A variation on a theme of Caffarelli and Vasseur

A VARIATION ON A THEME OF CAFFARELLI AND VASSEUR

ALEXANDER KISELEV AND FEDOR NAZAROV
Q>>^ , Dedicated to Nina Nikolaevna Uraltseva

o
o

I Abstract. Recently, using DiGiorgi-type techniques, CafTarelli and Vasseur [II showed that a

^jQi certain class of weak solutions to the drift diffusion equation with initial data in gain Holder

^ I continuity provided that the BMO norm of the drift velocity is bounded uniformly in time. We

show a related result: a uniform bound on BMO norm of a smooth velocity implies uniform
bound on the norm of the solution for some (3 > 0. We use elementary tools involving
control of Holder norms using test functions. In particular, our approach offers a third proof
of the global regularity for the critical surface quasi-geostrophic (SQG) equation in addition to
[5] and [T].

<

<

(N
>

^ ■ the latter interpretation in mind, let us recall the definition of the BMO norm:

1. Introduction

In the preprint [1], Caffarelli and Vasseur proved that certain weak solutions of the drift
diffusion equation with (— A)^/^ dissipation gain Holder regularity provided that the velocity
u is uniformly bounded in the BMO norm. The proof uses DiGiorgi-type iterative techniques.
(yr) [ The goal of this paper is twofold. First, we wanted to provide additional intuition for the
C^I . Caffarelli- Vasseur theorem by presenting an elementary proof of a related result. Secondly, we
^ . think that, perhaps, the method of this paper may prove useful in other situations.

I Everywhere in this manuscript, our setting for the space variable will be (i— dimensional

O ' torus, T*^. Equivalently, we may think of the problem set in with periodic initial data. With

uh(

BMO = supsgRd^ / \f{x) -Jb\ dx. (1.1)

s> ; \^\ J

rS B
H ;

^ Here B stands for a ball in R*^, \B\ for its volume and f ^ for the mean of the function / over
B.

Theorem 1.1. Assume that 9{x,t), u{x,t) are C°°{T'^ x [0,T]) and such that

et = {u-v)e-{-Ay/^e (1.2)

holds for any t > 0. Assume that the velocity u is divergence free and satisfies a uniform hound
||'w("5 '^)||bmo ^ B fort G [0,T]. Then there exists (3 = P{B,d) > such that

\\9{x,t)\\c,^j.)<C{B,9{x,0)) (1.3)

for any t G [0, T].

Remark. In fact, we get control of Holder continuity in terms of just norm of 9q if we are
willing to allow time dependence in (11. 3p . Namely, the following bound is also true:

\\9{xMcnf^)<c{B,\\9{xMW)^H^.tY'~^- (i-4)

Department of Mathematics, University of Wisconsin, Madison, WI 53706, USA; email: kise-
lev@math.wisc.edu, nazarov@math.wisc.edu.

1

2

ALEXANDER KISELEV AND FEDOR NAZAROV

Thus uniform bound on the BMO norm of u imphes uniform bound on a certain Holder norm
of 9. The dimension d is arbitrary.

Our result is different from [T|. For one thing, [Ij contains a local regularization version,
which we do not attempt here. However our proof is simpler, and is quite elementary. At the
expense of extra technicalities, it can be extended to more general settings.

Theorem 11.11 can be used to give a third proof of the global regularity of the critical surface
quasi-geostrophic (SQG) equation, which has been recently established in [5] and [1]. We discuss
this in Section [51 Throughout the paper, we will denote by C different constants depending on
the dimension d only.

2. Preliminaries

First, we need an elementary tool to characterize Holder-continuous functions. Define a
function VL{x) on T'^ by

= { j_ ui > 1 /9 (2-1)

(thinking of M'^ picture, fl is defined as above on a unit cell and continued by periodicity). Let
> 1 be a parameter to be fixed later.

Definition 2.1. We say that a function (f defined on T*^ belongs to Ur{T'^) if

\Mx)\\l^<^, (2.2)

if{x)dx = (2.3)
\\>p(x)\\l' < 1 (2.4)
/ M.)|n(.-..)<ix<.V. to. =„,„e x„ . T^. (2.5)

Observe that the classes Ur are invariant under shift. We will write f{x) G a'Ur{T'^) if
f{x)/a G Ur(T'^). The choice of the exponent 1/2 in (12. ip and (12. 5p is arbitrary and can be
replaced with any positive number less than 1 with the appropriate adjustment of the range of
/3 in Lemma 12.21 below.

The classes Ur can be used to characterize Holder spaces as follows. Let us denote

11/11 l/(^)

||/||c/3(Td) = sup^y^jd — , _ ,^ — , (2.6)

\x y\

omitting the commonly included on the right hand side term. The seminorm (12. 6p

is sufficient for our purposes since 6 remains bounded automatically, and moreover we could
without loss of generality restrict consideration to mean zero 6, invariant under evolution, for
which (12. 6p is equivalent to the usual Holder norm.

Lemma 2.2. A bounded function 6{x) is in C^{T'^), < f3 < 1/2, if and only if there exists a
constant C such that for every < r < 1,

9{x)ip{x) dx
fd

< Cr^ (2.7)

for all ip EUr- Moreover,

\e\\cP(^Td) < C(/5)sup^g^^^ o<^<ir

-/3

6{x)(p{x) dx
jd

(2.8)

A VARIATION ON A THEME OF CAFFARELLI AND VASSEUR 3

Remark. The lemma holds for each fixed A in (12.21) . It will be clear from the proof that the
constant C in (12. 8p does not depend on A provided that A was chosen sufficiently large.

Proof. Assume first that 9 G C^. Consider any ip EUr, and observe that

X — Xq \^\(f{x)\ dx.

Ijd. Jjd Jjd

Using Holder inequality, we get

/ \x — XQ\^\'^{x)\dx < I / |v9(a;)| (ix I I / |x — Xo|"^^^|v5(a;) | 1

<c( [ \^{x)\dx^ ( [ n{x~xo)\^{x)\dx^ (2.9)

Due to (12.41) and (12. Sp . the right hand side of (12. 9p does not exceed r^.

For the converse, consider a periodization 6p of 6 in W^. Recall a well known characterization
of Holder continuous functions in (see e.g. [6j):

eeC ^ ||^||L-<g, ||A,-(0)|u<g2-«, vj. (2.10)

Moreover, if the right hand side of (I2.10p is satisfied, then H^Ucs ^ CQ.
Here Aj are the Littlewood-Paley projections:

A,(^) =9*^2-^,

where ^t(a;) = t~'^^{x/t), and $(0 = viO - ^(20, with t] G C^, < r]{^) < 1, r]{^) = 1
if 1^1 < 1 and ri{^) = 0, |^| > 2. Observe that \& is in the Schwartz class S, f^^"^ dx = 0,
J^a dx < C, J^a |a;|^/2|^(x)| dx < C and < C. Let us define

then ^j{x) G W2-j (T ) if c is sufficiently small (independently of j). Moreover,

/ 0{x)-^j{x -y)dx = c 6p{x)'^2-j{x - y)dx. (2.11)

By assumption, the left hand side in (12. lip does not exceed Q2~^^. Thus, by the criterion (I2.10p .
6p is and so is 6'. The remark after (I2.10p implies that (12.80 is true. □

The proof of Theorem 11.11 which we will outline in the beginning of the next section, relies
on transfer of evolution on the test function. Here is an elementary lemma that allows us to
do that. Let (p\x, s) be the solution of

ipl = -{u{x,t-s)-V)ip'-i-Ay/^ip\ ip\x,0) = ip{x). (2.12)

Lemma 2.3. Let 6^0, e C7'^(T"'), and let e{x,t) be the solution of ([L2]) with 9{x,0) = 9o{x).
Then we have

9{x,t)(p{x) dx = / 6o{x)(p (x^t) dx.
jd Jjd

Proof. We claim that for < s < t, the expression

e{x,t - s)<^\x,s)dx (2.13)

remains constant. A direct computation using (I2.12p . (II. 2p and the fact that u is divergence
free shows that the s-derivative of (12.130 is zero. Substituting s = and s = t into (I2.13P
proves the lemma. □

alexander kiselev and fedor nazarov
3. The proof of the main result

Let us outline our plan for the proof of Theorem ll.il Conceptually, the proof is quite simple:
integrate the solution against a test function from Ur, transfer the evolution on the test function
and prove estimates on the test function evolution.

The key to the proof of Theorem 11.11 is the following result.

Theorem 3.1. Letv{x,s) G C°°(T'^ x [0,T]) be divergence free d- dimensional vector field, and
let ip{x, s) solve

= -(v ■ V)V' - (-A)^/V, 0) = ^{x).

Assume that

maXs(z[o^T]\\vi-,s)\\BMO < B.

(3.1)

Then the constant A = A{B, d) in (12. 2p can be chosen so that the following is true.

Suppose ip G Wr(T'^), < r < 1. Then there exist constants 5 and i^T > 0, which depend only
on B and dimension d, such that

tp{x, s) G

r

5/K

r + Ks

if r + Ks < 1 and il){x, s) G r^/^Ui(T'^) otherwise.

(3.2)

Let us assume Theorem 13.11 is true and prove Theorem 11.11
Proof of Theorem \l.l[ Let (3 = 5/ K. Since 9{x, 0) is smooth, we have

Yd

9{x, 0)ip{x) dx

< C{9{x,0)y

(3.3)

for all (p{x) G W^(T°'), < r < 1. But by Lemma EJl

9{x,t)(p{x) dx = / 9{x,0)(p (x^t) dx.
jd Jjd

By Theorem 13. 11 V5*(x, t) belongs to ( ^^j^f- ) ^ Ur+Kt (T'^) if r + Kt < 1, and to r'^Wi(T'^) otherwise.
Then (13. 3 p implies that

jd

9{x, t)ip{x) dx

< C{9{x,0)y,

for all ip{x) G Ur{T'^), < r < 1.

(3.4)

□

Observe that C{9{x, 0)) will depend only on the norm of 9{x, 0) if we are willing to allow
time dependence in (13. 4p :

9{x, t){p{x) dx

9{x, 0)^\x,t) dx

jd

< \\9{x,0)\\Liy\x,t)l

< A{B,dymm{l,r + Kty-^ < C{B,d, ||^(x, 0)||ii)min(l, t)"

This proves the bound (ll.4p in the Remark after Theorem 11.11
Thus it remains to prove Theorem 13. 1[

a variation on a theme of caffarelli and vasseur 5

4. The Evolution of the Test Function

The proof of Theorem 13.11 is based on the following lemma, which looks at what happens
over small time increments.

Lemma 4.1. Under assumptions of Theorem \3.1\ we can choose A = A{B,d) so that the
following is true. There exist positive 6, K and 7 (dependent only on B and d) such that for
allO<s< -fr, ifi'{x,0) E Ur{T'^), < r < 1, then

ilj{x,s)e fl - - (4.1)

The estimate (14. ip is valid as long as r + Ks < 1; otherwise the solution just remains in Ui.

Proof. We have to check four conditions. First, the equation for ip preserves the mean zero
property, so that /^^ ^J^x, s) dx = Q for all s.

Next, let us consider the L°° norm. Set M(s) = ||'?/'(-, s)||loo. Consider any point Xq where
the maximum or minimum value is achieved. Without any loss of generality, we can assume
Xq = 0, V(0,s) = M(s). Then

Here we used the well-known formula for the fractional Laplacian (see e.g. [3J). Since s) ||Li(Td) <
1 (see the argument below on the norm monotonicity) , it is clear that the contribution to the
right hand side of f l4.2p from the central period cell is maximal when ip{y) is the characteristic
function of a ball of radius cM{s)~^^'^ centered at the origin. This gives us the estimate

dsipsiO, s) < -C M{s)\y\~'^-Uy<-CiM{s)^ + C2rM{s)<-CM{s)'^, (4.3)

The argument is valid for all sufficiently large M(s), which is the only situation we need to
consider provided A was chosen large enough. The same bound holds for any point xq where
M{s) is attained and by continuity in some neighborhoods of such points. So, we have (14. 3 p in
some open set U. Due to smoothness of ip, away from U we have

ma.x\ilj{x,T)\ < M{t)

for every r during some period of time after s. Thus we obtain that

^M(s) < -CM^(s), M(0) < Ar-'^. (4.4)
ds

This is valid for all times while M{s) remains sufficiently large. Solving (14. 4p . we get an estimate

Mis) < < Ar-''(1 - CA'/^'r-h)

for all sufficiently small s. This implies

-, s) IIloo < Ar-\l - CA^''^r-h), (4.5)

for all sufficiently small s < 7(^4, d)r. Observe that 7 is independent oi ip 01 v other than
through the value of A, which will be chosen below depending on the value of B only. The
estimate (14. 5 p agrees with the properties of the (1 — y)Wr+A's(T°') class provided that

5 + dK < CA^I'^. (4.6)

ALEXANDER KISELEV AND FEDOR NAZAROV

Next, we consider the concentration condition J^^il^x — Xo)\ijj{x) \ dx < r^l"^. Consider
x{s) G T'^ satisfying

v{y,s)dy, x(0) = Xq. (4.7)

X [s]

t\ JBr{x{s))

Here Br{x) stands for the ball of radius r centered at x, and \Br \ is its volume. We will estimate
Jj.a^{x — x{s))\ip{x, s)\ dx. Let us write i/j{x) = ip+{x) — i/j^{x), where ip±{x) > and have
disjoint support. Let us denote '?/'±(a;, s) the solutions of (13.11) with ip±{x,0) = ip±{x). Then
due to linearity and maximum principle, \ip{x, s) \ = \ip+{x, s) —ip-{x, s)\ < ip+{x, s) + ■?/'_ (x, s),
and so

il{x — x{s))\il>{x, s)\ dx < / il{x — x{s))^p+{x, s) dx + / il{x — x{s))^jj-{x, s) dx. (4.8)

fd

Let us estimate the first integral on the right hand side of (14. Sp . the second can be handled the
same way. We have

Q{x — x{s))4'^ dx

{n{x - x{s)) {{-V ■ V)ip+ - (-A)^/V+) - V(^](x - x(s))) ■ x'{s)ip+) dx

jd

< c

V(^](x - x{s))) ■ {v - vb,(x{s))) ip+ dx

-A)i/2^(x-x(s))^+ dx

\x — X[S

jd

VBrix{s))\\i'+\dx + / \x-x{s)\ ^^'^\ip+\dx

(4.9)

We used the divergence free condition on v and (14. 7p in the second step, and estimated |Vfi(x —
Xo)\ < C\x — xq\~^^'^, \{—AY^'^n{x — xq)\ < C\x — XqI^^"^. Let us consider the two integrals in
(14.91) . Since H^/^+Hii < 1/2 and ||?/'+||loo < y4r~°', the integral J^^ a; (s)|~^/^| t/'+I is maximal
when tp^ is a characteristic function of a ball centered at x{s) of radius crA"^/"^. This gives
an upper bound of Cr~^^'^A^^'^'^ for this integral. To estimate the first integral in (14.91) . split
T'^ = U^^^Ek, where

Ek = {x: r2'^"^ < |2;-x(s)| <r2'=}nr. A; > 0, Eo = Br{x{s)).
Recall (see e.g. [6]) that for any BMO function /, any ball B, and any 1 < p < oo.

11/ - /i?IUf(B) < Cp\B\^^^\\J WBMO-

(4.10)

By Holder's inequality.

\x - x{s)\ ^^^\v - vb,.{x{s))\\^+\ dx <

-1/2 I

'Br(x{s))

ja; - X[S)\ ^'"||Lf(B.(x(s))||^^ -^JB,(x(s))||L-{iJ,(x(s)))||^+||L9(B,(x(s))),

where + + q^^ = 1. Now

IIV^+IU^ <

Using ( 14.10p . we also see that

Finally, for any p < 2d,

\\v - VB,{xis))\\L-{Brix(s))) < C{z,d)r-B.
\x - x{s)\''^/'^\\LP{Br{xis)) < C{p,d)rp~^.

A VARIATION ON A THEME OF CAFFARELLI AND VASSEUR

2d

Taking z very large, and p very close to 2(i, we find that for any q > 2d^^ have

JBrixis))

(4.11)

where a is any number greater than ^. Furthermore, for A; > 0,

\x - x{s)\-'^/^\v -VBr{x{s))\\'ip+\ dx < C2-^l^r-^l^ / \v -vb,^x{s))\\'^A^) \ dx <

fj2-k/2^-y2 I I \v-VB^^,(ccis))\\^+{x)\dx+ \vB^,^,ix(s)) -VBr{x(s))\\i^+{x)\dxU.12)

\JB^^,{xis)) Jb^^,(x{s)) J

Recall that (see, e.g., [6])

\vB^2kixis)) -VBrix(s))\ < Ck\\v\\BMO-

Therefore the last integral in fl4.12p does not exceed CkB. The first integral can be estimated

by

where q is any number greater than 1. Thus in particular

f \x- x{s)\-^'^\v -VB,ix(s))\\iJ+{x)\dx < CE2-=^'=/^5(A;2-'=/^ + A^^^^y-^^^ (4.13)
if g = Adding (HTT]) and (KT^ . we obtain

\x - x{s)\''/^\v - VB^ixis))\\Mx, s)\dx< CBA'/^''r~'/\

provided that A is large enough (the exponent for A can be anything greater then ^). Coming
back to (USD and (gSD, we see that

[ |x-x(s)|i/2|^(x,s)|rfx <r^/2^Csr-i/2^Ai/2d^^^3/4d)_ ^4_^4)

This is consistent with the (1 — ^)Ur+Ksi'^'^) class if

(^1 - (r + Ksf/' > + Csr-'/^A'/^'' + BA'/^").
Provided that 7 is chosen sufficiently small, this condition reduces to

y - 5 > C( ^ 5^3/4d) ^

Finally, we consider the norm. Recall (see e.g. [3]) that for a function ip{x),

(-A)VV.(.) = Ihn J] / ^ti^^l^dy. (4.16)

e-+o ^-^ J \x ~ y — n\'^^^

"e^Vn|x-y|>e

Let S be the set where ipi^x, s) = 0, and define

D± = {x E T'^\±ip{x,s) > 0}.

The sets S and D± depend on s, but we will omit this in notation to save space. Due to (13. ip
and incompressibility of f , we have

dsm;s)hr= [ ^ip^{-vVi;{x,s)-i-AY/'i;{x,s)) dx +

ALEXANDER KISELEV AND FEDOR NAZAROV

^tp^(-A)'/'^P(x,s)dx+ [ |(-A)VV(x,.)|dx. (4.17)

The integral over S is of course nonzero only if the Lebesgue measure of S is positive. Substi-
tuting fl4.16p into fl4.17l) and symmetrizing with respect to x, y we get

-- lim

2 e^O

/ (

^Pix,s) ipiy,s)

{x,s)\ \ij{y,s)\

((Td\S)x(T<*\S))n|a;-s/|>e

E

T'i\S

' 'ix,s)

i){x,s)

g \x — y — n\'^^^

dy \ dx +

E

\x — y —

fd\s \x-y-n\

-dy

dx.

Observe that the expression under the first integral in (14.181) is non-negative for all x, y, and it
is positive if ip{x,s) and ip{y,s) have different signs. Also, observe that

E

jd\s \x-y-n

d+l

dy

4^{y,s)

Id+ \x-y-n\

-dy

'^{y,s)

D_ \x-y-n

d+l

dy

Therefore, the combined contribution of the last line in (14. ISp over every cell is less than or
equal to zero. Leaving only the central cell contributions in (14.181) . we get

ip{x,s)

dy

\x — y\^^

-dydx+ I ^lj{y,s)

dx

\x - y\

-dxdy -{A.19)

D-

D4

X, S]

D+UD-

i^{y,s)

Id- F - y\

-dy

Without loss of generality, we can assume that 1 > \\ip{-, s)\\li > 9/10 for every s we consider,
since otherwise the L} condition is already satisfied. Also, due to (I4.14p we can assume that
Vl{x — x{s))\il){x, s)\dx < ^T^^'^ provided that the time interval [0,7r] that we consider is
sufficiently small, with 7 = 7(v4, B). These two bounds imply that /Trdn|a;-a'(s)|<4oor s)\dx >
4/5. The mean zero condition leads to

±

i){x,s)dx > 3/10.

(4.20)

'D±n{|x-x(s)|<400r}

Let us denote D± = D±f^{\x- x{s)\ < 400r}, S = S f\ {\x - x{s)\ < 400r}. Observe that if
xeS, then, by fICTD .

±

i'jy^s)

d+l

D± \x — y

dy>±

^{y,s)

D± F ~ y\

-dy> Cr

This implies that due to cancelation in the last term of (14.190 . we can estimate the last line of
(I4.19P from above by —C\S\r~'^~^ . Reducing the integration in the second line of (I4.19P to D±,

we obtain
5.||^(-,s)|Ui<-Cr-'^-^

ip{x, s) dx + \D^

D+

D-

i){x, s) dx +\S\] < -cr-\ (4.21)

where c is a fixed positive constant. Here in the last step we used (I4.20p and \D^ \ + \D_ \ + \S\>
Cr'^. The estimate (lOTD is consistent with (1 - ^)W,.+x,(T'^) class if 5 < c.

A VARIATION ON A THEME OF CAFFARELLI AND VASSEUR

9

It remains to observe that, if A = A{B, d) is chosen sufficiently large, one can indeed find
K and 5 so that the conditions fl4.6l) . fl4.15l) and the 5 < c condition arising from the L} norm
estimate are all satisfied. It is also clear from the proof that (14.11) then holds for all s < 7(-B, d)r.
The only restriction from above on the value of r comes from the L°° norm condition, which
has to be consistent with and concentration conditions. For convenience, we chose to cap
the value of r at 1. □

The proof of Theorem 13.11 is now straightforward.

Proof of Theorem \3.1[ From Lemma [4.11 it follows that for any s > 0, ip{x, s) G f{s)Ur+Ksi^'^)
provided that f'{s) > — ^.^j^^ fis). Solving this differential equation, we obtain that the factor

/(■s) = {^^y^^ is acceptable. □

Theorem 11.11 provides an alternative path to the proof of existence of global regular solutions
to the critical surface quasi-geostrophic equation:

where 6 : — is a periodic scalar function, and -Ri and R2 are the usual Riesz transforms
in M^. Indeed, the local existence and uniqueness of smooth solution starting from periodic
initial data is known (see e.g. [1]). The L°° norm of the solution does not increase due to
the maximum principle (see e.g. [3]), which implies uniform bound on the BMO norm of the
velocity. Since the local solution is smooth, one can apply Theorem 1 1.1[ This, similarly to [1],
implies a uniform bound on some Holder norm of the solution 6. This improvement over the
L°° control is sufficient to show the global regularity (see |1] or [2] for slightly different settings
which can be adapted to our case in a standard way).

One can pursue a number of generalizations of Theorem 11.11 for instance reducing assump-
tions on smoothness of solution, velocity, or initial data. However we chose to present here the
case with the most transparent proof containing the heart of the matter. As follows from the
proof, the role of the BMO space is mainly the right scaling: the BMO is the most general
function space for which (14.101) is available. The BMO scaling properties are of course also
crucial for the proof of |lj to work.

Acknowledgement. Research of AK has been supported in part by the NSF-DMS grant
0653813. Research of FN has been partially supported by the NSF-DMS grant 0501067. AK
thanks for hospitality the Department of Mathematics of the University of Chicago, where part
of this work was carried out.

[1] L. CafFarcUi and A. Vasseur, Drift diffusion equations with fractional diffusion and the quasi-geostrophic

equation, arXiv:math/0608447, 25 pages
[2] P. Constantin and J. Wu, Regularity of Holder continuous solutions of the supercritical quasi-geostrophic

equation, arXiv:math/0701592, 12 pages
[3] A. Cordoba and D. Cordoba, A maximum principle applied to quasi-geostrophic equations, Commun. Math.

Phys. 249 (2004), 511-528

[4] H. Dong, Higher regularity for the critical and super- critical dissipative quasi-geostrophic equations,

arXiv:math/0701826, 18 pages
[5] A. Kiselev, F. Nazarov and A. Volberg, Global well-posedness for the critical 2D dissipative quasi-geostrophic

equation, Inventiones Math. 167 (2007) 445-453
[6] E. Stein, Harmonic Analysis, Princeton University Press, 1993

5. The Critical SQG equation and further discussion

References