Processing math: 100%
Research article Special Issues

Distributed optimal control of a nonstandard nonlocal phase field system

  • We investigate a distributed optimal control problem for a nonlocal phase field model of viscous Cahn-Hilliard type. The model constitutes a nonlocal version of a model for two-species phase segregation on an atomic lattice under the presence of diffusion that has been studied in a series of papers by P. Podio-Guidugli and the present authors. The model consists of a highly nonlinear parabolic equation coupled to an ordinary differential equation. The latter equation contains both nonlocal and singular terms that render the analysis difficult. Standard arguments of optimal control theory do not apply directly, although the control constraints and the cost functional are of standard type. We show that the problem admits a solution, and we derive the first-order necessary conditions of optimality.

    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

    Related Papers:

    [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
  • We investigate a distributed optimal control problem for a nonlocal phase field model of viscous Cahn-Hilliard type. The model constitutes a nonlocal version of a model for two-species phase segregation on an atomic lattice under the presence of diffusion that has been studied in a series of papers by P. Podio-Guidugli and the present authors. The model consists of a highly nonlinear parabolic equation coupled to an ordinary differential equation. The latter equation contains both nonlocal and singular terms that render the analysis difficult. Standard arguments of optimal control theory do not apply directly, although the control constraints and the cost functional are of standard type. We show that the problem admits a solution, and we derive the first-order necessary conditions of optimality.


    1. Introduction

    Let ΩR3 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)=β12Q|ρρQ|2dxdt+β22Q|μμQ|2dxdt+β32Q|u|2dxdt (1.1)

    subject to the state system

    (1+2g(ρ))tμ+μg(ρ)tρΔμ=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

    uUad:={uH1(0,T;L2(Ω)):0uumaxa.e.inQ and uH1(0,T;L2(Ω))R}. (1.6)

    Here, \gianni{β1,β2,β30 and R>0 are given constants, with β1+β2+β3>0, } and the threshold function umaxL(Q) is nonnegative}. Moreover, ρQ,μQL2(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)+(1ρ)log(1ρ)),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(|yx|)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 βi0 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:={wH2(Ω):nw=0a.e.onΓ}.

    We have the dense and continuous embeddings WVHHVW, where u,vV=(u,v)H and u,wW=(u,w)H for all uH, vV, and wW.

    In the following, we will make repeated use of Young’s inequality

    abδa2+14δb2for all a,bRand δ>0, (1.10)

    as well as of the fact that for three dimensions of space and smooth domains the embeddings VLp(Ω), 1p6, and H2(Ω)C0(¯Ω) are continuous and (in the first case only for 1p<6) compact. In particular, there are positive constants ˜Ki, i=1,2,3, which depend only on the domain Ω, such that

    v6˜K1vVvV, (1.11)
    vwHv6w3˜K2vVwVv,wV, (1.12)
    v˜K3vH2(Ω)vH2(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.


    2. Problem statement and results for the state system

    Consider the optimal control problem (1.1)-(1.6). We make the following assumptions on the data:

    (A1) F=F1 + F2, where F1C3(0,1) is convex, F2C3[0,1], and

    limr0F1(r)=,limr1F1(r)=+. (2.1)

    (A2) ρ0V, F(ρ0)H, μ0W, where μ00\, a.\, e. in Ω,

    inf{ρ0(x):xΩ}>0,sup{ρ0(x):xΩ}<1. (2.2)

    (A3) gC3[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+vLp(Qt)) (2.4)

    for every vLp(Q) and t(0,T].

    (iii) For every v,wL1(0,T;H) and t(0,T], it holds that

    t0B[v](s)B[w](s)6dsCBt0v(s)w(s)Hds. (2.5)

    (iv) It holds, for every vL2(0,T;V) and t(0,T], that

    B[v]L2(0,t;H)CB(1+vL2(0,t;V)). (2.6)

    (v) For every vH1(0,T;H), we have tB[v]L2(Q) and

    tB[v]L2(Q)CB(1+tvL2(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 ¯vL2(Q) and t(0,T] the following properties:

    DB[ˉv](w)Lp(Qt)CBwLp(Qt)wLp(Q),p[2,6], (2.8)
    (DB[ˉv](w))L2(Qt)CBwL2(0,t;V)wL2(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,β30, β1+β2+β3>0, and R>0. (2.10)
    ρQ,μQL2(Q),umaxL(Q) and umax0 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)CBvwL2(Qt)v,wL2(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 \, vL2(0,T;V),

    |QtB[v]vdxds|CB(1+v2L2(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)˜CBwL2(0,T;V)ˉvL2(Q),wL2(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)|C1rα,|k(r)|C2rβ,r>0.

    In this case, we have 2α<3 and thus, for all v,wL1(0,T;H)\, and t(0,T],

    t0B[v](s)B[w](s)6dsC1t0(Ω|Ω|yx|α|v(y,s)w(y,s)|dy|6dx)1/6dsC3t0(Ω|(Ω|yx|2αdy)1/2v(s)w(s)H|6dx)1/6dsC4t0v(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 vL2(0,T;V), we find, since 6β/5<3, that

    B[v]2L2(0,t;H)C2Qt|Ω|yx|β|v(y,s)|dy|2dxdsC5Qt(Ω|yx|6β/5dy)5/3v(s)26dxdsC6t0v(s)2Vds.

    Finally, since the operator B is linear in this case, we have DB[¯v]=B for every ˉvL2(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 μ0C(¯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 uUad a unique solution (ρ,μ) such that

    ρH2(0,T;H)W1,(0,T;L(Ω))H1(0,T;V), (2.14)
    μW1,(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 K1>0, which depend only on the given data, such that for every uUad the corresponding solution (ρ,μ) satisfies

    0<ρρρ<1,0μμ,a.e.inQ, (2.16)
    μW1,(0,T;H)H1(0,T;V)L(0,T;W)L(Q)+ρH2(0,T;H)W1,(0,T;L(Ω))H1(0,T;V)K1. (2.17)

    PROOF: In the following, we denote by Ci>0,iN, constants which may depend on the data of the control problem (CP) but not on the special choice of uUad. First, we note that in [18, Thms. 2.1, 2.2] it has been shown that under the given assumptions there exists for u0 a unique solution (ρ,μ) with the properties

    0<ρ<1,μ0,a.e.inQ, (2.18)
    ρL(0,T;V), tρL6(Q), (2.19)
    μH1(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 uUad. We thus infer that (1.2)-(1.5) enjoys for every uUad a unique solution satisfying (2.18)-(2.20); more precisely, there is some C1 > 0 such that

    μH1(0,T;H)L(0,T;V)L(Q)L2(0,T;W2,3/2(Ω))+ρL(0,T;V)+tρL6(Q)C1uUad. (2.21)

    Moreover, invoking (2.18), and (2.4) for p=+∞, we find that

    B[ρ]L(Q)C2uUad,

    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 uUad,

    0<ρinf{ρ0(x):xΩ}sup{ρ0(x):xΩ}ρ<1,F(ρ)+B[ρ]μg(ρ)0if0<ρρ,F(ρ)+B[ρ]μg(ρ)0ifρρ<1.

    Therefore, multiplying (1.3) by the positive part (ρρ)+ of ρρ, and integrating over Q, we find that

    0=Qtρ(ρρ)+dxdt+Q(F(ρ)+B[ρ]μg(ρ))(ρρ)+dxdt12Ω|(ρ(t)ρ)+|2dx,

    whence we conclude that (ρρ)+=0, and thus ρρ, 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

    μL2(0,T;W)C3uUad. (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

    tρL(Q)C5uUad. (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)ρB[ρ]+g(ρ)μ+μg(ρ)ρ (2.24)

    is bounded in L2(Q), so that

    tρL2(0,T;V)C6uUad. (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

    ρttL2(Q)C7uUad. (2.26)

    At this point, we observe that Eq. (1.2) can be written in the form

    atμ+μtaΔμ=b,with a:=1+2g(ρ),b:=u+μg(ρ)tρ,

    where, thanks to the above estimates, we have, for every uUad,

    aL(Q)+taL(Q)+bL(Q)C8, (2.27)
    t2aL2(Q)=2g(ρ)ρ2t+g(ρ)ρttL2(Q)C9, (2.28)
    tbL2(Q)=ut+μtg(ρ)ρt+μg(ρ)ρ2t+μg(ρ)ρttL2(Q)C10. (2.29)

    Since also μ0W, we are thus in the situation of [15, Thm. 3.4], whence we obtain that tμL(0,T;H)L2(0,T;V) and μL(0,T;W). Moreover, a closer look at the proof of [15, Thm. 3.4] reveals that we also have the estimates

    tμL(0,T;H)L2(0,T;V)+μL(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 VL6(Ω), we can without loss of generality (by possibly choosing a larger K1) assume that also

    max0i3F(i)(ρ)L(Q)+max0i3g(i)(ρ)L(Q)+μL(0,T;L6(Ω)3)+tμL2(0,T;V)+B[ρ]H1(0,T;L2(Ω))L(Q)L2(0,T;V)K1uUad. (2.31)

    According to Theorem 2.1, the control-to-state mapping S:Uadu(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 uiUad, 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 K2>0, which depends only on the data of the problem, such that, for every t(0,T],

    ρ1ρ2H1(0,t;H)L(0,t;L6(Ω))+μ1μ2H1(0,t;H)L(0,t;V)L2(0,t;W)K2u1u2L2(0,t;H). (2.32)

    PROOF: Taking the difference of the equations satisfied by (ρi,μi), i=1, 2, and setting u:=u1u2, ρ:=ρ1ρ2, μ:=μ1μ2, we first observe that we have almost everywhere in Q the identities

    (1+2g(ρ1))tμ+g(ρ1)tρ1μΔμ+2(g(ρ1)g(ρ2))tμ2=u(g(ρ1)g(ρ2))tρ1μ2g(ρ2)μ2tρ, (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, u2Uad; 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))μtμ+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|2dxds3j=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:=2Qt(g(ρ1)g(ρ2))tμ2μdxdsct0tμ2(s)6μ(s)3ρ(s)2dsγt0μ(s)2Vds+cγt0tμ2(s)2Vρ(s)2Hds, (2.37)

    where it follows from (2.31) that the mapping stμ2(s)2V belongs to L1(0, T). Next, we see that

    I2:=Qt(u(g(ρ1)g(ρ2))tρ1μ2)μdxdscQt(|u|+|ρ|)|μ|dxdscQt(u2+ρ2+μ2)dxds. (2.38)

    Finally, Young’s inequality yields that

    I3:=Qtg(ρ2)μ2ρtμdxdsγQtρ2tdxds+cγQtμ2dxds. (2.39)

    Combining (2.36)-(2.39), and recalling that g(ρ1) is nonnegative, we have thus shown the estimate

    12μ(t)2H+(1γ)t0μ(s)2VdsγQtρ2tdxds+cQtu2dxds+c(1+γ1)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

    (1γ)Qtρ2tdxds+12ρ(t)2HcγQt(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

    μL(0,t;H)L2(0,t;V)+ρH1(0,t;H)cuL2(0,t;H)t(0,T]. (2.42)

    STEP 2: Next, we multiply (2.34) by ρ|ρ| and integrate over Qt. We obtain

    13ρ(t)333j=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|dxdsct0ρ(s)33ds. (2.44)

    Moreover, invoking (2.42), Hälder’s inequality, as well as the global bounds once more,

    J2:=Qtμg(ρ1)ρ|ρ|dxdsct0μ(s)6ρ(s)2ρ(s)3dst0ρ(s)33ds+ct0μ(s)3/2Vρ(s)3/2Hdst0ρ(s)33ds+cρ3/2L(0,t;H)μ3/2L3/2(0,t;V)t0ρ(s)33ds+cρ3/2L(0,t;H)μ3/2L2(0,t;V)t0ρ(s)3ds+cu3L2(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|dxdsct0ρ(s)3ρ(s)2B[ρ1](s)B[ρ2](s)6dscsup0stρ(s)3ρL(0,t;H)t0ρ(s)Hds16sup0stρ(s)33+cu3L2(0,t;H). (2.46)

    Combining the estimates (2.43)-(2.46), and invoking Gronwall’s lemma, we can easily infer that

    ρL(0,t;L3(Ω))cuL2(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+12μ(t)2H5j=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)tρ1μμtdxdsγQtμ2tdxds+cγQtμ2dxdsγQtμ2tdxds+cγu2L2(0,t;H). (2.49)

    Next, thanks to the mean value theorem, and employing (2.31) and (2.47), we find that

    K2:=2Qt(g(ρ1)g(ρ2))tμ2μtdxdscQt|ρ||tμ2||μt|dxdsct0ρ(s)3tμ2(s)6μt(s)2dsγQtμ2tdxds+cγρ2L(0,t;L3(Ω))t0tμ2(s)2VdsγQtμ2tds+cγu2L2(0,t;H). (2.50)

    Moreover, we infer that

    K3:=QtuμtdxdsγQtμ2tdxds+cγu2L2(0,t;H), (2.51)

    as well as, invoking the mean value theorem once more,

    K4:=Qt(g(ρ1)g(ρ2))tρ1μ2μtdxdscQt|ρ||μt|dxdsγQtμ2tdxds+cγu2L2(0,t;H), (2.52)

    and, finally, using (2.42) and Young’s inequality,

    K5:=Qtg(ρ2)μ2ρtμtdxdscQt|ρt||μt|dxdsγQtμ2tdxds+cγρ2H1(0,t;H)γQtμ2tdxds+cγu2L2(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)cuL2(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

    ΔμL2(0,t;H)c(μtL2(0,t;H)+μL2(0,t;H)+ρtL2(0,t;H)+ρL2(0,t;H)+uL2(0,t;H))+cρtμ2L2(0,t;H)cuL2(0,t;H), (2.55)

    where the last summand on the right-hand side was estimated as follows:

    Qt|ρ|2|tμ2|2dxdsct0tμ2(s)26ρ(s)23dscρ2L(0,t;L3(Ω))t0tμ2(s)2Vdscu2L2(0,t;H).

    Invoking standard elliptic estimates, we have thus shown that

    μL2(0,t;W)cuL2(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

    16ρ(t)663j=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|4dxdsct0ρ(s)66ds. (2.58)

    Moreover, from (2.54) and Hälder’s and Young’s inequalities we conclude that

    L2:=Qtg(ρ1)μρ|ρ|4dxdsct0μ(s)6ρ(s)56dscμL(0,t;V)ρ5L5(0,t;L6(Ω))cμ6L(0,t;V)+cρ6L5(0,t;L6(Ω))cu6L2(0,t;H)+ct0ρ(s)66ds. (2.59)

    Finally, we employ (2.5) and (2.42) to infer that

    L3:=Qt(B[ρ1]B[ρ2])ρ|ρ|4dxdsct0B(ρ1](s)B[ρ2](s)6ρ(s)56dscsup0stρ(s)56t0ρ(s)Hds112sup0stρ(s)66+cu6L2(0,t;H). (2.60)

    Combining the estimates (2.57)-(2.60), and invoking Gronwall’s lemma, then we readily find the estimate

    ρL(0,t;L6(Ω))cuL2(0,t;H),

    which concludes the proof of the assertion.


    3. Directional differentiability of the control-to-state mapping

    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

    uX:=uH1(0,T;H)+uL(Q)uX,(ρ,μ)Y:=ρH1(0,T;H)+μL(0,T;H)+μL2(0,T;V)(ρ,μ)Y,

    and consider the control-to-state operator S as a mapping between UadX and Y. Now let ˉuUad 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)ξtΔη=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, hX must satisfy ˉu+ˉλhUad 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 hL2(Q), and it is uniquely solvable under this weaker assumption.

    THEOREM 3.1: Suppose that the general hypotheses (A1)-(A5) are satisfied and let hL2(Q). Then, the linearized problem (3.1)-(3.4) has a unique solution (ξ,η) satisfying

    ξH1(0,T;H)L(0,T;L6(Ω)), (3.5)
    ηH1(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

    12Ω|η(t)|2dx+t0η(s)2Vds+12 Ω|ξ(t)|2dx+Qt|ξt|2dxds3j=1Hj,

    where the terms Hj are defined and estimated as follows. We have

    H1:=Qt2g(ˉρ)ˉutξηdxdsct0ˉut(s)3ξ(s)2η(s)6ds12t0η(s)2Vds+ct0ˉut(s)2Vξ(s)22ds,

    and we notice that the function sˉut(s)2V belongs to L1(0, T), by (2.30) for ˉu. Next, we easily have the estimate

    H2:=Qt(ηˉug(ˉρ)ˉρtξˉug(ˉρ)ξt)ηdxds14Qt|ξt|2dxds+cQt(|ξ|2+|η|2)dxds.

    Finally, recalling (2.8), it is clear that

    H3:=Qt((ξ+ˉug(ˉρ)F(ˉρ))ξDB[ˉρ](ξ)+g(ˉρ)η)ξtdxds14Qt|ξt|2dxds+cQt(|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,μεC(ˉQ) as specified below. For every ε(0,1), it holds that

    ρρερ in ˉQ and ρεtL(Q)+μεH1(0,T;L3(Ω))L(Q)C, (3.8)

    for some constants ρ,ρ(0,1) and C>0; as ε0, we have

    ρεˉρ,ρεtˉρt,μεˉμ,in Lp(Q), for every p<+ and a.e.in Q,andμεtˉμtin 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)ξεtΔηε=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|CRεQt(|v|2+|w|2)ds, (3.15)

    for some constant CRε, and every v,wL2(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)+aεηε.

    As aεα:=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)dxα2w2V (3.16)

    for every wV 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+Vηεt(t),wV+VQε(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.at(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

    VQε(t)(ˆv,ˆw),(v,w)V=Ω(ˆvv+aε(t)ˆww+(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

    RεiL(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 aεijL(Q). Therefore, by virtue of (2.8), we see that

    Qt(Rε1(v,w)v+Rε1(v,w)w)dxdscQt(|v|2+|w|2)dxds+cDB[ˉρ](v)2L2(Qt)cQt(|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., ξεH1(0,T;H)andηεH1(0,T;V)L2(0,T;V).

    On the other hand, this solution has to satisfy

    tηε,w+Ωaεηεwdx=Ωϕεw dx a.e.in (0, T), forevery wV,

    for some ϕεL2(Ω). From standard elliptic regularity, it follows that ηεH1(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

    12Ω|ξε(t)|2dx+Qt|ξεt|2dxds+12Ω|ηε(t)|2dx+t0ηε(s)2Vds3j=1Ij,

    where the terms Ij are defined and estimated as follows. In view of (2.8), we first infer that

    I1:=Qt(ξεF(ˉρ)ξεDB[ˉρ](ξε)+ˉμg(ˉρ)ξε+g(ˉρ)ηε)ξεtdxds14Qt|ξεt|2dxds+cQt(|ξε|2+|ηε|2)dxds.

    Next, we have

    I2:=Qt(ηεˉμg(ˉρ)ˉρtξεˉμg(ˉρ)ξεt+h)ηεdxds14Qt|ξεt|2dxds+cQt(|ξε|2+|ηε|2)dxds+c.

    Finally, by virtue of the H¨older and Sobolev inequalities, we have

    I3:=Qt2g(ˉρ)μεtξεηεdxdsct0μεt(s)3ξε(s)2ηε(s)6ds12t0ηε(s)2Vds+ct0ηε(s)23ξε(s)22ds.

    At this point, we recall all the inequalities we have proved, notice that (3.8) implies that the function sμεt(s)23 is bounded in L1(0, T), and apply the Gronwall lemma. We obtain

    ξεH1(0,T;H)+ηεL(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

    ξεδtδΔξεδ+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

    bδL(Q)c, and bδb a.e.in Q as δ0. (3.24)

    This problem has a unique weak solution ξεδH1(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 AδL(V,V) defined by

    Aδv,w:=Ω(δvw+vw)dxfor every v,wV,

    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, fεδL2(0,T;V), so that AδfεδL2(0,T;V), and we can consider the unique solution ζεδH1(0,T;V)L2(0,T;V) to the problem

    ζεδt+Aδζεδ=Aδfεδ and ζεδ(0)=0

    Now, A1δζεδ satisfies

    (A1δζεδ)t+Aδ(A1δζεδ)=A1δAδfεδ=fεδ and (A1δζεδ)(0)=0,

    so that a comparison with (3.25) shows that ξεδ=A1δζεδ, by uniqueness. Since ζεδL(0,T;H), and A1δ(H)=W by elliptic regularity, we deduce that ξεδL(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

    16Ω|ξεδ(t)|6dx+5δQt|ξεδ|4|ξεδ|2dxds=3j=1Kj,

    where the terms Kj are defined and estimated as follows. First, recalling (3.24), we deduce that

    K1:=Qtbδξεδ(ξεδ)5dxdscQt|ξε|6dxds.

    On the other hand, Hälder’s inequality, and assumption (2.8) with p=6, imply that

    K2:=QtL(ξεδ)(ξεδ)5dxdscLξεδL6(Qt)(ξεδ)5L6/5(Qt)cξεδL6(Qt)ξεδ5L6(Qt)=cQt|ξεδ|6dxds.

    Finally, also invoking Sobolev’s inequality, we see that

    K3:=Qtfε(ξεδ)5dxdsct0ηε(s)6(ξεδ(s))56/5dsct0ηε(s)6ξεδ(s)56dsct0ηε(s)V(1+ξεδ(s)66)ds.

    Collecting the above estimates, and noting that the function sηε(s)V is bounded in L1(0, T) by (3.20), we can apply the Gronwall lemma to conclude that

    ξεδL(0,T;L6(Ω))c. (3.26)

    At this point, we quickly show that ξεδ converges to ξε as δ0, at least for a subsequence. Indeed, one multiplies (3.23) first by ξεδ, and then by ξεt, and proves that

    ξεδH1(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

    ξεL(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+12Ω|ηε(t)|2dx3j=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)μεtξεηεtdxdsct0μεt(s)3ξε(s)6ηεt(s)2dsct0μεt(s)3ηεt(s)2ds14Qt|ηεt|2dxds+cT0μεt(s)23ds14Qt|ηεt|2dxds+c.

    Next, using (3.8) for ρεt and (3.20), we obtain that

    L2:=Qtg(ρε)ρεtηεηεtdxds14Qt|ηεt|2dxds+cQt|ηε|2dxds14Qt|ηεt|2dxds+c.

    Finally, in view of (3.20), we have

    L3:=Qt(ˉμg(ˉρ)ˉρtξεˉμg(ˉρ)ξεt+h)ηεtdxds14Qt|ηεt|2dxds+cQt(|ξε|2+|ξεt|2+1)dxds14Qt|ηεt|2dxds+c.

    By collecting the above estimates, we conclude that

    ηεtL2(0,T;H)+ηεL2(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

    ηεL2(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

    ξεξweaklystarinH1(0,T;H)L(0,T;L6(Ω)),ηεη weaklystarinH1(0,T;H)L(0,T;V)L2(0,T;W),

    as ε0, 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 ˉuUad be given and (ˉρ,ˉμ)=S(ˉu). Moreover, let hX be a function such that ˉu+ˉλhUad 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+λhUad for 0<λˉλ, since Uad is convex. We put, for any such λ,

    uλ:=ˉu+λh,(ρλ,μλ):=S(uλ),yλ:=ρλˉρλξ,zλ:=μλˉuλη.

    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

    ρλˉρH1(0,t;H)L(0,t;L6(Ω))+μλˉμH1(0,t;H)L(0,t;V)L2(0,t;W)K2λhL2(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|ρλˉρ|2a.e.in Q, (3.31)
    |g(ρλ)g(ˉρ)λg(ˉρ)ξ|c|yλ|+c|ρλˉρ|2a.e.in Q, (3.32)
    |g(ρλ)g(ˉρ)λg(ˉρ)ξ|c|yλ|+c|ρλˉρ|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,vL2(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)cyλL2(Qt)+R(λhL2(Qt)), (3.34)

    with a function R : (0; +∞)→(0; +∞) satisfying limσ0R(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=limλ0S(ˉu+λh)S(ˉu)λ(ξ,η)Yλ=limλ0yλH1(0,T;H)+zλL(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<λˉλ the pair (zλ,yλ) is a strong solution to the system

    (1+2g(ˉρ))zλt+g(ˉρ)ˉρtzλ+ˉμg(ˉρ)zλtΔzλ=2(g(ρλ)g(ˉρ))(μλtˉμt)2ˉμt(g(ρλ)g(ˉρ)λg(ˉρ)ξ)ˉμˉρt(g(ρλ)g(ˉρ)λg(ˉρ)ξ)ˉμ(g(ρλ)g(ˉρ))(ρλtˉρt)(μλˉμ)[(g(ρλ)g(ˉρ))ˉρt+g(ρλ)(ρλtˉρt)]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)+(μλˉμ)(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|2dxdsc7j=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)ytzdxdsγQty2tdxds+cγQtz2dxds. (3.41)

    Moreover, we have, by H¨older’s and Young’s inequalities and (3.30),

    I2:=2Qt(g(ρλ)g(ˉρ))(μλtˉμt)zdxdsct0ρλ(s)ˉρ(s)6μλt(s)ˉμt(s)2z(s)3cρλˉρL(0,t;L6(Ω))μλˉμH1(0,t;H)zL2(0,T;V)γz2L2(0,T;V)+cγλ4. (3.42)

    Next, we employ (3.32), the Hälder and Young inequalities, and (3.30), to infer that

    I3:=2Qtˉμt(g(ρλ)g(ˉρ)λg(ˉρ)ξ)zdxdscQt|ˉμt|(|y|+|ρλˉρ|2)|z|dxdsct0ˉμt(s)6(y(s)2z(s)3+ρλ(s)ˉρ(s)26z(s)2)dsγt0z(s)2Vds+cγt0ˉμt(s)2Vy(s)2Hds+ct0ˉμt(s)2Vz(s)2Hds+cρλˉρ4L(0,t;V)γt0z(s)2Vds+(1+cγ)t0ˉμt(s)2V(y(s)2H+z(s)2H)ds+cλ4, (3.43)

    where we observe that, in view of (2.17), the mapping sˉμt(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)zdxdscQt(|y|+|ρλˉρ|2)|z|dxdscQt(y2+z2)dxds+ct0ρλ(s)ˉρ(s)24z(s)2dscQt(y2+z2)dxds+cλ4. (3.44)

    In addition, arguing similarly, we have

    I5:=Qtˉμ(g(ρλ)g(ˉρ))(ρλtˉρt)zdxdsct0ρλ(s)ˉρ(s)6ρλt(s)ˉρt(s)2z(s)3dscρλˉρL(0,t;L6(Ω))ρλˉρH1(0,t;H)zL2(0,t;V)γt0z(s)2Vds+cγλ4, (3.45)

    as well as

    I6:=Qtˉρt(μλˉμ)(g(ρλ)g(ˉρ))zdxdscQt|μλˉμ||ρλˉρ||z|dxdsct0ρλ(s)ˉρ(s)6μλ(s)ˉμ(s)3z(s)2dscQtz2dxds+cλ4. (3.46)

    Finally, we find that

    I7:=Qt(μλˉμ)g(ρλ)(ρλtˉρt)zdxdsct0μλ(s)ˉμ(s)6ρλt(s)ˉρt(s)2z(s)3dscμλˉμL(0,t;V)ρλˉρH1(0,t;H)zL2(0,t;V)γt0z(s)2Vds+cγλ4. (3.47)

    In conclusion, combining the estimates (3.40)-(3.47), and choosing γ=18, we have shown that

    12zλ(t)2H+12t0zλ(s)2Vds18Qt|yλt|2dxds+cλ4+ct0(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+12y(t)2H6j=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:=QtyytdxdsγQty2tdxds+cγQty2dxds. (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)ytdxdscQt(|y|+|ρλˉρ|2)|yt|dxdsct0(y(s)2+ρλ(s)ˉρ(s)24)yt(s)2dsγQty2tdxds+cγQty2dxds+cγλ4. (3.51)

    By the same token, this time invoking (3.33), we find that

    J3:=Qtˉμ(g(ρλ)g(ˉρ)λg(ˉρ)ξ)ytdxdsγQty2tdxds+cγQty2dxds+cγλ4. (3.52)

    Moreover, we obviously have

    J4:=Qtg(ˉρ)zytdxdsγQty2tdxds+cγQtz2dxds. (3.53)

    Also, using (3.30) and the global bounds once more, we obtain that

    J5:=Qt(μλˉμ)(g(ρλ)g(ˉρ))ytdxdsct0μλ(s)ˉμ(s)6ρλ(s)ˉρ(s)3yt(s)2dsγQty2tdxdx+cγλ4. (3.54)

    Finally, invoking (3.34) and Young’s inequality, we have the estimate

    J6:=Qt(B[ρλ]B[ˉρ]λDB[ˉρ](ξ))ytdxdsB[ρλ]B[ˉρ]λDB[ˉρ](ξ)L2(Qt)ytL2(Qt)γQty2tdxds+cγy2L2(Qt)+cγ(R(λhL2(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

    14Qt|yλt|2dxds+12yλ(t)2Hc(t0yλ(s)2Hds+λ4+(R(λhL2(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+zλ2L2(0,t;V)+yλ(t)2H+yλt2L2(0,t;H)c1Z(λ)+c2t0(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(λhL2(Ω)))2. (3.58)

    Recalling that the mapping sˉμt(s)2V belongs to L1(0, T), we conclude from Gronwall’s lemma that, for every t(0,T],

    yλ2H1(0,t;H)+zλ2L(0,t;H)L2(0,t;V)c1Z(λ)exp(c2T0(1+ˉμt(s)2V)ds)cZ(λ). (3.59)

    Since limλ0Z(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 ˉuUad is a solution to the control problem (CP) with associated state (ˉρ,ˉμ)=S(ˉu). Then we have, for every vUad,

    β1Q(ˉρρQ)ξdxdt+β2Q(ˉμμQ)ηdxdt+β3Qˉu(vˉu)dxdt0, (3.60)

    where (ξ,η) denotes the (unique) solution to the linearized system (3.1)-(3.4) for h=vˉu.

    PROOF: Let vUad be arbitrary. Then h=vˉu is an admissible direction, since ˉu+λhUad for 0<λ1. For any such λ, we have

    0J(ˉu+λh,S(ˉu+λh))J(ˉu,S(ˉu))λJ(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 λ0. For the second summand, we obtain from Theorem 3.2 that

    limλ0J(ˉu,S(ˉu+λh)J(ˉu,S(ˉu))λ=β1Q(ˉρρQ)ξdxdt+β2Q(ˉμμQ)ηdxdt,

    whence the assertion follows.


    4. Existence and first-order necessary conditions of optimality

    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 ˉuUad.

    PROOF: Let {un}nNUad be a minimizing sequence for (CP), and let {(ρn,μn)}nN 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,

    unˉuweakly star in H1(0,T;H)L(Q),ρnˉρweakly star in H2(0,T;H)nW1,(0,T;L(Ω))H1(0,T;V),μnˉμweakly star in W1,(0,T;H)H1(0,T;V)L(0,T;W).

    Clearly, we have that ˉuUad 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 ˉρρ 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)tμng(ˉρ)tˉμweakly in L1(Q),μng(ρn)tρnˉμg(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, ˉμL(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 ˉuUad 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))ptg(ˉρ)ˉρtpΔpg(ˉρ)q= β2(ˉμμQ)in Q, (4.3)
    qt+F(ˉρ)qˉμg(ˉρ)q+g(ˉρ)(ˉμtpˉμpt)+DB[ˉρ](q)= β1(ˉρρQ)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)dxdtv,wL2(Q). (4.7)

    As, for every vL2(Q), the restriction of DB[ˉρ](v) to Qt depends only on vjQt, it follows that, for every wL2(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)CBwL2(Qt)wL2(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 ˉuUad 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

    pH1(0,T;H)L(0,T;V)L2(0,T;W)andqH1(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+12p(t)2V+12Ω|q(t)|2dxQt(pg(ˉρ)ˉρtpg(ˉρ)q)ptdxds+Qt((ˉμg(ˉρ)F(ˉρ))q+ˉμg(ˉρ)ptDB[ˉρ](q))qdxdsQtg(ˉρ)ˉμtpqdxds12Qt|pt|2dxds+cQt(p2+q2)dxds+cQt|ˉμt||p||q|dxds.

    The last integral is estimated as follows: employing the Hölder, Sobolev and Young inequalities, we have

    Qt|ˉμt||p||q|dxdsTtˉμt(s)3p(s)6q(s)2dscTt(ˉμt(s)2Vp(s)2V+q(s)2H)ds.

    As the function sˉμt(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,μεC(ˉQ) satisfying (3.8)-(3.9) and look for a solution (pε,qε) to the following problem:

    (1+2g(ρε))pεtg(ˉρ)ˉρtpεΔpεg(ˉρ)qε=β2(ˉμμQ)in Q, (4.10)
    qεtεΔqε+F(ˉρ)qεˉμg(ˉρ)qε+g(ρε)(μεtpεμεpεt)+DB[ˉρ](qε)=β1(qερQ)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,qεH1(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):VV; moreover, (,)H denotes the inner product in H (equivalent to the usual one) that one has chosen, the embedding HV 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 HV. 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

    φεMε,|ψε|Mε,|φε|Mε, and |φε|Mε, ae. 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

    pεtφεΔpε+λε1pελε1pεφεg(ˉρ)ˉρtpεφεg(ˉρ)qε=φεβ2(ˉμμQ),qεtεΔqε+φεΔpε+λε2qελε2qε+F(ˉρ)qεˉμg(ˉρ)qε+g(ρε)μεtpε+ψε(g(ˉρ)ˉρtpε+g(ˉρ)qε+β2(ˉμμQ))+DB[ˉρ](qε)=β1(ˉρρQ).

    By observing that

    φεΔpε=div(φεpε)+φεpε,

    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(ˉμμQ)(t)vdxΩqεt(t)wdx+aεt(t;(pε(t),qε(t)),w)+Ω(Rε1(pε,qε))(t)wdx=Ωψε(t)β2(ˉμμQ)(t)wdx+Ωβ1(ˉρρQ)(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,wV,

    aε1(t;ˆv,v):=Ω(φε(t)ˆvv+(φε(t)ˆv)v+λε1ˆvv)dx,aε2(t;(ˆv,ˆw),w):=Ω(εˆwwψε(t)ˆvw(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|2Mε|v||v|+λε1v2)dxΩ(α|v|2α2|v|2Mε2αv2+λε1v2)dx.

    Therefore, the choice λε1:=α2+M2ε2α yields

    aε1(t;v,v)α2vv2 for every vV and t[0,T].

    Next, we deal with aε2. We have, for every v,wV and t[0,T],

    aε2(t;(v,w),w)Ω(ε|w|2Mε|v||w|Mε|v||w|+λε2w2)dxΩ(ε|w|2ε2|w|2M2ε2ε|v|2M2ε2ε|v|2ε2|w|2+λε2w2)dx=Ω(ε2|w|2+(λε2ε2)w2M2εε|v|2)dx,

    and the choice λε2:=ε leads to

    aε2(t;(v,w),w)ε2w2VM2εεv2V.

    Therefore, if we choose λε such that λεα2M2εεε2, then we obtain

    λεaε1(t;v,v)+aε2(t;(v,w),w)ε2(v2V+w2V)

    for every (v, w)∈V and t[0,T]. Hence, if we define aε:[0,T]×V×VR 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((λεφεψε)(t)β2(ˉμμQ)(t)v+β1(ˉρρQ)(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 HV :

    V(ˆv,ˆw),(v,w)V=((ˆv,ˆw),(v,w))H=λεVˆv,vV+Vˆw,wV

    for every (ˆv,ˆw)H and (v,w)V, provided that the embedding HV 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,qεL2(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

    12pε(t)2V+Qt|pεt|2dxds+12Ω|qε(t)|2dx+εQt|qε|2dxdscQt|pε||pεt|dxds+cQt|qε||pεt|dxds+cQt|qε|2dxds+cQt|μεt||pε||qε|dxds+Qt|DB[ˉρ](qε)||qε|dxds+cpε2L2(Qt)+c.

    Just two of the terms on the right-hand side need some treatment. We have

    Qt|μεt||pε||qε|dxdsTtμεt(s)3pε(s)6qε(s)2dscTtpε(s)2Vds+cTtμεt(s)23qε(s)22ds,

    and we observe that the function sμεt(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|dxdsCBqε2L2(Qt).

    Therefore, we can apply the backward version of Gronwall’s lemma to obtain that

    pεH1(0,T;H)L(0,T;V)+qεH1(0,T;H)+ε1/2qεL2(0,T;V)c. (4.17)

    By comparison in (4.10), we see that Δpε is bounded in L2(Q). Hence,

    pεL2(0,T;W)c. (4.18)

    STEP 3: We multiply (4.11) by qεt and integrate over Qt. We obtain

    Qt|qεt|2dxds+ε2Ω|qε(t)|2dxcΩ|qε||qεt|dxds+cTtμεt(s)3pε(s)6qεt(s)2ds+cQt|pεt||qεt|dxds+Qt|DB[ˉρ](qε)||qεt|dxds.

    Thanks to (4.8) once more, we deduce that

    12Qt|qεt|2dxds+ε2Ω|qε(t)|2dxcQt|qε|2dxds+cTtμεt(s)23pε(s)2Vds+cQt|pεt|2dxds.

    Thus, (3.8) and (4.17) imply that

    qεtL2(0,T;H)+ε1/2qεL(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,

    pεpweakly star in H1(0,T;H)L(0,T;V)L2(0,T;W),qεqweakly in H1(0,T;H),εqε0strongly 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 ˉuUad 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)(vˉu)dxdt0vUad. (4.20)

    PROOF: We fix vUad and choose h=vˉu. 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((vˉu)pβ2(ˉμμQ)ηβ1(ˉρρQ)ξ)dxdt

    Now, we observe that the expression on the left-hand side coincides with

    Qt{(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(ˉρρQ)η+β2(ˉμμQ)ξ)dxdt=Q(vˉu)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 β13p is nothing but the L2(Q) orthogonal projection of ˉu onto Uad.


    Acknowledgements

    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.
  • This article has been cited by:

    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
  • Reader Comments
  • © 2016 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(5330) PDF downloads(1219) Cited by(10)

Article outline

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog