A Consecutive Quasilinearization Method for the Optimal Boundary Control of Semilinear Parabolic Equations

Optimal boundary control of semilinear parabolic equations requires efficient solution methods in applications. Solution methods bypass the nonlinearity in different approaches. One approach can be quasilinearization (QL) but its applicability is locally in time. Nonetheless, consecutive applications of it can form a new method which is applicable globally in time. Dividing the control problem equivalently into many finite consecutive control subproblems they can be solved consecutively by a QL method. The proposed QL method for each subproblem constructs an infinite sequence of linear-quadratic optimal boundary control problems. These problems have solutions which converge to any optimal solutions of the subproblem. This implies the uniqueness of optimal solution to the subproblem. By merging solutions to the subproblems, the solution of original control problem is obtained and its uniqueness is concluded. This uniqueness result is new. The proposed consecutive quasilinearization method is numerically stable with convergence order at least linear. Its consecutive feature prevents large scale computations and increases machine applicability. Its applicability for globalization of locally convergent methods makes it attractive for designing fast hybrid solution methods with global convergence.


Introduction
The solution methods for the optimal control of nonlinear systems pass from nonlinearity to linearity in different * Corresponding author.
approaches.For example the gradient methods modify iteratively the previous approximate solution by linearly seeking a suitable direction thorough solving a linear problem [1].The SQP methods seek the optimal solution by linearizing the optimality systems using some version of Newton's method [1] [2].Our approach in this respect is to linearize the state equation through a quasilinearization method.
Quasilinearization method for nonlinear equations has its origin in the theory of dynamic programming and has important features in common with Newton's method especially its form [3]. For a formal explanation let Y and Z be ordered Banach spaces, :  In the convex case N , Equation (1.1) can be written as where the right hand side is quasilinear.Then starting from 0 y Y ∈ thorough the quasilinearization method, a sequence of linear equations is defined , L y N y N y y y which produces the sequence of approximate solutions { } n y in Y , converging to y , the solution of (1.2) or (1.1); see [3] [4].This method has the following features.1) { } n y is monotonic.This stems from positivity and inverse positivity of L and 1 L − . 2)the convergence is globally in the sense that 0 y can be any lower solution of (1.1), i.e.
3) The rate of convergence is quadratic.For details on these features refer to [3] [5].There are some extensions, refinements and generalizations to the quasilinearization method which preserve the above features but relax the convexity assumption on N ; for a complete survey see [5]- [7].Quasilinearization method has intimate connection with the theory of positive and monotone operators, maximum operation and differential inequalities; confer [8], Sec.4.33; [4] [9].
In order to introduce the proposed consecutive quasilinearization method for optimal control problems let U be a Banach space,

J y t u t Ly t N v t N v t y t v t By t u t u U
Starting from 0 y , let ( ) be the optimal solution of (1.5) with 1 n v y − = . Then the sequence ( ) { } , n n y u converges to the solution of (1.4) with the following features: 1) The convergence is occurred for 1 < T T , for some 1 > 0 T .2) The convergence is globally in the sense that 0 y can be chosen any element in a subspace of Y .3) The rate of convergence is at least linear but it is not necessarily super-linear or quadratic.Here the sequence { } n y or { } n u is not necessarily monotonic even when N is convex or concave.For the case 1 > T T the optimal control problem is decomposed into many finite optimal control subproblems each on a time interval with length less than some 2 T and then the above method be applied to each of them consecutively.Here 2 1 < T T is such that the stability is preserved.The optimal boundary control problem which is investigated has the standard quadratic objective of tracking type and a state constraint comprised of a semilinear parabolic equation with mixed boundary type.For such control problems, due to lack of convexity of the solution set, there is no general uniqueness result based on the optimality theory of optimal control problems [1] [2] [10].However, a uniqueness result for such problems is obtained here as a by-product of the convergence of proposed consecutive quasilinearization method.
The organization of paper is as follows.Section 2 introduces the state equation and some estimates concerning solution of linear initial-boundary value problems.Section 3 proves the existence of an optimal solution.Section 4 introduces the quasilinearization method and proves its convergence for 1 < T T .Section 5 explains how to apply the quasilinearization method consecutively to the optimal boundary control problem when 1 > T T .Also the uniqueness of optimal solution is stated there.In Section 6 the error and stability analysis of consecutive quasilinearization method is investigated.Section 7 presents a numerical example concerning the obtained results.

The State Equation
Let × ∂Ω .Consider the control system described by the semilinear parabolic initial-boundary value problem: , , , f t x y t x is the system nonlinearity and is the normal derivative of y associated with A wherein ( ) is the outward unit normal to Ω ∂ .The following assumptions are imposed on the system and data: • (A1) A is a secod order differential operator in divergence form: , : , , where ( ) ∈ Ω and A is uniformly elliptic, i.e. for every ( ) , ,   satisfies Caratheodory's condition, i.e. f is measurable on Q and continuous on  , and the Nemytskii's operator : , , , , , , , is bounded and continuous.A sufficient condition for that is for some ( ) , , , a.e. , , , , , , a.e. , , , ( ) Ω are used in the paper.Identifying H with its dual H ′ results in the evolution triples , where the embeddings are dense, continuous and compact.The standard solution space of parabolic problems and its norm is defined as 0, ; 0, ; : 0, ; 0, ; , : .
The continuous embeddings and are well-known and the latter is compact.For detailed definitions and properties of the above spaces refer to [11]- [14].
The bilinear form associated with Equation (1.6) is defined as follows: [ ] ( ) where .
By Assumption (A1) the coefficients in (0.12) are bounded for a.e.t .Thus [ ] [ ] be the duality pairing between V and its dual V ′ , and ( ) be the inner product of ( ) [ ] Norm estimates concerning solution of problem (1.11) are common in the literature of linear initial-boundary value problems [12]- [14].Next theorem states some of them clarifying the time dependency quality of their constants.Its proof has been included due to lack of suitable reference, on the best of our knowledge, for the form stated here.
To obtain estimates (1.14) and (1.15) let y be the weak solution of (1.11).Then Definition 2 with ( ) where the equation is proved in Ch.III, Proposition 2.1 [11] and the inequalities are obtained by Cauchy's inequality.The continuous embedding yields

y t t y t y t B t y t y t y t y t y t h t y t g t
for a.e.

By employing the inequality
The estimate (1.15) is a consequence of (1.20) with ( ) ( ) The last assertion of Theorem is proved in Proposition 3.3 of [10].
We also meet the backward form of problem (1.11), i.e. the linear final-boundary value problem, , , ( )

Proof 2
The substitution ( ) ( ) in (0.21) yields the following equivalent problem to the problem (0.11) in the forward form , the assertions of theorem and the estimates (1.22)-(1.24)are verified.Thus, the existence of an optimal solution ( )

The Optimality System
for the problem (1.26) can be deduced from Theorem 1.45 of [1].
Proof 5 As ad U is bounded, utilizing Theorem 2.1 (or Theorem 3.1 in [10]) it is deduced ( ) . Lemma 1 The optimality condition (0.29) can be written in the equivalent form bellow: for a.e. ( ) , t x ∈ Σ .Proof 6 Refer to Lemma 1.12 in [1].( i u 's and i p 's do not necessarily satisfy an optimality system).
Proof 7 Let ( ) , t x ∈ Σ and ( ) , satisfy (0.30) at ( ) , t x .Then one of the three cases below occurs for ( ) Similarly one of the three such cases occurs for ( ) , u p at ( ) , t x .Let the first case be occurred for ( ) , u p at ( ) , t x .Then one of the three cases below must be considered for

a t x a t x u t x u t x p t x p t x a t x u t x u t x u t x p t x p t x a t x b t x u t x u t x p t x p t x
As you see each of the three cases above satisfies (0.31) at ( ) , t x .In a similar argument for each of the two other cases of ( ) , u p at ( ) , t x , three relations as the above can be written proving that each of them satisfy (0.31) at ( )
Referring to Theorem 4, p belongs to ( ) Consequently employing Assumption (A3) and the mean value theorem, it is obtained .
Owing to Corollary 1 and the continuous embeddings .
Now combining (1.41), (1.45) and (1.46) results in , where the second inequality is obtained using the mean value theorem.
where n Y , .
Next employing the estimates (1.45) and (1.46) result in T being determined in Theorem 7. Then (1.50) is obtained from (1.52) by repeatedly employing the estimate (1.47).
Corollary 3 Suppose in the quasilinearization method in Theorem 7 instead of the accurate initial value 0 y the approximate initial value 0 y +  is used.Let 2 T be as in Corollary 2. Then for  .
Next employing the estimate (1.45) and (1.46) result in where ( ) ( ) ( ) Now in order to conclude (0.54) we need an estimate like (1.47).(1.47) is for the case ( ) ( ) Then employing the estimates (1.45) and (1.46) result in T being determined after (1.48).Referring to (1.48), without loss of generality, it is considered Employing repeatedly (0.57) yields where the last inequality is obtained from (1.59)

Application to the Optimal Boundary Control Problems and the Uniqueness
The proposed quasilinearization method in Theorem 7 is convergent on the time intervals [ ] T being determined in Theorem 7. In order to apply the quasilinearization method to the optimal control problem (1.26) up to an arbitrary final time T it is possible to decompose the problem into many finite optimal control problems each on an interval with length less than 1 T .In order to follow such an approach let 2 1 T T ≤ 1 1 In order to preserve the stability, T2 is chosen as in Corollary 2 (also confer Section 6). , , ; .
, ; with the norm identity , and replacing X by V yields that W be normisomorphic to the closed subspace with the norm identity =  , satisfy consecutively the following initial-boundary value problems and vice versa: : . Consequently, the optimal control problem (1.26) is equivalent to the consecutive optimal control subproblems ( , satisfies 1.60 , min 2 2 . Therefore, solving the optimal control problem (1.26) is equivalent to consecutively solving the optimal control subproblems (1.61).Furthermore, the proposed quasilinearization method in Theorem 7 is applicable to each optimal control subproblem in (1.61).In fact the substitution whereby the quasilinearization method will be applicable to it.Moreover, as a consequence of Theorem 7 the solution of optimality system of i -th subproblem in (1.61) is unique.Thus, in view of Theorem's 4 and 5, each subproblem in (1.61) has a unique optimal solution.Consequently by the equivalence between problem (1.26) and consecutive subproblems (1.61) it can be stated Theorem 8 Optimal boundary control problem (1.26) under Assumptions (A1)-(A3) has unique optimal boundary control solution and optimal state solution.
Note that the uniqueness could not be established thorough the optimality theory of optimal control problems which was used for stating the existence in Section 3.This is due to lack of convexity of the solution set of problem (1.26).
An issue concerning the above consecutive process is the relation between ( ) , , y p u , the solution of optimality system of problem (1.26), and ( ) , , i i i y p u , the solution of optimality system of i -th subproblem in (1.61)   ( ) is not necessarily zero; confer Also it is not possible in general to construct p from i p 's; however, after obtaining i y 's, p can be computed from (1.28).

Error Analysis
By the consecutive quasilinearization method in Section 5, the optimal control problem (1.26) is solved through m consecutive optimal control subproblems (1.61).Each subproblem is solved by the quasilinearization method in Theorem 7 which is an iterative method with infinite iterations.In applications it is implemented up to a finite iterations, thereby producing error.Consequently, during solving each subproblem there exists an error production and an error propagation.
Y t presents the accumulated error consists of the production errors and the propagation errors in the consecutive implementation of m quasilinearization method, when the implementation is up to N iteration on each subproblem.In the estimate (1.67) the term C T decrease when T decrease (or m increase).Consequently it may a trade off be necessary between size of m (the number of subproblems) and N (the number of required iterations in the implementation of quasilinearization method) in order to have the desired total error in the consecutive quasilinearization method.

Numerical Example
A typical example is presented reflecting the obtained results in the previous sections in applications.Consider the optimal control problem (1.26) with the following data: without discretization of time.The linear optimality systems (1.62)-(1.64)are solved by the semismooth Newton's method [16] or Section 2.5 in [1], and the implementation is done with MATLAB software.is in some sense the step size of time discretization, its increment yields more accurate approximation to the optimal objective value.

Conclusions
A consecutive quasilinearization method was proposed for the optimal boundary control problems with quadratic objective of tracking type and a semilinear parabolic equation with mixed boundary as the state constraint; cf.(1.26) and (1.32).The proposed method divides the control problem equivalently into many finite consecutive subproblems through partitioning the time interval into subintervals; cf.Section 5 and (1.61).Then subproblems are solved consecutively by a quasilinearization method (hence the name of proposed method).Finally the optimal solution of control problem is obtained by consecutively merging optimal solutions of subproblems.The quasilinearization method for each subproblem constructs an infinite sequence of linear-quadratic optimal boundary control problems of form (1.38).The sequence of solutions to the optimality systems of these linear problems converges to any solutions of the optimality system of subproblem; confer Theorem 7 and Section 5.This implies the uniqueness of solution to the optimality system of a subproblem, hence the uniqueness of optimal solution to the original control problem; confer Theorem 8.This uniqueness result is new, on the best of our knowledge, in the class of optimal control problems with state constraint of semilinear parabolic equation type.
The convergence of quasilinearization method for each subproblem depends on the time interval length of the subproblem, 2 T , and there is a bound on 2 T which the convergence occurs, 2 1 T T ≤ , 1 T being determined in Theorem 7. In comparison with methods which require the fully discretization of original control problem, cf.Chapter 2 in [1], [2] and [17], 2 T can be considered as the time discretization step length.In this view the con- secutive feature of proposed method replaces the large scale computations in fully discrete methods by the consecutive small scale computations in the subproblems, hence increasing the machine applicability of method.Specially in quasilinearization method in solving the sequence of linear-quadratic control problems the time discretization can be avoided by choosing 2 T enough small , cf.Section 7. In comparison with superlinear methods which are locally convergent, as different versions of Newton's method and/or Lagrange-SQP methods (Chapter 2 in [1], and [2]), the consecutive quasilinearization method is globally convergent and its convergence order is at least linear, cf.Theorem 7.For example Table 1 of Section 7 presents a cubic convergence rate.Thereby the consecutive quasilinearization method is very suitable for the globalization of locally convergent methods by applying it to find a starting solution for those methods.
The quasilinearization method for subproblems has infinite iterations, but in applications it is implemented up to a finite iteration.Therefore its consecutive application on the subproblems produces and propagates errors.However choosing 1 2

T T ≤
guarantees the numerical stability, cf.Section 6.The imposed boundedness assumptions on the nonlinearity of problem and the admissible controls are necessary for the convergence proof, cf.Assumption (A3), Section 3 and proof of Theorem 7. As the investigated control problem here also has optimal solution with much weaker boundedness assumptions, cf.[10], application of consecutive quasilinearization method in this case requires new convergence proof.

:
linear boundary operator.Consider the following optimal boundary control problem: t belongs to a time interval [ ] 0,T .For v Y ∈ consider the following approximation to (1.4):

10 )
By Assumption (A1) the coefficients in (0.10) are bounded.This results in the boundedness of [ ] bounded when T varies boundedly and bounded when T varies boundedly and 25) satisfies all the assumptions which problem (1.11) satisfies.Therefore the assertions of Theorem 2 and the estimates (1.13)-(1.15)are valid for w .Since
34)By Theorem 5 and Theorem 4 optimality system (1.32)-(1.34)has at least one solution.of optimality system (0.32)-(0.34).Then there exists a elements are the unique solution of the following linear optimality systems and there exists 1 > 0 T such that this sequence converges, at least linearly, to ( ) optimality system (1.32)-(1.34)has a unique solution.

3 C 9
T are as in(1.39),(1.48)and (1.53), respectively.Proof The proof follows the lines of proof of Theorem 7. As n Y satisfies (0.40) the estimate (0.15) in Theorem 2 yields

5 C
T are as in(1.48),(1.53) and (1.59), respectively.Proof 10 The proof follows the lines of proof of Theorem 7. As n Y satisfies (1.40) in Q with be the solution provided by the quasilinearization method at iteration n for the i-th optimality system, i.e. one which satisfies (1.35)-(1.37) on i Q .For the first subproblem the quasi- linearization method starts with the accurate initial value 48) and thereafter.Since m is fixed, by increasing the number of iterations N, the total accumulated error in H. Therefore, the proposed consecutive quasilinearization method in Section 5 is stable.Furthermore,

1 a = − , 1 b
= .Setting 2 1 T m = , the consecutive quasilinearization method is implemented on the m consecutive subproblems (1.61) with the optimality systems (1.62)-(1.64).The corresponding states n i y , costates n i p and controls n i u are approximated by the elements and boundary elements of continuous linear finite element spaces on [ ] an estimate is obtained following the lines which (1.47) obtained.As n Y satisfies (1.40) with . ( )

Table 1
presents the values of Table2presents the optimal objective values of problem when the consecutive quasilinearization method is implemented with different number of subproblems but fixed number of iterations in each quasilinearization These values present at least a linear rate of convergence in the quasilinearization method as it was deduced from (1.45)-(1.47).

Table 1 .
The difference between iterations in the quasilineariztion method for the forth subproblem at t = t 4 when m = 15 and the number of iterations is N = 10.

Table 2 .
The optimal objective values with different number of subproblems, m, but fixed number of iterations in the quasilinearization method, i.e.N = 5.
method, i.e. with different m's and fixed N. As 2 1 T m =