Observability for heat equations Kim Dang PHUNG Yangtze Center of Mathematics, Sichuan University, Chengdu 610064, China. Université d’Orléans, Laboratoire de Mathématiques - Analyse, Probabilités, Modélisation - Orléans, CNRS FR CNRS 2964, 45067 Orléans cedex 2, France. E-mail: kim_dang_phung@yahoo.fr Abstract This talk describes di¤erent approaches to get the observability for heat equations without the use of Carleman inequalities. Contents 1 The heat equation and observability 2 2 Our motivation 3 3 Our strategy 3.1 Proof of "Hölder continuous dependence from one ) "Sum of Laplacian eigenfunctions" . . . . . . 3.2 Proof of "Hölder continuous dependence from one ) "Observability" . . . . . . . . . . . . . . . . . 3.3 Proof of "Hölder continuous dependence from one ) "Re…ned Observability" . . . . . . . . . . . . 4 point in time" . . . . . . . . . point in time" . . . . . . . . . point in time" . . . . . . . . . 7 4 What I hope 4.1 Logarithmic convexity method . . . . . . . . . . . . . . . . . . . 4.2 Weighted logarithmic convexity method . . . . . . . . . . . . . . 9 9 10 4 4 5 What I can do 12 5.1 The frequency function . . . . . . . . . . . . . . . . . . . . . . . . 12 5.2 The frequency function with weight . . . . . . . . . . . . . . . . . 14 5.3 The heat equation with space-time potential . . . . . . . . . . . . 19 This talk was done when the author visited School of Mathematics & Statistics, Northeast Normal University, Changchun, China, (July 4-21, 2011). 1 6 What already exists 21 6.1 Monotonicity formula . . . . . . . . . . . . . . . . . . . . . . . . 21 6.1.1 Proof of Lemma B . . . . . . . . . . . . . . . . . . . . . . 22 6.1.2 Proof of Lemma A . . . . . . . . . . . . . . . . . . . . . . 23 6.1.3 Proof of Lemma C . . . . . . . . . . . . . . . . . . . . . . 24 6.2 Quantitative unique continuation property for the Laplacian . . . 28 6.3 Quantitative unique continuation property for the elliptic operator @t2 + . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 6.4 The heat equation and the Hölder continuous dependence from one point in time . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 1 The heat equation and observability We consider the heat equation in the solution u = u(x; t) 8 u=0 in (0; +1) , < @t u u=0 on @ (0; +1) , : u ( ; 0) 2 L2 ( ) , (1.1) living in a bounded open set in Rn , n 1, either convex or C 2 and connected, with boundary @ . It is well-known that the above problem is well-posed and have a unique solution u 2 C [0; T ] ; L2 ( ) \ L2 0; T ; H01 ( ) for all T > 0. The observability problem consists in proving the following estimate Z 2 ju (x; T )j dx C Z 0 T Z ! 2 ju (x; t)j dxdt for some constant C > 0 independent on the initial data. Here, T > 0 and ! is a non-empty open subset in . 2 In the literature, two ways allow to prove such observability estimate. One is due to the work of Fursikov and Imanuvilov based on global Carleman inequalities (see [FI]). The other proof is established by Lebeau and Robbiano (see [LR]. See [Le] for an english version). We resume the Lebeau-Robbiano strategy as follows. Z TZ 2 2 ku ( ; T )kL2 ( ) C ju (x; t)j dxdt 0 ! * controllability in …nite and in…nite dimension * X 2 j=1;::;N for any faj g where (ej ; ary conditions (0 < 1 jaj j j) C Ce p N Z X 2 aj ej (x) dx ! j=1;::;N solves the eigenvalue problem with Dirichlet boundbeing the corresponding eigenvalues). 2 * For any > 0 and any non-trivial ' 2 C01 ( (0; T )), there are C > 0 and 2 (0; 1), such that for any w 2 H 2 ( (0; T )) with @t2 + w = f and wj@ = 0, it holds 1 kwkH 1 ( ( ;T )) C kwkH 1 ( kf kL2 ( (0;T )) (0;T )) + k'wkL2 ( (0;T )) The above interpolation inequality is proved using Carleman inequalities. Recently, a shortcut of the Lebeau-Robbiano strategy is given in [M]. 2 Our motivation In application to bang-bang control (see [W]), we need the following re…ned observability estimate from measure set in time. Z Z ku ( ; T )kL2 ( ) C ju (x; t)j dxdt E ! for some constant C > 0 independent on the initial data. Here, E (0; T ) is a measurable set of positive measure and ! is a non-empty open subset in . Further, we want to be able to extend the proof to heat equations with space-time potentials. The approach describes in this talk is linked to parabolic quantitative unique continuation (see [BT], [Li], [L], [P], [EFV], [K], [KT] and references therein). 3 3 Our strategy We are able to prove that Hölder continuous dependence from one point in time =) Observability from a measure set in time. More precisely, ku ( ; t)kL2 ( C ku ( ; 0)kL2 ( ) ku ( ; T )kL2 ( 3.1 bj e jt P j=1;::;N 2 0 @C j=1;::;N bj e X j=1;::;N p which implies when t = 1= P j=1;::;N + Z Z C ju (x; t)j dxdt ) E bj e j= p 2 N jt ! ej (x). Then it holds for any t > 0 1 0 Z 2A B C=t jbj j @Ce 11 2 X jt bj e ! j=1;::;N C ej (x) dxA N Ce2 p CeC N P j=1;::;N p This gives the desired estimate with bj e 3.2 8t > 0 Proof of "Hölder continuous dependence from one point in time" ) "Sum of Laplacian eigenfunctions" Choose u (x; t) = X 1 eC=t ku ( ; t)kL2 (!) ) N R ! j= p bj e P j= p j=1;::;N bj e N 2 N j= p 2 N 1 ej (x) dx = aj . Proof of "Hölder continuous dependence from one point in time" ) "Observability" Let ` 2 [0; T ) and `1 2 (`; T ]. Let z > 1. Introduce the decreasing sequence f`m gm 1 , which converges to ` given by `m+1 = ` + 1 (`1 zm 4 `) . . Then 1 (z 1) (`1 `) > 0 . zm We start with the following interpolation estimate. Let ! be a nonempty open subset of . For any 0 t1 < t2 T , `m ku ( ; t2 )kL2 ( C1 t C2t e 2 1 ku ( ; t2 )kL2 (!) + " ku ( ; t1 )kL2 ( " ) Let 0 < `m+2 < `m+1 ku ( ; t)kL2 ( `m+1 = t < `m < T . We get C1 t e " ) 8" > 0 . ) C2 `m+2 ku ( ; t)kL2 (!) + " ku ( ; `m+2 )kL2 ( 8" > 0 . ) Recall that ku ( ; `m )kL2 ( ku ( ; t)kL2 ( ) ) . Therefore, ku ( ; `m )kL2 ( C1 t e " ) C2 `m+2 ku ( ; t)kL2 (!) + " ku ( ; `m+2 )kL2 ( ) 8" > 0 . Integrating it over t 2 (`m+1 ; `m ), it gives (`m `m+1 ) ku ( ; `m )kL2 ( R `m C2 `m+2 C1 `m+1 " e ) ku ( ; t)kL2 (!) dt `m+1 `m+1 ) ku ( ; `m+2 )kL2 ( ) 8" > 0 . +" (`m That is ku ( ; `m )kL2 ( h ) 1 `1 zm `z 1 i C1 C2 " e h 1 z m+1 `1 ` z 1 +" ku ( ; `m+2 )kL2 ( ) i R `m `m+1 8" > 0 . ku ( ; t)kL2 (!) dt Then ku ( ; `m )kL2 ( Take d = 2C2 h dz m+2 1 C1 z C2 R `m 1 z m+1 `1 ` z 1 +" ku ( ; `m+2 )kL2 ( i ) R `m ku ( ; t)kL2 (!) dt 8" > 0 . `m+1 . It gives ku ( ; `m )kL2 ( "1+ e ) dz m+2 8" > 0 . ku ( ; `m+2 )kL2 ( ) , then ( +1)dz m+2 1 C1 z C2 i h ku ( ; t)kL2 (!) dt `m+1 dz m+2 e ) 1 1 `1 ` z(z 1) " e Take " = e 1 1 C1 2C2 " z C2 e R `m `m+1 ku ( ; `m )kL2 ( ) ku ( ; t)kL2 (!) dt . 5 e (2+ )dz m+2 ku ( ; `m+2 )kL2 ( ) Take z = q +2 +1 , then e (2+ q )dz m +1 C1 +2 C2 ku ( ; `m )kL2 ( R `m (2+ )dz m+2 e ) ku ( ; t)kL2 (!) dt . `m+1 ku ( ; `m+2 )kL2 ( ) Change m to 2m0 and sum the above from m0 = 1 to in…nity give the desired result with E = (0; T ). When E = (0; T ), we take ` = 0 and `1 = T , and get 2 (2+ )2C2 4 T1 e e (2+ P )dz 2 e q m0 1 P 0 qm Finally, r r +2 +1 +2 +1 1 3 5 ku ( ; T )kL2 ( ku ( ; `2 )kL2 ( ) (2+ )dz 2m 0 ) ku ( ; `2m0 )kL2 ( R `2m0 +1 C1 1 +2 C2 `2m0 +1 ku ( ; t)kL2 (!) R +1 C1 `2 +2 C2 0 ku ( ; t)kL2 (!) dt . ku ( ; T )kL2 ( r ) + 1 C1 2C2 T1 (2+ e + 2 C2 ) p e ) (2+ )dz 2m 0 +2 ku ( ; `2m0 +2 )kL2 ( ) dt p p +2[ +2+ +1] Z T ku ( ; t)kL2 (!) dt . 0 In particular, ku ( ; 1)kL2 ( ) r + 1 C1 2C2 (2+ e + 2 C2 r + 1 C1 2C2 (2+ e + 2 C2 ) p p p +2[ +2+ +1] Z 1 ku ( ; t)kL2 (!) dt 0 and ku ( ; m)kL2 ( for any m ) ) p p +2[ +2+ +1] p Z m m 1 1. Now, take T such that M < T ku ( ; t)kL2 (!) dt M + 1 for some M 1, T 2 ku ( ; T )kL2 ( ) M ku ( ; M )kL2 ( ) PM ku ( ; m)kL2 ( q m=1 p +1 C1 2C2 (2+ ) +2 C2 e q p +1 C1 2C2 (2+ ) +2 C2 e q p +1 C1 2C2 (2+ ) +2 C2 e ) p p +2[ +2+ +1] Rm m=1 m 1 ku ( ; t)kL2 (!) p p R +2[ +2+ +1] M ku ( ; t)kL2 (!) dt 0 p p R T +2[ +2+ +1] ku ( ; t)kL2 (!) dt . 0 We conclude that for any T 1, r 1 + 1 C1 2C2 T1 (2+ ku ( ; T )kL2 ( ) e T + 2 C2 6 ) p PM p p +2[ +2+ +1] Z 0 dt T ku ( ; t)kL2 (!) dt , and for any T > 1 ku ( ; T )kL2 ( 2 T ) r + 1 C1 2C2 (2+ e + 2 C2 Combining the case T ku ( ; T )kL2 ( ) p p p +2[ +2+ +1] Z T ku ( ; t)kL2 (!) dt . 0 1 and the case T > 1, we get the following observability. r Z 3 + 1 C1 4C2 (1+ T1 )(2+ )2 T ku ( ; t)kL2 (!) dt e T + 2 C2 0 ) for any T > 0. 3.3 Proof of "Hölder continuous dependence from one point in time" ) "Re…ned Observability" In this proof, C denotes a positive constant which may change of value from line to line. Let ` 2 [0; T ) and `1 2 (`; T ]. Let z > 1. Introduce the decreasing sequence f`m gm 1 , which converges to ` given by `m+1 = ` + 1 (`1 zm `) . Then 1 (z 1) (`1 `) > 0 . zm We start with the following interpolation estimate. Introduce ! eb! any 0 t1 < t2 T , `m ku ( ; t2 )kL2 ( ) `m+1 = C1 t C2t e 2 1 ku ( ; t2 )kL2 (e!) + " ku ( ; t1 )kL2 ( " ) 8" > 0 . By Nash and Poincare inequality, ku ( ; t2 )kL2 (e!) C3 ku ( ; t2 )kL1 (!) + " kru ( ; t2 )kL2 ( "n=2 8" > 0 . ) By an energy method, kru ( ; t2 )kL2 ( C4 ) 1=2 (t2 t1 ) ku ( ; t1 )kL2 ( ) Therefore, from the above three estimate, denoting = C5 = C1 C3 C1 C4 p C2 1+ n n + , 2 2 n=2 , C6 = (n + 1) C2 , 7 . . For we get ku ( ; t2 )kL2 ( C5 t C6t e 2 1 ku ( ; t2 )kL1 (!) + " ku ( ; t1 )kL2 ( " ) Let 0 < `m+2 < `m+1 ku ( ; t)kL2 ( t < `m < T . We get C5 t e " ) 8" > 0 . ) C6 `m+2 ku ( ; t)kL1 (!) + " ku ( ; `m+2 )kL2 ( 8" > 0 . ) Recall that ku ( ; `m )kL2 ( ku ( ; t)kL2 ( ) . ) Therefore, ku ( ; `m )kL2 ( C5 t e " ) C6 `m+2 ku ( ; t)kL1 (!) + " ku ( ; `m+2 )kL2 ( ) 8" > 0 . Integrating it over t 2 (`m+1 ; `m ), it gives (`m `m+1 ) ku ( ; `m )kL2 ( R `m C6 `m+2 C5 `m+1 " e ) ku ( ; t)kL1 (!) dt `m+1 ) ku ( ; `m+2 )kL2 ( ) 8" > 0 . +" (`m `m+1 That is ku ( ; `m )kL2 ( h ) 1 `1 zm `z 1 i C5 C6 " e h 1 z m+1 `1 ` z 1 +" ku ( ; `m+2 )kL2 ( ) i R `m `m+1 8" > 0 . ku ( ; t)kL1 (!) dt Then ku ( ; `m )kL2 ( Take d = 2C6 h dz m+2 1 C5 z C6 q e +" ku ( ; `m+2 )kL2 ( i ) R `m ku ( ; t)kL1 (!) dt 8" > 0 . `m+1 . It gives "1+ e ) dz m+2 8" > 0 . ku ( ; `m )kL2 ( ) e ku ( ; `m+2 )kL2 ( (2+ )dz m+2 ku ( ; t)kL1 (!) dt . `m+1 ) ku ( ; `m+2 )kL2 ( ) then (2+ )dz m 1 C5 z C6 1 z m+1 `1 ` z 1 , then R `m +2 +1 , h ku ( ; t)kL1 (!) dt `m+1 ( +1)dz m+2 1 C5 z C6 i ku ( ; `m )kL2 ( R `m dz m+2 e Take z = ) 1 1 `1 ` z(z 1) " e Take " = e 1 1 C5 2C6 " z C6 e R `m `m+1 ku ( ; `m )kL2 ( e ) ku ( ; t)kL1 (!) dt . 8 (2+ )dz m+2 ku ( ; `m+2 )kL2 ( ) Change m to 2m0 and sum the above from m0 = 1 to in…nity give the desired result with E = (0; T ). By using a property of density points, we are able to replace (0; T ) by a measurable set of positive measure (see [PW2]). For simplicity, when E = (0; T ), we take ` = 0 and `1 = T , and get 2 (2+ )2C6 4 T1 r r +2 +1 +2 1 3 5 +1 ku ( ; T )kL2 ( ) e (2+ )dz 2 e ku ( ; `2 )kL2 ( ) h 0 P (2+ )dz 2m ku ( ; `2m0 )kL2 ( m0 1 e R P 1 C5 `2m0 m0 1 z C6 `2m0 +1 ku ( ; t)kL1 (!) dt R 1 C5 T z C6 0 ku ( ; t)kL1 (!) dt . ) e (2+ )dz 2m 0 +2 ku ( ; `2m0 +2 )kL2 ( That is ku ( ; T )kL2 ( 4 ) r + 2 C5 2C6 T1 (2+ e + 1 C6 ) p p p +2[ +2+ +1] Z T 0 Z ! ju (x; t)j dxdt . What I hope Recall the following two energy identities. For t 2 (0; T ], Z Z 1 d 2 2 ju (x; t)j dx + jru (x; t)j dx = 0 , 2 dt Z Z 1 d 2 2 jru (x; t)j dx + j u (x; t)j dx = 0 . 2 dt 4.1 Logarithmic convexity method Recall that if t 7!logf is a convex function, then for any t1 < t2 and any 2 (0; 1), logf ( t2 + (1 ) t1 ) logf (t2 ) + (1 ) logf (t1 ) which implies f ( t2 + (1 1 ) t1 ) [f (t2 )] [f (t1 )] . 00 2 Further, when f 2 C , logf is a convex function if and only if (logf ) 2 f 00 f (f 0 ) 0. 2 f 9 = ) i Now take t = T , t1 = 0, t2 = T and f (t) = 00 (logf (t)) = 1 f2 4 R 2 j u (x; t)j dx R R 2 ju (x; t)j dx. Therefore, 2 ju (x; t)j dx 2 0 by Cauchy-Schwarz inequality, R 2 2 jru (x; t)j dx which implies Z Z 2 ju (x; t)j dx t=T 2 ju (x; T )j dx Z 1 t=T 2 . ju (x; 0)j dx This estimate has similar form than a quantitative Hölder continuous dependence. We refer to [FEF] for more advanced computation with this technique. 4.2 Weighted logarithmic convexity method We would like reproduce similar computation than the previous subsection but with Z 2 f (t) = ju (x; t)j G (x; t) dx , where G is a suitable weighted function. Step 1 .- Make appear Br .- Let Br be the ball of radius r and center 2 x0 , and contained in . Denote m0 = max jx x0 j . Let > 0. Take G (x; t) = x2 e jx T x0 j2 t+ R . If logf is a convex function, then 2 R ju (x; t)j e 2 ju (x; T )j e 1 "(1 )= 1 "(1 jx T )= R R x0 j2 t+ jx dx x0 j2 t=T dx jx 2 ju (x; T )j e x0 j2 2 jx x0 j2 2 jx x0 j2 ju (x; T )j e R 2 ju (x; 0)j e dx dx + " On the other hand, = R R R R ju (x; T )j e 2 Br ju (x; T )j e jx 2 Br ju (x; T )j dx + e 2 dx x0 j2 ju (x; T )j dx + e Br dx + r2 R r2 " R R nBr e m0 T t 10 R R jx x0 j2 T+ 1 t=T dx 2 ju (x; 0)j e jx x0 j2 T+ 2 ju (x; 0)j e 2 nBr jx x0 j2 T+ ju (x; T )j e jx 1 dx dx x0 j2 8" > 0 . dx 2 ju (x; T )j dx 2 ju (x; t)j e jx T x0 j2 t+ dx . Take t = T =2 i.e., R = 1=2 and 2 ju (x; T =2)j e jx x0 j2 T =2+ dx > 0 such that 1 " R + 12 +" It implies R 2 ju (x; T =2)j dx for any " > 0. Take " = R 2 ju (x; T )j dx 1 2 1 "(1 e )= r2 m0 e T =2 = 12 . Then 2 Br R ju (x; T )j dx 2 ju (x; T =2)j e R 2 ju (x; 0)j e jx x0 j2 T =2+ jx x0 j2 T+ dx dx 8" > 0 . m0 R jx x0 j2 2 e T =2 ju (x; T =2)j e T =2+ dx m0 m0 R R 2 2 2e T =2 1" Br ju (x; T )j dx + 2e T =2 " ju (x; 0)j dx , R hR 2 ju (x; T =2)j dx 2 ih m0 2e T =2 i 2 ju (x; 0)j dx R 1 . Then ju (x; T =2)j dx i1=2 hR i1=2 hR m0 2 2 4e T =2 Br ju (x; T )j dx ju (x; 0)j dx . Step 2 .- Get good sign for a derivative. The choice of the weighted function G can be seen as follows. The quantity Z Z 2 2 (@t ) ju (x; t)j G (x; t) dx + ju (x; t)j (@t + ) G (x; t) dx where G 2 C 1 , has two expressions Z Z d 2 2 2 ju (x; t)j G (x; t) dx @ ju (x; t)j G (x; t) ju (x; t)j @ G (x; t) d dt @ and 2 Z 2 jru (x; t)j G (x; t) dx + Z 2 ju (x; t)j (@t + ) G (x; t) dx . When (@t + ) G (x; t) = 0 and u = 0 on @ , Z Z 1 d 2 2 ju (x; t)j G (x; t) dx + jru (x; t)j G (x; t) dx = 0 . 2 dt This can be compared with the …rst energy identity. The natural choice for G is given by jx x0 j2 1 G (x; t) = e 4(T t+ ) , n=2 (T t + ) which satis…es (@t + ) G (x; t) = 0 and G (x; T ) = Of course, we will need to compute d dt 11 R 1 n=2 2 e jx x0 j2 4 . jru (x; t)j G (x; t) dx. Conclusion .- There are two parts: "Make appear Br " and "Get good sign for a derivative". With the weighted logarithmic convexity method, "Make appear Br " was easy but "Get good sign for a derivative" is di¢ cult and actually unsolved. Extension .- The above computation can be extended to the heat equation with second member. We consider the heat equation in the solution U = U (x; t) with a suitable second member F = F (x; t) 8 U =F in (0; T ) , < @t U U =0 on @ (0; T ) , : U ( ; 0) 2 L2 ( ) . The two energy identities become as follows. For t 2 (0; T ], Z Z Z 1 d 2 2 jU (x; t)j dx + jrU (x; t)j dx = F (x; t) U (x; t) dx , 2 dt Z Z Z 1 d 2 2 jrU (x; t)j dx + j U (x; t)j dx = F (x; t) U (x; t) dx . 2 dt Further, R R 2 2 1 d jU (x; t)j G (x; t) dx + jrU (x; t)j G (x; t) dx R2 dt = U (x; t) F (x; t) G (x; t) dx . 5 What I can do 5.1 The frequency function De…ne for t 2 (0; T ], R N (t) = R 2 jru (x; t)j dx 2 ju (x; t)j dx , whenever Z 2 ju (x; t)j dx 6= 0 . Then the …rst energy identity becomes as follows. Z Z 1 d 2 2 ju (x; t)j dx + N (t) ju (x; t)j dx = 0 . 2 dt On the other hand, we study the sign N 0 (t). d N (t) = 2 dt R 2 j u (x; t)j dx R 2 R ju (x; t)j dx + 2 ju (x; t)j dx 12 2 R 2 jru (x; t)j dx 2 0 by Cauchy-Schwarz inequality. Therefore, for t 2 [ ; T ], N (T ) N (t) N( ) . Backward uniqueness uses this technique when u ( ; 0) 2 H01 ( ) with N (t) N (0). We have to solve Z Z 1 d 2 2 ju (x; t)j dx + N (0) 0 ju (x; t)j dx . 2 dt That is d 2tN (0) e dt 0 Integrating over (0; t), Z 2 ju (x; 0)j dx Z 2 ju (x; t)j dx e2tN (0) Z . 2 ju (x; t)j dx . In our problem, we use N (T ) N (t). We have to solve Z Z 1 d 2 2 ju (x; t)j dx + N (T ) ju (x; t)j dx 0 . 2 dt That is d 2tN (T ) e dt Z 2 ju (x; t)j dx 0. Integrating over (0; T ), 2T N (T ) e Z 2 ju (x; T )j dx That is 2T N (T ) R log R Z 2 ju (x; 0)j dx . 2 ju (x; 0)j dx 2 ju (x; T )j dx . Extension .- The above computation can be extended to the heat equation with second member. We consider the heat equation in the solution U = U (x; t) with a suitable second member F = F (x; t) 8 U =F in (0; T ) , < @t U U =0 on @ (0; T ) , : U ( ; 0) 2 L2 ( ) . The following inequality holds (see [BT]). R 2 jF (x; t)j dx d N (t) R . 2 dt jU (x; t)j dx 13 Indeed, d dt N (t) = 2 [ R But j U j2 R R F 2 U dx] R R R 2 R jrU j2 dx F U dx jU j2 dx+( jrU j2 dx) R 2 ( jU j2 dx) R 2 jrU j dx F U dx R . 2 jrU j dx 2 R R R 2 2 1 = jrU j dx 12 F U dx F U dx 4 R R R 2 2 1 = U U dx 12 F U dx F U dx 4 R R 2 2 1 = U + 12 F U dx F U dx 4 R R R 2 2 2 1 U + 2 F dx jU j dx 14 F U dx by Cauchy-Schwarz inequality. Finally, = 5.2 Let R R R 2 2 jrU j dx R F U (x; t) dx R 2 jrU j dx R 2 2 U + 21 F dx jU j dx R R R R 2 2 2 j U j dx jU j dx + U F dx jU j dx + 2 1 2F dx R 2 jU j dx . The frequency function with weight > 0. Recall that for t 2 [0; T ] G (x; t) = 1 (T e n=2 t+ ) jx x0 j2 4(T t+ ) . De…ne for t 2 (0; T ], R Z 2 jru (x; t)j G (x; t) dx 2 N (t) = R , whenever ju (x; t)j dx 6= 0 . 2 ju (x; t)j G (x; t) dx Step 1 .- Get good sign for a derivative. We claim that if star-shaped w.r.t. x0 , then d N (t) dt That is d [(T dt 1 N (t) . t+ T t + ) N (t)] 0. Integrating over (t; T ), N (T ) (T + ) N (t) . 14 is convex or Recall that 1 d 2 dt Z 2 ju (x; t)j G (x; t) dx + Z 2 jru (x; t)j G (x; t) dx = 0 . That is 1 d 2 dt Z 2 ju (x; t)j G (x; t) dx + N (t) Z 2 ju (x; t)j G (x; t) dx = 0 . Therefore, Z Z d 2 N (T ) 2 2 ju (x; t)j G (x; t) dx + ju (x; t)j G (x; t) dx dt T+ That is 2t N (T ) d e T+ dt Z 2 ju (x; t)j G (x; t) dx Integrating over (0; T =2), Z T N (T ) 2 T + ju (x; T =2)j G (x; T =2) dx e But R R 2 ju (x; T )j dx 0. 2 ju (x; 0)j G (x; 0) dx . 2 ju (x; T =2)j dx m0 R 2 ju (x; T =2)j e e 2T 2 where m0 = max jx Z 0. jx x0 j2 4(T =2+ ) dx . x0 j . Therefore, x2 T e N (T ) T+ R R R ju(x;0)j2 G (x;0)dx 2 ju(x;T =2)dx R =2)j G (x;T ju(x;0)j2 dx jx x0 j2 ) ju(x;T =2)j2 e 4(T =2+ m0 R e 2T ju(x;0)j2 dx R . ju(x;T )j2 dx That is m0 N (T ) 1+ On the other hand, n 4 n 4 m0 1+ T Finally, n + N (T ) 4 T e 2T log R n +1 4 e1+ 2T log R R 2 ju (x; 0)j dx T 15 . 2 ju (x; T )j dx R 2 ju (x; 0)j dx 2 ju (x; T )j dx m0 1+ dx e1+ 2T log R R . 2 ju (x; 0)j dx 2 ju (x; T )j dx . Now, we prove the claim saying that if is convex or star-shaped w.r.t. x0 , then d 1 N (t) N (t) . dt T t+ x0 First, by an integration by parts and by using rG (x; t) = 2(Tx t+ ) G (x; t), we have R R R 2 jruj G dx = ( uuG + ruurG ) dx + @ @ uuG d R x0 = @t u 2(Tx t+ ) ru uG dx . Next, we compute R d dt 2 d dt R jruj G dx = 2 jruj G dx. R 2 2ru@t ruG + jruj @t G dx R R 2 = 2 ( u@t uG + ru@t urG ) dx jruj G dx R R 2 x x0 = 2 j@t uj G dx + @t u T t+ ruG dx R R 2 2 x0 + r jruj rG dx + @ jruj 2(Tx t+ G d . ) But R R 2 2 r jruj rG dx = @i j@j uj @i G dx R 2 = 2 R@j u@ij u@i G dx 2 2 = 2 R @j u@i u@i G + @j u@i u@ij G dx +2 @ @j u@i u j @i G d i h R (xi x0i )(xj x0j ) @j (xi x0i ) + = 2 G dx ururG + @j u@i u 2 2(T t+ ) 4(T t+ ) R 2 x x0 j@ uj (T t+ ) G d @ R R R 2 x x0 x0 ruG dx + (T 1t+ ) jruj G dx 2 = @t u Tx t+ 2(T t+ ) ru R 2 x0 j@ uj (Tx t+ G d . ) @ Therefore, R d dt = G dx 2 jruj G dx R 2 x0 j@t uj G dx + 2 @t u Tx t+ ruG dx R 2 1 + (T t+ ) jruj G dx R R 2 x x0 2 j@ uj (T t+ ) G d + @ jruj 2(Tx @ 2 R 2 2 R x0 t+ ) x x0 2(T t+ ) G d . That is d dt R 2 jruj G dx = 2 R + (T R @ @t u R 1 t+ ) x x0 2(T t+ ) 2 2 ru G dx jruj G dx x0 j@ uj 2(Tx t+ G d . ) 16 2 2 ru G dx Finally, we compute d dt N (t) d dt = = + + + R jruj2 G dx R 1 2 juj G dx) R 1 22 juj2 G dx) ( R R d jruj2 G dx dt juj2 G dx juj2 G dx R 2 2 ( juj G dx) 2 R R x0 @t u 2(Tx t+ 2 ) ru G dx R (T 2 1 2 juj G dx) R ( 2 ( ( R R 1 2 juj G dx) R 2 + (T 1t+ ) N + (t) 1 2 juj2 G dx) R ( = d dt N 2 2 jruj G dx 2 x x0 2(T t+ ) 2 2 R jruj G dx 2 @ (t) R j@ uj @ R 1 22 ( juj G dx) 1 (T t+ ) N (t) R 1 t+ ) R R x x0 2(T t+ ) @t u x x 0 j@ uj2 2(T t+ G d ) R 2 juj G dx x x0 2(T t+ ) @t u 2 juj G dx R 2 G d juj G dx 2 ru G dx R R R 2 ju (x; T )j e 2 Br ju (x; T )j e jx x0 j2 4 jx ju (x; T )j dx + On the other hand, = R R jx 2 2 x0 j ju (x; T )j e 2 2 juj G dx 2 x0 ) 0 for convex . 2 Br R ru uG dx by Cauchy-Schwarz inequality and the fact that (x domain . Step 2 .- Make appear Br 2 juj G dx jx dx R 2 dx + nBr ju (x; T )j e R 2 2 jx x0 j ju (x; T )j e x0 j2 4 1 r2 x0 j 4 jx jx x0 j2 4 x0 j2 4 dx dx . 2 dx jx x0 j2 4 (x x0 ) ju (x; T )j ( 2 ) re dx R R jx x0 j2 jx x0 j2 2 2 4 4 = 2 @ ((x x0 ) ) ju (x; T )j e d + 2 n ju (x; T )j e dx R jx x0 j2 4 +4 (x x0 ) u (x; T ) ru (x; T ) e dx by integration by parts R jx x0 j2 2 4 2 n ju (x; T )j e dx R R jx x0 j2 jx x0 j2 2 2 2 2 1 4 4 16 jru (x; T )j e dx + 21 jx x0 j ju (x; T )j e dx , +2 by Cauchy-Schwarz inequality. Therefore, R jx x0 j2 2 4 ju (x; T )j e dx R 2 ju (x; T )j dx Br R R jx x0 j2 jx x0 j2 2 2 1 4 4 + r2 4 n ju (x; T )j e dx + 16 2 jru (x; T )j e dx 2 R R jx x0 j 2 2 n 4 ju (x; T )j dx + 16 ju (x; T )j e dx . r2 4 + N (T ) Br 17 By step 1, if is convex, R jx x0 j2 2 4 ju (x; T )j e dx R 2 ju (x; T )j dx Br + 16 r2 n 4 +1 Take T m0 R e1+ 2T ju(x;0)j2 dx R ju(x;T )j2 dx 2 r 6 T + 4T 2 + 4 16 in order that Then log 2 = 16 r2 1+ " n +1 4 R ju (x; T )j dx = m0 4 m0 4 T 2 e 1+ R r (n+4)m0 R ju(x;0)j2 dx R ju(x;T )j2 dx 2 ju (x; 0)j dx 2 ju (x; T )j dx 2 2 m0 2T # jx x0 j 4 2 dx . 31=2 7 5 = 1 . 2 m0 R jx x0 j2 2 4 e4 dx ju (x; T )j e m0 R 2 4 ju (x; T )j dx . 2e Br B r @ T + 4T 2 + 4 16 1+ + 1 log e1+ 2T log R But m0 4 n 4 2 2 0 2 ju (x; T )j e T m0 1+ R T 2( n 4 +1 1=2 )log 2 r2 +1 m 1+ 0 R 2T e ju(x;0)j2 dx R ju(x;T )j2 dx m0 e1+ 2T (n + 4) log R 31=2 1 5 C A 1 R ju(x;0)j2 dx ju(x;T )j2 dx . Finally, R 2 2 ju (x; T )j dx m0 R e1+ 2T ju(x;0)j2 dx R ju(x;T )j2 dx That is Z 2 ju (x; T )j dx m 1+ 2T0 e Z 1=2 2 ( n4 +1) r 1+ (n+4)m 0 2 ju (x; 0)j dx 2m0 r2 +1 2 Z Br R 2 Br ju (x; T )j dx . 1 2 ju (x; T )j dx where n 4 +1 1+ r2 (n+4)m0 1=2 = 1+ n 4 +1 1+ r2 (n+4)m0 18 +1 2m0 r2 . 1=2 +1 2m0 r2 , It implies R 2 ju (x; T )j dx m0 R R 2 2 1 ju (x; T )j dx " ju (x; 0)j dx 2e(1+ 2T ) 1 Br " =(1 ) m0 R R 2 2 1 2e(1+ 2T ) ju (x; T )j dx + " ju (x; 0)j dx 8" > 0 , Br " where 1 = n = +1 4 1 r2 1+ (n + 4) m0 ! 1=2 +1 2m0 . r2 Conclusion .- There are two parts: "Make appear Br " and "Get good sign for a derivative". With the frequency function with weight, "Make appear Br " is unusual "Get good sign for a derivative" is more usual. Now, recall that ku ( ; T )kL2 ( C1 C2 e T ku ( ; T )kL2 (!) + " ku ( ; 0)kL2 ( " ) ) 8" > 0 , implies ku ( ; T )kL2 ( Take ! = Br , 3 T ) r then using the fact that r2 whenever 5.3 )2 Z T ku ( ; t)kL2 (!) dt . 0 = , m0 C1 = 2e and C2 = 2 ku ( ; T )kL2 ( + 1 C1 4C2 (1+ T1 )(2+ e + 2 C2 ) C(n) n = +1 4 r2 1+ (n + 4) m0 m0 = max jx x0 j , x2 1=2 ! +1 m20 r2 2 1 1 C(n) m0 (1+ T1 )(1+ m20 )3 r e T m0 Z 0 T ku ( ; t)kL2 (Br ) dt , is convex. The heat equation with space-time potential The above computation can be extended to the heat equation with second member. We consider the heat equation in the solution U = U (x; t) with a suitable second member F = F (x; t) 8 U =F in (0; T ) , < @t U U =0 on @ (0; T ) , : U ( ; 0) 2 L2 ( ) . 19 The following inequality holds (see [PW]). If d N (t) dt T is convex, R 2 jF (x; t)j G (x; t) dx 1 . N (t) + R 2 t+ jU (x; t)j G (x; t) dx Consider the heat equation with potentials a 2 L1 ( n L1 ( (0; T )) . 8 < @t v v + av + b rv = 0 v=0 v ( ; 0) 2 L2 ( ) : in on @ (0; T )) and b 2 (0; T ) , (0; T ) , . Take F = av + b rv. Then d N (t) dt T 1 N (t) + kakL1 ( t+ (0;T )) + kbkL1 ( (0;T )) N (t) , whenever is convex. We are able to reproduce the above treatment to get a re…ned observability estimate (see [PW2]). In order to drop the convexity hypothesis, a local study is necessary (see next section). Take w = u where 2 C01 (BR ) with 0 1 and = 1 in BR=2 . Then 8 < @t w Take F = w= : d N (t) dt 2r ru T 2r ru u w=0 w ( ; 0) 2 L2 (BR ) in BR (0; T ) , on @BR (0; T ) , . u. Then 1 N (t) + 2 t+ We need to bound R R R 2 2 j2r ruj G dx j uj G dx + 2 . R R 2 2 j uj G dx j uj G dx R 2 2 j2r ruj G dx j uj G dx + R R 2 2 j uj G dx j uj G dx by an easy function on time in order to reproduce the above treatment. Here, we can choose t small to do so. Actually, it is not yet done but ideas already may be set up in [EFV]. 20 6 What already exists Here, we recall most of the material from the works of I. Kukavica [Ku2] and L. Escauriaza [E] for the elliptic equation and its application to the heat equation. In the original paper dealing with doubling property and frequency function, N. Garofalo and F.H. Lin [GaL] study the monotonicity property of the following quantity R 2 r B0;r jrv (y)j dy . R 2 jv (y)j d (y) @B0;r However, it seems more natural in our context to consider the monotonicity properties of the frequency function (see [Ze]) de…ned by R 6.1 2 B0;r 2 jrv (y)j r2 jyj R 2 jv (y)j dy B0;r dy . Monotonicity formula Following the ideas of I. Kukavica ([Ku2], [Ku], [KN], see also [E], [AE]), one obtains the following three lemmas. Detailed proofs are given in [Ph3]. Lemma A .Let D RN +1 , N 1, be a connected bounded open set D with yo 2 D and Ro > 0. If v = v (y) 2 H 2 (D) is a such that Byo ;Ro solution of y v = 0 in D, then (r) = R 2 Byo ;r jrv (y)j R Byo ;r r2 2 jy yo j 2 dy jv (y)j dy is non-decreasing on 0 < r < Ro , and Z d 1 2 ln jv (y)j dy = (N + 1 + dr r Byo ;r (r)) . Lemma B .Let D RN +1 , N 1, be a connected bounded open set such that Byo ;Ro D with yo 2 D and Ro > 0. Let r1 , r2 , r3 be three real numbers such that 0 < r1 < r2 < r3 < Ro . If v = v (y) 2 H 2 (D) is a solution of y v = 0 in D, then Z Byo ;r2 2 jv (y)j dy Z Byo ;r1 2 jv (y)j dy 21 ! Z Byo ;r3 2 jv (y)j dy !1 , 1 where = 1 r2 r1 ln 1 ln r2 r1 + 1 ln 2 (0; 1). r3 r2 The above two results are still available when we are closed to a part of the boundary @ under the homogeneous Dirichlet boundary condition on , as follows. Lemma C .Let D RN +1 , N 1, be a connected bounded open set with boundary @D. Let be a non-empty Lipschitz open subset of @D. Let ro , r1 , r2 , r3 , Ro be …ve real numbers such that 0 < r1 < ro < r2 < r3 < Ro . Suppose that yo 2 D satis…es the following three conditions: i). Byo ;r \ D is star-shaped with respect to yo 8r 2 (0; Ro ) , ii). Byo ;r D 8r 2 (0; ro ) , iii). Byo ;r \ @D 8r 2 [ro ; Ro ) . If v = v (y) 2 H 2 (D) is a solution of y v = 0 in D and v = 0 on , then Z Z 2 Byo ;r2 \D jv (y)j dy 2 Byo ;r1 jv (y)j dy ! Z 2 Byo ;r3 \D jv (y)j dy !1 1 where 6.1.1 = 1 1 ln r2 r1 ln r2 r1 + 1 ln r3 r2 2 (0; 1). Proof of Lemma B Let H (r) = Z 2 Byo ;r jv (y)j dy . By applying Lemma A, we know that d 1 lnH (r) = (N + 1 + dr r (r)) . Next, from the monotonicity property of , one deduces the following two inequalities Rr (r) 2) ln H(r = r12 N +1+ dr H(r1 ) r (N + 1 + (r2 )) ln rr21 , Rr (r) 3) ln H(r = r23 N +1+ dr H(r2 ) r (N + 1 + (r2 )) ln rr32 . Consequently, ln H(r2 ) H(r1 ) ln rr12 ln (N + 1) + 22 (r2 ) H(r3 ) H(r2 ) ln rr32 , and therefore the desired estimate holds 1 H (r2 ) (H (r1 )) (H (r3 )) , 1 1 1 1 where = 6.1.2 Proof of Lemma A ln r2 r1 ln + r2 r1 . r3 r2 ln We introduce the following two functions H and D for 0 < r < Ro : R 2 H (r) = By ;r jv (y)j dy , o R 2 2 D (r) = jrv (y)j r2 jy yo j dy . By ;r o RrR 2 First, the derivative of H (r) = 0 S N jv ( s + yo )j N d d (s) is given by R 2 H 0 (r) = @By ;r jv (y)j d (y). Next, recall the Green formula o R = R jvj @ Gd (y) 2 Byo ;r jvj Gdy @ 1 r Byo ;r o yv H 0 (r) H(r) = N +1 r + jvj Gdy . 2 r2 jvj jy r 2 2 jy yo j dy jy yo j 2 2 yo j 2r, dy dy . = 0, H 0 (r) = that is Gd (y) yo j where Gj@Byo ;r = 0, @ Gj@Byo ;r = Byo ;r Consequently, when 2 jvj 2 R 2 1 (N + 1) jvj dy + 2r Byo ;r R N +1 1 = r H (r) + r By ;r div (vrv) r2 R o 2 = N r+1 H (r) + 1r By ;r jrvj + v v = R @Byo ;r R 2 We apply it with G (y) = r2 jy and G = 2 (N + 1). It gives H 0 (r) R 2 @Byo ;r 1 D(r) r H(r) N +1 1 H (r) + D (r) , r r (A.1) the second equality in Lemma A. Now, we compute the derivative of D (r). D0 (r) = d dr R r2 RrR 0 2 SN (rv)j s+yo N d d (s) 2 r2 (rv)jrs+yo rN d (s) 2 RrR N = 2r 0 S N (rv)j s+yo d d (s) R 2 = 2r By ;r jrvj dy . SN o 23 (A.2) On the other hand, we have by integrations by parts that R R 2 2 2r By ;r jrvj dy = N r+1 D (r) + 4r By ;r j(y yo ) rvj dy o o R 2 1 yo ) v r2 jy yo j dy . r By ;r rv (y (A.3) o Therefore, R 2 2 (N + 1) By ;r jrvj r2 jy yo j dy o R R 2 2 = 2r2 By ;r jrvj dy 4 By ;r j(y yo ) rvj dy o R o 2 +2 By ;r (y yo ) rv v r2 jy yo j dy , o and this is the desired estimate (A.3). Consequently, from (A.2) and (A.3), we obtain, when y v = 0, the following formula Z N +1 4 2 D0 (r) = D (r) + j(y yo ) rvj dy . (A.4) r r Byo ;r The computation of the derivative of 0 (r) = (r) = 1 [D0 (r) H (r) H 2 (r) D(r) H(r) gives D (r) H 0 (r)] , which implies using (A.1) and (A.4) that Z 1 2 H 2 (r) 0 (r) = 4 j(y yo ) rvj dyH (r) r Byo ;r 2 ! D (r) 0, indeed, thanks to an integration by parts and using Cauchy-Schwarz inequality, we have 2 R D2 (r) = 4 By ;r vrv (y yo ) dy R o R 2 2 4 By ;r j(y yo ) rvj dy jvj dy Byo ;r o R 2 4 By ;r j(y yo ) rvj dy H (r) . o Therefore, we have proved the desired monotonicity for the proof of Lemma A. 6.1.3 and this completes Proof of Lemma C Under the assumption Byo ;r \ @D for any r 2 [ro ; Ro ), we extend v by zero in Byo ;Ro nD and denote by v its extension. Since v = 0 on , we have 8 in Byo ;Ro , < v = v1D v=0 on Byo ;Ro \ @D , : rv = rv1D in Byo ;Ro . 24 Now, we denote r = Byo ;r \D, when 0 < r < Ro . In particular, when 0 < r < ro . We introduce the following three functions: R 2 H (r) = jv (y)j dy , r R 2 2 D (r) = jrv (y)j r2 jy yo j dy , r = Byo ;r , r and D (r) 0. H (r) Our goal is to show that is a non-decreasing function. Indeed, we will prove that the following equality holds d 1 d lnH (r) = (N + 1) lnr + (r) . (C.1) dr dr r Therefore, from the monotonicity of , we will deduce (in a similar way than in the proof of Lemma A) that (r) = ln H(r2 ) H(r1 ) ln rr21 ln (N + 1) + (r2 ) and this will imply the desired estimate ! Z Z 2 2 jv (y)j dy jv (y)j dy Byo ;r1 r2 Z H(r3 ) H(r2 ) ln rr32 , 2 r3 jv (y)j dy !1 , 1 where = 1 ln 1 r2 r1 ln r2 r1 + 1 ln . r3 r2 R 2 First, we compute the derivative of H (r) = By ;r jv (y)j dy. o R 2 H 0 (r) = S N jv (rs + yo )j rN d (s) R 2 1 = r S N jv (rs + yo )j rs srN d (s) R 2 = 1r By ;r div jv (y)j (y yo ) dy o R 2 2 = 1r By ;r (N + 1) jv (y)j + r jv (y)j (y yo ) dy o R = N r+1 H (r) + 2r r v (y) rv (y) (y yo ) dy . Next, when yv (C.2) = 0 in D and vj = 0, we remark that Z D (r) = 2 v (y) rv (y) (y yo ) dy , (C.3) r indeed, R 2 2 jrvj r2 jy yo j dy h i h R R 2 = r div vrv r2 jy yo j dy vdiv rv r2 r R R 2 = v v r2 jy yo j dy vrv r r2 jy r r because on @Byo ;r , r = jy yo j and vj = 0 R = 2 r vrv (y yo ) dy because y v = 0 in D . r 25 jy 2 2 yo j yo j dy i dy Consequently, from (C.2) and (C.3), we obtain H 0 (r) = N +1 1 H (r) + D (r) , r r (C.4) and this is (C.1). On another hand, the derivative of D (r) is D0 (r) = 2r = 2r Here, when 2r R yv R0 2 SN (rv)j N d d (s) s+yo (C.5) 2 jrv (y)j dy . r = 0 in D and vj = 0, we will remark that 2 r RrR R N +1 4 r D (r) + r Byo ;r R 2 + 1r \By ;r j@ vj r2 o jrv (y)j dy = indeed, 2 j(y yo ) rv (y)j dy jy 2 yo j (y yo ) d (y) (C.6) R 2 2 (N + 1) r jrvj r2 jy yo j dy R 2 2 = r div jrvj r2 jy yo j (y yo ) dy R 2 2 r jrvj r2 jy yo j (y yo ) dy r R 2 2 = \By ;r jrvj r2 jy yo j (y yo ) d (y) R o 2 2 @yi jrvj r2 jy yo j (yi yoi ) dy r R 2 2 = \By ;r jrvj r2 jy yo j (y yo ) d (y) R o 2 2rv@yi rv r2 jy yo j (yi yoi ) dy Rr 2 2 +2 r jrvj jy yo j dy , R and = + + + = R R R R R R r @yj v@y2i yj v r2 r @yj (yi r @yj (yi r r 2 jy yo j yoi ) @yj v@yi v r yoi ) @yj v@yi v r2 (yi yoi ) @y2j v@yi v (yi yoi ) @yj v@yi v@yj r2 \Byo ;r j (yi 2 (yi r 2 jy 2 2 26 2 jy jy yo j 2 jy 2 yo j 2 yo j yoi ) @yj v@yi v r + r jrvj r2 jy yo j dy +0 because y v = 0 in D R 2 2 j(y yo ) rvj dy . r yoi ) dy dy dy dy 2 yo j dy jy yo j 2 d (y) Therefore, when = yv (N + 1) R \Byo ;r 2 R = 0 in D, we have R 2 r2 jrvj r 2 r2 jrvj 2 jy jy yo j 2 yo j @yj v j ((yi R\Byo ;r 2 R 2 +2r jruj dy 4 r dy (y yo ) yoi ) @yi v) r r 2 j(y = +2r R 2 r 2 \Byo ;r R 2 r r2 jrvj j@ vj r 2 jrvj dy 4 R jy r ) 2 jy 2 2 jy yo ) rvj dy . By using the fact that vj = 0, we get rv = (rv (N + 1) R d (y) 2 yo j on and we deduce that (y yo ) d (y) 2 j(y yo ) rvj dy , and this is (C.6). Consequently, from (C.5) and (C.6), when vj = 0, we have D0 (r) = N +1 4 r D (r) + r R + 1r \By ;r j@ o R r 2 vj j(y r2 (r) = yv = 0 in D and 2 yo ) rv (y)j dy 2 jy The computation of the derivative of 0 d (y) dy 2 yo j yo j yo j (y yo ) (r) = D(r) H(r) gives 1 [D0 (r) H (r) H 2 (r) d (y) . (C.7) D (r) H 0 (r)] , which implies from (C.4) and (C.7), that H 2 (r) 0 (r) 1 r = 4 R + H(r) r r R j(y 2 yo ) rv (y)j dy H (r) \Byo ;r 2 j@ vj r2 jy 2 yo j D2 (r) (y yo ) d (y) Thanks to (C.3) and Cauchy-Schwarz inequality, we obtain that 0 4 Z r j(y 2 yo ) rv (y)j dy H (r) D2 (r) . The inequality 0 (y yo ) on holds when Byo ;r \ D is star-shaped with respect to yo for any r 2 (0; Ro ). Therefore, we get the desired monotonicity for which completes the proof of Lemma C. 27 6.2 Quantitative unique continuation property for the Laplacian Let D RN +1 , N 1, be a connected bounded open set with boundary @D. Let be a non-empty Lipschitz open part of @D. We consider the Laplacian in D, with a homogeneous Dirichlet boundary condition on @ : 8 in D , < yv = 0 v=0 on , (D.1) : v = v (y) 2 H 2 (D) . The goal of this section is to describe interpolation inequalities associated to solutions v of (D.1). Theorem D .Let ! be a non-empty open subset of D. Then, for any D, there exist C > 0 D1 D such that @D1 \ @D b and D1 n( \ @D1 ) and 2 (0; 1) such that for any v solution of (D.1), we have Z 2 D1 jv (y)j dy Z C ! 2 jv (y)j dy Z 1 2 jv (y)j dy D . Or in a equivalent way, Theorem D’.Let ! be a non-empty open subset of D. Then, for any D1 D such that @D1 \ @D b and D1 n( \ @D1 ) D, there exist C > 0 and 2 (0; 1) such that for any v solution of (D.1), we have Z D1 Z 1 2 jv (y)j dy C 1 " ! 2 jv (y)j dy + " Z D 2 jv (y)j dy 8" > 0 . Proof of Theorem D .- We divide the proof into two steps. Step 1 .- We apply Lemma B, and use a standard argument (see e.g., [Ro]) which consists to construct a sequence of balls chained along a curve. More precisely, we claim that for any non-empty compact sets in D, K1 and K2 , such that meas(K1 ) > 0, there exists 2 (0; 1) such that for any v = v (y) 2 H 2 (D), solution of y v = 0 in D, we have Z K2 2 jv (y)j dy Z K1 2 jv (y)j dy Z D 1 2 jv (y)j dy . (D.2) Step 2 .- We apply Lemma C, and choose yo in a neighborhood of the part such that the conditions i, ii, iii, hold. Next, by an adequate partition of D, we deduce from (D.2) that for any D1 D such that @D1 \ @D b 28 and D1 n( \ @D1 ) D, there exist C > 0 and 2 (0; 1) such that for any v = v (y) 2 H 2 (D) such that y v = 0 on D and v = 0 on , we have Z D1 2 jv (y)j dy C Z Z 2 ! jv (y)j dy 1 2 D jv (y)j dy . This completes the proof. 6.3 Quantitative unique continuation property for the elliptic operator @t2 + In this section, we present the following result. Theorem E .Let be a bounded open set in Rn , n 1, either convex 2 or C and connected. We choose T2 > T1 and 2 (0; (T2 T1 ) =2). Let f 2 L2 ( (T1 ; T2 )). We consider the elliptic operator of second order in (T1 ; T2 ) with a homogeneous Dirichlet boundary condition on @ (T1 ; T2 ), 8 2 in (T1 ; T2 ) , < @t w + w = f w=0 on @ (T1 ; T2 ) , (E.1) : w = w (x; t) 2 H 2 ( (T1 ; T2 )) . Then, for any ' 2 C01 ( (T1 ; T2 )), ' 6= 0, there exist C > 0 and such that for any w solution of ( E.1), we have R T2 T1 + C R 2 R T2 R T1 R T2 T1 2 (0; 1) R jw (x; t)j dxdt 1 2 jw (x; t)j dxdt 2 j'w (x; t)j dxdt + R T2 R T1 2 jf (x; t)j dxdt . Proof .- First, by a di¤erence quotient technique and a standard extension at fT1 ; T2 g, we check the existence of a solution u 2 H 2 ( (T1 ; T2 )) solving @t2 u + u=0 u=f in on @ (T1 ; T2 ) , (T1 ; T2 ) [ fT1 ; T2 g , such that kukH 2 ( (T1 ;T2 )) c kf kL2 ( (T1 ;T2 )) , for some c > 0 only depending on ( ; T1 ; T2 ). Next, we apply Theorem D with D = (T1 ; T2 ), (T1 + ; T2 ) D1 , y = (x; t), y = @t2 + , and v = w u. 29 6.4 The heat equation and the Hölder continuous dependence from one point in time In this section, we presents the following result. Theorem F .or C 2 and connected. there are C > 0 and 8 < Let be a bounded open set in Rn , n 1, either convex Let ! be a nonempty open subset of , and T > 0. Then 2 (0; 1) such that any solution to @t u u=0 in u=0 on @ : u ( ; 0) 2 L2 ( ) , satis…es, for any to 2 (0; T ), Z 2 ju (x; to )j dx C C e to Z (0; T ) , (0; T ) , Z 1 2 ju (x; 0)j dx ! (F.1) 2 ju (x; to )j dx . Proof .- We divide the proof into three steps. Step 1 .- Let 1 , 2 , and e1 ,e2 , be the eigenvalues and eigenfunctions 1 of the Laplacian in H ( ), constituting an orthonormal basis in L2 ( ). 0 R P 2 For any uo = u ( ; 0) = uo ej dx, the solution j ej in L ( ) with j = j 1 P jt u to (F.1), can be written as u (x; t) = . Let to 2 (0; T ). We j ej (x) e j 1 introduce (see [L] or [CRV]) the function X p j to w (x; t) = ch j ej (x) e jt (x) j 1 where solves X j ej (x) e j to , j 1 2 C01 (!), = 1 in ! e b !. Recall that cht = (et + e t ) =2. Then, w 8 2 in (0; T ) , @ w+ w= f > > < t w=0 on @ (0; T ) , w = @ w = 0 in ! e f0g , > t > : w = w (x; t) 2 H 2 ( (0; T )) , where f = u ( ; to ) 2 H02 ( ). We denote by w the extension of w 8 2 < @t w + w = f 1 (0;T ) w=0 : w=0 to zero in ! e in in ! e on @ ( T; 0). Then, w solves (0; T ) [ ! e ( T; 0) , (0; T ) . ( T; 0) , Now, we construct D, an open connected set in RN +1 , satisfying the following six conditions: 30 i). ( ;T ) D with 2 (0; T =2) ; ii). @ ( ;T ) @D ; iii). D (0; T ) [ ! e ( T; 0) ; iv). there is an nonempty open ! o b D \ ! e ( To ; 0) with To 2 (0; T ) ; v). D 2 C 2 when is of class C 2 and connected ; vi). D is convex with an adequate choice of the pair ( ; To ) when is convex . In particular, w 2 H 2 (D). Step 2 .- We claim that there is g 2 H 2 ( H (D) such that 2 @t2 g + g=0 g= f1 (0;T ) ( T; T )) \ H01 ( in ( T; T ) , on @ ( ( T; T )) = @ ( T; T )) ( T; T ) [ f Tg , and kgkL2 (D) kf kL2 ( (0;T )) . (F.2) Indeed, we will proceed in six substep when is of class C 2 and connected (the case where is convex is well-known because ( T; T ) is then convex). We de…ne h = f 1 (0;T ) 2 L2 ( ( T; T )). Substep 1: recall that h 2 L2 ( ( T; T )) implies the existence of such g 2 H01 ( ( T; T )). Substep 2: thanks to the interior regularity theorem of elliptic systems, for any D0 b ( T; T ), g 2 H 2 (D0 ). Substep 3: thanks to the boundary regularity theorem of elliptic systems, but far for f T; T g, g is locally in H 2 because is of class C 2 . Substep 4: we extend the solution at t = T as follows. Let h (x; t) = h (x; t) when (x; t) 2 ( T; T ) and h (x; t) = h (x; 2T t) when (x; t) 2 (T; 3T ). Then h 2 L2 ( ( T; 3T )). Let g (x; t) = g (x; t) when (x; t) 2 [ T; T ) and g (x; t) = g (x; 2T t) when (x; t) 2 [T; 3T ]. Then, g solves @t2 g + g=0 g=h in ( T; 3T ) , on @ ( ( T; 3T )) = @ ( T; 3T ) [ f T; 3T g . By applying the boundary regularity theorem of elliptic systems as in substep 3, we get g 2 H 2 ( (0; 2T )). In particular, g 2 H 2 ( (0; T )). Substep 5: we extend in a similar way the solution at t = T in order to conclude that g 2 H 2 ( ( T; 0)). 1 Substep 6: Finally, we multiply @t2 g + g = h by ( ) g and integrate by parts over ( T; T ), to get RT 2 2 k@t gkH 1 ( ) dt + kgkL2 ( ( T;T )) R TTR 1 = 0 f (x) ( ) g (x; t) dxdt RT R 1 = 0 f (x) ( )( ) g (x; t) dxdt because f 2 H02 ( ) kf kL2 ( (0;T )) kgkL2 ( ( T;T )) by Cauchy-Schwarz . 31 This gives the desired inequality (F.2). Step 3 .- Finally, we apply Theorem D with y = @t2 + , v = w where = @ ( ;T ) and ( =2 + T =4; 3T =4 =2) D1 that @D1 \ @D b and D1 n( \ @D1 ) D in order that Z Z 1 D1 2 jw gj dy 1 " C !o 2 jw gj dy + " Z D 2 gj dy jw g in D D such 8" > 0 , which implies Z Z 1 2 D1 jwj dy C 1 " 2 D jgj dy + " Z 2 D jwj dy 8" 2 (0; 1) , where we used the fact that w = 0 in ! o . By (F.2), we conclude that there exist C > 0 and 2 (0; 1) such that R 3T =4 =2 =2+T =4 R 2 jw (x; t)j dxdt Consequently, there exist RT R 1 1 " C RT R +" 0 RT R 0 2 jf (x)j dxdt 2 jw (x; t)j dxdt 2 (0; T =2), C > 0 and 8" > 0 . 2 (0; 1) such that 1 RT R 2 jf (x)j dxdt C 1" RT R 0 2 +" 0 jw (x; t)j dxdt 8" > 0 . 2 jw (x; t)j dxdt On the other hand, the following four inequalities hold. Z T 0 Z T 0 Z Z 2 jw (x; t)j dxdt P j 1 2( 2 je 2 jf (x)j dxdt j to 2T X T Z ! 2( 2 je j to 2 j (x) u (x; to )j dx , p jT ) + 2T jT ) 2 ! j 1 p Z = j (x) u (x; to )j dx , p P 2( j to 2 jT ) e j p T fj 1; j to g p P 2( j to 2 jT ) + e j p T fj 1; j > to g 2 2 Tto P 2 e j , j 1 RT R P j 1 j ej (x) e j to ch p 2 jt dxdt 2 RT +2T 32 R ! R 2 jw (x; t)j dxdt 2 j (x) u (x; to )j dx , We deduce from the above …ve inequalities that for any " > 0, P 2 2 j to P 2 2( j to p j ) (T 2 ) e (T 2 ) e j 1 j j j 1 RT 2T C R 1 " P j ej j 1 1 T2 R (x) e j to ch 2 p 2 jt dxdt j (x) u (x; to )j dx ! P 2 R 2 j + ! j (x) u (x; to )j dx ! +4T " e2 to j 1 R 2 +2T ! j (x) u (x; to )j dx . Finally, there is C > 0 such that for any to > 0, R 2 ju (x; to )j dx = P 2 2 je j 1 C 1 " j to R 1 C 2 ! ju (x; to )j dx + "e to which implies the desired estimate of Theorem F, Z 2 ju (x; to )j dx C e C(1 to ) Z 1 2 ju (x; 0)j dx R 2 ju (x; 0)j dx 8" 2 (0; 1) , Z ! 2 ju (x; to )j dxdt . References [AE] V. Adolfsson and L. Escauriaza, C 1; domains and unique continuation at the boundary, Comm. Pure Appl. Math. 50 (1997) 935-969. [BT] C. Bardos, L. Tartar, Sur l’unicité retrograde des équations paraboliques et quelques questions voisines, Arch. Rational Mech. Anal. 50 (1973) 10-25. [CRV] B. Canuto, E. Rosset, S. Vessella, Quantitative estimates of unique continuation for parabolic equations and inverse initial-boundary value problems with unknown boundaries, Trans. Amer. Math. Soc. 354 (2001) 491535. [E] L. Escauriaza, Doubling property and linear combinations of eigenfunctions, manuscript. [EFV] L. Escauriaza, F.J. Fernandez, S. Vessella, Doubling properties of caloric functions, Applicable analysis 85 (1-2) (2006) 205-223. 33 [FI] A.V. Fursikov, O.Yu. Imanuvilov, Controllability of evolution equations. Lecture Notes Series, 34. Seoul National University, Research Institute of Mathematics, Global Analysis Research Center, Seoul, 1996. [FEF] X-L. Feng, L. Eldén, C-L Fu, Stability and regularization of a backward parabolic PDE with variable coe¢ cients, Journal of Inverse and Ill-Posed Problems 18 (2010) 217-243. [GaL] N. Garofalo and F H. Lin, Monotonicity properties of variational integrals, Ap -weights and unique continuation, Indiana Univ. Math. J. 35 (1986) 245-268. [K] C. Kenig, Quantitative unique continuation, logarithmic convexity of gaussian means and Hardy’s uncertainty principle, Proc. of Symp. of Pure Math., volume in celebration of V. Mazya’s 70 birthday, 79, 2008, p.207-227. [KT] H. Koch, D. Tataru, Carleman estimates and unique continuation for second order parabolic equations with non smooth coe¢ cients, Comm. Partial Di¤erential Equations 34 (4) (2009) 305-366. [Ku] I. Kukavica, Level sets for the stationary Ginzburg-Landau equation, Calc. Var. 5 (1997) 511-521. [Ku2] I. Kukavica, Quantitative uniqueness for second order elliptic operators, Duke Math. J. 91 (1998) 225-240. [KN] I. Kukavica, K. Nyström, Unique continuation on the boundary for Dini domains, Proc. of Amer. Math. Soc. 126 (1998) 441-446. [Le] M. Léautaud, Spectral inequalities for non-selfadjoint elliptic operators and application to the null-controllability of parabolic systems, Journal of Functional Analysis 258 (2010) 2739–2778. [LR] G. Lebeau, L. Robbiano, Contrôle exact de l’équation de la chaleur, Comm. Part. Di¤. Eq. 20 (1995) 335-356. [Li] F.H. Lin, Remarks on a backward parabolic problem, Methods and Applications of Analysis 10 (2) (2003) 245-252. [L] F.H. Lin, A uniqueness theorem for the parabolic equations, Comm. Pure Appl. Math. 43 (1990) 127-136. [M] L.Miller, A direct Lebeau-Robbiano strategy for the observability of heatlike semigroups, Discrete and Continuous Dynamical Systems Series B 14 (4) (2010) 1465-1485. [PW] K. D. Phung , G. Wang , Quantitative unique continuation for the semilinear heat equation in a convex domain, Journal of Functional Analysis 259 (5) (2010) 1230-1247. 34 [PW2] K. D. Phung , G. Wang, An observability for parabolic equations from measurable set in time and its applications, in preparation. [P] C.C. Poon, Unique continuation for parabolic equations, Comm. Partial Di¤erential Equations 21 (3-4) (1996) 521-539. [Ph3] K.-D. Phung, Observation et stabilisation d’ondes: géométrie et coût du contrôle, Habilitation à diriger des recherches, (2007). [Ro] L. Robbiano, Théorème d’unicité adapté au contrôle des solutions des problèmes hyperboliques, Comm. Part. Di¤. Eq. 16 (1991) 789-800. [W] G. Wang, L1 -null controllability for the heat equation and its consequences for the time optimal control problem, SIAM J. Control Optim. 47 (2008) 1701-1720. [Ze] S. Zelditch, Local and global analysis of eigenfunctions on Riemannian manifolds, arXiv:09033420. 35