INSTITUTE OF MATHEMATICS POLISH ACADEMY OF SCIENCES
WARSZAWA 1992
ASYMPTOTIC EXPANSION OF THE HEAT KERNEL FOR A CLASS OF HYPOELLIPTIC OPERATORS
A L E X A N D E R L O P A T N I K O V
Space Research Institute, Russian Academy of Sciences Profsoyuznaya 84/32, Moscow, 117810 Russia
Introduction. Let M be a smooth compact manifold without boundary, let dx be a fixed positive smooth density on M , and let X
1, . . . , X
lbe smooth real vector fields on M , i.e. in a local coordinate system, X
j= P
ni=1
a
ij∂
xi. We will consider operators A of the form (m is even)
(1) (−1)
m/2l
X
j=1
X
jm+ X
|α|<m
a
α(x)X
αwhere α = (α
1, . . . , α
k), X
α= X
α1. . . X
αk, |α| = k and the a
αare smooth functions. The well-known example is the sum of the squares of vector fields,
(2) −
l
X
j=1
X
j2+ X
0+ c(x) .
A result of H¨ ormander [7, 14] states that this operator is hypoelliptic if the vector fields X
1, . . . , X
land all their commutators [X
i1, [X
i2. . . [X
is−1, X
is] . . .], s ≤ r, up to length r span the tangent space to M at each point. We recall that an operator A is said to be hypoelliptic on M if for any open set U ⊂ M and distributions u, f on U satisfying Au = f , f ∈ C
∞(U ) implies u ∈ C
∞(U ). In [17] for the operator (2) it was shown (with m = 2) that
kuk
m/r≤ C(kAuk
0+ kuk
0) ,
for all u ∈ C
∞(M ), where k · k
sdenotes the norm in the usual Sobolev space H
s(M ). For the operator (1) this estimate and hypoellipticity were proved in [6, 16].
We assume that the operator (1) is formally selfadjoint and positive, that is, (Au, v) = (u, Av) and (Au, u) ≥ 0 for all u, v ∈ C
∞(M ). It is easy to show that
[309]
under our assumption A is an unbounded selfadjoint operator on the domain D
A= {u ∈ H
m/r(M ) : Au ∈ L
2(M )} and has discrete spectrum λ
j→ ∞. Let U (x, y, t) be the kernel of the operator exp(−tA),
U (x, y, t) =
∞
X
j=1
e
−λjtϕ
j(x)ϕ
j(y) .
Here ϕ
j(x) is a complete orthonormal set of eigenfunctions of A with eigenvalues {λ
j}. U (x, y, t) is a fundamental solution for the operator L = ∂
t+ A and so it is called the heat kernel for L.
We denote by V
k(x) the subspace of T
x(M ) spanned by X
1, . . . , X
land all their commutators of length ≤ k and let ν
k(x) = dim V
k(x) (ν
0= 0). We say that H¨ ormander’s condition (of order r) holds if
ν
r(x) = dim M = n for all x ∈ M . We will also use the condition introduced by M´ etivier [13]:
ν
k(x) = ν
k= const , 1 ≤ k ≤ r, for all x ∈ M . Our main results are the following.
Theorem 1. If H¨ ormander’s condition holds then the heat kernel for the op- erator (1) has the following asymptotic expansion as t → +0 :
(3) U (x, x, t) =
∞
X
j=−q(x)
c
j(x)t
j/m+
∞
X
j=0
d
j(x)t
j/mln(t) where q(x) = P
rk=1
(ν
k(x)−ν
k−1(x))k, and c
j(x), d
j(x) are some functions on M . We remark that in general this expansion is not uniform in x ∈ M and the functions c
j, d
jare not continuous.
Theorem 2. In the M´ etivier case the asymptotics in Theorem 1 is uniform in x ∈ M , c
j(x), d
j(x) ∈ C
∞(M ), and as t → +0,
(4) tr exp(−tA) =
∞
X
j=−q
c
jt
j/m+
∞
X
j=0
d
jt
j/mln(t) where q = P
rk=1
(ν
k− ν
k−1)k.
It is also possible to find the leading coefficient c
−q(x)(x) explicitly (see below).
For elliptic operators this result is well known; in this case q = dim M and in
addition all d
j= 0. For the operator (2) our results were obtained independently
by G. Ben Arous [2, 3], who used probabilistic methods. Formula (4) was also
proved in [19] for r = 2, and related results were obtained in [1, 18]. D. Jerison and
A. S´ anchez-Calle [8, 9, 15] estimated the kernel U (x, y, t) in terms of the metric
associated with the operator A. From the asymptotics of the heat kernel it is easy
to find the first term of the asymptotics of the spectral function of A [10–12] (for
second order operators in the M´ etivier case this was done by a different method
by G. M´ etivier [13]). To prove Theorems 1, 2 we use the method developed in [3, 7, 8].
1. Dilations and homogeneity. In this section we recall some definitions and propositions connected with homogeneous structures (see [4, 5, 15, 17] for details). Let e
1, . . . , e
nbe a basis in R
nand let 0 = ν
0< ν
1< . . . < ν
rbe integers.
We write [j] = k if ν
k−1< j ≤ ν
k. We define a group of linear automorphisms δ
sof R
nby
δ
s(e
j) = s
[j]e
j, 1 ≤ j ≤ n .
We also consider a homogeneous norm k · k with respect to δ
ssuch that kuk = 0 ⇔ u = 0, kδ
s(u)k = skuk .
For example we can take kuk = ( P
nj=1
|u
j|
2/[j])
1/2. This norm satisfies the fol- lowing inequalities:
ku + vk ≤ C(kuk + kvk), C
1|u| ≤ kuk ≤ C
2|u| for |u| ≤ C , where | · | is the usual euclidean norm in R
n. The number q = P
rk=1
(ν
k− ν
k−1)k is called the homogeneous dimension of the space. It is easy to see that R
n= L
rk=1
V
k, V
kis spanned by the vectors e
jfor [j] = k, and q = P
rk=1
k dim V
k. A function f is homogeneous of degree λ if f ◦ δ
s= s
λf for all s > 0. A distribution v is homogeneous of degree λ if hv, ϕ ◦ δ
si = s
Q−λhv, ϕi. A function k(u) is said to be a kernel of type λ if it is smooth away from the origin and homogeneous of degree −Q + λ. A differential operator T is homogeneous of degree λ if T (f ◦ δ
s) = s
λ(T f ) ◦ δ
sfor all s > 0. For example, the function u
α= u
α11. . . u
αnnis homogeneous of degree [α] = P
nj=1
α
j[j], the operator u
α∂/∂u
jis homogeneous of degree [j] − [α]. Let U be a neighborhood of the origin in R
n. We define the function space
C
m∞(U ) = {f (u) ∈ C
∞(U ) : |f (u)| = O(kuk
m), u → 0} . A differential operator T = P
|α|≤k
a
α(u)∂
uαfrom C
∞(U ) to C
∞(U ) is said to have degree at most p at 0 whenever T (C
m∞(U )) ⊂ C
m−p∞(U ) for all m ∈ N. For such an operator it is possible to define an operator b T ,
T = b X
|α|≤k
X
[β]≤[α]−p
(∂
βua
α(0)u
β/β!)∂
uα,
which is homogeneous of degree p. The operator T − b T has degree at most p − 1 at 0.
Let g be a free nilpotent Lie algebra of step r with l generators, g = L
r k=1g
k, and [g
i, g
j] = g
i+jif i + j ≤ r, [g
i, g
j] = 0 if i + j > r. Using the exponen- tial mapping we can identify g and the corresponding Lie group G; the group multiplication in g will be given by the Campbell–Hausdorff formula
u · v = u + v + [u, v] + . . . , u, v ∈ g .
Following [15] we now define a function class F
λ: k ∈ F
λif (i) k ∈ C
∞(g \ 0), k(u) = 0 for kuk > 1,
(ii) |P k(u)| ≤ C
s(1 + kuk
λ−Q−s), for all left-invariant differential operators P homogeneous of degree s.
We will also use another class HF
λ. A function k is in HF
λif k ∈ F
λand (i) if λ < Q then k(u) = b k(u) + g(u), where b k(u) is a kernel of type λ, g(u) ∈ C
∞(g),
(ii) if λ ≥ Q then (i) holds for the function P k for all left-invariant differential operators P of degree s, λ − s < Q.
Lemma 1. 1) If k is a kernel of type λ, ϕ ∈ C
0∞(g) and ϕ(u) = 1 for kuk ≤ 1/2, then ϕk ∈ HF
λand P ϕk ∈ HF
λ−sfor P homogeneous of degree s.
2) If k ∈ HF
α, h is a kernel of type β, 0 < β < Q, α > 0 and ϕ ∈ C
0∞(g) with ϕ(u) = 1 for kuk ≤ 1/2 then ϕ(k ∗ h) ∈ HF
α+β.
P r o o f. We have ϕk = b k + g with b k = k and g = (1 − ϕ)k(u) ∈ C
∞(g) since k(u) is a kernel of type λ, so ϕk ∈ HF
λ. For P ϕk we observe that since ϕ = 1 in a neighborhood of the origin, P ϕk − ϕP k ∈ C
0∞; since P k is a kernel of type λ − s, we have P ϕk ∈ HF
λand 1) is proved.
For 2) we observe that if α + β < Q then by definition k(u) = b k(u) + g(u), so k ∗h(u) = R
k(v)h(v
−1u) dv = R
b k(v)h(v
−1u) dv + R
g(v)h(v
−1u) dv = I
1+I
2. I
1is a kernel of type α + β by the result of Folland [4], I
2(u) is a smooth function, and using the same arguments as in Lemma 3 of [15] one can see that g = I
2satisfies the required estimate. In the case α + β ≥ Q we have P (k ∗ h) = (P k) ∗ h, P k is a kernel of type λ−s, and λ−s+β < Q, so as was shown before ϕP (k ∗h) ∈ HF
α+β−sand by definition k ∗ h ∈ HF
α+β.
We say that a function k is in SF
λif for any s ∈ N with s > λ, k(u) =
s
X
j=0
k
j(u) + q
s(u) , where k
j∈ HF
λ+jand q
s∈ F
s(g).
2. Lifting of vector fields. Let L(M ) be the Lie algebra of smooth real vector fields on M . There exists a partial homomorphism µ : g → L(M ), that is, µ is linear and for all a ∈ g
i, b ∈ g
jwe have µ([a, b]) = [µ(a), µ(b)] if i + j ≤ r.
Write µ
x(a) = µ(a)|
x, x ∈ M . We now define
H
k(x) = {a ∈ g
k: µ
x(a) ∈ V
k−1(x)}, 1 ≤ k ≤ r , H(x) =
r
M
k=1
H
k(x) . We select S
k(x) such that g
k= H
k(x) ⊕ S
k(x), and set S(x) = L
rk=1
S
k(x). As
was shown in [5], H(x) is a subalgebra in g, dim S
k(x) = ν
k(x) − ν
k−1(x) and
dim S(x) = dim M . Obviously q(x) is the homogeneous dimension of S(x), and q(x) + β(x) = Q, where β(x) is the homogeneous dimension of H(x).
We now change the local coordinate system in a neighborhood of x ∈ M so that in the new coordinates the vector fields X
1, . . . , X
lhave degree at most one.
It is easy to see that S(x) = g/H(x); let γ be a projection from g to S(x). The essential result in this situation is
Theorem 3 (Helffer–Nourrigat [5]). For any x ∈ M there exists a diffeo- morphism Θ
x: U → ω, where U is a neighborhood of 0 in S(x) and ω is a neighborhood of x in M , so that if µ(a) = X then
1) ( d Θ
−1x)
∗X = b X, Xf (u) = b d dt
t=0f (γ(u · ta));
2) (Θ
x(0))
∗(0) = µ
x|
S(x).
In the M´ etivier case Θ
xis smooth in x ∈ M .
We introduce coordinates (u, v) in g so that u ∈ S(x), v ∈ H(x). If µ(a) = X
i(1 ≤ i ≤ l) we define a left-invariant vector field Y
ion g by Y
if (u, v) = (d/dt)|
t=0f ((u, v) · ta). Consequently,
Y
i(f · γ) = d dt
t=0f (γ((u, v)) ◦ ta) = ( b X
if ) ◦ γ .
Let R
i= X
i− b X
i, 1 ≤ i ≤ l. By Theorem 3 the vector fields R
ihave degree at most 0 at 0. If we now define e X
i= Y
i+ R
ithen we obtain
Lemma 2. e X
i(f ◦ γ) = (X
if ) ◦ γ for 1 ≤ i ≤ l.
3. Construction of the fundamental solution. We will consider two dif- ferential operators connected with L:
L = (−1) e
m/2l
X
j=1
X e
jm+ X
|α|<m
a
α(x) e X
α+ ∂
∂t , L = (−1) b
m/2l
X
j=1
Y
jm+ ∂
∂t . Lemma 3. The operator b L is hypoelliptic.
P r o o f. For m = 2 this follows directly from H¨ ormander’s theorem. In the case m > 2 it can be shown by using a criterion of hypoellipticity by Helffer–Nourrigat [6] (see also [12]).
On g
0= g × R
1we define dilations by δ
s(ξ, t) = (δ
s(ξ), s
mt), ξ ∈ g
0, t ∈ R
1.
Then Q
0= Q + m is the homogeneous dimension of g
0. For g
0we can define the
spaces F
λ, HF
λ, SF
λas in the previous section. It is clear that Lemma 1 is true
in this situation. The operator b L is homogeneous on g
0of degree m. By a result
of G. B. Folland [4] we can find a kernel k(ξ, t) of type m which is a fundamental
solution for b L, that is,
(5) Lk = δ(ξ, t) b
in the sense of distributions, where δ is the delta distribution on g
0.
We denote by U, U
1neighborhoods of the origin in S(x), and by V, V
1neigh- borhoods of the origin in H(x) which are sufficiently small and satisfy U b U
1, V b V
1. Let ϕ ∈ C
0∞(U
1), ϕ = 1 on U , ψ ∈ C
0∞(V
1), ψ = 1 on V , and
%(t) ∈ C
0∞(−2, 2), %(t) = 1 for |t| < 1. We now define k
0(ξ, t) = ϕψ%k(ξ, t) .
From the definitions of the operators b L and e L we see that L = b e L + R ,
where R has degree at most m − 1. Consequently, for any s ∈ N, the operator e L can be written in the form
L = b e L +
s
X
i=1
R
i+ Q
s,
where the R
iare homogeneous operators of degree m − i and Q
shas degree at most m − s − 1 at 0. Using (5) and Lemma 1 we obtain
(6) Lk e
0(ξ, t) = ϕψ% · δ +
s
X
i=1
ϕψ%kr
i+ q
sfor ξ ∈ U
1× V
1, t ∈ (−2, 2), where r
i∈ HF
i, q
s∈ F
s+1.
Lemma 4. Given s ∈ N there exists a function K
s(ξ, t) ∈ SF
msuch that LK e
s= ϕψ% · δ + H
s, H
s∈ SF
s.
P r o o f. We use induction on s. For s = 0 we set K
0(ξ, t) = k
0(ξ, t), and the statement of the lemma follows from (6). Assume that it is true for s − 1; then we have
LK e
s−1= ϕψ% · δ + H
s−1, H
s−1∈ SF
s−1.
We now define K
s(ξ, t) by K
s= K
s−1− a(ξ, t)k
0∗ H
s−1, where a(ξ, t) ∈ C
0∞(g
0), supp a ⊂ U
1× V
1× (−2, 2) and a ≡ 1 in supp H
s−1. We have
LK e
s= ϕψ% · δ + H
s−1− aH
s−1+ H
s,
where H
s= a(ξ, t) b Lk
0∗ H
s−1− e L(a(ξ, t)k
0∗ H
s−1). By Lemma 1 it is clear that K
s(ξ, t) ∈ SF
m, H
s∈ SF
sand the proof is finished.
By Sobolev’s embedding theorem for any p ∈ N there exists s so that SF
λ⊂ C
p(g). From this fact and the previous lemma
LK e
s= ϕψ% · δ + H
s, H
s∈ C
s(g) .
We now want to construct a fundamental solution for the original operator L. Set p
s(u, t) = R
K
s(u, v, t) dv, h
s(u, t) = R
H
s(u, v, t) dv .
Lemma 5. Lp
s= ϕ% · δ + h
s. P r o o f. Let R = L − e L. Then
Lp
s= ϕ% · δ + h
s− R
RK
sdv + R
RH
sdv .
The operator R is selfadjoint and acts only in the v variables so it is easy to see that R RK
sdv = 0 and R RH
sdv = 0, and the lemma is proved.
Using the second property of the map Θ from Theorem 3 one can show that in the original coordinate system (y) in some small neighborhood ω of the point x ∈ M ,
Lv(x)p
s= δ + h
s,
where v(x) = |det (µ
x|
S(x))|. This formula and Lemma 5 imply that (7) U (x, x, t) = v(x)p
s(0, t) + g(t) ,
where g(t) ∈ C
s(ω × (−1, 1)). By construction, p
s(0, t) =
s
X
j=1
R kj(0, v, t) dv + . . .
If k
j∈ HF
λfor λ < β(x) then by definition of this class
R kj(0, v, t) dv = R
b k
j(0, v, t) dv + R
g
j(0, v, t) dv .
Consequently, p
j(0, t) = p b
s(0, t) + g(t), where p b
s(0, t) is homogeneous of degree λ − β(x) and g ∈ C
∞(−1, 1).
If λ > β(x) then we have
∂
tap
j(0, t) = R
∂
tak
j(0, v, t) dv + R
∂
tac(0, v, t) dv = p b
j(0, t) + g(t) . The function p b
j(0, t) is homogeneous of degree (j −Q)/m−a. If (j −Q)/m−a 6∈ Z then b p
j= c
jt
(j−Q)/m−a, and after integrating over t we obtain
(8) p
j(0, t) = c
jt
(j−Q)/m+ g(t) ,
g ∈ C
∞(−1, 1). If (j − Q)/m − a ∈ Z then p b
j= c
jt
−1and so in this case (9) p
j(0, t) = c
jt
(j−Q)/mln(t) + d
jt
(j−Q)/m+ g(t) ,
g ∈ C
∞(−1, 1). From (7)–(9) it follows that for any s ∈ N
U (x, x, t) −
s
X
j=−q(x)
c
j(x)t
j/m−
s
X
j=0
d
j(x)t
j/mln(t)
< C
st
(s+1)/mand the proof of Theorem 1 is finished.
From the proof of Theorem 1 one can find the leading coefficient c
−q(x)ex- plicitly. It is clear that
c
−q(x)= v(x) · R
k(0, v, 1) dv ,
where v(x) = |det (µ
x|
S(x))| and k(u, v, t) is a fundamental solution for the oper- ator b L.
For the proof of Theorem 2 we observe that in the M´ etivier case Θ
xis smooth in x ∈ M and so the asymptotic formula of Theorem 1 is uniform in x. Conse- quently, to obtain the statement of Theorem 2 we just integrate this formula over the manifold M .
References
[1] R. B e a l s and N. S t a n t o n, The heat equation for the ∂-Neumann problem I , Comm.
Partial Differential Equations 12 (4) (1987), 407–413.
[2] G. B e n A r o u s, Noyau de la chaleur hypoelliptique et g´eom´etrie sous-riemannienne, in:
Lecture Notes in Math. 1322, Springer, 1988, 1–16.
[3] —, D´eveloppement asymptotique du noyau de la chaleur hypoelliptique sur la diagonale, Ann. Inst. Fourier (Grenoble) 39 (1) (1989), 73–99.
[4] G. B. F o l l a n d, Subelliptic estimates and function spaces on nilpotent Lie groups, Ark.
Mat. 13 (1975), 161–207.
[5] B. H e l f f e r et J. N o u r r i g a t, Approximation d’un syst`eme de champs de vecteurs et applications `a l’hypoellipticit´e, ibid. 17 (2) (1979), 237–254.
[6] —, —, Hypoellipticit´e maximale pour des op´erateurs polynˆomes de champs de vecteurs, Progr. Math. 58, Birkh¨auser, 1985.
[7] L. H ¨o r m a n d e r, Hypoelliptic second order differential equations, Acta Math. 119 (1967), 147–171.
[8] D. J e r i s o n and A. S ´a n c h e z - C a l l e, Estimates for the heat kernel for a sum of squares of vector fields, Indiana Univ. Math. J. 35 (4) (1986), 835–854.
[9] —, —, Subelliptic second order differential operators, in: Lecture Notes in Math. 1287, Springer, 1988, 46–77.
[10] A. M. L o p a t n i k o v, Asymptotic behaviour of the spectral function for operators con- structed from vector fields, Vestnik Moskov. Univ. Ser. I Mat. Mekh. 1987 (3), 76–78 (Moscow Univ. Math. Bull. 42 (3) (1987), 70–72).
[11] —, Asymptotic behaviour of the spectral function for a class of hypoelliptic operators, Vestnik Moskov. Univ. Ser. I Mat. Mekh. 1988 (1), 92–94.
[12] —, Spectral asymptotics for a class of hypoelliptic operators, manuscript, Moscow 1987, 94 pp., VINITI no. 2530 B87.
[13] G. M ´e t i v i e r, Fonction sp´ectrale et valeurs propres d’une classe d’op´erateurs non ellip- tiques, Comm. Partial Differential Equations 1 (5) (1976), 467–519.
[14] O. O l e˘ın i k and E. R a d k e v i c h, Second Order Equations with Nonnegative Characteristic Form, Amer. Math. Soc., Providence, R.I., 1973.
[15] A. S ´a n c h e z - C a l l e, Fundamental solutions and geometry of the sum of squares of vector fields, Invent. Math. 78 (1984), 143–160.
[16] L. P. R o t h s c h i l d, A criterion for hypoellipticity of operators constructed from vector fields, Comm. Partial Differential Equations 4 (1989), 645–699.
[17] L. P. R o t h s c h i l d and E. M. S t e i n, Hypoelliptic differential operators and nilpotent Lie groups, Acta Math. 137 (3–4) (1976), 247–320.
[18] M. T a y l o r, Noncommutative microlocal analysis, Mem. Amer. Math. Soc. 313 (1984).
[19] T. T a y l o r, A parametrix for step two hypoelliptic diffusion equations, Trans. Amer. Math.
Soc. 296 (1) (1986), 191–215.
[20] N. T. V a r o p o u l o s, Analysis on Lie groups, J. Funct. Anal. 76 (1988), 346–410.