Thuy 2017
Thuy 2017
Thuy 2017
DOI 10.1007/s40306-017-0224-1
Abstract In this paper, we study the Mordukhovich coderivative and the local metric reg-
ularity in Robinson’s sense of the solution map to a parametric dynamic programming
problem with linear constraints and convex cost functions. By establishing abstract results
on the coderivative and the local metric regularity of the solution map to a parametric vari-
ational inequality, we obtain the Mordukhovich coderivative and the local metric regularity
in Robinson’s sense of the solution map to a parametric discrete optimal control problem.
1 Introduction
A wide variety of problems in discrete optimal control problem can be posed in the follow-
ing form.
Determine a pair (x, u) of a path x = (x0 , x1 , . . . , xN ) ∈ Rm × Rm × · · · × Rm and a
control vector u = (u0 , u1 , . . . , uN−1 ) ∈ Rn × Rn × · · · × Rn , which minimize the cost
N−1
f (x, u, μ) = hk (xk , uk , μk ) + hN (xN , μN ) (1)
k=0
1 School of Applied Mathematics and Informatics, Hanoi University of Science and Technology, 1
Dai Co Viet, Hanoi, Vietnam
L. Q. Thuy, N. T. Toan
coderivative and the local metric regularity of solution maps. We then apply the obtained
results to the problem (1)–(4).
The paper is organized as follows. In Section 2, we recall some notions and facts of varia-
tional analysis and generalized differentiation that we shall use in this paper. Section 3 gives
a formula for computing the Mordukhovich coderivative of the solution map. Some suffi-
cient conditions under which the solution map is locally-metrically regular in Robinson’s
sense are given in Section 4.
In this section, we recall some notions and facts of variational analysis and generalized
differentiation, which will be used in the sequel. The reader can find these notions and facts
in [4, 9, 11–14, 16, 17].
Let E1 and E2 be finite-dimensional Euclidean spaces and let F : E1 ⇒ E2 be a multi-
function. The effective domain, denoted by domF , and the graph of F , denoted by gphF ,
are defined as
domF := {z ∈ E1 : F (z) = ∅},
and
gphF := {(z, v) ∈ E1 × E2 : v ∈ F (z)}.
One says that F has a locally closed graph around (x0 , y0 ) ∈ gphF if there exists a closed
ball B in E1 × E2 of positive radius with center (x0 , y0 ) such that B ∩ gphF is a closed
subset of E1 × E2 . The norm in the product space is given by
(z, v) = z + v.
Let E be a finite-dimensional Euclidean space, a nonempty closed set in E, and z̄ ∈ .
The set ⎧ ⎫
⎨
z ∗ , z − z̄ ⎬
(z̄; ) := z∗ ∈ Z ∗ : lim sup
N ≤0
⎩ z − z̄ ⎭
z−
→z̄
is called the Fréchet normal cone to at z̄, and the set
)
N (z̄; ) := lim N(z;
z−
→z̄
is called the Mordukhovich normal cone to at z̄. Hence,
we put
K(w) = {z = (x, u) ∈ Z : xk+1 = Ak xk + Bk uk + wk , x0 = c,
αk ≤ uk ≤ βk , k = 0, 1, . . . , N − 1} . (5)
Then, the problem (1)–(4) can be formulated in a simpler form:
min {f (z, μ) : z ∈ K(w)} . (6)
Note that, for each fixed couple (μ, w), (6) is a programming problem under linear con-
straints. If f is convex and differentiable in z, then z is a solution of the problem if and only
if
0 ∈ fz (z, μ) + N (z; K(w)) ,
where N (z; K(w)) is the normal cone to K(w) at z in the sense of the convex analysis.
Putting φ(z, μ) = fz (z, μ), we get
0 ∈ φ(z, μ) + N (z; K(w)), (7)
which is called a parametric variational inequality.
Put ϒ = M × W . Throughout this paper, we assume that z̄ = (x̄, ū) is a solution of the
¯ 2 ¯
problem at ῡ = (μ̄, w̄), that is z̄ = (x̄, ū) ∈ S(μ̄, w̄), and symbols h̄k , ∂∂uhkk , ∂u∂ k h∂xk k , etc.,
stand, respectively, for
∂hk ∂ 2 hk
hk (x̄k , ūk , μ̄k ), (x̄k , ūk , μ̄k ), (x̄k , ūk , μ̄k ), etc.,
∂uk ∂uk ∂xk
where
⎡ ∂ 2 h̄k ∂ 2 h̄k ∂ 2 h̄k
⎤
...
∂xk1 ∂u1k ∂xk1 ∂u2k ∂xk1 ∂uM
⎢ k ⎥
⎢ ∂ 2 h̄k ∂ 2 h̄k ∂ 2 h̄k ⎥
∂ 2 h̄k ⎢ ∂xk2 ∂u1k ∂xk2 ∂u2k
...
∂xk2 ∂uM ⎥
=⎢
⎢ .. .. .. ..
k ⎥
⎥
∂xk ∂uk ⎢ . . . . ⎥
⎣ ⎦
∂ 2 h̄ k ∂ 2 h̄ k ∂ 2 h̄ k
...
∂xkS ∂u1k ∂xkS ∂u2k ∂xkS ∂uM
k
∂ 2 f (z̄, μ̄)
:M→Z
∂μ∂z
is surjective.
The Mordukhovich Coderivative and the Local Metric Regularity...
Here, μ̄ = (μ̄0 , μ̄1 , . . . , μ̄N ), x̄ = (x̄0 , x̄1 , . . . , x̄N ), ū = (ū0 , ū1 , . . . , ūN−1 ), and
Mk0 , Xk0 , Uk0 are convex neighborhoods of μ̄k , x̄k , ūk , respectively, and
⎡ ∂ 2 h̄0
⎤
∂μ0 ∂x0 0 0 ... 0 0
⎢ ∂ 2 h̄1 ⎥
⎢ 0 0 ... 0 0 ⎥
⎢ ∂μ1 ∂x1 ⎥
⎢ .. .. .. .. .. .. ⎥⎡ ⎤
⎢ ⎥ μ0
⎢ . . . . . . ⎥
⎢ ∂ 2 h̄N −1 ⎥ ⎢ μ1 ⎥
⎢ 0 0 0 ... 0 ⎥⎢ ⎥
2 ⎢ ∂μN −1 ∂xN −1 ⎥ ⎢ μ2 ⎥
∂ f (z̄, μ̄) ⎢ ∂ 2 h̄N ⎥⎢ ⎥
(μ0 , μ1 , . . . , μN ) = ⎢ 0 0 0 ... 0 ⎥ ⎢ .. ⎥
∂μ∂z ⎢ ∂μN ∂xN ⎥⎢ . ⎥
⎢ ∂ 2 h̄0 ⎥⎢ ⎥
⎢ 0 0 ... 0 0 ⎥ ⎣ μN−1 ⎦
⎢ ∂μ0 ∂u0
⎥
⎢ 0 ∂ 2 h̄1
0 ... 0 0 ⎥ μN
⎢ ∂μ1 ∂u1 ⎥
⎢ .. .. .. .. .. .. ⎥
⎢ ⎥
⎣ . . . . . . ⎦
∂ 2 h̄N −1
0 0 0 ... ∂μN −1 ∂uN −1 0
∂ 2 h̄0 ∂ 2 h̄1 ∂ 2 h̄N−1
= μ0 , μ1 , . . . , μN−1 ,
∂μ0 ∂x0 ∂μ1 ∂x1 ∂μN−1 ∂xN−1
∂ 2 h̄N ∂ 2 h̄0 ∂ 2 h̄1
μN , μ0 , μ1 , . . . ,
∂μN ∂xN ∂μ0 ∂u0 ∂μ1 ∂u1
∂ 2 h̄N−1
μN−1 .
∂μN−1 ∂uN−1
It is clear that local properties of the solution map S around the point
depend on the structure of the set gphG := {(z, υ, y) ∈ Z × ϒ × Z : y ∈ G(z, υ)} around
the point (z̄, ῡ, 0Z ).
The distance from z ∈ Z to a subset ⊂ Z is defined by
uk ≤ βk and − uk ≤ −αk .
L. Q. Thuy, N. T. Toan
Define
⎡ ⎤
w0
⎢ −w0 ⎥
⎢ ⎥
⎢ w1 ⎥
⎢ ⎥
⎡ ⎤ ⎢ −w1 ⎥
x0 ⎢ ⎥
⎢ .. ⎥
⎢ x1 ⎥ ⎢ . ⎥
⎢ ⎥ ⎢ ⎥
⎢ .. ⎥ ⎢ wN−1 ⎥
⎢ . ⎥ ⎢ ⎥
⎢ ⎥ ⎢ −wN−1 ⎥
⎢ xN ⎥ ⎢ ⎥
⎢
z=⎢ ⎥, b(w) = ⎢ c ⎥, (8)
⎥ ⎢ ⎥
⎢ u0 ⎥ ⎢ −c ⎥
⎢ u1 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ β0 ⎥
⎢ . ⎥ ⎢ ⎥
⎣ .. ⎦ ⎢ −α0 ⎥
⎢ ⎥
⎢ β1 ⎥
uN−1 ⎢ ⎥
⎢ −α1 ⎥
⎢ ⎥
⎢ .. ⎥
⎣ . ⎦
−αN−1
and
⎡ ⎤
−A0 Im 0 0 ... 0 0 −B0 0 0 ... 0
⎢ A0 −Im 0 0 ... 0 0 B0 0 0 ... 0 ⎥
⎢ ⎥
⎢ 0 −A1 Im 0 ... 0 0 0 −B1 0 ... 0 ⎥
⎢ ⎥
⎢ 0 A1 −Im 0 ... 0 0 0 B1 0 ... 0 ⎥
⎢ ⎥
⎢ .. .. .. .. .. .. .. .. .. .. .. .. ⎥
⎢ . . . . . . . . . . . . ⎥
⎢ ⎥
⎢ 0
⎢ 0 0 0 . . . −AN−1 Im 0 0 0 . . . −BN−1 ⎥
⎥
⎢ 0
⎢ 0 0 0 . . . AN−1 −Im 0 0 0 . . . BN−1 ⎥⎥
C=⎢ 0 0 0 0 0 0 0 0 0 ⎥,
⎢ Im ... ... ⎥ (9)
⎢ −Im 0 0 0 ... 0 0 0 0 0 ... 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 ... 0 0 In 0 0 ... 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 ... 0 0 −In 0 0 ... 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 ... 0 0 0 In 0 ... 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 ... 0 0 0 −In 0 ... 0 ⎥
⎢ ⎥
⎢ . .. .. .. .. .. .. .. .. .. .. .. ⎥
⎣ .. . . . . . . . . . . . ⎦
0 0 0 ... 0 0 0 0 0 0 . . . −In
where Im and In denote the m × m and n × n unit matrices, respectively. Then, we have
K(w) = {z ∈ X × U : Cz ≤ b(w)}.
Let us put = R2(N+1)m+2Nn , P = M × and define a mapping K1 : ⇒ Z by
K1 (b) = {z ∈ Z : Cz ≤ b} ∀b ∈ (10)
where the matrix C is defined by (9). Thus, we have K(w) = K1 (b(w)) for all w ∈ W ,
where b(w) is defined by (8). We denote by D1 the effective domain of K1 and D the
effective domain of K. It is clear that
D = {w ∈ W : K(w) = ∅} = {w ∈ W : b(w) ∈ D1 }.
In the sequel, we shall need the following lemma.
The Mordukhovich Coderivative and the Local Metric Regularity...
Lemma 1 ([9, Theorem 2.2]) The set-valued map K1 : ⇒ Z defined by (10) is Lipschitz
on D1 , i.e., there exists a constant l > 0, independent of b, such that
K1 (b1 ) ⊆ K1 (b2 ) + lb1 − b2 BZ
for all b1 , b2 ∈ D1 , where BZ stands for the closed unit ball in Z.
Notice that the set-valued map K is also Lipschitz on D by [19, Corollary 2.2 ].
Given a set ⊂ Z, the set
∗ = {z∗ ∈ Z :
z∗ , z ≤ 0 ∀z ∈ }
is called the polar cone of . The tangent cone to at z̄ ∈ denoted by T (z̄; ) is defined
by
T (z̄; ) = N (z̄; )∗ = v ∈ Z :
z∗ , v ≤ 0 ∀z∗ ∈ N (z̄; ) .
From now on, we shall write b̄ instead of b(w̄). For each (μ, b) ∈ P = M × , consider
the problem of finding z = z(μ, b), which satisfies the equation
0 ∈ φ(μ, z) + N (z; K1 (b)) . (11)
Let us denote by S1 (μ, b) the solution set of (11) corresponding to
(μ, b) ∈ M × .
It is clear that S(μ, w) = S1 (μ, b(w)) for all (μ, w) ∈ M × D, where S(μ, w) is the
solution set of (7), which is also the solution map of the problem (1)–(4).
Notice that C = (cij )p×q , where
p = 2(N + 1)m + 2N n and q = (N + 1)m + N n.
Put
T = {0, 1, . . . , p} = T0 ∪ T1 ,
where
T0 = {1, 2, . . . , 2(N + 1)m},
T1 = {2(N + 1)m + 1, 2(N + 1)m + 2, . . . , 2(N + 1)m + 2N n}.
Let us denote by Ci the i-th row of matrix C. For a fixed element z ∈ K1 (b), the set of
active indices at z is given by
I (z, b) = {i ∈ T : Ci z = (b)i } , (12)
where (b)i is the i-th component of b. Here, the vector b consists of 2(N + 1)m + 2N n
components, and the vector z consists of (N + 1)m + N n components. For convenience, we
assume that
βi = b̂2(N+1)m+2in+1 , b̂2(N+1)m+2in+2 , . . . , b̂2(N+1)m+2in+n ,
−αi = b̂2(N+1)m+(2i+1)n+1 , b̂2(N+1)m+(2i+1)n+2 , . . . , b̂2(N+1)m+(2i+1)n+n ,
i = 0, 1, . . . , N − 1.
Here, b̂k are fixed for all
k ∈ {2(N + 1)m + 1, 2(N + 1)m + 2, . . . , 2(N + 1)m + 2N n}.
L. Q. Thuy, N. T. Toan
Thus, we have
where
T1 (z̄, b̄) = {i ∈ T1 : Ci z̄ = (b̂)i }.
For every subset I ⊂ T , we put I¯ = T \I . Let CI (resp., CI¯ ) be the matrix composed by the
rows Ci , i ∈ I (resp., the rows Ci , i ∈ I¯) of C.
The following lemma gives formulas of the normal cone and the tangent cone of K1 (b).
Its proof can be found in [14, Lemma 3.1].
Lemma 2 Let K1 (b) be defined by (10), z ∈ K1 (b) and I (z, b) be defined by (12). Then,
the following representations are fulfilled:
(i)
N (z; K1 (b)) = y ∈ Z : y ∈ pos{CiT : i ∈ I (z, b)} ,
where
⎧ ⎫
⎨ ⎬
pos CiT : i ∈ I (z, b) = λi CiT , λi ≥ 0 ;
⎩ ⎭
i∈I (z,b)
(ii)
T (z; K1 (b)) = {v ∈ Z : Ci v ≤ 0 ∀i ∈ I (z, b)} .
and
BQ,P = z ∈ Z : Ci z = 0 ∀i ∈ P , Ci z ≤ 0 ∀i ∈ QP ,
where span stands for the linear subspace generated by . We have the following result
from [4, Lemma 3.3].
Lemma 3 If P ⊂ Q ⊂ T then
(BQ,P )∗ = AQ,P .
and assume that 2 is the graph of F2 . For each (z̄, b̄, z̄∗ ) ∈ 2 , we get
z̄∗ ∈ F2 (z̄, b̄) = N z̄; K1 (b̄) .
The Mordukhovich Coderivative and the Local Metric Regularity...
By Lemma 2, we have
z̄∗ = λi CiT for some λi ≥ 0, i ∈ I = I (z̄, b̄).
i∈I (z̄,b̄)
Put
(z̄, b̄, z̄∗ ) = (λi )i∈I : z̄∗ = λi CiT , λi ≥ 0 for all i ∈ I ,
i∈I
⎧ ⎫
⎨ ⎬
Iˆ1 (z̄, b̄, z̄∗ ) = i ∈ I : z̄∗ = λj CjT for some λj ≥ 0, j ∈ I {i} ,
⎩ ⎭
j ∈I {i}
and
∗ I if z̄∗ = 0 and |I | = 1,
I1 (z̄, b̄, z̄ ) =
Iˆ1 (z̄, b̄, z̄ ) otherwise.
∗
Assume that (λi )i∈I ∈
(z̄, b̄, z̄∗ ), we put K = {i ∈ I : λi > 0}. From [16, Theorem 3.1],
[16, Lemma 4.3] and [16,
Lemma 4.4], we obtain the following formula for computing the
Fréchet normal cone N̂ (z̄, b̄, z̄∗ ); 2 .
and
J (z, b, z∗ ) if z∗ = 0,
Î (z, b, z∗ ) = (14)
J (z, b, z∗ ) ∪ {∅} if z∗ = 0.
L. Q. Thuy, N. T. Toan
Using similar arguments as in the proof of [16, Theorem 4.1], we obtain the following
formula for computing the Mordukhovich normal cone to 2 at (z̄, b̄, z̄∗ ).
Theorem 1 Assume that (z̄, b̄, z̄∗ ) ∈ 2 , I = I (x̄, b̄) and Î = Î (z̄, b̄, z̄∗ ). Then
⎧
⎨
N (z̄, b̄, z̄∗ ); 2 ⊂ (z∗ , b∗ , v) : (z∗ , v) ∈ AQ,P × BQ,P ,
⎩
P ⊂Q⊂I,P ∈Î
⎫
⎬
z∗ = − CiT bi∗ , bQ̄
∗ ∗
= 0, bQP ≤0 .
⎭
i∈Q
The following theorem is our first main result, which provides a formula for computing the
Mordukhovich coderivative of the solution map to the problem (1)–(4).
Theorem 3 Suppose that z̄ = (x̄, ū) is a solution of the problem (1)–(4) corresponding
to parameter (μ̄, w̄), z∗ ∈ Z and the assumptions (H1)–(H2) are satisfied. Then, for each
(μ∗ , w ∗ ) ∈ D ∗ S(μ̄, w̄, z̄)(z∗ ) there exist y ∗ ∈ Z and index sets P , Q with P ⊂ Q ⊂
I, P ∈ Î such that the following conditions are satisfied:
T
∗ ∂ 2 f (z̄, μ̄)
μ = y∗, (15)
∂μ∂z
⎛
T ⎞
2
⎝−z∗ − ∂ f (z̄, μ̄) y ∗ , −y ∗ ⎠ ∈ AQ,P × BQ,P , (16)
∂z2
T
∂ 2 f (z̄, μ̄)
z∗ + y∗ = CiT b∗ (w ∗ )i , (17)
∂z2
i∈Q∩T0
and
b∗ (w ∗ )Q̄∩T = 0; b∗ (w ∗ )(Q\P )∩T ≤ 0, (18)
0 0
The Mordukhovich Coderivative and the Local Metric Regularity...
where I is defined by (13) and Î is defined by (14) for z̄∗ = − ∂f (z̄, μ̄)
∂z . Here,
T
∂ 2 f (z̄, μ̄)
(x0∗ , x1∗ , . . . , xN
∗
, u∗0 , u∗1 , . . . , u∗N−1 )
∂μ∂z
⎡ ⎤
x0∗
⎡ ⎤⎢ x1∗
⎥
∂ 2 h̄0 ∂ 2 h̄0 ⎢ ⎥
0 ... 0 0 0 ... 0 ⎢ ⎥
..
⎢ ∂μ0 ∂x0 ∂μ0 ∂u0
⎥⎢ ⎥
.
⎢ 0 ∂ 2 h̄1
... 0 0 0 ∂ 2 h̄1
... 0 ⎥⎢ ∗ ⎥
⎢ ∂μ1 ∂x1 ∂μ1 ∂u1 ⎥⎢ x ⎥
⎢ .. .. .. .. .. .. .. .. .. ⎥ ⎢ N−1 ⎥
=⎢
⎢ . . . . . . . . .
⎥ ⎢ x∗ ⎥
⎥ ⎢ N∗ ⎥
⎢ ⎥⎢ u ⎥
⎢ 0 0 ...
∂ 2 h̄N −1
0 0 0 ... ∂ 2 h̄N −1 ⎥ ⎢ 0∗ ⎥
⎣ ∂μN −1 ∂xN −1 ∂μN −1 ∂uN −1 ⎦⎢ u ⎥
∂ 2 h̄N ⎢ 1 ⎥
0 0 ... 0 0 0 ... 0 ⎢ . ⎥
∂μN ∂xN ⎣ .. ⎦
u∗N−1
∂ 2 h̄0 ∗ ∂ 2 h̄0 ∗ ∂ 2 h̄1 ∗ ∂ 2 h̄1 ∗
= x0 + u0 , x1 + u ,...,
∂μ0 ∂x0 ∂μ0 ∂u0 ∂μ1 ∂x1 ∂μ1 ∂u1 1
∂ 2 h̄N−1 ∗ ∂ 2 h̄N−1 ∗ ∂ 2 h̄N ∗
x + u , x ,
∂μN−1 ∂xN−1 N−1 ∂μN−1 ∂uN−1 N−1 ∂μN ∂xN N
T
∂ 2 f (z̄, μ̄)
(x0∗ , x1∗ , . . . , xN ∗
, u∗0 , u∗1 , . . . , u∗N−1 )
∂z2
⎡ 2 2 h̄
⎤
∂ h̄0
0 ... 0 0 ∂x∂ 0 ∂u 0
0 ... 0
⎢ ∂x02 0 ⎥
⎢ 2
∂ h̄1 2
∂ h̄1 ⎥⎡ ⎤
⎢ 0 ... 0 0 0 ∂x1 ∂u1 . . . 0 ⎥ x0∗
⎢ ∂x12 ⎥
⎢ . . . . . . . . . ⎥ ⎢ x∗ ⎥
⎢ . .. .. .. .. .. .. .... ⎥⎢ 1 ⎥
⎢ . ⎥⎢ . ⎥
⎢ ⎥ ⎢ .. ⎥
⎢ 0 2
∂ h̄N −1 2
0 . . . ∂xN∂ −1h̄N∂u−1N −1 ⎥⎢
⎢ 0 ... 0 0 ⎥ ⎢ x∗ ⎥
⎥ ⎢ N−1 ⎥
2
⎢ ∂xN −1
⎢ ∂ 2 h̄N ⎥⎢ ∗ ⎥
=⎢ 0 0 ... 0 0 0 ... 0 ⎥ ⎢ xN ⎥
⎢ 2 ∂xN 2
⎥⎢ ∗ ⎥
⎢ ∂ h̄0 ⎥ ⎢ u0 ⎥
⎢ ∂u0 ∂x0 0 ... 0 0 ∂ 2 h̄0
0 ... 0 ⎥⎢ ∗ ⎥
⎢ ∂u20 ⎥ ⎢ u1 ⎥
⎢ ⎥⎢ . ⎥
⎥⎣ . ⎥
2 2
⎢ 0 ∂ h̄1 ∂ h̄1
. ⎦
⎢ ∂u1 ∂x1 . . . 0 0 0 ... 0 ⎥
∂u21
⎢ . ⎥ ∗
⎢ . .
.. .
.. .
.. .
.. .
.. .
.. .
.. ⎥ uN−1
⎢ . ⎥
⎣ 2
∂ h̄ 2 ⎦
0 0 . . . ∂uN −1N∂x−1N −1 0 0 0 . . . ∂ h̄2N −1
∂uN −1
∂ 2 h̄0 ∗ ∂ 2 h̄0 ∗ ∂ 2 h̄1 ∗ ∂ 2 h̄1 ∗ ∂ 2 h̄N−1 ∗
= 2
x0 + u0 , 2
x1 + u1 , . . . , 2
xN−1
∂x0 ∂x0 ∂u0 ∂x1 ∂x1 ∂u1 ∂xN−1
∂ 2 h̄N−1 ∂ 2 h̄N ∗ ∂ 2 h̄0 ∗ ∂ 2 h̄0 ∗ ∂ 2 h̄1 ∗ ∂ 2 h̄1 ∗
+ u∗N−1 , x ,
2 N ∂u ∂x 0
x + u0 , x + u1 ,
∂xN−1 ∂uN−1 ∂xN 0 0 ∂u20 ∂u1 ∂x1 1 ∂u21
∂ 2 h̄N−1 ∗ ∂ 2 h̄N−1 ∗
..., x + uN−1 , (19)
∂uN−1 ∂xN−1 N−1 ∂u2N−1
L. Q. Thuy, N. T. Toan
and
b∗ (w ∗ ) = [w0∗ , 0, w1∗ , 0, . . . , wN−1
∗
, 0, 0, 0, 0, 0, 0, 0, . . . , 0]T ,
T0 = {1, 2, . . . , m, 2m + 1, 2m + 2, . . . , 2m + m, . . . , 2(N − 1)m + 1,
2(N − 1)m + 2, . . . , 2(N − 1)m + m} .
Hence, we get
φ(zk , μk ) − zk∗ , z − zk ≥ 0 ∀z ∈ K1 (bk ).
Fix any z ∈ K1 (b). By Lemma 1, K1 is Lipschitz continuous with the Lipschitz constant
l. Hence for each k, there exists zk ∈ K1 (bk ) such that
z − zk ≤ lb − bk .
This implies that zk → z . Since
φ(zk , μk ) − zk∗ , zk − zk ≥ 0
and letting k → ∞, we get
φ(z, μ) − z∗ , z − z ≥ 0.
As z ∈ K1 (b) is arbitrary, we obtain (z, μ, b, z∗ ) ∈ gphF. Hence, F has a closed graph.
Similarly, we can show that 2 , S1 and S have closed graphs.
By changing the order of components and using Lemma 5, we see that for all z∗ ∈ Z =
R (N+1)m+Nn , one has
D ∗ F (z̄, μ̄, b̄, 0Z )(z∗ )
= ∇φ1 (z̄, μ̄, b̄)T z∗ + D ∗ 2 z̄, μ̄, b̄, −φ1 (z̄, μ̄, b̄) (z∗ )
= ∇z φ(z̄, μ̄)T z∗ , ∇μ φ(z̄, μ̄)T z∗ , 0 + D ∗ 2 z̄, μ̄, b̄, −φ1 (z̄, μ̄, b̄) (z∗ )
% &
= ∇z φ(z̄, μ̄)T z∗ , 0 + D ∗ F2 z̄, b̄, −φ(z̄, μ̄) (z∗ ) × {∇μ φ(z̄, μ̄)T z∗ }, (21)
where 0 and 0Z are zero elements in the spaces and Z, respectively. We now show that
the implication
0 ∈ ∇φ1 (z̄, μ̄, b̄)T z∗ + D ∗ 2 z̄, μ̄, b̄, −φ1 (z̄, μ̄, b̄) (z∗ ) ⇒ z∗ = 0 (22)
is valid. Indeed, from (21) and condition
0 ∈ ∇φ1 (z̄, μ̄, b̄)T z∗ + D ∗ 2 z̄, μ̄, b̄, −φ1 (z̄, μ̄, b̄) (z∗ ),
we get
∇μ φ(z̄, μ̄)T z∗ = 0. (23)
The assumption (H2) and [11, Lemma 1.18] imply that the adjoint mapping
T
∂ 2 f (z̄, μ̄)
∇μ φ(z̄, μ̄) =
T
:Z→M
∂μ∂z
is injective. Hence, we obtain from (23) that z∗ = 0. Consequently, (22) is justified. By
Lemma 6, we have
D ∗ S1 (μ̄, b̄, z̄)(z∗ ) (24)
⊂ (μ∗ , b∗ ) ∈ M × : ∃y ∗ ∈ Z with −z∗ − z 1 (z̄, μ̄, b̄)T y ∗ ,
μ∗ − μ 1 (z̄, μ̄, b̄)T y ∗ , b∗ ∈ D ∗ 2 z̄, μ̄, b̄, −1 (z̄, μ̄, b̄) (y ∗ ) ,
for any z∗ ∈ Z. Take any (μ∗ , w ∗ ) ∈ D ∗ S(μ̄, w̄, z̄)(z∗ ), we will prove that
∗ ∗ ∗
μ , b (w ) ∈ D ∗ S1 (μ̄, b̄, z̄)(z∗ ),
with
b∗ (w ∗ ) = [w0∗ , 0, w1∗ , 0, . . . , wN−1
∗
, 0, 0, 0, 0, 0, 0, 0, . . . , 0]T .
L. Q. Thuy, N. T. Toan
Since (μ∗ , w ∗ ) ∈ D ∗ S(μ̄, w̄, z̄)(z∗ ), we have (μ∗ , w ∗ , −z∗ ) ∈ N ((μ̄, w̄, z̄); gphS). So,
gphS
there exist sequences (μk , wk , zk ) −−→ (μ̄, w̄, z̄), and
(μ∗k , wk∗ , −zk∗ ) ∈ N̂ ((μk , wk , zk ); gphS)
such that
(μ∗k , wk∗ , −zk∗ ) → (μ∗ , w ∗ , −z∗ ).
From (μ∗k , wk∗ , −zk∗ ) ∈ N̂ ((μk , wk , zk ); gphS), we get
μ∗k , μ − μk +
wk∗ , w − wk +
−zk∗ , z − zk
lim sup ≤ 0. (25)
gphS μ − μk + w − wk + z − zk
(μ,w,z)−−→(μk ,wk ,zk )
gphS1
Take any (μ, b, z) −−−→ (μk , b(wk ), zk ). Without loss of generality, we can assume that
b = [w0 , b0 , w1 , b1 , . . . , wN−1
, bN−1 , b2Nm+1 , b2Nm+2 , . . . , bp ]T ,
with p = 2(N + 1)m + 2N n and
w = [w0 , w1 , . . . , wN−1
]T → wk = [w0k , w1k , . . . , w(N−1)k ]T .
Since
(μ, b,z) ∈ gphS1 , (μ, b, z) → (μk , b(wk ), zk ) and the closedness of gphS1 , we get
μ, b(w ), z ∈ gphS1 . For
b∗ (w ∗ )k = [w0k
∗ ∗
, 0, w1k ∗
, 0, . . . , w(N−1)k , 0, 0, 0, 0, 0, 0, 0, . . . , 0]T ,
we have
μ∗k , μ − μk +
b∗ (w ∗ )k , b − b(wk ) +
−zk∗ , z − zk
μ − μk + b − b(wk ) + z − zk
' ∗
μk , μ − μk +
wk∗ , w − wk +
−zk∗ , z − zk
=
μ − μk + w − wk + z − zk
(
μ − μk + w − wk + z − zk
×
μ − μk + b − b(wk ) + z − zk
μ∗k , μ − μk +
wk∗ , w − wk +
−zk∗ , z − zk
≤ .
μ − μk + w − wk + z − zk
By (25), we get
μ∗k , μ − μk +
b∗ (w ∗ )k , b − b(wk ) +
−zk∗ , z − zk
lim sup
gphS1 μ − μk + b − b(wk ) + z − zk
(μ,b,z)−−−→(μk ,b(wk ),zk )
μ∗k , μ − μk +
b∗ (w ∗ )k , b − b(wk ) +
−zk∗ , z − zk
≤ lim sup
gphS1 μ − μk + w − wk + z − zk
(μ,b,z)−−−→(μk ,b(wk ),zk )
μ∗k ,μ −μk +
b∗ (w ∗ )k , b(w ) − b(wk ) +
−zk∗ , z − zk
= lim sup
gphS1 μ − μk + w − wk + z − zk
(μ,b(w ),z)−−−→(μk ,b(wk ),zk )
μ∗k , μ − μk +
wk∗ , w − wk +
−zk∗ , z − zk
= lim sup ≤ 0.
gphS μ − μk + w − wk + z − zk
(μ,w ,z)−−→(μk ,wk ,zk )
So,
μ∗k , b∗ (w ∗ )k , −zk∗ ∈ N̂ ((μk , b(wk ), zk ); gphS1 ) .
The Mordukhovich Coderivative and the Local Metric Regularity...
gphS
Since (μk , wk , zk ) −−→ (μ̄, w̄, z̄) and (μ∗k , wk∗ , −zk∗ ) → (μ∗ , w ∗ , −z∗ ), we have
gphS1
(μk , b(wk ), zk ) −−−→ (μ̄, b̄, z̄) and μ∗k , b∗ (w ∗ )k , −zk∗ → μ∗ , b∗ (w ∗ ), −z∗ .
Hence,
μ∗ , b∗ (w ∗ ), −z∗ ∈ N (μ̄, b̄, z̄); gphS1 .
Thus, (μ∗ , b∗ (w ∗ )) ∈ D ∗ S1 (μ̄, b̄, z̄)(z∗ ). Note that
1 (z, μ, b) = (z, μ), 2 (z, μ, b) = F2 (z, b).
By (24) and Theorem 2, we have μ∗ = μ (z̄, μ̄)T y ∗ and there exist index sets P ⊂ Q ⊂
I, P ∈ Î such that
−z∗ − z (z̄, μ̄)T y ∗ , −y ∗ ∈ AQ,P × BQ,P ,
−z∗ − z (z̄, μ̄)T y ∗ = − CiT b∗ (w ∗ )i ,
i∈Q
∗ ∗ ∗ ∗
b (w )Q̄ = 0; and b (w )Q\P ≤ 0,
and
E = R3 × {0} × {0}.
Indeed, it is easy to see that z̄ = (x̄, ū) is a solution of the problem corresponding to
(μ̄, w̄) and the assumption (H1) is satisfied. From the definition of f , we have
⎡ ⎤
1 1 0 0
∂ 2 f (z̄, μ̄) ⎣
= 0 0 2 0⎦.
∂μ∂z 0 1 0 0
So,
∂ 2 f (z̄, μ̄)
(μ01 , μ02 , μ11 , μ12 ) = (μ01 + μ02 , 2μ11 , μ02 ),
∂μ∂z
∂ 2 f (z̄,μ̄)
which implies that the map ∂μ∂z is surjective. Hence, the assumption (H2) is also
satisfied. Define
⎡ ⎤ ⎡ ⎤
1 −1 1 −1
⎡ ⎤ ⎢ −1 ⎥ ⎢ 1 −1 1 ⎥
0 ⎢ ⎥ ⎢ ⎥
⎢ 0 ⎥ ⎢ 1 0 0 ⎥
z̄ = ⎣ 1 ⎦ , b̄ = ⎢ ⎥
⎢ 0 ⎥, and C = ⎢
⎢ −1
⎥.
⎢ ⎥ ⎢ 0 0 ⎥
⎥
0 ⎣ 1 ⎦ ⎣ 0 0 1 ⎦
1 0 0 −1
We have
K1 (b̄) = z ∈ R3 : Ci z ≤ b̄i = {0} × [0, 2] × [−1, 1],
and
I = I (z̄, b̄) = i ∈ T : Ci z̄ = b̄i = {1, 2, 3, 4}.
Hence,
F2 (z̄, b̄) = N z̄; K1 (b̄) = pos CiT : i ∈ I = R3 .
So, BQ,P = {0R3 }. From the condition (16) of Theorem 3, we get y ∗ = (0, 0, 0). Substi-
tuting y ∗ = (0, 0, 0) into (15), we obtain μ∗ = (0, 0, 0, 0). By (18), we have w∗ ≤ 0.
Hence,
(μ∗ , w ∗ ) ∈ {0R4 } × R− .
∗ ∗
Since y = (0, 0, 0), w ≤ 0 and the condition (17), we obtain
z∗ = (−w ∗ , w ∗ , −w ∗ ) ∈ R+ × R− × R+ .
For P = {2} or P = {3} or P = {4} or P = {2, 3} or P = {2, 4}, using similar arguments
as in Case 1 (a), we can show that
(μ∗ , w ∗ ) = ((0, 0, 0, 0), a) ∈ {0R4 } × R− if z∗ = (−a, a, −a) ∈ R+ × R− × R+ .
(b) P = {1}, we have
AQ,P = span{C1 } + pos{Ci : i = 2, 3, 4} = R3 .
So, BQ,P = {0R3 }. From the condition (16), we have y ∗ = (0, 0, 0). Substituting y ∗ =
(0, 0, 0) into (15), we obtain μ∗ = (0, 0, 0, 0). By (18), we get w∗ ∈ R. Hence,
(μ∗ , w ∗ ) = ((0, 0, 0, 0), a) ∈ {0R4 } × R if z∗ = (−a, a, −a) ∈ R3 .
For P = {1, 3} or P = {1, 4}, using similar arguments as in Case 1 (a), we can show that
(μ∗ , w ∗ ) = ((0, 0, 0, 0), a) ∈ {0R4 } × R if z∗ = (−a, a, −a) ∈ R3 .
For P = {3} or P = {4}, using similar arguments as in Case 4 (b), we can show that
(μ∗ , w ∗ ) ∈ ({0} × R− × R+ × {0}) × R− if z∗ ∈ R+ × R− × R+ .
(b) P = ∅,
(μ∗ , w ∗ ) ∈ R2+ × R− × {0} × {0} if z∗ ∈ R− × R+ × R− .
(b) P = ∅,
AQ,P = pos{Ci : i ∈ Q} = R+ × {0} × {0}.
So, BQ,P = R− × R × R. From the condition (16), we have
y ∗ = (y1∗ , y2∗ , y3∗ ) ∈ R+ × R × R.
Since y ∗ ∈ R+ × R × R and (15), we get
μ∗ = (y1∗ , y1∗ + y3∗ , 2y2∗ , 0) ∈ R+ × R × R × {0}.
By (18), we have w ∗ = 0. From (16) and
T
∂ 2 f (z̄, μ̄)
(y1∗ , y2∗ , y3∗ ) = (2y1∗ + 2y3∗ , 0, 2y1∗ + 2y3∗ ) ∈ R × {0} × R,
∂z2
T
∂ 2 f (z̄,μ̄)
we obtain −z∗ ∈ ∂z2
y ∗ + AQ,P = R × {0} × R. Hence,
(μ∗ , w ∗ ) ∈ R+ × R2 × {0} × {0} if z∗ ∈ R × {0} × R.
(b) P = ∅,
AQ,P = pos{Ci : i ∈ Q} = R− × {0} × {0}.
So, BQ,P = R+ × R × R. From the condition (16), we have
y ∗ = (y1∗ , y2∗ , y3∗ ) ∈ R− × R × R.
Since y ∗ ∈ R− × R × R and (15), we get
μ∗ = (y1∗ , y1∗ + y3∗ , 2y2∗ , 0) ∈ R− × R × R × {0}.
By (18), we have w ∗ = 0. From (16) and
T
∂ 2 f (z̄, μ̄)
(y1∗ , y2∗ , y3∗ ) = (2y1∗ + 2y3∗ , 0, 2y1∗ + 2y3∗ ) ∈ R × {0} × R,
∂z2
T
∂ 2 f (z̄,μ̄)
we obtain −z∗ ∈ ∂z2
y ∗ + AQ,P = R × {0} × R. Hence,
(μ∗ , w ∗ ) ∈ R− × R2 × {0} × {0} if z∗ ∈ R × {0} × R.
In the previous section, we have obtained a result on the Mordukhovich coderivative of the
solution map to the problem (1)–(4). In this section, we continue to study properties of the
solution map. Namely, we want to investigate the local metric regularity in Robinson’s sense
of the solution map to the problem (1)–(4). The following theorem is our second main result.
Theorem 4 Suppose that z̄ = (x̄, ū) is a solution of the problem (1)–(4) corresponding to
parameter (μ̄, w̄) and the assumptions (H1), (H2) are satisfied. If for any
(z∗ , b∗ ) ∈ R(N+1)m+Nn × R2(N+1)m+2Nn , one has (z∗ , b∗ ) = (0, 0)
whenever
⎛
T ⎞
2
⎝ ∂ f (z̄, μ̄) z∗ , z∗ ⎠ ∈ AQ,P × BQ,P , (27)
∂z2
T
∂ 2 f (z̄, μ̄)
z∗ = − CiT bi∗ , (28)
∂z2
i∈Q
∗ ∗
bQ̄ = 0 and bQP ≤ 0, (29)
Proof of Theorem 4 From the proof of Theorem 3, we obtain that the functions F and 2
have closed graphs and
D ∗ F (z̄, μ̄, b̄, 0Z )(z∗ )
% &
= ∇z φ(z̄, μ̄)T z∗ , 0 + D ∗ F2 z̄, b̄, −φ(z̄, μ̄) (z∗ ) × {∇μ φ(z̄, μ̄)T z∗ } (32)
for all z∗ ∈ Z = R (N+1)m+Nn . Also from the proof of Theorem 3, it follows that the
condition (30) is satisfied. We now show that the condition
{(μ∗ , b∗ ) ∈ P : (0, μ∗ , b∗ ) ∈ D ∗ F (ȳ)(z∗ )} = {0} with ȳ = (z̄, μ̄, b̄, 0Z ) (33)
z∗ ∈Z
is valid. Indeed, assume that (μ∗ , b∗ ) ∈ P satisfying (0, μ∗ , b∗ ) ∈ D ∗ F (ȳ)(z∗ ) for some
z∗ ∈ Z. By (32), we have
μ∗ = ∇μ φ(z̄, μ̄)T z∗ , (34)
and
(0, b∗ ) ∈ ∇z φ(z̄, μ̄)T z∗ , 0 + D ∗ F2 z̄, b̄, −φ(z̄, μ̄) (z∗ ).
The latter is equivalent to
−∇z φ(z̄, μ̄)T z∗ , b∗ ∈ D ∗ F2 z̄, b̄, −φ(z̄, μ̄) (z∗ ).
with
i0 are neighborhoods of w̄i (i = 0, 1, . . . , N − 1),
2(N−1)+1 2(N−1)+2
0 is a neighborhood of c, 0 is a neighborhood of − c,
2(N−1)+3 2(N−1)+4
0 is a neighborhood of β0 , 0 is a neighborhood of − α0 ,
2(N−1)+5 2(N−1)+6
0 is a neighborhood of β1 , 0 is a neighborhood of − α1 ,
..
.
4N−1
0 0 is a neighborhood of − αN−1 .
is a neighborhood of βN−1 , 4N
Put
W0 = 00 ×10 × · · · × N−1
0 .
We now take any z ∈ Z0 , μ ∈ M0 , w ∈ W0 such that
dist (0; (z, μ) + N (z; K(w))) < λ.
Then b(w) ∈ 0 and
dist (0; (z, μ) + N (z; K1 (b(w)))) = dist (0; (z, μ) + N (z; K(w))) < λ.
By (35), we have
dist (z; S1 (μ, b(w))) ≤ γ dist (0; (z, μ) + N (z; K1 (b(w)))) .
This is equivalent to
dist (z; S (μ, w)) ≤ γ dist (0; (z, μ) + N (z; K(w))) .
Thus, we have proven that there exist constants γ > 0, λ > 0, and neighborhoods Z0 of z̄,
M0 of μ̄, W0 of w̄ such that
dist (z; S (μ, w)) ≤ γ dist (0; (z, μ) + N (z; K(w)))
whenever z ∈ Z0 , μ ∈ M0 , w ∈ W0 , dist (0; (z, μ) + N (z; K(w))) < λ.
Hence, the map M × W (μ, w) → S(μ, w) is locally-metrically regular in Robinson’s
sense around the point (μ̄, w̄, z̄, 0Z ). The proof of theorem is complete.
Acknowledgements This research is funded by Vietnam National Foundation for Science and Technology
Development (NAFOSTED) under grant number 101.01-2015.04 and by the Vietnam Institute for Advanced
Study in Mathematics (VIASM).
References
1. Arutyunov, A.V., Marinkovich, B.: Necessary optimality conditions for discrete optimal control prob-
lems. Mosc. Univ. Comput. Math. Cybern. 1, 38–44 (2005)
2. Bertsekas, D.P.: Dynamic Programming and Optimal Control, vol. I. Springer, Berlin (2005)
3. Gabasov, R., Mordukhovich, B.S., Kirillova, F.M.: The discrete maximum principle. Doklady Akademii
Nauk SSSR 213, 19–22 (1973). (Russian; English transl. in Soviet Math. Dokl. 14, 1624–1627 (1973))
4. Henrion, R., Mordukhovich, B.S., Nam, N.M.: Second-order analysis of polyhedral systems in finite
dimensions with applications to robust stability of variational inequalities. SIAM J. Optim. 20, 2199–
2227 (2010)
5. Larson, R.E., Casti, J.: Principles of Dynamic Programming, vol. I. Marcel Dekker, New York (1982)
6. Larson, R.E., Casti, J.: Principles of Dynamic Programming, vol. II. Marcel Dekker, New York (1982)
7. Lian, Z., Liu, L., Neuts, M.F.: A discrete-time model for common lifetime inventory systems. Math.
Oper. Res. 30, 718–732 (2005)
The Mordukhovich Coderivative and the Local Metric Regularity...
8. Malanowski, K.: Differential sentivity of solutions of convex constrained optimal control problems for
discrete systems. J. Optim. Theory Appl. 53, 429–449 (1987)
9. Mangasarian, O.L., Shiau, T.-H.: Lipschitz continuity of solutions of linear inequalities, programs and
complementarity problems. SIAM J. Control Optim. 25, 583–595 (1987)
10. Mordukhovich, B.S.: Difference approximations of optimal control system. Prikladaya Matematika I
Mekhanika 42, 431–440 (1978). (Russian; English transl. in J. Appl. Math. Mech. 42, 452–461 (1978))
11. Mordukhovich, B.S.: Variational Analysis and Generalized Differentiation I. Basis Theory. Springer,
Berlin (2006)
12. Mordukhovich, B.S.: Variational Analysis and Generalized Differentiation II. Applications. Springer,
Berlin (2006)
13. Mordukhovich, B.S.: Generalized differential calculus for nonsmooth and set-value mappings. J. Math.
Anal. Appl. 183, 250–288 (1994)
14. Nam, N.M.: Coderivatives of normal cone mappings and Lipschitzian stability of parametric variational
inequalities. Non. Anal. 73, 2271–2281 (2010)
15. Pindyck, R.S.: An aplication of the linear quaratic tracking problem to economic stabilization policy.
IEEE Tran. Automatic Con. 17, 287–300 (1972)
16. Quy, N.T.: New results on linearly perturbed polyhedral normal cone mapping. J. Math. Anal. Appl. 381,
352–364 (2011)
17. Rockafellar, R.T., Wets, R.J.-B.: Variational Analysis. Springer, Berlin (1998)
18. Seeger, A.: Subgradient of optimal-value function in dynamic programming: The case of convex system
without optimal paths. Math. Oper. Res. 21, 555–575 (1996)
19. Toan, N.T., Kien, B.T.: Continuity properties of the solution map to a parametric discrete optimal control
problem. J. Non. Conv. Anal. 12, 635–650 (2011)
20. Toan, N.T., Yao, J.-C.: Mordukhovich subgradients of the value function to a parametric discrete optimal
control problem. J. Glob. Optim. 58, 595–612 (2014)
21. Toan, N.T., Ansari, Q.H., Yao, J.-C.: Second-order necessary optimality conditions for a discrete optimal
control problem. J. Optim. Theory Appl. 165(3), 812–836 (2015)
22. Toan, N.T., Thuy, L.Q.: Second-order necessary optimality conditions for a discrete optimal control
problem with mixed constraints. J. Glob. Optim. 64, 533–562 (2016)
23. Tu, P.N.V.: Introductory Optimization Dynamics. Springer, Berlin (1991)
24. Yen, N.D., Yao, J.-C.: Point-based sufficient conditions for metric regularity of implicit multifunctions.
Non. Anal. 70, 2806–2815 (2009)