AN APPROACH TO SECOND-ORDER NECESSARY CONDITIONS FOR MULTIPARAMETER
DISCRETE INCLUSIONS
AURELIAN CERNEA
We study an optimization problem given by a multiparameter discrete inclusion that describes the Roesser model. An approach concerning second-order optima- lity conditions is proposed.
AMS 2000 Subject Classification: 49J30.
Key words: tangent cone, discrete inclusion, necessary optimality conditions.
1. INTRODUCTION Consider the problem
(1.1) minimize g(xNN)
over the solutions of the multiparameter discrete inclusion
(1.2) xij ∈Fij(xij−1, xi−1j, xi−1j−1), i= 0,1, . . . , N, j = 0,1, . . . , N, with endpoint constraints of the form
(1.3) xNN ∈XN,
where Fij(·,·,·) : Rn×Rn×Rn → P(Rn), i= 0,1, . . . , N, j = 0,1, . . . , N are given set-valued maps,XN ⊂Rn and g(·) :Rn →Rare given.
There are several papers devoted to first-order necessary optimality con- ditions for this problem ([6], [7], [8], [9] etc.). The aim of the present paper is to develop an approach concerning second-order necessary optimality conditions for problem (1.1)–(1.3). The general idea is to consider our problem as the problem of minimizing the terminal payoff on the intersection of the (known) target set with the (unknown) reachable set and to use a general result from nonsmooth analysis. This general (abstract) optimality condition was formu- lated for the first time by Zheng [10], but this result (Theorem 2.2 below) is, in fact, an obvious consequence of Theorems 6.3.1, 6.6.2, 4.7.4, Proposition 6.2.4 and Corollary 4.3.5 in [1].
REV. ROUMAINE MATH. PURES APPL.,52(2007),5, 529–538
In order to apply the general abstract optimality conditions (namely, Theorem 2.2 below) we must check a certain constraint qualification, so we are naturally led to study first and second order approximations of the reachable set along optimal solutions.
Let us mention that this idea has been already used in [3], [4], [5] and [10] to obtain second-order necessary optimality conditions for problems given by differential inclusions, hyperbolic differential inclusions and other classes of discrete inclusions.
The paper is organized as follows. In Section 2 we present the notation and definitions to be used in the sequel while in Section 3 we present our main results.
2. PRELIMINARIES
Since the reachable set that appears in optimization problems is, in gen- eral, neither a differentiable manifold nor a convex set, its infinitesimal prop- erties may be characterized only by tangent cones in a generalized sense, ex- tending the classical concepts of tangent cones in Differential Geometry and Convex Analysis, respectively. From the rather large number of tangent cones in the literature (e.g. [1]) we only use the following concepts.
LetX⊂Rn and x∈cl(X) (the closure of X).
Definition 2.1. (a) The quasitangent (intermediate) cone to X at x is defined as
QxX ={v∈Rn; ∀sm→0+, ∃vm→v: x+smvm∈X};
(b) the second-order quasitangent setto X at x relative to v ∈QxX is defined as
Q2(x,v)X={w∈Rn; ∀sm→0+, ∃wm→w: x+smv+s2mwm ∈X}; (c)Clarke’s tangent cone to X at x is defined as
CxX=
v∈Rn; ∀(xm, sm)→(x,0+), xm ∈X, ∃ym ∈X : ym−xm sm →v
.
For equivalent definitions and for several properties of these cones we refer to [1]. We recall that unlikeQxX, Clarke’s tangent cone CxX is convex and one hasCxX⊂QxX.
We denote byC+ the positive dual cone of C ⊂Rn, namely, C+={q ∈Rn; q, v ≥0, ∀v∈C}.
The negative dual cone ofC⊂Rn isC−=−C+.
As was often remarked, the geometric interpretation of the classical (Fr´echet) derivative suggests the possibility of the introduction of general- ized differentiability concepts corresponding to each type of tangent cone (to the graph, to the epigraph or to the subgraph of the function) but, of course, not all these concepts are equally important. In what follows, for a mapping g(·) : X ⊂ Rn → R which is not differentiable, we shall only use the first and second order uniform lower Dini derivative. We refer to [1] for the main properties of such derivatives:
D↑g(x;v) = lim inf
(v,θ)→(v,0+)
g(x+θv)−g(x)
θ ,
D↑2g(x, v;w) = lim inf
(w,θ)→(w,0+)
g(x+θv+θ2w)−g(x)−θD↑g(x;v)
θ2 .
Wheng(·) is of class C2 one has
D↑g(x, v) =g(x)Tv, D2↑g(x, v;w) =g(x)Tz+1
2vTg(x)v.
The key tool in the proof of our main result is the abstract optimality condition, below.
Theorem 2.2 ([10]). Let g : Rn → R be a locally Lipschitz function at z ∈ Rn and S1, S2 nonempty subsets of Rn containing z. If z solves the minimization problem
minimize g(x) over allx∈S1∩S2 and also satisfies the constraint qualification
(CzS1)−∩(CzS2)+={0}, then we have the first-order necessary condition
D↑g(z;v) ≥0 ∀v∈QzS1∩QzS2.
Furthermore, if equality holds for some v0, then we have the second-order necessary condition
D↑2g(z, v0;w)≥0 ∀w∈Q2(z,v0)S1∩Q2(z,v0)S2.
Corresponding to each type of tangent cone, sayτxX, one may introduce (e.g., [1]) a set-valued directional derivative of a multifunction G(·) : X ⊂ Rn → P(Rn) (in particular of a single-valued mapping) at a point (x, u) ∈ graph(G) as
τuG(x, ξ) ={ν ∈Rn; (ξ, ν)∈τ(x,u)graph(G)}, ν∈τxX.
This first-order derivative may be characterized, equivalently, by graphτuG(x,·) =τ(x,u)(graphG(·)).
If the set-valued mapG(·) is Lipschitz, i.e., there existsL >0 such that G(x1)⊂G(x2) +Lx1−x2B ∀x1, x2 ∈X,
whereB denotes the closed unit ball inRn, then the first order quasitangent derivative is given by (e.g., [1])
QuG(x;ξ) =
ν ∈Rn; lim
θ→0+
1
θd(u+θν, G(x+θξ)) = 0
.
Similarly, one may define (e.g., [1]) second-order directional deivatives of a set-valued mapG(·). For example the second-order quasitangent derivative of G at (x, u) relative to (y, v) ∈ Q(x,u)(graph(G(·)) is the set-valued map Q2(u,v)G(x, y,·) defined by
graphQ2(u,v)G(x, y;.) =Q2((x,u),(y,v))(graphG(·)).
We recall that a set-valued map A(·) : Rn → P(Rn) is called a closed (respectively, convex) process if graph(A(·)) is a closed (respectively, con- vex) cone.
For the basic properties of convex processes we refer to [1], but we shall only use here the above definition.
If G(·) : X ⊂ Rn → P(Rn) is a given set-valued map and (x, u) ∈ graph(G), as a closed convex process one may take the Clarke directional derivative A(·) =CuG(x,·) of G(·) at (x, u).
The adjoint processA∗ :Rn→ P(Rn) of the closed convex process Ais defined by
A∗(p) ={q ∈Rn;q, v ≤ p, v ∀(v, v)∈graphA(·)}. Denote by SF the solution set of inclusion (1.2), i.e.,
SF :=
x= (x0, x1, . . . , xN); xi= (xi0, xi1, . . . , xiN), xij ∈Rn, i, j = 0,1, . . . , N, x is a solution of (1.2), xij = 0 if i <0 or j <0
. and byRFN :={xNN; x∈SF}the reachable set of inclusion (1.2).
In what follows we consider a solution x = (x0, x1, . . . , xN) ∈ SF of problem (1.1)–(1.3) and shall assume the following hypotheses.
Hypothesis2.3. i)XN ⊂Rn is a closed set.
ii) The set-valued maps Fij(·) have nonempty closed values ∀i, j ∈ {0, 1, . . . , N} and there exists L >0 such that Fij(·) is Lipschitz with Lipschitz constant L,∀i, j∈ {0,1, . . . , N}.
Hypothesis 2.4. There exists a family Aij(·) : R3n → P(Rn), i, j = 0,1, . . . , N, of closed convex processes such that
Aij(u, v, w) ⊂QxijFij((xij−1, xi−1j, xi−1j−1); (u, v, w))
∀u, v, w∈Rn,∀i, j∈ {0,1, . . . , N}.
Let A0 ⊂Qx00F00(0,0,0) be a closed convex cone. With problem (1.2) we associate the linearized problem
(2.1) wij∈Aij(wij−1, wi−1j, wi−1j−1), w00∈A0, i, j= 0,1, . . . , N, i+j >0 with the boundary conditionswij = 0 fori <0 or j <0.
Denote bySAthe solution set of inclusion (2.1) and byRNA the reachable set of inclusion (2.1).
In the study of our optimization problem we need the duality result below.
Lemma 2.5 ([9]). Assume that Hypothesis 2.4 is verified and let r(·) : Rn→ P(Rn) be the set-valued map defined by
r(α) :={wNN; w= (w0, . . . , wN)is a solution of (2.1), w00=α}. Then, for all b∈Rn,
r∗(b) =
u101+u210+u311; (u1ij, u2ij, u3ij)∈A∗ij(u1ij+1+u2i+1j+u3i+1j+1), i, j= 0,1, . . . , N, 0< i+j <2N, (u1NN, u2NN, u3NN)∈A∗NN(b)
.
Corollary 2.6. Assume that Hypothesis 2.4 is satisfied and A0 ⊂ Qx00F00(0,0,0) is a closed convex cone. Then
(RNA)+⊂
q∈Rn; ∃u1ij, u2ij, u3ij ∈Rn such that(u1ij, u2ij, u3ij)∈
∈A∗ij(u1ij+1+u2i+1j+u3i+1j+1), i, j = 0,1, . . . , N, 0< i+j <2N, (u1NN, u2NN, u3NN)∈A∗NN(q), u101+u210+u311∈A+0
.
Proof. Obviously,RNA =r(A0), where r(·) was defined in the statement of Lemma 2.5. Therefore, (RNA)+= (r(A0))+ =r∗−1(A+0). If b∈(RNA)+ then it follows thatr∗(b)∈A+0 and the corollary now follows from Lemma 2.5.
Remark 2.7. Hypothesis 2.4 that appears in Lemma 2.5, is satisfied if we take
Aij(u, v, w) :=CxijFij((xij−1, xi−1j, xi−1j−1); (u, v, w))
∀u, v, w∈Rn,∀i, j∈ {0,1, . . . , N}.
3. THE MAIN RESULTS
We prove first an approximation of the reachable setRNF at xN.
Theorem3.1. Assume that Hypothesis2.3is satisfied and denote byRNQ the reachable set of the discrete inclusion
(3.1) wij∈QxijFij((xij−1, xi−1j, xi−1j−1); (wij−1, wi−1j, wi−1j−1)), i+j >0 w00∈Qx00F00(0,0,0), wij = 0 for i <0 or j <0.
ThenRNQ ⊂QxNNRNF.
Proof. Let ¯w∈RNQ and sk→0+. It follows that there exists a solution (w0, w1, . . . , wN) to (3.1) such that ¯w=wNN.
In particular, w00 ∈ Qx00F00(0,0,0) and, therefore, there exists w00k → w00 such thatx00+skw00k ∈F00(0,0,0) ∀k∈N.
On the other hand,w01∈Qx01F01((x00,0,0); (w00,0,0)) and by the defi- nition of the quasitangent derivative ofF01 there exists ( ˜w01k ,w˜k00,w˜k−11,w˜−10k )
→(w01, w00,0,0) such that
x01+skw˜k01∈F01(x00+skw˜k00, skw˜k−11, skw˜k−10) ∀k∈N.
By the Lipschitz property of the set-valued map F01(·), we can write x01+skw˜01k ∈F01(x00+skwk00,0,0)+skL(w˜k00−wk00+w˜k−11+w˜k−10)B
∀k∈N. Thus, there exists b1k∈B such that
x01+sk[ ˜w01k −Lb1k(w˜00k −w00k +w˜−11k +w˜k−10)]∈F01(x00+skw00k ,0,0)
∀k∈N, and if we definew01k := ˜w01k −Lb1k(w˜k00−w00k +w˜−11k +w˜k−10), then we have w01k →w01 ask→ ∞and
x01+skwk01∈F01(x00+skw00k ,0,0) ∀k∈N.
By repeating this construction forj= 2, . . . , N we find that there exists w0jk ∈Rn such that w0jk →w0j ask→ ∞ and
x0j+skwk0j ∈F0j(x0j−1+skw0j−1k ,0,0) ∀k∈N, ∀j= 2, . . . , N.
Next, we repeat this construction for i ∈ {1,2, . . . , N} fixed and j ∈ {0,1, . . . , N} and find that there exists wkij ∈ Rn such that wkij → wij as k→ ∞ and
xij+skwkij∈Fij(xij−1+skwkij−1, xi−1j+skwki−1j, xi−1j−1+skwi−1j−1k )∀k∈N. In particular, for sk → 0+ there exists wkNN → wNN such that xNN + skwkNN ∈RNF, i.e., w=wNN ∈QxNNRNF and the proof is complete.
Another first-order approximation of the reachable setRFN atxN can be obtained in terms of the variational inclusion defined by the Clarke derivative of the set valued map.
Theorem3.2. Assume that Hypothesis2.3is satisfied and denote byRNC the reachable set of the discrete inclusion
(3.2) wij∈CxijFij((xij−1, xi−1j, xi−1j−1); (wij−1, wi−1j, wi−1j−1)), i+j >0 w00∈Cx00F00(0,0,0), wij = 0 for i <0 or j <0.
ThenRNC ⊂CxNNRNF.
The proof of Theorem 3.2 can be done using the same arguments em- ployed to prove Theorem 3.1.
In order to apply Theorem 2.2 for our problem (1.1)–(1.4) we need to know the second-order quasitangent set to the reachable setRFN at xNN.
Theorem 3.3. Assume that Hypothesis 2.3 is satisfied, let y = (y0, y1, . . . , yN) satisfy (3.1) and let R2Q denote the reachable set of the discrete inclusion
(3.3)
vij ∈Q2(x
ij,yij)Fij((xij−1, xi−1j, xi−1j−1),(yij−1, yi−1j, yi−1j−1); (vij−1, vi−1j, vi−1j−1)), i+j >0
v00∈Q2(x
00,y00)F00(0,0,0), vij = 0 for i <0 or j <0.
ThenR2Q ⊂Q2(x
NN,yNN)RNF.
Proof. Letv ∈R2Q and tk →0+. It follows that there exists a solution (v0, v1, . . . , vN) to (3.3) such thatv=vNN.
In particular,v00∈Q2(x
00,y00)F00(0,0,0) and therefore there existsvk00→ v00 such that x00+tky00+t2kvk00∈F00(0,0,0).
On the other hand, v01 ∈ Q2(x
01,y01)F01((x00,0,0),(y00,0,0); (v00,0,0)) and by the definition of the second-order quasitangent derivative ofF01 there exists (˜vk01,˜vk00,v˜−11k ,˜vk−10)→(v01, v00,0,0) such that
(x00,0,0, x01) +tk(y00,0,0, y01) +t2k(˜v00k ,˜vk−11,v˜−10k ,v˜k01)∈graphF01(·)
∀k∈N.
Using the Lipschitz property of the set-valued mapF01(·), we can write x01+tky01+t2kv˜01k ∈F01(x00+tky00+t2kv˜00k , t2kv˜−11k , t2k˜vk−10)⊂
⊂F01(x00+tky00+t2kv00k ,0,0) +Lt2k(v00k −v˜k00+˜vk−11+v˜−10k )B.
Thus, there exists b1k∈B such that
x01+tky01+t2k[˜vk01−Lb1k(vk00−v˜00k +˜vk−11+˜v−10k )]∈
∈F01(x00+tky00+t2kv00k ,0,0)
∀k∈Nand if we define
vk01:= ˜v01k −Lb1k(vk00−v˜00k +v˜−11k +˜vk−10),
then we have v01k → v1 and
x01+tky01+t2kvk01∈F01(x00+tky00+t2kv00k ,0,0) ∀k∈N.
By repeating this construction forj= 2, . . . , N, we find that there exists v0jk ∈Rn such that vk0j →v0j and
x0j+tky0j+t2kv0jk ∈F0j(x0j+tky0j+t2kv0j−1k ,0,0)
∀k∈N,∀j= 2, . . . , N.
Next, we repeat this construction for i∈ {1,2, . . . , N} fixed andj ∈ {0, 1, . . . , N}and we find that there existsvkij ∈Rnsuch thatvijk →vij ask→ ∞ and
xij +tkyij +t2kvkij ∈Fij(xij−1+tkyij−1+t2kvkij−1, xi−1j+tkyi−1j+t2kvi−1jk , xi−1j−1+tkyi−1j−1+t2kvi−1j−1k ) ∀k∈N.
In particular, fortk→0+xNN +tkyNN +t2kvNNk ∈RFN, i.e.,v=vNN ∈ Q2(x
NN,yNN)RNF and the proof is complete.
We are now able to prove our main result.
Theorem3.4. Assume that Hypothesis 2.3 is satisfied. Let g(·) :Rn→ R be a locally Lipschitz function, A0 ⊂Qx00F00(0,0,0) a closed convex cone, x = (x0, x1, . . . , xN) ∈ SF an optimal solution to problem (1.1)–(1.3) and assume that the constraint qualification below is satisfied:
(3.4)
−q∈Rn; ∃u1ij, u2ij, u3ij ∈Rn such that(u1ij, u2ij, u3ij)∈
∈(CxijFij((xij−1, xi−1j, xi−1j−1); (·,·,·)))∗(u1ij+1+u2i+1j+u3i+1j+1), i, j= 0,1, . . . , N, 0< i+j <2N,
(u1NN, u2NN, u3NN)∈(CxNNFNN((xNN−1, xN−1N, xN−1N−1); (·,·,·)))∗(q), u101+u210+u311∈A+0
∩(CxNNXN)+={0}.
Then we have the first-order necessary condition
(3.5) D↑g(xNN;yNN)≥0 ∀yNN ∈RNQ ∩QxNNXN.
Furthermore, if equality holds for some yNN, then we have the second-order necessary condition
(3.6) D↑2g(xNN, yNN;wNN)≥0 ∀wNN ∈R2Q∩Q2(x
NN,yNN)XN. Proof. According to Theorem 3.2, RCN ⊂CxNNRNF. Hence
(3.7) (CxNNRFN)+⊂(RNC)+.
It follows from Lemma 2.6, applied with Aij(u, v, w) := CxijFij((xij−1, xi−1j, xi−1j−1); (u, v, w)),∀u, v, w∈Rn,∀i, j∈ {0,1, . . . , N}, that
(3.8)
(RNC)+⊂
q; ∃u1ij, u2ij, u3ij ∈Rn such that (u1ij, u2ij, u3ij)∈
∈(CxijFij((xij−1, xi−1j, xi−1j−1); (·,·,·)))∗(u1ij+1+u2i+1j +u3i+1j+1), i, j= 0,1, . . . , N, 0< i+j <2N,
(u1NN, u2NN, u3NN)∈(CxNNFNN((xNN−1, xN−1N, xN−1N−1); (·,·,·)))∗(q), u101+u210+u311∈A+0
.
Therefore, from (3.4), (3.7) and (3.8) we deduce that (3.9) (CxNNRNF)−∩(CxNNXN)+={0}.
From Theorem 3.1 we have
(3.10) RNQ ⊂QxNNRNF
while from Theorem 3.3 we have
(3.11) R2Q⊂Q2(xNN,y
NN)RNF.
If x = (x0, x1, . . . , xN) ∈ SF is an optimal solution to problem (1.1)–
(1.3), then we have
g(xNN) = min{g(z); z∈RFN∩XN}. So, we apply Theorem 2.2 withS1=RNF andS2=XN.
Condition (3.9) assures that the constraint qualification in Theorem 2.2 is satisfied. Hence from (3.10) and Theorem 2.2 we obtain (3.5) while from (3.11) and Theorem 2.2 we obtain (3.6).
REFERENCES
[1] J.P. Aubin and H. Frankowska,Set-valued Analysis. Birkh¨auser, Basel, 1990.
[2] A. Cernea, On some second-order necessary conditions for differential inclusion prob- lems. Lect. Notes Nonlinear Anal.2(1998), 113–121.
[3] A. Cernea,Some second-order necessary conditions for nonconvex hyperbolic differential inclusion problems. J. Math. Anal. Appl.253(2001), 616–639.
[4] A. Cernea,Second-order necessary conditions for differential-difference inclusion prob- lems. Nonlinear Anal.62(2005), 963–974.
[5] A. Cernea, Second-order necessary conditions for discrete inclusions with end point constraints. Discuss. Math. Differ. Incl. Control Optim.25(2005), 47–58.
[6] A. Cernea, Minimum principle for a class of discrete inclusions. Math. Reports8(58) (2006),4, 391–399.
[7] A. Cernea,Minimum principle and controllability for multiparameter discrete inclusions via derived cones. Discrete Dyn. Nat. Soc.2006(2006), 1–12.
[8] E.H. Mahmudov, Necessary conditions for optimal control problems described by dif- ferential inclusions with distributed parameters. Soviet. Math. Dokl.38(1989), 467–471.
[9] H.D. Tuan and Y. Ishizuka,On controllability and maximum principle for discrete in- clusions. Optimization 34(1995), 293–316.
[10] H. Zheng,Second-order necessary conditions for differential inclusion problems. Appl.
Math. Optim.30(1994), 1–14.
Received 18 September 2006 University of Bucharest
Faculty of Mathematics and Computer Science Str. Academiei 14
010014 Bucharest, Romania