Electronic Journal of Differential Equations, Vol. 2009(2009), No. 83, pp. 1–22. ISSN: 1072-6691. URL: http://ejde.math.txstate.edu or http://ejde.math.unt.edu ftp ejde.math.txstate.edu OPTIMAL BOUNDARY CONTROL FOR A TIME DEPENDENT THERMISTOR PROBLEM VOLODYMYR HRYNKIV Abstract. An optimal control of a two dimensional time dependent thermistor problem is considered. The problem consists of two nonlinear partial differential equations coupled with appropriate boundary conditions which model the coupling of the thermistor to its surroundings. Based on physical considerations, an objective functional to be minimized is introduced and the convective boundary coefficient is taken to be the control. Existence of solutions to the state system and existence of the optimal control are proven. To characterize this optimal control, the optimality system consisting of the state and adjoint equations is derived. 1. Introduction Thermistor is a thermally sensitive resistor whose electrical conductivity changes drastically by orders of magnitude as the temperature reaches a certain threshold. Thermistors are used as temperature control elements in a wide variety of military and industrial equipment ranging from space vehicles to air conditioning controllers. They are also used in the medical field for localized and general body temperature measurement, in meteorology for weather forecasting as well as in chemical industries as process temperature sensors [11, 13]. We consider the two-dimensional time-dependent thermistor problem ut − ∆u − σ(u)|∇ϕ|2 = 0 ∇ · (σ(u)∇ϕ) = 0 in Ω × (0, T ), in Ω × (0, T ), ∂u + βu = 0 on ΓN × (0, T ), ∂n u = 0 on ΓD × (0, T ), ϕ = ϕ0 u = u0 (1.1) on ∂Ω × (0, T ), on Ω × {0}, where ϕ(x, t) is the electric potential, u(x, t) is the temperature, and σ(u) is the electrical conductivity. Here n denotes the outward unit normal and ∂/∂n = n · ∇ is the normal derivative on ∂Ω. The elliptic-parabolic system (1.1) describes the 2000 Mathematics Subject Classification. 49J20, 49K20. Key words and phrases. Optimal control; thermistor problem; elliptic systems. c 2009 Texas State University - San Marcos. Submitted December 10, 2008. Published July 10, 2009. 1 2 V. HRYNKIV EJDE-2009/83 heating of a conductor produced by an electric current. It is the consequence of Ohm’s law and Fourier’s law I = −σ(u)∇ϕ, Q = −k(u)∇u, (1.2) coupled with the conservation laws of energy and current ρc ∂u + ∇ · Q = I · E, ∂t ∇ · I = 0, (1.3) (1.4) where E denotes the electric field, ρ is the density of the conductor, and c is its heat capacity. Since the nonlinearity k(u) varies only slightly with u, it is of secondary importance, and therefore we can assume k(u) = 1. As the density ρ of the material and its specific heat c typically are constants, we assume that ρc = 1 (see [1]). A more detailed discussion of equations (1.1) can be found in [1, 2, 7, 9, 17]. Boundary conditions show how the thermistor is connected thermally and electrically to its surroundings. To describe boundary conditions for the temperature we decompose the boundary ∂Ω into two disjoint sets Γ̄D ∪ Γ̄N = ∂Ω. The Robin boundary condition for temperature represents the physically significant case of convection and/or radiation in situations where u is not too different from the ambient temperature. It is precisely this boundary condition that will be used as a control. It has been observed that large temperature gradients may cause the thermistor to crack. Numerical experiments indicate (see [7, 23]) that low values of the heat transfer coefficient β lead to small temperature variations. On the other hand, low values of the heat transfer coefficient result in high operating temperatures of a thermistor which are also undesirable. This motivated us to take the heat transfer coefficient as a control and to consider the optimal control problem of minimizing the heat transfer coefficient while keeping the operating temperature of the thermistor reasonably low. These physical considerations lead us to the following objective functional Z TZ Z J(β) = u dx dt + β 2 ds dt. 0 Ω ΓN ×(0,T ) Denoting the set of admissible controls by UM = {β ∈ L∞ (∂Ω × (0, T )) : 0 < λ ≤ β ≤ M }, the optimal control problem is Find β ∗ ∈ UM such that J(β ∗ ) = min J(β). β∈UM (1.5) Henceforth we use the standard notation for Sobolev spaces, we denote k · kp = k · kLp (Ω) for each p ∈ [1, ∞]; other norms will be clearly marked. We also denote Q := Ω × (0, T ) and ΓTN := ΓN × (0, T ). Analysis of both the steady-state and time-dependent thermistor equations with different types of boundary and initial conditions has received a great deal of attention. See [1, 2, 3, 4, 7, 9, 17, 18, 19, 22] for existence of weak solutions, existence/nonexistence of classical solutions, uniqueness and related regularity results in different settings with various assumptions on the coefficients. We mention in passing that few authors consider a thermistor problem which consists of two parabolic equations coupled with some boundary conditions (see [12, EJDE-2009/83 OPTIMAL CONTROL FOR A THERMISTOR PROBLEM 3 15]). This fully parabolic system comes from (1.2), (1.3), and the time dependent version of charge conservation (1.4), i.e., ∂q + ∇ · I = 0. ∂t (1.6) However, the vast majority of the existing literature on time dependent thermistor deals with the elliptic-parabolic system (1.1). This reflects a feasible thermistor scenario where the time derivative in (1.6) is dropped considering that the relaxation time for the potential is very small, hence giving the equation (1.4) (see [6, 17]). The first optimal control paper on a thermistor problem is the paper by Lee and Shilkin [12] where a fully parabolic system is considered and the source is taken to be the control. Optimal control of a nonlocal elliptic-parabolic case is studied in [5], where the applied potential is taken to be the control. Optimal control of a steady state thermistor problem is considered in [10], where the heat transfer coefficient is the control. The motivation for our work is threefold. First, the results from [10] are extended to the time dependent elliptic-parabolic thermistor system. As it was mentioned before, such an elliptic-parabolic thermistor model represents a reasonably realistic situation where we neglect the time derivative in (1.6) by taking into consideration the difference in time scale with the thermal processes. One of the technical difficulties in dealing with such an elliptic-parabolic system lies in the fact that there is no information on time derivative of the potential ϕ, and as a result one cannot use directly the standard compactness result from [16] to obtain strong convergence of sequences of potentials in appropriate spaces. To circumvent this difficulty one needs to incorporate the structure of the system. This is different from [12] in that a fully parabolic system and a different type of control were considered there. Secondly, we present a complete and self-contained derivation of the optimality system, whereas in [5] the optimality system was given only for the special case of constant σ(u). Thirdly, the choice of the convective boundary condition as a control for this time dependent problem seems to be quite appropriate from point of view of practical applications. In section 2 we derive a priori estimates under the assumption of small boundary data and prove existence of solutions to the state system. In section 3 we prove existence of an optimal control. The optimality system is derived and an optimal control is characterized in section 4. 2. A priori estimates and existence of solutions to the state system We make the following assumptions: • • • • • Ω ⊂ R2 is a bounded domain with smooth boundary; σ(s) ∈ C 1 (R), 0 < C1 ≤ σ(s) ≤ C2 for all s ∈ R; σ(s) is Lipschitz: |σ(s1 ) − σ(s2 )| ≤ K |s1 − s2 | for all s1 , s2 ∈ R; 1,∞ ϕ0 ∈ W (∂Ω × (0, T )); Extending ϕ0 to the whole domain Ω × (0, T ) and using the same notation, 1,∞ i.e., ϕ0 ∈ W (Q), we assume that kϕ0 kW 1,∞ (Q) is sufficiently small. The following lemma guarantees a better regularity for solutions of equations in divergence form. It is taken from [20]. More information about this lemma can be found in [14]. 4 V. HRYNKIV EJDE-2009/83 Lemma 2.1. Let Ω ⊂ Rn be a bounded domain with a smooth boundary. Assume that g ∈ (L2 (Ω))n and a ∈ C(Ω̄) with minΩ̄ a > 0. Let u be the weak solution of the following problem −∇ · (a∇u) = ∇ · g u=0 in Ω on ∂Ω. Then for each p > 2 there exists a positive constant c∗ depending only on n, Ω, a, and p such that if g ∈ (L2 (Ω))n then we have k∇ukp ≤ c∗ (kgkp + k∇uk2 ). Let VD := {v ∈ H 1 (Ω) : v = 0 on ΓD }. By a weak solution to (1.1) we mean a pair (u, ϕ) such that u ∈ L2 (0, T ; VD (Ω)), ϕ − ϕ0 ∈ L2 (0, T ; H01 (Ω)) and Z T Z Z Z hut , vi + ∇u · ∇v dx dt + βuv ds dt = σ(u)|∇ϕ|2 v dx dt, 0 Q ΓT Q N (2.1) Z σ(u)∇ϕ · ∇w dx dt = 0, Q for all v ∈ L2 (0, T ; H 1 (Ω)) ∩ L∞ (Q) and w ∈ L2 (0, T ; H01 (Ω)), where h·, ·i denotes duality bracket between VD0 and VD , with VD0 being the dual of VD . The quadratic term on the right hand side of the first equation in (2.1) can be dealt with as in [21, 17], and we obtain Z T Z Z hut , vi dt + ∇u · ∇v dx dt + βuv ds dt 0 ΓT Q ZN = (ϕ0 − ϕ)σ(u)∇ϕ · ∇v dx dt + (σ(u)∇ϕ · ∇ϕ0 ) v dx dt, Q Q Z σ(u)∇ϕ · ∇w dx dt = 0, Z (2.2) Q for all v ∈ L (0, T ; H (Ω)) ∩ L∞ (Q) and w ∈ L2 (0, T ; H01 (Ω)). By taking p = r in Lemma 2.1, we can say that for each r > 2, ϕ(·, t) ∈ W 1,r (Ω) for a.e. t ∈ (0, T ). Therefore, since ∇ϕ(·, t) ∈ Lr (Ω) and ∇ϕ0 (·, t) ∈ L2 (Ω) a.e. t ∈ (0, T ), it follows 0 that there exists s0 such that ∇ϕ(·, t) · ∇ϕ0 (·, t) ∈ Ls (Ω) and 1 1 1 = + . (2.3) s0 2 r Next, since Ω ⊂ R2 it follows v(·, t) ∈ H 1 (Ω) ⊂⊂ Ls (Ω) for s ∈ [1, ∞), conjugate of s0 in (2.3): 1 1 + = 1. (2.4) 0 s s Also, by the weak maximum principle (see [17, formula (2.14)]) 2 1 kϕkL∞ (Q) ≤ sup |ϕ0 |. (2.5) ΓD Note that r > 2 is chosen from Lemma 2.1, then s0 and s are determined and satisfy s0 < 2 < s. First, we show existence of solutions to the state system. Define W (0, T ) := {v ∈ L2 (0, T ; VD ) : vt ∈ L2 (0, T ; VD0 )}. Theorem 2.2. There exists a weak solution to the state system (1.1). EJDE-2009/83 OPTIMAL CONTROL FOR A THERMISTOR PROBLEM 5 We show existence of solutions to the state system by using the Schauder’s fixed point theorem (see, for example [8]). Let R : W (0, T ) → W (0, T ) be the operator defined as follows R := S ◦ T : ϑ 7→ u, with T : L2 (0, T ; H 1 (Ω)) ∩ L∞ (Q) 3 ϑ 7→ ϕ ∈ L2 (0, T ; H01 (Ω)) where for a given ϑ, the function ϕ solves uniquely Z σ(ϑ)∇ϕ · ∇w dx dt = 0, (2.6) Q 2 (0, T ; H01 (Ω)), ∞ for all w ∈ L and where S is defined as follows S : L2 (0, T ; H01 (Ω)) × 2 1 L (0, T ; H (Ω)) ∩ L (Q) 3 (ϕ, ϑ) 7→ u ∈ L2 (0, T ; H 1 (Ω)) ∩ L∞ (Q), with u being the unique solution to Z Z Z T hut , vi dt + ∇u · ∇v dx dt + βuv ds dt 0 Q ΓT N (2.7) Z Z = (ϕ0 − ϕ)σ(ϑ)∇ϕ · ∇v dx dt + (σ(ϑ)∇ϕ · ∇ϕ0 ) v dx dt, Q Q for all v ∈ L2 (0, T ; H 1 (Ω)) ∩ L∞ (Q). Standard theory implies that there exists a unique solution to (2.6), (2.7) which means that the map R is well-defined. Before we proceed with the proof of the theorem, we need to derive a priori estimates. Lemma 2.3. Given ϑ ∈ L2 (0, T ; H 1 (Ω)) ∩ L∞ (Q), let ϕ = T (ϑ) and u = S(ϕ, ϑ) be solutions satisfying (2.6) and (2.7). Then there exist constants c1 , c2 , c3 , and c4 independent of ϑ such that kϕkL2 (0,T ;H 1 (Ω)) ≤ c1 , (2.8) kukL2 (0,T ;VD (Ω)) ≤ c2 , (2.9) kukL2 (0,T ;L2 (ΓN )) ≤ c3 , (2.10) kut kL2 (0,T ;VD0 (Ω)) ≤ c4 . (2.11) Proof of Lemma 2.3. Taking w = ϕ − ϕ0 in (2.6) and using properties of σ, we obtain (2.8). Taking v = u in (2.7), we have Z T Z Z βu2 ds dt hut , ui dt + |∇u|2 dx dt + 0 ΓT N Q Z Z (ϕ0 − ϕ)σ(ϑ)∇ϕ · ∇u dx dt + = (σ(ϑ)∇ϕ · ∇ϕ0 ) u dx dt. Q Q Using properties of β we obtain Z Z Z TZ Z TZ u2 u2 2 − + |∇u| dx dt + λ u2 ds dt Ω×{T } 2 Ω×{0} 2 0 Ω 0 ΓN Z TZ Z TZ ≤ (ϕ0 − ϕ)σ(ϑ)∇ϕ · ∇u dx dt + (σ(ϑ)∇ϕ · ∇ϕ0 ) u dx dt. 0 Ω We can write Z Ω×{T } Z ≤ Ω 0 u2 + 2 Z u20 + C3 2 T Z 0 |∇u|2 dx dt + λ Ω Z T Z 0 T Ω Z u2 ds dt ΓN Z Z T Z |∇ϕ| · |∇u| dx dt + C4 0 Ω |∇ϕ| · |u| dx dt 0 Ω 6 V. HRYNKIV ≤ C5 + C6 Z Z ≤ C5 + C8 EJDE-2009/83 T { k∇ϕk2 · k∇uk2 + k∇ϕk2 · kuk2 } dt 0 T k∇ϕk22 0 1 dt + 2 T Z 0 Z |∇u|2 dx dt, Ω where we used Poincare inequality. We obtained Z Z Z Z TZ u2 1 T |∇u|2 dx dt + λ + u2 ds dt 2 0 Ω Ω×{T } 2 0 ΓN Z T ≤ C5 + C8 k∇ϕk22 dt ≤ C9 , (2.12) 0 where the last inequality in (2.12) follows from (2.8) (also, see the estimate (2.15) on p. 245 of [17]). This implies estimates (2.9) and (2.10). Using the PDE and the previous estimates one can show (2.11). This ends the proof of Lemma 2.3. We continue the proof of Theorem 2.2. Let W0 := {v ∈ W (0, T ) : kvkL2 (0,T ;VD ) ≤ c2 , kvkL2 (0,T ;VD0 ) ≤ c4 , v(x, 0) = u0 (x, 0)}. We show that the operator R : W0 → W0 is continuous with respect to weak convergence in W (0, T ). Then, since W0 is weakly compact, it will follow that R has a fixed point in W0 . Let {ϑj } ⊂ W0 w be a sequence such that ϑj * ϑ in W (0, T ), and let uj = u(ϑj ) and ϕj = ϕ(ϑj ) be the corresponding sequences of solutions. A priori estimates imply that on a subsequence w in L2 (0, T ; VD (Ω)), ϑj * ϑ ∂ϑj w ∂ϑ * in L2 (0, T ; VD0 (Ω)), ∂t ∂t w uj * u in L2 (0, T ; VD (Ω)), ∂uj w ∂u * in L2 (0, T ; VD0 (Ω)), ∂t ∂t w ϕj * ϕ in L2 (0, T ; H 1 (Ω)), w uj * u 2 2 in L (0, T ; L (ΓN )). (2.13) (2.14) (2.15) (2.16) (2.17) (2.18) Note that by the standard compactness result (see [16]), convergence in (2.13), (2.14), (2.15), and (2.16), imply that on a subsequence, not relabeled, we have s in L2 (Q), (2.19) s 2 (2.20) ϑj → ϑ uj → u s in L (Q), σ(ϑj ) → σ(ϑ) in L2 (Q). (2.21) Moreover, it can be shown that s ϕj → ϕ s ∇ϕj → ∇ϕ in L2 (Q), in L2 (Q). (2.22) (2.23) Note that (2.22) and (2.23) can be obtained by incorporating the structure of the PDE system as we do not have information on ϕt . For the proof of (2.22) and (2.23) the reader is referred to [1], p. 1132-1133. EJDE-2009/83 OPTIMAL CONTROL FOR A THERMISTOR PROBLEM We have Z Z Z T h(uj )t , vi dt + ∇uj · ∇v dx dt + 0 Q βuj v ds dt ΓT N Q Z = 7 Z (2.24) (ϕ0 − ϕj )σ(ϑj )∇ϕj · ∇v dx dt + (σ(ϑj )∇ϕj · ∇ϕ0 ) v dx dt, Q Z σ(ϑj )∇ϕj · ∇w dx dt = 0, Q for all v ∈ L (0, T ; H (Ω)) ∩ L∞ (Q) and w ∈ L2 (0, T ; H01 (Ω)). It is immediate that Z T Z T hujt , vi dt → hut , vi dt, 0 0 Z Z ∇uj · ∇v dx dt → ∇u · ∇v dx dt, Q Q Z Z βuj v ds dt → βuv ds dt as j → ∞. 2 1 ΓT N ΓT N We show Z Z (ϕ0 − ϕj )σ(ϑj )∇ϕj · ∇v dx dt → (ϕ0 − ϕ)σ(ϑ)∇ϕ · ∇v dx dt. Q (2.25) Q We estimate the difference between the sequence terms and the limit Z Z (ϕ0 − ϕj )σ(ϑj )∇ϕj · ∇v dx dt − (ϕ0 − ϕ)σ(ϑ)∇ϕ · ∇v dx dt Q Q Z ≤ σ(ϑj )(ϕ0 − ϕj )∇ϕj · ∇v − σ(ϑ)(ϕ0 − ϕj )∇ϕ · ∇v dx dt Q Z σ(ϑ)(ϕ0 − ϕj )∇ϕ · ∇v − σ(ϑ)(ϕ0 − ϕ)∇ϕ · ∇v dx dt := A + B, + Q where B := w R Q σ(ϑ)(ϕ−ϕj )∇ϕ·∇v dx dt → 0 by the weak convergence ϕj * ϕ in 0 L2 (0, T ; Ls (Ω)) (by Lemma 2.1 and the fact that σ(ϑ)∇ϕ · ∇v ∈ L2 (0, T ; Ls (Ω))). For term A, we have Z A:= (ϕ0 − ϕj )(σ(ϑj ) − σ(ϑ))∇ϕj · ∇v dx dt Q Z + (ϕ0 − ϕj )σ(ϑ)(∇ϕj − ∇ϕ) · ∇v dx dt = A1 + A2 . Q Note that Z (ϕ0 − ϕj )(σ(ϑj ) − σ(ϑ))∇ϕj · ∇v dx dt A1 : = Q Z T Z ≤ 2M̃ K |ϑj − ϑ| · |∇ϕj | · |∇v| dx dt 0 Ω T Z ≤C kϑj − ϑks k∇ϕj kr k∇vk2 dt 0 Z ≤ C sup k∇ϕj kr t T kϑj − ϑks k∇vk2 dt 0 8 V. HRYNKIV EJDE-2009/83 ≤ Ckϑj − ϑkL2 (0,T ;Ls (Ω)) k∇vkL2 (Q) → 0 by the Simon compactness result [16], i.e., W (0, T ) ⊂⊂ L2 (0, T ; Ls (Ω)). Also, note sup k∇ϕj kr ≤ Cϕ0 , (2.26) t by [10, formula (2.9)], and Cϕ0 here denotes a constant which depends on the s boundary data ϕ0 . Hence on a subsequence ϑj → ϑ in L2 (0, T ; Ls (Ω)). In the above, C is a generic constant and M̃ := supQ |ϕ0 |. For A2 , we write Z A2 ≤ (ϕ0 − ϕ)σ(ϑ)(∇ϕj − ∇ϕ) · ∇v dx dt Q Z + (ϕ − ϕj )σ(ϑ)(∇ϕj − ∇ϕ) · ∇v dx dt → 0 Q by (2.23). Similarly, it can be shown that Z Z σ(ϑj )(∇ϕj · ∇ϕ0 )v dx dt → σ(ϑ)(∇ϕ · ∇ϕ0 )v dx dt, Q Q Z Z σ(ϑj )∇ϕj · ∇w dx dt → σ(ϑ)∇ϕ · ∇w dx dt. Q Q Letting j → ∞ in (2.24), we conclude by the Schauder theorem that R has a fixed point in W0 and hence (u, ϕ) with ϕ = T (u) solves (1.1). 3. Existence of an optimal control Now we proceed to the proof of existence of an optimal control. Theorem 3.1. There exists a solution to the optimal control problem (1.5). Proof. Let {βn }∞ n=1 ⊂ UM be a minimizing sequence, lim J(βn ) = inf J(β). n→∞ β∈UM Let un = u(βn ) and ϕn = ϕ(βn ) be the corresponding solutions to Z T Z Z hunt , vi dt + ∇un ∇v dx dt + βn un v ds dt 0 ΓT N Q Z = Q Z (ϕ0 − ϕn )σ(un )∇ϕn ∇v dx dt + (σ(un )∇ϕn · ∇ϕ0 ) v dx dt, Q Z σ(un )∇ϕn · ∇w dx dt = 0, (3.1) Q for all v ∈ L (0, T ; H (Ω)) ∩ L∞ (Q) and w ∈ L2 (0, T ; H01 (Ω)). Note that the existence of an optimal control can be derived from the proof of Theorem 2.2 and the estimates (2.8), (2.9), (2.10), and (2.11). In passing to the limit as n → ∞ in (3.1), we only need to check the term with βn . We have (in addition to convergences from the previous section): 2 1 βn * β ∗ w∗ in L∞ (∂Ω × (0, T )), s in L2 (0, T ; L2 (∂Ω)). un → u∗ EJDE-2009/83 OPTIMAL CONTROL FOR A THERMISTOR PROBLEM 9 This implies Z Z βn un v ds → ΓT N β ∗ u∗ v ds as n → ∞. ΓT N Letting n → ∞ in (3.1), we obtain that (u∗ , ϕ∗ ) is a weak solution associated with β ∗ , i.e., u∗ = u(β ∗ ) and ϕ∗ = ϕ(β ∗ ). Using weak lower semicontinuity of J(β) with respect to L2 norm one can show that the infimum is achieved at β ∗ . 4. Derivation of the optimality system To characterize an optimal control, we need to derive an optimality system which consists of the original state system coupled with an adjoint system. To obtain the necessary conditions for the optimality system, we differentiate the objective functional with respect to the control. Since the objective functional also depends on u which depends on the control β and is coupled to ϕ through the PDE (1.1), we will need to differentiate u and ϕ with respect to control β. Theorem 4.1 (Sensitivities). If the boundary data ϕ0 are sufficiently small; i.e., if kϕ0 kW 1,∞ (Q) is small enough, then the mapping β 7→ (u, ϕ) is differentiable in the following sense: u(β + ε`) − u(β) w * ψ1 in L2 (0, T ; H01 (Ω)), ε ϕ(β + ε`) − ϕ(β) w * ψ2 in L2 (0, T ; H01 (Ω)) as ε → 0 ε (4.1) for any β ∈ UM and ` ∈ L∞ (∂Ω × (0, T )) such that (β + ε`) ∈ UM for small ε. Moreover, the sensitivities, ψ1 ∈ L2 (0, T ; H01 (Ω)) and ψ2 ∈ L2 (0, T ; H01 (Ω)), satisfy −(ψ1 )t + ∆ψ1 + σ 0 (u)|∇ϕ|2 ψ1 + 2σ(u)∇ϕ · ∇ψ2 = 0 0 ∇ · [σ (u)ψ1 ∇ϕ + σ(u)∇ψ2 ] = 0 in Ω × (0, T ), in Ω × (0, T ), ∂ψ1 + βψ1 + `u = 0 on ΓN × (0, T ), ∂n ψ1 = 0 on ΓD × (0, T ), ψ2 = 0 ψ1 = 0 (4.2) on ∂Ω × (0, T ), on Ω × {0}. Proof. Recall that we denoted: u = u(β) and ϕ = ϕ(β). Now we also denote uε = u(β ε ), ϕε = ϕ(β ε ), where β ε := β + ε`. The weak formulation for (uε , ϕε ) is Z T huεt , vi dt + 0 Z ∇uε ∇v dx dt + β ε uε v ds dt ΓT N Q Z = Z Z (ϕ0 − ϕε )σ(uε )∇ϕε ∇v dx dt + (σ(uε )∇ϕε · ∇ϕ0 ) v dx dt, Q Q Z ε ε σ(u )∇ϕ · ∇w dx dt = 0, Q (4.3) 10 V. HRYNKIV EJDE-2009/83 for all v ∈ L2 (0, T ; H 1 (Ω)) ∩ L∞ (Q) and w ∈ L2 (0, T ; H01 (Ω)). Similarly, for (u, ϕ) Z T Z Z hut , ṽi dt + ∇u∇ṽ dx dt + βuṽ ds dt 0 Q ΓT ZN = (ϕ0 − ϕ)σ(u)∇ϕ∇ṽ dx dt + (σ(u)∇ϕ · ∇ϕ0 ) ṽ dx dt, Q Q Z σ(u)∇ϕ · ∇w̃ dx dt = 0, Z (4.4) Q for all ṽ ∈ L (0, T ; H (Ω)) ∩ L∞ (Q) and w̃ ∈ L2 (0, T ; H01 (Ω)). Take test functions (uε − u)/ε and (ϕε − ϕ)/ε, subtract the corresponding equations, and divide by ε: Z T ε Z ut − ut uε − u uε − u uε − u , dt + ∇ dx dt ∇ ε ε ε ε 0 Q Z uε − u uε − u ds dt + β ε ε ΓT N Z uε − u =− `uε ds dt ε ΓT N (4.5) Z h i 1 uε − u ε ε ε + (ϕ0 − ϕ )σ(u )∇ϕ − (ϕ0 − ϕ)σ(u)∇ϕ · ∇ dx dt ε Q ε Z h i uε − u 1 σ(uε )∇ϕε − σ(u)∇ϕ · ∇ϕ0 dx dt, + ε Q ε Z h ϕε − ϕ i ϕε − ϕ 1 σ(uε )∇ϕε · ∇ − σ(u)∇ϕ · ∇ dx dt = 0. ε Ω ε ε 2 1 We present the detailed derivation of the L2 (0, T ; H 1 (Ω)) estimate for (ϕε − ϕ)/ε. Since (ϕε − ϕ)/ε ∈ L2 (0, T ; H01 (Ω)) it follows from Poincaré’s inequality that it is enough to derive a bound on k∇(ϕε − ϕ)/εkL2 (Q) . From the second equation in (4.5) we obtain Z Z ϕε ϕε − ϕ ϕ ϕε − ϕ ε σ(u )∇ ·∇ dx dt = σ(u)∇ ·∇ dx dt. (4.6) ε ε ε ε Q Q Taking into account (4.6) we can write Z ϕε − ϕ 2 σ(u) ∇ dx dt ε Q Z ϕε − ϕ ϕε − ϕ = σ(u)∇ ·∇ dx dt ε ε Q Z Z ϕε ϕε − ϕ ϕ ϕε − ϕ = σ(u)∇ ·∇ dx dt − σ(u)∇ ·∇ dx dt (4.7) ε ε ε ε Q Q Z Z ϕε ϕε − ϕ ϕε ϕε − ϕ = σ(u)∇ ·∇ dx dt − σ(uε )∇ ·∇ dx dt ε ε ε ε Q Q Z σ(u) − σ(uε ) ε ϕε − ϕ ∇ϕ · ∇ dx dt. = ε ε Q Using (4.6) and (4.7) we obtain Z TZ ϕε − ϕ 2 C1 ∇ dx dt ε 0 Ω EJDE-2009/83 OPTIMAL CONTROL FOR A THERMISTOR PROBLEM Z T 11 ϕε − ϕ 2 dx dt ε Ω σ(u) − σ(uε ) ε ϕε − ϕ ∇ϕ · ∇ dx dt ε ε Z σ(u) ∇ ≤ Z0 = Q T ϕε − ϕ uε − u ε · k∇ϕ k · ∇ dt r s 2 ε ε 0 Z T uε − u ϕε − ϕ ≤ K sup k∇ϕε kr · ∇ dt s 2 ε ε t 0 uε − u ϕε − ϕ ≤ Cϕ0 ∇ , 2 s L (0,T ;L (Ω)) L2 (0,T ;L2 (Ω)) ε ε where we used (2.26). We can write ϕε − ϕ uε − u uε − u ∇ ≤ C ≤ C̃ . 2 2 s L (Q) L (0,T ;L (Ω)) L2 (0,T ;H01 (Ω)) ε ε ε Z ≤K (4.8) Since L2 (0, T ; H 1 (Ω)) ⊂ L2 (0, T ; Ls (Ω)), it follows by Poincaré’s inequality ϕε − ϕ ε L2 (0,T ;H 1 (Ω)) ≤ Cϕ0 uε − u ε L2 (0,T ;H 1 (Ω)) . (4.9) Now we proceed to the derivation of H 1 -norm estimate of (uε − u)/ε using (4.5): Z Z Z 1 uε − u 2 u ε − u 2 uε − u 2 ∇ dx dt + λ dx + ds dt 2 Ω×{T } ε ε ε ΓT Q N Z uε − u ≤ − `uε ds dt ε ΓT Z Nh i uε − u 1 (ϕ0 − ϕε )σ(uε )∇ϕε − (ϕ0 − ϕ)σ(u)∇ϕ ∇ dx dt + ε Q ε Z h i 1 uε − u + σ(uε )∇ϕε − σ(u)∇ϕ ∇ϕ0 dx dt ε Q ε Z uε − u = − `uε ds dt + C + D . ε ΓT N We have Z Z Z 1 uε − u 2 u ε − u 2 u ε − u 2 dx + ∇ dx dt + λ ds dt 2 Ω×{T } ε ε ε Q ΓT N Z Z uε − u uε − u ε ≤ |`| · |u | · ds dt + |C| + |D| ≤ M |uε | · ds dt + |C| + |D| ε ε ΓT ΓT N N uε − u + |C| + |D| L2 (0,T ;L2 (ΓN )) ε ε u −u ≤ c8 kuε kL2 (0,T ;H 1 (Ω)) · + |C| + |D|. L2 (0,T ;L2 (ΓN )) ε We estimate C term first Z σ(uε )∇ϕε − σ(u)∇ϕ uε − u |C| ≤ |ϕ0 − ϕ| · · ∇ dx dt ε ε Q Z uε − u ϕ − ϕε σ(uε )|∇ϕε | · ∇ dx dt + ε ε Q ≤ c7 kuε kL2 (0,T ;L2 (ΓN )) · 12 V. HRYNKIV EJDE-2009/83 σ(uε ) − σ(u) uε − u · |∇ϕε | · ∇ dx dt ε ε Q Z ϕε − ϕ uε − u + 2M̃ σ(u) · ∇ · ∇ dx dt ε ε Q Z uε − u ϕ − ϕε + C2 · |∇ϕε | · ∇ dx dt ε ε Q Z uε − u uε − u ≤ c2 · |∇ϕε | · ∇ dx dt ε ε Q Z uε − u ϕε − ϕ · ∇ dx dt + c3 ∇ ε ε Q Z uε − u ϕ − ϕε + c4 · |∇ϕε | · ∇ dx dt. ε ε Q Z ≤ 2M̃ Note that c2 , c3 , c5 , and c6 depend on the data ϕ0 . We continue estimating |C|: Z Z uε − u uε − u ϕε − ϕ uε − u |C| ≤ c9 · ∇ dx dt + c3 ∇ · ∇ dx dt ε ε ε ε Q Q Z uε − u ϕε − ϕ + c10 · ∇ dx dt ε ε Q Z T Z T uε − u uε − u ϕε − ϕ ≤ c9 · ∇ dt + c ∇ 3 2 2 2 ε ε ε 0 0 Z T ε ε ε u − u u − u ϕ −ϕ dt + c10 · ∇ dt × ∇ 2 2 2 ε ε ε 0 Z T Z T uε − u uε − u ϕε − ϕ ≤ c11 ∇ · ∇ dt + c ∇ 3 2 2 2 ε ε ε 0 0 Z T uε − u ϕε − ϕ uε − u dt + c10 · ∇ dt × ∇ 2 2 2 ε ε ε 0 uε − u uε − u ϕε − ϕ ∇ ≤ c11 ∇ · + c ∇ 3 L2 (Q) L2 (Q) L2 (Q) ε ε ε uε − u ϕε − ϕ uε − u × ∇ + c10 · ∇ L2 (Q) L2 (Q) L2 (Q) ε ε ε uε − u 2 uε − u 2 uε − u 2 ≤ c11 ∇ + c12 ∇ + c13 ∇ L2 (Q) L2 (Q) L2 (Q) ε ε ε uε − u 2 , ≤ c14 ∇ L2 (Q) ε where we used Poincaré’s inequality on (uε − u)/ε and the Cauchy inequality. Similarly, for D we obtain Z Z uε − u uε − u ϕε − ϕ uε − u |D| ≤ c5 · |∇ϕε | · dx dt + c6 · dx dt ∇ ε ε ε ε Q Q uε − u 2 ≤ c15 ∇ . L2 (Q) ε Notice that c14 and c15 depend on ϕ0 . Hence, we obtain Z Z uε − u 2 uε − u 2 uε − u 2 1 dx + ∇ + λ ds dt 2 (Q) L 2 Ω×{T } ε ε ε ΓT N EJDE-2009/83 OPTIMAL CONTROL FOR A THERMISTOR PROBLEM 13 uε − u uε − u 2 + c ∇ 16 2 2 L (0,T ;L (ΓN )) L2 (Q) ε ε ε ε λ u −u 2 u − u 2 ≤ c17 kuε k2L2 (0,T ;H 1 (Ω)) + . 2 (0,T ;L2 (Γ )) + c16 ∇ L L2 (Q) N 2 ε ε Observe that kuε k2L2 (0,T ;H 1 (Ω)) is bounded by (2.9). Take ϕ0 small enough to get Z 1 λ uε − u 2 u ε − u 2 uε − u 2 ≤ c18 , dx + k ∇ 2 (Q) + L L2 (0,T ;L2 (ΓN )) 2 Ω×{T } ε ε 2 ε ≤ c8 kuε kL2 (0,T ;H 1 (Ω)) · where k := 1 − c16 > 0. Collecting the estimates and using the pde (4.3) to get the estimate of the time derivative quotient, gives uε − u ≤ c, L2 (0,T ;H01 (Ω)) ε uε − u ≤ c, L2 (0,T ;L2 (ΓN )) ε uεt − ut ≤ c, L2 (0,T ;H −1 (Ω)) ε ε ϕ −ϕ ≤ c, L2 (0,T ;H 1 (Ω)) ε where the generic constant c > 0 does not depend on ε. Claim: we can derive a better estimate on (ϕε − ϕ)/ε than we have. This better estimate is needed to pass to the limit as ε → 0 in the derivation of the sensitivity equations. Indeed, we can write for a.e. t ∈ (0, T ) ∇ · (σ(uε )∇ϕε ) = 0 ∇ · (σ(u)∇ϕ) = 0 in Ω, in Ω. Subtract these equations as well as appropriate boundary conditions to obtain i h σ(uε ) − σ(u)) ϕε − ϕ ∇ · σ(u)∇ ∇ϕε in Ω, =∇· ε ε (4.10) ε ϕ −ϕ = 0 on ∂Ω. ε By Lemma 2.1, for a.e. t ∈ (0, T ), we have k∇ϕε k2r ≤ c, where r > 2 and ϕε solves ∇ · (σ(uε )∇ϕε ) = 0 ε ϕ = ϕ0 in Ω, on ∂Ω. Also, since Ω ⊂ R2 it follows that H 1 (Ω) ⊂ Lq (Ω) for any q ∈ [1, ∞). Thus, we can write for a.e. t ∈ (0, T ), uε − u uε − u ≤ ≤ c. (4.11) 2r H 1 (Ω) ε ε Therefore, returning to (4.10), we obtain σ(uε ) − σ(u) ϕε − ϕ ϕε − ϕ ≤ c∗ ∇ϕ r + ∇ . (4.12) ∇ r 2 ε ε ε We can estimate the first term on the right hand side of (4.12), Z σ(uε ) − σ(u) σ(uε ) − σ(u) r r ∇ϕ r = · |∇ϕε |r dx ε ε Ω Z uε − u r · |ϕε |r dx ≤C ε Ω 14 V. HRYNKIV ≤C EJDE-2009/83 uε − u ε 2r 2r + k∇ϕε k2r 2r ≤ c. Therefore, ϕε − ϕ ϕε − ϕ ≤ C, ≤ C, r r ε ε where a generic constant C > 0 is independent of ε. This implies that ∇ s ϕε → ϕ s ∇ϕε → ∇ϕ (4.13) in Lr (Ω), in Lr (Ω) as ε → 0. Moreover, by reiterating this argument we can show that for a.e. t ∈ (0, T ) ϕε − ϕ ϕε − ϕ ≤ C, ≤ C. (4.14) ∇ 2r 2r ε ε This proves the claim. Now we can proceed to the derivation of sensitivity equations, starting with Z T ε Z Z ut − ut uε − u uε − u , v dt + ∇ ∇v dx dt + β v ds dt ε ε ε 0 Q ΓT N Z Z h i 1 ε `u v ds dt + =− (ϕ0 − ϕε )σ(uε )∇ϕε − (ϕ0 − ϕ)σ(u)∇ϕ · ∇v dx dt ε Q ΓT ZN h i 1 + σ(uε )∇ϕε − σ(u)∇ϕ · ∇ϕ0 v dx dt, ε Q Z h i 1 σ(uε )∇ϕε − σ(u)∇ϕ · ∇w dx dt = 0. ε Ω (4.15) The following convergence are immediate Z T ε Z T E ut − ut , v dt → (ψ1 )t , v dt, ε Z 0 ε Z0 u − u ∇ ∇v dx dt → ∇ψ1 · ∇v dx dt, ε Q Q Z Z uε − u β v ds dt → βψ1 v ds dt, ε ΓT ΓT N N Z Z ε `u v ds dt → `uv ds dt. ΓT N ΓT N We write the second term on the right hand side of (4.15) in the form Z h i 1 (ϕ0 − ϕε )σ(uε )∇ϕε · ∇v − (ϕ0 − ϕ)σ(u)∇ϕ · ∇v dx dt ε Q Z h i 1 (ϕ0 − ϕε ) σ(uε )∇ϕε − σ(u)∇ϕ · ∇v dx dt = ε Q Z 1 + (ϕ − ϕε )σ(u)∇ϕ · ∇v dx dt ε Q = G + F. EJDE-2009/83 OPTIMAL CONTROL FOR A THERMISTOR PROBLEM 15 As far as the term F is concerned, we have that σ(u)∇ϕ · ∇v ∈ L2 (Q) and (ϕ − w ϕε )/ε * −ψ2 in L2 (Q), and therefore Z F →− ψ2 σ(u)∇ϕ · ∇v dx dt Q as ε → 0. Now we consider terms coming from G. We write Z h i 1 G= (ϕ0 − ϕε ) σ(uε )∇ϕε − σ(u)∇ϕ · ∇v dx dt ε Q Z h i 1 (ϕ0 − ϕε ) σ(uε ) − σ(u) ∇ϕε · ∇v dx dt = ε Q Z h i 1 + (ϕ0 − ϕε )σ(u) ∇ϕε − ∇ϕ · ∇v dx dt = G1 + G2 ε Q w Using (4.14) and the fact that ∇(ϕε − ϕ)/ε * ∇ψ2 in L2 (Q), one can show Z G2 → (ϕ0 − ϕ)σ(u)∇ψ2 · ∇v dx dt as ε → 0. (4.16) Q To illustrate terms from G1 , observe that σ(uε ) − σ(u) s 0 → σ (u)ψ1 in L2 (0, T ; Ls (Ω)) as ε → 0. ε Indeed, since N = 2 we have H01 (Ω) ⊂⊂ Lq (Ω) for any q ∈ [1, ∞), L2 (0, T ; H01 (Ω)) ⊂ L2 (0, T ; Ls (Ω)), and uε − u ε L2 (0,T ;Ls (Ω)) ≤c uε − u ε L2 (0,T ;H01 (Ω)) ≤ C. (4.17) Therefore, uεt − ut ≤ C, (4.18) L2 (0,T ;H −1 (Ω)) ε where (4.18) can be derived from the PDE. Since 1s + 1r + 12 = 1 and r > 2, 2r we have that s = r−2 > 2. If we define W̃ := {v : v ∈ L2 (0, T ; Ls (Ω)), vt ∈ 2 −1 L (0, T ; H (Ω))}, then, because H01 (Ω) ⊂⊂ Ls (Ω) ⊂ H −1 (Ω), it follows from [16] that W̃ ⊂⊂ L2 (0, T ; Ls (Ω)). Hence, the estimates (4.18) imply that on a subsequence uε − u − ψ1 L2 (0,T ;Ls (Ω)) → 0 as ε → 0, ε and therefore, σ(uε ) − σ(u) − σ 0 (u)ψ1 L2 (0,T ;Ls (Ω)) → 0 as ε → 0. ε Now, we are ready show Z σ(uε ) − σ(u) ε def G1 = (ϕ0 − ϕε ) ∇ϕ · ∇v dx dt ε Q Z → (ϕ0 − ϕ)σ 0 (u)ψ1 ∇ϕ · ∇v dx dt. (4.19) Q To show this convergence, we estimate Z Z σ(uε ) − σ(u) ε (ϕ0 − ϕε ) ∇ϕ · ∇v dx dt − (ϕ0 − ϕ)σ 0 (u)ψ1 ∇ϕ · ∇v dx dt ε Q Q 16 V. HRYNKIV EJDE-2009/83 n σ(uε ) − σ(u) o (ϕ0 − ϕ) ∇ϕε − σ 0 (u)ψ1 ∇ϕ · ∇v dx dt ε Q Z ε σ(u ) − σ(u) (ϕ − ϕε ) + − σ 0 (u)ψ1 + σ 0 (u)ψ1 ∇ϕε · ∇v dx dt ε Q Z n h σ(uε ) − σ(u) i (ϕ0 − ϕ) ≤ − σ 0 (u)ψ1 ∇ϕε · ∇v ε Q o + (ϕ0 − ϕ)σ 0 (u)ψ1 ∇(ϕε − ϕ) · ∇v dx dt Z σ(uε ) − σ(u) (ϕ − ϕε ) − σ 0 (u)ψ1 ∇ϕε · ∇v + ε Q Z ≤ + (ϕ − ϕε )σ 0 (u)ψ1 ∇ϕε · ∇v dx dt Z n σ(uε ) − σ(u) o (ϕ0 − ϕ) − σ 0 (u)ψ1 ∇ϕε · ∇v dx dt ≤ ε Q Z + (ϕ0 − ϕ)σ 0 (u)ψ1 ∇(ϕε − ϕ) · ∇v dx dt Q Z + (ϕ − ϕε ) Q Z + σ(uε ) − σ(u) − σ 0 (u)ψ1 ∇ϕε · ∇v dx dt ε (ϕ − ϕε )σ 0 (u)ψ1 ∇ϕε · ∇v dx dt := I + II + III + IV. Q We deal with the term I first. We have Z TZ σ(uε ) − σ(u) I ≤ kϕ − ϕ0 kL∞ (Q) − σ 0 (u)ψ1 · |∇ϕε | · |∇v| dx dt ε 0 Ω Z T σ(uε ) − σ(u) − σ 0 (u)ψ1 s · k∇ϕε kr · k∇vk2 dt ≤ kϕ − ϕ0 kL∞ (Q) ε 0 Z T σ(uε ) − σ(u) ≤ kϕ − ϕ0 kL∞ (Q) sup k∇ϕε kr − σ 0 (u)ψ1 s · k∇vk2 dt ε t 0 σ(uε ) − σ(u) ≤C − σ 0 (u)ψ1 L2 (0,T ;Ls (Ω)) · k∇vkL2 (Q) → 0 as ε → 0, ε by (4.19). Similarly, we obtain Z TZ II := (ϕ0 − ϕ)σ 0 (u)ψ1 ∇(ϕε − ϕ) · ∇v dx dt 0 Z ≤C Ω T kψ1 ks · k∇(ϕε − ϕ)kr · k∇vk2 dt 0 ≤ Ckψ1 kL2 (0,T ;Ls (Ω)) · k∇vkL2 (Q) · sup k∇(ϕε − ϕ)kr → 0 as ε → 0, t where we used (4.13). Analogously, it can be shown that III, IV → 0 as ε → 0. One can show that the third term on the right hand side of (4.15) satisfies Z h i 1 σ(uε )∇ϕε − σ(u)∇ϕ · ∇ϕ0 v dx dt ε Q Z Z → σ 0 (u)ψ1 ∇ϕ · ∇ϕ0 v dx dt + σ(u)∇ψ2 · ∇ϕ0 v dx dt as ε → 0. Q Q EJDE-2009/83 OPTIMAL CONTROL FOR A THERMISTOR PROBLEM 17 Finally, the equation for ϕ in (4.15) can be shown to satisfy Z h i 1 σ(uε )∇ϕε · ∇w − σ(u)∇ϕ · ∇w dx dt ε Q Z Z → σ 0 (u)ψ1 ∇ϕ · ∇w dx dt + σ(u)∇ψ2 · ∇w dx dt as ε → 0. Q Q Letting ε → 0 in (4.15), we obtain Z Z Z T (ψ1 )t , v dt + ∇ψ1 ∇v dx dt + 0 Z = (βψ1 + `u)v ds dt ΓT N Q Z 0 (ϕ0 − ϕ)σ (u)ψ1 ∇ϕ · ∇v dx dt + (ϕ0 − ϕ)σ(u)∇ψ2 · ∇v dx dt Q Z Z − ψ2 σ(u)∇ϕ · ∇v dx dt + σ 0 (u)ψ1 ∇ϕ · ∇ϕ0 v dx dt Q Q Z + σ(u)∇ψ2 · ∇ϕ0 v dx dt, Q Z (σ 0 (u)ψ1 ∇ϕ + σ(u)∇ψ2 ) · ∇w dx dt = 0 Q (4.20) Q for all v ∈ L (0, T ; H 1 (Ω)) ∩ L∞ (Q) and w ∈ L2 (0, T ; H01 (Ω)). This leads to the sensitivity system (4.2). The weak formulation for (4.2) is given by Z T Z Z − (ψ1 )t , v dt + ∇ψ1 · ∇v dx dt + (βψ1 + `u)v ds dt 0 Q ΓT N (4.21) Z Z 0 2 = σ (u)ψ1 |∇ϕ| v dx dt + 2 σ(u)∇ϕ · ∇ψ2 v dx dt, Q Q Z (σ 0 (u)ψ1 ∇ϕ + σ(u)∇ψ2 ) · ∇w dx dt = 0 (4.22) 2 Q for all v ∈ L (0, T ; H (Ω)) ∩ L∞ (Q) and w ∈ L2 (0, T ; H01 (Ω)) which can be shown to be equivalent to (4.20). 2 1 To characterize the optimal control, we need to introduce adjoint functions and ∂p + β ∗ p = 0 on ΓN × (0, T ), the adjoint operator associated with ψ1 and ψ2 . Using ∂n q = ψ2 = 0 on ∂Ω × (0, T ), p = 0 on ΓD × (0, T ), p(x, T ) = 0, and the notation L for the operator in the sensitivity system gives: Z ψ1 p q L dx dt ψ2 Q Z Z Z = p (−ψ1 )t dx dt + p ∆ψ1 dx dt + pσ 0 (u)|∇ϕ|2 ψ1 dx dt Q Q Q Z Z Z +2 pσ(u)∇ϕ · ∇ψ2 dx dt + q∇ · [σ(u)∇ψ2 ] dx dt + q∇ · [σ 0 (u)ψ1 ∇ϕ] dx dt Q Q Q Z Z Z 0 2 = ψ1 pt dx dt + ψ1 ∆p dx dt + ψ1 σ (u)|∇ϕ| p dx dt Q Q Q Z Z −2 ψ2 ∇ · [pσ(u)∇ϕ] dx dt − ψ1 σ 0 (u)∇ϕ · ∇q dx dt Q Q 18 V. HRYNKIV EJDE-2009/83 Z + ψ2 ∇ · [σ(u)∇q] dx dt p ψ1 ψ2 L∗ dx dt, q Q Z = Q where L∗ p pt + ∆p + σ 0 (u)|∇ϕ|2 p − σ 0 (u)∇ϕ · ∇q = . q ∇ · [−2pσ(u)∇ϕ + σ(u)∇q] Thus, the adjoint system is given by pt + ∆p + σ 0 (u)|∇ϕ|2 p − σ 0 (u)∇ϕ · ∇q = 1 in Ω × (0, T ), ∇ · [−2pσ(u)∇ϕ + σ(u)∇q] = 0 in Ω × (0, T ), ∂p + β ∗ p = 0 on ΓN × (0, T ), ∂n p = 0 on ΓD × (0, T ), (4.23) q = 0 on ∂Ω × (0, T ), p=0 on Ω × {T }, where the nonhomogeneous term “1” comes from differentiating the integrand of J(β) with respect to the state u. Theorem 4.2. Let kϕ0 kW 1,∞ (Q) be sufficiently small. Then, given an optimal control β ∗ ∈ UM and corresponding states u, ϕ, there exists a solution (p, q) ∈ L2 (0, T ; H01 (Ω)) × L2 (0, T ; H01 (Ω)) to the adjoint system (4.23). Furthermore, β ∗ can be explicitly characterized as: up ,λ ,M . (4.24) β ∗ (x, t) = min max − 2 Proof. The weak formulation of (4.23) is Z T Z Z hpt , vi dt − ∇p · ∇v dx dt − β ∗ p v ds dt 0 Q ΓN ×(0,T ) Z Z Z 0 2 0 + σ (u)|∇ϕ| p v dx dt − σ (u)(∇ϕ · ∇q)v dx dt = v dx dt, Q Q Q Z Z 2 pσ(u)∇ϕ · ∇w dx dt − σ(u)∇q · ∇w dx dt = 0 Q 2 (4.25) Q 1 ∞ for all v ∈ L (0, T ; H (Ω)) ∩ L (Q) and w ∈ L2 (0, T ; H01 (Ω)). We show that the solution to the adjoint system (4.23) exists. Define the map F : L2 (0, T ; H 1 (Ω)) × L2 (0, T ; H01 (Ω)) 7→ L2 (0, T ; H 1 (Ω)) × L2 (0, T ; H01 (Ω)) with F (V, W ) = (p, q) as follows: pt + ∆p + σ 0 (u)|∇ϕ|2 V − σ 0 (u)∇ϕ · ∇W = 1 ∇ · [−2V σ(u)∇ϕ + σ(u)∇q] = 0 in Ω × (0, T ), ∂p + β ∗ p = 0 on ΓN × (0, T ), ∂n p = 0 on ΓD × (0, T ), q = 0 on ∂Ω × (0, T ), p=0 on Ω × {T }, in Ω × (0, T ), (4.26) EJDE-2009/83 OPTIMAL CONTROL FOR A THERMISTOR PROBLEM The weak formulation of (4.26) is Z T Z Z hpt , Θi dt − ∇p · ∇Θ dx dt − 0 Z β ∗ pΘ ds dt ΓT N Q 0 19 Z 2 Z 0 σ (u)|∇ϕ| V Θ dx dt − σ (u)(∇ϕ · ∇W )Θ dx dt = Θ dx dt, Q Q Z Z 2 V σ(u)∇ϕ · ∇Ψ dx dt = σ(u)∇q · ∇Ψ dx dt + Q Q 2 Q 1 ∞ for all Θ ∈ L (0, T ; H (Ω)) ∩ L (Q) and Ψ ∈ L2 (0, T ; H01 (Ω)). We use the Banach fixed point theorem. Let F (V1 , W1 ) = (p1 , q1 ) and F (V2 , W2 ) = (p2 , q2 ). We show the contraction property (4.27) kp1 − p2 kH 1 + kq1 − q2 kH01 ≤ δ kV1 − V2 kH 1 + kW1 − W2 kH01 for some 0 < δ < 1. Indeed, for F (V1 , W1 ) = (p1 , q1 ), we have Z T Z Z β ∗ p1 Θ ds dt h(p1 )t , Θi dt − ∇p1 · ∇Θ dx dt − 0 Z 0 + Q ΓT N Q Z 2 Z 0 σ (u)|∇ϕ| V1 Θ dx dt − σ (u)(∇ϕ · ∇W1 )Θ dx dt = Θ dx dt, Q Q Z Z 2 V1 σ(u)∇ϕ · ∇Ψ dx dt = σ(u)∇q1 · ∇Ψ dx dt, Q Q and similarly, for F (V2 , W2 ) = (p2 , q2 ). Let us denote p̄ := p1 − p2 , q̄ := q1 − q2 , W̄ := W1 − W2 , and V̄ := V1 − V2 . Take Θ = p̄, then equations for p1 and p2 give Z T Z Z h(p̄)t , p̄i dt + σ 0 (u)|∇ϕ|2 V̄ p̄ dx dt − σ 0 (u)∇ϕ · ∇W̄ p̄ dx dt 0 Q Q Z Z 2 ∗ 2 = |∇p̄| dx dt + β p̄ ds dt. ΓT N Q Thus, we obtain Z Z Z 1 p̄2 dx + |∇p̄|2 dx dt + λ p̄2 ds dt 2 Ω×{0} Q ΓT N Z Z 0 2 0 ≤ σ (u)|∇ϕ| V̄ p̄ dx dt + σ (u)∇ϕ∇W̄ p̄ dx dt . Q Q We have Z σ 0 (u)|∇ϕ|2 V̄ p̄ dx dt ≤ C sup k|∇ϕ|2 kr Q t Z T kV̄ ks · kp̄k2 dt 0 ≤ Cϕ kV̄ kL2 (0,T ;Ls (Ω)) kp̄kL2 (Q) ≤ Cϕ kV̄ kL2 (0,T ;Ls (Ω)) kp̄kL2 (0,T ;H 1 (Ω)) , and Z Q σ 0 (u)∇ϕ · ∇W̄ p̄ dx dt ≤ C̃ϕ k∇W̄ kL2 (Q) kp̄kL2 (0,T ;H 1 (Ω)) . 20 V. HRYNKIV EJDE-2009/83 The above three inequalities give kp̄kL2 (0,T ;H 1 (Ω)) ≤ Cϕ kV̄ kL2 (0,T ;H 1 (Ω)) + C̃ϕ kW̄ kL2 (0,T ;H 1 (Ω)) , (4.28) where the constants Cϕ and C̃ϕ depend on the boundary data ϕ0 . A similar estimate can be derived for q̄ kq̄kL2 (0,T ;H 1 (Ω)) ≤ Cϕ0 kV̄ kL2 (0,T ;H 1 (Ω)) . (4.29) Adding (4.28) and (4.29) we get kp̄kL2 (0,T ;H 1 (Ω)) + kq̄kL2 (0,T ;H 1 (Ω)) ≤ cϕ (kV̄ kL2 (0,T ;H 1 (Ω)) + kW̄ kL2 (0,T ;H 1 (Ω)) ). Now, choosing ϕ0 so that kϕ0 kW 1,∞ (Q) is small, we obtain δ < 1. This proves (4.27), the contraction property, which gives the desired fixed point of the map and therefore the existence of solutions to the adjoint system. Recall that for variation ` ∈ L∞ (ΓTN ), with β ∗ + ε` ∈ UM for ε small, the weak formulation of the sensitivity system (4.2) is given by equations (4.21) and (4.22). Now we are ready to characterize the optimal control. Since the minimum of J is achieved at β ∗ and for small ε > 0, β ∗ +ε` ∈ UM , and denoting uε = u(β ∗ +ε`), ϕε = ϕ(β ∗ + ε`), we obtain J(β ∗ + ε`) − J(β ∗ ) 0 ≤ lim ε ε→0+ Z Z = ψ1 dx dt + 2β ∗ ` ds dt ΓT Q N Z 1 ψ1 ψ 2 = dx dt + 2β ∗ ` ds dt 0 T Q ΓN Z Z Z = − h(ψ1 )t , pi dt − ∇p∇ψ1 dx dt − Z Q Z 0 β ∗ p ψ1 ds dt ΓT N Q Z ψ1 σ 0 (u)∇ϕ · ∇q dx dt Z Z Z +2 pσ(u)∇ϕ · ∇ψ2 dx dt − σ(u)∇q · ∇ψ2 dx dt + 2 ψ1 σ (u)|∇ϕ| p dx dt − + Q Q Q = n Z − Z ∇p · ∇ψ1 dx dt − 2β ∗ ` ds dt ΓT N Q β ∗ p ψ1 ds dt ΓT N Q Z Z o ψ1 σ 0 (u)|∇ϕ|2 p dx dt + 2 pσ(u)∇ϕ · ∇ψ2 dx dt Q Q Z h Z i Z 0 + − ψ1 σ (u)∇ϕ · ∇q dx dt − σ(u)∇q · ∇ψ2 dx dt + + Q Z = Q Z ` up ds dt + ΓT N 2β ∗ ` ds dt ΓT N 2β ∗ ` ds dt, ΓT N where we integrated by parts and used the sensitivity system (4.21), (4.22) with test functions p, q. Thus, we obtain Z `(2β ∗ + up) ds ≥ 0. (4.30) ΓT N We use this inequality with three cases to characterize β ∗ : EJDE-2009/83 OPTIMAL CONTROL FOR A THERMISTOR PROBLEM 21 (i) Take the variation ` to have support on the set {x ∈ ΓTN : λ < β ∗ (x, t) < M }. The variation `(x, t) can be of any sign, therefore we obtain 2β ∗ + up = 0 whence β ∗ = − up 2 . (ii) On the set {(x, t) ∈ ΓN × (0, T ) : β ∗ (x, t) = M }, the variation must satisfy `(x, t) ≤ 0 and therefore we get 2β ∗ + up ≤ 0 implying M = β ∗ (x, t) ≤ − up 2 . (iii) On the set {(x, t) ∈ Γ × (0, T ) : β ∗ (x, t) = λ}, the variation must satisfy `(x, t) ≥ 0. This implies 2β ∗ + up ≤ 0 and hence λ = β ∗ (x) ≥ − up 2 . Combining cases (i), (ii), and (iii) gives us the explicit characterization (4.24) of the optimal control β. Substituting (4.24) into the state system (1.1) and the adjoint equations (4.23) we obtain the optimality system: ut − ∆u − σ(u)|∇ϕ|2 = 0 in Ω × (0, T ), ∇ · (σ(u)∇ϕ) = 0 0 in Ω × (0, T ), 0 2 pt + ∆p + σ (u)|∇ϕ| p − σ (u)∇ϕ · ∇q = 1 in Ω × (0, T ), ∂p ∂n ∂u ∂n ∇ · [−2pσ(u)∇ϕ + σ(u)∇q] = 0 in Ω × (0, T ), up + min max − , λ , M p = 0 on ΓN × (0, T ), 2 up + min max − , λ , M u = 0 on ΓN × (0, T ), 2 u = p = 0 on ΓD × (0, T ), u = u0 ϕ = ϕ0 on Ω × {0}, on ∂Ω × (0, T ), q = 0 on ∂Ω × (0, T ), p=0 on Ω × {T }, Note that existence of solution to the above optimality system follows from the existence of solution to the state system (1.1) and Theorem 4.2. References [1] S. N. Antontsev and M. Chipot, The thermistor problem: existence, smoothness, uniqueness, and blow up, SIAM J. Math. Anal., 25 (1994), pp. 1128–1156. [2] G. Cimatti, Existence of weak solutions for the nonstationary problem of the Joule heating of a conductor, Ann. Mat. Pura Appl. (4), 162 (1992), pp. 33–42. [3] G. Cimatti and G. Prodi, Existence results for a nonlinear elliptic system modeling a temperature dependent electrical resistor, Ann. Mat. Pura Appl. (4), 63 (1989), pp. 227–236. [4] G. Cimatti, A bound for the temperature in the thermistor’s problem, IMA J. Appl. Math., 40 (1988), pp. 15–22. [5] G. Cimatti, Optimal control for the thermistor problem with a current limiting device, IMA J. Math. Control Inform., 24 (2007), pp. 339–345. [6] G. Cimatti, Existence of periodic solutionsfor a thermally dependent resistor in a R-C circuit, Int. J. Engng Sci., vol. 32, no. 1 (1994), pp. 69–78. [7] A. C. Fowler, I. Frigaard, and S. D. Howison, Temperature surges in current-limiting circuit devices, SIAM J. Appl. Math., 52 (1992), pp. 998–1011. [8] D. Gilbarg and N. S. Trudinger, Elliptic partial differential equations of second order, 2nd edition, Springer-Verlag, Berlin, 1983. [9] S. D. Howison, J. F. Rodrigues, and M. Shillor, Stationary solutions to the thermistor problem, J. Math. Anal. Appl., 174 (1993), pp. 573–588. [10] V. Hrynkiv, S. Lenhart, and V. Protopopescu, Optimal control of a convective boundary condition in a thermistor problem, SIAM J. Control Optim., 47 (2008), pp. 20-39 22 V. HRYNKIV EJDE-2009/83 [11] K. Kwok, Complete guide to semiconductor devices, McGraw-Hill, New York, 1995. [12] H-C. Lee and T. Shilkin, Analysis of optimal control problem for the two-dimensional thermistor system, SIAM J. Control Optim., 44 (2005), pp. 268–282. [13] E. D. Maclen, Thermistors, Electrochemical Publications, Glasgow, 1979. [14] J. F. Rodriguez, Obstacle problems in mathematical physics, North-Holland Math Studies #134, Elsevier Science Publishers B.V., Amsterdam, 1987. [15] J. F. Rodriguez, A nonlinear parabolic system arising in thermomechanics and in thermomagnetism, Math. Models Methods Appl. Sci., vol. 2, no. 3, (1992), pp. 271-281. [16] J. Simon, Compact sets in the space Lp (0, T ; B), Ann. Mat. Pura Appl. (4) , 146 (1987), pp. 65–96. [17] P. Shi, M. Shillor, and X. Xu, Existence of a solution to the Stefan problem with Joule’s heating, J. Differential Equations, 105 (1993), pp. 239–263. [18] X. Xu, Local regularity theorems for the stationary thermistor problem, Proc. Roy. Soc. Edinburgh, 134A (2004), pp. 773–782. [19] X. Xu, Exponential integrability of temperature in the thermistor problem, Differential Integral Equations, 17 (2004), pp. 571–582. [20] X. Xu, Local and global existence of continuous temperature in the electrical heating of concuctors, Houston Journal of Mathematics, 22 (1996), pp. 435–455. [21] X. Xu, A degenerate Stefan-like problem with Joule’s heating, SIAM J. Math. Anal., 23 (1992), pp. 1417–1438. [22] G. Yuan, Regularity of solutions of the nonstationary thermistor problem, Appl. Anal., 53 (1994), pp. 149–164. [23] S. Zhou and D. R. Westbrook, Numerical solution of the thermistor equations, J. Comput. Appl. Math., 79 (1997), pp. 101–118. Volodymyr Hrynkiv Department of Computer and Mathematical Sciences, University of Houston-Downtown, One Main St, Houston, TX 77002, USA E-mail address: HrynkivV@uhd.edu