Research

Carleman estimates and unique continuation property for abstract elliptic equations

Veli B Shakhmurov

Author affiliations

Department of Electronics Engineering and Communication, Okan University, Akfirat Beldesi, Tuzla, 34959, Istanbul, Turkey

Boundary Value Problems 2012, 2012:46  doi:10.1186/1687-2770-2012-46

 Received: 26 January 2012 Accepted: 23 April 2012 Published: 23 April 2012

This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

The unique continuation theorems for elliptic differential-operator equations with variable coefficients in vector-valued Lp-space are investigated. The operator-valued multiplier theorems, maximal regularity properties and the Carleman estimates for the equations are employed to obtain these results. In applications the unique continuation theorems for quasielliptic partial differential equations and finite or infinite systems of elliptic equations are studied.

AMS: 34G10; 35B45; 35B60.

Keywords:
Carleman estimates; unique continuation; embedding theorems; Banach-valued function spaces; differential operator equations; maximal Lp-regularity; operator-valued Fourier multipliers; interpolation of Banach spaces

1 Introduction

The aim of this article, is to present a unique continuation result for solutions of a differential inequalities of the form:

P ( x , D ) u ( x ) E V ( x ) u ( x ) E , (1)

where

P ( x ; D ) u = i , j = 1 n a i j 2 u x i x j + A u + k = 1 n A k u x k ,

here aij are real numbers, A = A (x), Ak = Ak (x) and V (x) are the possible linear operators in a Banach space E.

Jerison and Kenig started the theory of Lp Carleman estimates for Laplace operator with potential and proved unique continuation results for elliptic constant coefficient operators in [1]. This result shows that the condition V Ln/2,loc is in the best possible nature. The uniform Sobolev inequalities and unique continuation results for second-order elliptic equations with constant coefficients studied in [2]. This was latter generalized to elliptic variable coefficient operators by Sogge in [3]. There were further improvement by Wolff [4] for elliptic operators with less regular coefficients and by Koch and Tataru [5] who considered the problem with gradients terms. A comprehensive introductions and historical references to Carleman estimates and unique continuation properties may be found, e.g., in [5]. Moreover, boundary value problems for differential-operator equations (DOEs) have been studied extensively by many researchers (see [6-18] and the references therein).

In this article, the unique continuation theorems for elliptic equations with variable operator coefficients in E-valued Lp spaces are studied. We will prove that if n 1 p - 1 p 2 , 1 μ = 1 p - 1 p , 1 p + 1 p | = 1 , V Lμ (Rn; L(E)), p, μ∈ (1, ∞) and u W p 2 ( R n ; E ( A ) , E ) satisfies (1), then u is identically zero if it vanishes in a nonempty open subset, where W p 2 ( R n ; E ( A ) , E ) is an E-valued Sobolev-Lions type space. We prove the Carleman estimates to obtain unique continuation. Specifically, we shall see that it suffices to show that if w ( x ) = x 1 + x 1 2 2 , then

e t w u L p | ( R n ; E ) C e t w L ( ε x , D ) u L p ( R n ; E ) , 1 p + 1 p | = 1 , | α | 1 t ( 1 + 1 n | α | ) e t w D α u L p ( R n ; E ) + e t w A u L p ( R n ; E ) C e t w L ( ε x , D ) u L p ( R n ; E ) .

In the Hilbert space L2 (Rn; H), we derive the following Carleman estimate

| α | 2 t 3 2 | α | e t w D α u L 2 ( R n ; H ) + e t w A u L 2 ( R n ; H ) C e t w L 0 u L 2 ( R n ; H ) .

Any of these inequalities would follow from showing that the adjoint operator Lt (x; D) = etwL (x; D) e-tw satisfies the following relevant local Sobolev inequalities

u L p | ( R n ; E ) C L t u L p ( R n ; E ) , 1 p + 1 p | = 1 , | α | 1 t ( 1 + 1 n | α | ) D α u L p ( R n ; E ) A u L p ( R n ; E ) C L t u L p ( R n ; E ) ,

uniformly to t, where L0t = etwL0e-tw. In application, putting concrete Banach spaces instead of E and concrete operators instead of A, we obtain different results concerning to Carleman estimates and unique continuation.

2 Notations, definitions, and background

Let R and C denote the sets of real and complex numbers, respectively. Let

S φ = { ξ C , | arg ξ | φ } { 0 } , φ [ 0 , π ) .

Let E and E1 be two Banach spaces, and L (E, E1) denotes the spaces of all bounded linear operators from E to E1. For E1 = E we denote L (E, E1) by L (E). A linear operator A is said to be a φ-positive in a Banach space E with bound M > 0 if D (A) is dense on E and

( A + ξ I ) 1 L ( E ) M ( 1 + | ξ | ) 1

with λ Sφ, φ ∈ (0, π], I is identity operator in E. We will sometimes use A + ξ or Aξ instead of A + ξI for a scalar ξ and (A + ξI)-1 denotes the inverse of the operator A + ξI or the resolvent of operator A. It is known [19, §1.15.1] that there exist fractional powers Aθ of a positive operator A and

E ( A θ ) = { u D ( A θ ) , u E ( A θ ) = A θ u E + u < , - < θ < } .

We denote by Lp (Ω; E) the space of all strongly measurable E-valued functions on Ω with the norm

u L p = u L p ( Ω ; E ) = Ω u ( x ) E p d x 1 / p , 1 p < .

By Lp,q (Ω) and W p , q l ( Ω ) let us denoted, respectively, the (p, q)-integrable function space and Sobolev space with mixed norms, where 1 ≤ p, q < ∞, see [20].

Let E0 and E be two Banach spaces and E0 is continuously and densely embedded E.

Let l be a positive integer.

We introduce an E-valued function space W p l ( Ω ; E 0 , E ) (sometimes we called it Sobolev-Lions type space) that consist of all functions u Lp (Ω; E0) such that the generalized derivatives D k l u = l u x k l L p ( Ω ; E ) are endowed with the

u W p l ( Ω ; E 0 , E ) = u L p ( Ω ; E 0 ) + k = 1 n D k l u L p ( Ω ; E ) < , 1 p < .

The Banach space E is called an UMD-space if the Hilbert operator ( H f ) ( x ) = lim ε 0 | x - y | > ε f ( y ) x - y dy is bounded in Lp (R, E), p ∈ (1, ) (see e.g., [21,22]). UMD spaces include, e.g., Lp, lp spaces and Lorentz spaces Lpq, p, q ∈ (1, ).

Let E1 and E2 be two Banach spaces. Let S (Rn; E) denotes a Schwartz class, i.e., the space of all E-valued rapidly decreasing smooth functions on Rn. Let F and F-1denote Fourier and inverse Fourier transformations, respectively. A function Ψ ∈ Cm (Rn; L (E1, E2)) is called a multiplier from Lp (Rn; E1) to Lq (Rn; E2) for p, q ∈ (1, ) if the map u → Ku = F-1 Ψ (ξ) Fu, u S (Rn; E1) is well defined and extends to a bounded linear operator

K : L p ( R n ; E 1 ) L q ( R n ; E 2 ) .

We denote the set of all multipliers from Lp (Rn; E1) to Lq (Rn; E2) by M p q ( E 1 , E 2 ) . For E1 = E2 = E and q = p we denote M p q ( E 1 , E 2 ) by Mp (E). The Lp-multipliers of the Fourier transformation, and some related references, can be found in [19, § 2.2.1-§ 2.2.4]. On the other hand, Fourier multipliers in vector-valued function spaces, have been studied, e.g., in [23-28].

A set K L (E1, E2) is called R-bounded [22,23] if there is a constant C such that for all T1, T2, . . . , Tm K and u1,u2, . . . , um E1, m N

0 1 j = 1 m r j ( y ) T j u j E 2 d y C 0 1 j = 1 m r j ( y ) u j E 1 d y ,

where {rj} is a sequence of independent symmetric {-1, 1}-valued random variables on [0,1]. The smallest C for which the above estimate holds is called a R-bound of the collection K and denoted by R (K).

Let

U n = { β = ( β 1 , β 2 , , β n ) , β i { 0 , 1 } , i = 1 , 2 , , n } , ξ β = ξ 1 β 1 ξ 2 β 2 ξ n β n , | ξ β | = | ξ 1 | β 1 | ξ 2 | β 2 | ξ n | β n .

For any r = (r1, r2, . . . , rn), ri ∈ [0, ) the function ()r, ξ Rn will be defined such that

( i ξ ) r = ( i ξ 1 ) r 1 ( i ξ n ) r n , ξ 1 , ξ 2 , , ξ n 0 , 0 , ξ 1 , ξ 2 , , ξ n = 0 ,

where

( i t ) ν = | t | ν exp i π 2 sign  t , t ( - , ) , ν [ 0 , ) .

Definition 2.1. The Banach space E is said to be a space satisfying a multiplier condition with respect to p, q ∈ (1, ) (with respect to p if q = p) when for Ψ ∈ C(n) (Rn; L (E1, E2)) if the set

ξ | β | + 1 p - 1 q D β Ψ ( ξ ) : ξ R n \ 0 , β U n

is R-bounded, then Ψ M p q ( E 1 , E 2 ) .

Definition 2.2. The φ-positive operator A is said to be a R-positive in a Banach space E if there exists φ ∈ [0, π) such that the set

L A = { ξ ( A + ξ I ) - 1 : ξ S φ }

is R-bounded.

Remark 2.1. By virtue of [29] or [30] UMD spaces satisfy the multiplier condition with respect to p ∈ (1, ).

Note that, in Hilbert spaces every norm bounded set is R-bounded. Therefore, in Hilbert spaces all positive operators are R-positive. If A is a generator of a contraction semigroup on Lq, 1 ≤ q ≤ ∞ [31], A has the bounded imaginary powers with ( A i t ) L ( E ) C e ν | t | , ν < π 2 or if A is a generator of a semigroup with Gaussian bound in E ∈ UMD then those operators are R-positive (see e.g., [24]).

It is well known (see e.g., [32]) that any Hilbert space satisfies the multiplier condition with respect to p ∈ (1, ). By virtue of [33] Mikhlin conditions are not sufficient for operator-valued multiplier theorem. There are however, Banach spaces which are not Hilbert spaces but satisfy the multiplier condition (see Remark 2.1).

Let H k = { Ψ h M p q ( E 1 , E 2 ) , h = ( h 1 , h 2 , , h n ) K } be a collection of multipliers in M p q ( E 1 , E 2 ) . We say that Hk is a uniform collection of multipliers if there exists a constant M > 0, independent on h K, such that

F - 1 Ψ h F u L q ( R n ; E 2 ) M u L p ( R n ; E 1 )

for all h K and u S (Rn; E1).

We set

C b ( Ω ; E ) = u C ( Ω ; E ) , lim | x | u ( x )  exists .

In view of [17, Theorem A0], we have

Theorem 2.0. Let E1 and E2 be two UMD spaces and let

Ψ C ( n ) ( R n \ 0 ; L ( E 1 , E 2 ) ) for  p , q ( 1 , ) .

If

R ξ | β | + 1 p - 1 q D ξ β Ψ h ( ξ ) : ξ R n \ 0 , β U n K β <

uniformly with respect to h K then Ψh (ξ) is a uniformly collection of multipliers from Lp (Rn; E1) to Lq (Rn; E2).

Let

χ = | α | + n 1 p - 1 q l , α = ( α 1 , α 2 , , α n ) .

Embedding theorems in Sobolev-Lions type spaces were studied in [13-18,32,34]. In a similar way as [17, Theorem 3] we have

Theorem 2.1. Suppose the following conditions hold:

(1) E is a Banach space satisfying the multiplier condition with respect to p, q ∈ (1, ) and A is a R-positive operator on E;

(2) l is a positive and αk are nonnegative integer numbers such that 0 ≤ μ ≤ 1 - ϰ, t and h are positive parameters.

Then the embedding

D α W p l ( R n ; E ( A ) , E ) L q ( R n ; E ( A 1 - χ - μ ) )

is continuous and there exists a positive constant Cµ such that for

u W p l ( R n ; E ( A ) , E )

the uniform estimate holds

D α u L q ( R n ; E ( A 1 - χ - μ ) ) C μ h μ u W p l ( R n ; E ( A ) , E ) + h - ( 1 - μ ) u L p ( R n ; E ) .

Moreover, for u W p l ( R n ; E ( A ) , E ) the following uniform estimate holds

A 1 - χ - μ u L p ( R n ; E ) C μ h μ u W p l ( R n ; E ( A ) , E ) + h - ( 1 - μ ) u L p ( R n ; E ) .

3 Carleman estimates for DOE

Consider at first the equation with constant coefficients

L 0 u = k = 1 n D k 2 u + A u = f ( x ) , (2)

where D k = i k and A is the possible unbounded operator in a Banach space E.

Let w ( x ) = x 1 + x 1 2 2 and t is a positive parameter.

Remark 3.1. It is clear to see that

e t w L 0 [ e - t w u ] = L 0 t ( x , D ) u = e t w k = 1 n D k 2 ( e - t w u ) + e - t w A u = k = 1 n D k 2 u + A u + 2 t w 1 u x 1 + [ - t 2 w 1 2 + t ] u , (3)

where w 1 = w x 1 . Let L0t (x, ξ) is the principal operator symbol of L0t (x, D) on the domain B0, i.e.,

L 0 t ( x , ξ ) = ξ 1 2 - 2 i ξ 1 w 1 t + A + | ξ | | 2 - t 2 w 1 2 = G t ( x , ξ ) B t ( x , ξ ) ,

where

G t ( x , ξ ) = ξ 1 - i A + | ξ | | 2 1 2 + t w 1 , B t ( x , ξ ) = ξ 1 + i A + | ξ | | 2 1 2 - t w 1 , | ξ | | 2 = k = 2 n ξ k 2 .

Our main aim is to show the following result:

Remark 3.2. Since Q(ξ) ∈ S (φ) for all φ ∈ [0, π) due to positivity of A, the operator function A + ||2, ξ Rn is uniformly positive in E. So there are fractional powers of A+||2 and the operator function A + | ξ | | 2 1 2 is positive in E (see e.g., [19, §1. 15.1]).

First, we will prove the following result.

Theorem 3.1. Suppose A is a positive operator in a Hilbert space H. Then the following uniform Sobolev type estimate holds for the solution of Equation (3)

| α | 2 t 3 2 | α | e t w D α u L 2 ( R n ; H ) + e t w A u L 2 ( R n ; H ) C e t w L 0 u L 2 ( R n ; H ) . (4)

By virtue of Remark 3.1 it suffices to prove the following uniform coercive estimate

| α | 2 t 3 2 | α | D α u L 2 ( R n ; H ) + A u L 2 ( R n ; H ) C L 0 t u L 2 ( R n ; H ) (5)

for u W 2 2 ( R n ; H ( A ) , E ) .

To prove the Theorem 3.1, we shall show that L0t (x, D) has a right parametrix T, with the following properties.

Lemma 3.1. For t > 0 there are functions K = Kt and R = Rt so that

L 0 t ( x , D ) K ( x , y ) = δ ( x - y ) + R ( x , y ) , x , y B 0 , (6)

where δ denotes the Dirac distribution. Moreover, if we let T = Tt be the operator with kernel K, i.e.,

T f ( x ) = B 0 K ( x , y ) f ( y ) d y , f C 0 ( B 0 ; E ) ,

and R is the operator with kernel R (x, y), then for large t > 0, the adjoint of these operators satisfy the following estimates

| α | 2 t 2 | α | D α T * f L 2 ( B 0 ; H ) C f L 2 ( B 0 ; H ) , A T * f L 2 ( B 0 ; H ) C f L 2 ( B 0 ; H ) , (7)

t 1 2 R * f L 2 ( B 0 ; H ) C f L 2 ( B 0 ; H ) , (8)

t 1 2 D ν R * f L 2 ( B 0 ; H ) C | α | | ν | 1 D α f L 2 ( B 0 ; H ) , 1 | ν | 2. (9)

Proof. By Remark 3.2 the operator function A + | ξ | | 2 1 2 is positive in E for all ξ Rn. Since tw1 + 1 S(φ), due to positivity of A, for φ [ π 2 , π ) the factor G t ( x , ξ ) = - i A + | ξ | | 2 1 2 + w 1 t + i ξ 1 has a bounded inverse G t - 1 ( x , ξ ) for all ξ Rn, t > 0 and

G t 1 ( x , ξ ) B ( H ) C ( 1 + | t w 1 + i ξ 1 | ) 1 . (10)

Therefore, we call Gt (x, ξ) the regular factor. Consider now the second factor

B t ( x , ξ ) = i A + | ξ | | 2 1 2 - ( w 1 t + i ξ 1 ) .

By virtue of operator calculus and fractional powers of positive operators (see e.g., [19, §1.15.1] or [35]) we get that - [tw1 + 1] ∉ S (φ) for ξ1 = 0 and tw1 = ||, i.e., the operator Bt (x, ξ) does not has an inverse, in the following set

Δ t = { ( x , ξ ) B 0 × R n : ξ 1 = 0 , | ξ | | = t w 1 } .

So we will called Bt the singular factor and the set Δt call singular set for the operator function Bt. The operator B t - 1 cannot be bounded in the set Δt. Nevertheless, the operator B t - 1 , and hence L 0 t - 1 , can be bounded when (x, ξ) is sufficiently far from Δt. For instance, if we define

Γ t = ( x , ξ ) B 0 × R n : | ξ | | t 4 , 4 t , | ξ 1 | t 4 ,

by properties of positive operators we will get the same estimate of type (10) for the singular factor Bt. Hence, using this fact and the resolvent properties of positive operators we obtain the following estimate

L 0 t 1 ( x , ξ ) B ( E ) C ( 1 + | ξ | 2 + t 2 ) 1 when  ( x , ξ ) c Γ t , (11)

where the constant C is independent of x, ξ, t and cΓt denotes the complement of Γt.

Let β C 0 ( R ) such that, β(ξ) = 0 if | ξ | 1 4 , 4 and β (ξ) = 0 near the origin. We then define

β 0 ( ξ ) = β 0 t ( ξ ) β 0 ( ξ ) = 1 - β ( | ξ | | / t ) β ( 1 - ξ 1 / t )

and notice that β0 (ξ) = 0 on Γt. Hence, if we define

K 0 ( x , y ) = ( 2 π ) - n R n β 0 ( ξ ) e i ( ( x - y ) , ξ ) L 0 t - 1 ( y , ξ ) d ξ (12)

and recall (11), then by [31] it follows from standard microlocal arguments that

L 0 t ( x , D ) K 0 ( x , y ) = ( 2 π ) - n R n β 0 ( ξ ) e i ( ( x - y ) , ξ ) d ξ + R 0 t ( x , y ) ,

where R0t belongs to a bounded subset of S-1 which is independent of t. Since operator R 0 t * also has the same property, it follows that for all f C 0 ( B 0 ; H )

D ν R 0 t * f L 2 ( B 0 ; H ) C | α | | ν | 1 D α f L 2 ( B 0 ; H ) , 1 | ν | 2.

By reasoning as in [31] we get that tR0t belongs to a bounded subset of S0. So, we have the following estimate

t D ν R 0 t * f L 2 ( B 0 ; H ) C f L 2 ( B 0 ; H ) .

Moreover, the Remark 3.2, positivity properties of A and, (11) and (12) imply that, the operator functions | α | 2 β 0 ( ξ ) t 2 - | α | ξ α L 0 t - 1 ( x , ξ ) and β 0 ( ξ ) A L 0 t - 1 ( x , ξ ) are uniformly bounded. Then, if we let T0 be the operator with kernel K0 (x, y), by using the Minkowski integral inequality and Plancherel's theorem we obtain

| α | 2 t 2 | α | D α T 0 f L 2 ( B 0 ; H ) C f L 2 ( B 0 ; H ) , A T 0 f L 2 ( B 0 ; H ) C f L 2 ( B 0 ; H ) .

For inverting L0t (x, D) on the set Γt we will require the use of Fourier integrals with complex phase. Let β1 (ξ) = 1 - β0 (ξ). We will construct a Fourier integral operator T1 with kernel

K 1 ( x , y ) = ( 2 π ) - n R n β 1 ( ξ ) e i Φ ( x , y , ξ ) L 0 t - 1 ( y , ξ ) d ξ (13)

so that the analogs of (16) and the estimates (7)-(9) are satisfied. Since G t - 1 ( x , ξ ) is uniformly bounded on Γt, we should expect to construct the phase function Φ in (13) using the factor Bt (x, ξ). Specifically, we would like Φ to satisfy the following equation

B t ( x , Φ x ) = B t ( y , ξ ) , y B 0 , ( x , ξ Γ t ) . (14)

The Equation (14) leads to complex eikonal equation (i.e., a non-linear partial differential equation with complex coefficients).

( A + | Φ x | ( x , y , ξ ) | 2 ) 1 2 - [ w 1 ( x ) t + i Φ x 1 ( x , y , ξ ) ] = A + | ξ | | 2 1 2 - ( w 1 ( y ) t + i ξ 1 ) . (15)

Since w1 (x) = 1 + x1, w1 (y) = 1 + y1, we have

Φ = ( x - y , ξ ) + ( x 1 - y 1 ) 2 ξ 1 2 ( 1 + y 1 ) + i ( x 1 - y 1 ) 2 | ξ | | 2 ( 1 + y 1 ) (16)

is a solution of (15). To use this we get

L 0 t ( x , D ) e i Φ ( x , y , ξ ) = e i Φ L 0 t ( x , Φ x ) + e i Φ 2 Φ x 1 2 .

Next, if we set

r ( x , y , ξ ) = G t ( y , ξ ) - G t ( x , ξ ) = - i [ w 1 ( y ) - w 1 ( x ) ] t (17)

then it follows from L0t (x, ξ) = Gt (x, ξ)Bt (x, ξ) and (14) that

L 0 t ( x , Φ x ) = L 0 t ( y , ξ ) + B t ( y , ξ ) r ( x , y , ξ ) . (18)

Consequently, (16)-(18) imply that

( 2 π ) n L 0 t ( x , D ) K 1 ( x , y ) = R n β 1 ( ξ ) e i Φ d ξ + R n β 1 ( ξ ) r ( x , y , ξ ) G t - 1 ( y , ξ ) e i Φ d ξ R n β 1 ( ξ ) A L 0 t - 1 ( y , ξ ) e i Φ d ξ + R n β 1 ( ξ ) 2 Φ x 1 2 L 0 t - 1 ( y , ξ ) e i Φ d ξ . (19)

By reasoning as in [3] we obtain that the first and second summands in (19) belong to a bounded subset of S0. So, we see that the equality (5) must hold. Now we let K (x, y) = K0 (x, y) + K1 (x, y) and R (x, y) = R0 (x, y) + R1 (x, y), where

R 1 ( x , y ) = R 10 ( x , y ) + R 11 ( x , y ) , R 10 ( x , y ) = R n β 1 ( ξ ) r ( x , y , ξ ) G t - 1 ( y , ξ ) e i Φ d ξ , R 11 ( x , y ) = R n β 1 ( ξ ) 2 Φ x 1 2 L 0 t - 1 ( y , ξ ) e i Φ d ξ , T 0 f ( x ) = B 0 K 0 ( x , y ) f ( y ) d y , T 1 f ( x ) = B 0 K 1 ( x , y ) f ( y ) d y .

Due to regularity of kernels, by using of Minkowski and Hölder inequalities we get the analog estimate as (7) and (9) for the operators T0 and R10. Thus, in order to finish the proof, it suffices to show that for f L2 (B0; E) one has

| α | 2 t 2 | α | D α T 1 * f L 2 ( B 0 ; H ) + A T 1 * f L 2 ( B 0 ; H ) C f L 2 ( B 0 ; H ) , (20)

t 1 2 R 11 * f L 2 ( B 0 ; H ) C f L 2 ( B 0 ; H ) , (21)

t 1 2 | | D ν R 11 * f L 2 ( B 0 ; H ) C | α | | ν | 1 D α f L 2 ( B 0 ; H ) , 1 | ν | 2. (22)

However, since R1,1 tT1, we need only to show the following

t 3 / 2 T 1 * f L 2 ( B 0 ; H ) C f L 2 ( B 0 ; H ) . (23)

By using the Minkowski inequalities we get

T 1 * f L 2 ( R n 1 ; E ) 1 4 1 4 B 0 K 1 * ( x , y ) f ( y ) d y | d y 1 ,

where K 1 * ( x , y ) = K ̄ 1 ( y , x ) . The estimates (13) and (16) imply that

K 1 * ( x , y ) = ( 2 π ) - n R n - 1 e i x - y m ( x 1 , y 1 , ξ | ) d ξ | ,

where

m ( x 1 , y 1 , ξ | ) = - β 1 ( ξ ) e i ( x 1 - y 1 ) ξ 1 + ( x 1 - y 1 ) 2 ( i | ξ | | - ξ 1 ) / 2 ( 1 + x 1 ) L 0 t - 1 ( y , ξ ) d ξ 1 .

Consequently, it follows from Plancherel's theorem that

R n - 1 K 1 * ( x , y ) f ( y ) d y | sup ξ | | m ( x 1 , y 1 , ξ | ) | R n - 1 | f ( y ) | 2 d y | 1 2 . (24)

Note that for every N we have

e i [ ( x 1 - y 1 ) 2 | ξ | | / 2 ( 1 + x 1 ) ] C N [ 1 + t ( x 1 - y 1 ) 2 ] - N on supp  β 1 .

Since A is a positive operator in E, we have

L 0 t 1 ( x , ξ ) B ( E ) 1 + | 2 i ξ 1 w 1 t + | ξ | 2 t 2 w 1 2 | 1

when - 2 i ξ 1 w 1 t + A + | ξ | 2 - t 2 w 1 2 S ( φ ) . Then by using the above estimate it not easy to check that

- β 1 ( ξ ) e i ξ 1 ( x 1 - y 1 ) - ( x 1 - y 1 ) 2 / 2 ( 1 + x 1 ) L 0 t - 1 ( y , ξ ) d ξ 1 = O ( t - 1 ) ,

i.e.,

| m ( x 1 , y 1 , ξ | ) | C t - 1 [ 1 + t ( x 1 - y 1 ) 2 ] - 1 .

Moreover, it is clear that

- ( 1 + t x 1 ) - 1 d x 1 = O t - 1 2 .

Thus from (24) by using the above relations and Young's inequality we obtain the desired estimate

T 1 * f L 2 ( B 0 ; H ) C t - 1 1 + t ( x 1 - y 1 ) 2 f ( y 1 , ) L 2 d y 1 d x 1 C t - 3 / 2 f L 2 ( R n ; H ) .

Moreover, by using the estimate (10) and the resolvent properties of the positive operator A we have

A T 1 * f L 2 ( B 0 ; H ) C f L 2 ( B 0 ; H ) .

The last two estimates then, imply the estimates (20)-(22).

Proof of Theorem 3.1: The estimates (7)-(9) imply the estimate (5), i.e., we obtain the assertion of the Theorem 3.1.

4 Lp-Carleman estimates and unique continuation for equation with variable coefficients

Consider the following DOE

L ( x , D ) u = i , j = 1 n a i j ( x ) D i j 2 u + A u = f ( x ) , x R n , (25)

where D k = i k and A is the possible unbounded operator in a Banach space E and aij are

real-valued smooth functions in Bε = {x Rn, |x| < ε}.

Condition 4.1. There is a positive constant γ such that i , j = 1 n a i j ( x ) ξ i ξ j γ | ξ | 2 for all ξ Rn, x B 0 = { x R n , | x | < 1 4 } .

The main result of the section is the following

Theorem 4.1. Let E be a Banach space satisfies the multiplier condition and A be a R-positive operator in E. Suppose the Condition 4.1 holds, n ≥ 3, p = 2 n n + 2 and p' is the conjugate of p, w = x 1 + x 1 2 2 and aij C(Bε). Then for u C 0 ( Bε; E(A)) and > 0 , 1 t < 1 2 the following estimates are satisfied:

e t w u L p | ( R n ; E ) C e t w L ( ε x , D ) u L p ( R n ; E ) , 1 p + 1 p | = 1 , (26)

α 1 t 1 + 1 n - α e t w D α u L p ( R n ; E ) + e t w A u L p ( R n ; E ) (27)

C e t w L ( ε x , D ) u L p ( R n ; E ) .

Proof. As in the proof of Theorem 3.1, it is sufficient to prove the following estimates

v L p | ( R n ; E ) C L t ε x , D v L p ( R n ; E ) , 1 p + 1 p | = 1 , (28)

| α | 1 t ( 1 + 1 n - | α | ) D α v L p ( R n ; E ) + A v L p ( R n ; E ) C L t ( ε x , D ) v L p ( R n ; E ) (29)

where,

L t ε x , D = e t w L ε x , D e - t w = L ε x , D + 2 t w 1 x 1 - ( t w 1 ) 2 - t 2 , w 1 = w x 1 .

Consequently, since w1 ≃ 1 on Bε, it follows that, if we let Qt (εx, D) be the differential operator whose adjoint equals

Q t * ( ε x , D ) = w 1 - 2 L ( ε x , D ) + 2 t w 1 - 1 x 1 - t 2 ,

then it suffices to prove the following

v L p | ( R n ; E ) C Q t ( ε x , D ) v L p ( R n ; E ) , 1 p + 1 p | = 1 , | α | t ( 1 + 1 n - | α | ) | | D α v | | L p ( R n ; E ) + | | A v | | L p ( R n ; E ) C | | Q t ( ε x , D ) v | | L p ( R n ; E ) , v C 0 ( B ε ; E ( A ) ) . (30)

The desired estimates will follow if we could constrict a right operator-valued parametrix T, for Qt* (εx, D) satisfying Lp estimates. these are contained in the following lemma.

Lemma 4.1. For t > 0 there are functions K = Kt and R = Rt, so that

Q t * ε x , D K x , y = δ x - y + R x , y , x , y B ε , (31)

where δ denotes the Dirac distribution. Moreover, if we let T = Tt be the operator with kernel K (x, y) and R be the operator with kernel R (x, y), then if ε and 1 t are sufficiently small, the adjoint of these operators satisfy the following uniform estimates

T * f L p 1 ( R n ; E ) C f L p ( R n ; E ) , 1 p + 1 p | = 1 , (32)

| α | 1 t ( 1 + 1 n - | α | ) D α T * f L p ( R n ; E ) C f L p ( R n ; E ) , (33)

A T * f L p ( R n ; E ) C f L p ( R n ; E ) ,

t 1 n R * f L q R n ; E C f L q R n ; E , q = p , p | , (34)

t - 1 + 1 n R * f L p R n ; E C f L p R n ; E , f C 0 B ε ; E . (35)

Proof. The key step in the proof is to find a factorization of the operator-valued symbol Q t * ε x , ξ that will allow to microlocally invert Q t * ε x , D near the set where Q t * ε x , ξ vanishes. Note that, after making a suitable choice of coordinates, it is enough to show that if L (x, D) is of the form

L x , D = D 1 2 + i , j = 2 n a i j D i D j , D j = 1 i x j

therefore, we can expressed Q t * ε x , ξ as

Q t * ε x , ξ = B t ε x , ξ G t ε x , ξ , (36)

where

B t x , ξ = w 1 - 1 ξ 1 + i A + w 1 - 1 a ε x , ξ | - t , G t x , ξ = w 1 - 1 ξ 1 - i A + w 1 - 1 a ε x , ξ | + t ,

where

a x , ξ | = i , j = 2 n a i j x ξ i ξ j .

The ellipticity of Q(x, D) and the positivity of the operator A, implies that the factor Gt (x, ξ) never vanishes and as in the proof of Theorem 3.1 we get that

G t - 1 ε x , ξ B H C 1 + | w 1 - 1 a ε x , ξ | | 1 2 + | t + w 1 - 1 ξ 1 | - 1 , (37)

x B ε , ξ R n ,

i.e., the operator function Gt (εx, ξ) has uniformly bounded inverse for (x, ξ) ∈ Bε ×Rn. One can only investigate the factor Bt (εx, ξ). In fact, if we let

Δ t = x , ξ B ε × R n : ξ 1 = 0 , | ξ | | = t w 1 ,

then the operator function Bt (x, ξ) is not invertible for (x, ξ) ∈ Δt. Nonetheless, Bt (εx, ξ) and Q t * ε x , ξ can be have a bounded inverse when (x, ξ) is sufficiently far away. For instance, if we define

Γ t = x , ξ B ε × R n : | ξ | | t 4 , 4 t , | ξ 1 | t 4 ,

by properties of positive operators we will get the same estimate of type (37) for the singular factor Bt. Hence, we using this fact and the resolvent properties of positive operators we obtain the following estimate

Q t * - 1 ε x , ξ B E C 1 + | ξ | | + | t + w 1 - 1 ξ 1 | - 1 when x , ξ c Γ t . (38)

As in § 3, we can use (38) to microlocallity invert Q t * ε x , D away from Γt . To do this, we first fix β C 0 R as in § 3. We then define

β 0 = β 0 t = 1 - β ξ | / t β 1 - ξ 1 / t .

It is clear that β0 (ξ) = 0 on Γt. Consequently, if we define

K 0 x , y = 2 π - n R n β 0 ξ e i x - y , ξ Q t * - 1 ε y , ξ d ξ (39)

and recall (37), then we can conclude that standard microlocal arguments give that

Q t * ε x , D K 0 x , y = 2 π - n R n β 0 ξ e i x - y , ξ d ξ + R 0 x , y , (40)

where R0 belongs to a bounded subset of S-1 that independent of t. Since the adjoint operator R 0 * also is abstract pseudodifferential operator with this property, by reasoning as in [31, Theorem 6] it follows that

R 0 * f L p R n ; E C f L p R n ; E , f C 0 B ε ; E , (41)

t R 0 * f L q R n ; E C f L q R n ; E , f C 0 B ε ; E , (42)

q = p , p , 1 p + 1 p = 1 .

Moreover, the positivity properties of A and the estimate (38) imply that the operator functions α 2 β 0 ξ t 2 - α ξ α Q t * - 1 ε x , ξ and β 0 ξ A Q t * - 1 ε x , ξ are uniformly bounded. Next, let T0 be the operator with kernel K0. Then in a similar way as in [31] we obtain that

α 1 t 2 - α D α T 0 * f L p R n ; E C f L p R n ; E , (43)

A T 0 * f L p R n ; E C f L p R n ; E

which also the first estimate is stronger than the corresponding inequality in Lemma 4.1. Finally, since T0 S-2 and 1 p - 1 p | = 2 n it follows from imbedding theorem in abstract Sobolev spaces [17] that

T 0 * f L p | R n ; E C f L p R n ; E , f C 0 B ε ; E . (44)

Thus, we have shown that the microlocal inverse corresponding to cΓt, satisfies the desired estimates.

Let β1 (ξ) = 10 (ξ). To invert Q t * ε x , D for (x, ξ) ∈ Γt, we have to construct a Fourier integral operator T1, with kernel

K 1 x , y = 2 π - n R n β 1 ξ e i Φ x , y , ξ Q 0 t * - 1 ε y , ξ d ξ , (45)

such that the analogs of (39) and (32)-(35) are satisfied. For this step the factorization (36) of the symbol Q t * ε y , ξ will be used. Since the factor Gt (εx, ξ) has a bounded inverse for (x, ξ) ∈ Γt, the previous discussions show that we should try to construct the phase function in (46) using the factor Bt (εx, ξ). We would like Φ (x, y, ξ) to solve the complex eikonal equation

B t ε x , Φ x = B t ε y , ξ , x , y B ε , ξ supp  β 1 , (46)

Since Bt (εx, Φx) - Bt (εy, ξ) is a scalar function (it does not depend of operator A ), by reasoning as in [3, Lemma 3.4] we get that

Φ ( x , y , ξ ) = ϕ ( x , y , ξ ) + ψ ( x , y , ξ ) ,

where ϕ is real and defined as

ϕ x , y , ξ = x 1 - y 1 ξ 1 + O x - y 2 ξ ,

while

ψ x , y , ξ = x 1 - y 1 ξ 1 + O x 1 - y 1 2 ξ

and

Im ψ x , y , ξ c x 1 - y 1 2 ξ , c > 0 . (47)

Then we obtain from the above that

Q t * ε x , D e i Φ x , y , ξ = e i Φ Q t * ε x , Φ x + e i Φ w 1 - 2 L ε x , D Φ .

Next, if we set

r x , y , ξ = G t ε y , ξ - G t ε x , ξ = w 1 - 1 y ξ 1 - i a ε y , ξ - w 1 - 1 x ξ 1 - i a ε x , ξ (48)

then it follows from (36) and (48) that

e i Φ Q t * ε x , Φ x = e i Φ Q t * ε y , ξ + e i Φ B t ε y , ξ r x , y , ξ + O t - N (49)

for every N when β1 (ξ) ≠ 0. Consequently, (49), (50) imply that

2 π n Q t * ε x , D K 1 x , y = β 1 ξ e i Φ d ξ + β 1 ξ r x , y , ξ G t - 1 ε y , ξ e i Φ d ξ

w 1 - 2 β 1 ( ξ ) Q t * - 1 ( ε y , ξ ) ( L ( ε x , D ) Φ ) e i Φ d ξ + O ( t - N ) . (50)

By reasoning as in Theorem 3.1 we obtain from (51) that

Q t * ( ε x , D ) K 1 ( x , y ) = ( 2 π ) - n β 1 ( ξ ) e i ( x - y , ξ ) d ξ + R 10 ( x , y ) + R 11 ( x , y ) ,

where

R 11 ( x , y ) = ( 2 π ) - n w 1 - 2 β 1 ( ξ ) Q t * - 1 ( ε y , ξ ) ( L ( ε x , D ) Φ ) e i Φ d ξ (51)

while R10 belongs to a bounded subset of S-1 and tR10 belongs to a bounded subset of S0. In view of this formula, we see that if we let K (x, y) = K0 (x, y) + K1 (x, y) and R (x, y) = R0 (x, y)+R1 (x, y), where R1 = R10 +R11, then we obtain (31). Moreover, since R10 satisfies the desired estimates, we see from Minkowski inequality that, in order to finish the proof of Lemma 4.1, it suffices to show that for f C 0 ( B ε ; E )

T 1 * f L p | ( R n ; E ) C f L p ( R n ; E ) , (52)

| α | 1 t ( 1 + 1 n | α | ) D α T 1 * f L p ( R n ; E ) C f L p ( R n ; E ) , (53)

t 1 n R 11 * f L q ( R n ; E ) C f L q ( R n ; E ) , q = p , p | , (54)

t 1 + 1 n R 11 * f L p ( R n ; E ) C f L p ( R n ; E ) , (55)

where 1 p + 1 p | = 1 .

To prove the above estimates we need the following prepositions for oscillatory integral in E-valued Lp spaces which generalize the Carleson and Sjolin result [36].

Preposition 4.1. Let E be Banach spaces and A C 0 ( R n , L ( E ) ) . Moreover, suppose Φ ∈ Csatisfies | ∇Φ| ≥ γ > 0 on supp A. Then for all λ > 1 the following holds

e i λ Φ ( x ) A ( x ) d x L ( E ) C N λ - N , N = 1 , 2 ,

where CN-depends only on γ if Φ and A (x) belong to a bounded subset of Cand C(Rn, L (E)) and A is supported in a fixed compact set.

Proof. Given x0 ∈ supp A. There is a direction ν Sn-1 such that |(ν, ∇Φ)| ≥ γ 2 on some ball centered at x0. Thus, by compactness, we can choose a partition of unity φ j C 0 consisting of a finite number of terms and corresponding unit vectors νj such that j = 1 m φ j ( x ) = 1 on supp A and | ( ν j , Φ ) | γ 2 on supp φj. For Aj = φjA it suffices to prove that for each j

e i λ Φ ( x ) A j ( x ) d x L ( E ) C N λ - N , N = 1 , 2 , .

After possible changing coordinates we may assume that νj = (1, 0, . . . , 0) which means that Φ x 1 γ 2 on supp φj. If let L ( x ; D ) = Φ x 1 - 1 1 i λ x 1 , then L ( x ; D ) e i λ Φ ( x ) = e i λ Φ ( x ) . Consequently, if L * = x 1 1 i λ Φ x 1 - 1 is a adjoint, then

e i λ Φ ( x ) A ( x ) d x = e i λ Φ ( x ) ( L * ) N A j ( x ) d x .

Since our assumption imply that (L*)N Aj (x) = O (λ-N), the result follows.

Preposition 4.2. Suppose Φ ∈ Cis a phase function satisfying the non-degeneracy condition det 2 Φ x i x j 0 on the support of

A ( x , y ) C 0 ( R n × R n , L ( E ) ) .

Then for T λ f = R n e i λ Φ ( x , y ) A ( x , y ) f ( y ) d x , λ > 0 the following estimates hold

T λ f L p ( R n ; E ) C λ - n - 1 p f L p ( R n ; E ) , 1 p 2 , T λ f L p ( R n ; E ) C λ - n p f L p ( R n ; E ) , 1 p + 1 p = 1 .

Proof. In view of [3, Remark 2.1] we have

| x [ Φ ( x , y ) - Φ ( x , z ) ] | | y - z | (56)

where |y - z| is small. By using a smooth partition of unity we can decompose A (x, y) into a finite number of pieces each of which has the property that (57) holds on its support. So, by (57) we can assume

| x [ Φ ( x , y ) - Φ ( x , z ) ] | C | y - z | (57)

on supp A for same C > 0. To use this we notice that

T λ f 2 2 = K λ ( y , z ) f ( y ) f ̄ ( z ) d y d z ,

where

K λ ( y , z ) = R n