On Second-Order Periodic Elliptic Operators in Divergence Form

Download as pdf or txt
Download as pdf or txt
You are on page 1of 41

On second-order periodic elliptic operators in divergence form

A.F.M. ter Elst1, Derek W. Robinson2 and Adam Sikora2


Abstract
We consider second-order, strongly elliptic, operators with complex coe cients in divergence form on Rd. We assume that the coe cients are all periodic with a common period. If the coe cients are continuous we derive Gaussian bounds with the correct small and large time asymptotic behaviour on the heat kernel and all its Holder derivatives. Secondly if the coe cients are Holder continuous we prove that the rst-order derivatives of the kernel satisfy good Gaussian bounds and the rst-order Riesz transforms are bounded on the Lp-spaces with p 2 h1 1i. Then we establish that the second-order derivatives exist and satisfy good bounds if, and only if, the coe cients are divergencefree or if, and only if, the second-order Riesz transforms are bounded. Finally if the third-order derivatives exist with good bounds then the coe cients must be constant.

March 1999 AMS Subject Classi cation: 35K05, 35B27, 35B40, 47A58.

Home institutions:

1. Department of Mathematics and Computing Science Eindhoven University of Technology P.O. Box 513 5600 MB Eindhoven The Netherlands

2. Centre for Mathematics and its Applications School of Mathematical Sciences Australian National University Canberra, ACT 0200 Australia

1 Introduction
We analyze the asymptotic behaviour of complex second-order strongly elliptic operators with periodic coe cients in divergence form acting on L2(Rd). We establish that if the coe cients of the operators are continuous then the corresponding semigroups have bounded kernels which, together with their Holder derivatives, satisfy good large time Gaussian bounds. Subsequently we establish a detailed hierarchy of asymptotic properties for periodic operators with Holder continuous coe cients. Auscher, McIntosh and Tchamitchian AMT] proved the existence of Gaussian-type bounds for the kernels of complex operators with Holder continuous coe cients but these bounds contain an extra factor with polynomial growth in time. Subsequently Auscher Aus] extended these results to operators with uniformly continuous coe cients. We demonstrate that periodicity of the coe cients is su cient to remove the growth factor and to give Gaussian bounds for the kernel and its Holder derivatives for both small and large times. Moreover, the kernels of periodic operators with Holder continuous coe cients have a remarkable 1 2 1-structure. The kernels are automatically once-di erentiable with the derivatives satisfying good Gaussian bounds, they are twice-di erentiable with the second derivatives satisfying good asymptotic bounds if, and only if, the coe cients are divergence-free and existence of the third derivatives with the canonical behaviour implies the coe cients must be constant and hence the kernel is in nitely often di erentiable. This hierarchy can also be characterized by the boundedness of the Riesz transforms the rst-order transforms are always bounded, boundedness of the second-order transforms is equivalent to the coe cients being divergence-free and boundedness of the third, or higher, order transforms implies the coe cients are constant. The theory of complex second-order elliptic operators in divergence form di ers substantially from the theory of their real counterparts. The complex theory shares the di culties encountered with systems of real operators in divergence form or with single real operators in non-divergence form. In 1968 De Giorgi Gio] gave an example of a system of real operators in divergence form which did not have the advantageous regularity features of the Nash{De Giorgi theory of single real operators with measurable coe cients. Several other authors subsequently gave re ned examples of this nature (see, for example, Gia], Section II.3). Then in 1981 Fabes and Kenig FaK] constructed a class of real second-order elliptic operators in non-divergence form on Rd such that the associated `parabolic measures' were singular with respect to Lebesgue measure. The corresponding semigroups are contractive on L1(Rd) but do not have bounded kernels. These examples are surprising since the operators have uniformly continuous coe cients and it is known that Holder continuity of the coe cients is su cient to ensure that the kernel exists and is at least twice di erentiable (see Fri], Chapter 9). The construction of the Fabes{Kenig examples was based on ideas of Modica, Mortola MoM] which in turn used a suggestion of De Giorgi (see also MMS] Bau]). The theory of complex divergence form operators presents di erent structural difculties. There are no problems in one or two dimensions the principal regularity results of the Nash{De Giorgi theory are still valid AMT]. But in 1985 Maz'ya, Nazarov and Plamenevsk i MNP] gave counterexamples for the complex theory in higher dimensions. Their ideas have been extended subsequently by various authors ACT] Dav] and result in examples of complex divergence form operators which generate continuous semigroups 1

on Lp(Rd) if, and only if, j1=p ; 1=2j is small. In particular the semigroups do not have bounded kernels. These examples require d 5 and the status of the complex theory for d = 3 and d = 4 is not clear. Nevertheless the reductio ad absurdum argument given in AMT], Remark 4.24, indicates that smoothness of the coe cients alone is not su cient to obtain a satisfactory description in terms of bounded kernels. This conclusion is con rmed by our analysis of periodic operators. Our approach to Gaussian bounds is through the homogenization theory of periodic systems BLP] ZKO]. The pointwise Gaussian-type bounds of Auscher, McIntosh and Tchamitchian AMT], or Auscher Aus], act as important a priori estimates. We use a Bloch{Zak decomposition and, following Zhikov Zhi], we exploit L2-spectral estimates to prove that the kernel of the homogenized system provides a good asymptotic approximation to the semigroup kernel. Careful handling of the approximation, and interpolation with Auscher's estimates, yields Gaussian bounds, for operators with continuous coe cients, both on the kernel and on its Holder derivatives. A similar tactic works to provide detailed regularity properties for operators with Holder continuous coe cients. In the sequel H denotes the maximal accretive operator on the complex Hilbert space L2(Rd) associated with the sectorial form (see, for example, Kat] ReS])

h(') =

d X k l=1

(@k ' ckl@l')

(1)

where the @T = @=@xk denote the usual partial derivatives and the domain of h is D(h) = k d ) = d D(@k ). The matrix C = (ckl ) of complex-valued coe cients ckl 2 L1 (Rd ), L2 1(R k=1 which is not necessarily symmetric, is assumed to satisfy the ellipticity condition <C = 2;1 (C + C ) I > 0 (2) in the sense of d d-matrices over Cd , uniformly over Rd, where C = (ckl) with ckl = clk . The ellipticity condition ensures that the form h is sectorial with vertex at the origin. The least upper bound C of the satisfying this condition is called the ellipticity constant. The operator H generates a continuous holomorphic semigroup S on L2(Rd). The action of S is determined by a distribution kernel K . Our derivation of boundedness and smoothness properties of K is based on continuity and periodicity of the coe cients ckl.

Theorem 1.1 Assume the coe cients ckl of H are continuous and periodic with a common
period. Then the kernel K of the semigroup generated by H has the following bounds. I. There exist b c > 0 such that
jKt (x y )j

c Gb t(x ; y) for all x y 2 Rd and t > 0 where Gb t (x) = t;d=2e;bjxj2t;1 . II. For all 2 h0 1i and > 0 there exist b c > 0 such that jhj + jk j jKt (x ; h y ; k ) ; Kt (x y )j c G (x ; y ) jx ; y j + t1=2 b t for all x y h k 2 Rd and all t > 0 with jhj + jk j jx ; y j + t1=2.
2

Although we have stated these bounds for real positive t they can be extended to all complex t in an open subsector of the sector of holomorphy. A similar remark is true for all subsequent bounds of a similar nature. The passage from real t to complex z = tei corresponds to replacing H by ei H and for small the latter operator is of the same type as H . The situation with Holder continuous coe cients is more detailed and introduces a new feature, boundedness of the Riesz transforms. For p 2 1 1] de ne Lp 1(Rd) = Td=1 D(@k ), k where @k is the partial derivative in Lp(Rd). Theorem 1.2 Assume the coe cients ckl of H are Holder continuous and periodic with a common period. Then K is once-di erentiable and there exist b c > 0 such that j(@xk Kt )(x y )j + j(@yk Kt )(x y )j c t;1=2Gb t (x ; y ) for all k 2 f1 : : : dg, x y 2 Rd and all t > 0. Moreover, if p 2 h1 1i then D(H 1=2 ) = Lp 1(Rd ) and there exist cp c0p > 0 such that cp max k@k 'kp kH 1=2'kp c0p max k@k 'kp
for all ' 2 Lp 1(Rd ).
1 k d 1 k d

The uniform regularity properties given by this theorem are optimal as the next result shows. Increased regularity of the kernel for all t is characterized by a divergence-free property of the coe cients. In the fourth condition of the theorem U denotes the action of translations on the spaces Lp(Rd). Theorem 1.3 Assume the coe cients ckl of H are Holder continuous and periodic with a common period. The following conditions are equivalent. I. Pd=1 @k ckl = 0, in the sense of distributions, for each l 2 f1 : : : dg. k II. There exist c > 0 such that j(@xk Kt )(x ; h y ) ; (@xk Kt )(x y )j c t;1=2 (jhjt;1=2) t;d=2 uniformly for all k 2 f1 : : : dg, x y h 2 Rd with jhj 1 and all t 1. III. K is twice-di erentiable in the rst variable and there exist b c > 0 such that j(@xk @xl Kt )(x y )j c t;1 Gb t (x ; y ) for all k l 2 f1 : : : dg, x y 2 Rd and t > 0. IV. There are p 2 1 1] and cp > 0 such that k(I ; U (h))@k St kp!p cp t;1=2 (jhjt;1=2) (3) uniformly for all k 2 f1 : : : dg, all jhj 1 and all t 1. V. If p 2 h1 1i then D(H ) = Lp 2(Rd) and there exist cp c0p > 0 such that cp 1maxd k@k @l'kp kH'kp c0p 1maxd k@k @l'kp kl kl
for all ' 2 Lp 2(Rd ).

Further regularity only occurs for all t in the simple case that the coe cients are constant. Then of course the kernel is Gaussian and the Riesz transforms of all orders are bounded.

Theorem 1.4 Assume the coe cients ckl of H are Holder continuous and periodic with a
common period. The following conditions are equivalent.

I. The coe cients ckl are constant. II. K is twice-di erentiable in the rst variable and there exist c > 0 such that
j(@xk @xl Kt )(x ; h y ) ; (@xk @xl Kt )(x y )j c t;1 (jhjt;1=2) t;d=2

uniformly for all k l 2 f1 : : : dg, x y h 2 Rd with jhj 1 and all t 1. III. There are p 2 1 1] and cp > 0 such that k(I ; U (h))@k @l St kp!p cp t;1 (jhjt;1=2) uniformly for all k l 2 f1 : : : dg, all jhj 1 and all t 1.

Although the regularity properties in these theorems are stated in terms of the rst variable of the kernel one can readily draw similar conclusions for the second variable by symmetry. Interchange of the variables corresponds, by duality, to replacement of C by its transpose C t, i.e., the matrix with coe cients ctkl = clk . For example, under the conditions of Theorem 1.2 the derivative with respect to the secondP variable is Holder continuous Pd t = d @k clk = 0 as distributions with the canonical behaviour if, and only if, k=1 @k ckl k=1 for each l 2 f1 : : : dg. In fact one can deduce joint regularity properties from separate regularity properties by using the semigroup property

Ks+t (x y) = Rd dz Ks (x z)Kt(z y) : For example, it follows in Theorem 1.2 that K is once-di erentiable jointly with respect to both variables. In Section 2 we summarize some basic elements of decomposition theory and homogenization theory for operators on L2(Rd) with periodic coe cients and derive the essential spectral estimates. In Section 3 we prove Theorem 1.1. This follows by establishing that the kernel of the homogenized operator gives the best asymptotic approximation to K . We also deduce some results on rst-order corrections to the asymptotic series for the kernel. In Section 4 we prove Theorems 1.2, 1.3 and 1.4. The conditions of Theorem 1.3 correspond to the rst-order corrector, in the sense of homogenization theory, vanishing. The conditions of Theorem 1.4 correspond to the rst and second order correctors vanishing.

2 Periodic systems
The proof of Theorem 1.1 relies on the a priori estimates of Auscher in combination with various Hilbert space properties of periodic systems which we develop in this section. Throughout the section we suppose that H is the maximal accretive operator associated 4

with the sectorial form (1) with the complex measurable coe cients satisfying the ellipticity assumption (2). We further assume that the coe cients are all periodic with a common period. Then, for simplicity, we choose units such that

ckl(x + n) = ckl(x)

(4)

for all k l 2 f1 : : : dg, x 2 Rd and n 2 Zd. No continuity of the coe cients is required in this section.

2.1 Decomposition theory

Decomposition theory for operators with periodic coe cients has a long history dating back to Bloch Blo]. We follow the description in BJR] which is based on the Zak formalism commonly used in wavelet theory (see, for example Dau]) although we then pass to the unitarily equivalent Bloch prescription. If U denotes the unitary action of Rd by left translations on L2(Rd), i.e., (U (y)')(x) = '(x ; y) for all ' 2 L2(Rd) and y 2 Rd, then U (x)D(h) = D(h) for all x 2 Rd and the periodicity (4) of the coe cients gives the invariance property h(U (n)') = h(') for all ' 2 D(h) and all n 2 Zd . Hence U (n)D(H ) = D(H ) and U (n)H = HU (n). Moreover,

U (n)St = StU (n)


on L2(Rd) for all n 2 Zd and all t > 0. Next we introduce versions of H and S on L2(Id) where I = 0 1]. Let @k , for k 2 f1 : : : dg, denote the partial derivatives on L2(Id) corresponding to periodic boundary conditions, '(u1 : : : 0 : : : ud) = '(u1 : : : 1 : : : ud) for almost every u 2 Id, where the 0 and 1 are in the k-th position. Secondly, for all 2 ; ]d de ne H as the maximal accretive operator on L2(Id) associated with the sectorial form

h (') =

d X

where D(h ) = Td=1 D(@k ) = L2 1(Id). k f The H are unitarily equivalent to versions H of H with -periodic boundary conditions, i.e., '(u1 : : : 0 : : : ud) = ei k '(u1 : : : 1 : : : ud) for all k 2 f1 : : : dg and almost every u 2 Id , where the 0 and 1 are in the k-th position. The unitarily equivalent description was used in BJR], Section 2, and is related to the present description as follows. Let V ( ) be the unitary operator on L2(Id ) de ned by (V ( )')(u) = eiu '(u) for u 2 Id. If ' 2 L2 1(Id) then V ( )' satis es the -periodic boundary conditions. Moreover,

k l=1

((@k + i k )' ckl(@l + i l)')

(5)

h (') =

d X

k l=1

(@k V ( )' ckl@lV ( )') :

f f Hence H = V ( ) H V ( ) where the H are the operators used in BJR].

One has a decomposition of L2(Rd),

; as a direct integral of copies of L2(Id) indexed by 2 ; ]d (for a description of the formalism of integral decompositions see, for example, Dix], Chapter II,Ror BrR], Chapter IV). The isomorphism is given as follows. De ne : L2(Rd) ! (2 );d ; ]d d L2(Id ) by (') ( ) = ' , where X i(n;u) ' (u) = e '(u ; n) d n2Z for all ' 2 L2(Rd), almost every 2 ; ]d and almost every u 2 Id. By Fubini's and Parseval's theorems one veri es that is a unitary operator. The inverse of is given by Z ;1 (' ) 2 ; ]d (u ; n) = (2 );d d e;i(n;u) ' (u) ; ]d R for all (' ) 2 ; ]d 2 (2 );d ; ]d d L2(Id) , n 2 Zd and almost every u 2 Id. It follows from BJR] that the sectorial form h, the operator H and the semigroup S decompose in the same manner. Explicitly, if (') = (' ) 2 ; ]d then
]d

L2

(Rd) = (2

);d

d L2(Id)

h(') = (2 );d
and

;
Z

]d

d h (' )

d H ; ]d in the sense of direct integral decompositions of closed sectorial forms. Similarly d S (6) ; ]d t where S denotes the semigroup generated by H on L2(Id) (see BJR], Theorem 2.3). If S has a bounded integrable kernel then the semigroup decomposition gives a similar kernel decomposition. We will use this fact extensively in Section 3. Moreover, we need the following estimate on crossnorms. Note that the rst crossnorm is on Rd, whilst the second one is on Id. Moreover, both can be in nite. Lemma 2.1 Let T be a bounded operator on L2(RdR) with a decomposition ;1 T = (2 );d R ; ]d d T on the direct integral space (2 );d ; ]d d L2(Id) . Then
kT k1!1 (2 );d ;
Z
]d

H ;1 = (2 );d

St ;1 = (2 );d

d kT k1!1 :
= ( ( ))( ), from which the lemma

Proof R If ' 2 Cc(Rd) then k' k1


follows.

j(2 );d ; ]d d ( T ' )j j(2 ) R ; ]d d kT k1!1, where ' = ( ('))( ) and

k'k1 for all ;d R

2 ; ]d. Therefore j( T')j = ; ]d d k k1 kT k1!1 k' k1 k k1 k'k1 (2 );d

In this subsection we derive some basic spectral properties of the operators H associated with the sectorial forms (5). It then follows that for large t the semigroup is determined by the small components S in the periodic decomposition. Moreover, for small j j the large t behaviour of the S is determined by the lowest eigenvalue and its eigenprojector. It follows immediately from the de nition (5) that the family h of sectorial forms extends to a (multi-variable) holomorphic family of type (a), with 2 Cd (see Kat], VIIx4.2). We denote the extension by h and the associated m-sectorial operator by H . Then the family H , 2 Cd , is a (multi-variable) holomorphic family of operators of type (B ) by Kat], Theorem VII-4.2. (Although we have a d-parameter extension of the formalism of Kato this does not a ect any of the results we adopt from the standard theory.) Each H has a compact resolvent. This follows because H0 has compact resolvent by BJR], Lemma 2.1, and the family H are a holomorphic family of type (B ). Then every H has compact resolvent for all 2 Cd (see Kat], Theorem 4.3). For all 2 0 ] set

2.2 Spectral estimates

B =f 2 ;
Then B is closed in ; ]d.

]d : k2f1 ::: dg j k j max

g :

Lemma 2.2 If 2 0 1] then


Re h (')
for all ' 2 L2 1(Id ) and 2 B .
C (1 ;

)kr'k2 + 2

Cj

j2 k'k2 2

Proof PLet (cn)n2Z 2 Cc (Zd) and consider the trigonometric polynomial ' given by
'(u) =
n2Zd cn e
2 in u .

Then

Re h (') = Re

Z
I

du d

d X k l=1 d X

(@k + i k )' (u) ckl(u) (@l + i l)' (u) (@k + i k )' (u) =
2

C d du I k=1 d XX k=1 n2Zd

d X X k=1 n2Zd

jcn j2 (2 nk + k )2 :

But if m 2 Z and 2 R with j j Re h (')


C

then (2 m + )2 4 2m2(1 ; ) + 2. Hence


C (1 ; ) d X k=1

jcn j2 4 2 n2 (1 ; ) + j k j2 = k

k@k 'k2 + 2

Cj

j2 k'k2 : 2

Since the trigonometric polynomials are dense in L2 1(Id) the lemma follows by a limiting argument. 2 The lemma has four immediate corollaries. Corollary 2.3 If 2 ; ]d then kSt k2!2 e; C j j2t for all t > 0.

Corollary 2.4 If 2 ;

]d and is an eigenvalue of H then Re 7

C j j2.

These statements follow immediately since Re h (') and 2 ; ]d.

C j j2 k'k2 2

for all ' 2 L2 1(Id)

Corollary 2.5 There exists a c > 0 such that


Re h (') c Re h0(')
for all ' 2 L2 1(Id) and 2 ;
jh0(') ; h (')j

]d.

Proof Let kC k = supu2Id k ckl(u) k2!2. Then it follows from the de nition of h that
2kC k kr'k2 j j k'k2 + kC k j j2 k'k2 2 " C kr'k2 + ";1 ;1kC k2 j j2 k'k2 + kC k j j2 k'k2 2 C 2 2 ;1 (";1 C1 kC k2 + kC k) Re h (') ; " Re h0(') + C

for all " > 0 and ' 2 L2 1(Id), where, in the last step, we have used Lemma 2.2. The corollary follows by taking " = 1=2. 2 Clearly the operator H0 and its adjoint H0 have eigenvalue 0 with eigenfunction 1 , the identity function. But it follows from the lemma that this eigenvalue is simple, i.e., it has multiplicity one.

Corollary 2.6 If 6= 0 is an eigenvalue of H0 then Re > 0. Moreover, 0 is a simple


eigenvalue of H0.

Proof If is a non-constant eigenfunction of the operator H0 with eigenvalue then


Re k k2 = Re h0( ) 2 For all > 0 set
C kr

k2 > 0. This implies both statements. 2

Then B C is open in Cd .

B C = f 2 Cd : k2f1 ::: dg j k j < g : max


0

> 0 and a holomorphic function 0 : B C ! C such 0 that 0(0) = 0 and 0 ( ) is a simple eigenvalue of H for all 2 B C . Moreover, 0 ( ) is 0 the unique eigenvalue of H with j j "0 for all 2 B C . 0

Proposition 2.7 There exist "0

Proof The rst part follows by Corollary 2.6 and the proof of Theorem VII-1.7 in Kat].

By Corollary 2.6 there exists an "0 > 0 such that 0 is the unique eigenvalue of H0 with j j "0. Moreover, the dimension of the eigenspace equals one. But these properties are stable under perturbation. It then follows from the proof of Kat], Theorem VII-1.7, that for all su ciently small j j the eigenvalue 0( ) is the unique eigenvalue of H with j j < "0 . 2 Let "0 0 and 0 be as in Proposition 2.7. For 2 B C de ne the projection 0

P0( ) = (2 i);1

;"0

d ( I ; H );1

where ;"0 = fz 2 C : jzj = "0g. Then the map 7! P0( ) is holomorphic near = 0. This again follows by the proof of Kat], Theorem VII-1.7. Therefore we may assume that , the map 7! P0( ) is holomorphic on B C and that 0 0 kP0 ( ) ; P0 (0)k2!2 1=2 for all 2 B C . Note that P0 (0) is the orthogonal projection onto the constant functions. 0 For all 2 B C de ne ' = P0( )1 . Then k' ; 1 k2 = k(P0( ) ; P0(0))1 k2 1=2, 1 1 0 so ' 6= 0. Since P0 ( ) projects onto the one-dimensional eigenspace corresponding to the eigenvalue 0( ) it follows that ' is an eigenfunction of H with eigenvalue 0( ). Obviously the map 7! ' is holomorphic from B C into L2(Id). But ' 2 D(H ) 0 D(h ) = L2 1(Id) for each 2 B C . We next show that the map 7! ' is also holomorphic 0 from B C into L2 1(Id). 0

Lemma 2.8 For all k 2 f1 : : : dg the map 7! @k ' is holomorphic from B C into L2(Id). Proof Obviously the map 7! ( @k' ) = ;(@k ' ) is holomorphic on B C for all 2 L2 1(Id ) and since L2 1 (Id ) is dense in L2 (Id ) it su ces to show that the map 7! @k ' from B C into L2(Id) is locally bounded (see Kat], Remark III-1.38). But
0 0 0

eRe 0( )k@k S1 k2!2k' k2 so it reduces the problem to showing that 7! k@k S1 k2!2 is locally bounded. Since Re(' H ') 2;1 C kr'k2 ; kC k j j2(1 + 2 ;1 kC k)k'k2 C 2 2 for all ' 2 D(H ) and 2 Cd it follows by a straightforward argument that the map 7! k@k S1 k2!2 is locally bounded on Cd (cf. the proof of Rob], Lemma III.4.4). 2
k@k ' k2 = ke
0

( )@ S ' k k 1 2

Proposition 2.9 There exists a > 0 such that


for all 2 B2;1 0

Re(' H ') and ' 2 (I ; P0( ))(D(H )).


1 0

k'k2 2

and

Proof Let 2 B2; and ' 2 (I ; P0( ))(D(H )). Then ' 2 D(H ) D(h ) = L2 1(Id)

Re(' H ') c Re h0(') = c Re h0((I ; P0(0))') where c is the constant of Corollary 2.5 and we have used H0 P0(0) = H0 P0(0) = 0. But the self-adjoint operator Re H0 associated with the quadratic form Re h0 has a discrete spectrum with simple lowest eigenvalue 0 and eigenfunction 1 (see Corollary 2.6). So if 1 = min( (Re H0 )nf0g) then Re(' H ') c 1k(I ; P0(0))'k2 = c 1 k'k2 ; (' P0(0)') : 2 2 Since kP0( ) ; P0(0)k2!2 2;1 for all 2 B C one has 0 Re(' H ') c 1 2;1 k'k2 ; Re(' P0( )') = 2;1 c 1 k'k2 2 2 for all 2 B2;1 0 and ' 2 (I ; P0( ))(D(H )). 9

Proposition 2.10 There exists a > 0 such that


kSt (I ; P0 ( ))'k2

e; tk(I ; P0 ( ))'k2
1

uniformly for all t > 0, ' 2 L2 (Id ) and 2 B2;1 0 .

Since the latter subspace is invariant under the adjoint of S it follows that H is invariant under S . Moreover, the restriction T of S to H is a continuous, holomorphic, semigroup with generator H jH\D(H ). But then it follows from Proposition 2.9 that Re(' H ') k'k2 for all ' 2 H \ D(H ). So kTtkH!H e; t for all t > 0, as required. 2 2

Proof Let > 0 be as in Proposition 2.9 and 2 B2; . The subspace H = (I ; P0( ))(L2(Id)) is the orthogonal complement of the one-dimensional subspace P0( ) L2(Id).
0

2.3 Homogenization

The asymptotic behaviour of periodic operators and the associated semigroups is often determined by the related, homogenized system (see BLP] or ZKO]). Next we de ne the homogenized operator corresponding to H and its decomposition. To this end we need a boundedness result for the Riesz transforms associated with H0. First the self-adjoint operator Re H0 associated with the quadratic form Re h0 has a discrete spectrum with simple lowest eigenvalue 0 and eigenfunction 1 . Hence by Kat], Theorem III-6.17, the operator Re H0 leaves the orthogonal complement L? (Id) = (I ; 2 P0(0))(L2(Id)) of the constant functions in L2(Id) invariant and the restriction of Re H0 to the space L? (Id) is invertible. With a slight abuse of notation we denote this inverse by 2 (Re H0);1 . In this section we shall use similar notation. Then the operator (Re H0);1=2 = ((Re H0);1 )1=2 is also bounded on L?(Id). But since 2 Re h0(') it follows that @k (Re H0);1=2 is bounded (has bounded closure) and (Re H0 );1=2@l extends to a bounded operator, on L? (Id). Hence @k (Re H0);1@l extends to a bounded operator, 2 on L? (Id), for all k l 2 f1 : : : dg. This last result has an analogue with Re H0 replaced 2 by H0. Note that this does not require any continuity of the coe cients. Lemma 2.11 The inverse H0;1 is bounded on L?(Id ) and for all k l 2 f1 : : : dg the forms 2
2 C kr'k2

xl(') = (' H0;1@l') xkl(') = (@k ' H0;1@l')

with D(xl ) = D(xkl ) = L2 1(Id ) have bounded extensions to L2 (Id). If Xl and Xkl denote the bounded operators associated with the bounded forms xl and xkl on L2(Id) then Xl and Xkl map L2(Id) into L? (Id). Moreover, XlL2(Id) L2 1(Id) and 2 Xkl = ;@k Xl. Proof It follows from Corollary 2.6 that H0;1 is bounded on L?(Id ). But if ' 2 L2 1(Id) 2 then @k ' 2 L? (Id) for each k 2 f1 : : : dg. Hence the xl and xkl are well-de ned. Then, 2

10

however, it follows from Kat], Theorem VI-3.2, that there is a bounded symmetric operator B on L? (Id) such that 2 H0;1 = (Re H0);1=2(I + iB );1(Re H0);1=2 : (7) Therefore and

xl(') = (' (Re H0 );1=2(I + iB );1(Re H0);1=2@l')

xkl(') = ((Re H0);1=2@k ' (I + iB );1(Re H0);1=2@l') : Since k(I + iB );1k2!2 1 it follows that jxl (')j k(Re H0 );1=2k2!2 k@l (Re H0 );1=2k2!2 k'k2 2 and jxkl (')j k@k (Re H0 );1=2 k2!2k@l (Re H0 );1=2k2!2 k'k2 2 ?(Id). Hence the xl and xkl extend to bounded where the operator norms are taken on L2 forms on L2(Id). The associated bounded operators map L2 1(Id ) into L? (Id) and since 2 L?(Id) is closed in L2(Id) it implies that the operators map L2(Id) into L?(Id ). 2 2 2
In the sequel we also use the notation Xl = Xl (C ) and Xkl = Xkl (C ) for the operators associated with the forms xl and xkl since it is sometimes necessary to track the dependence on the matrix C of coe cients. Hence Xl (C ) and Xkl (C ) will denote the operators corresponding to the adjoint operator. c The homogenization H of H is de ned to be the elliptic operator on L2(Rd) with b constant coe cients cij associated with the form
d ^ (') = X (@k ' ckl@l') b h k l=1

where the coe cients are de ned by


b ckl = Z
I

du ckl(u) ; d

d X m n=1

(ckm Xmn cnl) :

c This de nition of H coincides with that of BLP], pages 16 and 184, or that of BBJR]. b b Note that the homogenized matrix C = (cij ) satis es the ellipticity condition (2) with c = C (see, for example, BLP] Chapter 1, Theorem 3.2). Let K be the kernel of the b generated by the operator H . c semigroup S c Since the coe cients of H are constant it can be decomposed in a similar manner to H . c Then the components H , 2 ; ]d, in the decomposition are the maximal accretive operators on L2(Id) associated with the sectorial forms
d ^ (') = X ((@k + i k )' ckl (@l + i l)') b h k l=1

with D(^ ) = L2 1(Id). h Subsequently, we need the following lemma. 11

Lemma 2.12 Let ' 2 L2 1(Id) and

:::

2 L2 (Id ). Suppose that

h0( ') =

d X k=1

(@k

k)

for all 2 L2 1 (Id). Then there is an a 2 C such that ' = a 1 ; Pd=1 Xk k . k ? Proof Since h0( 1 ) = 0 we may assume that ' 2 L? (Id) \ L2 1(Id ). If 2 L2 (Id) then 2 = H0 ;1 2 L2 1(Id ) by (7). So

( ') = h0( ') =

d X k=1

(@k

k) = ;

d X k=1

(Xk

k) = ;

d X k=1

( Xk k )

for all 2 L2 1(Id) from which the statement of the lemma follows.

3 Uniformly continuous coe cients


In this section we prove Theorem 1.1. Our reasoning is a variation of the argument of Zhikov Zhi] who considered the asymptotic behaviour of real symmetric operators with periodic measurable coe cients. (Zhikov's argument is described in Section 2.6 of ZKO].) In the real symmetric case one has good Gaussian bounds on the kernel and Zhikov used spectral estimates to obtain an asymptotic approximation of the kernel. Throughout this section we assume that the coe cients are continuous. Then the coe cients are uniformly continuous and one has the Gaussian-type bounds of Aus], Theorem 4.3,
jKt (x y )j c (1 + t)N Gb t (x ; y )

(8)

for some N 2 N and b c > 0 uniformly for all x y 2 Rd and t > 0. We will combine these with a variation of Zhikov's spectral estimates to obtain improved bounds. In this subsection we prove the rst statement of Theorem 1.1. First note that by the decomposition theory of Section 2 the semigroups S are given by bounded integral kernels K which are related to K by X i(n;u+v) X i(n;u+v) Kt(u v + n) = e Kt(u ; n v) (9) Kt (u v) = e d d n2 Z n2Z for all u v 2 Id and 2 ; ]d. (This formula di ers from the similar one given in BJR], Theorem 2.4, by the presence of the factor e;i(u;v) which arises because of the current unitarily equivalent description.) The sums in (9) are convergent for all 2 Cd and u v 2 Id, so one can use (9) to de ne a function Kt for all 2 Cd . Then we know that Kt is the kernel of St if 2 ; ]d and we next argue that this extends to all 2 Cd . It immediately follows from (8), (9) and the Jacobi identity, ScD] Satz I.10.4, that the K satisfy bounds similar to those of K : there exist c ! > 0 such that jKt (u v )j c (1 + t)N (1 ^ t);d=2 e;bku;vk2 t;1 e!(Im )2 t 12

3.1 Gaussian bounds

for all 2 Cd, u v 2 Id and t > 0, where for all x 2 Rd. For all 2 Cd and t > 0 let Tt denote the operator on L2(Id) which has kernel Kt . If ' 2 Cc(Id) and t > 0 then it follows again from the Gaussian estimates (8) that 7! ( Tt ') is a holomorphic function from Cd into C. But it equals the entire function 7! ( St ') for all 2 ; ]d. Hence ( Tt ') = ( St ') for all 2 C. This implies that Tt = St and therefore K is the kernel of S for all 2 C. It now follows from the well known estimates for the norm of the operator of convolution with a Gaussian that there exist c ! > 0 such that kSt kp!q c (1 + t)N (1 ^ t);d(1=p;1=q)=2e!(Im )2 t (10) uniformly for all 1 p q 1, t > 0 and 2 Cd, where N is as in (8). Recall that 7! P0( ) is a holomorphic function from B C into L(L2 (Id)) since the H 0 form a holomorphic family of operators on L2(Id). But the semigroup estimates give the following stronger conclusion.
kxk = minfjx ; nj : n 2 Zd g

Lemma 3.1 The map 7! P0 ( ) is holomorphic from B C into L(L1(Id) L1(Id)). Proof First note that P0( ) = e2 ( )S1 P0( )S1 . Hence
0 0

e2 Re 0( )kS1 k1!2kP0( )k2!2kS1 k2!1 : Therefore 7! kP0( )k1!1 is locally uniformly bounded. Secondly, let ' 2 L1(Id) \ L2(Id) and set '( ) = P0( )'. Then 7! '( ) is a holomorphic function from B C into L2(Id) and with derivative @'( )=@ j 2 L2(Id) for 0 all j . Moreover, for all 2 B C one has 0
kP0 ( )k1!1

( '( + ) ; '( ) ;
d Y

d X j =1

@ '( )) = (2 i);d Z d : : : d ( '( )) 1 d j @j ;d

d d d Y k ;1 ; Y ( j ; j );1 ; X ( j ; j ; j) ( j ; j );1 j =1 j =1 k=1 ( k ; k ) j =1

for all 2 Cd with j j small enough, where ; is a positively-oriented circle around the origin with radius 2;1( 0 + maxk2f1 ::: dg j j j). One readily deduces that there is an M > 0 such that d X @ '( ))j M j j2k k1 supd k'( )k1 j( '( + ) ; '( ) ; j 2; j =1 @ j M j j2 supd kP0( )k1!1 k k1k'k1 2; uniformly for all ' 2 L1(Id) \ L2(Id) and all small j j. Then @'( )=@ j 2 L1(Id) for all j and d X @ 2 k'( + ) ; '( ) ; j @ j '( )k1 M j j supd kP0( )k1!1 k'k1 2; j =1 13

uniformly for all ' 2 L1(Id) \ L2(Id) and all small j j. So 7! '( ) is holomorphic from B C into L1 (Id) and the desired conclusion follows immediately since L1(Id) \ L2(Id) is 0 dense in L1(Id). 2 After these preliminaries we come to the rst important step in the derivation of the Gaussian bounds. We prove that for large t the essential contribution to the kernel comes from the small components in the periodic decomposition and these contributions are dominated by the lowest eigenstate. For 2 h0 0i de ne Z ;1 (2 );d d P0( ) : P0(B ) = Then P0(B ) is a projection on L2(Rd).
B

Lemma 3.2 There exist > 0 and for all 2 h0 0i a c > 0 such that
kSt(I ; P0 (B ))k1!1

c e; t

uniformly for all t 1.

Proof Let > 0 be as in Proposition 2.10 and 2 h0 0]. Then it follows from (6) and
d kSt (I ; P0( ))k1!1 ; ]d nB B for all t > 0. We estimate the two terms separately. It follows from Corollary 2.3, the estimates (10) and the semigroup property that there exists a c > 0 such that kSt k1!1 c e;2;1 C 2t uniformly for all t 1 and 2 ; ]dnB . So Z ;d (2 ) d kSt k1!1 c e;2;1 C 2 t d nB ; ] for all t 1. Alternatively, for the integral over B we use the semigroup property. Then kSt (I ; P0 ( ))k1!1 kSt=3k1!2kSt=3(I ; P0 ( ))k2!2kSt=3k2!1 for all 2 ; ]d and t > 0. But kP0( ) ; P0 (0)k2!2 1 for all 2 B C . Hence 0 kSt=3(I ; P0 ( ))k2!2 2e;3;1 t uniformly for all t > 0 and 2 B , by Proposition 2.10. Again using the estimates (10) it follows that there is a c0 > 0 such that ;1 kSt (I ; P0 ( ))k1!1 c0 e;4 t uniformly for all t 1 and 2 B . Combination of these estimates completes the proof.2 The second important step in the proof of Gaussian bounds is the comparison of the c kernel K with the corresponding kernel K of the homogenized system. Now the semigroup b generated by the homogenized operator H can be approximated as in Lemma 3.2 but c S b c with the eigenvalue ^0( ) and eigenprojector P0( ) corresponding to H . Hence to compare c b K and K one must compare the eigenvalues 0 , ^0 and the eigenprojectors P0, P0. The next proposition gives the key eigenvalue comparison. c Let "0, ^0 and ^0 be the parameters of Proposition 2.7 corresponding to the operator H . ^
14 Lemma 2.1 that Z kSt (I ; P0 (B ))k1!1 (2 );d

d kSt k1!1 + (2 );d

Proposition 3.3 There exists a c > 0 such that j 0( ) ; ^0( )j


C 2 B2;1 ( 0 ^ ^0 ) .

c j j3 uniformly for

b ^0( ) = ( C ) for all 2 B C by the last statement of Proposition 2.7. ^0 The maps 0 and 7! ' are holomorphic on B C , so by Lemma 2.8 for all 2 J (d) = 0 S1 f1 : : : dgn there are 2 C and ' 2 L2 1 (Id ) such that n=0

c c b 1 Proof Since H is an operator with constant coe cients one has H 1 = ( C )1 . So

2J (d) 2J (d) for all 2 B C . The second series converges both in L2(Id) and in L2 1(Id). Then = 0 0 and ' = 1 if j j = 0, since 0(0) = 0 and '0 = 1 , where j j = n if = (i1 : : : in) 2 J (d). But for all 2 L2 1(Rd) one has
X X

0(

)=

and ' =

'

2J (d)

2J (d)

' ) = 0( ) ( ' ) = ( H ' )


=
d X X k l=1

((@k + i k ) ckl(@l + i l) ' ) 2J (d)


m

(11)

for all 2 B C . 0 Let m 2 f1 : : : dg. Then comparing coe cients of


(m) (

gives

1 ) = h0( '(m)) +
d X k=1 d X k=1

d X k=1

i (@k ckm1 ) :

Taking = 1 gives

(m)

= 0. Moreover,

h0( '(m)) = ;i
for all 2 L2 1(Id ) and hence

(@k ckm1 )

(12)

@l'(m) = ;i

Xlk ckm
m n

(13) in (11)

for all l 2 f1 : : : dg, by Lemmas 2.11 and 2.12. Next let m n 2 f1 : : : dg with m 6= n. Then comparing coe cients of gives
(m n) (

1) +

(n m) (

1 ) = h0( '(m n)) + h0( '(n m)) + +


d X l=1 d X l=1

(i cml@l'(n)) +

d X k=1 d X k=1

(@k ckmi'(n)) (@k ckni'(m))

(i cnl@l'(m)) +

+ ( cmn1 ) + ( cnm1 ) 15

for all 2 L2 1(Id ). Substituting = 1 one nds


(m n) +

1 1 (n m) = ;i (1 cml @l'(n) ) ; i (1 cnl @l '(m) ) + =;


l=1 d X k l=1

d X

d X

Z
I

(1 cml Xlk ckn ) ; 1

l=1 d X

c + d mn

Z
Id

cnm

= cmn + cnm ^ ^ by (13). Similarly,


(n n)

k l=1

(1 cnl Xlk ckm) + Id cmn + Id cnm 1


Z
Id

=;

d X

for all n 2 f1 : : : dg. So


2 BC

k l=1

(1 cnl Xlk ckn ) + 1


X

cnn = cnn ^

0(

) = ^0( ) +

for all and the proposition follows. 2 0 ^ ^0 The above calculation also gives the rst-order term in the Taylor expansion of the holomorphic function 7! P0( ). For m 2 f1 : : : dg de ne
m

jj

d X k=1

Xk (C )ckm 2 L2

(Id)

and

t m

d X k=1

Xk (C t)ctkm 2 L2(Id)

where C t = (ctkm) with ctkm = cmk denotes the transpose matrix. These functions are the rst-order correctors of homogenization theory ( BLP], ZKO]). For 2 B C and 0 m 2 f1 : : : dg let P0(m)( ) = @P0( )=@ m.

Corollary 3.4 If m 2 f1 : : : dg then


P0(m)(0) ' = i
for all ' 2 L2(Id ).

t 2 L1 (Id ). m

Moreover,

t m P0 (0) ' ; i P0 (0) ( m ')

( ' )('y ') : ( P0( )') = ('y ' ) Hence we must calculate the coe cient of m on the right hand side and identify it with the expression for ( P0(m)(0)'). But it follows from (12) and Lemma 2.12 that

Proof If ' = P0 ( )1 and 'y = P0( ) 1 then 1

' =1 +i 'y = 1 + i

d X m=1 d X m=1

d X k=1

Xk (C )ckm + am 1 + O(j j2) = 1 + i

d X m=1

m1

m + am 1

+ O(j j2)

for some a 2 Cd. Similarly


m d X k=1 d X Xk (C t) ctkm + ay 1 + O(j j2 ) = 1 + i m m=1 m t + ay 1 m m

+ O(j j2)

16

for some ay 2 Cd . Hence


d X ('y ' ) = 1 + i (am ; atm) m + O(j j2) m=1

because (1 Xk (C )ckm) = 0 = (Xk (C t)ctkm 1 ). Therefore 1

') + ( 1 )(i tm ') (14) for all ' 2 L2(Id), by a simple calculation. The am, ay give no contribution to the rst m order terms. Next, the map 7! P0( ) is holomorphic from B C into L(L1 (Id) L1(Id)), by Lemma 3.1. 0 So P0(m)( ) 2 L(L1 (Id) L1(Id)) for all 2 B C . Therefore m = ;i P0(m)(0)1 2 L1 (Id) 1 0 t has mean value zero. Replacing C by C t gives the same for t . Note that this since m m implies that d X 2 ' =1 +i (15) m m + O(j j )
1 m)(1 both in the L2-, and the L1 -sense. The last statement rephrases (14). 2 The most important step in the proof of Gaussian bounds is to show that the homogc enized kernel K gives a rst-order asymptotic approximation to K .
m=1

( P0(m)(0)') = ( i

Theorem 3.5 There exists a c > 0 such that


c kKt ; Kt k1

c t;1=2 t;d=2

uniformly for all t 1.


c b 1 b Proof Since H 1 = ( C )1 if follows that P0( ) is a (possibly non-orthogonal) projecc b b tion onto the constant functions. But similarly (H ) 1 = ( C ) 1 and hence P0( ) is b also a projection onto the constant functions. Therefore P0( ) = P0(0) is the orthogonal C. projection onto the constant functions for all 2 B ^0 Next, by Lemma 3.2 there exist c > 0 such that b b kSt (I ; P0 (B ))k1!1 c e; t and kSt (I ; P0 (B ))k1!1 c e; t

uniformly for all t 1, where = 2;1 ( 0 ^ ^0). Then b b b kSt ; Stk1!1 2c e; t + kSt P0 (B ) ; St P0 (B )k1!1

for all t 1. Hence it follows from Lemma 2.1 that one needs to estimate ke; 0( )tP0( ) ; e;^0( )tP0(0)k1!1 for all 2 B . Obviously one has the decomposition ke; 0 ( )tP0 ( ) ; e;^ 0 ( )t P0 (0)k1!1 je; 0 ( )t ; e;^ 0( )t j kP0 (0)k1!1 + je; 0( )tj kP0( ) ; P0(0)k1!1 (16) and we estimate the two terms separately. 17

The rst term is easily handled with the eigenvalue estimates of Proposition 3.3. It follows from this proposition, Corollary 2.4 and a Duhamel estimate that
je;
0

for a suitable c > 0 uniformly for all t > 0 and 2 B . Next we consider the second term in (16). It follows from Lemma 3.1 that there is a c > 0 such that kP0 ( ) ; P0 (0)k1!1 c j j for all 2 B . Hence one deduces from Corollary 2.4 that there is a c0 > 0 such that ke; 0 ( )t P0 ( ) ; e;^ 0 ( )tP0 (0)k1!1 c0 (j j + t j j3)e; C j j2 t (17) uniformly for all t > 0 and 2 B . Then Z bt P0 (B )k1!1 c0 d (j j + t j j3)e; C j j2 t b kSt P0 (B ) ; S

Z ( )t ; e;^ 0 ( )t j = jt 1 ds e;t(s 0 ( )+(1;s)^ 0 ( )) ( ^ ( ) ; ( 0 0 0 Z1 c t ds e;t C j j2 j j3 = c t j j3 e;t C j j2 0

))j

c0

where c00 = c0 RRd d e; C j j2 (j j + j j3). It now follows by combination of these estimates that b kSt ; St k1!1 2c e; t + c00 t;1=2 t;d=2

d (j j + t j j3)e; C j j2t = c00 t;1=2 t;d=2 d

uniformly for all t 1. Thus one can now choose a c0 > 0 such that c b kKt ; Kt k1 = kSt ; St k1!1 c0 t;1=2 t;d=2

(18)

uniformly for all t 1. 2 c By interpolation one deduces Gaussian bounds for the di erence of K and K , together with a multiplicative factor which is almost t;1=2.

Corollary 3.6 For all " > 0 there exist b c > 0 such that
uniformly for all t 1 and x y 2 Rd.

c jKt (x y ) ; Kt (x ; y )j c t;(1;")=2 Gb t (x ; y )

Proof By Theorem 3.5 there exists a c > 0 such that

c jKt (x y ) ; Kt (x ; y )j c t;1=2 t;d=2

c uniformly for all t 1 and x y 2 Rd. Next K satis es the bounds (8) and because H is a c strongly elliptic operator with constant coe cients K satis es Gaussian bounds c jKt (x ; y )j c0 Gb t (x ; y ) :

18

Therefore there exist b c0 > 0 and N 2 N such that c jKt (x y ) ; Kt (x ; y )j c0 tN t;d=2 e;bjx;yj2 t;1 uniformly for all t 1 and x y 2 Rd . Then
c c c jKt (x y ) ; Kt (x ; y )j = jKt (x y ) ; Kt (x ; y )j" jKt (x y ) ; Kt (x ; y )j1;"

(c0)" c1;"t;(1;"(1+2N ))=2 t;d=2 e;b"jx;yj2 t;1

for all " 2 h0 1], t 1 and x y 2 Rd. The proof of the corollary is complete. 2 Finally the Gaussian bounds of Theorem 1.1.I are an immediate consequence. Since c K satis es Gaussian bounds uniformly for t > 0 the bounds follow for large t from Corollary 3.6. But for small t the bounds are well known.

3.2 Holder estimates

In this subsection we prove the second statement of Theorem 1.1, the Gaussian bounds on the Holder derivatives of the kernel K . The proof is similar in outline to the proof of Gaussian bounds in Subsection 3.1. We begin with the bounds of Auscher, Aus] Theorem 4.3, which establish that for all 2 h0 1i there exist b c > 0 and N 2 N such that jKt (x ; h y ) ; Kt (x y )j c (jhjt;1=2) (1 + t)N Gb t (x ; y ) (19) for all h x y 2 Rd and t > 0 with jhj t1=2. These bounds rely on the uniform continuity of the coe cients. We use the spectral estimates to improve these bounds by removing the growth factor. The tactic is to estimate the di erence between the Holder derivative of K c and the Holder derivative of K .

Theorem 3.7 There exists a c > 0 such that


uniformly for all t 1 and x y h 2 Rd .

c c j(Kt ; Kt )(x ; h y ) ; (Kt ; Kt )(x y )j c t;(1; )=2(jhjt;1=2) t;d=2

Proof The basic idea of the proof is to prove rst uniform bounds
b k(I ; U (h))(St ; St )k1!1

c jhj t;1=2 t;d=2

(20)

uniformly for all h 2 Rd and t 1 with jhj 1 and then remove the restriction on jhj. Again we use decomposition theory to isolate the principal contribution which comes from the lowest eigenstate. This is an elaboration of the arguments of Subsection 3.1 but the decomposition of translations introduces an additional term. Fix 2 h0 1i. If = 2;1 ( 0 ^ ^0), as before, then by Lemma 3.2 there exist c > 0 such that kSt(I ; P0(B ))k1!1 c e; t uniformly for all t 1. Then there is a c0 > 0 such that k(I ; U (h))St(I ; P0 (B ))k1!1 k(I ; U (h))S1 k1!1 kSt;1(I ; P0(B ))k1!1 c0 jhj e; t 19

b uniformly for all h 2 Rd with jhj 1 and t 2. Since similar estimates are valid for S we will have proved (20) once we can show that there is a c > 0 such that b b k(I ; U (h))(St P0 (B ) ; St P0 (B ))k1!1 c jhj t;1=2 t;d=2

uniformly for all h 2 Rd and t 1 with jhj 1. If ' 2 L2(Id) and h 2 Rd de ne L(h)' 2 L2(Id) by (L(h)')(u) = (e2 i(u;h)) where 2 L2(Td ) is given by (e2 iu ) = '(u). Then L(h) is a unitary operator and it follows that the operator U of translations on Rd has the decomposition

U (h) ;1 = (2 );d
for all h 2 Rd. Therefore

]d

d e;i hL(h)

(21)

b b (I ; U (h))(StP0(B ) ; StP0(B )) ;1 Z ;d d (1 ; e;i h )(e; 0( )tP0 ( ) ; e;^0( )tP0(0)) = (2 )

+ (2 );d

d e;i h (I ; L(h))(e; 0( )tP0( ) ; e;^0( )tP0 (0))

for all h 2 Rd and t > 0. Then by Lemma 2.1


bb k(I ; U (h))(St P0 (B ) ; StP0 (B ))k1!1

(2 );d

+ (2 );d

d j1 ; e;i hj ke; 0( )tP0( ) ; e;^0( )tP0(0)k1!1


Z
B

d k(I ; L(h))(e; 0 ( )tP0( ) ; e;^0( )tP0(0))k1!1 : (22)

The rst term on the right hand side is the additional contribution from the decomposition of the translation. If c0 is as in (17) then Z ;d d j1 ; e;i hj ke; 0( )tP0( ) ; e;^0( )tP0(0)k1!1 (2 )
B

(2 );dc0
Z

for all h 2 Rd and t > 0 with jhj t1=2 and c00 = c0 RRd d e; C j j2 j j(j j + j j3). The second term on the right hand side of (22) can be estimated by the reasoning of Subsection 3.1. First we need some preliminary estimates comparable to (10). If one extends the function u 7! Kt (u v) to a periodic function on Rd for all v 2 Id then X i(n;x+v) Kt (x v) = e Kt(x ; n v) n2Zd 20

c0 jhj Rd d j j (j j + t j j3)e; C j j2t = c00 jhj t;(d+2)=2 c00t;1=2 (jhjt;1=2) t;d=2

d jhj j j (j j + t j j3)e; C j j2t

for all 2 Cd , t > 0, x 2 Rd and v 2 Id. Using the bounds (19) and the Gaussian bounds of Theorem 1.1.I it follows that there are b c ! > 0 and N 2 N such that X i(n;u+h+v) jKt (u ; h v ) ; Kt (u v )j je ; ei(n;u+v) j jKt (u ; n v )j n2Zd X i(n;u+h+v) + je j jKt (u ; h ; n v ) ; Kt (u ; n v )j d n2Z c jhj t1=2(1 ^ t);d=2e;bku;vk2 t;1 e!(Im )2t + c (jhjt;1=2) (1 + t)N (1 ^ t);d=2e;bku;h;vk2 t;1 e!(Im )2t uniformly for all 2 Cd , t > 0, h 2 Rd and u v 2 Id with jhj t1=2 ^ 1. Since (u v) 7! K (u ; h v) is the kernel of the operator L(h) St it follows that there are c ! > 0 such that (23) k(I ; L(h))St kp!q c (jhjt;1=2) (1 + t)N (1 ^ t);d(1=p;1=q)=2e!(Im )2 t uniformly for all 1 p q 1, 2 Cd , t > 0 and h 2 Rd with jhj t1=2 ^ 1. Next de ne the Banach space C = f' 2 C (Id ) : sup jhj; k(I ; L(h))'k1 < 1g 0<jhj 1 with norm k'kC = k'k1+sup0<jhj 1 jhj; k(I ;L(h))'k1. Note that '(u1 : : : 0 : : : ud) = '(u1 : : : 1 : : : ud) for all ' 2 C , u 2 Id and k 2 f1 : : : dg, where the 0 and 1 are in the k-th position. Then (23) together with (10) implies that St is a bounded operator from L1(Id) into C and kSt kL (Id )!C c (1 + t)N (1 ^ t);d(1=p;1=q)=2e!(Im )2 t
1

uniformly for all 1 p q 1, 2 Cd and t > 0. In particular, the family of operators is bounded locally uniform in . Arguing as in the proof of Lemma 3.1 it follows that the spectral projections 7! P0( ) form a holomorphic family of operators from B C into 0 L(L1 (Id ) C ). Hence there is a c > 0 such that kP0 ( ) ; P0 (0)kL1 (Id )!C c j j uniformly for all 2 B . In particular

c j j jhj for all 2 B and h 2 Rd with jhj 1. This gives estimates, as in the proof of Theorem 3.5, Z ;d d k(I ; L(h))(e; 0( )tP0( ) ; e;^0( )tP0(0))k1!1 c0 t;1=2 jhj t;d=2 (2 )
uniformly for all h 2 Rd and t 1 with jhj 1 which bound the second term in (22). Adding the contributions it follows that there exists a c > 0 such that b k(I ; U (h))(St ; St )k1!1 c t;1=2 jhj t;d=2 (24)
B

k(I ; L(h))(P0 ( ) ; P0 (0))k1!1

uniformly for all h 2 Rd and t 2 with jhj 1. But then it follows from (18) that (24) is also valid for all h 2 Rd with jhj 1 and by (19) also for t 2 1 2], with a possibly larger 21

value of c. So the estimates (24) are valid uniformly for all h 2 Rd and t 1 and the theorem follows. 2 With several applications of interpolation one can deduce rather sharp Gaussian bounds for the Holder derivatives of the di erence. Theorem 3.8 For all " 2 h0 1i and > 0 there exist b c > 0 such that jhj + jk j c c G (x;y) (25) j(Kt ; Kt )(x ; h y ; k ) ; (Kt; Kt )(x y )j c t;(1;")(1; )=2 jx ; y j + t1=2 b t uniformly for all t 1 and x y h k 2 Rd with jhj + jkj jx ; y j + t1=2. Proof Interpolation of the bounds of Theorem 3.7 with the bounds (19) of Auscher, c which are of course also valid for K , implies that for all " 2 h0 1i there exist b c > 0 such that c c j(Kt ; Kt )(x ; h y ) ; (Kt ; Kt )(x y )j c (jhjt;1=2) t;(1;")(1; )=2 Gb t (x ; y ) uniformly for all t 1 and h x y 2 Rd with jhj t1=2. These bounds are a special case of (25). To obtain the general bounds we use two additional arguments. First it follows from the last bounds that for all " 2 h0 1i there exists a c > 0 such that b k(I ; U (h))(St ; St)k2!1 c (jhjt;1=2) t;(1;")(1; )=2 t;d=4 (26) uniformly for all h 2 Rd and t 1 with jhj t1=2. But by Theorem 3.5 and increasing the value of c it follows that the bounds (26) are valid uniformly for all h 2 Rd and t 1. b The semigroup S satis es similar bounds: there exists a c > 0 such that b k(I ; U (h))St k2!1 c (jhjt;1=2) t;d=4 for all h 2 Rd and t > 0. So there is a c > 0 such that k(I ; U (h))St k2!1 c (jhjt;1=2) t;d=4 for all h 2 Rd and t 1. Then by the previous estimates and duality one deduces that for all " 2 h0 1i there exists a c > 0 such that b b k(I ; U (h))(St ; St)(I ; U (k ))k1!1 k(I ; U (h))St=2k2!1 k(St=2 ; St=2)(I ; U (k ))k1!2 b b + k(I ; U (h))(St=2 ; St=2)k2!1kSt=2(I ; U (k))k1!1 c (jhj jkj t;1) t;(1;")(1; )=2 t;d=2 for all h k 2 Rd and t 2. Together with the estimates of Theorem 3.7 and duality one establishes that for all " 2 h0 1i there exists a c > 0 such that c c j(Kt ; Kt )(x ; h y ; k ) ; (Kt ; Kt )(x y )j c (jhj jkj t;1) + (jhjt;1=2) + (jkjt;1=2) t;(1;")(1; )=2 t;d=2 uniformly for all h k x y 2 Rd and t 1. Finally we can apply Lemma 4.2 of ElR2]. This lemma allows one to interpolate between the last bounds and the bounds of Corollary 3.6 to obtain the bounds (25) for any slightly smaller . 2 Obviously Theorem 1.1.II is an immediate consequence of Theorem 3.8. 22

3.3 Asymptotic estimates

In the foregoing derivation of kernel bounds the estimate of Theorem 3.5 played a key role. The proof of this estimate can, in principle, be extended to obtain a complete asymptotic c c expansion of K in terms of K , the derivatives of K and the correctors of homogenization theory. In this subsection we discuss the rst-order term in this expansion. The zero-order term in the expansion was derived by rst making a spectral decomposition, secondly approximating the components in terms of an integral over small j j of the contribution coming from the eigenprojector P0( ) and the eigenvalue 0( ) and thirdly approximating the eigenprojector and eigenvalue by the lowest order terms in their holomorphic expansions. The rst-order term in the expansion is derived by the same process but arises from the next terms in the holomorphic expansions. In the case of the eigenprojector this is the linear term in and for the eigenvalue the term of order 3. The expansion for the eigenvalue has the form X 0( ) = ^0( ) + 2J (d) j j 3 for 2 B C^ ^0 as a consequence of Proposition 3.3. 0

Lemma 3.9 For all m 2 f1 : : : dg the correctors m and tm extend to continuous periodic functions on Rd . Proof It follows from (15) that P0( )1 = ' = 1 + i Pd =1 m m + O(j j2) in the L1 1 m sense. But the map 7! P0 ( ) is holomorphic from B C into L(L1 (Id) C ), by the proof of
Theorem 3.7. Hence function on Rd.
m

2 C for all m 2 f1 : : : dg and it extends to a continuous periodic

Proposition 3.10 There exists a c > 0 such that


X c jKt (x y ) ; Kt (x ; y ) ;
d m=1 t m (x) ; m (y )

c (@mKt )(x ; y) X c (@ Kt )(x ; y)j c t;1 t;d=2

; it

2J (d) j j=3

uniformly for all t 1 and x y 2 Rd.

Remark Each term occurring in the di erence estimated in the proposition satis es a Gaussian bound. Therefore one can interpolate between the uniform bound of the proposition and the Gaussian bounds to obtain a Gaussian bound for the di erence. Then, for each " > 0, one obtains a bound c t;(1;") Gb t(x ; y) for t 2 1 1i. Proof of Proposition 3.10 The proof is based on the same reasoning used to prove Theorem 3.5 and we only sketch the main ideas. First it follows from the previous proof that one has bounds b b b kSt ; St ; St P0 (B ) + St P0 (B )k1!1 c e; t (27)
23

b b for all t 1. To estimate StP0 (B ) ; StP0(B ) we now use the identity e; 0( )tP0( ) ; e;^0( ))tP0(0) = e;^0( )t (P0( ) ; P0(0)) + t(^0( ) ; 0( ))P0 (0)

+ te;^0( )t(^0( ) ; 0( ))(P0 ( ) ; P0 (0)) Z1 + t2(^0( ) ; 0( ))2 ds (1 ; s) e;^0( )t(1;s);


0

( )ts P

0(

which is readily veri ed with two applications of the Duhamel identity. The most signi cant contribution to the operator crossnorm comes from the term

Lt( ) = e;^0( )t

X (m) m P0 (0) ; t m=1 j j=3

d X

P0(0)

arising from the expansion of the the bracketed factor in the rst term, where P (m) is as in Corollary 3.4 and the as in the proof of Proposition 3.3. All other contributions are bounded by a multiple of t;1 t;d=2. But it follows from Corollary 3.4 and the observation b that P0(0) = P0( ) for j j < , that j j=3 where M tm denotes the multiplication operator with the function tm on the space L2(Id). If one di erentiates (21) then Z b b b @m St ;1 = (2 );d ; ]d d (i m St + @mSt ) and Z b b @mStP0(B ) ;1 = (2 );d d ie;^0( )t m P0(0)
m=1

Lt ( ) =

d X

i me;^0( )t

b m P0 (

b ) ; P0( )M

t m

; it

b (i ) e;^0( )tP0( )

b b b since @mP0( ) = @mP0(0) = 0. We next determine a decomposition for m @mStP0(B ) for

all m 2 f1 : : : dg. If ' 2 Cc(Rd) with (') = (' ) 2 ; ]d then (' m) ( ) = So Z bt P0 (B ) ;1 = (2 );d d e;^ 0 ( )t i m b m @m S for all t > 0. Similarly the terms with Therefore (27) gives an estimate
b kSt ; St ;
d X m=1 t m B

m'

for all 2 ;

]d. (28)

m P0 (0)

and third order terms in can be handled. ) ; it


X bb @ StP0 (B )k1!1 c t;1 t;d=2

j j=3 b uniformly for t 1. But then P0(B ) can be replaced by the identity because this only introduces exponential corrections. The resulting bounds translate directly into the uniform bounds on the kernels given in the proposition. 2 Under some quite general circumstances the rst-order correction in the asymptotic expansion of K are zero. De ne div C = 0 to mean that Pd=1 @k ckm = 0, in the sense of k distributions, for each m 2 f1 : : : dg. 24

b b b m @m St ; @m St M t P0 (B m

Lemma 3.11 I. The correctors m = 0 for all m 2 f1 : : : dg if, and only if, div C = 0. II. The correctors tm = 0 for all m 2 f1 : : : dg if, and only if, div C t = 0. III. If div C = div C t = 0 then = 0 for all 2 J (d) with j j = 3. Proof First d
(H0
m) = ;

for all 2 D(H0 ) \ L? (Id) and hence for all 2 D(H0 ) and m 2 f1 : : : dg. But D(H0 ) 2 is dense in L2 1(Id ). Hence if m = 0 for all m 2 f1 : : : dg then div C = 0. Conversely, if div C = 0 then (H0 m) = 0 for all 2 D(H0 ) since Cc1(Id) is dense in L2 1(Id) and L2 1(Id) D(H0 ). As H0 (L? (Id) \ D(H0 )) = L? (Id) and m 2 L? (Id) by Lemma 2.11 it 2 2 2 follows that m = 0. The second statement follows from the rst by replacing C with C t. b Next for 2 B 0 set R( ) = 0( ) ; ( C ). If div C t = 0 then
0(

k=1

(@k ckm)

) (1 P0 ( )1 ) = (1 H P0( )1 ) = (1 ( C ) P0 ( )1 ) 1 1 1 1 1 1

b R( ) = (1 ( C ) ; ( C ) P0( )1 )=(1 P0 ( )1 ) : 1 1 1 1 b Moreover, div C t = 0 implies that C is the mean of C over Id and hence (1 ( C ) ; 1 b ) P0 (0)1 ) = 0. Therefore ( C 1 b R( ) = (1 ( C ) ; ( C ) P0( ) ; P0(0) 1 )=(1 P0 ( )1 ) : 1 1 1

and hence

But if, in addition, div C = 0 then kP0( ) ; P0(0)k2!2 = O(j j2) by Corollary 3.4. Hence jR( )j = O(j j4 ). 2 Note that if div C t = 0 then the coe cients (k l m) of the third-order terms in R( ) are given by d X 1 1 (29) (k l m) = i (1 (ckl ; (1 ckl 1 ))Xn cnm ): where we have use the explicit form of the correctors given in Subsection 2.3 and Corollary 3.4. The last two results allow one to conclude that the asymptotic estimate (18) is usually optimal.
n=1

Proposition 3.12 The following conditions are equivalent. b I. There is a c > 0 such that kSt ; Stk1!1 c t;1 t;d=2 uniformly for t 1. b II. There are c " > 0 such that kSt ; Stk1!1 c t;(1+")=2 t;d=2 uniformly for t 1. III. div C = div C t = 0.
25

Proof It follows from Lemma 3.11 and Proposition 3.10 that III)I. But it is evident
that I)II. If I is valid then it follows from Proposition 3.10 that there is a c0 > 0 such that d X X t c c j (@ Kt)(x ; y)j c0 t;"=2 t;(d+1)=2 m(x) ; m (y ) (@m Kt )(x ; y ) + it m=1 2J (d) j j=3 for all t 1 and x y 2 Rd. Then
j
d X m=1 t 1=2 m(x) ; m (x ; t y ) d X m=1

c (@mK1)(y) + i

2J (d) j j=3

c (@ K1)(y)j X

= t(d+1)=2j

t 1=2 m (x) ; m (x ; t y )

c (@mKt)(t1=2y) + i t

c0 t;"=2 for all x y 2 Rd and t 1. Since the m and tm are periodic functions with mean value zero, integration with respect to x over Id gives X c j (@ K1)(y)j c0t;"=2 2J (d) j j=3 c for all y 2 Rd and t 1. Then P 2J (d) j j=3 (@ K1)(y) = 0 by taking the limit t ! 1. So d X t 1=2 0 ;"=2 c j m (x) ; m (x ; t y ) (@m K1 )(y )j c t
for all x y 2 and t 1. Now let x y0 2 Rd. Since tm has again mean value zero it follows by integration with respect to y over the set y0 + t;1=2Id that
j
d X m=1 m

2J (d) j j=3

c (@ Kt )(t1=2y)j

Rd

m=1

(x) td=2

y0 +t;1=2 Id m=1

c dy (@mK1)(y)
y0 +t;1=2 Id

d X d=2 Z t

dy

t (x ; t1=2y ) m

c c (@mK1 )(y) ; (@mK1)(y0) j

c0t;"=2 R c c c for all t > 0. But limt!1 td=2 y0 +t;1=2 Id dy (@mK1)(y) = (@mK1)(y0) and j(@mK1 )(y) ; c c c (@mK1)(y0)j kr@mK1 k1 jy ; y0j = d1=2kr@mK1k1 t;1=2 for all y 2 y0 + t;1=2Id. So taking the limit t ! 1 one deduces that
d X

for all x y0 2 Rd. Scaling then gives Pd =1 m t > 0. Then for all x 2 Rd one has
d X m=1 m (x) m

m=1

c m (x) (@m K1 )(y0 ) = 0 c m (x) (@m Kt )(y ) c m(x) Kt (y )

= 0 for all x y 2 Rd and


y

= lim t#0

dy d

d X

= ; lim Rd dy e t#0 26

m=1

@ e @ym

d yX c m (x) (@m Kt )(y ) = 0 m=1

and m = 0 for all m 2 f1 : : : dg. Similarly tm = 0 for all m 2 f1 : : : dg. Then Statement I is a consequence of Lemma 3.11. 2 Finally we note that the third order terms in the holomorphic expansion of 7! 0( ) vanish in the case of symmetric coe cients. We temporarily adopt the notation H (C ) = H , 0( C ), etc. to denote the dependence of the various element on the matrix of coefcients.

Lemma 3.13 There is a > 0 such that


0(

C ) = 0 (; C t )

for all 2 B C. Hence if C is symmetric, i.e., if C = C t, then 7! 0 ( C ) is an even function for small j j.

of H (C ) . Let ' (C ) be a corresponding eigenfunction. Then

Proof Since 0( C ) is an eigenvalue of H (C ) it follows that 0 ( C ) is an eigenvalue


H (C ) ' (C ) = 0( C ) ' (C ) :

Since H (C ) = H; (C t) one deduces that

H (C ) '; (C t) = 0(; C t) '; (C t) :


Thus 0(; C t) is an eigenvalue of H (C ) such that 0(; C t) ! 0 as j j ! 0. Hence, by simplicity of the eigenvalue, 0(; C t) = 0( C ) for all small j j (see Proposition 2.7).2 Despite this last result there are non-symmetric examples for which the third-order terms are non-zero. If C = I + "B then H is a perturbation of the Laplacian for small ". Now for d = 2 set ! e2 ix2 0 : B (x) = e;2 ix2 0 Then div B t = 0, (1 B 1 ) = 0 and div B (x) = (;2 ie;2 ix2 0). Since H0 = 4 2 if 1 ;2 ix2 it follows from (29) that the coe cient (1 1 1) = (2 );1"2 6= 0. (x) = e

4 Holder continuous coe cients


If the coe cients are Holder continuous of order 2 h0 1i then it follows from AMT] or ElR1] that the kernel is once-di erentiable in both variables and the derivatives are Holder continuous of order . Moreover, these derivatives satisfy Gaussian bounds but with an extra factor which grows polynomially with the t variable. In this section we establish the hierarchy of regularity estimates for periodic systems described in Theorems 1.2, 1.3 and 1.4 which are uniform in time.

27

In this subsection we prove Theorem 1.2. Suppose 2 h0 1i is such that ckl 2 C for all k l 2 f1 : : : dg. We begin with the derivatives. Since the statements concerning the derivative of the kernel with respect to the second variable follow by duality from similar statements with respect to the rst variable, we only need to consider the rst variable. We set @k = @=@xk, the derivative with respect to the rst variable in the k-th direction. The proofs then follow by a variation of the methods of Section 3. Lemma 4.1 For all 2 Cd with j j < 0 the projection P0( ) is bounded from L1(Id) into L1 1(Id ) and @k P0 ( ) 2 L(Id) C ) for all k 2 f1 : : : dg. Moreover, the map 7! @k P0( ) is holomorphic from B C into L(L1(Id) L1 (Id)) and from B C into L(L1(Id) C ) for all 0 0 k 2 f1 : : : dg.
j(@k Kt )(x y )j c t;1=2 (1 + t)N Gb t (x ; y ) for all k 2 f1 : : : dg, x y 2 Rd and t > 0. Using the identity (9) one sees that K is di erentiable, the derivative of St maps L1(Id) into L1(Id) and 7! k@k St k1!1 is locally

4.1 Derivatives on the kernel

that

Proof It follows from AMT], Theorem 4.15, that there exist b c > 0 and N 2 N such

uniformly bounded. Then the proof is a variation of the arguments given in the proof of Lemma 3.1. Finally, the Holder estimates follow by a small change of the argument starting from the kernel bounds j(@k Kt )(x y ) ; (@k Kt )(x ; h y )j c t;1=2(jhjt;1=2) (1 + t)N Gb t (x ; y ) of Theorem 4.15 in AMT]. 2 Obviously @k P0(0) = 0 for all k. For k l 2 f1 : : : dg let Pkl 2 L(L1(Id) L1 (Id)) be the partial derivatives of 7! @k P0( ). So
d X supd j j;2 k@k P0( ) ; l Pklk1!1 < 1 : 2C l=1 0<j j 2;1 0 Comparison with Corollary 3.4 identi es the operators Pkl in terms of the correctors l. Lemma 4.2 If k l 2 f1 : : : dg then l 2 L1 1(Id), @k l 2 C and Pkl = i (@k l) P0 (0). Proof It follows from (15) that

M = sup k2f1 ::: dg

P0( )1 = ' = 1 + i 1

d X l=1

l l + O(j

j2 )

in the L2-sense. But the map 7! @k P0 ( ) from B C into L(L2(Id ) L2(Id)) is also holo0 morphic for all k 2 f1 : : : dg and, moreover, l 2 L2 1(Id) for all l 2 f1 : : : dg, by Lemma 2.11. Hence d X @k P0( )1 = i l @k l + O(j j2) 1
l=1

28

in the L2-sense. On the other hand the map 7! @k P0( ) from B C into L(L1 (Id) L1(Id)) is 0 also holomorphic. Therefore @k l 2 L1 (Id) for all l 2 f1 : : : dg and hence k 2 L1 1(Id). Similarly, @k l 2 C . It follows from Corollary 3.4 that ( Pkl') = ( i @k l) (1 ') for all ' 2 L1(Id), 1 which implies the last statement. 2 We are now able to formulate an analogue of Theorems 3.5, 3.7 and 3.8 and Corollary 3.6 for the asymptotic behaviour of @k Kt for large t. Then the Gaussian bounds for @k Kt follow as before. Theorem 4.3 Assume the coe cients are Holder continuous and periodic with a common period. Then there exists a c > 0 such that d X c c j(@k Kt )(x y ) ; (@k Kt )(x ; y ) ; (@k l )(x) (@lKt )(x ; y )j c t;1 t;d=2
uniformly for all k 2 f1 : : : dg, t 1 and x y 2 Rd . Moreover, for all " 2 h0 1i there exist b c > 0 such that
c j(@k Kt )(x y ) ; (@k Kt )(x ; y ) ;
d X l=1 l=1

c (@k l)(x) (@lKt)(x ; y)j c t;(1;")=2 t;1=2 Gb t(x ; y)

uniformly for all k 2 f1 : : : dg, t 1 and x y 2 Rd . Proof Fix k 2 f1 : : : dg. Since


c (x y) 7! (@k Kt )(x y) ; (@k Kt)(x ; y) ; b is the kernel of the operator @k St ; @k St ; prove that there is a c > 0 such that b b k@k St P0 (B ) ; @k St P0 (B ) ;
d X l=1

d X l=1

c (@k l)(x) (@lKt )(x ; y)

Pd

b l=1 (@k l ) @l St

it su ces, by Lemma 3.2, to (30)

bb (@k l) @lStP0 (B )k1!1 c t;1 t;d=2

for all t 1, where = 2;1 ( 0 ^ ^0). If one di erentiates (21) then Z @k St ;1 = (2 );d d (i k St + @k St ) ; ]d and arguing as in (28) gives Z Z bb (@k l) @lStP0 (B ) ;1 = (2 );d d e;^0( )ti l (@k l) P0 (0) = (2 );d d e;^0( )t l Pkl B B for all t > 0, where we used Lemma 4.2 in the last step. Hence d X b b bb @k StP0(B ) ; @k StP0(B ) ; (@k l) @lStP0(B ) ;1 = (2 );d
Z
B l=1

d i k (e; 0( )tP0( ) ; e;^0( )tP0(0)) + e; 0( )t @k P0( ) ;


+ (e; 29
0

d X l=1

l Pkl

Arguing as before (30) is valid and the proof of Theorem 4.3 is complete.

d ( )t ; e;^ 0 ( )t ) X P l kl l=1

The estimates of Theorem 1.2 on the derivatives (@xk Kt)(x y) now follow from Theoc rem 4.3 and similar estimates on the derivatives of the Gaussian kernel Kt . Next we examine the second statement of Theorem 1.2 on the rst-order Riesz transforms. The proof of this result is simpli ed by noting that if one has the inclusion D(H 1=2) Lp 1(Rd) and the lower bounds

cp 1maxd k@k 'kp kH 1=2'kp k

(31)

for all ' 2 D(H 1=2) then it follows that Lq 1(Rd) D((H )1=2) and the upper bounds k(H )1=2 kq c0p max k@k kq
1 k d

are valid for all 2 Lq 1(Rd) with q dual to p. This follows starting from the identity ((H )1=2 ') = (H

H ;1=2') =

d X

k l=1

(@l ckl @k H ;1=2')

for all ' 2 D(H ;1=2 ) Lp(Rd) and 2 D(H ) Lq (Rd). Hence we need only establish the lower bounds (31) for H . Fix p 2 h1 1i. As the coe cients are Holder continuous, it is known (see ElR1], Theorem 4.6) that D(H 1=2) = Lp 1(Rd). Therefore the operators Rk " = @k ("I + H );1=2 are bounded on Lp(Rd) for all " > 0. The essence of the proof of (31) is to establish a uniform bound max sup kRk "kp!p < 1 : 1 k d This bound is obtained by rst observing that Z ;1 @k 1 dt e;"t t;1=2St = R(1) + R(2) Rk " = ;(1=2) k" k"
0

">0

with and

R(1) = ;(1=2);1 @k k"

Z1
0

dt e;(1+")tt;1=2St = @k ((1 + ")I + H );1


Z

But one has bounds


kR(1)kp!p k"

1 R(2) = ;(1=2);1 dt e;"t (1 ; e;t)t;1=2@k St k" 0

: c1 k@k (I + H );1=2kp!p

k@k (I + H );1=2 kp!p k(I + H )1=2 ((1+ ")I + H );1=2 kp!p

with c1 independent of " because H has a bounded H1 -holomorphic functional calculus. This is true for p = 2 since H is injective and maximal accretive ADM], Theorem G, and it then follows for p 2 h1 1i by DuR], Theorem 3.4. Fix 2 h0 1i. Next we use the estimates d bt ; X(@k l )@lSt kp!p c0 t;1=2(1 + t); =2 : b k@k St ; @k S
l=1

30

b These follow for t 1 from the bounds on @k St and @k St and they follow for t immediate corollary of Theorem 4.3. One deduces that Z1 kR(2)kp!p c0 ;(1=2);1 dt e;"t(1 ; e;t)t;1(1 + t); =2 k"
0 0

1 as an

Z d ;1 k 1 dt e;"t(1 ; e;t)t;1=2 @k St + X(@k l)@lSt kp!p : b b + ;(1=2)

But the rst term on the right hand side has a nite bound, uniform for all " > 0, since 2 h0 1i. Moreover, the second term on the right is the Lp -norm of the operator
d X c c c c @k ("I + H );1=2 ; ((1 + ")I + H );1=2 + (@k l) @l("I + H );1=2 ; @l((1 + ")I + H );1=2 l=1

l=1

c which is bounded, uniformly for " > 0, since the Riesz transforms of H are bounded. Thus (1) (2) Rk ", Rk " and Rk " are all norm bounded uniformly in ". It follows from the foregoing that max k@ 'k C k("I + H )1=2'kp 1 k d k p

for all " > 0 and ' 2 Lp 1(Rd) with C independent of ". Then (31) follows by taking the limit " ! 0.

4.2 Higher order derivatives of the kernel

The hierarchy of smoothness of the kernel together with optimal Gaussian bounds does not usually extend beyond one derivative. The situation is comparable with the scale of genuine Gaussian bounds for derivatives of the kernel of the Laplacian on groups with polynomial growth (see ERS]). In this subsection we prove the equivalence of the conditions in Theorem 1.3 by proving the implications III)II)IV)I)III and then I)V)IV. Finally we prove Theorem 1.4. Obviously the implication Theorem 1.3.III)II is trivial. Proof of Theorem 1.3.II)IV By (15) in Lemma 4.2 of ElR2] and Theorem 1.2 it follows that there are b c > 0 such that j(@k Kt )(x ; h y ) ; (@k Kt )(x y )j c t;1=2 Gb t (x ; y ) (32) uniformly for all x y h 2 Rd, t > 0 and k 2 f1 : : : dg with jhj t1=2. Interpolation with the bounds of Condition II establishes that for all " 2 h0 i there are b0 c0 > 0 such that j(@k Kt )(x ; h y ) ; (@k Kt )(x y )j c0 t;1=2 (jhjt;1=2) ;" Gb0 t(x ; y ) uniformly for all x y h 2 Rd, t 1 and k 2 f1 : : : dg with jhj 1. Then the bounds of Condition IV follow by integration of the kernel. We next show that the implication IV)I is a consequence of Theorem 4.3. Proof of Theorem 1.3.IV)I Suppose Condition IV is valid and that 2 h0 1i, cp > 0 and p 2 1 1] are such that (3) is satis ed. Fix k 2 f1 : : : dg. Then k(I ; L(h))@k St kp!p cp t;(1+ )=2 (33) 31

for all t 1 and h 2 Rd with jhj 1. We rst show that these bounds are also valid with p replaced by 2. The Gaussian bounds of Theorem 1.2 imply that there is a c > 0 such that k@k St kq!q c t;1=2 (34) uniformly for all q 2 1 1] and t > 0. Interpolating between (33) and (34) by taking q = 1 if p > 2 and q = 1 if p < 2 establishes that there are (probably di erent) c > 0 and 2 h0 1i such that k(I ; U (h))@k St k2!2 c t;(1+ )=2 uniformly for all t 1 and h 2 Rd with jhj 1. It follows from Theorem 4.3 that there is a c0 > 0 such that
b k@k St ; @k St ;
d X l=1

b (@k l) @lStk2!2 c0 t;(1+ )=2


d X l=1

uniformly for all t 1. So


b k(I ; U (h))@k St ; (I ; U (h))@k St ; (I ; U (h)) b uniformly for all t 1. Since St satis es bounds b k(I ; U (h))@l St k2!2
d X l=1

b (@k l) @lStk2!2 2c0 t;(1+ )=2

c00 t;1=2 (jhjt;1=2)

(35)

uniformly for all t > 0 and h 2 Rd it follows that


k(I ; U (h))
b (@k l) @lStk2!2 (c + 2c0 + c00 jhj ) t;(1+ )=2 (c + 2c0 + c00) t;(1+ )=2

for all t 1 and h 2 Rd with jhj 1. Using the bounds (35) once again, the boundedness of the @k l and the identity ((U (h) ; I )(' ) = U (h)' (U (h) ; I ) + (U (h) ; I )' one establishes that there is a c000 > 0 such that
k
d X l=1

b (I ; U (h))(@k l) @lStk2!2 c000 t;(1+ )=2

(36)

uniformly for all t 1 and h 2 Rd with jhj 1. Fix h 2 Rd with jhj 1 and set
l

= (I ; U (h))(@k l)

for all l 2 f1 : : : dg. We show that the periodic function l vanishes for all l. (The following elegant argument is due to Nick Dungey.) It follows from (36) and the semigroup property that
k
d X l=1

c l @l K2t k2 = k

d X l=1

bc l @l St Kt k2

c c000 t;(1+ )=2 kKtk2 c t;(1+ )=2 t;d=4

32

uniformly for all t 1 and for a suitable c > 0. Then by scaling


k
d X (t) c l @l K2 k2 l=1

c t; =2

for all t 1, where

Pd (t) c (t) 1= 2 l (x) = l (t x). In particular, limt!1 k l=1 l @l K2 k2 = 0. But d d X (t) c 2 X Z 1=2 c c 1=2 lim k @lK2k2 = tlim t!1 l=1 l !1 l m=1 Rd dx m(t x) l(t x) (@mK2)(x) (@lK2)(x) Z d X Z c c = du m (u) l(u) d dx (@mK2)(x) (@lK2)(x) Id R l m=1 Z d X 2 c = d d du dx l (u) (@lK2 )(x) I R l=1

(t because m) l(t) converges in the weak topology on L1(Rd) to its average value over c Id (see ZKO] page 5). So Pd=1 l(u) (@lK2 )(x) = 0 for almost P (u x) 2 Id Rd. every l c2 are continuous by Lemma 4.2 this implies that d=1 l (u) (@lK2 )(x) = 0 c Since l and K l d Rd . Arguing as at the end of the proof of Proposition 3.12 gives for all (u x) 2 I (I ; U (h))(@k l) = l = 0 for all l 2 f1 : : : dg. This is valid for all k 2 f1 : : : dg and h 2 Rd with jhj 1. But then @k l is constant. On the other hand, @k l 2 L? (Id), so 2 @k l = 0. But then l is constant and since l 2 L? (Id) one has l = 0 for all l 2 f1 : : : dg. 2 Then Condition I follows from Lemma 3.11.I. Before we can prove the remaining implications of Theorem 1.3 we need some preparation.

Lemma 4.4 If div C = 0 then L2 2(Rd) D(H ) and


H' = ;
for all ' 2 L2 2(Rd ).
d X k l=1

ckl @k @l'

Proof If 2 L2 1(Rd) and ' 2 L2 2(Rd) then


h( ') =
with for all
d X k l=1

(@k ckl@l') = ; )= )=
d X k=1

d X k l=1

( ckl@k @l') +

d X l=1

r( @l')

r( r(

(@k ckl ) + ( ckl@k )

2 L2 1(Rd ). But then

d XZ k=1 Rd

dx @k ( ) (x) ckl(x)

and since div C = 0 one has r(

) = 0. 33

Since H can be expressed in non-divergent form, and since the coe cients are Holder continuous, it follows that the kernel K is at least twice-di erentiable in the rst variable and the derivatives satisfy Gaussian bounds with an exponential growth in time (see the remark on page 255 in Fri]). Explicitly, there are b c > 0 and ! 0 such that j(@k @l Kt )(x y )j c t;1 e!t Gb t (x ; y ) (37) and (38) j(@k @l Kt )(x ; h y ) ; (@k @lKt )(x y )j c t;1 (jhjt;1=2) =2 e!t Gb t (x ; y ) for all h x y 2 Rd, k l 2 f1 : : : dg and t > 0 with jhj t1=2. This allows us to extend the conclusions of Lemmas 3.1 and 4.1. Lemma 4.5 For all 2 B C the projection P0( ) is bounded from L1(Id) into L1 2(Id) 0 and @k @l P0 ( ) is bounded from L1 (Id ) into C =2 for all k l 2 f1 : : : dg. Moreover, the map 7! @k @l P0 ( ) is holomorphic from B C into L(L1 (Id) L1 (Id)) and from B C into 0 0 L(L1 (Id ) C =2 ) for all k l 2 f1 : : : dg. Proof Using (9) and (37) one sees that K is twice-di erentiable and there are b c ! > 0 such that j(@k @l Kt )(u v )j c (j j2 + t;1 )(1 ^ t);d=2 e;bku;vk2 t;1 e!(1+(Im )2 )t for all 2 Cd , t > 0 and u v 2 Id. Arguing as in the proof of Lemma 3.1 the L1 2statements of the lemma follow. The Holder bounds follow from (38). 2 Next we examine the second-order terms in the holomorphic expansion of 7! P0( ). In general this is complicated but simpli cations occur because we are assuming div C = 0. Hence the correctors m = 0 by Lemma 3.11. As div C = 0 it also follows that cmn is the ^ average of cmn over Id and one can de ne ;1 ;1 c ^ 1 mn = 2 H0 (cmn + cnm ) ; (^mn + cnm )1 ; 2 for all m n 2 f1 : : : dg, where H0 1: L? (Id) ! L? (Id) is as in Subsection 2.3. Then it 2 follows from the proof of Proposition 3.3 that one can identify the second-order term in the expansion of the eigenvector ' as '(m n) = ; mn + amn for some amn 2 C. The second-order term 'ym n) in the expansion of 'y has a more complicated form which we ( will not require. But one veri es that ('y ; ('y 1 ) 1 ' ) ; 1 = O(j j3) and ( P0(mn)(0)') = ;( mn)(1 ') + ( 1 )('ym n) ; ('y 1 ) 1 ') 1 ( for all ' 2 L2(Id), where P0(mn)( ) = @ 2P0( )=(@ m @ n). As in Corollary 3.4 the constant amn gives no contribution. Then it follows that (@k @l P0(mn)(0)') = ;(@k @l mn)(1 ') 1 for 2 L2 2(Id ) \ L1 2(Id) and ' 2 L2 and 'ym n) also plays no role. Since 7! @k @lP0( ) is ( C into L(L1 (Id ) L1 (Id )) and from B C into L(L1 (Id ) C =2 ) it follows holomorphic from B 0 0 that mn 2 L1 2 and @k @l mn 2 C =2. Moreover, @k @lP0(mn)(0) = ;(@k @l mn)P0(0) 34

for all k l m n 2 f1 : : : dg. We extend the mn to periodic functions on C (Rd).

Proposition 4.6 If div C = 0 then there is a c > 0 such that


c j(@k @lKt )(x y ) ; (@k @l Kt )(x ; y ) +
d X m n=1

(@k @l

c mn )(x) (@m @n Kt )(x ; y )j

c t;3=2 t;d=2

uniformly for all k l 2 f1 : : : dg, x y 2 Rd and t 1. Proof Fix k l 2 f1 : : : dg. Then with = 2;1 ( 0 ^ ^0) one has bb @k @lStP0(B ) ; @k @lStP0(B ) ;1

= (2 );d

for all t > 0. If 2 B then b (i k + @k )(i l + @l)St P0( ) + k l St P0(0) = e; 0( )t ; k l P0( ) + i k @lP0( ) + i l @k P0( ) + @k @lP0( ) + e;^0( )t =
m n=1 d X ;^ 0 ( )t (mn) ; ( )t ;^ ( )t e m n @k @l P0 (0) ; k l(e 0 P0 ( ) ; e 0 P0 (0))

d (i k + @k )(i l + @l)St P0( ) +

b k l St P0 (0)

k l P0 (0)

+ i e;

( )t

k @lP0 (

) + i e;

( )t

l@k P0 (

) + (e; 0( )t ; e;^0( )t)@k @lP0( )

d ;^0( )t @k @lP0( ) ; X +e

m n=1

(mn) m n @k @l P0 (0)

for all t > 0. But @k P0(m)(0) = 0 for all k m 2 f1 : : : dg by Lemmas 3.11.I and 4.2 since div C = 0. So k@k P0( )k1!1 = O(j j2). Then also @k @lP0(m)(0) = 0 for all k l m 2 f1 : : : dg and therefore k@k @l P0 ( )k1!1 = O(j j2 ). Moreover,
k@k @l P0 ( ) ;
d X m n=1
(mn) m n @k @l P0 (0)k1!1

= O(j j3) :

So only the rst term gives a big contribution. Since @k @lP0(mn)(0) = ;(@k @l mn) P0(0) this b contribution comes from the kernel of the operator ;(@k @l mn) @m @nSt and the proposition follows by the earlier arguments. 2 Now we can complete the proof of the equivalence of the rst four conditions of Theorem 1.3. Proof of Theorem 1.3.I)III We rst show that one can replace the exponential factor e!t in the bounds (37) by a factor (1 + t). It follows from the semigroup property, the estimates (37) and the Gaussian bounds on the kernel that Z j(@k @l Kt )(x y )j c k dz Gb 1 (x ; z ) Gb t;1 (z ; y ) :
R

35

Hence

uniformly for all t c Since similar bounds are also valid for the kernel K there exist b c > 0 such that
c j(@k @l Kt )(x y );(@k @l Kt )(x;y )+
d X m n=1

j(@k @lKt )(x y )j c0 Gb0 t (x ; y ) 1 and x y 2 Rd .

(@k @l

c mn )(x) (@m@n Kt )(x;y )j

c (1+t) t;1 Gb t(x;y)

uniformly for all t > 0 and x y 2 Rd. Then by interpolation with the bounds of Proposition 4.6 one deduces that for all " 2 h0 1i there exist b c > 0 such that
c j(@k @l Kt )(x y ) ; (@k @lKt )(x ; y ) +
d X m n=1

(@k @l

c mn )(x) (@m @n Kt )(x ; y )j

c t;"=2 t;1 Gb t(x ; y)

(39)

uniformly for all t 1 and x y 2 Rd . This immediately implies Condition III. Proof of Theorem 1.3.I)V The proof is similar to the proof of the boundedness of the rst-order Riesz transforms in Theorem 1.2. Fix p 2 h1 1i. It follows from Lemma 4.4 that H can be expressed in non-divergence form. But then D(H ) = Lp 2(Rd) and one has estimates max k@ @ 'k c (kH'kp + k'kp) 1 kl d k l p

for all ' 2 Lp 2(Rd). (These estimates follow from similar estimates for operators with constant coe cients, the uniform continuity of the coe cients and a partition of the identity, by standard arguments.) Therefore the operators Rk l " = @k @l("I + H );1 are bounded on Lp(Rd) for each " > 0 and again it is necessary to obtain a norm bound uniform for small ". But Z1 Rk l " = @k @l dt e;"t St = R(1) " + R(2) " kl kl with
0

R(1) " kl
and

= @k @l

Z1
0

dt e;"t(2e;t ; e;2t)St = 2@k @l((1 + ")I + H );1 ; @k @l((2 + ")I + H );1 R(2) " = kl
Z1
0

dt e;"t(1 ; e;t)2@k @lSt :

Then kR(1) "kp!p is bounded uniformly in " because H has a bounded H1 -holomorphic kl functional calculus. Moreover, (39) together with the bounds of III, which are a consequence of I, give bounds
b k@k @l St ; @k @l St +
d X m n=1

(@k @l

b mn ) @m @n St kp!p

c t;1 (1 + t); =2

for some 2 h0 1i and all t > 0. Therefore one estimates easily that kR(2) "kp!p is also kl bounded uniformly in ". 36

It follows from these estimates that max k@ @ 'k c k("I + H )'kp 1 kl d k l p for all ' 2 D(H ) with c independent of ". Hence taking the limit " ! 0 one obtains max k@ @ 'k c kH'kp 1 kl d k l p for all ' 2 D(H ). But the complementary bounds c0 kH'kp 1maxd k@k @l'kp kl are elementary since H is in non-divergence form. Proof of Theorem 1.3.V)IV Suppose Condition V is valid and p 2 h1 1i. Then there is a c > 0 such that k@k @l'kp c kH'kp for all k l 2 f1 : : : dg and ' 2 Lp 2(Rd). But the operator H has a bounded H1 -functional calculus on Lp(Rd). Hence there is a c0 > 0 such that kHStkp!p c0 t;1 for all t > 0. Then k@k @lSt'kp c c0t;1 k'kp and Condition IV follows. For derivatives higher than second-order on the kernel, or the semigroup, the canonical small time behaviour cannot be expected to extend to large times except in the case that all coe cients are constant and derivatives of all orders are well-behaved. Proof of Theorem 1.4 The proof of the implication I)II is well known. \II)III". The proof is the similar to the proof of Theorem 1.3.II)IV although one more interpolation argument is necessary. It is a consequence of the Duhamel formula that Z ;1 '(x ; sek ) ; '(x) ; s;1 s du (@k ')(x ; uek) ; (@k ')(x) (@k ')(x) = ;s for any continuously di erentiable function ', all x 2 Rd, s > 0 and k 2 f1 : : : dg, where ek is the unit vector in the k-th direction. Setting '(x) = (@lKt)(x y) it follows from the bounds (32) and the bounds of Condition II that there are b c > 0 such that Z ;1 t;1=2 Gb t (x ; y) + c s;1 s du t;1(ut;1=2) t;d=2 j(@k @lKt )(x y )j c s 2c t;(d+1)=2e;2;1 bjx;yj2 t;1 uniformly for all t 1 and x y 2 Rd, if one chooses s = e;2;1bjx;yj2 t;1 1. Hence there are b0 c0 > 0 such that j(@k @l Kt )(x ; h y ) ; (@k @l Kt )(x y )j c0 (jhjt;1=2) Gb0 t (x ; y ) uniformly for all t 1 and h x y 2 Rd with jhj 1. Interpolation with the bounds of Condition II gives bounds j(@k @l Kt )(x ; h y ) ; (@k @lKt )(x y )j c00 t;(1;") (jhjt;1=2) G"b0 t (x ; y ) c00 t;1(jhjt;1=2) ;2" G"b0 t(x ; y) 37 = c t;(d+1)=2 s;1 e;bjx;yj t;1
2

+ (1 + );1 s t;(1+ )=2

uniformly for all " 2 h0 =2i, t 1 and h x y 2 Rd with jhj 1. Hence III is valid. \III)I". If Condition III is valid then k(I ;U (h))(I ;U (g))@lStkp!p cp t;(2+ )=2 jhj jgj for all g h 2 Rd, t 1 and l 2 f1 : : : dg with jhj 1. Arguing as in the proof of Theorem 1.3.IV)I it follows that (I ; U (h))(I ; U (g))@k l = 0 for all k l 2 f1 : : : dg and g h 2 Rd with jhj 1. Hence l = 0 for all l and div C = 0. Then one can apply Proposition 4.6 and argue as in the proof of the implication in Theorem 1.3.IV)I. One deduces that Condition I is valid.

Acknowledgements
The rst author is indebted to the Centre of Mathematics and its Applications at the ANU for hospitality and support whilst this work was carried out. All three authors would like to thank Nick Dungey for simplifying the proof of Theorem 1.3.

References
ADM] Albrecht, D., Duong, X. and McIntosh, A., Operator theory and harmonic analysis. In Instructional Workshop on Analysis and Geometry, Part III, vol. 34 of Proceedings of the Centre for Mathematics and its Applications. Australian National University, Canberra, 1996, 77{136. Aus] Auscher, P., Regularity theorems and heat kernel for elliptic operators. J. London Math. Soc. 54 (1996), 284{296. ACT] Auscher, P., Coulhon, T. and Tchamitchian, P., Absence de principe du maximum pour certaines equations paraboliques complexes. Colloq. Math. 71 (1996), 87{95. AMT] Auscher, P., McIntosh, A. and Tchamitchian, P., Heat kernels of second order complex elliptic operators and their applications. J. Funct. Anal. 152 (1998), 22{73. BBJR] Batty, C.J.K., Bratteli, O., J rgensen, P.E.T. and Robinson, D.W., Asymptotics of periodic subelliptic operators. J. Geom. Anal. 5 (1995), 427{443. Bau] Bauman, P., Equivalence of the Green's functions for di usion operators in Rn: a counterexample. Proc. Amer. Math. Soc. 91 (1984), 64{68. BLP] Bensoussan, A., Lions, J.L. and Papanicolaou, G., Asymptotic analysis for periodic structures, vol. 5 of Studies in Mathematics and its Applications. North-Holland, Amsterdam etc., 1978. Blo] Bloch, F., Uber die Quantenmechanik der Elektronen in Kristallgittern. Zeit. fur Physik 52 (1928), 555{600. BJR] Bratteli, O., J rgensen, P.E.T. and Robinson, D.W., Spectral asymptotics of periodic elliptic operators. Math. Z. (1999). To appear. 38

BrR] Bratteli, O. and Robinson, D.W., Operator algebras and quantum statistical mechanics, vol. 1. Second edition. Springer-Verlag, New York etc., 1987. Dau] Daubechies, I., Ten lectures on wavelets. CBMS Regional Conference Series in Mathematics 61. Amer. Math. Soc., Providence, 1992. Dav] Davies, E.B., Uniformly elliptic operators with measurable coe cients. J. Funct. Anal. 132 (1995), 141{169. Dix] Dixmier, J., Les algebres d'operateurs dans l'espace Hilbertien. Cahiers scienti ques 25. Gauthier-Villars, Paris, 1969. DuR] Duong, X.T. and Robinson, D.W., Semigroup kernels, Poisson bounds, and holomorphic functional calculus. J. Funct. Anal. 142 (1996), 89{129. ElR1] Elst, A.F.M. ter and Robinson, D.W., Second-order strongly elliptic operators on Lie groups with Holder continuous coe cients. J. Austr. Math. Soc. (Series A) 63 (1997), 297{363. ElR2] , Second-order subelliptic operators on Lie groups III: Holder continuous coe cients. Calc. Var. Partial Di erential Equations (1999). To appear. ERS] Elst, A.F.M. ter, Robinson, D.W. and Sikora, A., Riesz transforms and Lie groups of polynomial growth. J. Funct. Anal. 162 (1999), 14{51. FaK] Fabes, E. B. and Kenig, C. E., Examples of singular parabolic measures and singular transition probability densities. Duke Math. J. 48 (1981), 845{856. Fri] Friedman, A., Partial di erential equations of parabolic type. Prentice-Hall, Inc., Englewood Cli s, N.J., 1964. Gia] Giaquinta, M., Multiple integrals in the calculus of variations and nonlinear elliptic systems. Annals of Mathematics Studies 105. Princeton University Press, Princeton, 1983. Gio] Giorgi, E.D., Un esempio di estremali discontinue per un problema variazionale di tipo ellittico. Boll. Un. Mat. Ital. (4) 1 (1968), 135{137. Kat] Kato, T., Perturbation theory for linear operators. Second edition, Grundlehren der mathematischen Wissenschaften 132. Springer-Verlag, Berlin etc., 1984. MNP] Maz'ya, V.G., Nazarov, S.A. and Plamenevsk i, B.A., Absence of the De Giorgi-type theorems for strongly elliptic equations with complex coe cients. J. Math. Sov. 28 (1985), 726{739. MoM] Modica, L. and Mortola, S., Construction of a singular elliptic-harmonic measure. Manuscripta Math. 33 (1980), 81{98. MMS] Modica, L., Mortola, S. and Salsa, S., A nonvariational second order elliptic operator with singular elliptic measure. Proc. Amer. Math. Soc. 84 (1982), 225{ 230. 39

Rob] Robinson, D.W., Elliptic operators and Lie groups. Oxford Mathematical Monographs. Oxford University Press, Oxford etc., 1991. ReS] Reed, M. and Simon, B., Methods of modern mathematical physics IV. Analysis of operators. Academic Press, New York etc., 1978. ScD] Schempp, W. and Dreseler, B., Einfuhrung in die harmonische Analyse. Mathematische Leitfaden. Teubner, Stuttgart, 1980. Zhi] Zhikov, V.V., Spectral approach to asymptotic problems in di usion. Di . Equations 25 (1989), 33{39. ZKO] Zhikov, V.V., Kozlov, S.M. and Oleinik, O.A., Homogenization of di erential operators and integral functions. Springer-Verlag, Berlin, 1994.

40

You might also like