In this paper, we consider the fractional optimal control problem with the terminal and running state constraints. The fractional calculus of derivatives and integrals can be viewed as generalizations of their classical notions to any arbitrary real order. In our problem setup, the dynamical system (or state equation) is captured by the fractional differential equation in the sense of (left) Caputo with order α∈(0,1), and the objective functional is formulated by the Bolza form expressed as the left Riemann-Liouville fractional integral. In addition, there are terminal and running state constraints; while the former is described by initial and final states within a convex set, the latter is given by an explicit instantaneous inequality state constraint. We obtain the Pontryagin maximum principle for the problem of this paper. The proof is based on an application of the Ekeland variational principle and the spike variation, by which we develop fractional variational and duality analysis using fractional calculus and functional analysis techniques, together with the representation results on (RL and Caputo) linear fractional differential equations. In fact, due to the inherent complex nature of the fractional control problem and the presence of the terminal and running state constraints, our maximum principle is new in the optimal control problem, context and its detailed proof must be different from that of the existing literature. As an application, we consider the linear-quadratic fractional optimal control problem with terminal and running state constraints, for which the optimal solution is obtained using the maximum principle of this paper.
Citation: Jun Moon. The Pontryagin type maximum principle for Caputo fractional optimal control problems with terminal and running state constraints[J]. AIMS Mathematics, 2025, 10(1): 884-920. doi: 10.3934/math.2025042
[1] | Yuna Oh, Jun Moon . The infinite-dimensional Pontryagin maximum principle for optimal control problems of fractional evolution equations with endpoint state constraints. AIMS Mathematics, 2024, 9(3): 6109-6144. doi: 10.3934/math.2024299 |
[2] | Jun Moon . A Pontryagin maximum principle for terminal state-constrained optimal control problems of Volterra integral equations with singular kernels. AIMS Mathematics, 2023, 8(10): 22924-22943. doi: 10.3934/math.20231166 |
[3] | Ruiqing Shi, Yihong Zhang . Dynamic analysis and optimal control of a fractional order HIV/HTLV co-infection model with HIV-specific antibody immune response. AIMS Mathematics, 2024, 9(4): 9455-9493. doi: 10.3934/math.2024462 |
[4] | Xingxiao Wu, Lidong Huang, Shan Zhang, Wenjie Qin . Dynamics analysis and optimal control of a fractional-order lung cancer model. AIMS Mathematics, 2024, 9(12): 35759-35799. doi: 10.3934/math.20241697 |
[5] | Irmand Mikiela, Valentina Lanza, Nathalie Verdière, Damienne Provitolo . Optimal strategies to control human behaviors during a catastrophic event. AIMS Mathematics, 2022, 7(10): 18450-18466. doi: 10.3934/math.20221015 |
[6] | Regina S. Burachik, Bethany I. Caldwell, C. Yalçın Kaya . Douglas–Rachford algorithm for control- and state-constrained optimal control problems. AIMS Mathematics, 2024, 9(6): 13874-13893. doi: 10.3934/math.2024675 |
[7] | Zongmin Yue, Yitong Li, Fauzi Mohamed Yusof . Dynamic analysis and optimal control of Zika virus transmission with immigration. AIMS Mathematics, 2023, 8(9): 21893-21913. doi: 10.3934/math.20231116 |
[8] | Jie Liu, Zhaojie Zhou . Finite element approximation of time fractional optimal control problem with integral state constraint. AIMS Mathematics, 2021, 6(1): 979-997. doi: 10.3934/math.2021059 |
[9] | Asaf Khan, Gul Zaman, Roman Ullah, Nawazish Naveed . Optimal control strategies for a heroin epidemic model with age-dependent susceptibility and recovery-age. AIMS Mathematics, 2021, 6(2): 1377-1394. doi: 10.3934/math.2021086 |
[10] | Wei Ji . On the equilibrium strategy of linear-quadratic time-inconsistent control problems. AIMS Mathematics, 2025, 10(3): 5480-5494. doi: 10.3934/math.2025253 |
In this paper, we consider the fractional optimal control problem with the terminal and running state constraints. The fractional calculus of derivatives and integrals can be viewed as generalizations of their classical notions to any arbitrary real order. In our problem setup, the dynamical system (or state equation) is captured by the fractional differential equation in the sense of (left) Caputo with order α∈(0,1), and the objective functional is formulated by the Bolza form expressed as the left Riemann-Liouville fractional integral. In addition, there are terminal and running state constraints; while the former is described by initial and final states within a convex set, the latter is given by an explicit instantaneous inequality state constraint. We obtain the Pontryagin maximum principle for the problem of this paper. The proof is based on an application of the Ekeland variational principle and the spike variation, by which we develop fractional variational and duality analysis using fractional calculus and functional analysis techniques, together with the representation results on (RL and Caputo) linear fractional differential equations. In fact, due to the inherent complex nature of the fractional control problem and the presence of the terminal and running state constraints, our maximum principle is new in the optimal control problem, context and its detailed proof must be different from that of the existing literature. As an application, we consider the linear-quadratic fractional optimal control problem with terminal and running state constraints, for which the optimal solution is obtained using the maximum principle of this paper.
Fractional calculus of derivatives and integrals can be viewed as generalizations of their classocal ones to any real arbitrary order [1,2,3]. One important application of fractional calculus is a class of fractional differential equations, involving fractional derivatives (or integrals), which enables us to describe more general and extraordinary phenomena observed in real world. Indeed, various types of fractional differential equations and their applications have been studied in applied mathematics, science, engineering, and economics; see [1,2,3,4,5] and the references therein.
Along with fractional differential equations, fractional optimal control problems, i.e., optimal control for fractional differential equations in the sense of Riemann-Liouville (RL) and/or Caputo, have been studied extensively in the literature under different formulations; see [6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22] and the references therein. In particular, classical results on fractional optimal control problems were obtained in [6,7,8]. The time optimal control problem for fractional differential equations was studied in [9,10]. The existence of optimal solutions for fractional optimal control problems was studied in [11,20]. The fractional dynamic programming approach with the Hamilton-Jacobi-Bellman PDE was studied in [12,13]. The linear optimal control problem for fractional differential equations was considered in [14,23]. Applications and numerical approaches of various fractional optimal control problems were studied in [15,16,17,18,19,24].
The optimal control problems and their Pontryagin maximum principles for the fractional optimal control problems were studied in several different directions; see [20,25,26,27,28,29,30,31,32,33,34,35,36] and the references therein. Specifically, the researches in [25,26,27,28] considered the optimal control problem for RL and Caputo fractional differential equations without state constraints, which use the Euler-Lagrangian approach similar to [6,7]. In [34], the Lagrangian multiplier method was applied to obtain the necessary condition for optimality. However, it requires that the corresponding Hamiltonian is smooth, and does not know the nontriviality of the corresponding Lagrangian multipliers. Moreover, the terminal state constraint was not considered in [34]. In [29], a simple convex variation was applied to obtain the maximum principle for the RL fractional optimal control problem. The Caputo fractional optimal control problem without state constraints was studied in [30], where the sufficient condition was also obtained under the convexity assumption. In [31,32], the Caputo and Cucker-Smale multi-agent fractional control problems without state constraints were studied, where the corresponding control space was assumed to be compact. The authors in [36] studied the controllability and existence of optimal control for the unconstrained case in the infinite-dimensional case. In [35], the pseudospectral method for fractional optimal control problem was studied without the state constraints. In addition, [33] studied the singular fractional optimal control problem. Recently, the Caputo fractional optimal control problem with the terminal state constraint only was considered in [20]. However, the research in [20] did not consider the running state constraint. We also mention that the research in [22] studied the equivalent relationship between control problems for fractional differential equations and Volterra integral equations, and its infinite-dimensional case was treated in [21]. Notice that the research in [21,22] did not consider the state constraints in the corresponding maximum principles. {We note that [37,38] studied the nonfractional optimal control problems with state constraints. Recently, [39] studied the general numerical methods for the linear-quadratic fractional optimal control problem, where the Bolza-Riemann-Liouville objective functional was considered. Moreover, [39] provided several practical examples of linear-quadratic fractional optimal control problems such as the optimal drug scheduling of cancer chemotherapy problem and the minimum fuel optimal control problem. We also mention that the research in [40] considered the linear-quadratic fractional optimal control problem without state constraints, where instead of the open-loop optimal solutions, the closed-loop optimal solution was obtained in terms of the Riccati-type equation.
In this paper, we consider the fractional optimal control problem with terminal and running state constraints. The precise problem statement is given in Section 3. In our problem setup, the dynamic constraint is captured by the fractional differential equation in the sense of Caputo with order α∈(0,1), and the objective functional is formulated by the left RL fractional integral. In addition, there are terminal and running state constraints; while the former is described by initial and final states within a convex set, the latter is expressed by an explicit instantaneous inequality state constraint. Note that the general fractional optimal control problem with terminal and running state constraints has not been sufficiently studied in the existing literature, which we address in this paper. Indeed, the earlier results for fractional optimal control problems discussed above did not consider the case of both terminal and running state constraints. We also mention that the research in [39] studied the numerical approaches for solving the linear-quadratic fractional optimal control problem. Finally, in the existing literatures, some standing assumptions in the fractional optimal control problem (e.g., , the assumption of the control space being closed or convex) are different from those in our paper.
We prove the wellposedness (equivalently, the existence and uniqueness of the solution) of the Caputo fractional differential equation with order α∈(0,1) (see Theorem 3.1). Although similar versions of Theorem 3.1 are reported in the existing literature (e.g., [2,41]), we provide its different version and different proof. In particular, to prove Theorem 3.1, we first need to show the equivalent RL fractional integral representation of the Caputo fractional differential equation. Then, we use the Bielecki norm under an appropriate function space to apply the contraction mapping theorem. We also obtain several important estimates of the Caputo fractional differential equation using the fractional Gronwall's inequality.
We obtain the explicit representation results of left Caputo linear fractional differential equations with initial conditions and right RL fractional differential equations with terminal conditions, where the latter is also dependent on the variable of the Lebesgue-Stieltjes measure (see Section 4). In both cases, to obtain the desired representation formulas, a detailed analysis of the fundamental solution of the left and right RL fractional state-transition matrices has to be carried out, and the careful use of Fubini's formula is needed. While the representation result of left Caputo linear fractional differential equations was presented in ([42], Theorem 5.1), we provide its different version of the representation result in Lemma 5 using the definition of (left and right) RL fractional state-transition matrices. Furthermore, the representation result of right RL fractional differential equations with terminal conditions in Lemma 6 has not been reported in the existing literature. Note that the results in Section 4 are independent from our main result.
The main result of this paper is the Pontryagin maximum principle (see Theorem 5.1). Since the control space U is only a separable metric space and does not have any algebraic structure (see Assumption 1), we have to employ the spike perturbation technique. Due to the inherent complex nature of the fractional control problem and the presence of the terminal and running state constraints, our maximum principle and its detailed proof must be different from those of the existing literature (e.g., [20,29,31]). Specifically, in the proof (see Section 7), we need to formulate the penalized unconstrained fractional control problem, for which the Ekeland variational principle has to be applied. Then, to develop the variational and duality analysis via the spike variation, we have to use the fractional calculus and the intrinsic properties of distance functions (with functional analysis techniques), together with the fractional Gronwall's inequality, the representation results of (left and right) Caputo and RL linear fractional differential equations (see Theorem 3.1 and Section 4*), and the technical lemma (see Lemma C.1 in Appendix C†). Moreover, the proofs for the complementary slackness and the transversality condition are essentially required due to the presence of the terminal and running state constraints. We mention that our proof technique of the maximum principle for the fractional optimal control problem has not been presented in the existing literature.
*In Section 4, we prove the representation formulas of the solution to the left Caputo and right RL fractional differential equations, where the latter includes the Lebesgue-Stieltjes integral, in terms of the (left and right) RL fractional state-transition matrices. Note that the results in Section 4 have not been reported in the existing literature.
†We mention that the proof of Lemma C.1 in Appendix C requires a new technique compared with its nonfractional version in ([43], Corollary 3.8, Chapter 4).
As an application, we consider the linear-quadratic fractional optimal control problem with terminal and running state constraints (see Section 6 and Proposition 6.1). We obtain the optimal solution to the linear-quadratic problem using the maximum principle of this paper (see Theorem 5.1), and then discuss the computational method to obtain the unknown variables in the maximum principle (see Proposition 6.1 and Remark 6.1). Indeed, we mention that various numerical approaches for solving linear-quadratic fractional optimal control problems can be found in [39,44,45] and the references therein (see Remark 6.1).
Our maximum principle in Theorem 5.1 can be viewed as an extension of [20,29,31] in several different directions (see Remarks 2 and 4). In particular, the researches in [20,29,31] did not consider both terminal and running state constraints. In addition, the control spaces in [20,29,31] need to be compact or closed, whereas we assume only separability without convexity (see Assumption 3.1). Hence, the problem formulation of this paper is different from that of [20,29,31], and the new approaches and techniques have to be developed in this paper (particularly the proof of the maximum principle in Section 7 and the representation results in Section 4). Indeed, the maximum principle of this paper is new in the optimal control problem context and its proof requires to develop a new technique, both of which are not reported in the existing literature. Note also that recently, [46] obtained the maximum principle for optimal control of infinite-dimensional Caputo-type evolution equations under the end-point terminal state constraint only. Hence, the approach of [46] cannot be used to solve the problem with both terminal and running state constraints considered in our paper. We also mention that ([46], Theorem 3.1 and Remark 3.2) requires additional assumptions in the corresponding maximum principle due to the infinite-dimensional nature of the problem formulation. Finally, the example studied in this paper is different from that in [46]. Indeed, this paper considers the linear-quadratic problem with both terminal and running state constraints, whereas [46] studied the fractional heat equation with the terminal state constraint only.
The paper is organized as follows. Preliminaries on fractional calculus are given in Section 2. We formulate the problem and then provide the wellposedness result in Section 3. In Section 4, we present the representation results of (left and right) Caputo and RL linear fractional differential equations. The maximum principle of this paper is stated in Section 5. The application of the maximum principle to the state-constrained fractional linear-quadratic problem is studied in Section 6. The proof for the maximum principle is provided in Section 7. We conclude our paper in Section 8. In Appendix A, we provide the proof of the results in Section 3. In Appendix B, we give the variational analysis, which is needed in Section 7. In Appendix C, we provide the technical lemma used in Section 7.
In this section, we provide some preliminary results on fractional calculus. More detailed results on fractional calculus can be found in [1,2] and the references therein.
Let Rn be the n-dimensional Euclidean space, where ⟨x,y⟩:=x⊤y is the inner product and |x|:=⟨x,x⟩1/2 is the norm for x,y∈Rn. For A∈Rm×n, A⊤ denotes the transpose of A. Let In be an n×n identity matrix. Define 1S(⋅) by the indicator function of any set S. Let Γ be the Gamma function.
For any differentiable map f:Rn→Rl, let ∂xf:Rn→Rl×n be the partial derivative of f with respect to x∈Rn. Note that ∂xf=[∂xf⊤1⋯∂xf⊤l]⊤ with ∂xfj∈R1×n, and when l=1, ∂xf∈R1×n. For any differentiable map f:Rn×Rl→Rl, ∂xf:Rn×Rl→Rl×n for x∈Rn, and ∂yf:Rn×Rl→Rl×l for y∈Rl.
Let t0,tf∈[0,T] with t0<tf be fixed. Define the following spaces:
● Lp([t0,tf];Rn), p≥1: the space of functions ψ:[t0,tf]→Rn such that ψ is measurable and ‖ψ(⋅)‖Lp,n:=(∫tft0|ψ(t)|pRndt)1p<∞ (‖ψ(⋅)‖L∞,n:=esssupt∈[t0,tf]|ψ(t)|<∞ when p=∞);
● C([t0,tf];Rn): the space of functions ψ:[t0,tf]→Rn such that ψ is continuous and ‖ψ(⋅)‖∞,n:=supt∈[t0,tf]|ψ(t)|<∞;
● AC([t0,tf];Rn): the space of functions ψ:[t0,tf]→Rn such that ψ is absolutely continuous;
● BV([t0,tf];Rn): the space of functions ψ:[t0,tf]→Rn such that ψ is of bounded variation on [t0,tf].
The norm on BV([t0,tf];Rn) is defined by ‖ψ(⋅)‖BVn:=ψ(t0)+TV(ψ), where TV(ψ):=sup(tk)k{∑k|ψ(tk+1)−ψ(tk)|}<∞ with the supremum being taken by all partitions of [t0,tf]. Let NBV([t0,tf];Rn) be the space of functions ψ(⋅)∈BV([t0,tf];Rn) such that ψ(⋅)∈BV([t0,tf];Rn) is normalized, i.e., ψ(t0)=0 and ψ is left continuous. The norm on NBV([t0,tf];Rn) is defined by ‖ψ(⋅)‖NBVn:=TV(ψ). When ψ(⋅)∈NBV([t0,tf];R) is monotonically nondecreasing, we have ‖ψ(⋅)‖NBV:=‖ψ(⋅)‖NBV1=ψ(tf). Recall that both (BV([t0,tf];Rn),‖⋅‖BVn) and (NBV([t0,tf];Rn),‖⋅‖NBVn) are Banach spaces.
Definition 2.1 (i) For f(⋅)∈L1([t0,tf];Rn) and t∈[t0,tf], the left Riemann-Liouville (RL) fractional integral Iαt0+[f] of order α>0 is defined by Iαt0+[f](t):=∫tt0(t−s)α−1Γ(α)f(s)ds.
(ii) For f(⋅)∈L1([t0,tf];Rn) and t∈[t0,tf], the right RL fractional integral Iαtf−[f] of order α>0 is defined by Iαtf−[f](t):=∫tft(s−t)α−1Γ(α)f(s)ds.
Lemma 2.1. ([1], Lemma 2.1) For any f(⋅)∈L1([t0,tf];Rn), ‖Iαt0+[f](⋅)‖L1,n≤K‖f(⋅)‖L1,n and ‖Iαtf−[f](⋅)‖L1,n≤K‖f(⋅)‖L1,n, which implies Iαt0+[f](⋅),Iαtf−[f](⋅)∈L1([t0,tf];Rn).
Lemma 2.2. ([1], Lemma 2.3) For any f(⋅)∈L1([t0,tf];Rn) and α,β>0, the following semigroup property holds: Iαt0+[Iβt0+[f]](⋅)=Iα+βt0+[f](⋅)=Iβ+αt0+[f](⋅)=Iβt0+[Iαt0+[f]](⋅) and Iαtf−[Iβtf−[f]](⋅)=Iα+βtf−[f](⋅)=Iβ+αtf−[f](⋅)=Iβtf−[Iαtf−[f]](⋅).
Lemma 2.3. ([2], Theorem 2.6) For f(⋅)∈L∞([t0,tf];Rn), Iαt0+[f](⋅),Iαtf−[f](⋅)∈C([t0,tf];Rn).
Definition 2.2. (i) For f(⋅)∈L1([t0,tf];Rn), the left RL fractional derivative Dαt0+[f] of order α∈(0,1) is defined by Dαt0+[f](t):=ddt[I1−αt0+[f]](t), provided that I1−αt0+[f](⋅)∈AC([t0,tf];Rn). In this case, we have Dαt0+[f](⋅)∈L1([t0,tf];Rn). Let RLnDαt0+ be the set of functions f(⋅)∈L1([t0,tf];Rn) such that f admits the left RL fractional derivative Dαt0+[f] of order α∈(0,1).
(ii) For f(⋅)∈L1([t0,tf];Rn), the right RL fractional derivative Dαtf−[f] of order α∈(0,1) is defined by Dαtf−[f](t):=−ddt[I1−αtf−[f]](t), provided that I1−αtf−[f](⋅)∈AC([t0,tf];Rn). In this case, we have Dαtf−[f](⋅)∈L1([t0,tf];Rn). Let RLnDαtf− be the set of functions f(⋅)∈L1([t0,tf];Rn) such that f admits the right RL fractional derivative Dαtf−[f] of order α∈(0,1).
Remark 2.1. Note that when α=1, the notions of fractional integral and derivatives are reduced to those of usual ones [1,20] (see [1], Chapter 2).
Let Dαt0+[f(⋅)](⋅):=Dαt0+[f](⋅) and Dαtf−[f(⋅)](⋅):=Dαtf−[f](⋅), where (⋅) in the square bracket is used to emphasize the integration of f with respect to the other variable.
Definition 2.3. (i) For f(⋅)∈C([t0,tf];Rn), the left Caputo fractional derivative CDαt0+[f] of order α∈(0,1) is defined by CDαt0+[f](t):=Dαt0+[f(⋅)−f(t0)](t), where f(⋅)−f(t0)∈RLnDαt0+. In this case, we have CDαt0+[f](⋅)∈L1([t0,tf];Rn). Let CnDαt0+ be the set of functions f(⋅)∈C([t0,tf];Rn) such that f admits the left Caputo fractional derivative CDαt0+[f] of order α∈(0,1).
(ii) For f(⋅)∈C([t0,tf];Rn), the right Caputo fractional derivative CDαtf−[f] of order α∈(0,1) is defined by CDαtf−[f](t):=Dαtf−[f(⋅)−f(tf)](t), where f(⋅)−f(tf)∈RLnDαtf−. In this case, we have CDαtf−[f](⋅)∈L1([t0,tf];Rn). Let CnDαtf− be the set of functions f(⋅)∈C([t0,tf];Rn) such that f admits the right Caputo fractional derivative CDαtf−[f] of order α∈(0,1).
Consider the Rn-valued left Caputo fractional differential equation with order α∈(0,1):
{CDαt0+[X](t)=f(t,X(t),u(t)),t∈(t0,tf],X(t0)=X0∈Rn, | (3.1) |
where X(⋅)∈Rn is the state with the initial condition X0∈Rn, u:[t0,tf]→U is the control input with U being the control set, and f:[t0,tf]×Rn×U→Rn is the driver of the Caputo fractional differential equation. Let U:={u:[t0,tf]→U|uis measurable} be the space of admissible controls for (3.1). We sometimes use the notation X(⋅;X0,u):=X(⋅) to stress the dependency of (3.1) on the pair (X0,u(⋅))∈Rn×U.
Assumption 3.1. (i) (U,ρ) is a separable metric space, where ρ is the corresponding metric.
(ii) f(⋅,X,u)∈L∞([t0,tf];Rn), and it holds that supu∈U|f(t,x,u)|<∞ for (t,x)∈[t0,tf]×Rn. (X,u)↦f(t,X,u) is Lipschitz continuous, i.e., there is a constant L≥0 such that for any t∈[t0,tf] and (X,u),(X′,u′)∈Rn×U, |f(t,X,u)−f(t,X′,u′)|≤L(|X−X′|+ρ(u,u′)), and it holds that |f(t,0,u)|≤L(1+|X|).
(iii) X↦f(t,X,u) is continuously differentiable, where (t,X,u)↦∂Xf(t,X,u) is continuous, and (X,u)↦∂Xf(t,X,u) is Lipschitz continuous with the constant L≥0.
The following theorem shows the wellposedness of (3.1). Its proof can be found in Appendix A.
Theorem 3.1. Let Assumption 3.1 hold. Then, the following statements hold:
(i) The solution of (3.1), X(⋅)∈CnDαt0+, can be expressed as the left RL fractional integral form given by X(t)=X0+Iαt0+[f(⋅,X(⋅),u(⋅))](t);
(ii) For any (X0,u(⋅))∈Rn×U, (3.1) has a unique solution of X(⋅)∈CnDαt0+;
(iii) For any (X0,u(⋅)),(X′0,u′(⋅))∈Rn×U, we have
supt∈[t0,tf]|X(t;X0,u)−X′(t;X′0,u′)|≤b(tf)+∫tft0∞∑k=1(LΓ(α))kΓ(kα)(tf−s)kα−1b(s)ds, | (3.2) |
supt∈[t0,tf]|X(t;X0,u)|≤b′(tf)+∫tft0∞∑k=1(LΓ(α))kΓ(kα)(tf−s)kα−1b′(s)ds, | (3.3) |
where b(t):=|X0−X′0|+L∫tt0(t−s)α−1Γ(α)ρ(u(s),u′(s))ds and b′(t):=|X0|+L∫tt0(t−s)α−1Γ(α)ds.
The objective functional is given by the following left RL fractional integral with order β≥1:
J(X0;u(⋅))=Iβt0+[l(⋅,X(⋅),u(⋅)](tf)+m(X0,X(tf)), | (3.4) |
where by Definition 2.1,
Iβt0+[l(⋅,X(⋅),u(⋅)](tf)=∫tft0(tf−s)β−1Γ(β)l(s,X(s),u(s))ds. |
In (3.4), l:[t0,tf]×Rn×U→R is the running cost and m:Rn×Rn→R is the terminal cost. The fractional optimal control problem of this paper is formulated as follows:
(P)infu(⋅)∈UJ(X0;u(⋅)) |
subject to the terminal and running state constraints given by
{(X0,X(tf))∈F⊂R2n,Gi(t,X(t))≤0,∀t∈[t0,tf],i=1,…,q. | (3.5) |
Assumption 3.2. (i) l(⋅,X,u)∈L∞([t0,tf];R), and (X,u)↦l(t,X,u) is continuous. (X,X′)↦m(X,X′) is continuous.
(ii) X↦l(t,X,u) is continuously differentiable, where (t,X,u)↦∂Xl(t,X,u) is continuous, and (X,u)↦∂Xl(t,X,u) is Lipschitz continuous with the constant L≥0. (X,X′)↦m(X,X′) is continuously differentiable, where (X,X′)↦∂Xm(X,X′) and (X,X′)↦∂X′m(X,X′) are Lipschitz continuous with the constant L≥0.
(iii) F is a nonempty closed convex subset of R2n. (t,X)↦G(t,X):=[G1(t,X)⋯Gq(t,X)]⊤ with Gi:[t0,tf]×Rn→R, i=1,…,q, is continuous, where X↦G(t,X) is continuously differentiable with (t,X)↦∂XG(t,x) being bounded.
Remark 3.1. Assumptions similar to Assumptions 3.1 and 3.2 have been used in various optimal control problems and their maximum principles; see [21,22,43,47,48,49,50,51] and the references therein. Also, we extend [20,29,31]. In fact, the researches in [29,31] did not consider any state constraints, and [20] considered only the terminal state constraint. Furthermore, the control spaces in [20,29,31] need to be compact or closed, whereas we assume only separability without convexity (see Assumption 3.1). The fractional control problem with both terminal and running state constraints formulated in (P) has not been studied in the existing literature.
In this section, we provide the representation results of (RL and Caputo) linear fractional differential equations via RL fractional state-transition matrices. Note that the results of this section are independent from the main result of this paper in Theorem 5.1.
Definition 4.1. Let F(⋅)∈L∞([t0,tf];Rn×n). The left RL fractional state-transition matrix Π(⋅,τ)∈RLn×nDατ+ with t0≤τ≤t≤tf is defined by
Π(t,τ)=(t−τ)α−1Γ(α)In+∫tτ(t−r)α−1Γ(α)F(r)Π(r,τ)dr, | (4.1) |
which is the solution to the following left RL fractional differential equation:
{Dατ+[Π](t,τ)=F(t)Π(t,τ),t∈(τ,tf],I1−ατ+[Π](τ,τ)=In. |
Lemma 4.1. The left RL fractional state-transition matrix given in (4.1) can be written as the following backward fractional integration form:
Π(t,τ)=(t−τ)α−1Γ(α)In+∫tτ(r−τ)α−1Γ(α)Π(t,r)F(r)dr. | (4.2) |
Proof. By the right-hand-side of (4.2), let
ˆΠ(t,τ)=(t−τ)α−1Γ(α)In+∫tτ(r−τ)α−1Γ(α)Π(t,r)F(r)dr. | (4.3) |
Then, we need to show the equivalence of (4.1) and (4.3). Hence, we have to prove the following result:
∫tτ(t−r)α−1Γ(α)F(r)ˆΠ(r,τ)dr=∫tτ(r−τ)α−1Γ(α)Π(t,r)F(r)dr. | (4.4) |
Clearly, (4.4) holds when t=τ. Then, (4.4) follows by Fubini's formula ([52], Theorem 3.3, Chapter 6), since we can show that
∫tτ(t−r)α−1Γ(α)F(r)ˆΠ(r,τ)dr=∫tτ(t−r)α−1Γ(α)F(r)(r−τ)α−1Γ(α)Indr+∫tτ(t−r)α−1Γ(α)F(r)∫rτ(ν−τ)α−1Γ(α)Π(r,ν)F(ν)dνdr=∫tτ(t−r)α−1Γ(α)F(r)(r−τ)α−1Γ(α)Indr+∫tτ∫tr(t−ν)α−1Γ(α)F(ν)Π(ν,r)dν(r−τ)α−1Γ(α)F(r)dr=∫tτ(t−r)α−1Γ(α)F(r)(r−τ)α−1Γ(α)Indr+∫tτ(Π(t,r)−(t−r)α−1Γ(α)In)(r−τ)α−1Γ(α)F(r)dr=∫tτ(r−τ)α−1Γ(α)Π(t,r)F(r)dr. |
This completes the proof.
Indeed, we can see that (4.2) in Lemma 4.1 comes from the following definition:
Definition 4.2. Let F(⋅)∈L∞([t0,tf];Rn×n). The right RL fractional state-transition matrix Π(t,⋅)∈RLn×nDαt− with t0≤τ≤t≤tf is defined by
Π(t,τ)=(t−τ)α−1Γ(α)In+∫tτ(r−τ)α−1Γ(α)Π(t,r)F(r)dr, | (4.5) |
which is the solution to the following right RL fractional differential equation:
{Dαt−[Π](t,τ)=Π(t,τ)F(τ),τ∈[t0,t),I1−αt−[Π](t,t)=In. |
We state the representation result of left linear Caputo fractional differential equations with initial conditions.
Lemma 4.2. Let F(⋅)∈L∞([t0,tf];Rn×n) and G(⋅)∈L∞([t0,tf];Rn). Consider the following left Caputo fractional differential equation:
{CDαt0+[Q](t)=F(t)Q(t)+G(t),t∈(t0,tf],Q(t0)=Q0∈Rn. | (4.6) |
Then, (4.6) admits a unique solution Q(⋅)∈CnDαt0+. Furthermore, the solution of (4.6) is given by
Q(t)=Q0+∫tt0Π(t,s)(F(s)Q0+G(s))ds, | (4.7) |
where Π is the left RL fractional state-transition matrix associated with F in Definition 4.1.
Proof. As (4.6) is a linear Caputo fractional differential equation, the existence and uniqueness of the solution of Q(⋅)∈CnDαt0+ follows from Theorem 3.1. We now prove (4.7). By Theorem 3.1 and Definition 2.1, (4.6) is equivalent to
Q(t)=Q0+∫tt0(t−τ)α−1Γ(α)(F(τ)Q(τ)+G(τ))dτ. | (4.8) |
To show (4.7), we have to prove that (4.7) satisfies the integral representation in (4.8). Specifically, by comparing (4.7) and (4.8), we need to verify the following result:
∫tt0(t−τ)α−1Γ(α)(F(τ)Q(τ)+G(τ))dτ=∫tt0Π(t,s)(F(s)Q0+G(s))ds. | (4.9) |
Obviously, (4.9) holds when t=t0. For t∈(t0,tf], using (4.7), it follows that
∫tt0(t−τ)α−1Γ(α)(F(τ)Q(τ)+G(τ))dτ=∫tt0(t−τ)α−1Γ(α)F(τ)Q0dτ+∫tt0(t−τ)α−1Γ(α)G(τ)dτ+∫tt0(t−τ)α−1Γ(α)F(τ)∫τt0Π(τ,s)(F(s)Q0+G(s))dsdτ. | (4.10) |
By Fubini's formula (see [52], Theorem 3.3, Chapter 6),
∫tt0(t−τ)α−1Γ(α)F(τ)∫τt0Π(τ,s)(F(s)Q0+G(s))dsdτ=∫tt0∫tτ(t−s)α−1Γ(α)F(s)Π(s,τ)ds(F(τ)Q0+G(τ))dτ. |
Then, by Definition 4.1 and (4.10), it follows that
∫tt0(t−τ)α−1Γ(α)(F(τ)Q(τ)+G(τ))dτ=∫tt0(t−τ)α−1Γ(α)F(τ)Q0dτ+∫tt0(t−τ)α−1Γ(α)G(τ)dτ | (4.11) |
+∫tt0∫tτ(t−s)α−1Γ(α)F(s)Π(s,τ)ds(F(τ)Q0+G(τ))dτ=∫tt0Π(t,τ)(F(τ)Q0+G(τ))dτ. | (4.12) |
Hence, by (4.11), (4.9) is verified. This completes the proof.
The following result shows the representation result of right linear RL fractional differential equations with terminal conditions. Here, the corresponding RL fractional differential equation includes the variable of the Lebesgue-Stieltjes measure on [t0,tf]. Hence, its solution has to be written in terms of the integral of the Lebesgue-Stieltjes measure.
Lemma 4.3. Let F(⋅)∈L∞([t0,tf];Rn×n), G1(⋅),G2(⋅)∈L∞([t0,tf];Rn), and η(⋅)∈NBV([t0,tf];R). Consider the following right RL fractional differential equation:
{d[I1−αtf−[Q]](t)=−[F(t)⊤Q(t)+G1(t)]dt−G2(t)dη(t),I1−αtf−[Q](tf)=Qf, | (4.13) |
where dη denotes the Lebesgue-Stieltjes measure of η on ([t0,tf],B([t0,tf])).‡ Then, (4.13) admits a unique solution Q(⋅)∈BV([t0,tf];Rn). Furthermore, the solution of (4.13) is given by
‡B([t0,tf]) denotes the Borel σ-algebra generated by subintervals of [t0,tf].
Q(t)=Π(tf,t)⊤Qf+∫tftΠ(τ,t)⊤G1(τ)dτ+∫tftΠ(τ,t)⊤G2(τ)dη(τ), | (4.14) |
where Π is the right RL state-transition matrix associated with F defined in (4.5) of Definition 4.2.
Proof. The proof for the existence and uniqueness is similar to that of Theorem 3.1 and Lemma 4.2. To prove (4.14), by using the definition of Π in (4.5),
Q(t)=(tf−t)α−1Γ(α)Qf+∫tft(τ−t)α−1Γ(α)G1(τ)dτ+∫tft(τ−t)α−1Γ(α)G2(τ)dη(τ)+∫tft(τ−t)α−1Γ(α)F(τ)⊤Π(tf,τ)⊤Qfdτ+∫tft∫τt(r−t)α−1Γ(α)F(r)⊤Π(τ,r)⊤drG1(τ)dτ+∫tft∫τt(r−t)α−1Γ(α)F(r)⊤Π(τ,r)⊤drG2(τ)dη(τ). |
The multiple integrals in the above equation have to be analyzed. By Fubini's formula (see [52], Theorem 3.3, Chapter 6),
∫tft∫τt(r−t)α−1Γ(α)F(r)⊤Π(τ,r)⊤drG1(τ)dτ=∫tft(τ−t)α−1Γ(α)F(τ)⊤∫tfτΠ(r,τ)⊤G1(r)drdτ∫tft∫τt(r−t)α−1Γ(α)F(r)⊤Π(τ,r)⊤drG2(τ)dη(τ)=∫tft(τ−t)α−1Γ(α)F(τ)⊤∫tfτΠ(r,τ)⊤G2(r)drdη(τ). |
Hence, it follows that
Q(t)=(tf−t)α−1Γ(α)Qf+∫tft(τ−t)α−1Γ(α)(G1(τ)+F(τ)⊤Q(τ))dτ+∫tft(τ−t)α−1Γ(α)G2(τ)dη(τ). |
By Definitions 2.1 and 2.2, this is equivalent to
Q(t)=D1−αtf−[Qf](t)+Iαtf−[G1(⋅)+F(⋅)⊤Q(⋅)](t)+∫tft(τ−t)α−1Γ(α)G2(τ)dη(τ). |
As F(⋅)∈L∞([t0,tf];Rn×n) and G1(⋅),G2(⋅)∈L∞([t0,tf];Rn), from Lemmas 2.1 and 2.3 and Definitions 2.1 and 2.2 with ([2], Theorem 2.2) and Fubini's formula,
−I1−αtf−[Q](t)=−I1−αtf−[−ddt[Iαtf−[Qf]](⋅)](t)−I1tf−[G1(⋅)+F(⋅)⊤Q(⋅)](t)−I1−αtf−[∫tf⋅(τ−⋅)α−1Γ(α)G2(τ)dη(τ)](t)=−Qf−∫tft[G1(s)+F(s)⊤Q(s)]ds−∫tftG2(τ)dη(τ). | (4.15) |
Indeed, to show (4.15), let B the beta function. By Fubini's formula and change of integral variables with s=t−r(τ−t) (which implies ds=−(τ−t)dr), we can show that
I1−αtf−[∫tf⋅(τ−⋅)α−1Γ(α)G2(τ)dη(τ)](t)=∫tft∫tτ(s−t)−αΓ(1−α)(τ−s)α−1Γ(α)G2(τ)dsdη(τ)=1Γ(1−α)Γ(α)∫tft∫0−1(−r(τ−t))−α(τ−t+r(τ−t))α−1G2(τ)(−(τ−t))drdη(τ)=B(α,1−α)Γ(1−α)Γ(α)∫tftG2(τ)dη(τ)=∫tftG2(τ)dη(τ). |
Hence, (4.15) holds. Note that I1−αtf−[Q](tf)=Qf. As I1−αtf−[Q](⋅)∈AC([t0,tf];Rn), by Definitions 2.1 and 2.2, the differential form of (4.15) can be written by (4.13). We can easily observe that Q(⋅)∈BV([t0,tf];Rn). This completes the proof.
Assume that (¯u(⋅),¯X(⋅))∈U×CnDαt0+ is the optimal solution of (P), where ¯u(⋅)∈U is the optimal solution of (P) and ¯X(⋅):=¯X(⋅;¯X0,¯u)∈CnDαt0+ is the corresponding optimal state trajectory of (3.1) controlled by ¯u(⋅)∈U. Note that ¯X(⋅)∈CnDαt0+ satisfies the terminal and running state constraints in (3.6). For notational convenience, let
¯f(t):=f(t,¯X(t),¯u(t)),∂X¯f(t):=∂Xf(t,¯X(t),¯u(t))¯l(t):=l(⋅,¯X(t),¯u(t)),∂X¯l(t):=∂Xl(t,¯X(t),¯u(t))¯m:=m(¯X0,¯X(tf)),∂X0¯m:=∂X0m(¯X0,¯X(tf)),∂X¯m:=∂Xm(¯X0,¯X(tf))¯Gi(t):=Gi(t,¯X(t)),∂X¯Gi(t):=∂XGi(t,¯X(t)). |
We state the Pontryagin maximum principle of (P). The proof is provided in Section 7.
Theorem 5.1. Suppose that Assumptions 3.1 and 3.2 hold. Assume that the pair (¯u(⋅),¯X(⋅))∈U×CnDαt0+ is the optimal solution of (P). Then, there exist a tuple (λ,ξ,θ)∈R×R2n×NBV([t0,tf];Rq) and the adjoint vector p(⋅)∈BV([t0,tf];Rn), where ξ=(ξ1,ξ2) with ξ1,ξ2∈Rn, and θ(⋅)=(θ1(⋅),…,θq(⋅)) with θi satisfying θi(⋅)∈NBV([t0,tf];R), i=1,…,q, such that the following conditions are satisfied:
(i) Nontriviality condition: (λ,ξ,θ)≠0 with λ≥0, ξ=(ξ1,ξ2)∈NF(¯X0,¯X(tf)) and ‖θi(⋅)‖NBV=θi(tf)≥0 for i=1,…,q, where NF(¯X0,¯X(tf)) is the normal cone to the convex set F at (¯X0,¯X(tf))∈F defined in (7.1), and θi(⋅)∈NBV([t0,tf];R), i=1,…,q, is finite, nonnegative and monotonically nondecreasing on [t0,tf];
(ii) Adjoint equation: The adjoint vector p(⋅)∈BV([t0,tf];Rn) is the unique solution of the following right RL fractional differential equation: a.e. t∈[t0,tf],
d[I1−αtf−[p]](t)=−[∂X¯f(t)⊤p(t)+λ(tf−t)β−1Γ(β)∂X¯l(t)⊤]dt−q∑i=1∂X¯Gi(t)⊤dθi(t), | (5.1) |
which is equivalent to the following right RL fractional integral form:
I1−αtf−[p](t)=I1−αtf−[p](tf)+∫tft[∂X¯f(s)⊤p(s)+λ(tf−s)β−1Γ(β)∂X¯l(s)⊤]ds+∫tftq∑i=1∂X¯Gi(s)⊤dθi(s); |
(iii) Transversality condition:
I1−αtf−[p](t0)=−(ξ1+λ∂X0¯m⊤),I1−αtf−[p](tf)=ξ2+λ∂X¯m⊤; |
(iv) Complementary slackness condition:
∫tft0¯Gi(s)dθi(s)=0,∀i=1,…,q; |
(v) Hamiltonian minimization condition:
minu∈UH(t,¯X(t),p(t);u)=H(t,¯X(t),p(t);¯u(t)),a.e.t∈[t0,tf], |
where H:[t0,tf]×Rn×Rn×U→R is the Hamiltonian defined by
H(t,X,p;u):=⟨p,f(t,X,u)⟩+λ(tf−t)β−1Γ(β)l(t,X,u). | (5.2) |
Remark 5.1. The adjoint equation p(⋅)∈BV([t0,tf];Rn) in (5.1) is expressed as the right RL fractional differential (or integral) form that includes the Lebesgue-Stieltjes integral, where dθi, i=1,…,q, is the corresponding Lebesgue-Stieltjes measure on ([t0,tf],B([t0,tf])) with dθi(t)≥0 for t∈[t0,tf] by (i) of Theorem 5.1. When θi(⋅)∈AC([t0,tf];R), i=1,…,q, θi is differentiable a.e. t∈[t0,tf]. In this case, we have p(⋅)∈RLnDαtf−, which can be expressed as the following right RL fractional differential equation:
Dαtf−[p](t)=∂X¯f(t)⊤p(t)+λ(tf−t)β−1Γ(β)∂X¯l(t)⊤+q∑i=1∂X¯Gi(t)⊤dθi(t)dt,a.e.t∈[t0,tf]. |
Here, dθi(t)dt is the derivative of θi(⋅)∈AC([t0,tf];R) for i=1,…,q ([52], Theorem 3.11, Chapter 3).
Remark 5.2. (i) Without (3.5), Theorem 5.1 becomes similar to ([31], Theorem 3.1), where the control space of ([31], Theorem 3.1) is assumed to be compact. In this case, only (ii) and (v) of Theorem 5.1 are needed with λ=1, ξ=0, and θ=0, where the adjoint equation p(⋅)∈RLnDαtf− becomes
Dαtf−[p](t)=∂X¯f(t)⊤p(t)+(tf−t)β−1Γ(β)∂X¯l(t)⊤,a.e.t∈[t0,tf]. |
Note also that without the running state constraint, Theorem 5.1 is specialized to ([20], Theorem 3.12) with θ=0. However, our proof technique of Theorem 5.1 is different from that of ([20], Theorem 3.12), since we adopt the general measurable admissible controls.
(ii) Note that by Remark 2.1, when α=β=1 in Theorem 5.1, Theorem 5.1 is reduced to the maximum principle for the nonfractional optimal control problem with terminal and running state constraints (see [37,48,53] and the references therein for different versions of maximum principles for nonfractional optimal control problems with state constraints).
Remark 5.3. The maximum principle in Theorem 5.1 are of interest only when the running state constraint is nondegenerate in the sense that ∂X¯Gi(t)⊤≠0 whenever ¯Gi(t)=0 for all t∈[t0,tf] and i=1,…,q. A similar remark is given in ([54], page 330, Remarks (b)) for the classical state-constrained optimal control problem for ordinary state equations.
In this section, we study an application of Theorem 5.1 to the state-constrained fractional linear-quadratic control problem. The purpose of this section is to demonstrate the general applicability of Theorem 5.1 by providing a computation method of unknown variables (λ,ξ,θ) in Theorem 5.1.
Consider the following R2-valued left Caputo fractional differential equation with order α∈(0,1) and X(⋅)=[X1(⋅)X2(⋅)]⊤:
{CDαt0+[X1](t)=X2(t),t∈(t0,tf],CDαt0+[X2](t)=u(t). | (6.1) |
Let U=[−u′,u′]⊂R be the control space. The quadratic objective functional to be minimized is given by
(LQ-P)infu(⋅)∈UJ((a,b);u(⋅))=12∫tft0(tf−s)β−1Γ(β)u(s)2ds, | (6.2) |
subject to the terminal and running state constrains:
{(X1(t0),X2(t0))=(a,b),(X1(tf),X2(tf))=(0,0),X2(t)≤0,∀t∈[t0,tf]. | (6.3) |
We observe that (6.1)–(6.3) hold Assumptions 3.1 and 3.2. Therefore, we can apply Theorem 5.1 to obtain the optimal solution of (LQ-P).§
§We assume the existence of the optimal solutions of (LQ-P).
By Theorem 5.1, the adjoint equation p(⋅)=[p1(⋅)p2(⋅)]⊤∈R2 is given by
{d[I1−αtf−[p1]](t)=0⋅dt,d[I1−αtf−[p2]](t)=−p1(t)dt−dθ(t). | (6.4) |
The transversality condition (recall p(⋅)=[p1(⋅)p2(⋅)]⊤∈R2) is as follows:
{I1−αtf−[p](t0)=−ξ1=−[ξ11ξ12]∈R2,I1−αtf−[p](tf)=ξ2=[ξ21ξ22]∈R2. |
Due to the terminal state constraints in (6.1) and the definition of the normal cone in (7.1), ξ1 can be arbitrary, whereas ξ2 has to be determined based on the terminal condition of (6.4).
In view of Lemma 4.3 and ([1], Corollary 2.1 and Property 2.1, Chapter 2), the solution of the adjoint equation in (6.4) can be written as
{p1(t)=(tf−t)α−1Γ(α)ξ21,p2(t)=(tf−t)α−1Γ(α)ξ22+(tf−t)2α−1Γ(2α)ξ21+∫tft(τ−t)α−1Γ(α)dθ(τ). | (6.5) |
The complementary slackness condition with θ(⋅)∈NBV([t0,tf];R) is as follows (see also (7.22)):
∫tft0¯X2(s)dθ(s)=0. | (6.6) |
The Hamiltonian minimization condition is as follows:
minu∈UH(t,¯X(t),p(t);u)=H(t,¯X(t),p(t);¯u(t)),a.e.t∈[t0,tf], |
where H(t,X,p;u):=p1(t)X2(t)+p2(t)u(t)+λ(tf−t)β−1Γ(β)u(t)22. By (6.5) and the first-order optimality condition, the (candidate) optimal solution can be obtained by (note that λ≠0 by the nontriviablity condition)
¯u(t)=−1λΓ(β)(tf−t)1−βp2(t)=−ξ22λζ1(t)−ξ21λζ2(t)−1λζ3(t), | (6.7) |
where
ζ1(t)=Γ(β)(tf−t)α−βΓ(α),ζ2(t)=Γ(β)(tf−t)2α−βΓ(2α)ζ3(t)=Γ(β)(tf−t)1−β∫tft(τ−t)α−1Γ(α)dθ(τ). |
By Theorem 3.1 (and Lemma 2.2) and (6.7), the optimal state trajectory is as follows:
{¯X1(t)=a+bIαt0+[1](t)+I2αt0+[¯u(⋅)](t),¯X2(t)=b+Iαt0+[¯u(⋅)](t). | (6.8) |
When t=tf, the state trajectory must hold the terminal state constraint in (6.3), i.e.,
{¯X1(tf)=a+bIαt0+[1](tf)+I2αt0+[¯u(⋅)](tf)=0,¯X2(tf)=b+Iαt0+[¯u(⋅)](tf)=0. | (6.9) |
We study the computation method to obtain unknown variables (λ,ξ21,ξ22,θ). By (6.7), it should be λ≠0. Hence, by the nontriviality condition, we take λ=1. Recall from Theorem 5.1 that θ(⋅)∈NBV([t0,tf];R) is finite, nonnegative, and monotonically nondecreasing on [t0,tf]. Hence, θ has a countable number of jumps on [t0,tf] (see [52], Lemma 3.12), which implies that we may use the approximation of dθ([t0,tf])=∑Kk=0θk, where θi≥0, i=0,…,K, and K defines the K+1 disjoint intervals in [t0,tf] with t0≤t1≤t2≤⋯≤tK−1≤tK≤tK+1=tf. Here, you can choose K such that the algebraic equation given below (see (6.14)) admits a solution. Also, as θ(0)=0, we have θ0=0.
When t∈[tk,tk+1), k∈{0,…,K}, it follows that
{ζ3(t)=Γ(β)(tf−t)1−β∑Kj=kθj(tj+1−t)α−1Γ(α),t∈[tk,tk+1),ζ3(tf)=0. | (6.10) |
For t∈[tk,tk+1), k∈{0,…,K}, let
ˆζj3(t)=Γ(β)Γ(α)Γ(α)∫tt0(t−s)α−1(tf−s)1−β(tj−s)α−1ds,ˆζj3(tf)=0˜ζj3(t)=Γ(β)Γ(α)Γ(α)∫tt0(t−s)2α−1(tf−s)1−β(tj−s)α−1ds,˜ζj3(tf)=0. |
Then from (6.10), for t∈[tk,tk+1), k∈{0,…,K}, it holds that
Iαt0+[¯u(⋅)](t)=−ξ22Iαt0+[ζ1(⋅)](t)−ξ21Iαt0+[ζ2(⋅)](t)−K∑j=kθjˆζj+13(t)I2αt0+[¯u(⋅)](t)=−ξ22I2αt0+[ζ1(⋅)](t)−ξ21I2αt0+[ζ2(⋅)](t)−K∑j=kθj˜ζj+13(t). |
The complementary slackness condition in (6.6) becomes
∫tft0¯X2(s)dθ(s)=0⇔K∑k=0θk¯X2(tk)=0, | (6.11) |
where by (6.8)–(6.10) and the definition of ˆζ and ˜ζ (note that tK+1=tf),
{¯X1(tk)=a+bIαt0+[1](tk)−ξ22I2αt0+[ζ1(⋅)](tk)−ξ21I2αt0+[ζ2(⋅)](tk)−K∑j=kθj˜ζj+13(tk),k=0,…,K,¯X1(tK+1)=¯X1(tf)=a+bIαt0+[1](tf)−ξ22I2αt0+[ζ1(⋅)](tf)−ξ21I2αt0+[ζ2(⋅)](tf)=0,¯X2(tk)=b−ξ22Iαt0+[ζ1(⋅)](tk)−ξ21Iαt0+[ζ2(⋅)](tk)−K∑j=kθjˆζj+13(tk),k=0,…,K,¯X2(tK+1)=¯X2(tf)=b−ξ22Iαt0+[ζ1(⋅)](tf)−ξ21Iαt0+[ζ2(⋅)](tf)=0. | (6.12) |
Then from (6.9), (6.11) and (6.12), we arrive the following (linear and nonlinear) algebraic equations to compute ξ21, ξ22, and θk, k=1,…,K:
[I2αt0+[ζ1(⋅)](tf)I2αt0+[ζ2(⋅)](tf)Iαt0+[ζ1(⋅)](tf)Iαt0+[ζ2(⋅)](tf)]×[ξ22ξ21]=[a+bIαt0+[1](tf)b], | (6.13) |
and (note that θ0=0)
[b−ξ22Iαt0+[ζ1(⋅)](t1)−ξ21Iαt0+[ζ2(⋅)](t1)⋮b−ξ22Iαt0+[ζ1(⋅)](tK)−ξ21Iαt0+[ζ2(⋅)](tK)]⊤×[θ1⋮θK]−[∑Kj=1θjˆζj+13(t1)⋮θKˆζK+13(tK)]⊤×[θ1⋮θK]=0. | (6.14) |
Based on the above analysis, we get the following conclusion of this section:
Proposition 6.1. The pair (¯u(⋅),¯X(⋅))∈U×CnDαt0+ given in (6.7) and (6.8) is the optimal solution of (LQ-P), where (ξ2,θ) can be obtained by solving the algebraic equations in (6.13) and (6.14).
Remark 6.1. (i) In order to solve the algebraic equations in (6.13) and (6.14) and compute the optimal state trajectory ¯X (see (6.8)) as well as the adjoint equation in (6.5), we have to rely on numerical approaches. Note that to solve (6.13) and (6.14), we need to compute RL fractional integrals, for which [55,56,57] provide several related numerical schemes. Various numerical methods for solving fractional differential equations to obtain the optimal state trajectory ¯X (see (6.8)) as well as the adjoint equation in (6.5) can be found in [55,58,59] and the references therein.
(ii) We mention that [39] studied general numerical approaches for solving the linear-quadratic fractional control problem with state constraints. We believe that the approaches of [39] can be applied to solve the linear-quadratic problem in this section. Note that the researches in [44,45] also provided different numerical methods for solving fractional optimal control problems.
This section is devoted to prove Theorem 5.1.
Let dF:R2n→R+ be the standard Euclidean distance function to F defined by dF(x):=infy∈F|x−y| for x∈R2n. Note that dF(x)=0 when x∈F. Then the terminal state constraint in (3.5) can be written as (X0,X(tf))∈F⇔dF(X0,X(tf))=0. By the projection theorem ([60], Theorem 2.10), there is a unique PF(x)∈F with PF(x):R2n→F⊂R2n, the projection of x∈R2n onto F, such that dF(x)=infy∈F|x−y|=|x−PF(x)|. By ([60], Lemma 2.11), PF(x)∈F is the corresponding projection if and only if ⟨x−PF(x),y−PF(x)⟩≤0 for all y∈F. In view of ([60], Definition 2.37), we have x−PF(x)∈NF(PF(x)) for x∈R2n, where NF(x) is the normal cone to the convex set F at a point x∈F defined by
NF(x):={y∈R2n|⟨y,y′−x⟩≤0,∀y′∈F}. | (7.1) |
Lemma 7.1. ([61], page 167) and ([62], Proposition 2.5.4) dF(x)2 is Fréchet differentiable with the Fréchet differentiation of dF(x)2 at x∈R2n given by DdF(x)2(h)=2⟨x−PF(x),h⟩ for h∈R2n.
We define ψ(X(⋅)):=G(⋅,X(⋅))=[G1(⋅,X(⋅))⋯Gq(⋅,X(⋅))]⊤, where ψ:C([t0,tf];Rn)→C([t0,tf];Rq). Let S:=C([t0,tf];Rq−)⊂C([t0,tf];Rq), where Rq−:=R−×⋯×R−. Note that S is a nonempty closed convex cone of C([t0,tf];Rq) with a nonempty interior ([51], Appendix B) (see also ([48], Section 2.2)). Then, the running state constraint in (3.5) is equivalent to
ψ(X(⋅))∈S⇔Gi(t,X(t))≤0,∀t∈[t0,tf],i=1,…,q. | (7.2) |
As G is continuously differentiable, ψ is Fréchet differentiable with its Fréchet differentiation at X(⋅)∈C([t0,tf];Rn) given by Dψ(X(⋅))(w)=∂XG(⋅,X(⋅))w for w∈C([t0,tf];Rn) ([61], page 167). The normal cone to S at x∈S is defined by
NS(x):={κ∈C([t0,tf];Rq)∗|⟨κ,κ′−x⟩C∗q×Cq≤0,∀κ′∈S}, | (7.3) |
where ⟨⋅,⋅⟩C∗q×Cq:=⟨⋅,⋅⟩C([t0,tf];Rq)∗×C([t0,tf];Rq) stands for the duality paring between C([t0,tf];Rq) and C([t0,tf];Rq)∗ with C([t0,tf];Rq)∗ being the dual space of C([t0,tf];Rq).¶
¶Let (X,‖⋅‖X) be a Banach space. We denote (X∗,‖⋅‖X∗) by the dual space of (X,‖⋅‖X), where X∗ is the space of bounded linear functionals on X with the norm given by ‖ψ‖X∗:=supx∈X,‖x‖X≤1⟨ψ,x⟩X∗×X. Here, ⟨⋅,⋅⟩X∗×X denotes the usual duality paring between X and X∗, i.e., ⟨ψ,x⟩X∗×X:=ψ(x). Recall that (X∗,‖⋅‖X∗) is also a Banach space.
Remark 7.1. Note that (C([t0,tf];Rq),‖⋅‖∞,q) is a separable Banach space ([63], Theorem 6.6, page 140). Then, by ([43], Theorem 2.18, page 42), there exists a norm ‖⋅‖C([t0,tf];Rq) on C([t0,tf];Rq), equivalent to ‖⋅‖∞,q ([43], Definition 2.17, page 42), such that (C([t0,tf];Rq)∗,‖⋅‖C([t0,tf];Rq)∗) is strictly convex, i.e., ‖x‖C([t0,tf];Rq)∗=‖y‖C([t0,tf];Rq)∗=1 and ‖x+y‖C([t0,tf];Rq)∗=2 imply that x=y for x,y∈C([t0,tf];Rq)∗ ([43], Definition 2.12, page 41).
Let dS:C([t0,tf];Rq)→R+ be the distance function to S defined by dS(x):=infy∈S‖x−y‖C([t0,tf];Rq) for x∈C([t0,tf];Rq). Then, we note that (7.2) is equivalent to dS(ψ(X(⋅))=0⇔ψ(X(⋅))∈S.
Lemma 7.2. ([62], Proposition 2.4.1 and page 53) dS is nonexpansive, continuous, and convex.
We define the subdifferential of dS at x∈C([tf,tf];Rq) by ([62], pages 10, 27)
∂dS(x):={y′∈C([t0,tf];Rq)∗|⟨y′,y−x⟩C∗q×Cq≤dS(y)−dS(x),∀y∈C([t0,tf];Rq)}. | (7.4) |
By ([62], page 27), since dS is continuous, ∂dS(x) is a nonempty (∂dS(x)≠∅), convex, and weak–∗ compact subset of C([t0,tf];Rq)∗. From ([62], Proposition 2.1.2) and Lemma 7.2, ‖y′‖C([t0,tf];Rq)∗≤1 for all y′∈∂dS(x). Then, by ([43], Proposition 3.11, Chapter 4) (see also ([51], Proposition A.2) and ([48], Proposition 5, Appendix B.2)) and the fact that S⊂C([t0,tf];Rq) is a nonempty closed convex subset, we can conclude that ‖y′‖C([t0,tf];Rq)∗=1 for all y′∈∂dS(x) with x∉S.
Remark 7.2. We claim that ∂dS(x)={y′}, i.e., ∂dS(x) is a singleton, for x∈C([t0,tf];Rq)∖S. To show this, let us assume that ∂dS(x) is not a singleton. We take y′,y′′∈∂dS(x) with y′≠y′′. As ∂dS(x) is convex, we have ηy′+(1−η)y′′∈∂dS(x) for η∈[0,1]. Note that by the preceding discussion, ‖ηy′+(1−η)y′′‖C([t0,tf];Rq)∗=1 for η∈[0,1]. Then ‖y′‖C([t0,tf];Rq)∗=1 and ‖y′′‖C([t0,tf];Rq)∗=1 when η=1 and η=0, respectively. In addition, when η=12, we have ‖y′+y′′‖C([t0,tf];Rq)∗=2. This, together with the fact that (C([t0,tf];Rq)∗,‖⋅‖C([t0,tf];Rq)∗) is strictly convex by Remark 7.1, implies that y′=y′′, which contradicts our assumption y′≠y′′. Hence, y′=y′′∈∂dS(x) and ‖y′‖C([t0,tf];Rq)∗=1. Then we repeat this process for all elements of ∂dS(x), which implies ∂dS(x)={y′}, i.e., ∂dS(x) is a singleton, and ‖y′‖C([t0,tf];Rq)∗=1.
Lemma 7.3. dS is strictly Hadamard differentiable on C([t0,tf];Rq)∖S with the Hadamard differential DdS satisfying ‖DdS(x)‖C([t0,tf];Rq)∗=‖y′‖C([t0,tf];Rq)∗=1 for all x∈C([t0,tf];Rq)∖S. Also, dS(x)2 is strictly Hadamard differentiable on C([t0,tf];Rq)∖S with the Hadamard differential given by DdS(x)2=2dS(x)DdS(x) for x∈C([t0,tf];Rq)∖S. Finally, dS(x)2 is Fréchet differentiable on S with the Fréchet differential being DdS(x)2=0∈C([t0,tf];Rq)∗ for all x∈S.
Proof. Recall dS(x)=0 when x∈S, and dS(x)>0 for x∈C([t0,tf];Rq)∖S. Consider dS when x∈C([t0,tf];Rq)∖S. The nonexpansive property of dS from Lemma 7.2 implies that dS is Lipschitz continuous. In addition, by Remark 7.2, the subdifferential of dS is a singleton. Then we are able to invoke ([64], Theorem 3.54) to conclude that dS(x) is strictly Hadamard differentiable on C([t0,tf];Rq)∖S, where by Remark 7.2, ‖DdS(x)‖C([t0,tf];Rq)∗=1. Consequently, dS(x)2 is strictly Hadamard differentiable on C([t0,tf];Rq)∖S with DdS(x)2=2dS(x)DdS(x) for x∈C([t0,tf];Rq)∖S. When dS is on S, we have DdS(x)2=0∈C([t0,tf];Rq)∗ for x∈S due to the fact that dS(x)=0 for x∈S and dS is nonexpansive shown in Lemma 7.2.
Recall that (¯u(⋅),¯X(⋅))∈U×CnDαt0+ is the optimal solution of (P). For ϵ>0, define the penalized objective functional
Jϵ(X0;u(⋅))=(([J(X0;u(⋅))−J(¯X0;¯u(⋅))+ϵ]+)2+dF(X0,X(tf))2+dS(ψ(X(⋅)))2)12, | (7.5) |
where J is given in (3.4). Define the Ekeland metric by
ˆd((X0,u(⋅)),(˜X0,˜u(⋅)))=|X0−˜X0|+˜d(u(⋅),˜u(⋅)), | (7.6) |
where for u(⋅),˜u(⋅)∈U, ˜d(u(⋅),˜u(⋅)):=|{t∈[t0,tf]|u(t)≠˜u(t)}|. Note that (Rn×U,ˆd) is a complete metric space ([43], Proposition 3.10, Chapter 4), and Jϵ is a continuous functional on (Rn×U,ˆd) in view of Lemmas 7.1 and 7.2, Theorem 3.1.||
||Note that Theorem 3.1 with Assumption 3.1 (particularly supu∈U|f(t,x,u)|<∞ for (t,x)∈[t0,tf]×Rn) and the Arzeá–Ascoli theorem imply that all state trajectories X of the control system (3.1) is equicontinuous and bounded, and therefore relatively compact in the uniform topology. Then there is a subsequence of the state trajectories converging uniformly, which together with Assumption 3.2 and Lemmas 7.1 and 7.2 implies the continuity of Jϵ on (Rn×U,ˆd).
We observe that
Jϵ(X0,u(⋅))>0,∀(X0,u(⋅))∈Rn×U,Jϵ(¯X0;¯u(⋅))=ϵ≤inf(X0,u(⋅))∈Rn×UJϵ(X0,u(⋅))+ϵ. |
Then, by the Ekeland variational principle [65] (see also ([43], Corollary 2.2, Chapter 4)), there exists a pair (Xϵ0,uϵ(⋅))∈Rn×U such that
ˆd((Xϵ0,uϵ(⋅)),(¯X0,¯u(⋅)))≤√ϵ, | (7.7) |
and
{Jϵ(Xϵ0,uϵ(⋅))≤Jϵ(¯X0;¯u(⋅))=ϵ,Jϵ(Xϵ0,uϵ(⋅))≤Jϵ(X0;u(⋅))+√ϵˆd((Xϵ0,uϵ(⋅)),(X0;u(⋅))),∀(X0,u(⋅))∈Rn×U. | (7.8) |
From (7.8), we observe that the pair (Xϵ0,uϵ(⋅))∈Rn×U is the optimal solution to the unconstrained problem given by Jϵ(X0;u(⋅))+√ϵˆd((Xϵ0,uϵ(⋅)),(X0;u(⋅))).
For δ∈(0,1), define (with |E| being the Lebesgue measure of E)
Eδ={E⊂[t0,tf]||E|=δ(tf−t0)}. | (7.9) |
For any Eδ∈Eδ and u(⋅)∈U, we introduce the spike variation:
uϵ,δ(t):={uϵ(t),t∈[t0,tf]∖Eδ,u(t),t∈Eδ. | (7.10) |
Clearly uϵ,δ(⋅)∈U and ˜d(uϵ,δ(⋅),uϵ(⋅))≤|Eδ|=δ(tf−t0). Consider the variations Xϵ(⋅):=X(⋅;Xϵ0,uϵ) and Xϵ,δ(⋅):=X(⋅;Xϵ0+δa,uϵ,δ), where a∈Rn. Note that Xϵ(⋅) is the state trajectory of (3.1) under (Xϵ0,uϵ(⋅))∈Rn×U, and Xϵ,δ is the variational equation under (7.10) and the perturbed initial condition a.
We introduce (with Eϵ,δ(⋅):=Xϵ,δ(⋅)−Xϵ(⋅))
fϵ(s):=f(s,Xϵ(s),uϵ(s)),∂Xfϵ(s):=∂Xf(s,Xϵ(s),uϵ(s))ˆfϵ(s):=f(s,Xϵ(s),u(s))−f(s,Xϵ(s),uϵ(s))fϵ,δX(s):=∫10∂Xf(s,Xϵ(s)+rEϵ,δ(s),uϵ,δ(s))drlϵ(s):=l(s,Xϵ(s),uϵ(s)),∂Xlϵ(s):=∂Xl(s,Xϵ(s),uϵ(s))ˆlϵ(s):=l(s,Xϵ(s),u(s))−l(s,Xϵ(s),uϵ(s))lϵ,δX(s):=∫10∂Xl(s,Xϵ(s)+rEϵ,δ(s),uϵ,δ(s))drmϵ:=m(Xϵ0,Xϵ(tf)),∂X0mϵ:=∂X0m(Xϵ0,Xϵ(tf)),∂Xmϵ:=∂Xm(Xϵ0,Xϵ(tf))mϵ,δX0(tf):=∫10∂X0m(Xϵ0+rδa,Xϵ(tf)+rEϵ,δ(r))drmϵ,δX(tf):=∫10∂Xm(Xϵ0+rδa,Xϵ(tf)+rEϵ,δ(r))drˆf(s):=f(t,¯X(s),u(s))−f(t,¯X(s),¯u(s)),ˆl(s):=l(s,¯X(s),u(s))−l(t,¯X(s),¯u(s)). |
By (7.5) and (7.8), together with (7.6) and (7.10), we have
−√ϵ(|a|+tf)≤1Jϵ(Xϵ0+δa;uϵ,δ(⋅))+Jϵ(Xϵ0;uϵ(⋅))×1δ(([J(Xϵ0+δa;uϵ,δ(⋅))−J(¯X0;¯u(⋅))+ϵ]+)2−([J(Xϵ0;uϵ(⋅))−J(¯X0;¯u(⋅))+ϵ]+)2+dF(Xϵ0+δa,Xϵ,δ(tf))2−dF(Xϵ0,Xϵ(tf))2+dS(ψ(Xϵ,δ(⋅)))2−dS(ψ(Xϵ(⋅)))2). | (7.11) |
By Lemmas B.1 and B.2 in Appendix B, we introduce
Zϵ(t)=a+∫tt0(t−s)α−1Γ(α)(∂Xfϵ(s)Zϵ(s)+ˆfϵ(s))ds,ˆZϵ(tf)=∫tft0(tf−s)β−1Γ(β)(∂Xlϵ(s)Zϵ(s)+ˆlϵ(s))ds+∂X0mϵa+∂XmϵZϵ(tf). |
From Lemmas B.1 and B.2, Zϵ and ˆZϵ are the variational equations related to the pair (Xϵ0,uϵ(⋅))∈Rn×U in (7.8). In fact, by Theorem 3.1, Zϵ is equivalent to (B.2) in Lemma B.1.
We now consider the limit of δ↓0 in (7.11). By continuity of Jϵ on Rn×U, it follows that limδ↓0Jϵ(Xϵ0+δa;uϵ,δ(⋅))=Jϵ(Xϵ0;uϵ(⋅)). Then, the following facts hold:
(F.1) By Lemmas B.1 and B.2, we have
1Jϵ(Xϵ0+δa;uϵ,δ(⋅))+Jϵ(Xϵ0;uϵ(⋅))×1δ(([J(Xϵ0+δa;uϵ,δ(⋅))−J(¯X0;¯u(⋅))+ϵ]+)2−([J(Xϵ0;uϵ(⋅))−J(¯X0;¯u(⋅))+ϵ]+)2)→λϵˆZϵ(tf),as δ↓0, |
where since Jϵ(X0,u(⋅))>0 for (X0,u(⋅))∈Rn×U,
λϵ:=[J(Xϵ0;uϵ(⋅))−J(¯X0;¯u(⋅))+ϵ]+Jϵ(Xϵ0;uϵ(⋅))≥0. |
(F.2) From Lemmas 7.1 and B.1,
1δdF(Xϵ0+δa,Xϵ,δ(tf))2−dF(Xϵ0,Xϵ(tf))2Jϵ(Xϵ0+δa;uϵ,δ(⋅))+Jϵ(Xϵ0;uϵ(⋅))→⟨ξϵ1,a⟩+⟨ξϵ2,Zϵ(tf)⟩,as \delta \downarrow 0 , |
where with ξϵ1,ξϵ2∈Rn,
ξϵ=[ξϵ1ξϵ2]:={[Xϵ0Xϵ(tf)]−PF(Xϵ0,Xϵ(tf))Jϵ(Xϵ0;uϵ(⋅))∈NF(PF(Xϵ0,Xϵ(tf))),(Xϵ0,Xϵ(tf))∉F,0∈NF(PF(Xϵ0,Xϵ(tf))),(Xϵ0,Xϵ(tf))∈F. |
(F.3) By Theorem 3.1, and Lemmas 7.3 and B.1,
1δdS(ψ(Xϵ,δ(⋅)))2−dS(ψ(Xϵ(⋅)))2Jϵ(Xϵ0+δa;uϵ,δ(⋅))+Jϵ(Xϵ0;uϵ(⋅))→⟨μϵ,∂XG(⋅,Xϵ(⋅))Zϵ(⋅)⟩Cq×C∗q,as \delta \downarrow 0 , |
where as DdS(ψ(Xϵ(⋅))) is the subdifferential of dS at ψ(Xϵ(⋅))∈C([t0,tf];Rq) (see Lemma 7.3), by (7.3) and (7.4), μϵ∈NS(ψ(Xϵ(⋅))), and
μϵ:={dS(ψ(Xϵ(⋅)))DdS(ψ(Xϵ(⋅)))Jϵ(Xϵ0;uϵ(⋅))∈C([t0,tf];Rq)∗,ψ(Xϵ(⋅))∉S,0∈C([t0,tf];Rq)∗,ψ(Xϵ(⋅))∈S. |
Therefore, by (F.1)–(F.3), as δ↓0, (7.11) becomes for (a,u(⋅))∈Rn×U,
−√ϵ(|a|+tf)≤λϵˆZϵ(tf)+⟨ξϵ1,a⟩+⟨ξϵ2,Zϵ(tf)⟩+⟨μϵ,∂XG(⋅,Xϵ(⋅))Zϵ(⋅)⟩C∗q×Cq, | (7.12) |
where from Lemma 7.3 and the property of dF in Section 7.1,
|λϵ|2+|ξϵ|2+‖μϵ‖2C([t0,tf];Rq)∗=1. | (7.13) |
We consider the limit of ϵ↓0. We introduce
Z(t)=a+∫tt0(t−s)α−1Γ(α)(∂X¯f(s)Z(s)+ˆf(s))ds, | (7.14) |
ˆZ(tf)=∫tft0(tf−s)β−1Γ(β)(∂X¯l(s)Z(s)+ˆl(s))ds+∂X0¯ma+∂X¯mZ(tf), | (7.15) |
where by Lemma B.3, Z and ˆZ are variational equations related to the optimal solution ¯u(⋅)∈U. By Theorem 3.1, Z in (7.14) is equivalent to (B.4) in Lemma B.3.
Instead of taking the limit of ϵ↓0, let {ϵk} be a sequence of ϵ such that ϵk≥0 and ϵk↓0 as k→∞. We replace ϵ by ϵk. By (7.13), the sequences ({λϵk},{ξϵk},{μϵk}) are bounded for k≥0. Note also from (7.13) that the ball generated by ‖μϵk‖2C([t0,tf];Rq)∗≤1 is a closed unit ball in C([t0,tf];Rq)∗, which is weak–∗ compact by the Banach-Alaoglu theorem ([63], page 130). Then, by the standard compactness argument, we may extract a subsequence of {ϵk}, still denoted by {ϵk}, such that as k→∞,
({λϵk},{ξϵk},{μϵk})→(λ0,ξ0,μ0)=:(λ,ξ,μ), | (7.16) |
where the convergence of {μϵk}→μ (as k→∞) is understood in the weak–∗ sense [63].
We claim that the tuple (λ,ξ,μ) holds
λ≥0,ξ∈NF(PF(¯X0,¯X(tf))),μ∈NS(ψ(¯X(⋅))). | (7.17) |
Indeed, λ≥0 holds due to (F.1). ξ∈NF(PF(¯X0,¯X(tf))) follows from (F.2) and the property of limiting normal cones ([54], page 43). To prove μ∈NS(ψ(¯X(⋅))), μϵk∈NS(ψ(Xϵ(⋅))) in (F.3) means ⟨μϵk,z−ψ(Xϵk(⋅))⟩Cq∗×Cq≤0 for z∈S (see (7.3)). Then it follows that
0≥⟨μϵk,z−ψ(Xϵk(⋅))⟩C∗q×Cq≥⟨μ,z−ψ(¯X(⋅))⟩C∗q×Cq−‖ψ(Xϵk(⋅))−ψ(¯X(⋅))‖∞,q+⟨μϵk,z−ψ(¯X(⋅))⟩C∗q×Cq−⟨μ,z−ψ(¯X(⋅))⟩C∗q×Cq→⟨μ,z−ψ(¯X(⋅))⟩C∗q×Cq,ask→∞, |
where the convergence follows from (7.16) and Lemma B.3, together with the boundedness of {μϵk} by (7.13). This convergence is understood in the weak–∗ sense. This and (7.3) imply μ∈NS(ψ(¯X(⋅))).
By (7.13) and (7.16), together with Lemma B.3, it follows that
λϵkˆZϵk(tf)≤λˆZ(tf)+|ˆZϵk(tf)−ˆZ(tf)|+|λϵk−λ|ˆZ(tf)→λˆZ(tf),ask→∞,⟨ξϵk1,a⟩→⟨ξ1,a⟩,as k→∞,⟨ξϵk2,Zϵk(tf)⟩≤⟨ξ2,Z(tf)⟩+|Zϵk(tf)−Z(tf)|+|ξϵk2−ξ2||Z(tf)|→⟨ξ2,Z(tf)⟩,ask→∞, |
and similarly, together with the definition of the weak–∗ convergence,
⟨μϵk,∂XG(⋅,Xϵk(⋅))Zϵk(⋅)⟩C∗q×Cq≤⟨μ,∂X¯G(⋅)Z(⋅)⟩C∗q×Cq+‖Zϵk(⋅)−Z(⋅)‖∞,n+⟨μϵk,∂X¯G(⋅)Z(⋅)⟩C∗q×Cq−⟨μ,∂X¯G(⋅)Z(⋅)⟩C∗q×Cq→⟨μ,∂X¯G(⋅)Z(⋅)⟩C∗q×Cq,ask→∞. |
Therefore, as k→∞, (7.12) becomes for (a,u(⋅))∈Rn×U,
0≤λˆZ(tf)+⟨ξ1,a⟩+⟨ξ2,Z(tf)⟩+⟨μ,∂XG(⋅,¯X(⋅))Z(⋅)⟩C∗q×Cq, | (7.18) |
Note that (7.18) is the crucial inequality obtained from the Ekeland variational principle as well as the estimates of the variational equations in Lemmas B.1 and B.3.
Let μ=(μ1,…,μq)∈C([t0,tf];Rq)∗. For any z=(z1,…,zq)∈C([t0,tf];Rq), it holds that ⟨μ,z⟩C∗q×Cq=∑qi=1⟨μi,zi⟩C∗1×C1. Recall that μ∈NS(ψ(¯X(⋅))) from (7.17), where ψ(¯X(⋅))=[G1(⋅,¯X(⋅))⋯Gq(⋅,¯X(⋅))]⊤=(ψ1(¯X(⋅)),…,ψq(¯X(⋅)))∈S. Based on (7.3), this implies that
⟨μ,z−ψ(¯X(⋅))⟩C∗q×Cq=q∑i=1⟨μi,zi−ψi(¯X(⋅))⟩C∗1×C1≤0,∀z∈S. | (7.19) |
Take z in (7.19) as in the following cases (note that ¯Gi(⋅)=Gi(⋅,¯X(⋅))):
z=[¯G1(⋅)⋯2¯Gi(⋅)⋯¯Gq(⋅)]⊤∈Sz=[¯G1(⋅)⋯0∈C([t0,tf];R)⋯¯Gq(⋅)]⊤∈S. |
Then, (7.19) is equivalent to
⟨μi,ψi(¯X(⋅))⟩C∗1×C1=0,∀i=1,…,q, | (7.20) |
⟨μi,zi⟩C∗1×C1≥0,∀zi∈C([t0,tf];R+),i=1,…,q. | (7.21) |
For (7.21), by the Riesz representation theorem (see ([63], page 75 and page 382)), there is a unique θ(⋅)=(θ1(⋅),…,θq(⋅)) with θi(⋅)∈NBV([t0,tf];R), i.e., θi, i=1,…,q, are normalized functions of bounded variation on [t0,tf], such that we have ⟨μi,zi⟩C∗1×C1=∫tft0zi(s)dθi(s)≥0 for i=1,…,q, where every θi is finite, nonnegative, and monotonically nondecreasing on [t0,tf] with θi(0)=0. Moreover, for (7.20), the Riesz representation theorem leads to
⟨μi,ψi(¯X(⋅))⟩C∗1×C1=∫tft0¯Gi(s)dθi(s)=0⇔supp(dθi(⋅))⊂{t∈[t0,tf]|Gi(t,¯X(t))=0},∀i=1,…,q, | (7.22) |
where dθi, i=1,…,q, denotes the Lebesgue-Stieltjes measure on ([t0,tf],B([t0,tf])) with dθi(t)≥0 for t∈[t0,tf] due to the fact that θi(⋅)∈NBV([t0,tf];R) is nondecreasing and finite. Hence, the relation in (7.22) proves the complementary slackness condition in Theorem 5.1.
Recall (7.19), i.e., for any z∈S, ⟨μ,z−ψ(¯X(⋅))⟩C∗q×Cq=∑qi=1⟨μi,zi−ψi(¯X(⋅))⟩C∗1×C1≤0. By the Riesz representation theorem (see ([63], page 75 and page 382)) and the fact that θi(⋅)∈NBV([t0,tg];R), i=1,…,q, is finite, nonnegative, and monotonically nondecreasing on [t0,tf] with θi(0)=0, it follows that ‖μi‖C([t0,tf];R)∗=‖θi(⋅)‖NBV=θi(tf)≥0 for i=1,…,q.
By (7.17) and the fact that (¯X0,¯X(tf))∈F implies PF(¯X0,¯X(tf))=(¯X0,¯X(tf)) (see Section 7.1), we have ξ=(ξ1,ξ2)∈NF(¯X0,¯X(tf)).
From the fact that S=C([t0,tf];Rq−) has an nonempty interior, there are z′∈S and σ>0 such that z′+σz∈S for all z∈¯B(C([t0,tf];Rq),‖⋅‖C([t0,tf];Rq))(0,1) (the closure of the unit ball in C([t0,tf];Rq)). Then similar to (7.19) and the fact that μϵ∈NS(ψ(Xϵ(⋅))), it follows that σ⟨μϵ,z⟩C∗q×Cq≤⟨μϵ,ψ(Xϵ(⋅))−z′⟩C∗q×Cq for all z∈¯B(C([t0,t0];Rq),‖⋅‖C([t0,tf];Rq))(0,1). From (7.13) and the definition of the norm of the dual space (the norm of linear functionals on C([t0,tf];Rq) (see Section 7.1)), by taking the limit as ϵ↓0, we get
σ‖μ‖C([t0,tf];Rq)∗=σ√1−|λ|2−|ξ|2≤⟨μ,ψ(¯X(⋅))−z′⟩C∗q×Cq,∀z′∈S. |
Note that σ>0. When μ=0∈C([t0,tf];Rq)∗ and ξ=0, we have λ=1. When λ=0 and μ=0∈C([t0,tf];Rq)∗, we have |ξ|=1. When λ=0 and ξ=0, we have μ≠0∈C([t0,tf];Rq)∗. This implies (λ,ξ,θ)≠0, i.e., they cannot vanish simultaneously.
Based on the above discussion, it holds that (λ,ξ,θ)≠0, where λ≥0 (see (7.17)), ξ∈NF(¯X0,¯X(tf)), θi(⋅)∈NBV([t0,tg];R) is nondecreasing with ‖μi‖C([t0,tf];R)∗=‖θi(⋅)‖NBV=θi(tf)≥0 for i=1,…,q. This shows the nontriviality condition in Theorem 5.1.
From Lemma 4.3 and the transversality condition, the solution of the adjoint equation in (5.1) can be written as
p(t)=Π(tf,t)⊤(ξ2+λ∂X¯m⊤)+λ∫tftΠ(τ,t)⊤(tf−τ)β−1Γ(β)∂X¯l(τ)⊤dτ+∫tftΠ(τ,t)⊤q∑i=1∂X¯Gi(τ)⊤dθi(τ), | (7.23) |
where Π is the right RL fractional state-transition matrix associated with ∂X¯f (see Definition 4.2 and Lemma 4.1 in Section 4). Moreover, from Lemma 4.2 and Theorem 3.1, Z in (7.14) can be written as
Z(t)=a+∫tt0Π(t,s)(∂X¯f(s)a+ˆf(s))ds, | (7.24) |
where Π is the left RL state-transition matrix associated with ∂X¯f (see Definition 4.1).
By using (7.24), (7.15) and the Riesz representation theorem, (7.18) becomes
0≤⟨ξ1+λ∂X0¯m⊤,a⟩+⟨ξ2+λ∂X¯m⊤,a+∫tft0Π(tf,s)(∂X¯f(s)a+ˆf(s))ds⟩+∫tft0q∑i=1∂X¯Gi(s)(a+∫st0Π(s,τ)(∂X¯f(τ)a+ˆf(τ))dτ)dθi(s)+λ∫tft0(tf−s)β−1Γ(β)∂X¯l(s)(a+∫st0Π(s,τ)∂X¯f(τ)adτ)ds+λ∫tft0(tf−s)β−1Γ(β)∂X¯l(s)∫st0Π(s,τ)ˆf(τ)dτds+λ∫tft0(tf−s)β−1Γ(β)ˆl(s)ds. | (7.25) |
In (7.25), by Fubini's formula (see ([52], Theorem 3.3, Chapter 6)),
∫tft0q∑i=1∂X¯Gi(s)∫st0Π(s,τ)ˆf(τ)dτdθi(s)=∫tft0∫tfsq∑i=1∂X¯Gi(τ)Π(τ,s)ˆf(s)dθi(τ)ds,∫tft0(tf−s)β−1Γ(β)∂X¯l(s)∫st0Π(s,τ)ˆf(τ)dτds=∫tft0∫tfs(tf−τ)β−1Γ(β)∂X¯l(τ)Π(τ,s)ˆf(s)dτds. |
Then it follows from Lemma 4.1 in Section 4 that
∫tft0[∫tfsq∑i=1∂X¯Gi(τ)Π(τ,s)ˆf(s)dθi(τ)+(ξ⊤2+λ∂X¯m)Π(tf,s)ˆf(s)+λ∫tfs(tf−τ)β−1Γ(β)∂X¯l(τ)Π(τ,s)ˆf(s)dτ]ds=∫tft0⟨∫tfsq∑i=1Π(τ,s)⊤∂X¯Gi(τ)⊤dθi(τ)+λ∫tfs(tf−τ)β−1Γ(β)Π(τ,s)⊤∂X¯l(τ)⊤dτ+Π(tf,s)⊤(ξ2+λ∂X¯m⊤),ˆf(s)⟩ds. |
We may apply the same technique to the terms in (7.25) related to the initial variation a. Hence, by applying the above result and using (7.23) (with Lemma 4.1 in Section 4), (7.25) becomes
0≤⟨ξ1+λ∂X0¯m⊤,a⟩+⟨ξ2+λ∂X¯m⊤,a⟩+∫tft0q∑i=1∂X¯Gi(s)dθi(s)a+λ∫tft0(tf−s)β−1Γ(β)∂X¯l(s)dsa+∫tft0⟨p(s),∂X¯f(s)a⟩ds+∫tft0⟨p(s),ˆf(s)⟩ds+λ∫tft0(tf−s)β−1Γ(β)ˆl(s)ds. | (7.26) |
Below, we use (7.26) to prove the transversality condition, the nontriviality of the adjoint equation, and the Hamiltonian-like maximum condition of Theorem 5.1.
When u=¯u, (7.26) becomes
0≤⟨ξ1+λ∂X0¯m⊤,a⟩+⟨ξ2+λ∂X¯m⊤,a⟩+∫tft0q∑i=1∂X¯Gi(s)dθi(s)a+λ∫tft0(tf−s)β−1Γ(β)∂X¯l(s)dsa+∫tft0⟨p(s),∂X¯f(s)a⟩ds. | (7.27) |
Then, by the adjoint equation in (5.1), (7.27) can be written as
0≤⟨ξ1+λ∂X0¯m⊤,a⟩+⟨ξ2+λ∂X¯m⊤,a⟩+⟨∫tft0Dαtf−[p](s)ds,a⟩. | (7.28) |
Recall from (7.23) and Lemma 4.3 that
I1−αtf−[p](tf)=ξ2+λ∂X¯m⊤. | (7.29) |
Note that from Definition 2.2 and (5.1),
∫tft0Dαtf−[p](s)ds=I1−αtf−[p](t0)−I1−αtf−[p](tf)=I1−αtf−[p](t0)−(ξ2+λ∂X¯m⊤). |
Then (7.28) becomes 0≤⟨ξ1+λ∂X0¯m⊤+I1−αtf−[p](t0),a⟩. As this holds for a,−a∈Rn,
I1−αtf−[p](t0)=−(ξ1+λ∂X0¯m⊤). | (7.30) |
Hence, (7.29) and (7.30) show the transversality condition.
When a=0, (7.26) becomes 0≤∫tft0⟨p(s),ˆf(s)⟩ds+λ∫tft0(tf−s)β−1Γ(β)ˆl(s)ds. Then, by (5.2),
∫tft0H(s,¯X(s),p(s);¯u(s))ds≤∫tft0H(s,¯X(s),p(s);u(s))ds. |
Since U is separable by Assumption 3.1, there exists a countable dense set U0={ui,i≥1}⊂U. Moreover, there exists a measurable set Si⊂[t0,tf] such that |Si|=tf−t0 and any t∈Si is the Lebesgue point of H(t,¯X(t),p(t);u(t)), i.e., limτ↓012τ∫t+τt−τH(s,¯X(s),p(s);u(s))ds=H(t,¯X(t),p(t);u(t)) ([66], Theorem 5.6.2). We fix ui∈U0. For any t∈Si, define
u(s):={¯u(s),s∈[t0,tf]∖(t−τ,t+τ),ui,s∈(t−τ,t+τ). |
It then follows that
0≤limτ↓012τ∫t+τt−τ[H(s,¯X(s),p(s);ui)−H(s,¯X(s),p(s);¯u(s))]ds=H(t,¯X(t),p(t);ui)−H(t,¯X(t),p(t);¯u(t)). | (7.31) |
Note that (7.31) implies that
H(t,¯X(t),p(t);¯u(t))≤H(t,¯X(t),p(t);u),∀ui∈U0,t∈∩i≥1Si. | (7.32) |
Since ∩i≥1Si=[t0,tf] by the fact that U0 is countable, H is continuous in u∈U, and U is separable, (7.32) implies the Hamiltonian minimization condition. This is the end of the proof for Theorem 5.1.
In this paper, we obtained the Pontryagin maximum principle for the fractional optimal control problem with terminal and running state constraints (see Theorem 5.1). Indeed, our main result indicates the maximum principle leads to the necessary optimality conditions for the problem, which include the nontriviality condition, the transversality condition, the complement slackness condition, and the Hamiltonian minimization condition. Regarding the framework of general theory of control systems, the maximum principle of this paper provides a new approach for solving fractional optimal control problems with state constraints. Indeed, we need to develop the new proof of the maximum principle due to the inherent complex nature of the fractional control problem, the presence of the terminal and running state constraints, and the generalized standing assumptions.
There are several potential future research directions arising from this paper: (i) Generalization to other fractional differential equations such as the Caputo-Katugampola differential equation and systems driven by fractional Brownian motion [67]; (ii) studying the dynamic programming principle for the fractional control problem with state constraints; and (iii) the fractional control problem with delay and state constraints [36]. Furthermore, in engineering problems, instead of the open-loop optimal solution, the closed-loop optimal solution is more appropriate. Hence, one important future research problem is the characterization of the closed-loop optimal solution, particularly for the linear-quadratic problem. Note that we may have to extend techniques of [40] to the case of state constraints.
We should mention that similar to [20], our study has limitations in applications to real practical problems, including those for high-order systems. Recently, for the linear-quadratic case, various numerical methods for several practical fractional control applications were studied in [39] (e.g., the optimal drug scheduling of cancer chemotherapy problem and the minimum fuel optimal control problem; see [39,Section 6 and Example 9]). In addition, the tumor growth problem in fractional optimal control was considered in [45]. It is important to extend the major ideas of [39] to general nonlinear fractional optimal control problems to study practical applications.
The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this article.
This research was supported in part by the Korea Institute of Energy Technology Evaluation and Planning (KETEP) and the Ministry of Trade, Industry & Energy (MOTIE) of the Republic of Korea (RS-2024-00422103) and in part by Institute of Information & communications Technology Planning and Evaluation (IITP) grant funded by the Korea government (Ministry of Science and ICT, MSIT) (RS-2020-II201373).
The author would like to thank two anonymous reviewers for their careful reading and suggestions, which have helped to improve the paper.
All authors declare no conflicts of interest in this paper.
We prove Theorem 3.1.
We first show (3.1) implies the integral expression. As X(⋅)∈CnDαt0+, I1−αt0+[X(⋅)−a](⋅)∈AC([t0,tf];Rn). By Definition 2.3,
I1−αt0+[X(⋅)−X0](t)=I1−αt0+[X(⋅)−X0](t0)+I1t0+[CDαt0+[X](⋅)](t)=I1−αt0+[X(⋅)−X0](t0)+I1t0+[f(⋅,X(⋅),u(⋅))](t). |
As I1−αt0+[X(⋅)−X0](t0)=0 and f(⋅,X,u)∈L∞([t0,tf];Rn), by Lemmas 2.1–2.3, we have
Iαt0+[I1−αt0+[X(⋅)−X0](⋅)](t)=I1t0+[X(⋅)−X0](t)=Iαt0+[I1t0+[f(⋅,X(⋅),u(⋅))](⋅)](t)=I1t0+[Iαt0+[f(⋅,X(⋅),u(⋅))](⋅)](t). |
This implies I1t0+[X(⋅)−X0](t)=I1t0+[Iαt0+[f(⋅,X(⋅),u(⋅))](⋅)](t). Differentiating the above equality shows the desired expression.
We prove the integral expression implies (3.1). Note that by Lemma 2.3, we have Iαt0+[f(⋅,X(⋅),u(⋅)](⋅)∈C([t0,tf];Rn), which implies X(t0)=X0. Then, by Lemmas 2.1–2.3, it follows that I1−αt0+[X(⋅)−X0](t)=I1−αt0+[Iαt0+[f(⋅,X(⋅),u(⋅)](⋅)](t)=I1t0+[f(⋅,X(⋅),u(⋅)](t). Since f(⋅,X,u)∈L∞([t0,tf];Rn), it is clear that X(⋅)∈AC([t0,tf];Rn), which, together with Definition 2.3, implies that CDαt0+[X](t)=ddt[I1−αt0+[X(⋅)−X0]](t)=f(t,X(t),u(t)).
The proof for the existence and uniqueness is modified from ([41], Theorem 3). For t∈[t0,tf], define F[X(⋅)](t):=X0+Iαt0+[f(⋅,X(⋅),u(⋅)](t). Below, we show that F admits a unique fixed point on C([t0,tf];Rn). Then the existence and uniqueness of the solution in CnDαt0+ follows from (i) of Theorem 3.1.
Recall that ‖X(⋅)‖∞:=‖X(⋅)‖∞,n:=sups∈[t0,tf]|X(s)| for X(⋅)∈C([t0,tf];Rn). By a slight abuse of notation, for k∈N and X(⋅)∈C([t0,tf];Rn), define the Bielecki norm: ‖X(⋅)‖∞,k:=maxs∈[t0,tf]|e−k(s−t0)X(s)|. Note that ‖X(⋅)‖∞,k≤‖X(⋅)‖∞≤ek(tf−t0)‖X(⋅)‖∞,k.**
**This means that ‖⋅‖∞,k is equivalent to ‖⋅‖∞, and the space (C([t0,tf];Rn),‖⋅‖∞,k) is complete.
As f(⋅,X,u)∈L∞([t0,tf];Rn), we have F:C([t0,tf];Rn)→C([t0,tf];Rn). By Assumption 3.1, for any X(⋅),X′(⋅)∈C([t0,tf];Rn), it follows that
|e−k(t−t0)(F[X(⋅)](t)−F[X′(⋅)](t))|≤Le−k(t−t0)Γ(α)∫tt0(t−s)α−1|X(s)−X′(s)|ds. |
Note that |X(s)−X′(s)|≤ek(s−t0)‖X(⋅)−X′(⋅)‖∞,k for s∈[t0,tf]. Hence,
|e−k(t−t0)(F[X(⋅)](t)−F[X′(⋅)](t))|≤L‖X(⋅)−X′(⋅)‖∞,k1Γ(α)∫tt0(t−s)α−1ek(s−t)ds. |
By definition of the Gamma function Γ, ∫tt0(t−s)α−1e−k(t−s)ds≤∫∞0sα−1e−ksds=Γ(α)kα. This leads to ‖F[X(⋅)](⋅)−F[X′(⋅)](⋅)‖∞,k≤‖X(⋅)−X′(⋅)‖∞,kLkα. Then, by choosing k∈N such that Lkα<1, F:C([t0,tf];Rn)→C([t0,tf];Rn) is contraction. Note that this holds for any t∈[t0,tf], where Lkα<1 can be chosen independently of t. Therefore, by (i) of Theorem 3.1 and the contraction mapping theorem, (3.1) admits a unique solution of X(⋅)∈CnDαt0+.
Let X(⋅):=X(⋅;X0,u) and X′(⋅):=X(⋅;X′0,u′). By Assumption 3.1, we can show that |X(t)−X′(t)|≤|X0−X′0|+L∫tt0(t−s)α−1Γ(α)[|X(s)−X′(s)|+ρ(u(s),u′(s))]ds. By Gronwall's inequality (see [68,Theorem 1]) with b(t)=|X0−X′0|+L∫tt0(t−s)α−1Γ(α)ρ(u(s),u′(s))ds and z(t)=|X(t)−X′(t)|, it follows that |X(t)−X′(t)|≤b(t)+∫tt0∑∞k=1(LΓ(α))kΓ(kα)(t−s)kα−1b(s)ds. Hence, (3.3) follows. We can show (3.4) using a similar technique; thus completing the proof.
Lemma B.1. The following result holds:
supt∈[t0,tf]|Xϵ,δ(t)−Xϵ(t)δ−Zϵ(t)|=o(1), as δ↓0, | (B.1) |
where Zϵ is the variational equation related to the pair (Xϵ0,uϵ(⋅))∈Rn×U expressed by the left Caputo fractional differential equation (with ∂Xfϵ and ˆfϵ defined in Section 7.3):
{CDαt0+[Zϵ](t)=∂Xfϵ(t)Zϵ(t)+ˆfϵ(t),t∈(t0,tf],Zϵ(t0)=a∈Rn. | (B.2) |
Proof. Let Zϵ,δ(t):=Xϵ,δ(t)−Xϵ(t)δ, where by the Taylor series expansion, Zϵ,δ satisfies (with fϵ,δX defined in Section 7.3)
Zϵ,δ(t)=a+∫tt0(t−s)α−1Γ(α)(fϵ,δX(s)Zϵ,δ(s)+1Eδ(s)δˆfϵ(s))ds. |
Then (B.1) is equivalent to showing that limδ↓0supt∈[t0,tf]|Zϵ,δ(t)−Zϵ(t)|=0 for t∈[t0,tf]. Hereafter, K represents a generic constant, which may differ at different places.
By Theorem 3.1, it follows that supt∈[t0,tf]|Xϵ,δ(t)|≤K and supt∈[t0,tf]|Xϵ(t)|≤K. Then, by Theorem 3.1 and (7.10) (with some minor modification of the proof for Theorem 3.1),
supt∈[t0,tf]|Xϵ,δ(t)−Xϵ(t)|≤Kδ. | (B.3) |
By Theorem 3.1, (B.2) is equivalent to
Zϵ(t)=a+∫tt0(t−s)α−1Γ(α)(∂Xfϵ(s)Zϵ(s)+ˆfϵ(s))ds. |
As (B.2) is a linear Caputo fractional differential equation, by Theorem 3.1, (B.2) admits a unique solution. In addition, by Theorem 3.1, supt∈[t0,tf]|Zϵ(t)|≤K.
We have
Zϵ,δ(t)−Zϵ(t)=∫tt0(t−s)α−1Γ(α)fϵ,δX(s)(Zϵ,δ(s)−Zϵ(s))ds+∫tt0(t−s)α−1Γ(α)(fϵ,δX(s)−∂Xfϵ(s))Zϵ(s)ds+∫tt0(t−s)α−1Γ(α)(1Eδ(s)δ−1)ˆfϵ(s)ds. |
By Assumption 3.1, there is a constant K≥0 such that
|Zϵ,δ(t)−Zϵ(t)|≤|b1(t)+b2(t)|+K∫tt0(t−s)α−1Γ(α)|Zϵ,δ(s)−Zϵ(s)|ds, |
where
b1(t):=∫tt0|(t−s)α−1Γ(α)(fϵ,δX(s)−∂Xfϵ(s))Zϵ(s)|dsb2(t):=|∫tt0(t−s)α−1Γ(α)(1Eδ(s)δ−1)ˆfϵ(s)ds|. |
For b1, based on (7.10),
∫tt0|(t−s)α−1Γ(α)(fϵ,δX(s)−∂Xfϵ(s))Zϵ(s)|ds=∫[t0,t]∩Eδ|(t−s)α−1Γ(α)(fϵ,δX(s)−∂Xfϵ(s))Zϵ(s)|ds+∫[t0,t]∖Eδ|(t−s)α−1Γ(α)(fϵ,δX(s)−∂Xfϵ(s))Zϵ(s)|ds. |
In view of Assumption 3.1, and the fact that ˜d(uϵ,δ(⋅),uϵ(⋅))≤|Eδ|=δ(tf−t0) and supt∈[t0,tf]|Zϵ(t)|≤K by Theorem 3.1, the first term can be estimated by the dominated convergence theorem:
∫[t0,t]∩Eδ|(t−s)α−1Γ(α)(fϵ,δX(s)−∂Xfϵ(s))Zϵ(s)|ds=o(1),as δ↓0. |
Similarly, by (7.10) and (B.3), and the fact that ˜d(uϵ,δ(⋅),uϵ(⋅))≤|Eδ|=δ(tf−t0) and supt∈[t0,tf]|Zϵ(t)|≤K by Theorem 3.1, the second term can be estimated by the dominated convergence theorem:
∫[t0,t]∖Eδ|(t−s)α−1Γ(α)(fϵ,δX(s)−∂Xfϵ(s))Zϵ(s)|ds≤K∫[t0,t]∖Eδ(t−s)α−1Γ(α)|Xϵ,δ(s)−Xϵ(s)||Zϵ(s)|ds≤Kδ˜d(uϵ,δ(⋅),uϵ(⋅))=o(1),as δ↓0. |
These two estimates imply that for any t∈[t0,tf],
b1(t)=∫tt0|(t−s)α−1Γ(α)(fϵ,δX(s)−∂Xfϵ(s))Zϵ(s)|ds=o(1),as δ↓0. |
Then, it follows that limδ↓0supt∈[t0,tf]b1(t)=0. In addition, from Assumption 3.1 and Lemma C.1, it follows that supt∈[t0,tf]b2(t)≤δ=o(1).
By the preceding analysis, we have
|Zϵ,δ(t)−Zϵ(t)|≤K∫tt0(t−s)α−1Γ(α)|Zϵ,δ(s)−Zϵ(s)|ds+o(1),as δ↓0. |
In view of Fractional Gronwall's inequality (see [68], Theorem 1 and Corollary 2), it follows that supt∈[t0,tf]|Zϵ,δ(t)−Zϵ(t)|≤o(1)Eα[KΓ(α)(tf−t0)α]. This implies that as δ↓0, supt∈[t0,tf]|Zϵ,δ(t)−Zϵ(t)|→0, i.e., limδ↓0supt∈[t0,tf]|Zϵ,δ(t)−Zϵ(t)|=0, which shows (B.1). We complete the proof.
Lemma B.2. The following result holds: as δ↓0,
|J(Xϵ0+δa;uϵ,δ(⋅))−J(Xϵ0;uϵ(⋅))δ−ˆZϵ(tf)|=o(1),as δ↓0, |
where
ˆZϵ(tf)=∫tft0(tf−s)β−1Γ(β)(∂Xlϵ(s)Zϵ(s)+ˆlϵ(s))ds+∂X0mϵa+∂XmϵZϵ(tf). |
Proof. Recall Zϵ,δ(t)=Xϵ,δ(t)−Xϵ(t)δ. By the Taylor series expansion,
1δ(J(t0,Xϵ0+δa;uϵ,δ(⋅))−J(t0,Xϵ0;uϵ(⋅)))−ˆZϵ(tf)=∫tft0(tf−s)β−1Γ(β)lϵ,δX(s)(Zϵ,δ(s)−Zϵ(s))ds+∫tft0(tf−s)β−1Γ(β)(1Eδ(s)δ−1)ˆlϵ(s)ds+∫tft0(tf−s)β−1Γ(β)(lϵ,δX(s)−∂Xlϵ(s))Zϵ(s)ds+(mϵ,δX0(tf)−∂X0mϵ)a+mϵ,δX(tf)(Zϵ,δ(tf)−Zϵ(tf))+(mϵ,δX(tf)−∂Xmϵ))Zϵ(tf). |
Note that limδ↓0|Zϵ,δ(s)−Zϵ(s)|=0 for s∈[t0,tf] by Lemma B.1. From Assumption 3.2 and Lemma C.1, it follows that |∫tft0(tf−s)β−1Γ(β)(1Eδ(s)δ−1)ˆlϵ(s)ds|≤δ. Then using a similar proof technique of Lemma B.1, we can show the desired result. This completes the proof.
Lemma B.3. The following results hold:
(i) limϵ↓0ˆd((¯X0;¯u(⋅)),(Xϵ0,uϵ(⋅)))=0;
(ii) supt∈[t0,tf]|Zϵ(t)−Z(t)|=o(1) and |ˆZϵ(tf)−ˆZ(tf)|=o(1) as ϵ↓0, where Z is the left Caputo fractional differential equation given by
{CDαt0+[Z](t)=∂X¯f(t)Z(t)+ˆf(t),t∈(t0,tf],Z(t0)=a∈Rn, | (B.4) |
and ˆZ is given by
ˆZ(tf)=∫tft0(tf−s)β−1Γ(β)(∂X¯l(s)Z(s)+ˆl(s))ds+∂X0¯ma+∂X¯mZ(tf). |
Proof. Part (i) follows from (7.7). The proof for part (ii) is similar to that for Lemmas B.1 and B.2. We complete the proof.
Lemma C.1. For any α∈(0,1), δ∈(0,1) and g∈L1([t0,tf];R), there is an Eδ∈Eδ, where Eδ is defined in (7.9), such that
supt∈[t0,tf]|∫tt0(t−s)α−1Γ(α)(1Eδ(s)δ−1)g(s)ds|≤δ. |
Proof. Note that as g∈L1([t0,tf];R), there is a simple function g(n), which is an approximation of g in L1 such that g(n)→g in L1 as n→∞ ([66], Lemma 4.2.1). For any ϵ≥0, consider
supt∈[t0,tf]|∫tt0(t−s)α−1Γ(α)(1Eδ(s)δ−1)g(s)ds|≤supt∈[t0,tf]|∫tt0(t−s)α−1Γ(α)+ϵ(1Eδ(s)δ−1)g(n)(s)ds|+supt∈[t0,tf]|∫tt0(t−s)α−1Γ(α)(1Eδ(s)δ−1)g(n)(s)ds−∫tt0(t−s)α−1Γ(α)+ϵ(1Eδ(s)δ−1)g(n)(s)ds|+supt∈[t0,tf]|∫tt0(t−s)α−1Γ(α)(1Eδ(s)δ−1)(g(s)−g(n)(s))ds|≜supt∈[t0,tf]H(1)+supt∈[t0,tf]H(2)+supt∈[t0,tf]H(3). |
Note that \frac{(t-s)^{\alpha-1}}{\Gamma(\alpha) + \epsilon} g^{(n)} (s) is continuous in t with s < t \leq t_f and is an element of L^1 in both variables (t, s) with t_0 \leq s < t \leq t_f . Then, by ([43], Corollary 3.8, Chapter 4), there is an E_{\delta} \in \mathcal{E}_{\delta} such that
\begin{align*} \sup\limits_{t \in [t_0, t_f]} H^{(1)} = \sup\limits_{t \in [t_0, t_f]} \Bigl | \int_{t_0}^{t} \frac{(t-s)^{\alpha-1}}{\Gamma(\alpha) + \epsilon} \Bigl ( \frac{ \mathbb{1}_{E_{\delta}}(s)}{\delta} -1 \Bigr ) g^{(n)}(s) {\rm{d}} s \Bigr | \leq \delta. \end{align*} |
Also, from ([66], Lemma 4.2.1), as g^{(n)} is an approximation of g , for any t \in [t_0, t_f] ,
\begin{align*} H^{(3)} \leq C \Bigl (1 + \frac{1}{\delta} \Bigr ) \int_{t_0}^{t} \frac{(t-s)^{\alpha-1}}{\Gamma(\alpha)} |g(s) - g^{(n)}(s)| {\rm{d}} s \rightarrow 0, \; \text{as } ~~ n \rightarrow \infty .\end{align*} |
By the dominated convergence theorem, \sup_{t \in [t_0, t_f]} H^{(2)} = o(1) as \epsilon \downarrow 0 . Note that these estimates are independent of t \in [t_0, t_f] . Hence, by combining the preceding estimates, as \epsilon \downarrow 0 , the result follows. This completes the proof.
[1] | A. A. Kilbas, H. M. Srivastava, J. J. Trujillo, Theory and applications of fractional differential equations, Elsevier, 2006. |
[2] | K. Diethelm, The analysis of fractional differential equations, Springer, 2010. |
[3] | A. B. Malinowska, T. Odzijewicz, D. F. Torres, Advanced methods in the fractional calculus of variations, Springer, 2015. https://doi.org/10.1007/978-3-319-14756-7_3 |
[4] | Y. Zhou, Basic theory of fractional differential equations, World Scientific Publishing, 2014. |
[5] |
O. P. Agrawal, Some generalized fractional calculus operators and their applications in integral equations, Fract. Calc. Appl. Anal., 15 (2012), 700–711. https://doi.org/10.2478/s13540-012-0047-7 doi: 10.2478/s13540-012-0047-7
![]() |
[6] |
O. P. Agrawal, A general formulation and solution scheme for fractional optimal control problems, Nonlinear Dyn., 38 (2004), 323–337. https://doi.org/10.1007/s11071-004-3764-6 doi: 10.1007/s11071-004-3764-6
![]() |
[7] |
R. Almeida, D. F. Torres, Calculus of variations with fractional derivatives and fractional integrals, Appl. Math. Lett., 22 (2009), 1816–1820. https://doi.org/10.1016/j.aml.2009.07.002 doi: 10.1016/j.aml.2009.07.002
![]() |
[8] |
O. Agrawal, O. Defterli, D. Baleanu, Fractional optimal control problems with several state and control variables, J. Vib. Control, 16 (2010), 1967–1976. https://doi.org/10.1177/1077546309353361 doi: 10.1177/1077546309353361
![]() |
[9] |
R. K. Biswas, S. Sen, Free final time fractional optimal control problems, J. Franklin. I., 351 (2014), 941–951. https://doi.org/10.1016/j.jfranklin.2013.09.024 doi: 10.1016/j.jfranklin.2013.09.024
![]() |
[10] |
I. Matychyn, V. Onyshchenko, Time-optimal control of fractional-order linear systems, Fract. Calc. Appl. Anal., 18 (2015), 687–696. https://doi.org/10.1515/fca-2015-0042 doi: 10.1515/fca-2015-0042
![]() |
[11] |
R. Kamocki, On the existence of optimal solutions to fractional optimal control problems, Appl. Math. Comput., 235 (2014), 94–104. https://doi.org/10.1016/j.amc.2014.02.086 doi: 10.1016/j.amc.2014.02.086
![]() |
[12] |
M. I. Gomoyunov, Dynamic programming principle and Hamilton-Jacobi-Bellman equations for fractional-order systems, SIAM J. Control Optim., 58 (2020), 3185–3211. https://doi.org/10.1137/19M1279368 doi: 10.1137/19M1279368
![]() |
[13] |
M. I. Gomoyunov, Minimax solutions of Hamilton-Jacobi equations with fractional coinvariant derivatives, ESAIM: COCV, 28 (2022), 23. https://doi.org/10.1051/cocv/2022017 doi: 10.1051/cocv/2022017
![]() |
[14] |
R. Kamocki, M. Majewski, Fractional linear control systems with Caputo derivative and their optimization, Optim. Control Appl. Met., 36 (2015), 953–967. https://doi.org/10.1002/oca.2150 doi: 10.1002/oca.2150
![]() |
[15] |
Y. Ding, Z. Wang, H. Ye, Optimal control of a fractional-order HIV-Immune system with memory, IEEE T. Contr. Syst. T., 20 (2012), 763–769. https://doi.org/10.1109/TCST.2011.2153203 doi: 10.1109/TCST.2011.2153203
![]() |
[16] |
W. Li, S. Wang, V. Rehbock, Numerical solution of fractional optimal control, J. Optim. Theory Appl., 180 (2019), 556–573. https://doi.org/10.1007/s10957-018-1418-y doi: 10.1007/s10957-018-1418-y
![]() |
[17] |
C. Liu, W. Sun, X. Yi, Optimal control of a fractional smoking systems, J. Ind. Manag. Optim., 19 (2023), 2936–2954. https://doi.org/10.3934/jimo.2022071. doi: 10.3934/jimo.2022071
![]() |
[18] |
Z. Gong, C. Liu, K. L. Teo, S. Wang, Y. Wu, Numerical solution of free final time fractional optimal control problems, Appl. Comput. Math., 405 (2021), 126270. https://doi.org/10.1016/j.amc.2021.126270 doi: 10.1016/j.amc.2021.126270
![]() |
[19] |
P. Rahimkhani, Y. Ordokhani, Numerical investigation of distributed-order fractional optimal control problems via Bernstein wavelets, Optim. Contr. Appl. Met., 42 (2021), 355–373. https://doi.org/10.1002/oca.2679 doi: 10.1002/oca.2679
![]() |
[20] |
M. Bergounioux, L. Bourdin, Pontryagin maximum principle for general Caputo fractional optimal control problems with Bolza cost and terminal constraints, ESAIM: COCV, 26 (2020), 35. https://doi.org/10.1051/cocv/2019021 doi: 10.1051/cocv/2019021
![]() |
[21] |
X. L. Ding, I. Area, J. Nieto, Controlled singular evolution equations and Pontryagin type maximum principle with applications, Evol. Equ. Control The., 11 (2022), 1655–1679. https://doi.org/10.3934/eect.2021059 doi: 10.3934/eect.2021059
![]() |
[22] |
P. Lin, J. Yong, Controlled singular Volterra integral equations and Pontryagin maximum principle, SIAM J. Control Optim., 58 (2020), 136–164. https://doi.org/10.1137/19M124602X doi: 10.1137/19M124602X
![]() |
[23] |
M. I. Gomoyunov, Optimal control problems with a fixed terminal time in linear fractional-order systems, Arch. Control Sci., 30 (2020), 721–744. https://doi.org/10.24425/acs.2020.135849 doi: 10.24425/acs.2020.135849
![]() |
[24] |
H. Tajadodi, A. Khan, J. F. Gomez-Aguilar, H. Khan, Optimal control problems with Atangana-Baleanu fractional derivative, Optim. Contr. Appl. Met., 42 (2021), 96–109. https://doi.org/10.1002/oca.2664 doi: 10.1002/oca.2664
![]() |
[25] |
Z. D. Jelicic, N. Petrovacki, Optimality conditions and a solution scheme for fractional optimal control problems, Struct. Multidisc. Optim., 38 (2009), 571–581. https://doi.org/10.1007/s00158-008-0307-7 doi: 10.1007/s00158-008-0307-7
![]() |
[26] |
G. S. F. Frederico, D. F. Torres, Fractional optimal control in the sense of Caputo and the fractional Noether's theorem, Int. Math. Forum, 3 (2008), 479–493. https://doi.org/10.48550/arXiv.0712.1844 doi: 10.48550/arXiv.0712.1844
![]() |
[27] |
R. Almeida, D. F. Torres, Necessary and sufficient conditions for the fractional calculus of variations with Caputo derivatives, Commun. Nonlinear Sci., 16 (2011), 1490–1500. https://doi.org/10.1016/j.cnsns.2010.07.016 doi: 10.1016/j.cnsns.2010.07.016
![]() |
[28] |
T. L. Guo, The necessary conditions of fractional optimal control in the sense of Caputo, J. Optim. Theory Appl., 156 (2013), 115–126. https://doi.org/10.1007/s10957-012-0233-0 doi: 10.1007/s10957-012-0233-0
![]() |
[29] |
R. Kamocki, Pontryagin maximum principle for fractional ordinary optimal control problems, Math. Method. Appl. Sci., 37 (2014), 1668–1686. https://doi.org/10.1002/mma.2928 doi: 10.1002/mma.2928
![]() |
[30] |
S. Pooseh, R. Almeida, D. F. Torres, Fractional order optimal control problems with free terminal time, J. Ind. Manag. Optim., 10 (2014), 363–381. https://doi.org/10.3934/jimo.2014.10.363 doi: 10.3934/jimo.2014.10.363
![]() |
[31] |
H. M. Ali, F. L. Pereira, S. M. A. Gamma, A new approach to the Pontryagin maximum principle for nonlinear fractional optimal control problems, Math. Method. Appl. Sci., 39 (2016), 3640–3649. https://doi.org/10.1002/mma.3811 doi: 10.1002/mma.3811
![]() |
[32] |
R. Almeida, R. Kamocki, A. B. Malinowska, T. Odzijewicz, On the necessary optimality conditions for the fractional Cucker-Smale optimal control problem, Commun. Nonlinear Sci., 96 (2021), 105678. https://doi.org/10.1016/j.cnsns.2020.105678 doi: 10.1016/j.cnsns.2020.105678
![]() |
[33] |
S. S. Yusubov, E. N. Mahmudov, Optimality conditions of singular controls for systems with Caputo fractional derivatives, J. Ind. Manag. Optim., 19 (2023), 246–264. https://doi.org/10.3934/jimo.2021182 doi: 10.3934/jimo.2021182
![]() |
[34] |
S. Wang, W. Li, C. Liu, On necessary optimality conditions and exact penalization for a constrained fractional optimal control problem, Optim. Contr. Appl. Met., 43 (2022), 1096–1108. https://doi.org/10.1002/oca.2877 doi: 10.1002/oca.2877
![]() |
[35] |
M. Habibli, M. H. Nouri Skandari, Fractional Chebyshev pseudospectral method for fractional optimal control problems, Optim. Contr. Appl. Met., 40 (2019), 558–572. https://doi.org/10.1002/oca.2495 doi: 10.1002/oca.2495
![]() |
[36] |
M. M. Raja, V. Vijayakumar, A. Shukla, K. S. Nisar, N. Sakthivel, K. Kaliraj, Optimal control and approximate controllability for fractional integrodifferential evolution equations with infinite delay of order \gamma \in (1, 2), Optim. Contr. Appl. Met., 43 (2022), 996–1019. https://doi.org/10.1002/oca.2867 doi: 10.1002/oca.2867
![]() |
[37] |
M. de Pinho, M. Margarida, A. Ferreira, G. Smirnov, Optimal control problem with state constraints via penalty functions, Syst. Control Lett., 188 (2024), 105816. https://doi.org/10.1016/j.sysconle.2024.105816 doi: 10.1016/j.sysconle.2024.105816
![]() |
[38] |
E. S. Baranovskii, R. V. Brizitskii, Z. Y. Saritskaia, Optimal control problems for the reaction-diffusion-convection equation with variable coefficients, Nonlinear Anal. Real, 75 (2024), 103979. https://doi.org/10.1016/j.nonrwa.2023.103979 doi: 10.1016/j.nonrwa.2023.103979
![]() |
[39] |
I. Malmir, New pure multi-order fractional optimal control problems with constraints: QP and LP methods, ISA Trans., 153 (2024), 155–190. https://doi.org/10.1016/j.isatra.2024.08.003 doi: 10.1016/j.isatra.2024.08.003
![]() |
[40] |
I. Malmir, Novel closed-loop controllers for fractional linear quadratic time-varying systems, Numer. Algebr. Control, 14 (2024), 366–403. https://doi.org/10.3934/naco.2022032 doi: 10.3934/naco.2022032
![]() |
[41] |
L. Bourdin, Cauchy-Lipschitz theory for fractional multi-order dynamics: State-transition matrices, Duhamel formulas and duality theorems, Differ. Integral Equ., 31 (2018), 559–594. https://doi.org/10.57262/die/1526004031 doi: 10.57262/die/1526004031
![]() |
[42] |
M. I. Gomoyunov, On representation formulas for solutions of linear differential equations with Caputo fractional derivatives, Fract. Calc. Appl. Anal., 23 (2020), 1141–1160. https://doi.org/10.1515/fca-2020-0058 doi: 10.1515/fca-2020-0058
![]() |
[43] | X. Li, J. Yong, Optimal control theory for infinite dimensional systems, Birkhauser, 1995. https://doi.org/10.1007/978-1-4612-4260-4 |
[44] |
A. Lotfi, M. Dehghan, S. A. Yousefi, A numerical technique for solving fractional optimal control problems, Comput. Math. Appl., 62 (2011), 1055–1067. https://doi.org/10.1016/j.camwa.2011.03.044 doi: 10.1016/j.camwa.2011.03.044
![]() |
[45] |
N. Singha, Implementation of fractional optimal control problems in real-world applications, Fract. Calc. Appl. Anal., 23 (2020), 1783–1796. https://doi.org/10.1515/fca-2020-0088 doi: 10.1515/fca-2020-0088
![]() |
[46] |
Y. Oh, J. Moon, The infinite-dimensional Pontryagin maximum principle for optimal control problems of fractional evolution equations with endpoint state constraints, AIMS Math., 9 (2024), 6109–6144. https://doi.org/10.3934/math.2024299 doi: 10.3934/math.2024299
![]() |
[47] | J. Yong, X. Y. Zhou, Stochastic controls: Hamiltonian systems and HJB equations, Springer, 1999. https://doi.org/10.1007/978-1-4612-1466-3 |
[48] | L. Bourdin, Note on Pontryagin maximum principle with running state constraints and smooth dynamics-proof based on the Ekeland variational principle, 2016. https://doi.org/10.48550/arXiv.1604.04051 |
[49] |
P. Bettiol, L. Bourdin, Pontryagin maximum principle for state constrained optimal sampled-data control problems on time scales, ESAIM: COCV, 27 (2020), 51. https://doi.org/10.1051/cocv/2021046 doi: 10.1051/cocv/2021046
![]() |
[50] |
A. V. Dmitruk, N. P. Osmolovskii, Necessary conditions for a weak minimum in a general optimal control problem with integral equations on a variable time interval, Math. Control Relat. F., 7 (2017), 507–535. https://doi.org/10.3934/mcrf.2017019 doi: 10.3934/mcrf.2017019
![]() |
[51] |
L. Bourdin, G. Dhar, Optimal sampled-data controls with running inequality state constraints: Pontryagin maximum principle and bouncing trajectory phenomenon, Math. Program., 191 (2022), 907–951. https://doi.org/10.1007/s10107-020-01574-2. doi: 10.1007/s10107-020-01574-2
![]() |
[52] | E. M. Stein, R. Shakarchi, Real analysis: Measure theory, integration, and Hilbert spaces, Princeton: Princeton University Press, 2005. https://doi.org/10.1515/9781400835560 |
[53] |
R. F. Hartl, S. P. Sethi, R. G. Vickson, A survey of the maximum principle for optimal control problems with state constraints, SIAM Rev., 37 (1995), 181–218. https://doi.org/10.1137/1037043 doi: 10.1137/1037043
![]() |
[54] | R. Vinter, Optimal control, Springer Science & Business Media, 2000. |
[55] |
C. Li, A. Chen, J. Ye, Numerical approaches to fractional calculus and fractional ordinary differential equation, J. Comput. Phys., 230 (2011), 3352–3368. https://doi.org/10.1016/j.jcp.2011.01.030 doi: 10.1016/j.jcp.2011.01.030
![]() |
[56] |
M. Cai, C. Li, Numerical approaches to fractional integrals and derivatives: A review, Mathematics, 8 (2020), 43. https://doi.org/10.3390/math8010043 doi: 10.3390/math8010043
![]() |
[57] | C. Li, M. Cai, Theory and numerical approximations of fractional integrals and derivatives, Society for Industrial and Applied Mathematics, 2019. |
[58] |
F. M. Alharbi, A. M. Zidan, M. Naeem, R. Shah, K. Nonlaopon, Numerical investigation of fractional-order differential equations via \psi-Haar-Wavelet method, J. Funct. Space., 2021 (2021), 3084110. https://doi.org/10.1155/2021/3084110 doi: 10.1155/2021/3084110
![]() |
[59] | M. Altaf Khan, A. Atangana, Numerical methods for fractal-fractional differential equations and engineering, CRC Press, 2023. https://doi.org/10.1201/9781003359258 |
[60] | A. Ruszczynski, Nonlinear optimization, Princeton University Press, 2006. https://doi.org/10.1515/9781400841059 |
[61] | T. M. Flett, J. S. Pym, Differential analysis: differentiation, differential equations, and differential inequalities, 1980. |
[62] | F. H. Clarke, Optimization and nonsmooth analysis, Society for industrial and Applied Mathematics, 1990. |
[63] | J. B. Conway, A course in functional analysis, New York: Springer, 1985. https://doi.org/10.1007/978-1-4757-3828-5 |
[64] | B. S. Mordukhovich, Variational analysis and generalized differentiation I, Springer, 2006. https://doi.org/10.1007/3-540-31247-1 |
[65] | I. Ekeland, On the variational principle, J. Math. Anal. Appl., 47 (1974), 324–353. https://doi.org/10.1016/0022-247X(74)90025-0 |
[66] | V. I. Bogachev, Measure theory, Springer, 2000. https://doi.org/10.1007/978-3-540-34514-5 |
[67] |
Y. Sun, Maximum principle for mean-field controlled systems driven by a fractional Brownian motion, Optim. Control Appl. Method., 44 (2023), 3282–3305. https://doi.org/10.1002/oca.3039 doi: 10.1002/oca.3039
![]() |
[68] |
H. Ye, J. Gas, Y. Ding, A generalized Gronwall inequality and its application to a fractional differential equation, J. Math. Anal. Appl., 328 (2007), 1075–1081. https://doi.org/10.1016/j.jmaa.2006.05.061 doi: 10.1016/j.jmaa.2006.05.061
![]() |