As remarked above, we may take localizing functions to be independent of
$t,$
since were a derivative in
$t$
to land on such a localizer, one would be in the region where the operator was clearly elliptic and the analyticity of the solution
$u$
is well known. We denote such an Ehrenpreis type localizing function by
$\phi \left(x\right)={\phi}_{N}\left(x\right)$
subject to the usual growth of its derivatives:
$\left{D}^{\alpha}\phi \right\le {C}^{\left\alpha \right+1}{N}^{\left\alpha \right}$
for
$\left\alpha \right\le N,$
where the constant
$C$
is (universally) inversely proportional to the width of the band separating the regions where
$\phi \equiv 0$
and
$\phi \equiv 1.$
Next, since
$P$
is
${C}^{\infty}$
hypoelliptic we may assume that
$u$
is smooth and proceed to obtain estimates for
${D}_{t}^{p}u$
and
${D}_{{x}_{j}}^{p}u$
near
$0.$
The a priori estimate for
$P,$
while subelliptic, is more importantly maximal: for
$v\in {C}_{0}^{\infty},$
(2.1)
∥
D
t
v
∥
L
2
2
+
∑
1
2
∥
t
D
x
j
v
∥
L
2
2
+
∥
D
θ
(
x
)
v
∥
L
2
2
(
+
∥
v
∥
1
/
2
2
)
≤
C

〈
P
v
,
v
〉

+
C
∥
v
∥
L
2
2
.
Setting
$v=\phi {D}_{t}^{p}u$
, to begin with, we obtain
$$\parallel {D}_{t}\phi {D}_{t}^{p}u{\parallel}_{{L}^{2}}^{2}+{\sum}_{1}^{2}\parallel t{D}_{{x}_{j}}\phi {D}_{t}^{p}u{\parallel}_{{L}^{2}}^{2}+\parallel {D}_{\theta \left(x\right)}\phi {D}_{t}^{p}u{\parallel}_{{L}^{2}}^{2}(+\parallel \phi {D}_{t}^{p}u{\parallel}_{1/2}^{2})\le C\langle P\phi {D}_{t}^{p}u,\phi {D}_{t}^{p}u\rangle +C\parallel \phi {D}_{t}^{p}u{\parallel}_{{L}^{2}}^{2}\le C\langle \phi {D}_{t}^{p}Pu,\phi {D}_{t}^{p}u{\rangle}_{{L}^{2}}+C{\sum}_{1}^{4}\left\right([{X}_{j}^{2},\phi {D}_{t}^{p}]u,\phi {D}_{t}^{p}u\left)\right+C\parallel \phi {D}_{t}^{p}u{\parallel}_{{L}^{2}}^{2}.$$
Now crucial in the brackets are the quantities (recall that we may take
$\phi $
independent of
$t,$
and clearly to localize in
$x$
we may take it to be purely radial in
$({x}_{1},{x}_{2})$
i.e. we choose
$\phi $
to be constant on the integral curves of
${X}_{4}$
), so that
${X}_{4}\phi =0$
,
$$[{X}_{1},\phi {D}_{t}^{p}]=[{X}_{4},\phi {D}_{t}^{p}]=0,LocalAnalyticityforaClassofNonsymplecticSumsofSquares$$
and
$$[{X}_{j},\phi {D}_{t}^{p}]=t{\phi}^{\prime}{D}_{t}^{p}\underline{p}\phi {D}_{x}{D}_{t}^{p1},j=2,3.$$
In the first case, we may ignore the factor
$t$
and recognize the passage from one power of
${D}_{t}$
to a derivative on
$\phi $
as an acceptable swing, which, upon iteration, will lead to
${C}^{p+1}{N}^{p}\sim {C}^{p+1}p!$
when
$p\sim N.$
The second term takes two powers of
${D}_{t}$
(e.g.,
${X}_{1}$
from the estimate and one power of
${D}_{t}$
and produces a factor of
$p$
and a `bad' vector field
${D}_{x}.$
Iterating this will yield
$p!!{D}_{x}^{p/2}u\sim p{!}^{1/2}{D}_{x}^{p/2}u$
on the support of
$\phi .$
On the other hand, setting
$v=\phi {D}_{{x}_{j}}^{q}u,$
with perhaps
$q=p/2,$
or, better,
$v=\phi {\Delta}_{x}^{q/2}u$
, where we write
${\Delta}_{x}={\sum}_{j}{D}_{{x}_{j}}^{2}$
,
$$\parallel {D}_{t}\phi {\Delta}_{x}^{q/2}u{\parallel}_{{L}^{2}}^{2}+{\sum}_{1}^{2}\parallel t{D}_{{x}_{j}}\phi {\Delta}_{x}^{q/2}u{\parallel}_{{L}^{2}}^{2}+\parallel {D}_{\theta \left(x\right)}\phi {\Delta}_{x}^{q/2}u{\parallel}_{{L}^{2}}^{2}(+\parallel \phi {\Delta}_{x}^{q/2}u{\parallel}_{1/2}^{2})\le C\langle P\phi {\Delta}_{x}^{q/2}u,\phi {\Delta}_{x}^{q/2}u\rangle +C\parallel \phi {\Delta}_{x}^{q/2}u{\parallel}_{{L}^{2}}^{2}\le C\langle \phi {\Delta}_{x}^{q/2}Pu,\phi {\Delta}_{x}^{q/2}u{\rangle}_{{L}^{2}}+C{\sum}_{1}^{4}\langle [{X}_{j}^{2},\phi {\Delta}_{x}^{q/2}]u,\phi {\Delta}_{x}^{q/2}u\rangle +C\parallel \phi {\Delta}_{x}^{q/2}u{\parallel}_{{L}^{2}}^{2},$$
and now the crucial brackets are
$$[{X}_{1}^{2},\phi {\Delta}_{x}^{q/2}]=0,[{X}_{4}^{2},\phi {\Delta}_{x}^{q/2}]=0$$
and
$$[{X}_{j}^{2},\phi {\Delta}_{x}^{q/2}]=2{X}_{j}t{\phi}^{\prime}{\Delta}_{x}^{q/2}{t}^{2}{\phi}^{\left(2\right)}{\Delta}_{x}^{q/2},j=2,3$$
(where we have used rather heavily the fact that
${X}_{4}\phi =0$
since
$\phi $
depends only on
$x,$
and radially so, and that in fact
$[{D}_{\theta},{\Delta}_{x}]=0.$
) This last line leads to two kinds of terms, namely, for
$j=2,3$
, A. Bove, M. Derridj and D. S. Tartakoff
$$\langle 2{X}_{j}t{\phi}^{\prime}{\Delta}_{x}^{q/2}u,\phi {\Delta}_{x}^{q/2}u\rangle $$
and
$$\langle {t}^{2}{\phi}^{\left(2\right)}{\Delta}_{x}^{q/2}u,\phi {\Delta}_{x}^{q/2}u\rangle .$$
Morally, these terms show the correct gain to lead to analytic growth of derivatives, namely one must think of
$t{\Delta}^{1/2}$
as an
${X}_{j}$
with
$j=2$
or
$j=3,$
and so in the first term above one merely integrates by part noting that
${X}_{j}^{*}={X}_{j}$
and obtains, after a weighted Schwarz inequality, a small multiple of the left hand side of the a priori inequality and the square of a term with one derivative on
$\phi $
and
$t{\Delta}^{1/2}$
and
$q$
reduced by one, though one more commutator is required to make the order correct, and this will introduce another derivative on
$\phi $
and
$q$
again decreased by one unit, etc. The second term is of a different character, though the same observation reduces us essentially to
$$\langle X{\phi}^{\left(2\right)}{\Delta}_{x}^{(q1)/2}u,X\phi {\Delta}_{x}^{(q1)/2}u\rangle $$
in which instead of each copy of
$\phi $
receiving one derivative, we have two derivatives on one copy and none on the other. Fortunately, the Ehrenpreistype cutoff functions may be differentiated not merely
$N$
times with the usual growth but
$2N$
or
$3N$
with no change—so in the above inner product we include a factor
$CN$
with the copy of
$\phi $
which remains undifferentiated and a factor of
$(CN{)}^{1}$
with the other. The estimates work out just as before. Local Analyticity for a Class of Nonsymplectic Sums of Squares
3 Proof in the general case ( 1.6 )
The general case is not more complicated than the first, simplest case, with
${\Delta}_{t}={\sum}_{m=1}^{\ell}{D}_{{t}_{m}}^{2}$
requiring us to consider each
$t$
variable separately;
${\Delta}_{x}$
is replaced by
${\sum}_{i,j=1}^{k}{a}_{ij}(t,x){X}_{i}{X}_{j}$
and the square of the angular derivative by the sum
${\sum}_{i,j=1}^{s}{b}_{ij}(t,x){Y}_{i}{Y}_{j}$
. Thus we merely give a brief sketch of the proof.
We have an
a priori estimate of the form
(3.1)
∑
j
=
1
ℓ
∥
D
t
j
u
∥
2
+
∑
j
=
1
ℓ
∑
h
=
1
k
∥
t
j
X
h
u
∥
2
+
∑
j
=
1
s
∥
Y
s
u
∥
2
+
∥
u
∥
1
/
2
2
≤
C
(

〈
P
2
u
,
u
〉

+
∥
u
∥
2
)
.
Let us write, as introduced above,
$$\Lambda =apositivedefinitequadraticexpressioninthe{X}_{i}$$
and
$$A={\sum}_{i,j=1}^{k}{a}_{ij}(t,x){X}_{i}{X}_{j}.$$
Analogously to what has been done before we denote by
$\phi $
a cutoff function of Ehrenpreis type, constant on the integral manifold of the fields
${Y}_{1},\dots ,{Y}_{s}$
and independent of
$t$
.
The problem of estimating the growth rate of the derivatives of
$u$
then reduces to estimating
$$\parallel \phi {\Lambda}^{q/2}u\parallel ,$$
for every natural number
$q\le N$
, where
$\left{\partial}^{\alpha}\phi \right\le {C}^{1+\left\alpha \right}{N}^{\left\alpha \right}$
, for
$0\le \left\alpha \right\le 3N$
.
We have thus to examine the structure of the commutator
A. Bove, M. Derridj and D. S. Tartakoff
$$[{P}_{2},\phi {\Lambda}^{q/2}]=[{\Delta}_{t}+t{}^{2}A+B,\phi {\Lambda}^{q/2}],$$
where we wrote
$$B={\sum}_{i,j=1}^{s}{b}_{ij}(t,x){Y}_{i}{Y}_{j}.$$
The above quantity becomes:
$$[{\Delta}_{t}+t{}^{2}A+B,\phi {\Lambda}^{q/2}]=\frac{q}{2}\phi [{\Delta}_{t},\Lambda ]{\Lambda}^{q/21}+t{}^{2}[A,\phi ]{\Lambda}^{q/2}+\frac{q}{2}\leftt{}^{2}\phi \right[A,\Lambda ]{\Lambda}^{q/21}+\phi [B,{\Lambda}^{q/2}]={T}_{1}+{T}_{2}+{T}_{3}+{T}_{4},$$
modulo lower order terms whose treatment is easier. Let us look at each term in the above formula, denoting by `elliptic' any term which contains, in the inner product, two factors of the form maximally estimated by the operator, namely two factors each of the form
$Y,tX,or\partial /\partial t.$
Such terms will be subject to the a priori inequality (after an integration by parts) in a recursive manner and will cause little trouble. Local Analyticity for a Class of Nonsymplectic Sums of Squares
3.1
${T}_{1}$
.
Since the commutator appears in a scalar product, taking one
$t$
derivative to the other side, we have to estimate, for some coefficient
$a\left(x\right),$
$$q\langle \phi [{D}_{{t}_{s}},\Lambda ]{\Lambda}^{q/21}u,{D}_{{t}_{s}}\phi {\Lambda}^{q/2}u\rangle \sim q\langle \phi [{D}_{{t}_{s}},aX]{\Lambda}^{(q1)/2}u,{D}_{{t}_{s}}\phi {\Lambda}^{q/2}u\rangle .$$
But by ( 1.4 ), this bracket is elliptic, hence the factor of
$q$
balances the decrease in the exponent of
$\Lambda $
and will iterate analytically. A. Bove, M. Derridj and D. S. Tartakoff
3.2
${T}_{2}$
.
We have to estimate
$$\langle t{}^{2}[\Lambda ,\phi ]{\Lambda}^{q/2}u,\phi {\Lambda}^{q/2}u\rangle .$$
The commutator in the left hand side factor of the above scalar product leads to an expression of the type
$$2\langle t{}^{2}{X}_{i}{\phi}^{\prime}{\Lambda}^{q/2}u,\phi {\Lambda}^{q/2}u\rangle \sim 2\langle Z{\phi}^{\prime}{\Lambda}^{(q1)/2}u,Z\phi {\Lambda}^{q/2}u\rangle $$
with elliptic
$Z$
modulo (easier) lower order terms. Here we used just the form of
$\Lambda .$
A weighted Schwarz inequality shows that this term iterates analytically, since with the Ehrenpreistype localizing functions, a derivative on
$\phi $
balances a decrease in
$q$
. Local Analyticity for a Class of Nonsymplectic Sums of Squares
3.3
${T}_{3}$
.
We have to estimate the scalar product
$$\langle \frac{q}{2}t{}^{2}\phi [A,\Lambda ]{\Lambda}^{q/21}u,\phi {\Lambda}^{q/2}u\rangle .$$
Since
(3.2)
[
A
,
Λ
]
=
∑
i
,
j
=
1
k
∑
α
,
β
=
1
k
[
a
i
j
X
i
X
j
,
a
X
α
X
β
]
=
∑
a
~
X
2
[
X
,
X
]
=
∑
a
~
X
2
{
X
o
r
∂
∂
t
}
again modulo lower order terms, using assumption 3, part 1.
Now one of the
$X$
factors raises
$q/21$
to
$q/21/2=(q1)/2,$
with the factor of
$q$
balancing the decrease from
$q$
to
$q1,$
and the other two
$X$
's (or one
$X$
and one
$\frac{\partial}{\partial t}$
which is better, combine with
${t}^{2}$
to produce
${Z}^{2},$
with
$Z$
elliptic. Thus this inner product, as well, iterates analytically. A. Bove, M. Derridj and D. S. Tartakoff
3.4
${T}_{4}$
.
Since
$$\phi [B,{\Lambda}^{q/2}]\sim \frac{q}{2}\phi [B,\Lambda ]{\Lambda}^{q/21},$$
the estimate of
${T}_{4}$
boils down to computing the commutator
$[B,\Lambda ]$
and estimating the resulting terms. But this goes as before in view of the second part of assumption 3, since the bracket contains a product of three vector fields, two of which are elliptic and one serves to convert
${\Lambda}^{q/21}$
to
${\Lambda}^{(q1)/2}.$
This ends the proof of the Theorem in the general case. Local Analyticity for a Class of Nonsymplectic Sums of Squares
4 Remarks visavis the conjecture of Treves
The conjecture of Treves states, in this context, that the operator
$P$
should be analytic hypoelliptic at the origin if and only if all layers of the Poisson stratification are symplectic. The first of these layers is the characteristic manifold, which is patently nonsymplectic in all of these cases. And the operators are analytic hypoelliptic in the sense of germs at the origin.
The distinction is crucial. For to contradict Treves' conjecture, there would have to exist an open neighborhood
$V$
of the origin, in which one could have analytic data with a nonanalytic solution. And that may still well be the case. What we have shown is that for neighborhoods of the origin of a certain geometry relative to the operator, analytic data forces analyticity of the solution.
It is well known however that in the case of the operator
${P}_{2}$
there is in general propagation of the analytic wave front set (or rather of the analytic regularity) along the Hamilton leaves of the characteristic manifold (which are non trivial in this case), see e.g. [
1]
.
From the above proof we may see that the analyticity of the solution is forced, in some “adapted” open set, by a “global” phenomenon, that might be described by saying that the analytic singularities of the solution in the open set under consideration comes from points outside the open set lying on some Hamilton leaf of the characteristic manifold. This is actually prevented by the “large scale” geometry of the open set.
It can thus be asserted that, far from being in contradiction with Treves' conjecture, the present result is in complete agreement with it and points out the importance of the geometry of the nonsymplectic strata of the PoissonTreves stratification.
References

A. Bove, D. S. Tartakoff, Propagation of Gevrey Regularity for a Class of Hypoelliptic Equations , Trans. Amer. Math. Soc. 348 (1996), pp. 25332575.

A. Bove, D. S. Tartakoff, Optimal nonisotropic Gevrey exponents for sums of squares of vector fields , Comm. Partial Differential Equations, 22, 78, (1997), pp. 12631282.

N. Hanges, Analytic regularity for an operator with Treves curves , J. Functional Analysis, 210, (2004), pp. 295320.

D.S.Tartakoff, Local Analytic Hypoellipticity for
${\square}_{b}$
on NonDegenerate Cauchy Riemann Manifolds, Proc. Nat. Acad. Sci. U.S.A. 75 (1978), pp. 30273028.

D.S.Tartakoff, On the Local Real Analyticity of Solutions to
${\square}_{b}$
and the
$\overline{\partial}$
Neumann Problem, Acta Math. 145 (1980), pp. 117204.

F. Treves, Analytic Hypoellipticity of a Class of PseudoDifferential Operators with Double Characteristics and Application to the
$\overline{\partial}$
Neumann Problem, Comm. in P.D.E. 3 (67) (1978), pp. 475642.

F. Treves, Symplectic geometry and analytic hypoellipticity, in Differential equations: La Pietra 1996 (Florence), Proc. Sympos. Pure Math., 65, Amer. Math. Soc., Providence, RI, 1999, pp. 201219
Dipartimento di Matematica, Universita di Bologna, Piazza di Porta San Donato 5, 40127 Bologna, Italy Email address : Antonio.Bove@bo.infn.it 78350 Les Loges en Josas, FRANCE Email address : Makhlouf.Derridj@univrouen.fr Department of Mathematics, University of Illinois at Chicago, m/c 249, 851 S. Morgan St., Chicago IL 60607, USA Email address : dst@uic.edu