| + |ψu| eg dxdt
0 R3 10(T0 + t1 ) 10(T0 + t1 )
Z T0 Z
(t + t1 )2

t=T0
≤ (t + t1 )|L(ψu)|2 eg dxdt + (t + t1 ) + E(t)|t=0 .
0 R3 10(T0 + t1 )
Discarding some terms, we conclude that
Z T0 Z
t + t1 α
|∇u|2 + |u|2 eg dxdt
0 |x|≤r/2 10(T0 + t1 ) 10(T0 + t1 )
Z T0 Z
(4.17) ≤ (t + t1 )|L(ψu)|2 eg dxdt
0 R3
Z Z !
+ O T0 |∇(ψu)(T0 , x)|2 eg dx + α |u(0, x)|2 eg dx .
R3 |x|≤r
When |x| ≤ r/2, one has
|L(ψu)|2 = |Lu|2 ≤ 2T −2 |u|2 + 2T −1 |∇u|2
thanks to (4.4). By (4.12), (4.15) the contribution of this case is less than half of
the left-hand side of (4.17). When r/2 ≤ |x| ≤ r, we have from (4.16), (4.10) that
|L(ψu)|2 . T −2 |u|2 + T −1 |∇u|2 + r−4 |u|2 + r−2 |∇|2

. T −2 |u|2 + T −1 |∇u|2 .
Finally, L(ψu) vanishes for |x| > r. Putting all this together, we conclude that
Z T0 Z
t + t1 2 α 2
|∇u| + |u| eg dxdt
0 |x|≤r/2 T 0 + t 1 T 0 + t 1
Z T0 Z
. (t + t1 )(T −2 |u|2 + T −1 |∇u|2 ) eg dxdt
0 r/2≤|x|≤r
Z Z
+ T0 |∇(ψu)(T0 , x)|2 eg dx + α |u(0, x)|2 eg dx.
R3 |x|≤r
Restricting the left-hand integral to the region t0 ≤ t ≤ 2t0 and also bounding
t1 ≤ t0 ≤ T0 ≤ T in several places, we conclude that
Z 2t0 Z
t0 2 α 2
t0 |x|≤r/2 T0 T0
Z T0 Z
. (T −1 |u|2 + |∇u|2 ) eg dxdt
0 r/2≤|x|≤r
Z Z
2 g
+T |∇(ψu)(T0 , x)| e dx + α |u(0, x)|2 eg dx.
R3 |x|≤r
From elementary calculus we have the inequality

a b
− − b log t ≤ b log
t ae
for any a, b, t > 0 (the left-hand side attains its maximum when t = a/b). When
r/2 ≤ |x| ≤ r and 0 ≤ t ≤ T0 , we then have
|x|2

3
g≤− − α+ log(t + t1 ) + α log(T0 + t1 ) + α
4(t + t1 ) 2
4 α + 32

3
≤ α+ log + α log(T0 + t1 ) + α
2 e|x|2

3 32α
≤ α+ log 2 + α log(T0 + t1 ) + α
2 er
32α(T0 + t1 ) 3 32α
≤ α log + log 2
r2 2 er
and thus by (4.14)
Z T0 Z
−3/2 32α(T0 + t1 )
(T −1 |u|2 + |∇u|2 ) eg dxdt . t0 exp(α log )X.
0 r/2≤|x|≤r r2
When |x| ≤ r and t = T0 , then
3
g ≤ − log t0 + α
2
and ∇(ψu) is supported on the ball {|x| ≤ r} and obeys the estimate
|∇(ψu)| . |∇u| + r−1 |u| . T −1 |u| + |∇u|
thanks to (4.10), and hence by (4.13)
Z
−3/2
T |∇(ψu)(T0 , x)|2 eg dx . t0 exp(α)X.
R3
From (4.10), (4.15) we have log 32α(Tr20 +t0 ) ≥ 1. Thus

Z 2t0 Z
t0 2 α 2
t0 |x|≤r/2 T0 T0
Z
−3/2 32α(T0 + t1 )
. t0 exp(α log )X +α |u(0, x)|2 eg dx.
r2 |x|≤r
In the region t0 ≤ t ≤ 2t0 , |x| ≤ r/2, we have

|x|2 3 3t0
g≥− − log(3t0 ) − α log
4t 2 T0 + t1
36 TERENCE TAO
so that
−3/2 −|x|2 /4t 3t0
eg & t0 e exp −α log .
T0 + t1
Finally, when t = 0 and |x| ≤ r, we have
|x|2 3 t1
g≤− − log t1 − α log +α
4t1 2 T0 + t1
so that
−3/2 −|x|2 /4t1 e(T0 + t1 )
eg ≤ t1 e exp α log .
t1
We conclude that
Z 2t0 Z
t0 2 α 2
|∇u| + |u| dxdt
t0 |x|≤r/2 T0 T0
Z
96αt0 3et0 3/2 −3/2 2
. exp α log 2
X + α exp α log( ) t0 |u(0, x)|2 t1 e−|x| /4t1 dx.
r t1 |x|≤r
From (4.14) we have log 96αt

r2
0
≤ −1, while from (4.15), (4.12), (4.10), (4.14) we
−1 t0 t0
α
have T0 & T and T0 & r2 & α−1 . We conclude that
Z 2t0 Z
|∇u|2 + T −1 |u|2 dxdt

t0 |x|≤r/2
Z
3et0 3/2 −3/2 2
. α2 e−α X + exp α log( ) t0 |u(0, x)|2 t1 e−|x| /4t1 dx.
t1 |x|≤r
r2
From (4.14) we have α = O(r2 /t0 ) and α2 e−α . e− 500t0 , and the claim follows.
5. Main estimate
In this section we combine the estimates in Proposition 3.1 with the Carleman
inequalities from the previous section to obtain
Theorem 5.1 (Main estimate). Let t0 , T, u, p, A obey the hypotheses of Propo-
sition 3.1, and suppose that there exists x0 ∈ R3 and N0 > 0 such that
|PN0 u(t0 , x0 )| ≥ A−1
1 N0
j
where as before we set Aj := AC0 . Then
O(1)
T N02 ≤ exp(exp(exp(A6 ))).
Proof. After translating in time and space we may normalise (t0 , x0 ) = (0, 0).
Let T1 be an arbitrary time scale in the interval
A4 N0−2 ≤ T1 ≤ A−1
4 T.
By Proposition 3.1(v), there exists

−O(1) O(1) 1/2
(5.1) (t1 , x1 ) ∈ [−T1 , −A3 T1 ] × B(0, A4 T1 )]
and
O(1) −1/2
(5.2) N1 = A3 T1
such that
|PN1 u(t1 , x1 )| ≥ A−1
1 N1 .
From the Biot-Savart law we have

PN1 u(t1 , x1 ) = −∆−1 PN1 ∇ × P̃N1 ω(t1 , x1 ),
and hence by (2.2)
PN1 u(t1 , x1 ) . N1−1 kP̃N1 ω(t1 )kL∞ (B(x1 ,A1 /N1 )) + A−50
1 N1−1 kP̃N1 ω(t1 )kL∞ (R3 ) .
From (3.1), (2.3) one has
kP̃N1 ω(t1 )kL∞ (R3 ) . AN12
and thus we have
|P̃N1 ω(t1 , x01 )| & A−1 2
1 N1
O(1) 1/2
for some x01 = x1 + O(A1 /N1 ) = O(A4 T1 ). By Proposition 3.1(i), one has
∇P̃N1 ω = O(AN13 ); ∂t P̃N1 ω = O(AN14 )
and thus
(5.3) |P̃N1 ω(t, x)| & A−1 2
1 N1
for all (t, x) ∈ [t1 , t1 + A−2 −2 0 −2 −1

1 N1 ] × B(x1 , A1 N1 ). By Proposition 3.1(iii), there
is an interval
−O(1)
I 0 ⊂ [t1 , t1 + A−2 −2
1 N1 ] ∩ [−T1 , −A3 T1 ]
−O(1)
with |I 0 | = A3 T1 such that
O(1) −1/2 O(1)
u(t, x) = O(A3 T1 ), ∇u(t, x) = O(A3 T1−1 )
and
O(1) O(1) −3/2
(5.4) ω(t, x) = O(A3 T1−1 ), ∇ω(t, x) = O(A3 T1 )
0 3 0
on I × R . By shrinking I as necessary, we may thus assume that
−1/2
(5.5) |u(t, x)| ≤ C0 |I 0 |−1/2 ; |∇u(t, x)| ≤ C0−1 |I 0 |−1 .
On the other hand, from (5.3), (5.1), (5.2) one has
Z
−O(1) −1/2
|P̃N1 ω(t, x)|2 dx & A3 T1
O(1) 1/2
B(0,A4 T1 )
0
for all t ∈ I . From (2.2) and (5.5) this implies that
Z
−O(1) −1/2
(5.6) |ω(t, x)|2 dx & A3 T1
O(1) 1/2
B(0,A4 T1 )
0
for all t ∈ I .
1/2
Write I 0 := [t0 − T 0 , t0 ], and let x∗ ∈ R3 be any point with |x∗ | ≥ A5 T1 . We
0 3 0
apply Proposition 4.3 on the slab [0, T ] × R with r := A5 |x∗ |, t0 := T /2, and
t1 := A−4 0
5 T , and u replaced by the function
(t, x) 7→ ω(t0 − t, x∗ + x)
(so that the hypothesis (4.4) follows from the vorticity equation and (5.5)) to con-
clude that
Z . exp(−A5 |x∗ |2 /T 0 )X + (T 0 )3/2 exp(O(A35 |x∗ |2 /T 0 ))Y
where Z Z
X := ((T 0 )−1 |ω|2 + |∇ω|2 ) dxdt
I0 B(x∗ ,A5 |x∗ |)
38 TERENCE TAO
and Z
0 −3/2 4 2
/4T 0
Y := (T ) |ω(t0 , x)|2 e−A5 |x−x∗ | dx
B(x∗ ,A5 |x∗ |)
and Z t0 −T 0 /2 Z
2
/4(t0 −t)
Z := (T 0 )−1 |ω|2 e−|x−x∗ | dxdt.
t0 −T 0 B(x∗ ,A5 |x∗ |/2)
From (5.6) we have
−O(1)
Z & A3 exp(−|x∗ |2 /100T 0 )(T 0 )−1/2 .
From (5.4) we have
X . (T 0 )−2 A35 |x∗ |3 . A35 exp(|x∗ |2 /T 0 )(T 0 )−1/2
and hence the expression exp(−A5 |x∗ |2 /T 0 )X is negligible compared to Z. We
conclude that
Y & exp(−O(A35 |x∗ |2 /T 0 ))(T 0 )−2 .
Using (5.4), the contribution to Y outside of the ball B(x∗ , |x∗ |/2) is negligible,
thus Z
|ω(t0 , x)|2 dx & exp(−O(A35 |x∗ |2 /T 0 ))(T 0 )−1/2
B(x∗ ,|x∗ |/2)
and therefore
Z
|ω(t0 , x)|2 dx & exp(−O(A35 R2 /T 0 ))(T 0 )−1/2
B(0,2R)\B(0,R/2)
1/2
whenever R ≥ A5 T1 . A similar argument holds with t0 replaced by any time in
[t0 − T 0 /4, t0 ]. We conclude in particular that we have the Gaussian lower bound
Z −A−14 T1
Z
1/2
(5.7) |ω(t, x)|2 dxdt & exp(−A45 R2 /T1 )T1
−T1 B(0,2R)\B(0,R/2)
for any time scale T1 and spatial scale R with A4 N0−2 ≤ T1 ≤ A−1
4 T and R ≥
1/2
A5 T1 .
Now let T2 be a scale for which
(5.8) A24 N0−2 ≤ T2 ≤ A−1
4 T.
By Proposition 3.1(vi), there exists a scale
1/2 O(1) 1/2
(5.9) A6 T2 ≤ R ≤ exp(A6 )T2
such that on the cylindrical annulus
Ω := {(t, x) ∈ [−T2 , 0] × R3 : R ≤ |x| ≤ A6 R}
one has the estimates
− j+1 − j+2
(5.10) ∇j u(t, x) = O(A−2
6 T2
2
); ∇j ω(t, x) = O(A−2
6 T2
2
)
3
for j = 0, 1. We apply Proposition 4.2 on the slab [0, T2 /C0 ] × R with r− := 10R,
r+ := A6 R/10, and u replaced by the function
(t, x) 7→ ω(−t, x)
(so that the hypothesis (4.4) follows from the vorticity equation and (5.10)) to
conclude that
1/2 O(1)
Z 0 . exp(−A6 R2 /T2 )X 0 + exp(exp(A6 ))Y 0
where Z 0 Z
2|x|2
X :=0
e T2
(T2−1 |ω|2 + |∇ω|2 ) dxdt
−T2 /C0 10R≤|x|≤A6 R/10
and Z
Y 0 := |ω(0, x)|2 dx
10R≤|x|≤A6 R/10
and Z 0 Z
Z 0 := T2−1 |ω|2 dxdt.
−T2 /4C0 100R≤|x|≤A1 R/20
From (5.7) (with R replaced by 200R) we have
−1/2
Z 0 & exp(−A55 R2 /T2 )T2 .
Thus we either have
1/3 −1/2
(5.11) X 0 & exp(A6 R2 /T2 )T2
or
O(1) −1/2
(5.12) Y 0 & exp(− exp(A6 ))T2 .
Suppose for the moment that (5.11) holds. From the pigeonhole principle, we can
then find a scale
(5.13) 10R ≤ R0 ≤ A6 R/10
such that
Z 0 Z
2|x|2
1/4 −1/2
e T2
(T2−1 |ω|2 + |∇ω|2 ) dxdt & exp(A6 R2 /T2 )T2
−T2 /C0 R0 ≤|x|≤2R0
and thus
Z 0 Z
−1/2
(T2−1 |ω|2 + |∇ω|2 ) dxdt & exp(−10(R0 )2 /T2 )T2
−T2 /C0 R0 ≤|x|≤2R0
From (5.10) we see that the contribution to the left-hand side arising from those
times t in the interval [− exp(−20(R0 )2 /T2 )T2 , 0] is negligible, thus
Z − exp(−20(R0 )2 /T2 )T2 Z
−1/2
(T2−1 |ω|2 +|∇ω|2 ) dxdt & exp(−10(R0 )2 /T2 )T2 .
−T2 /C0 R0 ≤|x|≤2R0
Thus by a further application of the pigeonhole principle, one can locate a time
scale
(5.14) exp(−20(R0 )2 /T2 )T2 ≤ t0 ≤ T2 /C0
such that
Z −t0 Z
−1/2
(T2−1 |ω|2 + |∇ω|2 ) dxdt & exp(−10(R0 )2 /T2 )T2 .
−2t0 R0 ≤|x|≤2R0
1/2
Covering the annulus R0 ≤ |x| ≤ 2R0 by O(exp(O((R0 )2 /T2 )) balls of radius t0 ,
one can then find x∗ with R0 ≤ |x∗ | ≤ 2R0 such that
Z −t0 Z
−1/2
(5.15) (T2−1 |ω|2 + |∇ω|2 ) dxdt & exp(−O((R0 )2 /T2 ))T2 .
1/2
−2t0 B(x∗ ,t0 )
40 TERENCE TAO
1/4
Now we apply Proposition 4.3 on the slab [0, 1000t0 ]×R3 with r := C0 (t0 /T2 )1/2 R0 ≤
|x∗ |/10, t1 := t0 , and u replaced by the function
(t, x) 7→ ω(−t, x∗ + x)
(so that the hypothesis (4.4) follows from the vorticity equation and (5.5)) to con-
clude that
0 2
1/2 (R ) 3/2 1/2
(5.16) Z 00 . exp(−C0 )X 00 + t0 exp(O(C0 (R0 )2 /T2 ))Y 00
500T2
where Z 0 Z
X 00 := (t−1 2 2
0 |ω| + |∇ω| ) dxdt
−T2 B(x∗ ,|x∗ |/2)
and Z
−3/2 0 2
Y 00 := t0 |ω(0, x)|2 e−|x−x | /4t0
dx
B(x∗ ,|x∗ |/2)
and
Z [−t0 Z
2
Z 00 := (t−1 2 2 −|x−x∗ |
0 |ω| + |∇ω| )e
/4|t|
dxdt.
1/2
−2t0 B(x∗ ,t0 )
From (5.15), (5.14) one has

−1/2
Z 00 & exp(−O((R0 )2 /T2 ))T2 .
From (5.10), (5.14) one has
−1/2
X 00 . T2−1 t0−1 (R0 )3 . exp(O((R0 )2 /T2 ))T2 .
As C0 is large, the first term on the right-hand side of (5.16) can thus be absorbed
by the left-hand side, so we conclude that
1/2
Y 00 & exp(−O(C0 (R0 )2 /T2 ))T2−2
and hence
Z
1/2 3/2
|ω(0, x)|2 dx & exp(−O(C0 (R0 )2 /T2 ))T2−2 t0 .
R0 /2≤|x|≤2R0
Using the bounds (5.13), (5.9), (5.14), we conclude in particular that

Z
O(1) −1/2
(5.17) |ω(0, x)|2 dx & exp(− exp(A6 ))T2 .
2R≤|x|≤A6 R/2
Note that this bound is also implied by (5.12). Thus we have unconditionally
established (5.17) for any scale T2 obeying (5.8), and for a suitable scale R obeying
(5.9) and the bounds (5.10).
We now convert this vorticity lower bound (5.17) to a lower bound on the
O(1) 3/2
velocity. The annulus {2R ≤ |x| ≤ A6 R/2} has volume O(exp(exp(A6 ))T2 )
by (5.9), hence by the pigeonhole principle there exists a point x∗ in this annulus
for which
O(1)
|ω(0, x∗ )| & exp(− exp(A6 ))T2−1 .
Comparing this with (5.10), we see that
Z
O(1)
| ω(0, x∗ − ry)ϕ(y) dy| & exp(− exp(A6 ))T2−1
R3
for some bump function ϕ supported on B(0, 1), where r is a radius of the form
O(1) 1/2
r = exp(− exp(A6 ))T2 . Writing ω = ∇ × u and integrating by parts, we
conclude that
Z
O(1) −1/2
| u(0, x∗ − ry)∇ × ϕ(y) dy| & exp(− exp(A6 ))T2
R3
and hence by Hölder’s inequality
Z
O(1) −3/2
|u(0, x∗ − ry)|3 dy & exp(− exp(A6 ))T2
B(0,1)
or equivalently Z
O(1)
|u(0, x)|3 dx & exp(− exp(A6 )).
B(x∗ ,r)
We conclude that for any scale T2 obeying (5.8), we have
Z
O(1)
|u(0, x)|3 dx & exp(− exp(A6 )).
1/2 1/2
T2 ≤|x|≤exp(A7 )T2
Summing over a set of such scales T2 increasing geometrically at ratio exp(A7 ), we

conclude that if T ≥ A24 N0−2 , then
Z
O(1)
|u(0, x)|3 dx & exp(− exp(A6 )) log(T N02 ).
R3
Comparing this with (3.1), one obtains the claim.
6. Applications
Using the main estimate, we now prove the theorems claimed in the introduc-
tion.
We begin with Theorem 1.2. By increasing A as necessary we may assume that
A ≥ C0 , so that Theorem 5.1 applies. By rescaling it suffices to establish the claim
when t = 1, so that T ≥ 1. Applying Theorem 5.1 in the contrapositive, we see
that
−1
(6.1) kPN ukL∞ ∞ 3 ≤ A
t Lx ([1/2,1]×R ) 1 N
whenever N ≥ N∗ , where
7
N∗ := exp(exp(exp(AC0 ))).
We now insert this bound into the energy method. As before, we split u = ulin +unlin
on [1/2, 1] × R3 , where
ulin (t) := et∆ u(0)
and unlin := u − ulin , and similarly split ω = ω lin + ω nlin . From (2.5), (3.1) we have
(6.2) k∇j ulin kL∞ p 3 .j A
t Lx ([1/2,1]×R )
for all j ≥ 0 and 3 ≤ p ≤ ∞. We introduce the nonlinear enstrophy

Z
1
E(t) := |ω nlin (t, x)|2 dx
2 R3
for t ∈ [1/2, 1], and compute the time derivative ∂t E(t). From the vorticity equation
(3.40) and integration by parts we have
(6.3) ∂t E(t) = −Y1 (t) + Y2 (t) + Y3 (t) + Y4 (t) + Y5 (t)
42 TERENCE TAO
where
Z
Y1 (t) = |∇ω nlin (t, x)|2 dx
R3
Z
Y2 (t) = − ω nlin · (u · ∇)ω lin dx
R3
Z
Y3 (t) = ω nlin · (ω nlin · ∇)unlin dx
3
ZR
Y4 (t) = ω nlin · (ω nlin · ∇)ulin dx
R3
Z
Y5 (t) = ω nlin · (ω lin · ∇)unlin dx
R3
Z
Y6 (t) = ω nlin · (ω lin · ∇)ulin dx.
R3
From Hölder, (6.2), (3.1) we have

Y2 (t), Y6 (t) . A2 E(t)1/2 . A4 + E(t)
and similarly
Y4 (t), Y5 (t) . AE(t),
using Plancherel’s theorem to control
(6.4) k∇unlin kL2x (R3 ) . kω nlin kL2x (R3 ) .
For Y3 (t) we apply a Littlewood-Paley decomposition to all three factors to bound
X Z
Y3 (t) . PN1 ω nlin · (PN2 ω nlin · ∇)PN3 unlin dx
N1 ,N2 ,N3 R3
where N1 , N2 , N3 range over powers of two. The integral vanishes unless two of the
N1 , N2 , N3 are comparable to each other, and the third is less than or comparable
to the other two. Controlling the two highest frequency terms in L2x and the lower
one in L∞ x , and using the Littlewood-Paley localised version of (6.4), we conclude
that
X
Y3 (t) . kPN1 ω nlin kL2x (R3 ) kPN2 ω nlin kL2x (R3 ) kPN3 ω nlin kL∞ 3 .
x (R )
N1 ,N2 ,N3 :N1 ∼N2 &N3
From (3.1), (2.3), the quantity kPN3 ω nlin kL∞ 2

3 is bounded by O(AN ); for N3 ≥
x (R ) 3
−1 2
N∗ , (2.3) we have the superior bound O(A1 N3 ). We thus see that
X
−1 2
kPN3 ω nlin kL∞ 3 . A
x (R ) 1 N2 + AN∗
2
N3 .N2
and thus by Cauchy-Schwarz

X
Y3 (t) . kPN1 ω nlin k2L2x (R3 ) (A−1 2 2
1 N1 + AN∗ ).
N1
On the other hand, from Plancherel’s theorem we have

X
Y1 (t) ∼ kPN1 ω nlin k2L2x (R3 ) N12
N1
and X
E(t) ∼ kPN1 ω nlin k2L2x (R3 )
N1
and hence
Y3 (t) . A−1 2
1 Y1 (t) + AN∗ E(t).
Putting all this together, we conclude that
∂t E(t) + Y1 (t) . AN∗2 E(t) + A4 .
In particular, from Gronwall’s inequality we have
E(t2 ) . E(t1 ) + A4
whenever 1/2 ≤ t1 ≤ t2 ≤ 1 is such that |t2 − t1 | ≤ A−1 N∗−2 . On the other hand,
from a (slightly rescaled) version of (3.13) we have
Z 1
E(t) dt . A4
1/2
and hence on any time interval in [1/2, 1] of length A−1 N∗−2 there is at least one
time t with E(t) . A5 N∗2 . We conclude that
O(1)
E(t) . A5 N∗2 . N∗ ,
for all t ∈ [3/4, 1], which then also implies
Z 1
O(1)
Y1 (t) . N∗ .
3/4
Iterating this as in the proof of Proposition 3.1(iii) (or Proposition 3.1(vi)), we now
have the estimtes
O(1)
|u(t, x)|, |∇u(t, x)|, |ω(t, x)|, |∇ω(t, x)| . N∗
on [7/8, 1] × R3 . This gives Theorem 1.2.
Remark 6.1. More generally, one would expect in view of Theorem 5.1 that
any reasonable function space estimate obeyed by the linear heat equation with L3x
initial data will now also hold for classical solutions to Navier-Stokes obeying (1.2),
but with an additional loss of exp exp exp(AO(1) ) in the estimates. It seems likely
that a modification of the arguments above would be able to obtain such estimates,
particularly if one replaces the linear estimates (3.11) (or (6.2)) by more refined
estimates that involve the profile of the initial data u(0), and in particular on how
the Littlewood-Paley components kPN u(0)kL3x (R3 ) of the L3x norm of that data vary
with the frequency N . We will not pursue this question further here.
Now we prove Theorem 1.4. We may rescale T∗ = 1. Let c > 0 be a sufficiently
small constant, and suppose for contradiction that
ku(t)kL3x (R3 )
lim sup 1 c < +∞,
t→1+ (log log log 1−t )
thus we have
1
(6.5) ku(t)kL3x (R3 ) ≤ M (log log log(1000 + ))c
1−t
44 TERENCE TAO
for all 0 ≤ t < 1 and some constant M . Applying Theorem 1.2, we obtain (for c
small enough) the bounds
−1/10
(6.6) ku(t)kL∞ 3 , k∇u(t)kL∞ (R3 ) , kω(t)kL∞ (R3 ) , k∇ω(t)kL∞ (R3 ) .M (1 − t)
x (R ) x x x
(say) for all 1/2 ≤ t < 1. In particular, u is bounded in L2t L∞

x , contradicting the
classical Prodi-Serrin-Ladyshenskaya blowup criterion [Pr], [S2], [La]; one could
also use the Beale-Kato-Majda criterion [BKM] and (6.6) to obtain the required
contradiction. The claim follows.
References
[A] D. Albritton, Blow-up criteria for the Navier-Stokes equations in non-endpoint critical Besov
spaces, Anal. PDE 11 (2018), no. 6, 1415–1456.
[BS] T. Barker, G. Seregin, A necessary condition of potential blowup for the Navier–Stokes
system in half-space, Math. Ann. 369 (2017), 1327–1352.
[BKM] J. T. Beale, T. Kato, A. Majda, Remarks on the breakdown of smooth solutions for the
3-D Euler equations, Comm. Math. Phys. 94 (1984), no. 1, 61–66.
[B] J. Bourgain, Global wellposedness of defocusing critical nonlinear Schrödinger equation in the
radial case, J. Amer. Math. Soc. 12 (1999), no. 1, 145–171.
[CKN] L. Caffarelli, R. Kohn, L. Nirenberg, Partial regularity of suitable weak solutions of the
Navier-Stokes equations, Comm. Pure Appl. Math. 35 (1982), 771–831.
[C] C. Calderón, Existence of weak solutions for the Navier-Stokes equations with initial data in
Lp , Trans. Amer. Math. Soc. 318 (1990), 179–200.
[DD] H. Dong, D. Du, The NavierStokes equations in the critical Lebesgue space, Comm. Math.
Phys. 292 (2009), 811–827.
[ESS] L. Escauriaza, G. A. Seregin, V. Šverák, Backward uniqueness for parabolic equations, Arch.
Ration. Mech. Anal. 169 (2003), no. 2, 147157.
[ESS2] L. Escauriaza, G. A. Seregin, V. Šverák. L3,∞ -solutions of Navier-Stokes equations and
backward uniqueness, Uspekhi Mat. Nauk, 58 (2003), 3–44.
[GKP] I. Gallagher, G. Koch, F. Planchon, A profile decomposition approach to the L∞ 3
t (Lx )
Navier-Stokes regularity criterion, Math. Ann. 355 (2013), no. 4, 1527–1559.
[GKP2] I. Gallagher, G. S. Koch, and F. Planchon, Blow-up of critical Besov norms at a potential
NavierStokes singularity, Comm. Math. Phys. 343 (2016), 39–82.
[GIP] I. Gallagher, D. Iftimie, F. Planchon, Asymptotics and stability for global solutions to the
Navier-Stokes equations, Ann. Inst. Fourier (Grenoble) 53 (2003), no. 5, 1387–1424.
[K] T. Kato, Strong Lp -solutions of the Navier-Stokes equations in Rm with applications to weak
solutions, Math. Zeit. 187 (1984), 471–480.
[KK] C. Kenig, G. Koch, An alternative approach to regularity for the Navier-Stokes equations
in critical spaces, Ann. Inst. H. Poincar Anal. Non Linéaire 28 (2011), no. 2, 159–187.
[La] O. A. Ladyzhenskaya, On uniqueness and smoothness of generalized solutions to the Navier-
Stokes equations, Zapiski Nauchn. Seminar. POMI, 5 (1967), pp. 169–185.
[LSU] O. Ladyzenskaja, V. A. Solonnikov, N. N. Uralceva, Linear and quasilinear equations of
parabolic type, Translations of Mathematical Monographs, Amer. Math. Soc., 1968.
[Le] J. Leray, Sur le mouvement d’un liquide visqueux emplissant l’espace, Acta Math. 63 (1934),
193–248.
[Ph] N. C. Phuc, The Navier–Stokes equations in nonendpoint borderline Lorentz spaces, J. Math.
Fluid Mech. 17 (2015), 741–760.
[Pr] G. Prodi, Un teorema di unicit´ ‘ per le equazioni di Navier-Stokes, Ann. Mat. Pura Appl. 48
(1959), 173–182.
[S1] G. Seregin, A certain necessary condition of potential blow up for NavierStokes equations,
Comm. Math. Phys. 312 (2012), 833–845.
[SS] G. Seregin, V. Šverák, On global weak solutions to the Cauchy problem for the NavierStokes
equations with large L3-initial data, Nonlinear Anal. 154 (2017), 269–296.
[S2] J. Serrin, On the interior regularity of weak solutions of the Navier-Stokes equations, Arch.
Rational Mech. Anal. 9 (1962), 187–195.
[T] T. Tao, Localisation and compactness properties of the Navier-Stokes global regularity prob-
lem, Anal. PDE 6 (2013), no. 1, 25–107.
[WZ] W. Wang, Z. Zhang, Blow-up of critical norms for the 3-D Navier-Stokes equations, Sci.
China Math. 60 (2017), no. 4, 637–650.
UCLA Department of Mathematics, Los Angeles, CA 90095-1555

E-mail address: [email protected]

Quantitative Bounds For Critically Bounded Solutions To The Navier-Stokes Equations

Uploaded by

Copyright:

Available Formats

Quantitative Bounds For Critically Bounded Solutions To The Navier-Stokes Equations

Uploaded by

Document Information

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

Quantitative Bounds For Critically Bounded Solutions To The Navier-Stokes Equations

Uploaded by

Copyright:

Available Formats

Quantitative bounds for critically bounded solutions to the

arXiv:1908.04958v2 [math.AP] 10 Jul 2020

Abstract. We revisit the regularity theory of Escauriaza, Seregin, and Šverák

The Navier-Stokes system enjoys the scaling symmetry (u, p, T ) 7→ (uλ , pλ , λ2 T )

for some A ≥ 2. Then we have the derivative bounds

whenever 0 < t ≤ T , x ∈ R3 , and j = 0, 1, where ω := ∇ × u is the vorticity field.

Remark 1.3. It is not difficult to iterate using Schauder estimates in Hölder

for an absolute constant c > 0.

is bounded from below uniformly in n. Furthermore, by using a “bounded total

Figure 1. A schematic depiction of the main argument. Start-

Given a Schwartz function f : R3 → R, we define the Fourier transform

Lemma 2.1 (Multiplier theorem). Let N > 0, and let m : R3 → C be a smooth

whenever 1 ≤ p ≤ q ≤ ∞ and f : R3 → R is a Schwartz function. More generally, if

of the above estimate, whenever 1 ≤ p1 ≤ q1 ≤ ∞ and 1 ≤ p2 ≤ q2 ≤ ∞ are such

Proof. By homogeneity we can normalise M = 1; by scaling (or dimensional

for some A ≥ C0 . We adopt the notation

Then there exists (t2 , x2 ) ∈ [t0 − T, t1 ] × R3 and N2 ∈ [A−1

(v) (Iterated back propagation) Let x0 ∈ R3 and N0 > 0 be such that

Then for every A4 N0−2 ≤ T1 ≤ A−1

(t1 , x1 ) ∈ [t0 − T1 , t0 − A−1

As C0 is assumed large, any polynomial combination of A = A0 , A1 , . . . , Aj−1

From (3.7) and Duhamel’s formula one has

for any 3 ≤ p ≤ ∞ and j ≥ 0.

where the quantity (∇unlin ) · (u ⊗ u) is defined in coordinates as

and hence by Young’s inequality

Splitting u ⊗ u − unlin ⊗ unlin = ul ⊗ u + unlin ⊗ ul and using (3.1), (3.9), (3.11)

By Plancherel’s theorem this implies in particular that

From (2.3), Hölder’s inequality, and (3.11), (3.15) one has

kP≤N unlin ⊗ P>N unlin kL1t L2x ([−1/2,1]×R3 ) ,

× kP>N unlin kL2t L2x ([−1/2,1]×R3 )

and hence by (2.3), the triangle inequality, and Young’s inequality

Putting all this together, we conclude that

+ N −1 kP≤N unlin k2L2 L∞ ([−1/2,1]×R3 )

+ N 2 kP>N unlin k2L2 L2 ([−1/2,1]×R3 ) .

By (2.3) and Cauchy-Schwarz we have

Summing in N , and using the triangle inequality followed by (3.14), we conclude

From (3.9) and (2.3) we also have

which gives (ii).

and Hölder’s inequality, one concludes in particular that

and hence by (3.11)

also from Sobolev embedding and (3.19) one has

so from Hölder’s inequality we conclude that

Now we differentiate (3.7) to conclude that

From (3.22), (3.24), Leibniz and Hölder one has

for t ∈ [τ (0.4), τ (1)], and hence by Hölder’s inequality

Assume for contradiction that the claim fails, then we have

for all A−1

and hence in the range N ≥ A−1

(3.26) kPN ukL∞ L3/2 ([−A3 ,0]×B(0,A4 )) . A2 N −1 .

From (2.4), (3.1), and Hölder as before we have

and hence by (3.1)

Since P̃N (P≤N/100 u(t0 ) ⊗ P≤N/100 u(t0 )) vanishes, we can write

Arguing as before we have

gle inequality, Hölder, and (3.28) to control this contribution by

From (2.2), (3.1) one has

and hence by the pigeonhole principle we have

for some −A3 /4 ≤ t0 ≤ 0.

giving the required contradiction. This establishes (iv).

for all i = 1, . . . , n, with ti ∈ [t0 − T /2, t0 ] and Ni ≥ A3 T −1/2 for i = 0, . . . , n − 1