Proceeding Series of the Brazilian Society of Computational and Applied Mathematics

On the Continuous-Time Complementarity Problem

1
Moisés Rodrigues Cirilo do Monte
Instituto de Ciências Exatas e Naturais do Pontal - UFU, Ituiutaba - MG
2
Valeriano Antunes de Oliveira
Instituto de Biociências, Letras e Ciências Exatas - UNESP, São José do Rio Preto - SP

Resumo. This work deals with solving continuous-time nonlinear complementarity problems using
the variational inequality problem. A relation is set up so that a stationary point of an unconstrained
continuous-time programming problem is a solution for the continuous-time complementarity prob-
lem.
Palavras-chave. Complementarity, Variational Inequality, Continuous-time.

1 Introduction
Complementarity problems were firstly proposed as the question of finding an n−vector x which
satisfies the system of inequalities

x ≥ 0, Mx + b ≥ 0 and x> (M x + b) = 0 (1)

where M is an n × n matrix, b is an n−vector of real numbers and “>” denotes the transposition
of vectors and matrices. Such problems are elegant generalizations of certain linear programming,
quadratic programming, and game theory problems.
The importance of problem (1) lies in the fact that its form includes several problems by
appropriate choices of the vector b and the matrix M . As examples of applications, we can cite the
problem of the existence of solutions to linear programs (Cottle [2], Dorn [3]) that can be reduced
to a problem in the formate (1), the equilibrium point problem of bimatrix games (Lemke [8])
and the unloading problem for plane curves (Du Val [4]). For other examples of applications, see
Isac [6].
The formulation (1) was expanded to include a broader class of problems such as nonlinear
programming and was rewritten as the problem of to find an n−vector x which satisfies the system
of inequalities

x ≥ 0, f (x) ≥ 0 and x> f (x) = 0 (2)

n
where f is a mapping of R into itself. Among other authors who studied the formulation (2),
Cottle [2] gave sufficient conditions for the existence of x, and Karamardian [7] established sufficient
conditions for the existence of an unique solution.
Bodo and Hanson [1] extended the results of Karamardian [7] to the case where x is a bounded
measurable function which maps some finite interval into Rn , sufficient conditions for the existence
of a unique solution are given and applications to continuous linear and nonlinear programming
are presented.
1 [email protected].
2 [email protected].
2

In this work, we propose to solve the continuous-time complementarity problem presented by

Bodo and Hanson using the variational-type inequalities problem defined by Zalmai in [10]. In this
paper, Zalmai presents a generalized sufficiency criteria in continuous-time programming and uses
it to study the existence of a solution for the variational-type inequalities problem.
The text is organized as follows. In Section 2, we define the continuous-time complementarity
problem, the variational-type inequalities problem and we establish relationships between these
problems. In Section 3, using the Fischer-Burmeister function [5], we derive an unconstrained
equivalent problem in the sense that a stationary point of the unconstrained equivalent problem is
a solution of the continuous-time complementarity problem. Final comments are given in Section
4.

2 Variational-Type Inequality Problem

The continuous-time complementarity problem is to find x ∈ L∞ ([0, T ]; Rn ) such that, for a.e.
t ∈ [0, T ], we have that
x(t) ∈ K, f (x, t) ∈ K ◦ and x(t)> f (x, t) = 0 (3)
where K ⊂ Rn is a nonempty closed convex cone with vertex at 0, namely, if x ∈ K, αx ∈ K for
all α > 0. The polar cone K ◦ of K is given by K ◦ = {y ∈ Rn | y > x ≥ 0, ∀ x ∈ K}. L∞ ([0, T ]; Rn )
denotes the Banach space of all Lebesgue-measurable essentially-bounded n-dimensional vector
functions defined on the compact interval [0, T ] ⊂ R, with the norm k · k∞ defined by
kxk∞ = max ess sup |xi (t)|
1≤i≤n t∈[0,T ]

and f : L∞ ([0, T ], Rn ) × [0, T ] → Rn is a nonlinear function with f (·, t) Fréchet differentiable

throughout [0, T ] and f (x, ·) measurable for all x ∈ L∞ ([0, T ], Rn ). Define the following subset of
L∞ ([0, T ]; Rn ):
Ω = {x ∈ L∞ ([0, T ]; Rn ) | x(t) ∈ K a.e. t ∈ [0, T ]}.
Remark 2.1. In this text we also consider degenerate solutions for (3), in other words, solutions
such that xi (t) = 0 and fi (x, t) = 0 for some t ∈ [0, T ] and i ∈ I = {1, 2, . . . , n}.
Definition 2.1. The Variational-type Inequality Problem V IP (f, Ω) consists in finding x∗ ∈ Ω
such that, for a. e. t ∈ [0, T ],
Z T
f (x∗ , t)> (x(t) − x∗ (t)) dt ≥ 0,
0

for all x ∈ Ω, where f was previously defined.

Lemma 2.1. x∗ ∈ L∞ ([0, T ], Rn ) is a solution of the (3) if, and only if, x∗ ∈ L∞ ([0, T ], Rn ) is a
solution of the V IP (f, Ω).
Proof. If x∗ ∈ L∞ ([0, T ], Rn ) is a solution of the (3) then f (x∗ , t) ∈ K ◦ and we can conclude, for
a.e. t ∈ [0, T ], that
x(t)> f (x∗ , t) ≥ 0, for all x ∈ Ω. (4)
Using (4) and the hypothesis, for a.e. t ∈ [0, T ] and for all x ∈ Ω, we have that
f (x∗ , t)> (x(t) − x∗ (t)) = f (x∗ , t)> x(t) − f (x∗ , t)> x∗ (t)
= f (x∗ , t)> x(t)
≥ 0,
 3

resulting that x∗ ∈ Ω is a solution of V IP (f, Ω). Conversely, if x∗ ∈ Ω is a solution of the

V IP (f, Ω), then x∗ (t) ∈ K a.e. t ∈ [0, T ]. The inequality in Definition 2.1 holds for all x ∈ Ω.
Particularly, for x = 0 ∈ Ω and x = 2x∗ ∈ Ω we have that
Z T Z T
f (x∗ , t)> x∗ (t) dt ≤ 0 and f (x∗ , t)> x∗ (t) dt ≥ 0,
0 0

respectively, resulting in
Z T
f (x∗ , t)> x∗ (t) dt = 0. (5)
0
Statement: For all x ∈ Ω,
f (x∗ , t)> x(t) ≥ 0 a.e. t ∈ [0, T ].
Indeed, suppose that there exists x̃ ∈ Ω and a subset D ⊂ [0, T ], with positive measure, such that
f (x∗ , t)> x̃(t) < 0 for all t ∈ D. Define x̄ ∈ Ω given by
(
x̃(t) if t ∈ D,
x̄(t) =
0 if t ∈ [0, T ] \ D.

Then, using (5) and the definition of x̄, we have that

Z T Z T Z T
f (x∗ , t)> (x̄(t) − x∗ (t)) dt = f (x∗ , t)> x̄(t) dt − f (x∗ , t)> x∗ (t) dt
0
Z0 0
∗ >
= f (x , t) x̃(t) dt
D
< 0,

contradicting the fact that x∗ is a solution of the V IP (f, Ω). Therefore, by the above statement,
f (x∗ , t) ∈ K ◦ a.e. t ∈ [0, T ], that is, f (x∗ , t) ≥ 0 a.e. t ∈ [0, T ], resulting from (5) that

f (x∗ , t)> x∗ (t) = 0 a.e. t ∈ [0, T ].

Now, let x∗ ∈ Ω and define the auxiliary continuous-time problem

Z T
maximize P (x) = − f (x∗ , t)> x(t) dt
0
(6)
subject to x ∈ Ω.

Proposition 2.1. Let arbitrary x∗ ∈ Ω. If x∗ is a solution of the problem (3) then x∗ ∈ Ω is a

global maximum point of the problem (6) with P (x∗ ) = 0. Conversely, if x∗ is a global maximum
point of the problem (6), then x∗ is a solution of the problem (3).
Proof. As x∗ is a solution of the (3), x∗ ∈ Ω, f (x∗ , t) ∈ K ◦ a.e. t ∈ [0, T ] and
Z T
∗
P (x ) = − f (x∗ , t)> x∗ (t) dt = 0.
0

From f (x∗ , t) ∈ K ◦ a.e. t ∈ [0, T ] it follows, for allx ∈ Ω, follow that

Z T
P (x) = − f (x∗ , t)> x(t) dt ≤ 0.
0
4

Therefore, P (x) ≤ P (x∗ ) for all x ∈ Ω and x∗ is a global maximum point of the (6) with P (x∗ ) = 0.
Conversely, if x∗ is a global maximum point of the (6), then x∗ ∈ Ω and, for all x ∈ Ω, we have
that Z T
P (x) ≤ P (x∗ ) ⇔ f (x∗ , t)> (x(t) − x∗ (t)) dt ≥ 0,
0
implying that x∗ is a solution of the V IP (f, Ω). From Lemma 2.1 we have that x∗ is a solution of
the problem (3).
Remark 2.2. If x∗ is a global maximum point of the problem (6), noting that the gradient of in-
equality constraints are linearly independent, we have that the problem (6) satisfies all assumptions
of the Theorem 4.2 presented by do Monte and de Oliveira [9], that provides necessary optimal-
ity conditions for the nonlinear continuous-time optimization problem with equality and inequality
constraints.
The next example illustrates the Proposition 2.1.
Example 2.1. Consider the problem (3) with x : [0, 1] → R, f (x, t) = [x(t)]2 − x(t). If x∗
is a global maximum point of (6), then the Theorem 4.2 in [9] guarantees us that there exists
u∗ ∈ L∞ ([0, 1]; R) such that, for a.e. t ∈ [0, 1],
(i) −f (x∗ , t) + u∗ (t) = 0 ⇒ u∗ (t) = [x∗ (t)]2 − x∗ (t),
(ii) u∗ (t) ≥ 0, x∗ (t) ≥ 0 and u∗ (t)x∗ (t) = 0,
resulting that x∗ (t) = 0 or x∗ (t) = 1 for a.e. t ∈ [0, 1]. As the constraints in the problem (6) are
linear, we can conclude that the functions family {xρ } ⊂ L∞ ([0, 1]; R), 0 ≤ ρ ≤ 1, given by
(
0, if 0 ≤ t ≤ ρ,
xρ (t) =
1, if ρ < t ≤ 1,

are global minimizers of (6) and, by Proposition 2.1, solutions of (3). Note that x∗ ≡ 0 is a
degenerate solution because f (0, t) = 0 while x∗ ≡ 1 is a nondegenerate solution.

3 A Unconstrained Equivalent Problem

Let us consider K to be the positive octant of Rn . In this case, K = K ◦ . The Fischer-Burmeister
function ϕ : R2 → R is given by
p
ϕ(a, b) = a2 + b2 − a − b.
This function has the property that ϕ(a, b) = 0 ⇔ a ≥ 0, b ≥ 0, ab = 0. As the problem (6) has
no equality constraint, the Theorem 4.2 in [9] guarantees that there exists u∗ ∈ L∞ ([0, T ]; Rn ),
u∗ (t) ≥ 0 a.e. t ∈ [0, T ], such that kf (x∗ , t) − u∗ (t)k = 0 and ϕ(x∗i (t), u∗i (t)) = 0, for a.e. t ∈ [0, T ]
and i = 1, . . . , n. Consider the following unconstrained continuous-time problem:
Z T
maximize Q(x, u) = − F (x, u, t) dt (7)
0
n
where x and u are functions in L∞ ([0, T ]; R ) and, for a.e. t ∈ [0, T ],
F (x, u, t) = kf (x, t) − u(t)k2 + σ(x(t), u(t)),
where
n
X
σ(x(t), u(t)) = [ϕ(xi (t), ui (t))]2 a.e. t ∈ [0, T ].
i=1
 5

Remark 3.1. Note that the Fischer-Burmeister function is non-differentiable at (0, 0). But the
function σ above is a sum of the squared Fischer-Burmeister functions and the differential of ϕ2
at (0, 0) is zero.

Theorem 3.1. x∗ ∈ L∞ ([0, T ]; Rn ) is a solution of the problem (3) if, and only if, there exists u∗ in
L∞ ([0, T ]; Rn ) such that (x∗ , u∗ ) is a global optimal solution for the problem (7) with Q(x∗ , u∗ ) = 0.

Proof. By definition of F , if Q(x∗ , u∗ ) = 0 then F (x∗ , u∗ , t) = 0 a.e. t ∈ [0, T ]. Therefore,

σ(x∗ (t), u∗ (t)) 0 a.e. t ∈ [0, T ]

=
Xn
⇔ [ϕ(x∗i (t), u∗i (t))]2 = 0 a.e. t ∈ [0, T ]
i=1
⇔ ϕ(x∗i (t), u∗i (t)) = 0 a.e. t ∈ [0, T ], i = 1, . . . , n
⇔ x∗i (t) ≥ 0, u∗i (t) ≥ 0, x∗i (t)u∗i (t) = 0 a.e. t ∈ [0, T ], i = 1, . . . , n.

Then, for a.e. t ∈ [0, T ], we have that x∗ (t) ∈ K and

f (x∗ , t) − u∗ (t) = 0 ⇒ f (x∗ , t) = u∗ (t) ≥ 0 ⇒ f (x∗ , t) ∈ K ◦ ;

Xn
x∗ (t)> f (x∗ , t) = x∗ (t)> u∗ (t) = x∗i (t)u∗i (t) = 0.
i=1

Therefore, x∗ is a solution for the problem (3). Conversely, if x∗ is a solution of the problem (3),
by Proposition 2.1, x∗ is a global maximum point of the problem (6). Using the Remark 2.2, we
conclude that F (x∗ , u∗ , t) = 0 a.e. t ∈ [0, T ], in other words, (x∗ , u∗ ) is a global maximum point
of (7) with Q(x∗ , u∗ ) = 0.

Most of the algorithms used in the resolution of the problem (7) guarantee convergency only to
stationary points. A global minimum point is very hard to find. For this purpose, denote the n × n
jacobian matrix of f with respect to its first argument evaluated at x by Df (x, ·) and suppose that
its entries belong to L∞ ([0, T ]; R). We say that p = (x, u) is a stationary point of (7) if, and only
if, D[F (p, ·)] = 0.

Definition 3.1. The n×n matrix M (x, t), with elements mij (x, t) ∈ L∞ ([0, T ]; R), i, j = 1, . . . , n,
is positive definite at x∗ ∈ L∞ ([0, T ]; Rn ) if, for all y ∈ L∞ ([0, T ]; Rn ) and a.e. t ∈ [0, T ],

y(t)> M (x∗ , t)y(t) > 0 whenever y(t) 6= 0.

The next theorem relates stationary points of (7) to solutions of the problem (3).

Theorem 3.2. Let x∗ and u∗ be functions in L∞ ([0, T ]; Rn ). If (x∗ , u∗ ) is a stationary point of

(7) and Df (x∗ , t) is definite positive, then x∗ is a solution of (3).

Proof. For a.e. t ∈ [0, T ], let w∗ (t) = f (x∗ , t) − u∗ (t),

ϕ(x∗1 (t), u∗1 (t))

 

ϕ̃(x∗ (t), u∗ (t)) =  ..

,
 
.
ϕ(x∗n (t), u∗n (t))
 n  n
∂ϕ ∗ ∂ϕ ∗
J(t) = diag (x (t), u∗i (t)) and K(t) = diag (xi (t), u∗i (t)) .
∂xi i i=1 ∂ui i=1
6

If (x∗ , u∗ ) is a stationary point of (7), then D[F (x∗ , u∗ , t)] = 0 a.e. t ∈ [0, T ], that is,

Df (x∗ , t)w∗ (t) + J(t)ϕ̃(x∗ (t), u∗ (t)) = 0 (8)

−w∗ (t) + K(t)ϕ̃(x∗ (t), u∗ (t)) = 0 (9)

From (8) and (9) we obtain

w∗ (t)> Df (x∗ , t)w∗ (t) + w∗ (t)> J(t)ϕ̃(x∗ (t), u∗ (t)) = 0 (10)

and

w∗ (t)> = ϕ̃(x∗ (t), u∗ (t))> K(t), (11)

respectively. Using (10) and (11), we have that

w∗ (t)> Df (x∗ , t)w∗ (t) = −ϕ̃(x∗ , u∗ )> {K(t)J(t)} ϕ̃(x∗ (t), u∗ (t)). (12)

Noting that
∂ϕ ∗ ∂ϕ ∗
(xi (t), u∗i (t)) (x (t), u∗i (t)) ≥ 0 a.e. t ∈ [0, T ], i = 1, . . . , n,
∂xi ∂ui i
we have that the matrix K(t)J(t) is positive semi-definite for a.e. t ∈ [0, T ], implying that
w∗ (t)> Df (x∗ , t)w∗ (t) ≤ 0 a.e. t ∈ [0, T ]. But, by hypothesis, w(t)> Df (x∗ , t)w(t) > 0 for all
w ∈ L∞ ([0, T ]; Rn ), whenever w(t) 6≡ 0. Then

w∗ (t) = 0 a.e. t ∈ [0, T ]. (13)

By (13) and (9), for a.e. t ∈ [0, T ], we obtain

∂ϕ ∗
K(t)ϕ̃(x∗ (t), u∗ (t)) = 0 ⇔ ϕ(x∗i (t), u∗i (t)) (x (t), u∗i (t)) = 0, i = 1, . . . , n.
∂xi i
∂ϕ ∗
So, for almost every t and for all i = 1, . . . , n, we have that ϕ(x∗i (t), u∗i (t)) = 0 or (x (t), u∗i (t)) =
∂xi i
∂ϕ ∗
0. If (x (t), u∗i (t)) = 0, then x∗i (t) > 0 and u∗i (t) = 0, implying that ϕ(x∗i (t), u∗i (t)) = 0, i =
∂xi i
1, . . . , n. Therefore, in both cases,

ϕ̃(x∗ (t), u∗ (t)) = 0 a.e. t ∈ [0, T ]. (14)

Thereby, for a.e. t ∈ [0, T ], we have from (14) that x∗i (t) ≥ 0, u∗i (t) ≥ 0, x∗i (t)u∗i (t) = 0, i =
1, . . . , n, resulting from (13) that x∗ (t) ∈ K, f (x∗ , t) ∈ K ◦ and x∗ (t)> f (x∗ , t) = 0 for a.e. t ∈
[0, T ].
Example 3.1. Considering the Example 2.1, note that the solution xρ , 0 ≤ ρ ≤ 1, along with
the Lagrange multiplier uρ (t) = 0 a.e. t ∈ [0, 1] are such that (xρ , uρ ) is a stationary point for the
problem (7). Indeed, from (8) and (9), we have that
( )
2 xρ (t)
• (i) {2xρ (t) − 1}{[xρ (t)] − xρ (t) − uρ (t)} + p − 1 ϕ(xρ (t), uρ (t)) = 0;
[xρ (t)]2 + [uρ (t)]2
( )
u ρ (t)
• (ii) −{[xρ (t)]2 − xρ (t) − uρ (t)} + p − 1 ϕ(xρ (t), uρ (t)) = 0;
[xρ (t)]2 + [uρ (t)]2
 7

For all 0 ≤ ρ ≤ 1, with uρ (t) = 0 a.e. t ∈ [0, 1] and

(
0, if 0 ≤ t ≤ ρ,
xρ (t) =
1, if ρ < t ≤ 1,
we have that (i) and (ii) (iii) hold. Then, by Theorem 3.2, xρ is a solution for the problem (3),
for any 0 ≤ ρ ≤ 1. We remember that if xρ (t) = 0 and uρ (t) = 0, then D[(ϕ(xρ (t), uρ (t)))2 ] = 0.

4 Final Comments
In this work, we presented an approach to the resolution of the continuous-time complementarity
problem reformulating it as an equivalent unconstrained optimization problem. We prove that a
solution of the problem (3) is a global minimizer of the problem (6) with zero objective function
value and vice versa. As the problem (6) is a linear problem, a discretization approach can be used
to solve it.
Moreover, we showed that a stationary point of the problem (7) with zero objective function
value is a solution of the problem (3). Since the Fischer-Burmeister function squared is differen-
tiable at (0, 0), computational algorithm can be used to search for stationary points for the problem
(7), that is, solutions for the problem (3).

References
[1] Bodo, E. P. and Hanson, M. A. A Class of continuous nonlinear complementarity
problems, Journal of Optimization Theory and Applications, 24:243–262, 1978. DOI:
10.1007/BF00933280.
[2] Cottle, R. W. Nonlinear programs with positively bounded Jacobians. SIAM Journal on Ap-
plied Mathematics, 14:147–158, 1966. DOI: 10.1137/0114012.
[3] Dorn, W. S. Self-dual quadratic programs.Journal of the Society for Industrial and Applied
Mathematics, 9:51–54, 1961. DOI: 10.1137/0109006.
[4] Du Val, P. The unloading problem for plane curves.American Journal of Mathematics, 62:307–
311, 1940. DOI: 10.2307/2371454.
[5] Fischer, A. A special Newton-type optimization method.Optimization, 24:269–284, 1992. DOI:
10.1080/02331939208843795.
[6] Isac, G. Complementarity problems. Springer, 2006. ISSN: 0075-8434.
[7] Karamardian, S. The nonlinear complementarity problem with applications Part 1.Journal of
Optimization Theory and Applications, 4:87–98, 1969. DOI: 10.1007/BF00927414.
[8] Lemke, C. E. Bimatrix equilibrium points and mathematical programming. Management sci-
ence, 11:681–689, 1965. DOI: 10.1287/mnsc.11.7.681.
[9] do Monte, M. R. C. and de Oliveira, V. A. A full rank condition for continuous-time op-
timization problems with equality and inequality constraints,TEMA, 20:15–35, 2019. DOI:
10.5540/tema.2019.020.01.015.
[10] Zalmai, G. J. Generalized sufficiency criteria in continuous-time programming with application
to a class of variational-type inequalities, Journal of Mathematical Analysis and Applications,
153:331–355, 1990. DOI: 10.1016/0022-247X(90)90217-4.