arXiv:2005.14503v2 [math.FA] 21 Jan 2021
Observability and null-controllability for parabolic equations in L p -spaces
Clemens Bombach1, Dennis Gallaun2, Christian Seifert2,3, and Martin Tautenhahn1
1Technische Universit¨at Chemnitz, Fakult¨at f¨ur Mathematik, 09107 Chemnitz, Germany
2Technische Universit¨at Hamburg, Institut f¨ur Mathematik, 21073 Hamburg, Germany
3Technische Universit¨at Clausthal, Institut f¨ur Mathematik, 38678 Clausthal-Zellerfeld, Germany
Abstract
We study (approximate) null-controllability of parabolic equations inLp(Rd) and provide explicit bounds on the control cost. In particular we consider systems of the form ˙x(t) = −Apx(t) +1Eu(t), x(0) = x0 ∈ Lp(Rd), with interior control on a so-called thick set E ⊂ Rd, where p ∈ [1,∞), and where A is an elliptic operator of orderm∈NinLp(Rd). We prove null-controllability of this system via duality and a sufficient condition for observability. This condition is given by an uncertainty principle and a dissipation estimate. Our result unifies and generalizes earlier results obtained in the context of Hilbert and Banach spaces. In particular, our result applies to the casep= 1.
Mathematics Subject Classification (2020). 47D06, 35Q93, 47N70, 93B05, 93B07.
Keywords. Null-controllability, Banach space, non-reflexive,C0-semigroups, elliptic operators, observability estimate,Lp-spaces.
1. Introduction
We consider parabolic control systems on Lp(Rd),p∈[1,∞), of the form
˙
x(t) =−Apx(t) +1Eu(t), t∈(0, T], x(0) =x0 ∈Lp(Rd), (1) where −Ap is a strongly elliptic differential operator of order m ∈ N with constant coefficients, 1E: Lp(E) → Lp(Rd) is the embedding from a measurable set E ⊂ Rd to Rd,T > 0, and where u∈Lr((0, T);Lp(E)) with somer ∈[1,∞]. Hence, the influence of the control functionuis restricted to the subsetE. Note that we allow for lower order terms in the strongly elliptic differential operator. The focus of this paper is laid on null- controllability, that is, for any initial condition x0 ∈Lp(Rd) there is a control function u∈Lr((0, T);Lp(E)) such that the mild solution of (1) at timeT equals zero. We will also be concerned with the notion of approximate null-controllability, which means that for any ε >0 and any x0 ∈Lp(Rd) we can find a control functionu ∈Lr((0, T);Lp(E)) such that the mild solution of (1) at timeT has norm smaller thanε; in reflexive spaces,
these two notions agree (see [Car88]). By linearity, (approximate) null-controllability implies that any target state in the range of the semigroup generated by −Ap can be reached (up to an errorε) within timeT.
We will show that ifE is a so-called thick set, then the system is approximately null- controllable for p= 1 and null-controllable if p∈(1,∞). Note that the casep ∈(1,∞) is already covered by [GST20]. However, our new proof unifies the cases p = 1 and p ∈ (1,∞). Moreover, we provide explicit upper bounds on the control cost, i.e. on the norm of the control function u which steers the system (approximately) to zero at time T, which are explicit in terms of geometric properties of the thick set E and of the final time T. Controllability for systems on Lp(Ω), where Ω is a bounded domain and p∈[1,∞), has been studied earlier in the literature, for instance in [FPZ95] in the context of semilinear heat equations, which include as a special case approximate null- controllability for the linear heat equation. For further results in this direction, we refer to [FCZ00] and the survey article [Zua06]. In comparison, we obtain (approximate) null- controllability of linear differential operators of higher orders withu∈Lr((0, T);Lp(Rd)) where r∈[1,∞]. Note that from the physical point of view the case p= 1 is probably most interesting, since we can then interpret the states as heat densities (and its norms will be the total heat content).
An equivalent formulation of approximate null-controllability is final-state observabil- ity of the adjoint system to (1). This means that there is a constant Cobs≥0 such that for all ϕ∈Lp(Rd)′ we have
ST′ ϕ
Lp(Rd)′ ≤
CobsRT
0 k(St′ϕ)|EkrL′p(E)′dt1/r′
if r′∈[1,∞), Cobsess sup
t∈[0,T]
k(St′ϕ)|EkLp(E)′ if r′=∞,
where (St)t≥0 is theC0-semigroup generated by−Ap and r′∈[1,∞] is such that 1/r+ 1/r′ = 1. This equivalence follows from Douglas’ lemma, see [Dou66] for Hilbert spaces, and [Emb73,DR77,Har78,CP78,Car85,Car88,For14] for Banach spaces.
In Section 2 we formulate our results on final-state observability in Theorem 2.2 and (approximate) null-controllability in Theorem2.3. The proof of Theorem2.2rests on an abstract observability estimate stated in the appendix (see TheoremA.1) and is provided in Section3.
The main strategy we follow to prove observability has first been described in [LR95, LZ98, JL99] for the Hilbert space case (i.e. p = r = 2), and further studied, e.g., in [Mil10,TT11,WZ17,BPS18,NTTV20]. However, far less is known on its generalization to Banach spaces; to the best of our knowledge, we are only aware of [GST20]. Note that strong continuity of the semigroup (St)t≥0 is assumed there. However, being interested in approximate null-controllability in L1 requires observability in L∞, and there strong continuity of semigroups is rather rare [Lot85]. Theorem A.1 provides a generalization of [GST20] to not necessarily strongly continuous semigroups.
2. Observability and Null-controllability in Lp-Spaces
In order to formulate our main theorems we review some basic facts from Fourier analysis.
For details we refer, e.g., to the textbook [Gra14]. We denote by S(Rd) the Schwartz space of rapidly decreasing functions, which is dense in Lp(Rd) for all p ∈[1,∞). The space of tempered distributions, i.e. the topological dual space of S(Rd), is denoted by S′(Rd). Forf ∈ S(Rd) letFf:Rd→Cbe the Fourier transform of f defined by
Ff(ξ) :=
Z
Rd
f(x)e−iξ·xdx.
ThenF:S(Rd)→ S(Rd) is bijective, continuous and has a continuous inverse, given by F−1f(x) = 1
(2π)d Z
Rd
f(ξ)eix·ξdξ
for all f ∈ S(Rd). For u ∈ S′(Rd) the Fourier transform is again denoted by F and is given by (Fu)(φ) =u(Fφ) for φ∈ S(Rd). By duality, the Fourier transform is bijective on S′(Rd) as well.
Let m∈Nand
a(ξ) = X
|α|1≤m
aαξα, ξ∈Rd,
be a polynomial of degree m with coefficients aα ∈ C. We say that the polynomial a is strongly elliptic if there exist constants c > 0 and ω ∈ R such thata satisfies for all ξ∈Rdthe lower bound
Rea(ξ)≥c|ξ|m−ω. (2)
Note that strong ellipticity implies that m is even.
Given a strongly elliptic polynomial a and p ∈[1,∞], we define the associated heat semigroupS : [0,∞)→ L(Lp(Rd)) by
Stf =F−1e−taFf =F−1e−ta∗f. (3) Note that the second equality holds since e−ta∈ S(Rd). It is well known that the operator semigroup (St)t≥0 is strongly continuous if p∈[1,∞). For p=∞ the semigroup is the dual semigroup of a strongly continuous semigroup onL1(Rd) and hence it is only weak*- continuous in general. For details we refer, e.g., to [Are04]. By [TR96], the integral kernel kt = F−1e−ta satisfies the following heat kernel estimate: There exist c1, c2 > 0 such that for allx∈Rd and t >0 we have
|kt(x)| ≤c1eωtt−d/me−c2
|x|m t
m−11
. (4)
This implies that there is M ≥1 and ω∈R such that for allt≥0 we have
kStk ≤Meωt, t≥0. (5) In order to formulate our main result we introduce the notion of a thick subsetE ofRd.
Definition 2.1. Letρ∈(0,1] andL∈(0,∞)d. A set E⊂Rdis called (ρ, L)-thick if E is measurable and for all x∈Rd we have
E∩
×
d i=1(0, Li) +x
!
≥ρ
d
Y
i=1
Li.
Here,|·|denotes Lebesgue measure in Rd.
The following theorem yields a final-state observability estimate for (St)t≥0 on thick sets.
Theorem 2.2. Let m∈N, a:Rd→C a strongly elliptic polynomial of order m, c >0 and ω ∈ R as in (2), and (St)t≥0 as in (3). Let ρ ∈ (0,1], L ∈ (0,∞)d, E ⊂ Rd a (ρ, L)-thick set, p, r∈[1,∞], and T >0. Then we have for all f ∈Lp(Rd)
kSTfkLp(Rd)≤
Cobs
Z T 0
k(Stf)|EkrLp(E)dt 1/r
if r∈[1,∞), Cobsess sup
t∈[0,T]
k(Stf)|EkLp(E) if r=∞, where
Cobs = Ka
T1/r Kd
ρ
Kd(1+|L|1λ∗)
exp Km(|L|1ln(Kd/ρ))m/(m−1)
(cT)1/(m−1) +Kmax{ω,0}T
! .
Here, λ∗ = (2m+4max{ω,0}/c)1/m, K >0is an absolute constant, and Ka, Kd, Km >0 are constants depending only on the polynomial a, on d, or onm, respectively.
By duality, we thus obtain (approximate) null-controllability for (1).
Theorem 2.3. Let m∈N, a:Rd→C a strongly elliptic polynomial of order m, c >0 and ω ∈ R as in (2), and (St)t≥0 as in (3). Let ρ ∈ (0,1], L ∈ (0,∞)d, E ⊂ Rd a (ρ, L)-thick set, r∈[1,∞], and T >0.
(a) For any f ∈L1(Rd) and any ε >0 there exits u∈Lr((0, T);L1(E))with kukLr((0,T);L1(E))≤CobskfkL1(Rd)
such that
STf + Z T
0
ST−t1Eu(t)dt L1(Rd)
< ε.
(b) Letp∈(1,∞). Then for any f ∈Lp(Rd) there exits u∈Lr((0, T);Lp(E)) with kukLr((0,T);Lp(E))≤CobskfkLp(Rd)
such that
STf + Z T
0
ST−t1Eu(t)dt= 0.
Here, Cobs is as in Theorem 2.2 with r replaced byr′ where r′ ∈[1,∞] such that 1/r+ 1/r′ = 1.
Remark 2.4 (Discussion on observability and null-controllability). Forp∈[1,∞) let−Ap be the generator of theC0-semigroup (St)t≥0 onLp(Rd). Note that for allf ∈ S(Rd) we have
Apf = X
|α|1≤m
aα(−i)|α|∂αf.
Then, the statement of Theorem 2.2corresponds to a final-state observability estimate for the system
˙
x(t) =−Apx(t), t∈(0, T], x(0) =x0 ∈Lp(Rd), y(t) =x(t)|E, t∈[0, T].
Let us now turn to the discussion on null-controllability. For a measurable setE ⊂Rd and T >0 we consider the linear control problem
˙
x(t) =−Apx(t) +1Eu(t), t∈(0, T], x(0) =x0 ∈Lp(Rd)
where u ∈ Lr((0, T);Lp(E)) with r ∈ [1,∞]. The unique mild solution is given by Duhamel’s formula
x(t) =Stx0+Btu, where Btu= Z t
0
St−τ1Eu(τ)dτ.
Then, the statement (a) of Theorem2.3corresponds toapproximately null-controllability in timeT, that is, for allε >0 and andx0 ∈L1(Rd), there exists anu∈Lr((0, T);Lp(E)) such that kx(T)kLp(Rd) =kSTx0+BTukLp(Rd) < ε. The statement (b) of Theorem 2.3 corresponds to null-controllability in time T, that is, for all x0 ∈ Lp(Rd), p ∈ (1,∞), there exists an u∈ Lr((0, T);Lp(E)) such thatx(T) = 0. Note that in case p∈(1,∞) null-controllability and approximate null-controllability are equivalent, see, e.g., [Car88].
It is a standard duality argument that Theorem 2.3 follows from Theorem 2.2 by means of Douglas’ lemma. For the sake of completeness we give a short proof.
Proof of Theorem 2.3. Let p ∈ [1,∞), r ∈ [1,∞] and BT: Lr((0, T);Lp(E)) → Lp(Rd) be given by
BTu= Z T
0
ST−t1Eu(t)dτ.
Then, by [Vie05, Theorem 2.1] we have for all g∈Lp′(Rd) k(BT)′gkLr((0,T);Lp(E))′ = sup
τ∈[0,T]
k(ST′ −τg)|EkLp′(E)= sup
t∈[0,T]
k(St′g)|EkLp′(E)
if r= 1, and
k(BT)′gkLr((0,T);Lp(E))′ = Z T
0
k(St−τ′ g)|EkrL′
p′(E)dτ 1/r′
= Z T
0
k(St′g)|EkrL′
p′(E)dt 1/r′
if r ∈ (1,∞], where r′ ∈ [1,∞] is such that 1/r + 1/r′ = 1 and p′ ∈ (1,∞] is such that 1/p+ 1/p′ = 1. Since FSt′ = e−ta(−·)F, we have that (St′)t≥0 is associated to the symbol a(−·) which is strongly elliptic with the same constant c > 0. Moreover, since the associated heat kernel is given by (F−1e−ta)(−·), we have kS′tk ≤ Meωt with the same M and ω as in (5). Thus, Theorem 2.2 and the above equalities imply for all g∈Lp′(Rd)
kST′ gkLp′ ≤Cobsk(BT)′gkLr′((0,T);Lp′(E))=Cobsk(BT)′gkLr((0,T);Lp(E))′,
where Cobs is as in Theorem 2.2 with r replaced by r′. By Douglas’ lemma, see e.g.
[Har78,Car85,Car88], we conclude
{STf: kfkLp(Rd)≤1} ⊂ {BTu:kukLr((0,T);Lp(E))≤Cobs} if p= 1 and
{STf:kfkLp(Rd) ≤1} ⊂ {BTu: kukLr((0,T);Lp(E))≤Cobs} if p∈(1,∞).
By scaling, this implies the statement of the theorem.
3. Proof of Theorem 2.2
For the proof of Theorem 2.2 we apply the abstract observability estimate in Theo- remA.1. For this purpose, we define a familiy of operatorsPλ, and verify the uncertainty principle (13) and the dissipation estimate (14).
We start with defining the operators Pλ. Let η ∈ Cc∞([0,∞)) with 0 ≤ η ≤ 1 such that η(r) = 1 for r ∈[0,1/2] and η(r) = 0 for r ≥1. For λ > 0 we defineχλ:Rd →R by χλ(ξ) =η(|ξ|/λ). Since χλ ∈ S(Rd), we have F−1χλ ∈ S(Rd) ⊂L1(Rd) and for all p∈[1,∞] we definePλ:Lp(Rd)→Lp(Rd) byPλf = (F−1χλ)∗f. By Young’s inequality we have for all f ∈Lp(Rd)
kPλfkLp(Rd)=k(F−1χλ)∗fkLp(Rd)≤ kF−1χλkL1(Rd)kfkLp(Rd).
Moreover, the norm kF−1χλkL1(Rd) is independent of λ > 0. Indeed, by the scaling property of the Fourier transform and by change of variables we have for allλ >0
kF−1χλkL1(Rd)=|λ|dk(F−1χ1)(λ·)kL1(Rd)=kF−1χ1kL1(Rd). (6) Hence, for allλ >0 the operatorPλ is a bounded linear operator and the family (Pλ)λ>0 is uniformly bounded bykF−1χ1kL1(Rd).
Next, we observe that the uncertainty principle (13) is a consequence of the following Logvinenko–Sereda theorem from [Kov01], see also [LS74,Kov00] for predecessors.
Theorem 3.1 (Logvinenko–Sereda theorem). There exists K ≥ 1 such that for all p ∈ [1,∞], λ > 0, ρ ∈ (0,1], L ∈ (0,∞)d, (ρ, L)-thick sets E ⊂ Rd, and f ∈ Lp(Rd) satisfying suppFf ⊂[−λ, λ]d we have
kfkLp(Rd)≤d0ed1λkfkLp(E), where
d0 = eKdln(Kd/ρ) and d1 = 2|L|1ln(Kd/ρ). (7)
Concerning the dissipation estimate (14), we first consider the semigroups associated to powers of the Laplacian.
Proposition 3.2. Let m∈N be even, andGt:Lp(Rd)→Lp(Rd) be given by Gtf =F−1e−t|·|mFf .
Then for all p∈[1,∞], f ∈Lp(Rd), λ >0, and t≥0 we have k(Id−Pλ)GtfkLp(Rd)≤Km,de−2−m−2tλmkfkLp(Rd), where Km,d >0 is a constant depending only on m and d.
Proof. Let us set a=|·|m. Hence we have for all f ∈Lp(Rd) that (Id−Pλ)Gtf =F−1((1−χλ)e−ta)∗f,
and by Young’s inequality we obtain for allλ, t >0 and allf ∈Lp(Rd) k(Id−Pλ)GtfkLp(Rd) ≤ kF−1((1−χλ)e−ta)kL1(Rd)kfkLp(Rd).
For µ > 0 we define kµ: Rd → R by kµ = F−1((1−χµ)e−a). By substitution first in Fourier space, and then in direct space we obtain, using |t1/mξ|m = |t| |ξ|m, for all λ, t >0
kF−1((1−χλ)e−ta)kL1(Rd)
= Z
Rd
1 (2π)d
1 td/m
Z
Rd
eix·(t−1/mξ)(1−χt1/mλ(ξ))e−|ξ|mdξ
dx
= Z
Rd
1 (2π)d
Z
Rd
eiy·ξ(1−χt1/mλ(ξ))e−|ξ|mdξ
dy=kkt1/mλkL1(Rd). We denote by Km,d >0 constants which depend only on m and the dimension d. We allow these constants to change with each occurrence. By Young’s inequality and (6), we have
kF−1(χµe−a)kL1(Rd)=kF−1χµ∗ F−1e−akL1(Rd) ≤Km,d. Hence we find for allµ >0 the uniform bound
kkµkL1(Rd)≤ kF−1e−akL1(Rd)+kF−1(χµe−a)kL1(Rd)≤Km,d. (8) Next we show that the L1-norm of kµ decays even exponentially as µtends to infinity.
For this purpose, let now µ ≥ 1, α ∈ Nd
0 with |α|1 ≤ d+ 1, and denote by Mα the multiplication withxα. By differentiation properties of the Fourier transform we have
Mαkµ=MαF−1[(1−χµ)e−a] =F−1Dαξ[(1−χµ)e−a]
and hence for allx∈Rd
|xαkµ(x)|=
1 (2π)d
Z
Rd
eix·ξDξα[(1−χµ(ξ))e−|ξ|m]dξ
≤ 1 (2π)d
Z
Rd
Dξα[(1−χµ(ξ))e−|ξ|m]
dξ. (9) On the integrand of the right-hand side we apply the product rule and the triangle inequality to obtain
Dαξ[(1−χµ(ξ))e−|ξ|m] ≤ X
β∈Nd 0 β≤α
α β
Dξα−β(1−χµ(ξ))
Dβξe−|ξ|m
. (10)
For all β ∈Nd
0 and β≤α we have
|Dξβe−|ξ|m| ≤Km,d(1 +|ξ|)|β|1(m−1)e−|ξ|m≤Km,de−|ξ|m/2,
where for the last inequality we used that ξ 7→ (1 +|ξ|)|β|1(m−1)e−|ξ|m/2 is bounded on Rd. Since µ≥1, for all β ∈Nd
0,β ≤α we have
Dξα−β(1−χµ(ξ))
≤µ|β|1−|α|1(Dξα−βχ1)(ξ/µ)≤sup
γ≤α
sup
ξ∈Rd
|(Dγξχ1)(ξ)|1Rd\BRd(µ/2)(ξ) and hence
Dξα−β(1−χµ(ξ))
Dξβe−|ξ|m
≤Km,de−|ξ|m/21Rd\BRd(µ/2)(ξ)≤Km,de−|ξ|m/4e−µm/2m+2. Thus, (10) and|α|1 ≤d+ 1 imply for all ξ∈Rdthat
Dξα[(1−χµ(ξ))e−|ξ|m]
≤Km,de−|ξ|m/4e−µm/2m+2 X
β∈Nd 0 β≤α
α β
≤Km,de−|ξ|m/4e−µm/2m+2.
Hence, from (9), for all x∈Rdwe obtain
|xαkµ(x)| ≤Km,de−µm/2m+2 Z
Rd
e−|ξ|m/4dξ =Km,de−µm/2m+2. (11) In particular, forj∈ {1,2, . . . , d}andαj = (d+1)ej, whereej denotes thej-th canonical unit vector in Rd, we obtain |xj|d+1|kµ(x)| ≤ Km,de−µm/2m+2, hence kxkd+1∞ |kµ(x)| ≤ Km,de−µm/2m+2, and consequently for all x∈Rdand all µ≥1 we find
|x|d+1|kµ(x)| ≤Km,de−µm/2m+2. (12) From (11) withα= 0 and (12) we obtain for all µ≥1 that
kkµkL1(Rd)≤Km,de−µm/2m+2 Z
BRd(1)
dx+Km,de−µm/2m+2 Z
Rd\BRd(1)
|x|−d−1dx
≤Km,de−µm/2m+2.
From this inequality and (8) we obtain for all µ >0 that kkµkL1(Rd)≤Km,de−µm/2m+2.
We are now in the position to show the dissipation estimate for the general case.
Proposition 3.3. Let m ∈ N, a:Rd → C a strongly elliptic polynomial of order m, c > 0 and ω ∈ R as in (2), (St)t≥0 as in (3), and (Pλ)λ>0 as above. Then for all p∈[1,∞], f ∈Lp(Rd), λ >(2m+4max{ω,0}/c)1/m, and t≥0 we have
k(Id−Pλ)StfkLp(Rd)≤Kae−2−m−4ctλmkfkLp(Rd),
where Ka ≥ 1 is a constant depending only on the polynomial a (and therefore also on m and d).
Proof of Proposition 3.3. We introduce ˜a:Rd→C, ˜a(ξ) = (c/2)|ξ|m. Then ˜aand (a−˜a) are strongly elliptic polynomials of orderm∈N. Note that the semigroup associated to
˜
ais G(c/2)t for allt≥0, where (Gt)t≥0 is as in Proposition 3.2. Moreover, let Tt be the semigroup associated toa−˜a. Sincea=a−˜a+ ˜a, it follows thatSt=TtG(c/2)t. One can obtain a corresponding heat kernel bound for the kernel of the semigroup (Tt)t≥0 with the same growth rate ω as for (St)t≥0 as follows: Since m has to be even, by Young’s inequality for products there exists σ,σ, C˜ ≥0 such that
Re(a−a)(ξ˜ + iη) = Rea(ξ+ iη)−(c/2)|ξ+ iη|m
≥(3/4)c|ξ|m−σ|η|m−ω−(c/2)|ξ+ iη|m
≥(3/4)c|ξ|m−σ|η|m−ω−(c/2)
m/2
X
k=0
m/2 k
(|ξ|2)k(|η|2)m/2−k
≥(3/4)c|ξ|m−σ|η|m−ω−(c/2)(1 + 1/4)|ξ|m−C|η|m
≥(1/8)c|ξ|m−σ|η|˜ m−ω,
with yields (2) with a replaced by a−a˜ and ξ replaced by ξ+ iη. Then, arguing as in [TR96, Proposition 2.1], one can prove via a heat kernel bound as in (4) that there exists ˜M ≥1 such thatkTtk ≤M˜eωt for all t≥0. By Proposition 3.2and since Fourier multipliers commute, we obtain for allf ∈Lp(Rd)
k(Id−Pλ)StfkLp(Rd)=kSt(Id−Pλ)fkLp(Rd)
≤ kTtkkG(c/2)t(Id−Pλ)fkLp(Rd)
≤M K˜ m,de−t(2−m−2(c/2)λm−ω)kfkLp(Rd),
where Km,d > 0 is a constant depending only on m and d. Since λ > (2m+4max{ω, 0}/c)1/m, we have 2−m−2(c/2)λm−ω >2−m−2cλm/4 = 2−m−4cλm.
We can finally prove Theorem2.2.
Proof of Theorem 2.2. Let (Pλ)λ>0 be the family of operators defined at the beginning of this section. Then we have suppF(Pλf)⊂[−λ, λ]d for all λ >0 and allf ∈Lp(Rd).
Thus, Theorem3.1implies that for all f ∈Lp(Rd) and allλ >0 we have kPλfkLp(Rd)≤d0ed1λkPλfkLp(E),
where d0 and d1 are as in (7). Moreover, according to Proposition 3.3, for all λ > λ∗ and allf ∈Lp(Rd) we have
k(I−Pλ)StfkLp(Rd)≤d2e−d3λmtkfkLp(E),
where λ∗ = (2m+4max{ω,0}/c)1/m, d2 ≥ 1 depends only on the polynomial a, and whered3= 2−m−4c. Moreover, the functiont7→ k(Stf)|EkLp(E) is Borel-measurable for all f ∈Lp(Rd). Indeed, if p ∈[1,∞) the semigroup (St)t≥0 is strongly continuous and the measurability follows. If p=∞, measurability is a consequence of duality and the representation of the norm in L∞(E) by means of the Hahn–Banach theorem.
Hence we can apply Theorem A.1 with X =Lp(Rd), Y =Lp(E), C:X → Y given by the restriction map on E, and obtain that the statement of the theorem holds with Cobs replaced by
C˜obs = C1 T1/r exp
C2
Tm−11 +C3T
,
whereT1/r = 1 if r=∞, and C1 = (4M d0) maxn
4d2M2(d0+ 1)8/(e ln 2)
,e4d12λ∗o , C2 = 4 2·8m−m1dm1 /d3m−11
, C3 = max{ω,0} 1 + 10/(e ln 2)
,
with M as in (5). We denote by Kd, Km, and Ka positive constants which depend only on the dimensiond, on m, or on the polynomiala, respectively. A straightforward calculation shows that
C1 ≤Ka Kd
ρ
Kd(1+|L|1λ∗)
and C2 ≤ Km(|L|1ln(Kd/ρ))m/(m−1)
c1/(m−1) .
Thus we obtain C˜obs ≤ Ka
T1/r Kd
ρ
Kd(1+|L|1λ∗)
exp Km(|L|1ln(Kd/ρ))m/(m−1)
(cT)1/(m−1) +C3T
!
=:Cobs.
A. Sufficient criteria for observability in Banach spaces
We provide an abstract sufficient criteria for final-state observability in Banach space, which is a slight generalization of Theorem 2.1 in [GST20]. In particular, it does not assume strong continuity of the semigroup. For the proof we comment only on the necessary modifications compared to [GST20].
Theorem A.1. Let X and Y be Banach spaces, C:X →Y a bounded linear operator, (St)t≥0 a semigroup onX, M ≥1 and ω ∈R such that kStk ≤Meωt for all t≥0, and assume that for allx∈X the mapping t7→ kCStxkY is measurable. Further, letλ∗≥0, (Pλ)λ>λ∗ a family of bounded linear operators inX,r ∈[1,∞],d0, d1, d3, γ1, γ2, γ3, T >0 with γ1 < γ2, and d2≥1, and assume that
∀x∈X ∀λ > λ∗: kPλxkX ≤d0ed1λγ1kCPλxkY, (13) and
∀x∈X ∀λ > λ∗ ∀t∈(0, T /2] : k(Id−Pλ)StxkX ≤d2e−d3λγ2tγ3kxkX. (14) Then we have for all x∈X
kSTxkX ≤
Cobs
RT
0 kCStxkrY dt1/r
if r∈[1,∞), Cobsess supt∈[0,T]kCStxkY if r=∞, with
Cobs= C1 T1/r exp
C2 T
γ1γ3 γ2−γ1
+C3T
,
where T1/r = 1 if r =∞, and C1 = (4M d0) maxn
4d2M2(d0kCk+ 1)8/(e ln 2)
,e4d1(2λ∗)γ1o , C2 = 4 2γ1(2·4γ3)
γ1γ2
γ2−γ1dγ12/dγ31γ2−γ11 , C3 = max{ω,0} 1 + 10/(e ln 2)
.
The assumption in (13) is an abstract uncertainty principle (sometimes also called spectral inequality), while (14) is a dissipation estimate. Thus, Theorem A.1 can be rephrased that an uncertainty principle together with a dissipation estimate implies a final state observability estimate.
Remark A.2. In the situation of Theorem A.1, if we assume that t7→CStx is Bochner measurable, we can rewrite the statement of the theorem as
kSTxkX ≤CobskCS(·)xkLr((0,T);Y).
Proof of Theorem A.1. Since we do not assume the semigroup (St)t≥0 to be strongly continuous, we cannot apply [GST20, Theorem 2.1] directly. The strong continuity of (St)t≥0 was assumed in [GST20] in order to ensure that for all x ∈X and λ > λ∗ the functions
F(t) = Stx
X, Fλ(t) =
PλStx
X, Fλ⊥(t) =
(Id−Pλ)Stx X, G(t) =
CStx
Y, Gλ(t) =
CPλStx
Y, G⊥λ(t) =
C(Id−Pλ)Stx Y,
are measurable. The measurability of these six functions was used to obtain the estimate F(t)≤ 2Meω+Td0ed1λγ1
t
Z t t/2
G(τ)dτ +d2M2e5ω+T /4ed1λγ1
ed3λγ2(t/4)γ3 (d0kCk+ 1)F(t/4), where ω+ = max{0, ω}. Such an inequality implies the statement of the theorem by iteration, see [GST20]. Thus it suffices to show the last displayed inequality by assuming merely measurability of the mapping t7→G(t). Let t >0,τ ∈[t/2, t] andx∈X. Since F(τ)≤Fλ(τ) +Fλ⊥(τ), by our assumptions and by the semigroup property we obtain
F(τ)≤d0ed1λγ1Gλ(τ) +d2e−d3λγ2(τ /2)γ3F(τ /2).
Using Gλ(τ) ≤ G(τ) +G⊥λ(τ) ≤ G(τ) +kCkFλ⊥(τ), our assumption, ed1λγ1 ≥ 1, and F(τ /2)≤Meω+t/4F(t/4) we obtain
F(τ)≤d0ed1λγ1G(τ) + (d0kCk+ 1)d2e−d3λγ2(τ /2)γ3ed1λγ1Meω+t/4F(t/4).
Since F(t)≤Meω+tF(τ), we obtain
F(t)≤Meω+td0ed1λγ1G(τ) + (d0kCk+ 1)d2e−d3λγ2(τ /2)γ3ed1λγ1M2eω+5t/4F(t/4).
Since the mapping τ 7→ G(τ) is measurable by assumption, we can integrate this in- equality with respect to τ, and obtain the desired estimate.
References
[Are04] W. Arendt. Semigroups and evolution equations: functional calculus, reg- ularity and kernel estimates. In C. M. Dafermos and E. Feireisl, editors, Handbook of Differential Equations: Evolutionary Equations: Evolutionary Equations, volume 1, pages 1–85. Elsevier, Amsterdam, 2004.
[BPS18] K. Beauchard and K. Pravda-Starov. Null-controllability of hypoelliptic quadratic differential equations. J. ´Ec. polytech. Math., 5:1–43, 2018.
[Car85] O. Carja. A note on the continuity of the minimal time function for linear control systems. Boll. Unione Mat. Ital., 6(4A):293–302, 1985.
[Car88] O. Carja. On constraint controllability of linear systems in Banach spaces.
J. Optim. Theory Appl., 56(2):215–225, 1988.
[CP78] R. F. Curtain and A. J. Pritchard.Infinite dimensional linear systems theory.
Springer, Berlin, 1978.
[Dou66] R. G. Douglas. On majorization, factorization, and range inclusion of oper- ators on hilbert space. Proc. Amer. Math. Soc., 2(17):413–415, 1966.
[DR77] S. Dolecki and D. L. Russell. A general theory of observation and control.
SIAM J. Control Optim., 2(15):185–220, 1977.
[Emb73] M. R. Embry. Factorization of operators on Banach space. Proc. Amer.
Math. Soc., 38(3):587–590, 1973.
[FCZ00] E. Fern´andez-Cara and E. Zuazua. Null and approximate controllability for weakly blowing up semilinear heat equations. In Ann. Inst. H. Poincar´e Anal. Non Lin´eaire, volume 17, pages 583–616, 2000.
[For14] M. Forough. Majorization, range inclusion, and factorization for unbounded operators on Banach spaces. Linear Algebra Appl., 449:60–67, 2014.
[FPZ95] C. Fabre, J.-P. Puel, and E. Zuazua. Approximate controllability of the semilinear heat equation. Proc. Roy. Soc. Edinburgh Sect. A, 125(1):31–61, 1995.
[Gra14] L. Grafakos. Classical Fourier Analysis. Springer, New York, 2014.
[GST20] D. Gallaun, C. Seifert, and M. Tautenhahn. Sufficient criteria and sharp geometric conditions for observability in banach spaces. SIAM J. Control Optim., 58(4):2639–2657, 2020.
[Har78] R. E. Harte. Berberian-quigley and the ghost of a spectral mapping theorem.
Math. Proc. R. Ir. Acad., 78:63–68, 1978.
[JL99] D. Jerison and G. Lebeau. Nodal sets of sums of eigenfunctions. In M. Christ, C. E. Kenig, and C. Sadosky, editors, Harmonic analysis and partial dif- ferential equations, Chicago Lectures in Mathematics, pages 223–239. The University of Chicago Press, Chicago, 1999.
[Kov00] O. Kovrijkine. Some estimates of Fourier transforms. PhD thesis, California Institute of Technology, 2000.
[Kov01] O. Kovrijkine. Some results related to the Logvinenko-Sereda theorem. Proc.
Amer. Math. Soc., 129(10):3037–3047, 2001.
[Lot85] H. P. Lotz. Uniform convergence of operators onl∞and similar spaces.Math.
Z., 190(2):207–220, 1985.
[LR95] G. Lebeau and L. Robbiano. Contrˆole exact de l’´equation de la chaleur.
Comm. Partial Differential Equations, 20(1–2):335–356, 1995.
[LS74] V. N. Logvinenko and Ju. F. Sereda. Equivalent norms in spaces of entire functions of exponential type. Teor. Funkts., Funkts. anal. Prilozh., 20:102–
111, 1974.
[LZ98] G. Lebeau and E. Zuazua. Null-controllability of a system of linear thermoe- lasticity. Arch. Ration. Mech. Anal., 141(4):297–329, 1998.
[Mil10] L. Miller. A direct Lebeau-Robbiano strategy for the observability of heat- like semigroups. Discrete Contin. Dyn. Syst. Ser. B, 14(4):1465–1485, 2010.
[NTTV20] I. Naki´c, M. T¨aufer, M. Tautenhahn, and I. Veseli´c. Sharp estimates and homogenization of the control cost of the heat equation on large domains.
ESAIM Control Optim. Calc. Var., 26(54):26 pages, 2020.
[TR96] A. F. M. Ter Elst and D. W. Robinson. Elliptic operators on Lie groups.
Acta Appl. Math., 44(1–2):133–150, 1996.
[TT11] G. Tenenbaum and M. Tucsnak. On the null-controllability of diffusion equa- tions. ESAIM Control Optim. Calc. Var., 17(4):1088–1100, 2011.
[Vie05] A. Vieru. On null controllability of linear systems in Banach spaces. Systems Control Lett., 54(4):331–337, 2005.
[WZ17] G. Wang and C. Zhang. Observability inequalities from measurable sets for some abstract evolution equations. SIAM J. Control Optim., 55(3):1862–
1886, 2017.
[Zua06] E. Zuazua. Controllability and observability of partial differential equations:
Some results and open problems. In C. M. Dafermos and E. Feireisl, editors, Handbook of Differential Equations: Evolutionary Equations, volume 3, pages 527–621. Elsevier Science, 2006.