Citation: Pierluigi Colli, Gianni Gilardi, Jürgen Sprekels. Distributed optimal control of a nonstandard nonlocal phase field system[J]. AIMS Mathematics, 2016, 1(3): 225-260. doi: 10.3934/Math.2016.3.225
[1] | Zonghong Xiong, Wei Wei, Ying Zhou, Yue Wang, Yumei Liao . Optimal control for a phase field model of melting arising from inductive heating. AIMS Mathematics, 2022, 7(1): 121-142. doi: 10.3934/math.2022007 |
[2] | Wei Li, Guangying Lv . A fully-decoupled energy stable scheme for the phase-field model of non-Newtonian two-phase flows. AIMS Mathematics, 2024, 9(7): 19385-19396. doi: 10.3934/math.2024944 |
[3] | Noha M. Kamel, Hamdy M. Ahmed, Wafaa B. Rabie . Solitons unveilings and modulation instability analysis for sixth-order coupled nonlinear Schrödinger equations in fiber bragg gratings. AIMS Mathematics, 2025, 10(3): 6952-6980. doi: 10.3934/math.2025318 |
[4] | Abdelkader Moumen, Ammar Alsinai, Ramsha Shafqat, Nafisa A. Albasheir, Mohammed Alhagyan, Ameni Gargouri, Mohammed M. A. Almazah . Controllability of fractional stochastic evolution inclusion via Hilfer derivative of fixed point theory. AIMS Mathematics, 2023, 8(9): 19892-19912. doi: 10.3934/math.20231014 |
[5] | Qian Li, Zhenghong Jin, Linyan Qiao, Aichun Du, Gang Liu . Distributed optimization of nonlinear singularly perturbed multi-agent systems via a small-gain approach and sliding mode control. AIMS Mathematics, 2024, 9(8): 20865-20886. doi: 10.3934/math.20241015 |
[6] | Mahmoud Soliman, Hamdy M. Ahmed, Niveen Badra, Taher A. Nofal, Islam Samir . Highly dispersive gap solitons for conformable fractional model in optical fibers with dispersive reflectivity solutions using the modified extended direct algebraic method. AIMS Mathematics, 2024, 9(9): 25205-25222. doi: 10.3934/math.20241229 |
[7] | Tainian Zhang, Zhixue Luo . Optimal harvesting for a periodic competing system with size structure in a polluted environment. AIMS Mathematics, 2022, 7(8): 14696-14717. doi: 10.3934/math.2022808 |
[8] | Madeaha Alghanmi, Shahad Alqurayqiri . Existence results for a coupled system of nonlinear fractional functional differential equations with infinite delay and nonlocal integral boundary conditions. AIMS Mathematics, 2024, 9(6): 15040-15059. doi: 10.3934/math.2024729 |
[9] | Yongqiang Fu, Lixu Yan . Fully nonlocal stochastic control problems with fractional Brownian motions and Poisson jumps. AIMS Mathematics, 2021, 6(5): 5176-5192. doi: 10.3934/math.2021307 |
[10] | Yuanyuan Cheng, Yuan Li . A novel event-triggered constrained control for nonlinear discrete-time systems. AIMS Mathematics, 2023, 8(9): 20530-20545. doi: 10.3934/math.20231046 |
Let Ω⊂R3 denote an open and bounded domain whose smooth boundary Γ has the outward unit normal n; let T>0 be a given final time, and set Q:=Ω×(0,T) and Σ:=Γ×(0,T). We study in this paper distributed optimal control problems of the following form:
(CP) Minimize the cost functional
J(u,ρ,μ)=β12∫Q|ρ−ρQ|2dxdt+β22∫Q|μ−μQ|2dxdt+β32∫Q|u|2dxdt | (1.1) |
subject to the state system
(1+2g(ρ))∂tμ+μg′(ρ)∂tρ−Δμ=ua.e.inQ, | (1.2) |
∂tρ+B[ρ]+F′(ρ)=μg′(ρ)a.e.inQ, | (1.3) |
∂nμ=0a.e.onΣ, | (1.4) |
ρ(⋅,0)=ρ0,μ(⋅,0)=μ0,a.e.inΩ, | (1.5) |
and to the control constraints
u∈Uad:={u∈H1(0,T;L2(Ω)):0≤u≤umaxa.e.inQ and ‖u‖H1(0,T;L2(Ω))≤R}. | (1.6) |
Here, \gianni{β1,β2,β3≥0 and R>0 are given constants, with β1+β2+β3>0, } and the threshold function umax∈L∞(Q) is nonnegative}. Moreover, ρQ,μQ∈L2(Q) represent prescribed target functions of the tracking-type functional J. Although more general cost functionals could be admitted for large parts of the subsequent analysis, we restrict ourselves to the above situation for the sake of a simpler exposition.
The state system (1.2)-(1.5) constitutes a nonlocal version of a phase field model of Cahn-Hilliard type describing phase segregation of two species (atoms and vacancies, say) on a lattice, which was recently studied in [18]. In the (simpler) original local model, which was introduced in [25], the nonlocal term B[ρ] is replaced by the diffusive term-Δρ. On the other hand, it is important from the point of view of applications to have a nonlocal operator (see, e.g., the Introduction of [18]). Indeed, terms of double integral type are more natural in the free energy, whereas squared gradients can be seen as limiting situations of nonlocal contributions.
The local model has been the subject of intensive research in the past years; in this connection, we refer the reader to [4-7,9-12]. In particular, in [8] the analogue of the control problem (CP) for the local case was investigated for the special situation g(ρ)=ρ, for which the optimal boundary control problems was studied in [14].
The state variables of the model are the order parameter ρ and the chemical potential μ. While ρ can be interpreted as a volumetric density, the chemical potential μ plays as the coldness in the entropy imbalance (see [25, formulas (2) and (24)]). Hence, we must have 0≤ρ≤1 and μ > 0 almost everywhere in Q. The control function u on the right-hand side of (1.2) has the meaning of a microenergy source. We remark at this place that the requirement encoded in the definition of Uad, namely that u be nonnegative, is indispensable for the analysis of the forthcoming sections. Indeed, it is needed to guarantee the nonnegativity of the chemical potential μ.
The nonlinearity F is a double-well potential defined in the interval (0, 1), whose derivative F′ is singular at the endpoints ρ=0 and ρ=1: e. g., F=F1 + F2, where F2 is smooth and
F1(ρ)=ˆc(ρlog(ρ)+(1−ρ)log(1−ρ)),withaconstantˆc>0. | (1.7) |
The presence of the nonlocal term B[ρ] in (1.3) constitutes the main difference to the local model and introduces some difficulties in the mathematical analysis due to the lack of compacness and less regularity for the solution. Simple examples are given by integral operators of the form
B[ρ](x,t)=∫Ω×(0,t)k(t,s,x,y)ρ(y,s)dyds | (1.8) |
and purely spatial convolutions like
B[ρ](x,t)=∫Qk(|y−x|)ρ(y,t)dy, | (1.9) |
with sufficiently regular kernels.
Optimal control problems of the above type often occur in industrial production processes. For instance, consider a metallic workpiece consisting of two different component materials that tend to separate. Then a typical goal would be to monitor the production process in such a way that a desired distribution of the two materials (represented by the function ρQ) is realized during the time evolution in order to guarantee a wanted behavior of the workpiece; the deviation from the desired phase distribution is measured by the first summand in the cost J. The third summand of J represents the costs due to the control action u; the size of the factors βi≥0 then reflects the relative importance that the two conflicting interests“realize the desired phase distribution as closely as possible”and“minimize the cost of the control action”have for the manufacturer.
The state system (1.2)-(1.5) is singular, with highly nonlinear and nonstandard coupling. In particular, unpleasant nonlinear terms involving time derivatives occur in (1.2), and the expression F′(ρ) in (1.3) may become singular. Moreover, the nonlocal term B[ρ] is a source for possible analytical difficulties, and the absence of the Laplacian in (1.3) may cause a low regularity of the order parameter ρ. We remark that the state system (1.2)-(1.5) was recently analyzed in [18] for the case u=0 (no control); results concerning well-posedness and regularity were established.
The mathematical literature on control problems for phase field systems involving equations of viscous or nonviscous Cahn-Hilliard type is still scarce and quite recent. We refer in this connection to the works [2,3,16,17,21,28]. Control problems for convective Cahn-Hilliard systems were studied in [29,30], and a few analytical contributions were made to the coupled Cahn-Hilliard/Navier-Stokes system (cf. [19,20,22,23]). The very recent contribution [13] deals with the optimal control of a Cahn-Hilliard type system arising in the modeling of solid tumor growth.
The paper is organized as follows: in Section 2, we state the general assumptions and derive new regularity and stability results for the state system. In Section 3, we establish the directional differentiability of the control-to-state operator, and the final Section 4 brings the main results of this paper, namely, the derivation of the first-order necessary conditions of optimality.
Throughout this paper, we will use the following notation: we denote for a (real) Banach space X by ‖⋅‖X its norm and the norm of X×X×X, by X′ its dual space, and by ⟨⋅,⋅⟩X the dual pairing between X′ and X. If X is an inner product space, then the inner product is denoted by (⋅,⋅)X. The only exception from this convention is given by the Lp spaces, 1≤p≤∞, for which we use the abbreviating notation ‖⋅‖p for the norms. Furthermore, we put
H:=L2(Ω),V:=H1(Ω),W:={w∈H2(Ω):∂nw=0a.e.onΓ}. |
We have the dense and continuous embeddings W⊂V⊂H≅H′⊂V′⊂W′, where ⟨u,v⟩V=(u,v)H and ⟨u,w⟩W=(u,w)H for all u∈H, v∈V, and w∈W.
In the following, we will make repeated use of Young’s inequality
ab≤δa2+14δb2for all a,b∈Rand δ>0, | (1.10) |
as well as of the fact that for three dimensions of space and smooth domains the embeddings V⊂Lp(Ω), 1≤p≤6, and H2(Ω)⊂C0(¯Ω) are continuous and (in the first case only for 1≤p<6) compact. In particular, there are positive constants ˜Ki, i=1,2,3, which depend only on the domain Ω, such that
‖v‖6≤˜K1‖v‖V∀v∈V, | (1.11) |
‖vw‖H≤‖v‖6‖w‖3≤˜K2‖v‖V‖w‖V∀v,w∈V, | (1.12) |
‖v‖∞≤˜K3‖v‖H2(Ω)∀v∈H2(Ω). | (1.13) |
We also set for convenience
Qt:=Ω×(0,t) and Qt:=Ω×(t,T),for t∈(0,T). | (1.14) |
Please note the difference between the subscript and the superscript in the above notation.
About time derivatives of a time-dependent function v, we point out that we will use both the notations ∂tv,∂2tv and the shorter ones vt,vtt.
Consider the optimal control problem (1.1)-(1.6). We make the following assumptions on the data:
(A1) F=F1 + F2, where F1∈C3(0,1) is convex, F2∈C3[0,1], and
limr↘0F′1(r)=−∞,limr↗1F′1(r)=+∞. | (2.1) |
(A2) ρ0∈V, F′(ρ0)∈H, μ0∈W, where μ0≥0\, a.\, e. in Ω,
inf{ρ0(x):x∈Ω}>0,sup{ρ0(x):x∈Ω}<1. | (2.2) |
(A3) g∈C3[0,1] satisfies g(ρ)≥0 and g″(ρ)≤0 for all $\rho\in [0,1].
(A4) The nonlocal operator B:L1(Q)→L1(Q) satisfies the following conditions:
(i) For every t∈(0,T], we have
B[v]|Qt=B[w]|Qt whenever v|Qt=w|Qt. | (2.3) |
(ii) For all p∈[2,+∞], we have B(Lp(Qt))⊂Lp(Qt) and
‖B[v]‖Lp(Qt)≤CB,p(1+‖v‖Lp(Qt)) | (2.4) |
for every v∈Lp(Q) and t∈(0,T].
(iii) For every v,w∈L1(0,T;H) and t∈(0,T], it holds that
∫t0‖B[v](s)−B[w](s)‖6ds≤CB∫t0‖v(s)−w(s)‖Hds. | (2.5) |
(iv) It holds, for every v∈L2(0,T;V) and t∈(0,T], that
‖∇B[v]‖L2(0,t;H)≤CB(1+‖v‖L2(0,t;V)). | (2.6) |
(v) For every v∈H1(0,T;H), we have ∂tB[v]∈L2(Q) and
‖∂tB[v]‖L2(Q)≤CB(1+‖∂tv‖L2(Q)). | (2.7) |
(vi) B is continuously Fréchet differentiable as a mapping from L2(Q) into L2(Q), and the Fréchet derivative DB[¯v]∈L(L2(Q),L2(Q)) of B at ¯v has for every ¯v∈L2(Q) and t∈(0,T] the following properties:
‖DB[ˉv](w)‖Lp(Qt)≤CB‖w‖Lp(Qt)∀w∈Lp(Q),∀p∈[2,6], | (2.8) |
‖∇(DB[ˉv](w))‖L2(Qt)≤CB‖w‖L2(0,t;V)∀w∈L2(0,T;V). | (2.9) |
In the above formulas, CB, p and CB denote given positive structural constants. We also notice that (2.8) implicitely requires that DB[¯v](w)|Qt depends only on w|Qt. However, this is a consequence of (2.3).
The statements related to the control problem (CP) depend on the assumptions made in the Introduction. We recall them here.
(A5) J and Uad are defined by (1.1) and (1.6), respectively, where
β1,β2,β3≥0, β1+β2+β3>0, and R>0. | (2.10) |
ρQ,μQ∈L2(Q),umax∈L∞(Q) and umax≥0 a.e.inQ. | (2.11) |
Remark 1: In view of (2.8), for every t∈[0,T] it holds that
‖B[v]−B[w]‖L2(Qt)≤CB‖v−w‖L2(Qt)∀v,w∈L2(Q), | (2.12) |
that is, the condition (2.9) in [18] is fulfilled. Moreover, (2.4) and (2.6) imply that B maps L2(0, T; V) into itself and that, for all t∈(0,T]\, and \, v∈L2(0,T;V),
|∫Qt∇B[v]⋅∇vdxds|≤CB(1+‖v‖2L2(0,t;V)), |
which means that also the condition (2.10) in [18] is satisfied. Moreover, thanks to (2.8) and (2.9), there is some constant ˜CB>0 such that
‖DB[ˉv](w)‖L2(0,T;V)≤˜CB‖w‖L2(0,T;V)∀ˉv∈L2(Q),∀w∈L2(0,T;V). | (2.13) |
REMARK 2: We recall (cf. [18]) that the integral operator (1.9) satisfies the conditions (2.3) and (2.4), provided that the integral kernel k belongs to C1(0, +∞) and fulfills, with suitable constants C1 > 0, C2 > 0, 0<α<32, 0<β<52, the growth conditions
|k(r)|≤C1r−α,|k′(r)|≤C2r−β,∀r>0. |
In this case, we have 2α<3 and thus, for all v,w∈L1(0,T;H)\, and t∈(0,T],
∫t0‖B[v](s)−B[w](s)‖6ds≤C1∫t0(∫Ω|∫Ω|y−x|−α|v(y,s)−w(y,s)|dy|6dx)1/6ds≤C3∫t0(∫Ω|(∫Ω|y−x|−2αdy)1/2‖v(s)−w(s)‖H|6dx)1/6ds≤C4∫t0‖v(s)−w(s)‖Hds, |
with global constants Ci, 3≤i≤4; the condition (2.5) is thus satisfied. Also condition (2.6) holds true in this case: indeed, for every t∈(0,T] and v∈L2(0,T;V), we find, since 6β/5<3, that
‖∇B[v]‖2L2(0,t;H)≤C2∫Qt|∫Ω|y−x|−β|v(y,s)|dy|2dxds≤C5∫Qt(∫Ω|y−x|−6β/5dy)5/3‖v(s)‖26dxds≤C6∫t0‖v(s)‖2Vds. |
Finally, since the operator B is linear in this case, we have DB[¯v]=B for every ˉv∈L2(Q), and thus also (A4)(v) and (2.8)-(2.13) are fulfilled. Notice that the above growth conditions are met by, e. g., the three-dimensional Newtonian potential, where k(r)=c/r with some c≠0.
We also note that (A2) implies μ0∈C(¯Ω), and (A1) and (2.2) ensure that both F(ρ0) and F′(ρ0) are in L∞(Ω), whence in H. Moreover, the logarithmic potential (1.7) obviously fulfills the condition (2.1) in (A1).
We have the following existence and regularity result for the state system.
THEOREM 2.1: Suppose that (A1)-(A5) are satisfied. Then the state system (1.2)-(1.5) has for every u∈Uad a unique solution (ρ,μ) such that
ρ∈H2(0,T;H)∩W1,∞(0,T;L∞(Ω))∩H1(0,T;V), | (2.14) |
μ∈W1,∞(0,T;H)∩H1(0,T;V)∩L∞(0,T;W)⊂L∞(Q). | (2.15) |
Moreover, there are constants 0<ρ∗<ρ∗<1, μ∗>0, and K∗1>0, which depend only on the given data, such that for every u∈Uad the corresponding solution (ρ,μ) satisfies
0<ρ∗≤ρ≤ρ∗<1,0≤μ≤μ∗,a.e.inQ, | (2.16) |
‖μ‖W1,∞(0,T;H)∩H1(0,T;V)∩L∞(0,T;W)∩L∞(Q)+‖ρ‖H2(0,T;H)∩W1,∞(0,T;L∞(Ω))∩H1(0,T;V)≤K∗1. | (2.17) |
PROOF: In the following, we denote by Ci>0,i∈N, constants which may depend on the data of the control problem (CP) but not on the special choice of u∈Uad. First, we note that in [18, Thms. 2.1, 2.2] it has been shown that under the given assumptions there exists for u≡0 a unique solution (ρ,μ) with the properties
0<ρ<1,μ≥0,a.e.inQ, | (2.18) |
ρ∈L∞(0,T;V), ∂tρ∈L6(Q), | (2.19) |
μ∈H1(0,T;H)∩L∞(0,T;V)∩L∞(Q)∩L2(0,T;W2,3/2(Ω)). | (2.20) |
A closer inspection of the proofs in [18] reveals that the line of argumentation used there (in particular, the proof that μ is nonnegative) carries over with only minor modifications to general right-hand sides u∈Uad. We thus infer that (1.2)-(1.5) enjoys for every u∈Uad a unique solution satisfying (2.18)-(2.20); more precisely, there is some C1 > 0 such that
‖μ‖H1(0,T;H)∩L∞(0,T;V)∩L∞(Q)∩L2(0,T;W2,3/2(Ω))+‖ρ‖L∞(0,T;V)+‖∂tρ‖L6(Q)≤C1∀u∈Uad. | (2.21) |
Moreover, invoking (2.18), and (2.4) for p=+∞, we find that
‖B[ρ]‖L∞(Q)≤C2∀u∈Uad, |
and it follows from (2.21) and the general assumptions on ρ0, g, and F, that there are constants ρ∗,ρ∗ such that, for every u∈Uad,
0<ρ∗≤inf{ρ0(x):x∈Ω}≤sup{ρ0(x):x∈Ω}≤ρ∗<1,F′(ρ)+B[ρ]−μg′(ρ)≤0if0<ρ≤ρ∗,F′(ρ)+B[ρ]−μg′(ρ)≥0ifρ∗≤ρ<1. |
Therefore, multiplying (1.3) by the positive part (ρ−ρ∗)+ of ρ−ρ∗, and integrating over Q, we find that
0=∫Q∂tρ(ρ−ρ∗)+dxdt+∫Q(F′(ρ)+B[ρ]−μg′(ρ))(ρ−ρ∗)+dxdt≥12∫Ω|(ρ(t)−ρ∗)+|2dx, |
whence we conclude that (ρ−ρ∗)+=0, and thus ρ≤ρ∗, almost everywhere in Q. The other bound for ρ in (2.16) is proved similarly.
It remains to show the missing bounds in (2.17) (which then also imply the missing regularity claimed in (2.14)-(2.15)). To this end, we employ a bootstrapping argument.
First, notice that (A3) and the already proved bounds (2.21) and (2.16) imply that the expressions μg′(ρ)∂tρ and (1+2g(ρ))∂tμ are bounded in L2(Q). Hence, by comparison in (1.2), the same holds true for Δμ, and thus standard elliptic estimates yield that
‖μ‖L2(0,T;W)≤C3∀u∈Uad. | (2.22) |
Next, observe that (A1) and (2.16) imply that ‖F′(ρ)‖L∞(Q)≤C4, and comparison in (1.3), using (A3), yields that
‖∂tρ‖L∞(Q)≤C5∀u∈Uad. | (2.23) |
In addition, we infer from the estimates shown above, and using (2.6), that the right-hand side of the identity
∇ρt=−F″(ρ)∇ρ−∇B[ρ]+g′(ρ)∇μ+μg″(ρ)∇ρ | (2.24) |
is bounded in L2(Q), so that
‖∂tρ‖L2(0,T;V)≤C6∀u∈Uad. | (2.25) |
We also note that the time derivative ∂t(−F′(ρ)−B[ρ]+μg′(ρ)) exists and is bounded in L2(Q) (cf. (2.7)). We thus have
‖ρtt‖L2(Q)≤C7∀u∈Uad. | (2.26) |
At this point, we observe that Eq. (1.2) can be written in the form
a∂tμ+μ∂ta−Δμ=b,with a:=1+2g(ρ),b:=u+μg′(ρ)∂tρ, |
where, thanks to the above estimates, we have, for every u∈Uad,
‖a‖L∞(Q)+‖∂ta‖L∞(Q)+‖b‖L∞(Q)≤C8, | (2.27) |
‖∂t2a‖L2(Q)=2‖g″(ρ)ρ2t+g′(ρ)ρtt‖L2(Q)≤C9, | (2.28) |
‖∂tb‖L2(Q)=‖ut+μtg′(ρ)ρt+μg″(ρ)ρ2t+μg′(ρ)ρtt‖L2(Q)≤C10. | (2.29) |
Since also μ0∈W, we are thus in the situation of [15, Thm. 3.4], whence we obtain that ∂tμ∈L∞(0,T;H)∩L2(0,T;V) and μ∈L∞(0,T;W). Moreover, a closer look at the proof of [15, Thm. 3.4] reveals that we also have the estimates
‖∂tμ‖L∞(0,T;H)∩L2(0,T;V)+‖μ‖L∞(0,T;W)≤C11. | (2.30) |
This concludes the proof of the assertion.
REMARK 3: From the estimates (2.16) and (2.17), and using the continuity of the embedding V⊂L6(Ω), we can without loss of generality (by possibly choosing a larger K∗1) assume that also
max0≤i≤3‖F(i)(ρ)‖L∞(Q)+max0≤i≤3‖g(i)(ρ)‖L∞(Q)+‖∇μ‖L∞(0,T;L6(Ω)3)+‖∂tμ‖L2(0,T;V)+‖B[ρ]‖H1(0,T;L2(Ω))∩L∞(Q)∩L2(0,T;V)≤K∗1∀u∈Uad. | (2.31) |
According to Theorem 2.1, the control-to-state mapping S:Uad∋u↦(ρ,μ) is well defined. We now study its stability properties. We have the following result.
THEOREM 2.2: Suppose that (A1)-(A5) are fulfilled, and let ui∈Uad, i=1, 2, be given and (ρi,μi)=S(ui), i=1, 2, be the associated solutions to the state system (1.2)-(1.5). Then there exists a contant K∗2>0, which depends only on the data of the problem, such that, for every t∈(0,T],
‖ρ1−ρ2‖H1(0,t;H)∩L∞(0,t;L6(Ω))+‖μ1−μ2‖H1(0,t;H)∩L∞(0,t;V)∩L2(0,t;W)≤K∗2‖u1−u2‖L2(0,t;H). | (2.32) |
PROOF: Taking the difference of the equations satisfied by (ρi,μi), i=1, 2, and setting u:=u1−u2, ρ:=ρ1−ρ2, μ:=μ1−μ2, we first observe that we have almost everywhere in Q the identities
(1+2g(ρ1))∂tμ+g′(ρ1)∂tρ1μ−Δμ+2(g(ρ1)−g(ρ2))∂tμ2=u−(g′(ρ1)−g′(ρ2))∂tρ1μ2−g′(ρ2)μ2∂tρ, | (2.33) |
∂tρ+F′(ρ1)−F′(ρ2)+B[ρ1]−B[ρ2]=g′(ρ1)μ+(g′(ρ1)−g′(ρ2))μ2, | (2.34) |
as well as
∂nμ=0a.e.onΣ,μ(⋅,0)=ρ(⋅,0)=0a.e.inΩ. | (2.35) |
Let t∈(0,T] be arbitrary. In the following, we repeatedly use the global estimates (2.16), (2.17), and (2.31), without further reference. Moreover, we denote by c > 0 constants that may depend on the given data of the state system, but not on the choice of u1, u2∈Uad; the meaning of c may change between and even within lines. We establish the validity of (2.32) in a series of steps.
STEP 1: To begin with, we first observe that
(1+2g(ρ1))μ∂tμ+g′(ρ1)∂tρ1μ2=∂t((12+g(ρ1))μ2). |
Hence, multiplying (2.33) by μ and integrating over Qt and by parts, we obtain that
∫Ω(12+g(ρ1(t)))μ2(t)dx+∫Qt|∇μ|2dxds≤3∑j=1|Ij|, | (2.36) |
where the expressions Ij, j=1, 2, 3, defined below, are estimated as follows: first, we apply (A3), the mean value theorem, and Hälder’s and Young’s inequalities, to find, for every γ>0 (to be chosen later), that
I1:=−2∫Qt(g(ρ1)−g(ρ2))∂tμ2μdxds≤c∫t0‖∂tμ2(s)‖6‖μ(s)‖3‖ρ(s)‖2ds≤γ∫t0‖μ(s)‖2Vds+cγ∫t0‖∂tμ2(s)‖2V‖ρ(s)‖2Hds, | (2.37) |
where it follows from (2.31) that the mapping s↦‖∂tμ2(s)‖2V belongs to L1(0, T). Next, we see that
I2:=∫Qt(u−(g′(ρ1)−g′(ρ2))∂tρ1μ2)μdxds≤c∫Qt(|u|+|ρ|)|μ|dxds≤c∫Qt(u2+ρ2+μ2)dxds. | (2.38) |
Finally, Young’s inequality yields that
I3:=−∫Qtg′(ρ2)μ2ρtμdxds≤γ∫Qtρ2tdxds+cγ∫Qtμ2dxds. | (2.39) |
Combining (2.36)-(2.39), and recalling that g(ρ1) is nonnegative, we have thus shown the estimate
12‖μ(t)‖2H+(1−γ)∫t0‖μ(s)‖2Vds≤γ∫Qtρ2tdxds+c∫Qtu2dxds+c(1+γ−1)∫t0(‖μ(s)‖2H+(1+‖∂tμ2(s)‖2V)‖ρ(s)‖2H)ds. | (2.40) |
Next, we add ρ on both sides of (2.34) and multiply the result by ρt. Integrating over Qt, using the Lipschitz continuity of F′ (when restricted to [ρ∗,ρ∗]), (2.12) and Young’s inequality, we easily find the estimate
(1−γ)∫Qtρ2tdxds+12‖ρ(t)‖2H≤cγ∫Qt(ρ2+μ2)dxds. | (2.41) |
Therefore, combining (2.40) with (2.41), choosing γ>0 small enough, and invoking Gronwall’s lemma, we have shown that
‖μ‖L∞(0,t;H)∩L2(0,t;V)+‖ρ‖H1(0,t;H)≤c‖u‖L2(0,t;H)∀t∈(0,T]. | (2.42) |
STEP 2: Next, we multiply (2.34) by ρ|ρ| and integrate over Qt. We obtain
13‖ρ(t)‖33≤3∑j=1|Jj|, | (2.43) |
where the expressions Jj, 1≤j≤3, are estimated as follows: at first, we simply have
J1: = ∫Qt(−F′(ρ1)+F′(ρ2)+μ2(g′(ρ1)−g′(ρ2)))ρ|ρ|dxds≤c∫t0‖ρ(s)‖33ds. | (2.44) |
Moreover, invoking (2.42), Hälder’s inequality, as well as the global bounds once more,
J2:=∫Qtμg′(ρ1)ρ|ρ|dxds≤c∫t0‖μ(s)‖6‖ρ(s)‖2‖ρ(s)‖3ds≤∫t0‖ρ(s)‖33ds+c∫t0‖μ(s)‖3/2V‖ρ(s)‖3/2Hds≤∫t0‖ρ(s)‖33ds+c‖ρ‖3/2L∞(0,t;H)‖μ‖3/2L3/2(0,t;V)≤∫t0‖ρ(s)‖33ds+c‖ρ‖3/2L∞(0,t;H)‖μ‖3/2L2(0,t;V)≤∫t0‖ρ(s)‖3ds+c‖u‖3L2(0,t;H). | (2.45) |
In addition, condition (2.5), Hälder’s inequality, and (2.42), yield that
J3:=−∫Qt(B[ρ1]−B[ρ2])ρ|ρ|dxds≤c∫t0‖ρ(s)‖3‖ρ(s)‖2‖B[ρ1](s)−B[ρ2](s)‖6ds≤csup0≤s≤t‖ρ(s)‖3‖ρ‖L∞(0,t;H)∫t0‖ρ(s)‖Hds≤16sup0≤s≤t‖ρ(s)‖33+c‖u‖3L2(0,t;H). | (2.46) |
Combining the estimates (2.43)-(2.46), and invoking Gronwall’s lemma, we can easily infer that
‖ρ‖L∞(0,t;L3(Ω))≤c‖u‖L2(0,t;H)for all t∈(0,T]. | (2.47) |
Step 3: With the above estimates proved, the road is paved for multiplying (2.33) by μt. Integration over Qt yields that
∫Qt(1+2g(ρ1))μ2tdxds+12‖∇μ(t)‖2H≤5∑j=1|Kj|, | (2.48) |
where the expressions Kj, 1≤j≤5, are estimated as follows: at first, using the global bounds and Young’s inequality, we have for every γ>0 (to be specified later) the bound
K1:=−∫Qtg′(ρ1)∂tρ1μμtdxds≤γ∫Qtμ2tdxds+cγ∫Qtμ2dxds≤γ∫Qtμ2tdxds+cγ‖u‖2L2(0,t;H). | (2.49) |
Next, thanks to the mean value theorem, and employing (2.31) and (2.47), we find that
K2:=−2∫Qt(g(ρ1)−g(ρ2))∂tμ2μtdxds≤c∫Qt|ρ||∂tμ2||μt|dxds≤c∫t0‖ρ(s)‖3‖∂tμ2(s)‖6‖μt(s)‖2ds≤γ∫Qtμ2tdxds+cγ‖ρ‖2L∞(0,t;L3(Ω))∫t0‖∂tμ2(s)‖2Vds≤γ∫Qtμ2tds+cγ‖u‖2L2(0,t;H). | (2.50) |
Moreover, we infer that
K3:=∫Qtuμtdxds≤γ∫Qtμ2tdxds+cγ‖u‖2L2(0,t;H), | (2.51) |
as well as, invoking the mean value theorem once more,
K4:=−∫Qt(g′(ρ1)−g′(ρ2))∂tρ1μ2μtdxds≤c∫Qt|ρ||μt|dxds≤γ∫Qtμ2tdxds+cγ‖u‖2L2(0,t;H), | (2.52) |
and, finally, using (2.42) and Young’s inequality,
K5:=−∫Qtg′(ρ2)μ2ρtμtdxds≤c∫Qt|ρt||μt|dxds≤γ∫Qtμ2tdxds+cγ‖ρ‖2H1(0,t;H)≤γ∫Qtμ2tdxds+cγ‖u‖2L2(0,t;H). | (2.53) |
Now we combine the estimates (2.48)-(2.53) and choose > 0 appropriately small. It then follows that
μH1(0,t;H)∩L∞(0,t;V)≤c‖u‖L2(0,t;H). | (2.54) |
Finally, we come back to (2.33) and employ the global bounds (2.16), (2.17), (2.31), and the estimates shown above, to conclude that
‖Δμ‖L2(0,t;H)≤c(‖μt‖L2(0,t;H)+‖μ‖L2(0,t;H)+‖ρt‖L2(0,t;H)+‖ρ‖L2(0,t;H)+‖u‖L2(0,t;H))+c‖ρ∂tμ2‖L2(0,t;H)≤c‖u‖L2(0,t;H), | (2.55) |
where the last summand on the right-hand side was estimated as follows:
∫Qt|ρ|2|∂tμ2|2dxds≤c∫t0‖∂tμ2(s)‖26‖ρ(s)‖23ds≤c‖ρ‖2L∞(0,t;L3(Ω))∫t0‖∂tμ2(s)‖2Vds≤c‖u‖2L2(0,t;H). |
Invoking standard elliptic estimates, we have thus shown that
‖μ‖L2(0,t;W)≤c‖u‖L2(0,t;H). | (2.56) |
STEP 4: It remains to show the L∞(0, t; L6(Ω))-estimate for ρ. To this end, we multiply (2.34) by ρ|ρ|4 and integrate over Qt. It follows that
16‖ρ(t)‖66≤3∑j=1|Lj|, | (2.57) |
where quantities Lj, 1≤j≤3, are estimated as follows: at first, we obtain from the global estimates (2.17) and (2.31), that
L1:=∫Qt(−F′(ρ1)+F′(ρ2)+μ2(g′(ρ1)−g′(ρ2)))ρ|ρ|4dxds≤c∫t0‖ρ(s)‖66ds. | (2.58) |
Moreover, from (2.54) and Hälder’s and Young’s inequalities we conclude that
L2:=∫Qtg′(ρ1)μρ|ρ|4dxds≤c∫t0‖μ(s)‖6‖ρ(s)‖56ds≤c‖μ‖L∞(0,t;V)‖ρ‖5L5(0,t;L6(Ω))≤c‖μ‖6L∞(0,t;V)+c‖ρ‖6L5(0,t;L6(Ω))≤c‖u‖6L2(0,t;H)+c∫t0‖ρ(s)‖66ds. | (2.59) |
Finally, we employ (2.5) and (2.42) to infer that
L3:=−∫Qt(B[ρ1]−B[ρ2])ρ|ρ|4dxds≤c∫t0‖B(ρ1](s)−B[ρ2](s)‖6‖ρ(s)‖56ds≤csup0≤s≤t‖ρ(s)‖56∫t0‖ρ(s)‖Hds≤112sup0≤s≤t‖ρ(s)‖66+c‖u‖6L2(0,t;H). | (2.60) |
Combining the estimates (2.57)-(2.60), and invoking Gronwall’s lemma, then we readily find the estimate
‖ρ‖L∞(0,t;L6(Ω))≤c‖u‖L2(0,t;H), |
which concludes the proof of the assertion.
In this section, we prove the relevant differentiability properties of the solution operator S. To this end, we introduce the spaces
X:=H1(0,T;H)∩L∞(Q),Y:=H1(0,T;H)×(L∞(0,T;H)∩L2(0,T;V)), |
endowed with their natural norms
‖u‖X:=‖u‖H1(0,T;H)+‖u‖L∞(Q)∀u∈X,‖(ρ,μ)‖Y:=‖ρ‖H1(0,T;H)+‖μ‖L∞(0,T;H)+‖μ‖L2(0,T;V)∀(ρ,μ)∈Y, |
and consider the control-to-state operator S as a mapping between Uad⊂X and Y. Now let ˉu∈Uad be fixed and put (ˉρ,ˉu):=S(ˉu). We then study the linearization of the state system (1.2)-(1.5) at ˉu, which is given by:
(1+2g(ˉρ))ηt+2g′(ˉρ)ˉutξ+g′(ˉρ)ˉρtη+ˉug″(ˉρ)ˉρtξ+ˉug′(ˉρ)ξt−Δη=ha.e.in Q, | (3.1) |
ξt+F″(ˉρ)ξ+DB[ˉρ](ξ)=ˉug″(ˉρ)ξ+g′(ˉρ)ηa.e.in Q, | (3.2) |
∂nη=0a.e.on Σ, | (3.3) |
η(0)=ξ(0)=0a.e.in Ω. | (3.4) |
Here, h∈X must satisfy ˉu+ˉλh∈Uad for some ˉλ>0. Provided that the system (3.1)-(3.4) has for any such h a unique solution pair (ξ,η), we expect that the directional derivative δS(ˉu;h) of S at ˉu in the direction h (if it exists) ought to be given by (ξ,η). In fact, the above problem makes sense for every h∈L2(Q), and it is uniquely solvable under this weaker assumption.
THEOREM 3.1: Suppose that the general hypotheses (A1)-(A5) are satisfied and let h∈L2(Q). Then, the linearized problem (3.1)-(3.4) has a unique solution (ξ,η) satisfying
ξ∈H1(0,T;H)∩L∞(0,T;L6(Ω)), | (3.5) |
η∈H1(0,T;H)∩L∞(0,T;V)∩L2(0,T;W). | (3.6) |
PROOF: We first prove uniqueness. Since the problem is linear, we take h=0 and show that (ξ,η)=(0, 0). We add η and ξ to both sides of equations (3.1) and (3.2), respectively, then multiply by η and ξt, integrate over Qt, and sum up. By observing that
(1+2g(ˉρ))ηtη+g′(ˉρ)ˉρt|η|2=∂t[(12+g(ˉρ))|η|2], |
and recalling that g≥0, we obtain
12∫Ω|η(t)|2dx+∫t0‖η(s)‖2Vds+12 ∫Ω|ξ(t)|2dx+∫Qt|ξt|2dxds≤3∑j=1Hj, |
where the terms Hj are defined and estimated as follows. We have
H1:=−∫Qt2g′(ˉρ)ˉutξηdxds≤c∫t0‖ˉut(s)‖3‖ξ(s)‖2‖η(s)‖6ds≤12∫t0‖η(s)‖2Vds+c∫t0‖ˉut(s)‖2V‖ξ(s)‖22ds, |
and we notice that the function s↦‖ˉut(s)‖2V belongs to L1(0, T), by (2.30) for ˉu. Next, we easily have the estimate
H2:=∫Qt(η−ˉug″(ˉρ)ˉρtξ−ˉug′(ˉρ)ξt)ηdxds≤14∫Qt|ξt|2dxds+c∫Qt(|ξ|2+|η|2)dxds. |
Finally, recalling (2.8), it is clear that
H3:=∫Qt((ξ+ˉug″(ˉρ)−F″(ˉρ))ξ−DB[ˉρ](ξ)+g′(ˉρ)η)ξtdxds≤14∫Qt|ξt|2dxds+c∫Qt(|ξ|2+|η|2)dxds. | (3.7) |
Therefore, it suffices to collect these inequalities and apply Gronwall’s lemma in order to conclude that ξ=0 and η=0.
The existence of a solution is proved in several steps. First, we introduce an approximating problem depending on the parameter ε∈(0,1). Then, we show well-posedness for this problem and perform suitable a priori estimates. Finally, we construct a solution to problem (3.1)-(3.4) by letting ε tend to zero. For the sake of simplicity, in performing the uniform a priori estimates, we denote by c > 0 different constants that may depend on the data of the system but not on ε∈(0,1); the actual value of c may change within formulas and lines. On the contrary, the symbol cε stands for (different) constants that can depend also on ε. In particular, cε is independent of the parameter δ that enters an auxiliary problem we introduce later on.
STEP 1: We approximate ˉρ and ˉμ by suitable ρε,με∈C∞(ˉQ) as specified below. For every ε∈(0,1), it holds that
ρ∗∗≤ρε≤ρ∗∗ in ˉQ and ‖ρεt‖L∞(Q)+‖με‖H1(0,T;L3(Ω))∩L∞(Q)≤C∗, | (3.8) |
for some constants ρ∗∗,ρ∗∗∈(0,1) and C∗>0; as ε↘0, we have
ρε→ˉρ,ρεt→ˉρt,με→ˉμ,in Lp(Q), for every p<+∞ and a.e.in Q,andμεt→ˉμtin L2(0,T;L3(Ω)). | (3.9) |
In order to construct regularizing families as above, we can use, for instance, extension outside Q and convolution with mollifiers.
Next, we introduce the approximating problem of finding (ξε,ηε) satisfying
ξεt+F″(ˉρ)ξε+DB[ˉρ](ξε)=ˉμg″(ˉρ)ξε+g′(ˉρ)ηεa.e.in Q, | (3.10) |
(1+2g(ρε))ηεt+g′(ρε)ρεtηε+2g′(ˉρ)μεtξε+ˉμg″(ˉρ)ˉρtξε+ˉμg′(ˉρ)ξεt−Δηε=ha.e.in Q, | (3.11) |
∂nηε=0a.e.on Σ, | (3.12) |
ηε(0)=ξε(0)=0a.e.in Ω. | (3.13) |
In order to solve (3.10)-(3.13), we introduce the spaces
V:=H×V and H:=H×H, |
and present our problem in the form
ddt(ξ,η)+A(ξ,η)=f and (ξ,η)(0)=(0,0), |
in the framework of the Hilbert triplet (V, H, V′). We look for a weak solution and aim at applying [1, Thm. 3.2, p. 256]. To this end, we have to split Aε in the form Qε+Rε, where Q" is the uniformly elliptic principal part and the remainder Rε satisfies the requirements [1, (4.4)-(4.5), p. 259]. We notice at once that these conditions are trivially fulfilled whenever
Rε=(Rε1,Rε2)∈L(L2(0,T;H),L2(0,T;H)), | (3.14) |
|∫Qt(Rε1(v,w)v+Rε2(v,w)w)dxds|≤CRε∫Qt(|v|2+|w|2)ds, | (3.15) |
for some constant CRε, and every v,w∈L2(0,T;H) and t∈[0,T]. In order to present (3.10)-(3.13) in the desired form, we multiply (3.11) by aε:=1/(1+2g(ρε)) and notice that
−aεΔηε=−div(aεΔηε)+∇aε⋅∇ηε. |
As aε≥α:=1/(1+2supg) and ∇aε is bounded, we can fix a real number λε>0 such that
∫Ωaε(t)|∇w|2+(∇aε(t)⋅∇w)w+λε|w|2)dx≥α2‖w‖2V | (3.16) |
for every w∈V and t∈[0,T]. Next, we replace ξεt in (3.11) by using (3.10). Therefore, we see that a possible weak formulation of (3.10)-(3.12) is given by
∫Ωξεt(t)vdx+V′⟨ηεt(t),w⟩V+V′⟨Qε(t)(ξε,ηε)(t),(v,w)⟩V+∫Ω(Rε1(ξε,ηε)(t)v+Rε2(ξε,ηε)(t)w)dx=∫Ωaε(t)h(t)wdxfor a.a. t∈(0,T) and every (v,w)∈V, | (3.17) |
where the symbols ⟨⋅,⋅⟩ stand for the duality pairings and Qε and Rεi have the meaning explained below. The time-dependent operator Qε(t) from V into V′ is defined by
V′⟨Qε(t)(ˆv,ˆw),(v,w)⟩V=∫Ω(ˆvv+aε(t)∇ˆw⋅∇w+(∇aε(t)⋅∇ˆw)w+λεˆww)dx | (3.18) |
for every (ˆv,ˆw),(v,w)∈V and t∈[0,T]. By construction, the bilinear form given by the right-hand side of (3.18) is continuous on V×V, depends smoothly on time, and is V-coercive uniformly with respect to t (see (3.16)). The operators
Rεi∈L(L2(0,T;H),L2(0,T;H)) |
account for the term λεηε that has to be added also to the right-hand side of (3.11) and for all the contributions to the equations that have not been considered in the principal part. They have the form
(Rεi(v,w))(t)=aεi1(t)v+aεi2(t)w+aεi3(t)(DB[ˉρ](v))(t) | (3.19) |
for (v,w)∈L2(0,T;H), with some coefficients aεij∈L∞(Q). Therefore, by virtue of (2.8), we see that
∫Qt(Rε1(v,w)v+Rε1(v,w)w)dxds≤c∫Qt(|v|2+|w|2)dxds+c‖DB[ˉρ](v)‖2L2(Qt)≤c∫Qt(|v|2+|w|2)dxds, |
for every (v,w)∈L2(0,T;H) and every t∈[0,T]. Thus, the conditions (3.14)-(3.15) are fulfilled, and the result of [1] mentioned above can be applied. We conclude that the Cauchy problem for (3.17) has a unique solution (ξε,ηε) satisfying
(ξε,ηε)∈H1(0,T;V′)∩L2(0,T;V),i.e., ξε∈H1(0,T;H)andηε∈H1(0,T;V′)∩L2(0,T;V). |
On the other hand, this solution has to satisfy
⟨∂tηε,w⟩+∫Ωaε∇ηε⋅∇wdx=∫Ωϕεw dx a.e.in (0, T), forevery w∈V, |
for some ϕε∈L2(Ω). From standard elliptic regularity, it follows that ηε∈H1(0,T;H)∩L2(0,T;W).
In the next steps, besides of Young’s inequality, we make repeated use of the global estimates (2.16), (2.17), and (2.31), for ˉρ and ˉμ, without further reference.
STEP 2: For convenience, we refer to Eqs. (3.10)-(3.12) (using the language that is proper for strong solutions), but it is understood that they are meant in the variational sense (3.17). We add ξε and ηε to both sides of (3.10) and (3.11), respectively; then, we multiply the resulting equalities by ξεt and ηε, integrate over Qt, and sum up. By observing that
(1+2g(ρε))ηεtηε+g′(ρε)ρεt|ηε|2=∂t[(12+g(ρε))|ηε|2], |
and recalling that g≥0, we obtain
12∫Ω|ξε(t)|2dx+∫Qt|ξεt|2dxds+12∫Ω|ηε(t)|2dx+∫t0‖ηε(s)‖2Vds≤3∑j=1Ij, |
where the terms Ij are defined and estimated as follows. In view of (2.8), we first infer that
I1:=∫Qt(ξε−F″(ˉρ)ξε−DB[ˉρ](ξε)+ˉμg″(ˉρ)ξε+g′(ˉρ)ηε)ξεtdxds≤14∫Qt|ξεt|2dxds+c∫Qt(|ξε|2+|ηε|2)dxds. |
Next, we have
I2:=∫Qt(ηε−ˉμg″(ˉρ)ˉρtξε−ˉμg′(ˉρ)ξεt+h)ηεdxds≤14∫Qt|ξεt|2dxds+c∫Qt(|ξε|2+|ηε|2)dxds+c. |
Finally, by virtue of the H¨older and Sobolev inequalities, we have
I3:=−∫Qt2g′(ˉρ)μεtξεηεdxds≤c∫t0‖μεt(s)‖3‖ξε(s)‖2‖ηε(s)‖6ds≤12∫t0‖ηε(s)‖2Vds+c∫t0‖ηε(s)‖23‖ξε(s)‖22ds. |
At this point, we recall all the inequalities we have proved, notice that (3.8) implies that the function s↦‖μεt(s)‖23 is bounded in L1(0, T), and apply the Gronwall lemma. We obtain
‖ξε‖H1(0,T;H)+‖ηε‖L∞(0,T;H)∩L2(0,T;V)≤c. | (3.20) |
STEP 3: We would now like to test (3.10) by (ξε)5. However, this function is not admissible, unfortunately. Therefore, we introduce a suitable approximation. To start with, we consider the Cauchy problem
ˆξt+bˆξ+L(ˆξ)=fε and ˆξ(0)=0, | (3.21) |
where we have set, for brevity,
b:=F″(ˉρ)−ˉμg″(ˉρ),L:=DB[ˉρ], and fε:=g′(ˉρ)ηε. | (3.22) |
By (3.10), ˆξ:=ξε is a solution belonging to H1(0, T; H). On the other hand, such a solution is unique. Indeed, multiplying by ˆξ the corresponding homogeneous equation (i. e., fε is replaced by 0), and invoking (2.8) and Gronwall’s lemma, one immediately obtains that ˆξ=0. We conclude that ˆξ:=ξε is the unique solution to (3.21).
At this point, we approximate ξε by the solution to a problem depending on the parameter δ∈(0,1), in addition. Namely, we look for ξεδ satisfying the parabolic-like equation
ξεδt−δΔξεδ+bδξεδ+L(ξεδ)=fε, | (3.23) |
complemented with the Neumann boundary condition ∂nξεδ=0 and the initial condition ξεδ=0. In (3.23), bδ is an approximation of b belonging to C∞(¯Q) that satisfies
‖bδ‖L∞(Q)≤c, and bδ→b a.e.in Q as δ↘0. | (3.24) |
This problem has a unique weak solution ξεδ∈H1(0,T;V′)∩L2(0,T;V), as one easily sees by arguing as we did for the more complicated system (3.10)-(3.13) and applying [1, Thm. 3.2, p. 256].
We now aim to show that ξεδ is bounded. To this end, we introduce the operator Aδ∈L(V,V′) defined by
⟨Aδv,w⟩:=∫Ω(δ∇v⋅∇w+vw)dxfor every v,w∈V, |
and observe that Aδ is an isomorphism. Moreover, Eq. (3.23), complemented with the boundary and initial conditions, can be written as
ξεδt+Aδξεδ=fεδ:=fε−(1+bδ)ξεδ+L(ξεδ)andξεδ(0)=0 | (3.25) |
Now, by also accounting for (2.9), we notice that fε,ξεδ,bδξεδ, and L(ξεδ), all belong to L2(0, T; V). Hence, fεδ∈L2(0,T;V), so that Aδfεδ∈L2(0,T;V′), and we can consider the unique solution ζεδ∈H1(0,T;V′)∩L2(0,T;V) to the problem
ζεδt+Aδζεδ=Aδfεδ and ζεδ(0)=0 |
Now, A−1δζεδ satisfies
(A−1δζεδ)t+Aδ(A−1δζεδ)=A−1δAδfεδ=fεδ and (A−1δζεδ)(0)=0, |
so that a comparison with (3.25) shows that ξεδ=A−1δζεδ, by uniqueness. Since ζεδ∈L∞(0,T;H), and A−1δ(H)=W by elliptic regularity, we deduce that ξεδ∈L∞(0,T;W). Therefore, ξεδ is bounded, as claimed.
Consequently, (ξεδ)5 is an admissible test function, since it clearly belongs to the space L2(0, T; V). By multiplying (3.23) by (ξεδ)5 and integrating over Qt, we obtain that
16∫Ω|ξεδ(t)|6dx+5δ∫Qt|ξεδ|4|∇ξεδ|2dxds=3∑j=1Kj, |
where the terms Kj are defined and estimated as follows. First, recalling (3.24), we deduce that
K1:=−∫Qtbδξεδ(ξεδ)5dxds≤c∫Qt|ξε|6dxds. |
On the other hand, Hälder’s inequality, and assumption (2.8) with p=6, imply that
K2:=−∫QtL(ξεδ)(ξεδ)5dxds≤c‖Lξεδ‖L6(Qt)‖(ξεδ)5‖L6/5(Qt)≤c‖ξεδ‖L6(Qt)‖ξεδ‖5L6(Qt)=c∫Qt|ξεδ|6dxds. |
Finally, also invoking Sobolev’s inequality, we see that
K3:=∫Qtfε(ξεδ)5dxds≤c∫t0‖ηε(s)‖6‖(ξεδ(s))5‖6/5ds≤c∫t0‖ηε(s)‖6‖ξεδ(s)‖56ds≤c∫t0‖ηε(s)‖V(1+‖ξεδ(s)‖66)ds. |
Collecting the above estimates, and noting that the function s↦‖ηε(s)‖V is bounded in L1(0, T) by (3.20), we can apply the Gronwall lemma to conclude that
‖ξεδ‖L∞(0,T;L6(Ω))≤c. | (3.26) |
At this point, we quickly show that ξεδ converges to ξε as δ↘0, at least for a subsequence. Indeed, one multiplies (3.23) first by ξεδ, and then by ξεt, and proves that
‖ξεδ‖H1(0,T;H)∩L∞(0,T;V)≤cε, |
uniformly with respect to δ. Then, by weak compactness and (3.24) (which implies convergence of bδ to b in Lp(Q) for every p < +∞), it is straightforward to see that ξεδ converges to a solution ˆξ to the problem associated with (3.21). As ˆξ=ξε, we have proved what we have claimed. This, and (3.26), yield that
‖ξε‖L∞(0,T;L6(Ω))≤c. | (3.27) |
STEP 4: At this point, we can multiply (3.11) by ηεt and integrate over Qt. By recalling that g≥0, we obtain
∫Qt|ηεt|2dxds+12∫Ω|∇ηε(t)|2dx≤3∑j=1Lj, |
where each term Lj is defined and estimated below. First, by taking advantage of (3.27) and (3.8) for μεt, we have
L1:=−∫Qt2g′(ˉρ)μεtξεηεtdxds≤c∫t0‖μεt(s)‖3‖ξε(s)‖6‖ηεt(s)‖2ds≤c∫t0‖μεt(s)‖3‖ηεt(s)‖2ds≤14∫Qt|ηεt|2dxds+c∫T0‖μεt(s)‖23ds≤14∫Qt|ηεt|2dxds+c. |
Next, using (3.8) for ρεt and (3.20), we obtain that
L2:=−∫Qtg′(ρε)ρεtηεηεtdxds≤14∫Qt|ηεt|2dxds+c∫Qt|ηε|2dxds≤14∫Qt|ηεt|2dxds+c. |
Finally, in view of (3.20), we have
L3:=∫Qt(−ˉμg″(ˉρ)ˉρtξε−ˉμg′(ˉρ)ξεt+h)ηεtdxds≤14∫Qt|ηεt|2dxds+c∫Qt(|ξε|2+|ξεt|2+1)dxds≤14∫Qt|ηεt|2dxds+c. |
By collecting the above estimates, we conclude that
‖ηεt‖L2(0,T;H)+‖ηε‖L2(0,T;V)≤c. | (3.28) |
As a consequence, we can estimate Δηε in L2(Q), just by comparison in (3.11). Using standard elliptic regularity, we deduce that
‖ηε‖L2(0,T;W)≤c. | (3.29) |
STEP 5: At this point, we are ready to prove the existence part of the statement. Indeed, the estimates (3.20) and (3.27)-(3.29) yield that
ξε→ξweaklystarinH1(0,T;H)∩L∞(0,T;L6(Ω)),ηε→η weaklystarinH1(0,T;H)∩L∞(0,T;V)∩L2(0,T;W), |
as ε↘0, at least for a subsequence. By accounting for (3.9) and the Lipschitz continuity of g and g′, it is straightforward to see that (ξ,η) is a solution to problem (3.1)-(3.4). This completes the proof.
We are now prepared to show that S is directionally differentiable. We have the following result:
THEOREM 3.2: Suppose that the general hypotheses (A1)-(A5) are satisfied, and let ˉu∈Uad be given and (ˉρ,ˉμ)=S(ˉu). Moreover, let h∈X be a function such that ˉu+ˉλh∈Uad for some ˉλ > 0. Then the directional derivative δS(ˉu;h) of S at ˉu in the direction h exists in the space (Y,‖⋅‖Y), and we have δS(ˉu;h)=(ξ,η), where (ξ,η) is the unique solution to the linearized system (3.1)-(3.4).
PROOF: We have ˉu+λh∈Uad for 0<λ≤ˉλ, since Uad is convex. We put, for any such λ,
uλ:=ˉu+λh,(ρλ,μλ):=S(uλ),yλ:=ρλ−ˉρ−λξ,zλ:=μλ−ˉu−λη. |
Notice that (ρλ,μλ) and (ˉρ,ˉμ) fulfill the global bounds (2.16), (2.17), and (2.31), and that (yλ,zλ)∈Y 2 Y for all λ∈[0,ˉλ]. Moreover, by virtue of Theorem 2.2, we have the estimate
‖ρλ−ˉρ‖H1(0,t;H)∩L∞(0,t;L6(Ω))+‖μλ−ˉμ‖H1(0,t;H)∩L∞(0,t;V)∩L2(0,t;W)≤K∗2λ‖h‖L2(0,t;H). | (3.30) |
We also notice that, owing to (2.16) and the assumptions on F and g, it follows from Taylor’s theorem that
|F′(ρλ)−F′(ˉρ)−λF″(ˉρ)ξ|≤c|yλ|+c|ρλ−ˉρ|2a.e.in Q, | (3.31) |
|g(ρλ)−g(ˉρ)−λg′(ˉρ)ξ|≤c|yλ|+c|ρλ−ˉρ|2a.e.in Q, | (3.32) |
|g′(ρλ)−g′(ˉρ)−λg″(ˉρ)ξ|≤c|yλ|+c|ρλ−ˉρ|2a.e.in Q, | (3.33) |
where, here and in the remainder of the proof, we denote by c constants that may depend on the data of the system but not on λ∈[0,ˉλ]; the actual value of c may change within formulas and lines. Moreover, by the Fréchet differentiability of B (recall assumption (A4)(vi) and the fact that, for ¯v,v∈L2(Q), the restrictions of B[v] and DB[¯v](v) to Qt depend only on v|Qt), we have (cf. (3.30))
‖B[ρλ]−B[ˉρ]−λDB[ˉρ](ξ)‖L2(Qt)≤c‖yλ‖L2(Qt)+R(λ‖h‖L2(Qt)), | (3.34) |
with a function R : (0; +∞)→(0; +∞) satisfying limσ↘0R(σ)/σ=0. As we want to prove that δS(ˉu;h)=(ξ,η), according to the definition of directional differentiability, we need to show that
0=limλ↘0‖S(ˉu+λh)−S(ˉu)−λ(ξ,η)‖Yλ=limλ↘0‖yλ‖H1(0,T;H)+‖zλ‖L∞(0,T;H)∩L2(0,T;V)λ. | (3.35) |
To begin with, using the state system (1.2)-(1.5) and the linearized system (3.1)-(3.4), we easily verify that for 0<λ≤ˉλ the pair (zλ,yλ) is a strong solution to the system
(1+2g(ˉρ))zλt+g′(ˉρ)ˉρtzλ+ˉμg′(ˉρ)zλt−Δzλ=−2(g(ρλ)−g(ˉρ))(μλt−ˉμt)−2ˉμt(g(ρλ)−g(ˉρ)−λg′(ˉρ)ξ)−ˉμˉρt(g′(ρλ)−g′(ˉρ)−λg″(ˉρ)ξ)−ˉμ(g′(ρλ)−g′(ˉρ))(ρλt−ˉρt)−(μλ−ˉμ)[(g′(ρλ)−g′(ˉρ))ˉρt+g′(ρλ)(ρλt−ˉρt)]a.e.in Q, | (3.36) |
yλt=−(F′(ρλ)−F′(ˉρ)−λF″(ˉρ)ξ)−(B[ρλ]−B[ˉρ]−λDB[ˉρ](ξ))+g′(ˉρ)zλ+ˉμ(g′(ρλ)−g′(ˉρ)−λg″(ˉρ)ξ)+(μλ−ˉμ)(g′(ρλ)−g′(ˉρ))a.e.in Q, | (3.37) |
∂nzλ=0a.e. on Σ, | (3.38) |
zλ(0)=yλ(0)=0a.e.in Ω. | (3.39) |
In the following, we make repeated use of the mean value theorem and of the global estimates (2.16), (2.17), (2.31), and (3.30), without further reference. For the sake of a better readability, we will omit the superscript λ of the quantities yλ,zλ during the estimations, writing it only at the end of the respective estimates.
STEP 1: Let t∈(0,T] be fixed. First, observe that
∂t((12+g(ˉρ))z2)=(1+2g(ˉρ))zzt+g′(ˉρ)ˉρtz2. |
Hence, multiplication of (3.36) by z and integration over Qt yields the estimate
∫Ω(12+g(ˉρ(t)))z2(t)dx+∫Qt|∇z|2dxds≤c7∑j=1|Ij|, | (3.40) |
where the quantities Ij, 1≤j≤7, are specified and estimated as follows: at first, Young’s inequality shows that, for every λ>0 (to be chosen later),
I1:=−∫Qtˉμg′(ˉρ)ytzdxds≤γ∫Qty2tdxds+cγ∫Qtz2dxds. | (3.41) |
Moreover, we have, by H¨older’s and Young’s inequalities and (3.30),
I2:=−2∫Qt(g(ρλ)−g(ˉρ))(μλt−ˉμt)zdxds≤c∫t0‖ρλ(s)−ˉρ(s)‖6‖μλt(s)−ˉμt(s)‖2‖z(s)‖3≤c‖ρλ−ˉρ‖L∞(0,t;L6(Ω))‖μλ−ˉμ‖H1(0,t;H)‖z‖L2(0,T;V)≤γ‖z‖2L2(0,T;V)+cγλ4. | (3.42) |
Next, we employ (3.32), the Hälder and Young inequalities, and (3.30), to infer that
I3:=−2∫Qtˉμt(g(ρλ)−g(ˉρ)−λg′(ˉρ)ξ)zdxds≤c∫Qt|ˉμt|(|y|+|ρλ−ˉρ|2)|z|dxds≤c∫t0‖ˉμt(s)‖6(‖y(s)‖2‖z(s)‖3+‖ρλ(s)−ˉρ(s)‖26‖z(s)‖2)ds≤γ∫t0‖z(s)‖2Vds+cγ∫t0‖ˉμt(s)‖2V‖y(s)‖2Hds+c∫t0‖ˉμt(s)‖2V‖z(s)‖2Hds+c‖ρλ−ˉρ‖4L∞(0,t;V)≤γ∫t0‖z(s)‖2Vds+(1+cγ)∫t0‖ˉμt(s)‖2V(‖y(s)‖2H+‖z(s)‖2H)ds+cλ4, | (3.43) |
where we observe that, in view of (2.17), the mapping s↦‖ˉμt(s)‖2V belongs to L1(0, T). Likewise, utilizing (2.17), (3.33), (3.30), and the Hälder and Young inequalities, it is straightforward to deduce that
I4:=−∫Qtˉμˉρt(g′(ρλ)−g′(ˉρ)−λg″(ˉρ)ξ)zdxds≤c∫Qt(|y|+|ρλ−ˉρ|2)|z|dxds≤c∫Qt(y2+z2)dxds+c∫t0‖ρλ(s)−ˉρ(s)‖24‖z(s)‖2ds≤c∫Qt(y2+z2)dxds+cλ4. | (3.44) |
In addition, arguing similarly, we have
I5:=−∫Qtˉμ(g′(ρλ)−g′(ˉρ))(ρλt−ˉρt)zdxds≤c∫t0‖ρλ(s)−ˉρ(s)‖6‖ρλt(s)−ˉρt(s)‖2‖z(s)‖3ds≤c‖ρλ−ˉρ‖L∞(0,t;L6(Ω))‖ρλ−ˉρ‖H1(0,t;H)‖z‖L2(0,t;V)≤γ∫t0‖z(s)‖2Vds+cγλ4, | (3.45) |
as well as
I6:=−∫Qtˉρt(μλ−ˉμ)(g′(ρλ)−g′(ˉρ))zdxds≤c∫Qt|μλ−ˉμ||ρλ−ˉρ||z|dxds≤c∫t0‖ρλ(s)−ˉρ(s)‖6‖μλ(s)−ˉμ(s)‖3‖z(s)‖2ds≤c∫Qtz2dxds+cλ4. | (3.46) |
Finally, we find that
I7:=−∫Qt(μλ−ˉμ)g′(ρλ)(ρλt−ˉρt)zdxds≤c∫t0‖μλ(s)−ˉμ(s)‖6‖ρλt(s)−ˉρt(s)‖2‖z(s)‖3ds≤c‖μλ−ˉμ‖L∞(0,t;V)‖ρλ−ˉρ‖H1(0,t;H)‖z‖L2(0,t;V)≤γ∫t0‖z(s)‖2Vds+cγλ4. | (3.47) |
In conclusion, combining the estimates (3.40)-(3.47), and choosing γ=18, we have shown that
12‖zλ(t)‖2H+12∫t0‖zλ(s)‖2Vds≤18∫Qt|yλt|2dxds+cλ4+c∫t0(1+‖ˉμt(s)‖2V)(‖yλ(s)‖2H+‖zλ(s)‖2H)ds. | (3.48) |
STEP 2: Let t∈(0,T] be fixed. We add y to both sides of (3.37), multiply the resulting identity by yt, and integrate over Qt to obtain
∫Qty2tdxds+12‖y(t)‖2H≤6∑j=1|Jj|, | (3.49) |
where the terms Jj, 1≤j≤6, are specified and estimated as follows: at first, we have, for every γ>0 (to be specified later),
J1:=∫Qtyytdxds≤γ∫Qty2tdxds+cγ∫Qty2dxds. | (3.50) |
Then, we employ (2.17), (2.31), (3.30), and (3.31), as well as Hölder’s and Young’s inequalities, to obtain the estimate
J2:=−∫Qt(F′(ρλ)−F′(ˉρ)−λF″(ˉρ)ξ)ytdxds≤c∫Qt(|y|+|ρλ−ˉρ|2)|yt|dxds≤c∫t0(‖y(s)‖2+‖ρλ(s)−ˉρ(s)‖24)‖yt(s)‖2ds≤γ∫Qty2tdxds+cγ∫Qty2dxds+cγλ4. | (3.51) |
By the same token, this time invoking (3.33), we find that
J3:=∫Qtˉμ(g′(ρλ)−g′(ˉρ)−λg″(ˉρ)ξ)ytdxds≤γ∫Qty2tdxds+cγ∫Qty2dxds+cγλ4. | (3.52) |
Moreover, we obviously have
J4:=∫Qtg′(ˉρ)zytdxds≤γ∫Qty2tdxds+cγ∫Qtz2dxds. | (3.53) |
Also, using (3.30) and the global bounds once more, we obtain that
J5:=∫Qt(μλ−ˉμ)(g′(ρλ)−g′(ˉρ))ytdxds≤c∫t0‖μλ(s)−ˉμ(s)‖6‖ρλ(s)−ˉρ(s)‖3‖yt(s)‖2ds≤γ∫Qty2tdxdx+cγλ4. | (3.54) |
Finally, invoking (3.34) and Young’s inequality, we have the estimate
J6:=−∫Qt(B[ρλ]−B[ˉρ]−λDB[ˉρ](ξ))ytdxds≤‖B[ρλ]−B[ˉρ]−λDB[ˉρ](ξ)‖L2(Qt)‖yt‖L2(Qt)≤γ∫Qty2tdxds+cγ‖y‖2L2(Qt)+cγ(R(λ‖h‖L2(Q)))2. | (3.55) |
Thus, combining the estimates (3.49)-(3.55), and choosing γ=18, we have shown that, for every t∈(0,T], we have the estimate
14∫Qt|yλt|2dxds+12‖yλ(t)‖2H≤c(∫t0‖yλ(s)‖2Hds+λ4+(R(λ‖h‖L2(Q)))2). | (3.56) |
STEP 3: We now add the estimates (3.48) and (3.56). It follows that, with suitable global constants c1 > 0 and c2 > 0, we have for every t∈(0,T] the estimate
‖zλ(t)‖2H+‖zλ‖2L2(0,t;V)+‖yλ(t)‖2H+‖yλt‖2L2(0,t;H)≤c1Z(λ)+c2∫t0(1+‖ˉμt(s)‖2V)(‖yλ(s)‖2H+‖zλ(s)‖2H)ds, | (3.57) |
where we have defined, for λ>0, the function Z by
Z(λ):=λ4+(R(λ‖h‖L2(Ω)))2. | (3.58) |
Recalling that the mapping s↦‖ˉμt(s)‖2V belongs to L1(0, T), we conclude from Gronwall’s lemma that, for every t∈(0,T],
‖yλ‖2H1(0,t;H)+‖zλ‖2L∞(0,t;H)∩L2(0,t;V)≤c1Z(λ)exp(c2∫T0(1+‖ˉμt(s)‖2V)ds)≤cZ(λ). | (3.59) |
Since limλ↘0Z(λ)/λ2=0 (recall (3.34)), we have finally shown the validity of (3.35). This concludes the proof of the assertion.
We are now in the position to derive the following result.
COROLLARY 3.3: Let the general hypotheses (A1)-(A5) be fulfilled and assume that ˉu∈Uad is a solution to the control problem (CP) with associated state (ˉρ,ˉμ)=S(ˉu). Then we have, for every v∈Uad,
β1∫Q(ˉρ−ρQ)ξdxdt+β2∫Q(ˉμ−μQ)ηdxdt+β3∫Qˉu(v−ˉu)dxdt≥0, | (3.60) |
where (ξ,η) denotes the (unique) solution to the linearized system (3.1)-(3.4) for h=v−ˉu.
PROOF: Let v∈Uad be arbitrary. Then h=v−ˉu is an admissible direction, since ˉu+λh∈Uad for 0<λ≤1. For any such λ, we have
0≤J(ˉu+λh,S(ˉu+λh))−J(ˉu,S(ˉu))λ≤J(ˉu+λh,S(ˉu+λh))−J(ˉu,S(ˉu+λh))λ+J(ˉu,S(ˉu+λh)−J(ˉu,S(ˉu))λ. |
It follows immediately from the definition of the cost functional J that the first summand on the right-hand side of this inequality converges to ∫Qβ3ˉuhdxdt as λ↘0. For the second summand, we obtain from Theorem 3.2 that
limλ↘0J(ˉu,S(ˉu+λh)−J(ˉu,S(ˉu))λ=β1∫Q(ˉρ−ρQ)ξdxdt+β2∫Q(ˉμ−μQ)ηdxdt, |
whence the assertion follows.
In this section, we derive the first-order necessary conditions of optimality for problem (CP). We begin with an existence result.
THEOREM 4.1: Suppose that the conditions (A1)-(A5) are satisfied. Then the problem (CP) has a solution ˉu∈Uad.
PROOF: Let {un}n∈N⊂Uad be a minimizing sequence for (CP), and let {(ρn,μn)}n∈N be the sequence of the associated solutions to (1.2)-(1.5). We then can infer from the global estimate (2.17) the existence of a triple (ˉu,ˉρ,ˉμ) such that, for a suitable subsequence again indexed by n,
un→ˉuweakly star in H1(0,T;H)∩L∞(Q),ρn→ˉρweakly star in H2(0,T;H)∩nW1,∞(0,T;L∞(Ω))∩H1(0,T;V),μn→ˉμweakly star in W1,∞(0,T;H)∩H1(0,T;V)∩L∞(0,T;W). |
Clearly, we have that ˉu∈Uad and, by virtue of the Aubin-Lions lemma (cf. [24, Thm. 5.1, p. 58]) and similar compactness results (cf. [27, Sect. 8, Cor. 4]),
ρn→ ˉρstrongly in L2(Q), | (4.1) |
whence also ρ∗≤ ˉρ≤ρ∗ a. e. in Q and
B[ρn]→B[ˉρ]strongly in L2(Ω),Φ(ρn)→Φ(ˉρ)strongly in L2(Q),for Φ∈{F′,g,g′}, |
thanks to the general assumptions on B, F and g, as well as the strong convergence
μn→ ˉμstrongly in C0([0,T];C0(ˉΩ))=C0(ˉQ). | (4.2) |
From this, we easily deduce that
g(ρn)∂tμn→g(ˉρ)∂tˉμweakly in L1(Q),μng′(ρn)∂tρn→ˉμg′(ˉρ)∂tˉρweakly in L1(Q). |
In summary, if we pass to the limit as n→∞ in the state equations (1.2)-(1.3), written for the triple (un,ρn,μn), we find that (ˉρ,ˉμ) satisfies (1.2) and (1.3). Moreover, ˉμ∈L∞(0,T;W) satisfies the boundary condition (1.4), and it is easily seen that also the initial conditions (1.5) hold true. In other words, we have (ˉρ,ˉμ)=S(ˉμ), that is, the triple (ˉu,ˉρ,ˉμ) is admissible for the control problem (CP). From the weak sequential lower semicontinuity of the cost functional J it finally follows that ˉu, together with (ˉρ,ˉμ)=S(ˉμ), is a solution to (CP). This concludes the proof.
We now turn our interest to the derivation of first-order necessary optimality conditions for problem (CP). To this end, we generally assume in the following that the hypotheses (A1)-(A5) are satisfied and that ˉu∈Uad is an optimal control with associated state (ˉρ,ˉμ), which has the properties (2.16)-(2.31). We now aim to eliminate ξ and η from the variational inequality (3.60). To this end, we employ the adjoint state system associated with (1.2)-(1.5) for ˉu, which is formally given by:
−(1+2g(ˉρ))pt−g′(ˉρ)ˉρtp−Δp−g′(ˉρ)q= β2(ˉμ−μQ)in Q, | (4.3) |
−qt+F″(ˉρ)q−ˉμg″(ˉρ)q+g′(ˉρ)(ˉμtp−ˉμpt)+DB[ˉρ]∗(q)= β1(ˉρ−ρQ)in Q, | (4.4) |
∂np=0on Σ, | (4.5) |
p(T)=q(T)=0in Ω. | (4.6) |
In (4.4), DB[ˉρ]∗∈L(L2(Q),L2(Q)) denotes the adjoint operator associated with the operator DB[ˉρ]∈L(L2(Q),L2(Q)), thus defined by the identity
∫QDB[ˉρ]∗(v)wdxdt=∫QvDB[ˉρ](w)dxdt∀v,w∈L2(Q). | (4.7) |
As, for every v∈L2(Q), the restriction of DB[ˉρ](v) to Qt depends only on vjQt, it follows that, for every w∈L2(Q), the restriction of DB[ˉρ]∗(w) to Qt=Ω×(t, T) (see (1.14)) depends only on w|Qt. Moreover, (2.8) implies that
‖DB[ˉρ]∗(w)‖L2(Qt)≤CB‖w‖L2(Qt)∀w∈L2(Q). | (4.8) |
We also note that in the case of the integral operator (1.9) it follows from Fubini’s theorem that DB[ˉρ]∗=DB[ˉρ]=B.
We have the following existence and uniqueness result for the adjoint system.
THEOREM 4.2: Suppose that (A1)-(A5) are fulfilled, and assume that ˉu∈Uad is a solution to the control problem (CP) with associated state (ˉρ,ˉμ)=S(ˉu). Then the adjoint system (4.3)-(4.6) has a unique solution (p, q) satisfying
p∈H1(0,T;H)∩L∞(0,T;V)∩L2(0,T;W)andq∈H1(0,T;H). | (4.9) |
PROOF: Besides of Young’s inequality, we make repeated use of the global estimates (2.16)-(2.17) and (2.31) for ˉρ and ˉμ, without further reference. Moreover, we denote by c different positive constants that may depend on the given data of the state system and of the control problem; the meaning of c may change between and even within lines.
We first prove uniqueness. Thus, we replace the right-hand sides of (4.3) and (4.4) by 0 and prove that (p, q)=(0, 0). We add p to both sides of (4.3) and multiply by-pt. At the same time, we multiply (4.4) by q. Then we add the resulting equalities and integrate over Qt=Ω×(t, T). As g is nonnegative, and thanks to (2.8), we obtain that
∫Qt|pt|2dxds+12‖p(t)‖2V+12∫Ω|q(t)|2dx≤∫Qt(−p−g′(ˉρ)ˉρtp−g′(ˉρ)q)ptdxds+∫Qt((ˉμg″(ˉρ)−F″(ˉρ))q+ˉμg′(ˉρ)pt−DB[ˉρ](q))qdxds−∫Qtg′(ˉρ)ˉμtpqdxds≤12∫Qt|pt|2dxds+c∫Qt(p2+q2)dxds+c∫Qt|ˉμt||p||q|dxds. |
The last integral is estimated as follows: employing the Hölder, Sobolev and Young inequalities, we have
∫Qt|ˉμt||p||q|dxds≤∫Tt‖ˉμt(s)‖3‖p(s)‖6‖q(s)‖2ds≤c∫Tt(‖ˉμt(s)‖2V‖p(s)‖2V+‖q(s)‖2H)ds. |
As the function s↦‖ˉμt(s)‖2V belongs to L1(0, T), we can apply the backward version of Gronwall’s lemma to conclude that (p, q)=(0, 0).
The existence of a solution to (4.3)-(4.6) is proved in several steps.
STEP 1: We approximate ˉρ and ˉμ by functions ρε,με∈C∞(ˉQ) satisfying (3.8)-(3.9) and look for a solution (pε,qε) to the following problem:
−(1+2g(ρε))pεt−g′(ˉρ)ˉρtpε−Δpε−g′(ˉρ)qε=β2(ˉμ−μQ)in Q, | (4.10) |
−qεt−εΔqε+F″(ˉρ)qε−ˉμg″(ˉρ)qε+g′(ρε)(μεtpε−μεpεt)+DB[ˉρ]∗(qε)=β1(qε−ρQ)in Q, | (4.11) |
∂npε=∂nqε=0on Σ, | (4.12) |
pε(T)=qε(T)=0in Ω. | (4.13) |
We prove that this problem has a unique solution satisfying
pε,qε∈H1(0,T;H)∩L∞(0,T;V)∩L2(0,T;W). | (4.14) |
To this end, we present (4.10)-(4.12) as an abstract backward equation, namely,
−ddt(pε,qε)(t)+Aε(t)(pε,qε)(t)+(Rε(pε,qε))(t)=fε(t), | (4.15) |
in the framework of the Hilbert triplet (V, H, V′), where
V:=V×V and H:=H×H. |
Notice that (4.15), together with the regularity (4.14), means that
−((pεt,qεt)(t),(v,w))H+aε(t;(pε,qε)(t),(v,w))+((Rε(pε,qε))(t),(v,w))H=(fε(t),(v,w))Hfor every (v,w)∈V and a. a. t∈(0,T), | (4.16) |
where aε(t;⋅,⋅) is the bilinear form associated with the operator Aε(t):V→V′; moreover, (⋅,⋅)H denotes the inner product in H (equivalent to the usual one) that one has chosen, the embedding H⊂V′ being dependent on such a choice. In fact, we will not use the standard inner product of H, which will lead to a nonstandard embedding H⊂V′. We aim at applying first [1, Thm. 3.2, p. 256], in order to find a unique weak solution, as we did for the linearized problem; then, we derive the full regularity required in (4.14). We set, for convenience,
φε:=11+2g(ρε) and ψε:=μεg′(ρε)1+2g(ρε)=φε:μεg′(ρε), |
and choose a constant Mε such that
φε≤Mε,|ψε|≤Mε,|∇φε|≤Mε, and |∇φε|≤Mε, a. e. in Q. |
Moreover, we introduce three parameters λε,λε1,λε2, whose values will be specified later on. In order to transform our problem, we compute pεt from (4.10) and substitute in (4.11). Moreover, we multiply (4.10) by φε. Finally, we add and subtract the same terms for convenience. Then (4.10)-(4.11) is equivalent to the system
−pεt−φεΔpε+λε1pε−λε1pε−φεg′(ˉρ)ˉρtpε−φεg′(ˉρ)qε=φεβ2(ˉμ−μQ),−qεt−εΔqε+φεΔpε+λε2qε−λε2qε+F″(ˉρ)qε−ˉμg″(ˉρ)qε+g′(ρε)μεtpε+ψε(g′(ˉρ)ˉρtpε+g′(ˉρ)qε+β2(ˉμ−μQ))+DB[ˉρ]∗(qε)=β1(ˉρ−ρQ). |
By observing that
−φεΔpε=−div(φε∇pε)+∇φε⋅∇pε, |
and that the same holds true with ψε in place of φε, we see that the latter system, complemented with the boundary condition (4.12), is equivalent to
−∫Ωqεt(t)vdx+aεt(t;pε(t),v)+∫Ω(Rε1(pε,qε))(t)vdx=∫Ωφε(t)β2(ˉμ−μQ)(t)vdx−∫Ωqεt(t)wdx+aεt(t;(pε(t),qε(t)),w)+∫Ω(Rε1(pε,qε))(t)wdx=−∫Ωψε(t)β2(ˉμ−μQ)(t)wdx+∫Ωβ1(ˉρ−ρQ)(t)wdx |
for every (v, w)∈V and a. a. t∈(0,T), where the forms aεi are defined below and the operators Rεi account for all the other terms on the left-hand sides of the equations. We set, for every t∈[0,T] and ˆv,ˆw,v,w∈V,
aε1(t;ˆv,v):=∫Ω(φε(t)∇ˆv⋅∇v+(∇φε(t)⋅∇ˆv)v+λε1ˆvv)dx,aε2(t;(ˆv,ˆw),w):=∫Ω(ε∇ˆw⋅∇w−ψε(t)∇ˆv⋅∇w−(∇ψε(t)⋅∇ˆv)w+λε2ˆww)dx. |
Now, we choose the values of λεi and of the further parameter λε in such a way as to guarantee some coerciveness. Putting α :=1/(1 + 2 sup g), we have that
aε1(t;v,v)≥∫Ω(α|∇v|2−Mε|∇v||v|+λε1v2)dx≥∫Ω(α|∇v|2−α2|∇v|2−Mε2αv2+λε1v2)dx. |
Therefore, the choice λε1:=α2+M2ε2α yields
aε1(t;v,v)≥α2‖v‖v2 for every v∈V and t∈[0,T]. |
Next, we deal with aε2. We have, for every v,w∈V and t∈[0,T],
aε2(t;(v,w),w)≥∫Ω(ε|∇w|2−Mε|∇v||∇w|−Mε|∇v||w|+λε2w2)dx≥∫Ω(ε|∇w|2−ε2|∇w|2−M2ε2ε|∇v|2−M2ε2ε|∇v|2−ε2|w|2+λε2w2)dx=∫Ω(ε2|∇w|2+(λε2−ε2)w2−M2εε|∇v|2)dx, |
and the choice λε2:=ε leads to
aε2(t;(v,w),w)≥ε2‖w‖2V−M2εε‖v‖2V. |
Therefore, if we choose λε such that λεα2−M2εε≥ε2, then we obtain
λεaε1(t;v,v)+aε2(t;(v,w),w)≥ε2(‖v‖2V+‖w‖2V) |
for every (v, w)∈V and t∈[0,T]. Hence, if we define aε:[0,T]×V×V→R by setting
aε(t;(ˆv,ˆw),(v,w)):=λε1aε1(ˆv,v)+aε2(t;(ˆv,ˆw),w), |
then we obtain a time-dependent continuous bilinear form that is coercive on V (endowed whith its standard norm), uniformly with respect to t∈[0,T]. Moreover, aε depends smoothly on t, and (4.10)-(4.12) is equivalent to
−∫Ω(λεpεt(t)v+qεt(t)w)dx+aε(t;(pε(t),qε(t)),(v,w))+∫Ω{λε(Rε1(pε,qε))(t)v+(Rε2(pε,qε))(t)w}dx=∫Ω((λεφε−ψε)(t)β2(ˉμ−μQ)(t)v+β1(ˉρ−ρQ)(t)w |
for every (v, w)∈V and a. a. t∈(0,T). Therefore, the desired form (4.16) is achieved if we choose the scalar product in H as follows:
((ˆv,ˆw),(v,w))H:=∫Ω(λεˆvv+ˆww)dx for every (ˆv,ˆw),(v,w)∈H. |
Notice that this leads to the following nonstandard embedding H⊂V′ :
V′⟨(ˆv,ˆw),(v,w)⟩V=((ˆv,ˆw),(v,w))H=λεV′⟨ˆv,v⟩V+V′⟨ˆw,w⟩V |
for every (ˆv,ˆw)∈H and (v,w)∈V, provided that the embedding H⊂V′ is the usual one, i. e., corresponds to the standard inner product of H. As the remainder, given by the terms Rε1 and Rε2, satisfies the backward analogue of (3.14)-(3.15) (see also (4.8)), the quoted result of [1] can be applied, and problem (4.10)-(4.13) has a unique solution satisfying
(pε,qε)∈H1(0,T;V′)∩L2(0,T;V). |
Moreover, if we move the remainder of (4.15) to the right-hand side, we see that
−ddt(pε,qε)+Aε(pε,qε)∈L2(0,T;H). |
Therefore, by also accounting for (4.13), we deduce that (pε,qε)∈H1(0,T;H) as well as Aε(pε,qε)∈L2(0,T;H). Hence, we have that pε,qε∈L2(0,T;W), by standard elliptic regularity.
STEP 2: We add pε to both sides of (4.10) and multiply by −pεt. At the same time, we multiply (4.11) by qε. Then, we sum up and integrate over Qt. As g≥0, we easily obtain that
12‖pε(t)‖2V+∫Qt|pεt|2dxds+12∫Ω|qε(t)|2dx+ε∫Qt|∇qε|2dxds≤c∫Qt|pε||pεt|dxds+c∫Qt|qε||pεt|dxds+c∫Qt|qε|2dxds+c∫Qt|μεt||pε||qε|dxds+∫Qt|DB[ˉρ]∗(qε)||qε|dxds+c‖pε‖2L2(Qt)+c. |
Just two of the terms on the right-hand side need some treatment. We have
∫Qt|μεt||pε||qε|dxds≤∫Tt‖μεt(s)‖3‖pε(s)‖6‖qε(s)‖2ds≤c∫Tt‖pε(s)‖2Vds+c∫Tt‖μεt(s)‖23‖qε(s)‖22ds, |
and we observe that the function s↦‖μεt(s)‖23 belongs to L1(0, T), by (3.8). Moreover, the Schwarz inequality and (4.8) immediately yield that
∫Qt|DB[ˉρ]∗(qε)||qε|dxds≤CB‖qε‖2L2(Qt). |
Therefore, we can apply the backward version of Gronwall’s lemma to obtain that
‖pε‖H1(0,T;H)∩L∞(0,T;V)+‖qε‖H1(0,T;H)+ε1/2‖qε‖L2(0,T;V)≤c. | (4.17) |
By comparison in (4.10), we see that Δpε is bounded in L2(Q). Hence,
‖pε‖L2(0,T;W)≤c. | (4.18) |
STEP 3: We multiply (4.11) by −qεt and integrate over Qt. We obtain
∫Qt|qεt|2dxds+ε2∫Ω|∇qε(t)|2dx≤c∫Ω|qε||qεt|dxds+c∫Tt‖μεt(s)‖3‖pε(s)‖6‖qεt(s)‖2ds+c∫Qt|pεt||qεt|dxds+∫Qt|DB[ˉρ]∗(qε)||qεt|dxds. |
Thanks to (4.8) once more, we deduce that
12∫Qt|qεt|2dxds+ε2∫Ω|∇qε(t)|2dx≤c∫Qt|qε|2dxds+c∫Tt‖μεt(s)‖23‖pε(s)‖2Vds+c∫Qt|pεt|2dxds. |
Thus, (3.8) and (4.17) imply that
‖qεt‖L2(0,T;H)+ε1/2‖qε‖L∞(0,T;V)≤c. | (4.19) |
STEP 4: Now, we let ε tend to zero and construct a solution to (4.3)-(4.6). By (4.17)-(4.19) we have, at least for a subsequence,
pε→pweakly star in H1(0,T;H)∩L∞(0,T;V)∩L2(0,T;W),qε→qweakly in H1(0,T;H),εqε→0strongly in L∞(0,T;V), |
for some pair (p, q) satisfying the regularity requirements (4.9). By accounting for (3.9) and the Lipschitz continuity of g and g′, it is straightforward to see that (p, q) is a solution to problem (4.3)-(4.6). This completes the proof.
COROLLARY 4.3: Suppose that (A1)-(A5) are fulfilled, and assume that ˉu∈Uad is an optimal control of (CP) with associated state (ˉρ,ˉμ)=S(ˉu) and adjoint state (p, q). Then it holds the variational inequality
∫Q(p+β3ˉu)(v−ˉu)dxdt≥0∀v∈Uad. | (4.20) |
PROOF: We fix v∈Uad and choose h=v−ˉu. Then, we write the linearized system (3.1)-(3.4) and multiply the equations (3.1) and (3.2) by p and q, respectively. At the same time, we consider the adjoint system and multiply the equations (4.3) and (4.4) by −η and −ξ, respectively. Then, we add all the equalities obtained in this way and integrate over Q. Many terms cancel out. In particular, this happens for the contributions given by the Laplace operators, due to the boundary conditions (3.3) and (4.5), as well as for the terms involving DB[ˉρ] and DB[ˉρ]∗, by the definition of adjoint operator (see (4.7)). Thus, it remains
∫Q(2g′(ˉρ)ˉρtηp+(1+2g(ˉρ))ηtp+(1+2g(ˉρ))ηpt)dxdt+∫Q(ˉμtg′(ˉρ)ξp+ˉμg″(ˉρ)ˉρtξp+ˉμg′(ˉρ)ξtp+ˉμg′(ˉρ)ξpt)dxdt+∫Q(ξtq+ξqt)dxdt=∫Q((v−ˉu)p−β2(ˉμ−μQ)η−β1(ˉρ−ρQ)ξ)dxdt |
Now, we observe that the expression on the left-hand side coincides with
∫Q∂t{(1+2g(ˉρ))ηp+ˉμg′(ˉρ)ξp+ξq}dxdt. |
Thus, it vanishes, due to the initial and final conditions (3.4) and (4.6). This implies that
∫Q(β1(ˉρ−ρQ)η+β2(ˉμ−μQ)ξ)dxdt=∫Q(v−ˉu)pdxdt. |
Therefore, (4.20) follows from (3.60).
REMARK 4: The variational inequality (4.20) forms together with the state system (1.2)-(1.5) and the adjoint system (4.3)-(4.6) the system of first-order necessary optimality conditions for the control problem (CP). Notice that in the case β3>0 the function −β−13p is nothing but the L2(Q) orthogonal projection of ˉu onto Uad.
This work received a partial support from the GNAMPA (Gruppo Nazionale per l’Analisi Matematica, la Probabilità e loro Applicazioni) of INDAM (Istituto Nazionale di Alta Matematica) and the IMATI-C.N.R. Pavia for PC and GG.
[1] |
C. Baiocchi, Sulle equazioni di erenziali astratte lineari del primo e del secondo ordine negli spazi
di Hilbert, Ann. Mat Pura Appl. (4) 76 (1967), 233-304.
2. P. Colli, M. H. Farshbaf-Shaker, G. Gilardi, J. Sprekels, Optimal boundary control of a viscous Cahn–Hilliard system with dynamic boundary condition and double obstacle potentials, SIAM J. Control Optim. 53 (2015), 2696-2721. 3. P. Colli, M. H. Farshbaf-Shaker, G. Gilardi, J. Sprekels, Second-order analysis of a boundary control problem for the viscous Cahn–Hilliard equation with dynamic boundary conditions, Ann. Acad. Rom. Sci. Math. Appl. 7 (2015), 41-66. 4. P. Colli, G. Gilardi, P. Krejčí, P. Podio-Guidugli, J. Sprekels, Analysis of a time discretization scheme for a nonstandard viscous Cahn–Hilliard system, ESAIM Math. Model. Numer. Anal. 48 (2014), 1061-1087. 5. P. Colli, G. Gilardi, P. Krejčí, J. Sprekels, A continuous dependence result for a nonstandard system of phase field equations, Math. Methods Appl. Sci. 37 (2014), 1318-1324. 6. P. Colli, G. Gilardi, P. Krejčí, J. Sprekels, A vanishing di usion limit in a nonstandard system of phase field equations, Evol. Equ. Control Theory 3 (2014), 257-275. 7. P. Colli, G. Gilardi, P. Podio-Guidugli, J. Sprekels, Well-posedness and long-time behavior for a nonstandard viscous Cahn–Hilliard system, SIAM J. Appl. Math. 71 (2011), 1849-1870. 8. P. Colli, G. Gilardi, P. Podio-Guidugli, J. Sprekels, Distributed optimal control of a nonstandard system of phase field equations, Contin. Mech. Thermodyn. 24 (2012), 437-459. 9. P. Colli, G. Gilardi, P. Podio-Guidugli, J. Sprekels, Continuous dependence for a nonstandard Cahn–Hilliard system with nonlinear atom mobility, Rend. Sem. Mat. Univ. Politec. Torino 70 (2012), 27-52. 10. P. Colli, G. Gilardi, P. Podio-Guidugli, J. Sprekels, Global existence for a strongly coupled Cahn– Hilliard system with viscosity, Boll. Unione Mat. Ital. (9) 5 (2012), 495-513. 11. P. Colli, G. Gilardi, P. Podio-Guidugli, J. Sprekels, An asymptotic analysis for a nonstandard Cahn–Hilliard system with viscosity, Discrete Contin. Dyn. Syst. Ser. S 6 (2013), 353-368. 12. P. Colli, G. Gilardi, P. Podio-Guidugli, J. Sprekels, Global existence and uniqueness for a singular /degenerate Cahn–Hilliard system with viscosity, J. Di erential Equations 254 (2013), 4217- 4244. 13. P. Colli, G. Gilardi, E. Rocca, J. Sprekels, Optimal distributed control of a di use interface model of tumor growth, preprint arXiv:1601.04567 [math.AP] (2016), 1-32. 14. P. Colli, G. Gilardi, J. Sprekels, Analysis and optimal boundary control of a nonstandard system of phase field equations, Milan J. Math. 80 (2012), 119-149. 15. P. Colli, G. Gilardi, J. Sprekels, Regularity of the solution to a nonstandard system of phase field equations, Rend. Cl. Sci. Mat. Nat. 147 (2013), 3-19. 16. P. Colli, G. Gilardi, J. Sprekels, A boundary control problem for the pure Cahn–Hilliard equation with dynamic boundary conditions, Adv. Nonlinear Anal. 4 (2015), 311-325. 17. P. Colli, G. Gilardi, J. Sprekels, A boundary control problem for the viscous Cahn–Hilliard equation with dynamic boundary conditions, Appl. Math. Optim. 73 (2016), 195-225. 18. P. Colli, G. Gilardi, J. Sprekels, On an application of Tikhonov’s fixed point theorem to a nonlocal Cahn–Hilliard type system modeling phase separation, J. Di erential Equations 260 (2016), 7940- 7964. 19. S. Frigeri, E. Rocca, J. Sprekels, Optimal distributed control of a nonlocal Cahn–Hilliard/Navier– Stokes system in 2D, SIAM J. Control Optim. 54 (2016), 221-250. 20. M. Hinterm¨uller, T. Keil, D. Wegner, Optimal control of a semidiscrete Cahn–Hilliard–Navier– Stokes system with non-matched fluid densities, preprint arXiv:1506.03591 [math.AP] (2015), 1- 35. 21. M. Hinterm¨uller, D.Wegner, Distributed optimal control of the Cahn–Hilliard system including the case of a double-obstacle homogeneous free energy density, SIAM J. Control Optim. 50 (2012), 388-418. 22. M. Hinterm¨uller, D. Wegner, Optimal control of a semidiscrete Cahn–Hilliard–Navier–Stokes system, SIAM J. Control Optim. 52 (2014), 747-772. 23. M. Hinterm¨uller, D. Wegner, Distributed and boundary control problems for the semidiscrete Cahn–Hilliard/Navier–Stokes system with nonsmooth Ginzburg–Landau energies, Isaac Newton Institute Preprint Series No. NI14042-FRB (2014), 1-29. 24. J. L. Lions, “Quelques méthodes de résolution des problèmes aux limites non linéaires”, Dunod Gauthier-Villars, Paris, 1969. 25. P. Podio-Guidugli, Models of phase segregation and diffusion of atomic species on a lattice, Ric. Mat. 55 (2006), pp. 105-118. 26. E. Rocca, J. Sprekels, Optimal distributed control of a nonlocal convective Cahn–Hilliard equation by the velocity in three dimensions, SIAM J. Control Optim. 53 (2015), 1654-1680. 27. J. Simon, Compact sets in the space Lp(0; T; B) , Ann. Mat. Pura Appl. (4) 146 (1987), pp. 65-96. 28. Q.-F. Wang, S.-i. Nakagiri, Weak solutions of Cahn–Hilliard equations having forcing terms and optimal control problems, Mathematical models in functional equations (Japanese) (Kyoto, 1999), Sῡrikaisekikenkyῡsho Kōkyῡroku No. 1128 (2000), 172-180. 29. X. Zhao, C. Liu, Optimal control of the convective Cahn–Hilliard equation, Appl. Anal. 92 (2013), 1028-1045. 30. X. Zhao, C. Liu, Optimal control for the convective Cahn–Hilliard equation in 2D case, Appl. Math. Optim. 70 (2014), 61-82. |
1. | Pierluigi Colli, Gianni Gilardi, Jürgen Sprekels, Distributed optimal control of a nonstandard nonlocal phase field system with double obstacle potential, 2017, 6, 2163-2480, 35, 10.3934/eect.2017003 | |
2. | Pierluigi Colli, Jürgen Sprekels, 2017, Chapter 7, 978-3-319-64488-2, 151, 10.1007/978-3-319-64489-9_7 | |
3. | Tania Biswas, Sheetal Dharmatti, Manil T. Mohan, Pontryagin maximum principle and second order optimality conditions for optimal control problems governed by 2D nonlocal Cahn–Hilliard–Navier–Stokes equations, 2020, 40, 2196-6753, 127, 10.1515/anly-2019-0049 | |
4. | Pierluigi Colli, Gianni Gilardi, Jürgen Sprekels, Optimal boundary control of a nonstandard viscous Cahn–Hilliard system with dynamic boundary condition, 2018, 170, 0362546X, 171, 10.1016/j.na.2018.01.003 | |
5. | Gianni Gilardi, Jürgen Sprekels, Asymptotic limits and optimal control for the Cahn–Hilliard system with convection and dynamic boundary conditions, 2019, 178, 0362546X, 1, 10.1016/j.na.2018.07.007 | |
6. | Pierluigi Colli, Gianni Gilardi, Jürgen Sprekels, 2018, Chapter 11, 978-3-319-75939-5, 217, 10.1007/978-3-319-75940-1_11 | |
7. | Pierluigi Colli, Gianni Gilardi, Jürgen Sprekels, Optimal Distributed Control of a Generalized Fractional Cahn–Hilliard System, 2020, 82, 0095-4616, 551, 10.1007/s00245-018-9540-7 | |
8. | Pierluigi Colli, Gianni Gilardi, Jürgen Sprekels, Optimal Velocity Control of a Viscous Cahn--Hilliard System with Convection and Dynamic Boundary Conditions, 2018, 56, 0363-0129, 1665, 10.1137/17M1146786 | |
9. | Zuo Wang, Hong Xue, Yingnan Pan, Hongjing Liang, Adaptive neural networks event-triggered fault-tolerant consensus control for a class of nonlinear multi-agent systems, 2020, 5, 2473-6988, 2780, 10.3934/math.2020179 | |
10. | Qian Li, Zhenghong Jin, Linyan Qiao, Aichun Du, Gang Liu, Distributed optimization of nonlinear singularly perturbed multi-agent systems via a small-gain approach and sliding mode control, 2024, 9, 2473-6988, 20865, 10.3934/math.20241015 |