Antoni Carla (Auth.), F. Giannessi, A. Maugeri (Eds.) - Variational Inequalities and Network Equilibrium Problems (1995, Springer)
Antoni Carla (Auth.), F. Giannessi, A. Maugeri (Eds.) - Variational Inequalities and Network Equilibrium Problems (1995, Springer)
Antoni Carla (Auth.), F. Giannessi, A. Maugeri (Eds.) - Variational Inequalities and Network Equilibrium Problems (1995, Springer)
Edited by
F. Giannessi
University of Pis a
Pisa, Italy
and
A. Maugeri
University of Catania
Catania, Italy
Proceedings of the International School of Mathematics "G Stampacchia" 19th Course on Variational
Inequalities and Network Equilibrium Problems, held June 19—25, 1994, in Erice, Italy
109 8 7 6 5 4 3 2 1
No part of this book may be reproduced, stored in a retrieval system, or transmitted in any form or by any
means, electronic, mechanical, photocopying, microfilming, recording, or otherwise, without written
permission from the Publisher
PREFACE
This volume brings forth a set of papers presented at the conference on "Varia-
tional Inequalities and network equilibrium problems", held in Erice at the "G. Stam-
pacchia" School of the "E. Majorana" Centre for Scientific Culture in the period 19~25
June 1994.
The meeting was conceived to contribute to the exchange between Variational
Analysis and equilibrium problems, especially those related to network design. Most
of the approaches and viewpoints of these fields are present in the volume, both as
concerns the theory and the applications of equilibrium problems to transportation,
computer and electric networks, to market behavior, and to bi~level programming.
Being convinced of the great importance of equilibrium problems as well as of their
complexity, the organizers hope that the merging of points of view coming from differ-
ent fields will stimulate theoretical research and applications.
In this context Variational and Quasi~Variational Inequalities have shown them-
selves to be very important models for equilibrium problems. As a consequence in the
last two decades they have received a lot of attention both as to mathematical inves-
tigation and applications. The proof that the above mentioned equilibrium problems
can be expressed, in terms of Variational or Quasi~Variational Inequalities also in the
non~standard and non~symmetric cases, has been a crucial improvement. Alternatives
and complements to this approach are offered by: the use of the normal map; the
belonging of zero to a multifunction; the use of gap functions; the use of models with
unilateral constraints; the separation theory; the duality theory. All these approaches
deserve attention and may lead to enlarging the class of equilibrium problems which we
can deal with in the applications. To be able to widen the applicability of the models
obviously has an important feedback in the theoretical investigation: Calculus of Vari-
ations, Stability Theory, Separation Theory, Complementarity, Duality, Generalized
Monotonicity, Normal Manifold Theory are instances of fields whose growth has been
strongly influenced by the applications. The editors hope this volume will contribute
to further exchanges between researchers in theory and those in applications.
We express our sincere thanks to all those who took part in the conference. Their
invaluable discussions have made this book possible. Special mention should once
more be made of the" E. Majorana" Centre, which offered its facilities and stimulating
environment for the meeting. We are all indebted to the Mathematical Committee and
to the group for Functional Analysis and Applications (GNAFA) of the Italian National
Research Council (CNR), and to the Italian Ministry for University (MURST) for their
v
VI
finantial support. We thank Giuliana Cai (Pisa) for the fine typing. Finallly, we want
to express our special thanks to PLENUM Publishing Co. for having offered to publish
this volume and for their constant encouragement and cooperation.
1. Introduction ............................................................ 1
2. Separation .............................................................. 3
5. Conclusions ............................................................. 18
1. Introduction ............................................................ 21
1. Introduction ............................................................ 33
vii
Vlll
1. Introduction ............................................................ 45
1. Introduction ............................................................ 55
1. Introduction ............................................................ 69
1. Introduction ............................................................ 85
1. Introduction 101
2. Gap functions for quasi-variational inequalities. The weak case .......... 102
1. Introduction 123
3. Convergence............................................................ 147
Antoni Carla
Key words: Variational Inequality, separation of sets, Lagrangean function, gap func-
tion, duality.
1. INTRODUCTION
1
2
Let y E Dom f be fixed; consider the function .p(y; . ) : 3 --. R U {-oo}, with
.p(y; x) > 0, x E3
is impossible.
Now, another system is introduced to characterize a solution y of VI. More pre-
cisely, fix a vector space V and a function cp(y; ., . ) : 3 x V --. RU {-oo}, and assume
the function cp(y; ., . ) satisfies the following condition:
OEVo~V. (1.2)
Lemma 1.1 Let y E Domf be fixed. The system (Sy) is impossible if and only if the
following system (in the unknowns x, v);
cp(y;x,v) > 0, x E 3, v E Va
is impossible.
In fact, in terms of separation arguments, Lemma 1.1 becomes the following proposi-
tion, whose obvious proof shows Lemma 1.1 too.
Lemma 1.2 Let V be a vecto? space, cp(y; ., . ) : 3 X V --. R U {-oo} a function such
that (1.1) holds, and let Vo be a set which fulfils (1.2). Then y E Domf is a solution
of VI if and only if
qy)n1i=0. (1.3)
2. SEPARATION
The disjunction of K(y) and 1i is studied by means of their inclusion in level sets
of suitable functionals defined on R xV, depending on a parameter w which belongs
to a fixed class fl (wallows to describe the class of the functionals w(·, . ;w) ). The
class fulfils the following conditions:
VwEfl,
i) w( . , '; w) : R x V -+ R, w( u, V; w) = u + "Y( V; w)
ii) lev~o w(·, . ;w) ;21i.
Remark 2.1. When Vo = R+ the class of functionals considered above becomes that
introduced in [4J.
Remark 2.2. If the type of inclusion in ii) is reversed, a different kind of separation is
obtained. We call it " strong separation". By symmetry we call " weak " that defined
by i), ii); it is conceivable and interesting to investigate the strong case too.
From ii) it follows that (1.3) is true if there exists w E fl such that
that is
inf sup [cp(y; x, v) + "Y(v;w)] ~ O. (2.2)
",EO (z,v)EB><V
defined by
A(y;x,w):= inf [-cp(y;x,v)-"y(v;w)],
vEV
Proposition 2.1. Let Y E Domf be fixed. If there exists w E 0 which fulfils condition
(2.1), then (y,w), is a saddle point for for the Lagrangean A(Yi ., .) on B x n .
A(YiY,W) = 0i
hence, we achieve the inequalities
A(YiY,W) ~ A(YiY,W), Vw E O.
Remark 2.4. Let V = B and let It'(Yi ., .) : B x ::: -+ R U { -oo} be the function
defined by
It'(Yi x,v) = (F(y),y - x) + f(y) - f(x + v).
Moreover set Va = {OJ, 0 = B, ,),(ViW) = -(v,w). Then, the Lagrangean function
A(Yi ., . ) coincides with that considered in [2] to obtain the Dual Variational Inequality
proposed by Mosco [5].
Remark 2.5. The present approach can be applied to any problem which may be
reduced to the impossibility of a system like (Sy). This happens, in particular, for
extremum problems: using this approach it is possible to derive Lagrange and Fenchel
duality [1].
3. GAP FUNCTIONS
Using the previous approach we are able to consider a general class of gap func-
tions.
'f'(y;x,O) > 0;
then
sup [- A(y;x,w)] ~ 'f'(y;x,O) +,(O;w) ~ 'f'(y;x,O) > 0.
xES
Remark 3.1. The existence of an wEn, such that lev~o w(·, . ;w) ;2 K(y), is a
sufficient condition to prove that
Then, to prove the function 9 defined in (3.1) is a gap function, we need some assump-
tions about K(y), in particular about the function 'f'(y; " . ).
Remark 3.2. The function 9 defined in (3.1) collapses in those introduced in [3] when
the functions 'f'(y; . , . ) and ,( . ; w) are specialized.
(F(y),x - y) ~ 0, 'Ix E K,
sup [h(y)-{A,By-b)] .
• e1R~.
Vh(.)=.B
={
(F(y),y-x),ifBX~b+V,
<P ( yjX,V ) .
-00, otherwise,
Remark 4.1. Let y E K be fixed. The value g(y) of the gap function g defined in
(4.1), can be written as follows:
where the supremum is with respect to >.. This expression for 9 suggests to use Wolfe's
duality theory for extremum problems to find again 9 is a gap function.
Consider the following Complementarity Problem (in short, CP): find y E C such
that
(F(y), y) = 0, F(y) E C* ,
In the literature, among the gap functions associated with CP, there is the follow-
ing g : C -+ R with
g(y) = (F(y), y) .
The problem
(P) inf{g(y), yEC, F(y)EC*},
turns out to be equivalent to CPo
The function g and the problem (P) can be derived from the previous scheme. In
fact, CP coincides with the following VI: find y E C , such that
(F(y),x-y)~O, VxEC.
Remark 5.1. This VI is obtained from the one introduced in Sect. 1 setting f = oe,
the indicator function of C.
REFERENCES
[1] C. Antoni, "Dual problems and separation of sets. Lagrange and Fenchel duality". To appear.
[2] M. Castellani and G. Mastroeni, "On the duality theory for finite Variational Inequalities", this
volume.
[3] F. Giannessi, "Separation of sets and Gap Functions for Quasi-Variational Inequalities", this
volume.
[4] F. Giannessi, "Theorems of the alternative and optimality conditions". lou. Optim. Theory
Appl., 42 N. 3, pp. 331-365, 1984.
[5] U. Mosco, "Dual Variational Inequalities", Journal of Mathematical Analysis and Applications,
40, 202-206, 1972.
[6] R.T. Rockafellar, "Convex Analysis", Princeton University Press, Princeton, 1970.
TRAFFIC SCHEDULING IN TELECOMMUNICATION SYSTEMS AND
NETWORK FLOW
Bonuccelli Maurizio A.
1. INTRODUCTION
9
10
2. BASIC PROBLEM
line (row or column). Observe that when S has exactly one entry equal to 1 in each
line, it is a permutation matrix . Given this representation, a TSA can be seen as
a decomposition of D into switching matrices. So, we have the following minimum
length scheduling problem (also called ML TSA):
Given a nonnegative integer (traffic) matrix D, decompose it into the smallest
number of switching matrices, namely find 51, 52, ... , 5 L such that D = 2:r=o 5 i , and
L is minimum.
We shall first recall a classical linear algebra result. A doubly stochastic matrix is
a nonnegative matrix in which each line (i.e. row or column) has sum equal to 1.
The classical Birkhof f - vonN eumann theorem states that any doubly stochastic
matrix is a convex combination of permutation matrices (or, equivalently, the doubly
stochastic matrices are the convex envelope of the set of permutation matrices).
Birkhoff-von Neumann theorem could be the theoretical basis for our TSA prob-
lems. Unfortunately, it deals with non-integer matrices, while we are interested in
matrices with integer entries only. Thus, we need to generalize in some way that theo-
rem. Let us define pseudo doubly stochastic matrices as nonnegative integer matrices
with identical line sums. Notice that a doubly stochastic matrix can always be ob-
tained by a proper pseudo doubly stochastic one by dividing each entry by the line
sum (and viceversa, by multiplying each entry by the inverse of the smallest nonzero
entry).
Let us go back to our TSA problem. A lower bound L on the length of any schedule
for it can be readily obtained as follows:
Theorem 1. Let us assume that traffic matrix D is pseudo doubly stochastic. Then,
we can extract a permutation matrix from it, or equivalently the associated bipartite
graph has a perfect matching (each node is incident in one edge of the matching).
Proof. Let us consider the bipartite graph associated to D, with one edge joining
Ui E U with Vi E V whenever entry di,i 2: 0, instead of di,i parallel edges, and let us
assign a capacity equal to di,j to such an edge. Then, we add a source node s, and an
edge between s and each node Ui E U, as well as a sink node t, and an edge between
each node Vi E V and t. All these edges have capacity 1. It is easy to see that a
feasible flow in the above network corresponds to a matching (just pick up the (Ui, Vj)
edges with non zero flow). That matching is perfect if and only if all the nodes are
involved in positive flows. This is true if the cut separating s from all the other nodes
12
The above theorem can be used to build an optimal algorithm for our basic TSA
problem. In fact, if D is pseudo doubly stochastic, and S is a permutation matrix
extracted from it, then D'=D-S is again pseudo doubly stochastic. Besides, if D had
a lower bound equal to L, then D' has a lower bound L'=L-l. If we repeat the above
procedure untill we are left with a zero matrix, we have a decomposition of D into L
permutation matrices, an optimal TSA.
time complexity now is O(n 4 .S ), which is fully polynomial, independent of the size of
the input numbers [4].
The traffic matrix is rarely in pseudo doubly stochastic shape. What can we do
when the traffic matrix has a generic shape, and is not pseudo doubly stochastic?
Fortunately, it is possible to add some dummy traffic to the entries of the matrix
so that it is transformed into a pseudo doubly stochastic one. The resulting matrix
has the same lower bound on the schedule length of the original matrix, and the
transformation algorithm takes only O(n 2 ) time [4}. Essentially, the algorithm (called
filling) is as follows:
Filling Algorithm
for i:=l to N do
for j:=l to N do
di,j f- di,j + L - max { E7=1 di,j; E:'=1 di,j }
Obviously, the dummy traffic must be deleted from the schedule as soon as it is
produced.
When the traffic matrix D has a special structure, faster algorithms that take into
account such structure, can be used.
As we shall see, the above approach can be used to solve in polynomial time other,
more complex problems. For instance, the above algorithm can be used to optimally
solve the TSA problem in which the traffic matrix is not a square one, but has a num-
ber of rows different from that of the columns, and when there is a further constraint
on the total number of nonzero entries that can be present in the switching matrices.
This constraint takes into account the presence of the so-called transponders (fre-
quency shifters) on board of the satellite, thus limiting the number of simultaneous
communications [5J.
Observe that the same constraints, with ai=1 and (3j=l, hold for the basic problem.
An a(3 -switching matrix S is called complete whenever the above inequalities are tight.
A traffic matrix D is called normalized pseudo doubly stochastic (NPDS for short) if
there exists some integer number L such that:
L = max { r -k Ei=t di,; 1, for each i;f t Ei=t di,j 1, for each j j
Thus,
15
and so,
We consider now the case of general switching modes [7]. In this case, we have the
same constraints of the variable bandwidth systems, with the additional constraint that
simultaneous packet transmission between any I/O pair is not allowed. Mathematically,
this means that each af3-switching matrix must have only 0 or 1 entries: summarizing,
we have the constraints presented in the previous section, with the additional constraint
Si,; E {O, I}, for each i andj. An af3-switching matrix meeting this additional constraint
is called general af3-switching matrix, or Gaf3SM. Some of the properties already
established for variable bandwidth systems, carryover to the general switching modes
problem studied in this section, subject to some changes. For instance, the lower bound
to the minimum schedule length, L, is the same as before, with the following additional
inequality: L ~ { di ,;, for each i and j }. Thus, in this case we can have critical lines,
and critical entries, also (these being those entries di,;=L). The following tightness
parameters must be defined, in order to adapt the optimal algorithm shown above to
the present problem [7].
16
where Ai,j is the lower bound and Xi,i the capacity of the edge (ij).
Let w be such that s is in it, while t is not, in our bipartite network. Then, Hoffman's
inequality becomes the following:
A similar argument applies for the other three cases, namely when both s and tare
in Wj when t is in w, while s is notj nd finally when both sand t are not in w.
Like in the previous cases, an optimal algorithm can be constructed from this
property of the GO'flSM. It suffices to find a feasible circulation instead of a maximum
flow in the bipartite network [7J. Again, a multiplicity factor is in order, for the
algorithm to be polynomial time. Now, such a factor, J.L must meet the following
17
constraints [7]:
l
I J~l I,)
J1:'S (Li3J
i3 _- "n_ . )d"J)j , lOr each J. such th at
£ "n . aller t h an f3i;
L-i:l Si,i IS sm
J 1=1 I,)
J1 _
< l(KL-2:~=l2:;=l d"J)j , wh en "n
K-2::~,2:;=18'.J) L-i:l
"n
L-j:l
.
Si,j IS
11
sma er
th
an
K;
J1 :'S L - di,j if Si,j = OJ
jl :'S di,j if Si,j =1.
Such a multiplicity factor allows the algorithm to have a time complexity O( <P
n 2 ), where <p is the time complexity of your favourite algorithm for finding a feasible
circulation [1]. In fact, at most O(n 2 ) Gof3SM must be in the TSA since each time
we produce one such matrix, either jl = di,j (and so in the remaining traffic matrix,
D', we have at least one more zero entry), or jl is smaller than di,j, in which case D'
will have at least one more critical item (row, column, or entry). The details of the
algorithm and its properties for general switching modes can be found in [7].
A similar approach can be used to give a polynomial time optimal algorithm for
the so called hierarchical switching systems (HSS) [6]. Both the inputs and the
outputs are grouped. Input group Uh is the multiplex of Ph inputs, which can produce
up to qh parallel signals, while up to Xk parallel signals are demultiplexed into Yk
streams, corresponding to the Yk outputs in group Vk. Such an arrangement have been
proposed to better utilize the system when there is low traffic between the inputs and
the outputs. Now, we have a physical constraint, additional to the basic ones: at most
qh (Xk' respectively) entries can be selected out of the Ph rows (Yk columns, resp.) of
group Uh (Vk ).
The corresponding switching matrix, called hierarchical switching matrix, or HSM,
must satisfy the basic constraints, and the following two:
more strict delay constraint). So, we first derive an optimal schedule for V. In such
a schedule, some switching matrices will not contain all the possible non-zero entries,
and some D traffic can be scheduled in it. The problem consists in placing as much
D traffic as possible into the schedule for V, without violating the basic constraints
(at most one nonzero entry in each line of every switching matrix). This problem is
NP-complete even in very restricted cases. The NP-completeness proof is based on a
reduction of the well known timetable design problem to the TSA one [8]. An effective
heuristic for the incremental TSA problem have been proposed. Such an heuristic is
based on the Lagrangean Relaxation of the integer linear programming formulation of
the problem, consisting in a series of independent Hitchcock-Koopmans transportation
(namely, maximum profit flow) problems [8].
5. CONCLUSIONS
In this paper, we have presented several problems, called time slot assignment,
which are related to traffic scheduling in computer networks and telecommunication
systems based on interconnection networks. We have shown how network flow can be
used to efficiently solve such problems, whenever possible, namely when the problems
are not NP-complete.
Time slot assignment problems have simple expressions for the lower bound on the
schedule length. An interesting phenomenon has shown up to now in this setting:
whenever the lower bound can be achieved for any traffic matrix, the problem can be
solved by polynomial time network flow based algorithms. When it is possible to find a
traffic matrix for which any schedule is longer than the lower bound, NP-completeness
has always been established. An interesting question arises: is it possible to show that
the above phenomenon always holds, with no counter-examples?
Several problems are left open in this area. Two outstanding open problems are
the TSA of on line requests, and the scheduling of multicast communication. In the
first, instead of the traffic matrix gathering, packets must be scheduled as soon as they
are generated. This could speed up the overall system operation, and is strictly related
to the successfull ATM communication protocol. The second problem arises when the
same message must be sent from one source to multiple destinations, and is typical
of emerging applications, like teleconferencing. In this case, we cannot use a matrix
or a bipartite graph to represent the traffic to be scheduled: a hypergraph is a more
appropriate model for the problem.
REFERENCES
[I] R.K. Ahuja, T.L. Magnanti, and J.B. Orlin: "Network flows: theory, algorithms and applications",
Prentice Hall, New York, 1992.
[2] P. Barcaccia and M.A. Bonuccelli: "A polynomial time optimal algorithm for time slot assignment
of variable bandwidth systems", ACM-IEEE Trans. Networking., to appear.
[3] A.A. Bertossi, G. Bongiovanni, and M.A. Bonuccelli: "Time slot assignment in SS/TDMA systems
with intersatellite links", IEEE Trans. Commun., 35, pp. 602-608, 1987.
[4] G. Bongiovanni, D. Coppersmith, and C.K. Wong: "An optimal time slot assignment for an
SS/TDMA system with variable number of transponders", IEEE Trans. Commun., 29, pp.721-
726, 1981.
[5] G. Bongiovanni, D.T. Tang, and C.K. Wong: "A general multibeam satellite switching algorithm",
IEEE Trans. Commun., 29, pp.l025-1036, 1981.
19
[6] M.A. Bonuccelli: "A fast time slot assignment algorithm for TDM hierarchical switching systems" ,
IEEE Trans. Commun., 37, pp. 870-874, 1989.
[7] M.A. Bonuccelli: "A polynomial time optimal algorithm for satellite-switched time- division mul-
tiple access satellite communications with general switching modes", SIAM Journal on Discrete
Mathematics, 4, pp. 28-35, 1991.
[8) M.A. Bonuccelli, I.S. Gopal, and C.K.Wong: "Incremental time slot assignment in SS/TDMA
satellite systems", IEEE Trans. Commun., 39, pp. 1147-1156, 1991.
[9) S. Chalasani and A. Varma: "Efficient time slot assignment algorithms for SS/TDMA systems with
variable bandwidth beams", Proc. IEEE INFOCOM '91, Miami, pp. 658-667, April 1991.
[10) K.Y. Eng and A.S. Acampora: "Fundamental conditions governing TDM switching assignments in
terrestrial and satellite networks", IEEE Trans. Commun., 35, pp. 755-761, 1987.
[ll) A. Ganz and Y. Gao: "SS/TDMA scheduling for satellite clusters", IEEE Trans. Commun., 40,
pp. 597-603, 1992.
[12) M.R. Garey and D.S. Johnson: "Computers and intractability: a guide to the theory of NP-
completeness", Freeman, S. Francisco, 1979.
[13] I.S. Gopal, G. Bongiovanni, M.A. Bonuccelli, D.T. Tang, and C.K. Wong: "An optimal switching
algorithm for multibeam satellite systems with variable bandwidth" ,IEEE Trans. Commun., 30,
pp. 2475-2481, 1982.
[14) I.S. Gopal, M.A. Bonuccelli, and C.K.Wong, "Scheduling in multibeam satellites with interfering
zones" , IEEE Trans. Commun., 31, pp. 941-951, 1983.
[15) I.S. Gopal , D. Coppersmith and C.K. Wong: "Minimizing packet waiting time in a multibeam
satellite system", IEEE Trans. Commun., 30, pp. 305-316, 1982.
[16) I.S. Gopal and C.K. Wong, "Minimizing the number of switching in an SS/TDMA system", IEEE
Trans. Common., 33, pp. 497-501, 1985.
[17) T. Inukai: "An efficient SS/TDMA time slot assignment algorithm", IEEE Trans. Commun., 27,
pp. 1449-1455, 1979.
[18] J.L. Lewandowski, J.W.S. Liu, and C.L. Liu: "SS/TDMA time slot assignment with restricted
switching modes", IEEE Trans. Commun., 31, pp.149-154, 1983.
[19) Y.K. Tham, "Burst assignment for satellite-switched and earth-station frequency-hopping TDMA
networks", lEE Proceedings, 137, pp. 247-255, 1990.
[20) W.W. Wu, "On the efficacy of traffic assignment in on Digital Satellite Communication, Montreal,
P.Q., Canada, October 1978.
[21] W.W. Wu, "Elements of digital satellite communication", Computer Science Press, Rockville MD,
1985.
ON THE DUALITY THEORY FOR FINITE
DIMENSIONAL VARIATIONAL INEQUALITIES
1 Dipartimento di Matematica
Via Buonarroti, 2 -56127 Pisa
e-mail [email protected]
2Dipartimento di Matematica
Via Buonarroti, 2 -56127 Pisa
e-mail [email protected]
1. INTRODUCTION
Duality theory for optimization problems has been widely developed in the liter-
ature (see [2,5,6]). In [2] it has been proven that the dual problem can be obtained by
searching for a suitable separating function defined in the image space associated to
the optimization problem. Following this scheme, an optimality condition for a Vari-
ational Inequality (VI) has been stated [1] by means of separation arguments, in a
suitable parametric image space and a parametric dual problem has been associated to
the Variational Inequality. In this work we will show how it is possible to recover the
dual Variational Inequality introduced by Mosco [4], from the general parametric dual
problem and we will investigate the relations between the dual Variational Inequality
and a saddle point condition of a generalized lagrangean function associated to (VI).
21
22
are the Fenchel conjugates, in convex and concave sense respectively, of the functions
f and 9 and are defined on the dual space (RR)* ~ RR;
the subdifferential of the convex function f at Xo ERR is the set
In this section we summarize the duality scheme for Variational Inequalities in-
troduced in [1]. Consider the following Variational Inequality
is impossible.
Fix a vector space V and a function <t'(y;.,.) : RR xV -+ R that satisfies the
following property:
o E V- de!
= {V E V: <t'(y;x,v) = </l(y;x), "Ix E Dom<t'(y; ·,v)}. (2.1)
Then, y E Domf is a solution for (GVI) iff the system (PS)y in the unknowns x and
V:
<t'(y;x,v) > 0,
{
V E Yo
is impossible, where Yo is any subset such that
o E Vo ~ V.
23
It is easy to see that the system (PS}y is impossible iff the previous subsets are disjoint.
a function such that (2.1) holds. Then y is a solution for (GVI) iff
The disjunction of K(y) and'H is studied introducing the class of separation func-
tions depending on a parameter w E Q
w(·,·;w): RxV -+ R
The previous remark leads to introduce a dual problem associated to the system
(PS)y; let A(y;·,·) : R n xQ ~ R
Definition 2.1. Let y E Dom! be fixed. The dual problem associated to the system
(PS)y is defined by
sup inf A(y; x, w) (PD)y
"'EO xERn
24
In this section we will see how, considering the dual problem introduced by means
of the separation scheme, it is possible to recover the dual Variational Inequality in-
troduced by Mosco [4].
Consider the following generalized Variational Inequality:
Theorem 3.1. [4] If F has an inverse and f is a proper, convex and l.s.c. function,
then:
Lemma 3.1. (xo,w) ERR X RR is a saddle point for AFCii;x,w) if and only if Xo E
8f*( -F(fj)) and w = F(f}).
Proof. (xo,w) E R n x R n is a saddle point for AF(fj;x,w) {::? the following
inequalities are fulfilled:
a) (w,xo) + f*(-w)::; (w,xo) + f*(-w),
b) (F(y),xo) - (w,xo) ::; (F(fj),x) - (w,x),
From a), by the definition of subgradient, we obtain Xo E 8f*(-w); from b) we
get that it must be necessarily w = F(y). 0
25
Theorem 3.2. If P possesses an inverse and f is a proper convex and l.s.c. function,
then
(1) y E R n is a solution for (GVI) {::} (y, P(y» E R n x R n is a saddle point for
AF(Yj x, w).
(2) it* = -P(jj) E R n is a solution for (DGVI) {::} (fi,P(fi» E R n x R n is a saddle
point for AF(Y; x, w).
Proof.
(1) *) Y is a solution for (GVI) {::} -P(y) E af(y) {::} [5] Y E af*( -P(y» and for
Lemma 3.1 we have the thesis.
{=:) From Lemma 3.1 we know that y E af*( -P(y» {::} [5] - P(j}) E af(Y)·
p-l(_it*) E aJ*(it*).
Remark 3.1. Theorem 3.1 follows immediately from Theorem 3.2; moreover Theorem
3.1 can be proven, more generally, in a Hausdorff locally convex space [4].
Remark 3.2. It is possible to obtain directly Theorem 3.1, without considering a sad-
dle point condition for A F , starting from the definition of the dual problem associated
to the system (PS)y
sup inf AF(Y;X,w).
wEn rERR
We have that
It is interesting to consider the case in which the Variational Inequality has the
integrability property that is 3h E C1(Rn;R) convex such that P = "ilh.
26
Proposition 3.1. If h and f are convex functions and h E Cl(Rn; R) is such that
F = Vh then y is a solution of (GVI) if and only if y is a solution of the following
extremum problem (P)
min {h(x)
xER n
+ f(x)}.
Proof. We observe that:
It is possible to show that the dual Variational Inequality (DGVI) is equivalent (up
to a symmetry) to the classical first order optimality condition for the dual problem
(D).
Theorem 3.3. Suppose that F possesses an inverse and that there exists a convex
function h E C1(Rn; R) such that F = Vh; then u* is a solution for (DGVI) {:}
-u· is an optimal solution for (D).
Proof. From the relation (- f)*(x*) = -1*( -x*), putting 'Ij;*(x*) = 1*( -x*) we
obtain that (D) is equivalent to the problem:
x* is an optimal solution for (D) {:} 0 E 8('Ij;* + h*)(x*) = 8'1j;*(x*) + 8h(x*) {:}
Remark 3.3. We note that, given a saddle point (y, F(y» of AF, a solution of
(DGVI) can be obtained considering the vector -F(Y). If we want the saddle point
condition to be equivalent to the problems (GVI) and (DGVI), we need that F(y) is
a solution of (DGVI). This can be obtained putting z' = -u' in (DGVI), therefore
we have the following dual Variational Inequality:
where Q = {x E R n : g(x) E C}, g(x) := (gl(X), ... ,gm(x», with g;(x) nn-+
R, i = 1, .. , m and C E R m is a closed convex cone with vertex in the origin of nm.
Following the scheme introduced in Sect.2, consider the Lagrangean function
AL(y;x,w) obtained choosing l'(v,w) = (v,w), V = n = R m , Vo = C and
Proposition 4.1.
Proof. By definition
Since
Lemma 4.1. Suppose that 9 is a differentiable C-concave function and that 3xo E RR
such that g(xo) E intC. Then y E Q is a solution of (VI) <=> 3X E R m such that
(y, X) is a solution of (S).
Proof. We observe that y is a solution for (VI) iff Y is a global optimal solution of
the problem (Ply
min(F(y), x - y).
zEQ
Remark 4.1. If there exists hE C1(Rn, R) such that F = Vh, the system (S) is the
Kuhn-Tucker necessary condition of the problem (P), defined in section 3.
5 (F(y),x - y) - (X,g(x»)
that is
-(A, g(y») 5 -(X, g(y») 5 (F(y), x - y) - (X, g(x »). (4.2)
Ab absurdo if g(y) ¢ C = (C·)· there exists >. E C· such that (>., g(y») < o. Since
VOl > 0 we have 01>. E C· we obtain
Therefore g(y) E C and we have that (X, g(y») 2: 0 which implies, with (4.3)
(X, g(y») = O.
The second inequality in (4.2) is equivalent to
Theorem 4.1. Suppose that 9 is a differentiable C-concave function and that 3xo E
R n such that g(xo) E intC. Then jj is a solution of (VI) {::} 3X E C· such that
(jj,X) is a saddle point of AL(jjiX,>') on R n xC·.
Remark 4.2. From theorem 4.1, we can interpret Xas the solution of a (Lagrangean)
dual Variational Inequality (DVI) and the system (S) represents the relation between
the solutions jj and X.
The dual Variational Inequality (DVI) is defined in order to guarantee that Xis
a solution of (DVI) if and only if (jj, X) is a saddle point of AL(jji x, A).
Consider the set
The dual Variational Inequality (DVI) can be defined in the following way:
{
Find (y, X)_ E Q such that
(4.5)
(g(fl), >. - >.) 2: 0, V>' E C*.
Another way of formulating the system (S) by means of a Variational Inequality is the
following:
Find (y, X) E R n xC· such that
{ (4.6)
(r(fl, X), (x - y, >. - X») 2: 0, Vex, >.) E R n xC*.
where
Remark 4.3. In the hypotheses of Lemma 4.1 it is immediate to see that (fl,X) is a
solution of (4.5) or (4.6) if and only if it is a solution of the system (S).
Consider the Variational Inequality (VI) introduced in the previous section in the
hypothesys that the feasible set Q is a closed convex cone in R n containing the origin.
To this aim we define g(x) = x and C = Q. It is well known [3] that the problem (VI)
collapses into the generalized complementarity problem defined as follows:
w = F(fl), (F(y), y) = 0.
Therefore, since Q* is a convex cone, we can define the following dual of the problem
(GCP):
Find w E Q* such that
{ (DGCP)
(F-l(W),w) = 0, F-1(w) E Q
Remark 5.1. The same result would be obtained if we considered the Lagrangean
function AF defined in section 3, since the constraints of the feasible set Q are given
by the identity function g(x) = x. In fact (DGCP) has been obtained in [4] as dual of
the Variational Inequality (GVI) in which ¢>( x) is the indicatorfunction of the closed
convex cone Q.
REFERENCES
Cubiotti Paolo
Department of Mathematics
University of Messina
98166 Sant'Agata-Messina, Italy
e-mail: [email protected]
1. INTRODUCTION
33
34
Even though in the original formulation of Theorem 1 the values of the multifunction f3
were not supposed to be closed, the reader can verify that such requirement is necessary
in order to make the proof correct. Also, it is easily seen that Theorem 1 does not work
if f3 is not closed-valued. To this aim, it suffices to consider X = [0,1), f3(x) = )0, 1[
and <jJ(x) = {x}. Very recently, the following result was proved [2).
(i) the set <jJ(x) is nonempty and compact for each x E X, and convex for each x E K,
with x E f3(x);
(iii) f3 is a lower semicontinuous multifunction with closed graph, and, for every x E
X, f3( x) is a convex set which meets K;
Then GQVI (X, f3, <jJ) has at least one solution belonging to K X R n •
Given two metric spaces (5, p), (V, d) and a multifunction t/J : 5 - t 2 v, we say that
t/J is lower semicontinuous in 5 if the set t/J-(n) = {s E 5: 1jJ(s) n n:f:. 0} is open in 5
n
for each open ~ V. We recall that a nonempty-valued 1jJ is lower semicontinuous in
35
S if and only if for any sequence {sd in S converging to s and any v E .,p(s), there is
a sequence {vd converging to v such that Vk E .,p(Sk) for all kEN. We say that the
multifunction .,p is upper semicontinuous in S if .,p-(n) is closed in S for each closed
n ~ V. We say that .,p is continuous in S if it is both lower and upper semicontinuous.
The graph of .,p is the set {( s, v) E S x V : v E .,p( s)}. We recall that if the graph of
the multifunction .,p is closed, then .,p has closed values; if .,p is upper semi continuous
with nonempty closed values, then the graph of .,p is closed; if Y is compact and .,p has
closed graph, then .,p is upper semicontinuous. We say that .,p has open lower sections
if the set .,p-({ v}) is open in S for any v E V. It is clear from the definition that if.,p
has open lower sections, then it is lower semicontinuous. Given A, B ~ V, A, B # 0,
the Hausdorff distance dH(A, B) is defined by
for all s, u E S. If L < 1, the multifunction .,p is said to be a contraction. The reader
is referred to [3,4] for an excellent treatment of the basic facts about multifunctions.
If So E S, r > 0, we put B(so,r) = {s E S: peso,s) < r} and R(so,r) = {s E S:
(ii) the set 1>( x) is nonempty and compact for each x EX;
Then there exists at least one solution to GQVI (X, f3, <p) belonging to K x RR.
Proof. Fix kEN. By assumptions (iv) and (v) and Lemma 3.1 of [5], the multi-
function x -> f3( x) n D" is continuous on X n D" with nonempty closed convex values.
Since X n D" is convex and compact, the original proof of Theorem 2 (see [2], Proof
of Theorem 1) shows that there exists a vector x" EX n D" such that X" E f3(x,,) and
(3.1)
To see this, it suffices to observe that in the proof of Theorem 2 the convexity as-
sumption on the set <p( x) is needed only in the last step, in order to apply a minimax
result. Now, assumption (vi) implies x" E K. Since K is compact, the sequence {xA;}
admits a subsequence, still denoted by {xA;}, converging to a point x E K. Of course,
assumption (iv) implies x E f3(x). We claim that
To see this, we argue as in the last part of the proof of Theorem 1 of [2]. Assume that
there exists fj E f3( x) such that
inf (z,
zE</>(i:)
x- fj) > O.
Choose kEN in such a way that K ~ Die and y E Die. Since the multifunction
x -> f3(x) n Die is continuous on X n D k , there exists a sequence {vd converging to y
such that
Since assumption (iii) is actually satisfied for each y E span(X - X) (see [2]), the
multifunction ehas open lower sections and thus it is lower semicontinuous. Moreover,
37
concavity of y -+ inf (z, y) implies that each set ~(x) is convex and open in span(X -
zEq,(x)
X). By Proposition 3 of [2], there exists 0' > 0 such that
inf (z, w) > 0 for all x E B(x, 0') n X, w E B(x - g, 0') n span(X - X).
zEq,( x)
Choosing kEN such that k 2:: k and max{lIxk - xII, IIvk - gil} < 0'/2, we obtain
Since Vk E (3(Xk)nDk, this contradicts (3.1). Hence, (3.2) holds. Since by assumptions
(i) and (ii) the set ¢>(x) is compact and convex, Theorem 5 at p. 216 of [6] implies that
there is some z E ¢>( x) such that
Proposition 1. Let S be a closed convex subset of R", z E R" and x E riC S) such
that
(z, x - y) :::; 0, Ir/y E S. (3.3)
(x - v) + ..\(w - v) E S - v.
Remark 1. It can be easily checked that Theorem 4 can fail if any of the assertions (A),
(B) and (C) is removed from the statement. To see this, one can take n = 1, X = R,
f3(x) = R as regards assertion (A)j n = 2, X = [-1,1] x R, f3(u, v) = {u} x [-1,1]
as regards assertion (B)j X = Rn,f3(x) = {O} as regards assertion (C). Also, it should
be noticed that assertion (C) is contradicted, in particular, if int(f3( x)) =1= 0 for every
x E X, with x E f3(x). To see this, choose any w E R n \{O}, x E {x EX: x E f3(x)},
v E int(f3(x)). If we pick any A > 0 such that VA = v + Aw E f3(x), we get
Proof of Theorem 4. By the classical Fan-Kakutani fixed point theorem (see [7],
Theorem 1), it is easily seen that the set E = {x EX: x E f3(x)} is nonempty. Now,
assume that assertions (A) and (C) do not hold. Thus, there is some w E R n such
that for every x E E one has
( ) _{1jJ(x), ifxEE,
'Y x - Rn, if x E X\E,
and then to apply Theorem 3.1111 of [91. Now, if we apply Theorem 3 with K = E and
</>(x) = {f(x)}, we get that there exists a point x E E such that
Suppose that x E ri(f3( x)). Then, Proposition 1 implies that f( x) E (f3( x) - f3( x)).L,
and this is a contradiction since f( x) E 1jJ( x). In fact, the inclusion f( x) E (f3( x) -
f3(x )).L n (f3(x) - f3(x)) implies f(x) = 0, while the inclusion f(x) E {y E R n : (w, y) >
O} implies f( x) =1= O. Thus, the point x lies in 8r (f3( x)), as desired. 0
39
Given T > 0, two matrix functions A : [0, T] -+ R nxn and B : [0, T] -+ R nxm , a
nonempty compact set n ~ R m, we say that an absolutely continuous x(·) : [0, T] -+
R n is a solution of the linear control system
X' = A(t)x + B(t)u, u(t) E n (4.1)
For Xo ERn, we shall denote by AT(XO) the attainable set for the system (4.1) at time
T starting from Xo, namely, the set
Of course, each fixed point of the multifunction AT gives a periodic solution of problem
(4.1) in [O,T]. We shall denote by AT(xo) the attainable set at time T starting from
Xo for the relaxed problem
Before giving the main result of this section, we need some preliminaries.
Definition 1 (see [10], Definition 4.1.1 and Proposition 4.1.1). Let D be a nonempty
subset of R n , xED. The set
We recall that if x E int(D), then TD(X) = R,n. Moreover, when D is convex, we have
the following nice characterization (see 10, Proposition 5.1.1).
In the sequel we shall consider the tangent cone TD(X) when D is a closed ball. In this
occurrence, the next proposition provides a useful characterization.
Proposition 3. Let r be a positive real number, D the closed ball of radius r centered
at the origin in Rn, x E R n with IIxll = r.
Then we have TD(X) = {v E R n : (v, x) ::; O}.
Proof. If r = I, our claim is proved at p. 221 of [10]. The general case follows by an
easy application of Proposition 2. 0
40
The following result is well-known. Its proof follows at once from Sections 10 and 11
of [11] and from Theorem lA at p. 164 of [12].
Theorem 5. Let A : [0, T] -+ RRXR, B : [0, T] -+ RRxm be two matrix functions whose
entries are Lebesgue summable in [0, T], and let n ~ R m be a nonempty compact set.
Then, for each Xo E RR, the set AT( xo) is nonempty, compact, convex, and
where H(t) is the principal matrix solution of the matrix differential equation
H' = A(t)H.
The main result of this section is the following.
for all (t, x) E [0, T] X D". Then, at least one of the following assertions holds:
Of course, A and B have measurable and bounded entries in [0, T] and also satisfy
assumption (4.4). Moreover, the problem
where
F(t, x) = U{A(t)x+B(t)u}
uEIl
for each (t, x) E [0, T] x Rn. We recall that a solution of problem (4.5) is an absolutely
continuous x(·) : [0, T]-+ R n such that
At this point, it is not difficult to realize that all the assumptions of Theorem 7.1 of
[13] are satisfied. In particular, the measurability of the multifunction F(·,x) follows
from Theorem 8.2.8 of [3]; moreover, since for each fixed t E [0, T] the multifunction
F(t,.) is Lipschitzean with constant IIA(t)11 (the norm of A(t) as a linear operator from
R n into itself), we have that
x
finally, if (t,x,u) E [O,T] R n xn,
K is chosen in such a way that n~ B(O,K) (of
course, the last inclusion is in the space Rm) and
we easily obtain
hence F(t,x) ~ (1 + IIxII)B(O, c) for all (t,x) E [O,T] X Rn. Thus, for each Xo E Dk
there exists a solution x(·) of problem (4.5) in [0, T] satisfying
Consequently, we have
By Theorem 5, each set AT(XO) is nonempty, compact, convex, and (4.3) holds. Thus,
by Proposition 1.4.14 of [3], the multifunction AT : R n -+ 2Rn is continuous. In
particular, since each AT( xo) is closed, the graph of AT is closed. Thus, all the
assumptions of Theorem 4 are satisfied with X = R n and f3(x) = AT(X). In particular,
the set {x E R n : x E AT(X)} is nonempty. Now, assume that assertion (C) above
42
does not hold. By (4.3), this implies that there exist two points z, v E AT(O) such that
z =1= v. If we pick w = z - v and choose any Xo E {x E R n : x E AT(X)}, since the
vectors
Zo = H(T)xo + z, Vo = H(T)xo + v
are in AT(XO), we have
For such system the assumption (4.4) of Theorem 6 cannot be satisfied. In fact, if we
°
take any t E [O,T], f" > 0, X = (Xl,X2) E R2 with Xl > 0, X2 > and IIxll = fk, by
Proposition 3 we have
The next example shows that assertion (A) can not be removed from the statement of
Theorem 6.
°
All the assumptions of Theorem 6 are satisfied. In particular, one can take fie = k.
In fact, given kEN, t E [O,T], x E R2 with IIxll = k, we have E C,,(t,x) (see
Proposition 3). We have
AT(X) = x + [-T,T]2.
43
Since x E int(AT(x)) for all x E R2, assertions (B) and (C) are contradicted, as
claimed.
The next example shows that assertion (B) can not be removed from the statement of
Theorem 6.
{ xi = -Xl + U
x~ = -X2 + u, lul:5 1.
Again, all the assumptions of Theorem 6 are satisfied, with n = [-1,1). Again, one
can take fk = k. In fact, if t E [0, T), x E R n , IIxll = k, by Proposition 3 we have that
A(t)x = -x E TD.(X), hence 0 E Ck(t,X). For such system, we have
We easily get
{x E R2 : x E AT(Xn ~ B(O, v'2),
hence assertion (A) of Theorem 6 does not hold. Moreover, since the attainable sets
are line segments never reducing to a single point, assertion (e) is also contradicted,
as desired.
Now, let us observe that assertion (C) of Theorem 6 is very restrictive, and it is
contradicted in many significant situations. However, it can not be dropped from the
statement, as the following very simple example shows.
Xi = -Xl +U
{
x~ = -x2 + u, U E {O}.
All the assumptions of Theorem 6 are satisfied, with n = {O}. Once more, one can
take fk = k. In fact, if t E [0, T), x E R2, IIxll = k, by Proposition 3 we have that
A(t)x = -x E TD.(X), hence 0 E Ck(t,X). In this case we have
AT(X) = x {exp(-T)}
for all x ERn. Hence, the only fixed point of AT is the origin, we have AT(O) = {O}
and assertions (A) and (B) of Theorem 6 are contradicted.
Remark 2. We point out that Theorem 6 is significant in the case IIGII ~ 1, where
G : R n -+ R n is the linear operator defined by G(x) = H(T)x. In fact, for
IIGII < 1, Theorem 6 automatically works without assumption (4.4), giving a more
precise information; that is, at least one of assertions (B) and (e) holds. To see this,
observe that, by (4.3), the fact IIGII < 1 implies that the multifunction AT is a mul-
tivalued contraction with nonempty compact values, hence, by eovitz and Nadler's
44
theorem (see [14]), it admits at least one fixed point. Now, assume that assertion
(C) is false. As we have already noticed in the proof of Theorem 6, this implies that
the nonempty compact convex set AT(O) contains at least two points. Hence, the set
Or(AT(O» is nonempty. Since by (4.3) we have
for all x ERn, the multifunction x -+ Or( AT( x» also is a multivalued contraction with
nonempty compact values. Hence, it admits at least one fixed point and (B) holds,
as desired. Moreover, we point out that when IIGII < 1, by Theorem 1 of [15], the
set {x E R n : x E AT(X)} is compact, hence bounded. We have already showed (see
Example 1) how in the case IIGII ;::: 1 assumption (4.4) is necessary to make Theorem
6 true. Also, Example 2 shows that if IIGII ;::: 1 then assertion (A) of Theorem 6 is not
necessarily contradicted, even if assumption (4.4) is satisfied.
REFERENCES
[1] D. Cha.n and J.S. Pang, "The generalized quasi-variational inequality problem" , Math. Oper.
Res. 7, pp. 211-222, 1982.
[2] P. Cubiotti, "An existence theorem for generalized quasi-variational inequalities", Set-Valued
Anal. 1, pp. 81-87, 1993.
[3] J.P. Aubin and H. F'rankowska, "Set-Valued Analysis", Birkhauser, Basel, 1990.
[4] E. Klein and A.C. Thompson, "Theory of Correspondences", John Wiley & Sons, New York,
1984.
[5] N.D. Yen, "On a class of discontinuous vector-valued functions and the associated quasi-
variational inequalities", Optimization, to appear.
[6] J.P. Aubin, "Mathematical Methods of Game and Economic Theory", North-Holland, Amster-
dam, 1979.
[7] K. Fan, "Fixed-point a.nd minimax theorems in locally convex topological linear spaces", Proc.
Nat. Acad. Sci. U.S.A. 38, pp. 121-126, 1952.
[8] O. Naselli Ricceri, "On the covering dimension of the fixed point set of certain multifunctions",
Comment. Math. Unit!. Carolin. 32, pp. 281-286, 1991.
[9] E. Michael, "Continuous selections I", Ann. of Math. 63, pp. 361-382, 1956.
[10] J.P. Aubin and A. Cellina, "Differential Inclusions", Springer-Verlag, Berlin, 1984.
[11] H. Hermes and J.P. Lasalle, "Functional Analysis and Time Optimal Control", Academic Press,
New York, 1969.
[12] E.B. Lee and L. Markus, "Foundations of Optimal Control Theory", John Wiley & Sons, New
York,1967.
[13] C. Castaing, Moussaoui S.B. and A. Syam, "Multivalued differential equations on closed convex
sets in Banach spaces", Set- Valued Anal. 1, pp. 329-353, 1994.
[14] H. Covitz and S.B. Nadler, "Multivalued contraction mappings in generalized metric spaces",
Israel J. Math. 8, pp. 5-11, 1970.
[15] J. Saint Raymond, "Multivalued contractions", Set- Valued Anal., to appear.
GENERALIZED QUASI-VARIATIONAL INEQUALITIES
AND TRAFFIC EQUILIBRIUM PROBLEM
De Luca Marino
Dipartimento di Matematica
Universita di Palermo,
Via Archirafi, 34
90123 Palermo, Italy
e-mail: [email protected]
Abstract The model that expresses the traffic equilibrium problem in terms of Quasi-
Variational Inequalities is improved taking into account that: i) the cost function may
be discontinuous; ii) the cost function may be considered as a multifunction. Existence
theorems in such directions are given with examples and considerations, based on a
direct computational method, that justify this approach.
1. INTRODUCTION
45
46
with
K:= {F E R m + : ~F = pl.
Here, m is the number of paths for the I % pairs, CO : R m + ---+ R m + is the
path cost function, p E RI+ is the fixed demand and ~ is a 1X m matrix whose elements
are:
"I.. _
'l'JT -
{I, E
if Rr 'Rj
0, if Rr f. 'Rj r = 1,2, ... ,m, j = 1,2, ... ,1, (1.2)
where Rr is the r-th path and 'Rj is the set of those paths which link the % pair j.
A solution H of the Variational Inequality (1.1) is an equilibrium pattern flow in
the sense of J.G.Wardrop (1952), that is:
In this case, we have equilibrium costs Cj for every % pair j, obtained con-
sidering those paths on which the equilibri.um flows are greater than zero. Hence, we
can require that the demand p depends on this equilibrium costs or, more generally,
directly on the equilibrium pattern flow H.
Thus, if we put, for each H E R m +
It is more convenient for our purpose to rewrite the Q.V.I. (1.4) in a different
form. To this end, let us observe that from (1.3) we can derive the values of 1 variables
Fi (which we may suppose, without loss of generality, be the first 1 variables), because
the matrix ~ in (1.2) is such that in each column there is an unique entry which is 1.
Thus, we have:
m
Fi = Pi(H)- L
r=1+l
~irFr, i = 1,2, ... ,1 (2.1)
and, in particular:
m
Hi = Pi(H)- L
r=l+l
~irHr, i = 1,2, ... ,1. (2.2)
47
such that
O"i(iI) ~ 0, viI = (H'+l, ... , Hm) E E
and (2.2) is fulfilled when Hi = O"i(iI).
We put:
v) k(H) := {F E R m - I : JF ~ p(H)}.
Then, if we replace H with H, k with K, and so on, and put r(H) for r(H, H),
the Q.V.1. (1.4) is transformed in the following one: to find
In this section we are interested in the numerical evaluation of the solution for
Q.V.1. (2.3). Since the problem cannot be reduced to a convex minimization problem,
without making further assumptions, we cannot use the well-known algorithms of the
convex programming theory.
Some algorithms for the construction of the solution have been established, using
various techniques. We present here a direct method proposed in [2) for Variational
Inequalities, also generalized for Q.V.I. (see for more details [3)).
Let us start observing that every solution of the system
{ H E K(H) (2.4)
r(H) = 0
is a solution of (2.3), whereas any other solution H must belong to the boundary
BK(H) of K(H): instead, if H were an interior point (we observe that the interior of
K(H) is not empty), we should have r(H) = O.
To obtain a face of (m -I)-dimensional polyhedron K(H) it needs to equate some
constraint in v). So, let us describe a face « m - 1) - (h + k ))-dimensional in the
following way: setting Jk = {it,h, ... ,jd the set of indices such that
As before, using the equality constraints in (2.5), we can express k components Ft, of
.
F,li E I, i = 1,2, ... ,k. Denoting by L = I\{1},12, ... ,ld and, for r E L,
r(h,k) = {rr - rl, if 3i : ¢>j,r = 1
r rr otherwhise
we obtain the Q.V.I., equivalent to (2.6):
(2.8)
Obviously, a solution of (2.8) solves Q.V.1. (2.7) and the following theorem give neces-
sary and sufficient conditions to this solution also solve the original Q.V.1. (2.3).
r(h,k)(H!h,k») = 0,
then H~h,k) is a solution of Q. V.l. (2.9) if and only if
r E Sh
(2.10)
i = 1,2, ... , k .
A similar theorem, with slight modifications, gives necessary and sufficient condi-
tions when (2.9) is not fulfilled. IT the above conditions are not verified for each solution
of the system (2.8), or when (2.8) does not admit solutions, Q.V.1. (2.7) cannot have
solutions in the interior of Km-I-(h+A;). Then, we iterate the procedure, passing to a
face of less dimension, until to reach eventually the vertexes of K(H).
Remarks.
- from a computational point of view, the efficience of this direct method is con-
nected with the value of m - 1;
- the direct method can be very useful to study sensitivity analysis and paradoxes,
specially when costs and/or demands depend on parameter, as we shall show in
the next example.
Example (see [4]). Let us consider a network with four nodes PI. P2, Pa, P4 and
five links (Pl,P2),(PI,Pa),(P2,Pa),(P2,P4),(Pa,P4). We have only a travel demand
related to the OlD pair (PI. P4) that is connected by the following paths:
Rl = (P1,Pa) U (Pa,P4)
R2 = (PI. P2) U (P2, P4)
Ra = (P1,P2) U (P2,Pa) U (Pa,P4).
Assuming that the path costs are, respectively:
with
K(H) = {F E RH : FI + F2 + Fa = p(H)}.
Following the above procedure, we get
Fa = p(H) - Fl - F2 • (2.12)
(2.13)
So, there exists H E E such that ECK( H), and this is remarkable for the next existence
results. The Q.V.1. (2.11) becomes:
L fi(H)(Fi -
3
H E K(H), Hi) ~ 0, TlF E K(H) (2.14)
i=1
does not admit solutions; then we consider the face of 8K(H) corresponding to
(2.15)
has solutions such that f 2 (H) < 0, observing that in this case Th.(2.1) is automatically
satisfied because Sh = 0.
The vector H = (f, f' 0) satisfies (2.15) and the above condition. Hence H is a
solution of the Q.V.I.'s (2.14) and (2.11).
Now, if we increase the cost C 3 , considering a positive parameter J.L* > Jl the solution
H does not change but it results that the total cost
also does not change. Thus, we have the following paradox: the C08t on a path increa8e8
but the total cost is constant.
51
r: E -+ R,m
K: E -+ 2E;
We have:
Theorem 3.1. If E is compact and convex, if both rand K are continuous and if,
for every H E E, the set K(H) is non empty, closed and convex, then Q. V.I. (r, K)
(3.1) has at least one solution.
A recent result [5] improves Th.(3.1) by replacing the continuity of r with a more
general condition:
Theorem 3.2. Let E be compact and convex, and assume that : i) for every F E
E - E, the set
GF = {H E E: r(H)F S; O} (3.2)
Example (see [6]). Let us show by means of an example the utility of the assumption
i): let E be the set
It results:
It is easily to prove that the set GF defined by (3.2) is closed for ever FEE - E.
So, Th.(3.2) ensures existence of solutions for the Q.V.1. (3.1), nevertheless r is
discontinuous on E.
52
Also in this case, a very recent result [7] ensures the existence of solutions for the
generalized Quasi-Variational Inequality which replaces the Q.V.I.(r,K) (3.1).
In traffic network model, also this generalization can be useful. It seems more
convenient to associate at each flow distribution F not a single-valued cost distribution
but a closed convex subset of Rm.
The results above are the starting point for our result, but we must overcome the
difficulty that, in our case, the assumption
can be not true. As alredy pointed out, K(H) is not a subset of E for each H E E.
Then, we have the following theorem where K is our multifunction, that is
Theorem 3.3.
Let us suppose that E is a non empty convex, compact subset of R m and that
i) Pi : E -+ R+, i = 1,2, ... , I, are positive continuous functions;
ii) for each H E E, it results
iii) for each H E E,r(H) is non empty, convex, compact, and i') holds; then there
exist
HE K(H) n E, Z E r(H) : Z(F - H) ~ 0, \IF E K(H) n E.
Moreover, if
ivy either H is an interior relative point of K(H) n E or Z= 0, there exist
3. The assumption iv) implies existence for the generalized Quasi-Variational Inequal-
ity.
can be true for only a finite number of terms in the sequence H n , because of the
continuity of p. Let us call No this finite subset of N. Then, we set
Fr, ifr E Br,
F.rn = { F.r - Pi(H)-Pi(Hn)
#(c!,>
'f en d
' 1 r E i an n E
N\ .!YO,
l\T
0, otherwhise.
Thus, we have a sequence Fn whose components Frn are such that lim Frn = Fr and
n .... oo
Step 2. Also for K : H -+ K(H) n E the condition b) follows directly by i). For a),
let us observe that, thanks to ii), we get
K(H) n E = cl(K(H) n ri(E», '<IH E E.
On the other hand, the multifunction H -+ K(H) n ri(E) is l.s.c. by virtue of ii) and
a') .
Then we may apply the result of [7] to obtain a solution of the G.Q.V.1. (f, K).
Step 3.
- If HE ri(K(H) n E) then Z = 0 (it is enough to take F = H ± f), and our claim
is showedj
- If 2 = 0 it results
Z(F - H) = (2 + Z.L)(F - H) = 0
where Z.L is the normal component of Z to aff(E), and then our claim is ensured
also in this case. o
54
Example (see [6]). Let us show an application of Th. (3.3) to a discontinuos Q.V.I.
Let it be:
E = {H E R3 : HI = PI, 0 ~ H2 ~ ~2, H3 = P3}
K(H)={FER3 :FI =PI, O~F2~P2-H2, F3 =P3}
where Pi are positive numbers such that
PI - P3 + 2P2 > O.
Then, let it be :
rI(H) = PI + P2 - 2H2
r 2(H)= :F(H2)
r3(H) = P3 + P2 - 2H2
where :F(H2) is the function
if Pf ~ H2 ~ I!f
if H2 = Pf
if 0 ~ H2 ~ Pf.
The conditions of Th.(3.3) are fulfilled. By the direct method, the vector
P2
H = (PI, 4,P3)
is the solution of the Q.V.1. (r, K). Since H belongs to ri(K(H) n E), this vector is
also the solution of Q.V.1. (r, K).
REFERENCES
[1) P.T. Harker and J.S. Pang, "Finite-Dimensional Variational Inequality and Nonlinear Comple-
mentarity Problems: A Survey of Theory, Algorithms and Applications", Mathematical Pro-
gramming 48, 161-270, 1990.
[2) A. Maugeri, "Convex programming, Variational Inequalities and applications to the traffic equi-
librium problem", Appl. Math. Optim. 16, pp. 169-185, 1987.
[3) M. De Luca and A. Maugeri, "Quasi-Variational Inequalities and applications to equilibrium
problems with elastic demand", in "Nonsmooth Optimization and Related Topics", F.M. Clarke,
V.F. Dem'yanov and F.Giannessi (eds.), Ettore Majorana, International Science Series, Plenum
Press, New York, pp. 61-67, 1989.
[4) M. De Luca and A. Maugeri, "Quasi-Variational Inequalities and applications to the traffic
equilibrium problem: discussion of a paradox", J. Compo Appl. Math., 28, 163-171, 1989.
[5) P. Cubiotti, "Finite-dimensional Quasi-Variational Inequalities associated with discontinuous
functions", J. Optim. Theory Appl., 72, 577-582, 1992.
[6) M. De Luca and A. Maugeri, "Discontinuous quasi-variational inequalities and applications to
equilibrium problems", in "Nonsmooth Optimization. Methods and Applications", F. Giannessi
(ed.) Gordon £3 Breach Sc. Publ., 70-74, 1992.
[7) P. Cubiotti, "An existence theorem for generalized Quasi-Variational Inequalities", Set- Valued
Analysis, VoU, No.1, 81-87, 1993.
VECTOR VARIATIONAL INEQUALITY AND
GEOMETRIC VECTOR OPTIMIZATION
Abstract: The paper presents a Vector Variational Inequality which is closely con-
nected with the geometric vector inequality, introduced in [9]. This Vector Variational
Inequality can be obtained as a specialized vector Variational Inequality treated in
[1]. Relationships between the vector Variational Inequality and the geometric vector
optimization ~e shown, concerning a necessary condition and a sufficient condition for
properly efficient solutions of such special structured vector optimization problems.
Key Words: Vector Variational Inequality, geometric vector inequality, vector opti-
mization.
1. INTRODUCTION
55
56
For proving assertions concerning relationships between the geometric vector in-
equality (briefly: GVI), introduced in [9], and the following vector Variational In-
equality (briefly:VVI), we use the same notations as in [7] and modify those given in
[1],[10],[20] for spaces, mappings, sets, and cones.
We denote by ~(R~) the set of nonnegative (positive) real numbers and by R n
the n-dimensional Euclidean space. Let be given the index sets
(2.1b)
where
mo := 1, ml := no + 1, ... , m.+p := n.+p-l + 1, n.+p := n.
Moreover, let be
Sk ~ RI[kJl, k E J2+ p , open convex sets,
'+p
S := II Sk ~ Rn, S =10, (2.2)
k=O
'+p
T := Sk II Tk ~ Rn, (2.3)
k=O
Z := _ 1 _[ :
ZO Zo
:
1E R(S+I)Xp, E R~ Vk E J~ (2.6)
\I Z 11 2 • .
Zk
Zs .. , Zs
A(y) := [Alo(Y) 0]
A 2 (y)
where (2.9a)
and
'xk : T -+ ~, k E J~+p, non-negative functions such that (2.9c)
(2.11a)
(2.12)
3D = diag(d[~](x)hEJo.+p
with 0;01: d[k](X) E Tk 'Ilk E J2+ p such that
xES -+ { Yzx = Az(y)G(x) - EzV(y)
Vy = (I'odfol x), ... , I'8+P1.+p] ( x»T E T, I'k E ~ 'Ilk E J2+,-
The matrix Yz has the representation
Yz = EzY W = rD = diag(f'k£.'[k](x)hEJO•+p .
Furthermore, we recall to the well-known cone-convexity.
For the vector function G : S -+ R 8 +l+p involved in the Geometric Vector Inequality
(2.15), the following properties are proved in [7].
Theorem 2.1. Let the geometric vector inequality (2.15) be given. Then, for each
xES, any 'Yk E ~,k E J2+ p , and
(2.16)
59
holds:
(i) y E Tj
where
Here +00 (-00) means the imaginary point whose every" coordinate" is +00(-00).
Especially for the elements of R~ we define
Definition 2.4. Let C ~ Rs+I be a convex cone. The vector function G : S -+ R.~~+p
is said to be C-convex, if
(2.18)
Now we will introduce the Fenchel vector conjugate in the space R~~ .
(2.20)
where each G k, k E J2+ p , depends only on the k-th row vector W(k), k E J2+ p , of the
matrix W, W according to (2.10).
The following lemma contains an arithmetic rule for Fenchel vector conjugates.
Lemma 2.1. Let G : S -+ R"+1+p be R++I-convex and let EzG* be the Fenchel
vector conjugate of G. Then, for the function
where
Definition 2.9. Let the vector function F : S -+ R~~+P be R++1-convex and let
:F: S -+ R+~, where :F(x) = EzF(z). Then the set
Lemma 2.2. Let:F: S -+ R~~ with :F(x) := EzF(x) be given, where the vector
function F: S -+ R~~+P is properly R++1-convex . Then
Theorem 3.1. Let GVI be given according to (2.15). IT the VVI (see (2.14»
Setting in (3.2)
-A(y)G*(A -l(y)Y) =: V(y) (3.3)
it follows:
Yzx = Az(y)G(x) - EzV(y), xES, yET. o
Theorem 3.2. Let VVI and GVI be given according to (2.14) and (2.15), respectively.
If in GVI the matrix Yz is chosen according to
Proof. Because of (3.4) we deduce the equality in (2.15). Then, regarding (3.3), it
follows
From (3.6) we imply immediately the existence of a vector xES and a matrix Yz such
that
(-Yz, r - x) ~ :F(x) - :F(r)Vr E S.
( 4.1)
and assuming
because of
V(y) := Ez V(y), ( 4.4)
G(r) := (Go(r), ... G.(r)f,
._ (G(r»)
G(r).- G(r) , (4.5a)
G(r) := (G.+I(r), ... , G.+p(r))T
and
(4.5b)
where
Inequality (4.3), derived from (2.2), gives now motivation for establishing dual
geometric vector optimization problems [8]:
VO(Y)]
(P*) : V(y) := [ : + II zZll2 L Vk(y) -t v - max (4.8)
Vo(y) kEJO
For the vector optimization problems we define properly efficient solutions accord-
ing to [18).
Definition 4.1. A point xO E B (yO E B*) is called a properly efficient solu.tion of (P)
((P*», if there is azo E KO such that
Theorem 4.1. Let xO E RR be a properly efficient solution of (P) . Then the point
xO is a solution of the following VVI:
Then, there exists an element ZO E KO, such that each coordinate function
(4.10)
of
Recall that in (4.11) each function Fk, k E J2 , depends only on certain coordinates of
x E B , since
Fk(X) = >'k(yo)Gk(X),Gk(X) according to (2.4).
Thus, it follows from (4.11) for all k E J2 :
8:F.(x)
~
Vi E [kJ,
{
aFk(X):= 0
Vi f/. [kJ, (4.12)
ax;
-!L 8:F,(x) Vi E [IJ, I E J O•
zUII.oi: 8Xi
where
Since
Azo(yO)G at xO E B :
( 4.15)
Y~o E 8A z o(yO)G(xO).
To prove ( - Y~o, X - xO) If.J we use the relation
Azo(yO)G(x) - Azo(yO)G(xo)
be solvable, where B and T are given according to (4.7) and (2.3), respectively,
(4.17)
._ (G(x»)
G(x).- G(x) according to (2.4) . (4.18)
(4.19)
(4.20)
REFERENCES
[1] G.-Y. Chen, "Vector Variational Inequality and its application for multiobjective optimization".
Chinese Science Bulletin, 34, 17, 969-972, 1989.
[2] G.-Y. Chen, "Existence of solutions for a Vector Variational Inequality". An extension of
Hartmann-Stampacchia theorem". lou. Optim. Th. Appl., V. 74, No.3, pp. 445-456, 1992.
[3] G.-Y. Chen and G.M. Cheng, "Vector Variational Inequality and vector optimization". In
Sawaragi, Y., Inoue, K. and Nakayama, H. (eds.), Toward Interactive and Intelligent Deci-
sion Support Systems, Vol. 1 , Proceedings 1986. Lecture Notes in Econ.and Math. Syst., 285,
Springer-Verlag, Berlin·Heidelberg·New York, 1987.
[4] G.-Y. Chen, and X.-Q. Yang, "The vector complementarity problem and its equivalence with
the weak minimal element in ordered spaces". l.Math.Anal.Appl. 153, 136-158, 1990.
67
[5] J. Crank, "Free and Moving Boundary Problems". Clarendon Press, Oxford, 1984.
[6] S. Dafermos, "Exchange price equilibrium and Variational Inequalities". Math. Programming
46, 391-402, 1990.
[7] K.-H. Elster and R. Elster, "The geometric vector inequality and its properties". Report
3.200(735} Univ. of Pisa, Dept.of Mathematics (Optimization and OR). Forthcoming in Op-
timization, 1993.
[8] K.-H. Eelster, R. Elster and A. Gopfert, "On approaches to duality theory in geometric vector
optimi- zation", In "Methods of Operations Research", Anton Hain,Meisenheim/ Frankfurt
a.M., 60, 23-37, 1990.
[9] R. Elster, "Geometrische Vektoroptimierung". Dissertation B, Technische H och.chule M er.e-
burg, 1990.
[10] F. Giannessi, "Theorems of alternative, quadratic programs and complementarity problems". In
Cottle R.W., Giannessi F. and Lions J.-L.(eds.), "Variational Inequalities and Complementarity
Problems. Theory and Applications". J .Wiley & Sons, New York, pp.151-186, 1980.
[11] R. Glowinski, J .-1. Lions and R. Tremolieres, "Numerical Analysis of Variational Inequalities",
North-Holland, Amsterdam, 198!.
(12) P.T. Harker and J.-S. Pang, "Finite-dimensional Variational Inequality and nonlinear comple-
mentarity problems: A survey of theory, algorithms and applications" Mathem.Programming
48, 161-220, 1990.
(13) A. Kaplan and R. Tichatschke, "Stable Methods for Ill-Posed Variational Problems" . Akademie
Verlag, Berlin, 1990.
[14] O.G. Mancino and G. Stampacchia, "Convex programming and Variational Inequalities".
J.Optim. Theory and Appl. 9,1,3-23,1972.
[15] A. Maugeri, "Convex programming, Variational Inequalities,and applications to the traffic equi-
libriumproblem". Appl.Math.Optim., 16,169-185,1987.
[16) M.Aslam Noor, "Variational Inequalities in physical oceanography". In Rahmen, M. (ed.),
"Ocean Waves Engineering". Comput. Mechanics Publ.,U.K., pp.201-226, 1994.
[17] Y. Sawaragi, H. Nakayama and T. Tanino, "Theory of Multiobjective Optimization". Academic
Press Inc.,Orlando, 1985.
[18] K.P. Schonfeld, "Some duality theorems for the nonlinear vector maximum problem". Un-
ternehmen.forschung 14, 1, 51-63, 1970.
[19] C. Tammer, "A generalization of Ekeland's variational principle". Optimization 25,2-3, 129-141,
1992.
[20) YANG; X.-Q., "Vector Variational Inequality and its duality. Nonlinear Analysis". Theory,
Methods and Applications 21, 11, 869-877, 1993.
[21) P.-L. Yu, "Multiple-Criteria Decision Making". Plenum Press, New York and London, 1985.
TESTING A NEW CLASS OF ALGORITHMS
FOR NONLINEAR COMPLEMENTARITY PROBLEMS
2 Departamento de Matematica
Universidade de Coimbra
3000 Coimbra, Portugal
e-mail: [email protected]
Abstract: We investigate the numerical behavior of a new, simple algorithm for the
solution of nonlinear complementarity problems. The algorithm is based on a recently
proposed merit function which possesses some interesting theoretical properties. One of
the aims of the paper is to show that algorithms based on this merit function can be
viable also from the numerical point of view.
Key Words: Nonlinear complementarity problem, merit function, SCI function, global
convergence, superlinear convergence.
69
70
minimization of a new merit function, and to shortly review some recent results on
merit functions.
In the next section we give some basic definitions which are needed in the sequel. In
Sect. 3 we review the properties of some existing merit functions and we present the
new merit function, which is shown to possess some definite advantages over previous
merit functions. In Sect. 4 we describe a new algorithm for the solution of Problem
(NC) which is based on the new merit function. Finally, in Sect. 5 numerical results are
presented.
A few words about the notation. If M is an n x n matrix with elements M i ;, i =
1, ... n, and I and J are index sets such that I, J ~ {i = 1, ... n}, we denote by MIJ the
III x IJI submatrix of M consisting of elements Mij, i E I, j E J. If v is an n vector,
we denote by VI the subvector with components Vi, i E I. A superscript k indicates the
iteration number; finally, II . II indicates the euclidean norm.
2. BASIC DEFINITIONS
- A function F : lRn -+ lRn is a Po-function iff, for every x and y in lRn with x :F y,
there is an index i such that
- A function F : IRn -+ IRn is a P-function iff, for every x and y in IRn with x i: y,
there is an index i such that
- A function F : IRn -+ IRn is a uniform P-function iff there exists a positive constant
p such that, for every x and Y in IRn, there is an index i such that
- A function F : IRn -+ IRn is strictly monotone iff, for every x and y in IRn with xi: y,
(x - ynF(x) - F(y)] > O.
- A function F : IRn -+ IRn is strongly monotone iff there is a positive constant p such
that, for every x and y in IRn,
exists for any v E IRn, where 8F(y) is the generalized Jacobian of Clarke at y.
Semismooth functions lie between Lipschitz functions and CI functions. Note that
this class is strictly contained in the class of B-differentiable functions.
It is known that [20,28]:
(a) Continuously differentiable functions and convex functions are semismoothj also the
compositions of semismooth functions are semismooth.
(b) If a function F is semismooth at x, then F is directionally differentiable at x, and
the directional derivative F'(xj d) is equal to the limit (2.1).
We can now give the definition of SCI function.
Semi smooth functions can be viewed as functions which lie between CI and C2 func-
tions. Semismooth systems of equations form an important class, since they often occur
in practice and many of the classical methods for their solution (e.g. Newton's method)
can be extended to solve such problems [23,28,26]. Analogously, many classical results
concerning the minimization of C 2 functions can be extended to the minimization of
SCI functions (see e.g. [24,2] and references therein), which, in turn, play an important
role in many optimization problems. Under very mild differentiability assumptions on
F, the new merit function we will introduce in the next section is an SCI function.
Let C ~ IRn be given. A merit function for Problem (NC) is a nonnegative function
M : IRn -+ IR such that x is a solution of Problem (NC) iff x E C and M(x) = 0, i.e. iff
the global solutions of the problem
coincide with the solutions of Problems (NC). Usually either C = IRi. or C = IRn.
It is not difficult to find a merit function for Problem (NC), the challenging task
is to find a merit function which enjoys some properties which are useful from the
computational point of view. For example one could consider the merit function M(x) =
F(xyx whose global minimizers on the set C:= {xix ~ O,F(x) ~ O} are the solutions
of the complementarity problem (NC). But seeking these global minimizers is not easy
because, even in very simple cases, the structure of C may be very complicated and the
minimization problem can have stationary points which are not global solutions.
Properties which seem particularly attractive, and that we would like the merit func-
tion to enjoy, are the following:
• M is "smooth";
• every stationary point of Problem (PM) is a global solution of Problem (PM);
Mfa is a merit function with C = IRi., so that solving (NC) is equivalent to finding
the global solutions of the simply constrained minimization problem {min Mfa(x) : x E
IRi.}. Furthermore the merit function Mfa enjoys the following properties:
- Mfa is continuously differentiable;
73
- if the Jacobian of the map F is a positive definite matrix for every x, then every
stationary point of Problem (PM) is a global minimum point of Problem (PM) [8];
- if the map F is strongly monotone then the sets L(o:) are bounded [8].
Mm. is a merit function with C = IRn, so that solving (NC) is equivalent to finding the
unconstrained global solutions of the problem {min M m .( x)}. Furthermore the merit
function Mm. enjoys the following properties:
- if the map F is strongly monotone and globally Lipschitzian, then the sets L( 0:) are
bounded [35].
We note that on one hand the implicit Lagrangian merit function is simpler than
the regularized gap function, since it only requires an unconstrained minimization, but,
on the other hand, the condition to have bounded level sets is much stronger for the
implicit Lagrangian than for the regularized gap function.
We point out that most of the existing global algorithms for the solution of nonlinear
complementarity problems are based, in a more or less explicit way, on a merit function.
Unfortunately often the analyses of these algorithms are based on conditions on the limit
points of the sequence generated by the algorithm itself; this is most unfortunate because
it makes the properties of the algorithm depend on the behavior of the algorithm itself.
Furthermore this way of analyzing algorithms makes it extremely difficult to identify in
a clear way classes of problems (e.g. monotone, strictly monotone, etc) for which the
algorithms work. Thus in this short review we have only considered those merit functions
which are analyzed in the framework illustrated at the beginning of the section. The
interested reader is referred to [17,4,14,25] for further examples of merit functions.
¢>(a, b) := Va 2 + b2 - (a + b).
¢>(a, b) = 0 a 2: 0, b 2: 0, ab = 0; (3.1)
note also that ¢> is continuously differentiable everywhere but in the origin. The function
¢> has been introduced by Fischer in [5,6], since then it has attracted the attention of
74
many researchers and has shown to enjoy many favorable properties; for example it has
been used in [7,14,15,16,9,33,27).
In view of (3.1) it is obvious that the function
n
'l1(x):= LcP(x;,F;(x)?
;=1
is a merit function with C = m.n , and that solving (NC) is equivalent to finding the
unconstrained global solutions of the problem {min 'l1( x)}. The merit function 'l1 enjoys
the following properties [3]:
We remark that the theoretical properties of 'l1 are superior to those of the regularized
gap function and of the implicit Lagrangian. In fact, on one hand 'l1 allows to solve the
nonlinear complementarity problem by an unconstrained minimization, on the other
hand the conditions under which we can be sure that every stationary point of the merit
function is a global minimizer and that the level sets are bounded are substantially
weaker. Furthermore also the differentiability properties of 'l1 seem more interesting,
and actually the SCI property is very important from an algorithmic point of view.
The merit function 'l1 has also been independently introduced by Geiger and Kanzow
[9). Their results are however weaker than those reported above or simply of a different
nature. In particular they showed that every stationary point of the merit function is a
global minimum point if F is monotone, while the level sets of 'l1 are bounded if F is
strictly monotone. The analysis of the second order properties of 'l1 is cruder than ours
and, to define superlinear convergent algorithms for the solution of the complementarity
problem, they require the solutions to be nondegenerate, which is not the case for the
algorithm described in the next section. On the other hand Geiger and Kanzow describe
an interesting algorithm for the solution of strictly monotone complementarity problems
which does not require the evaluation of the Jacobian of F.
The merit function 'l1 can be used in several ways to define globally convergent
algorithms for the solution of nonlinear complementarity problems: for example one
could simply use an off-the-shelf unconstrained algorithm to minimize 'l1. In this section
we use the merit function in a different, but classical way. We first define a fast, local
algorithm for the solution of Problem (NC). Then we globalize this local algorithm by
performing an Armijo-type linesearch using the "local" direction, but reverting to the
antigradient when the "local" direction is not a good descent direction for the merit
function. Note that this scheme follows exactly the same lines used in the classical
stabilization scheme for Newton's method for the unconstrained minimization of a twice
continuously differentiable function. The crucial point will be to show that eventually
the gradient direction is never used and the stepsize of one is accepted, so that locally the
global algorithm coincides with the local one thus ensuring a fast asymptotic convergence
75
rate. To this end the SCI smoothness turns out to be a crucial property. We remark
that this is neither the only way to exploit the function IV, nor, possibly, the best one.
However, we note that the local algorithm enjoys several interesting properties and that
the overall global algorithm, in spite of its simplicity, performs surprisingly well.
xk+ I = xk + dk.
To motivate the local algorithm we first consider a simplified situation. Suppose that
x is a solution of Problem (NC), that x is nondegenerate and that we know the sets A
and N of variables which are 0 or positive at x
Then, to determine XN we would only need to solve the system of equations F;(XN' OA) =
0, i E N. Provided that V FNN(XN, OA) is nonsingular we could also apply Newton's
method to this system, thus setting x';tt
= x~ + d~, where d~ is the solution of the
following linear system
(4.1)
Obviously, in general we do not know the sets A and N, and furthermore we would like
to avoid the nondegeneracy assumption, which is often not met in practice. We then
define dk in two steps. At each iteration we first estimate the sets A and N, thus fixing
some of the components of dk , then we calculate the remaining part of dk by solving a
reduced linear system. We approximate the sets A and N by the sets Ak and Nk defined
by
where e is a fixed positive constant. It is very easy to check that the following result
holds.
Proposition 4.1 Suppose that x is a solution of Problem (NC). Then, for every fixed
e there exists a neighborhood n of x such that, for every xk belonging to n
l' ~ Ak ~ l' U ,8
a ~ N k ~ aU,8.
(4.2)
Step 0: Set k =0
Step 1: (stopping criterion) If the stopping criterion is satisfied stop.
Step 2: Calculate the "local direction" dk according to (4.2)-(4.3).
If system (4.3) is not solvable set dk = -V1lo(x k ).
Step 3: If
(4.5)
77
(4.6)
set d" = - Vw( x"). Find the smallest i = 0,1,2, ... such that
(4.7)
A few comments are in order. At Step 1 any reasonable stopping criterion can be
used. Note that in our case we can use classical measures of optimality, like the norm of
the vector of residuals, but also measures connected to the merit function as, for example,
the norm of the gradient of w. At Step 2 we try to calculate the "local" search direction
defined by (4.2)-(4.3). If this direction is not well defined we switch to the antigradient
of the merit function. Then we exploit the fact that, if the nonlinear complementarity
problem is solvable, the optimal value of W is o. So, if, for some constant u E (0,1),
test (4.5) is satisfied, we accept the stepsize of one. If this test is passed an infinite
number of times this will obviously lead to the function value tending to zero as desired.
Should test (4.5) not be satisfied, we perform in Step 4 a classicallinesearch procedure
to determine the step size. In this latter case we possibly switch to the antigradient, see
test (4.6), in order to ensure that the search direction is "sufficiently" downhill.
The aim of the acceptability test of Step 3 is twofold. On one hand it gives us one
more chance to accept the stepsize of one, on the other hand it makes it easier to prove
the superlinear converge rate of the algorithm (see [3]). A test close to (4.5) has been
proposed, with similar purposes, in [26].
The following result can be proved [3].
a. each accumulation point of the sequence {x"} generated by the algorithm is a station-
ary point of w;
h. if one of the limit points of the sequence {x"} is a b-regular solution of Problem (NC),
then {x"} -+ X;
c. if {x"} -+ if and if is an R-regular solution of Problem (NC) and each component F;
of F is S(Ji in a neighborhood of x, then:
1. eventually d" is always the "local" direction defined in the previous subsection
(i. e. the antigradient is never used eventually);
~. eventually the stepsize of one is always accepted so that x"+1 = x" + d" ;
9. the convergence rate is quadratic.
We note that in general what we can guarantee is that every limit point x (if any)
is a stationary point of w. If W(if) = 0 then x is also a solution of the nonlinear
complementarity problem. According to what seen in Sect. 3, we can be sure that every
limit point of the sequence generated by the algorithm is a solution of Problem (NC) is
F is a Po-function. If F is a uniform P-function, we can also guarantee the existence
78
of a limit point. Actually in this latter case, it is elementary to show that the whole
sequence converges to the unique solution of the complementarity problem.
In this section we have considered a basic algorithm; some simple modifications
which preserve all the properties described here, but which can be useful from the
computational point of view, will be illustrated in the next section.
5. NUMERICAL RESULTS
In this section we report on some numerical results obtained with the basic algo-
rithm described in the previous sections and with some variants of it. To motivate these
variants we first recall that it has been often observed in the field of nonlinear com-
plementarity algorithms that the linesearch test used to enforce global convergence can
lead to very small stepsizes; in turn this can bring to very slow convergence and even to
a numerical failure of the algorithm. To circumvent this problem many heuristics have
been used (see, e.g., [11,22,14]). Here we propose to substitute the linesearch test (4.7)
by the following non monotone linesearch:
(5.1)
where the Mk are nonnegative integers bounded above for any k. This kind of linesearch
has been first proposed in [10] and since then it has proved very useful in the uncon-
strained minimization of smooth functions. Adopting the same proof techniques used in
[10] it is easy to see that all the results described in the previous section still hold if we
substitute the linesearch (5.1) to (4.7) in the Global Algorithm.
A variant of test (4.5) in line with the nonmonotone approach and which still pre-
serves all the properties of the Global Algorithm can also be envisaged.
llt(x k + dk ) ~ u. max k llt(xk-i), (5.2)
J=O,l, ... ,M
where, again, Mk are nonnegative integers bounded above for any k. Exploiting this
ideas we have then tried three algorithms.
Algorithm A: A pure monotone algorithm. This is the basic Global Algorithm de-
scribed in the previous section.
We tried the algorithm on several test problems taken from the literature, we con-
sidered nonlinear (Table 1) but also linear problems (Table 2); the test set includes
problems that are not Po and problems that are not R-regular or b-regular at the solu-
tion. Some details on these problems, the starting points, and adequate references are
reported in the appendix. For each algorithm we report the number of iterations and
the number of function evaluations needed to reach a point where the euclidean norm of
the residual, min( xl:, F( xl:», is less than 10- 6 • The asterisk indicates that convergence
occurred to a different solution (see Appendix). By these results it is apparent that
the three algorithms seem to be robust. Algorithm A, as expected, requires in some
cases many functions evaluations because in some iterations many halvings are needed
to satisfy the acceptability test. Algorithm B behaves better in this respect, however
the nonmonotonicity is not always beneficial. On the whole the hybrid Algorithm C
seems the best one, since it combines the good features of the other two algorithms.
Algorithm C compares well also with the results reported in the literature (see, e.g.,
[4,9,11,14,15,22,31]) especially if one considers that the computational cost per iteration
is low.
APPENDIX
In this appendix we report some relevant data, the source and the starting points for
81
the test problems used in the previous section. The source reported is not necessarily
the original one.
Kojima-Shindo problem: See [22]. F(x) is not a Po-function. This problem has
two solutions: xl = (0.5V6, 0, 0, 0.5) and x 2 = (1,0,3,0); xl fails to be R-regular and
is degenerate. In the results of Table 1 convergence always occurred to Xl except in
the case indicated by an asterisk. The linearized complementarity problem at 0 has no
solution.
Starting points: (a) (0,0,0,0), (b) (I, I, I, 1).
Spatial price equilibrium problem: See [22]. F is a P-function and the unique
solution is therefore R-regular.
Starting points: (a) (0,0, ... ,0), (b) (I, 1, ... ,1).
Traffic equilibrium problem: See [22].
Starting points: (a) All the components are 0 except Xl, X2, X3, XlO, Xu, X20, X21l X22, X29, X30,
x40, X4S which are 1, X39,X42,X43,X46 which are 7, X4t,X47,X48,XSO which are 6, and X44
HS35 problem: This problem represents the KKT conditions for the 35th problem
in [12). The resulting F is monotone but not strictly monotone.
Starting points: (a) (0.5,0.5,0.5,0).
HS76 problem: This problem represents the KKT conditions for the 76th problem
in [12). The resulting F is monotone but not strictly monotone.
Starting points: (a) (0.5,0.5,0.5,0.5,0,0,0).
REFERENCES
[1] G. Auchmuty. "Variational principles for variational inequalities". Numerical Functional Analysis
and Optimization, 10, pp. 863-874, 1989.
[2] F. Facchinei. "Minimization of SCI functions and the Maratos effect". DIS Technical Report 10.94.
To appear in Operations Research Letters.
[3] F. Facchinei and J. Soares. "A new merit function for nonlinear complementarity problems and a
related algorithm". DIS Technical Report 15.94, 1994.
[4] M. Ferris and S. Lucidi. "Globally convergent methods for nonlinear equations". Journal of Opti-
mization Theory and Applications, 81(1), 1994.
[5] A. Fischer. "A special Newton-type optimization method". Optimization, 24, pp. 269-284, 1992.
[6] A. Fischer. "A special Newton-type method for positive semidefinite linear complementarity prob-
lems". Technical Report, Institute of Numerical Mathematics, Dresden University of Technology,
Dresden, Germany, 1992. To appear in Journal of Optimization Theory and Applications.
[7] A. Fischer. "On the local superlinear convergence of a Newton-type method for LCP under weak
conditions". Technical Report MATH-NM-07-1993, Institute of Numerical Mathematics, Dresden
University of Technology, Dresden, Germany, 1993.
[8] M. Fukushima. "Equivalent differentiable optimization problems and descent methods for asymmet-
ric variational inequality problems". Mathematical Programming, Series A, 53, pp. 99-110, 1992.
[9] C. Geiger and C. Kanzow. "On the resolution of monotone complementarity problems". Preprint
82, Institute of Applied Mathematics, University of Hamburg, Hamburg, Germany, 1994.
[10] L. Grippo, F. Lampariello and S. Lucidi. "A nonmonotone line-search technique for Newton's
method". SIAM Journal on Numerical Analysis, 23, pp. 707-716, 1986.
[11] P.T. Harker and B. Xiao. "Newton's method for the nonlinear complementarity problem: A B-
differentiable equation approach". Mathematical Programming, Series A, 48, pp. 339-357, 1990.
[12] W. Hock and K. Schittkowski. "Test examples for nonlinear programming codes". Lectures Notes
in Economics and Mathematical Systems 187, Springer -Verlag, Berlin, 1981.
[13] N.H. Josephy. "Newton's methods for generalized equations". MRC Technical Summary Report
1965, Mathematics Research Center, University of Wisconsin-Madison, Wisconsin, 1979.
[l4] C. Kanzow. "Some equation-based methods for the nonlinear complementarity problem". Preprint
63, Institute of Applied Mathematics, University of Hamburg, Hamburg, Germany, 1993.
83
Ferrari Paolo
Abstract. This paper considers the problem of multimodal transport networks with
elastic demand and capacity constraints. It is shown that, when capacity constraints
are active, there is not equilibrium in these networks. However an equilibrium solution
can be assured in any case if additional costs are imposed on network links. The paper
presents a model of multimodal networks where capacity constraints, and both demand
and cost functions, are symmetric: in this case the additional costs are the Lagrange
multipliers of capacity constraints. A method to calculate these additional costs is
proposed and is used in a simple numerical example.
1. INTRODUCTION
The hypothesis that capacity constraints are not binding is very often used in the
study of transport networks. In this hypothesis an equilibrium solution always exists
if demand is fixed, whereas some restrictions have to be imposed on both demand and
cost functions in order to assure the existence of equilibrium when demand is elastic.
But road capacity constraints are often active in real life: when this happens
there is not equilibrium in transport networks. However the existence of a network
equilibrium can be assured also when capacity constraints are active, if additional
costs are imposed on network links.
This paper, after a discussion about the problem of equilibrium in transport net-
works with capacity constraints, considers a model of multimodal transport network
where capacity constraints, and both demand and cost functions, are symmetric. It is
shown that in this case the additional costs that assure the existence of equilibrium
85
86
are the Lagrange multipliers of capacity constraints. A method to calculate this "road
pricing" is proposed and is used in a simple numerical example.
(2.1)
those of the sequence Tl, T2 ... T n ... : e.g. the probability that an individual receives a
given utility from making or not a trip in a given day period depends on the costs of
this trip in other day periods.
An individual characterized by the vector Sk associates to the vector x E Bk in
any epoch of the sequence Tl . .. Tn an utility given by:
(2.2)
where Vex, Sk) is the regression of U on x and Sk, and p is a random variable with zero
mean. If p is a Wei bull random variable and if the utilities associated to the various
alternatives are independent, the probability Pkj that the individual chooses the vector
Xj is given by the logit model (McFadden, 1975):
(2.3)
Let y, v, w be the vectors of attributes of the destination, of the mode and of the
path into which the vector x can be partitioned:
(2.4)
We assume that the utility regression V has the following additively separable
form (Domencich and McFadden, 1975):
(2.5)
It follows from Eq. (2.5) that an individual belonging to N k , who has chosen to go
to the destination j using the t mode, travels along the path from which he expects the
maximum utility, i.e. the minimum cost: thus this minimum cost becomes an attribute
of the transport mode chosen to go from k to j.
Assume that all individuals belonging to Nk have the same socioeconomic char-
acteristics. In this case we have:
(2.6)
If the choices of the individuals belonging to NI< are independent, the trips made
in the epoch under examination from k to the other centroids are the components of
a multinomial random vector. Thus the average Dt j of demand from k to j on mode
t in this epoch is given by:
(2.8)
D -N -
I< - I<
E E Dt
/ .... 1< tET
-N
1</ - I< exp[al<l +E E
exp[al<l
exp[a/ - .B.\~,I·
(2.9)
... kl 1#1< tETkl
If aj and .\~j Vj, t are constant during the sequence Tl •• •T n ••• , the same happens
for the probability given by Eq. (2.7); in this case the succession of the demand values
during the sequence is the realization of a stationary process whose average is given by
Eqs. (2.8) and (2.9). Moreover a multinomial random variable converges in probability
to its average; this means that if N" is large, the probability of a considerable shift
of demand from its average is negligible. For this reason we can consider the demand
value constant during all the sequence: Eqs. (2.8) and (2.9) give the demand function,
and its inverse, i.e. the vector of the averages .\~j Vk,j, t of the expected costs that
yield the demand, is obtained from Eqs. (2.8) and (2.9):
epoch such a situation is reached that the trip costs the individuals bear coincide or
are very close to those they expected: in this case the individuals will repeat their
choices in the successive epochs, giving rise to constant values of >"L
and of demand.
However this happens only if an equilibrium solution for the transport network exists.
d!v = the number of individuals who travel between i,j = w on mode t during an
unit time period within the epoch T under examination: if we assume that the
demand D!., given by Eq. (2.8) is uniformly distributed during T, d!v is obtained
dividing D!v by the duration of T.
t E T between all w E W. Hence forward we will omit, for the sake of brevity, the
words "during the unit of time" when we will refer to vector d and its components.
h = ( ... h t' ... )' E R~ = the vector of flows on all paths pEP
f = ( .. .p' ... )' = the vector of link flows generated by all modes.
Given:
1
we have f = Ah.
Let Bt be the incidence matrix between pairs wand paths travelled on mode t,
and
1
we have:
Bh - d = O. (3.1)
Let:
Ci(f) = the average of trip costs perceived by users who travel on link i E L; Ci(f) is
continuous function of f in 8.
C p( h) = L Ci(f)l!ip = the average of trip costs perceived by users who travel on path
iEL
pEP, where l!ip = 1 if p traverses i, 0 otherwise; Cp( h) is continuous function
of h in f!.
Alj = the inverse of the demand function, obtained by substituting in Eq. (2.10)
D!j = dljT
At = ( .. . A:" ... ),
A = ( .. . At ' .. . )'.
A vector [~] E f! is an equilibrium solution if and only if, for every pair w, every
(3.3)
where hp and d:" are the components of any vector [~] E f!.
The following demonstration of equivalence is a modification of that reported in
(Dafermos, 1982) in order to take into account the capacity constraints.
We note that (3.2) =} (3.3). Indeed, given a vector [~] E f!, (3.2) implies:
(3.4)
for every pair w, every mode t E Tw and every path pEP!. Inequality (3.4) follows
from the fact that, when hp > 0 for pEP!, we have Cp(h) - A:"(£1) = 0, so that (3.4)
is satisfied. When hp = 0 it follows that Cp(h) - A:" (d) 2: 0, and (3.4) is satisfied again
because hp 2: O. By summing (3.4) over all pEP!, the following relation holds:
f!. We show that [~] satisfies the equilibrium conditions (3.2). Indeed, for a fixed
92
pair Wn E Wand a mode tk E Twn , and given a path Pk E P!':., we consider a vector
[J] E 11 such that hp = hp Vp ::f. Pk, but hpk ::f. hpk ' As B h - d = Bh - d = 0, it
follows that d~ = d~ if (t, w) ::f. (tk' w n) and d!:;n = d!:;n - hpk + hpk ' In this case (3.3)
reduces to:
CpkCh). (h pk - hpk ) - A!:;Jd). (hpk - hpk ) ~ o. (3.6)
If hpk = 0, as hpk ~ 0 (3.6) implies:
As (3.7) and (3.8) hold for every w, t and p, we obtain (3.2). It is worth noting that
the condition that capacity constraints are strict inequalities, so that we can always
choose hpk > hpk , is essential in order to demonstrate the equivalence between (3.2)
and (3.3). On the other hand it is easy to verify that if capacity constraints are written
as :::;, the user equilibrium problem (3.2) and the Variational Inequality (3.3) are not
equivalent, because the existence of a solution of (3.3) would not imply the same for the
problem (3.2). Consider for example the simple netwotk represented in fig. 1, where a
pair of nodes (WI, W2) is connected by a link characterized by the cost function c(f).
The link flow is equal to the demand between the two nodes; let A( d) be the inverse
of demand function, and let H be the link capacity. No equilibrium solution exists,
because for any demand value in the closed interval [0, HI the link cost is less than A( d),
so that the equilibrium conditions (3.2) are not verified. If the capacity constraint
c, }..
W1 W2
0>--------..-<0
o f~ f H f,d
were written as f :::; H, the interval of feasible solution [0, H] would be closed, and the
point f =H would be a solution of the Variational Inequality (3.3); in fact we would
93
whereas H is not an equilibrium solution. On the other hand, if the capacity constraints
is I < H, the point I = H is not a feasible solution, and no point r E [0, H) is a
r
solution of the Variational Inequality: in fact a point I > II E [0, H) always exists,
so that:
[e(r) - AU·)]' U - r) < ° (3.10)
IT capacity constraints are strict inequalities, 0 is not closed, so that the Varia-
tional Inequality (3.3) does not always admit a solution (Kinderleher and Stampacchia,
1980, pp. 13-14). IT (3.3) has no solution, the transport network has no equilibrium
point belonging to 0: this means, from a practical point of view, that it has no equi-
librium solution at all.
It has been shown (Ferrari, 1994) that it is always possible to obtain an equilibrium
solution for a network with capacity constraints if one modifies in a suitable way the
link cost functions. In this section we illustrate this procedure when cost functions,
the inverse of demand functions and capacity constraints are symmetric. This means
that a link cost depends only on the flow travelling on this link:
(4.3)
where Hi is the capacity of link j, and I is the set of links whose capacity constraints
are considered in the network. The case of symmetry is interesting because, among
other things, the modification of link cost functions assumes a particular meaning.
Let X be the open set of vectors hERM which satisfy the constraints defined by
the generation capacity of centroids:
(4.4)
94
min[R(h):L,hpDjp~Hj
pEP
VjEI, h2:0, hEX] (4.5)
where
(E h,6;, f E h.6!...
R(h) = L, }O·EP ci(x)dx - L, }o·EP >'~(y)dy (4.6)
iEL 0 fET", 0
"'EW
(4.9)
crUd = CiUi) Vi ~ I
(4.10)
c';(f;) = Ci(f;) + Ui Vi E I
95
be the new cost functions. The modified path cost C;.(h), Pk E P!, is:
C;.(h) = 'L>:Cf;)D;Pk = L Ci(/i)Dip> + L UjDiP. = Cpo(h) + L UjDip.' (4.11)
iEL iEL iEI iEI
By substituting the expression (4.11) into Eq. (4.9), the latter becomes:
(4.12)
and coincides with the equilibrium definition (3.2). So, by using the cost functions
(4.10), the point h becomes an equilibrium solution.
The traditional method used to calculate an equilibrium solution solves the prob-
lem (4.5) neglecting the capacity constraints. Thus the problem solved is:
r " h 6· I: hpDip - Hi
pEP
[ ]
min[R(h) : h ~ 0, hE Xl (4.17)
tends to the solution h ofthe problem (4.5). If h, is the solution of the problem (4.17)
for a particular value of €, the KT condition for f sufficiently small can be written:
in hand Z<p are the Lagrange multipliers of the non-negativity constraints sP( h) :::; o.
The Pk-component of (4.18) is:
where d!w = L: h.pb~p. When f ..... 0 h • ..... h. By comparing Eq. (4.9) with
pEP
Eq. (4.19) when f ..... 0, and taking into account that there exists unique Lagrange
multipliers Ui ~ 0, ZPk ~ 0 that solve Eq. (4.9), we have that the lim U.i = Uj Vi E I .
• -0
Thus the solution of the problem (4.17) for f ..... 0 gives the solution h of the
problem (4.5) and at the same time the values of Lagrange multipliers Ui. If Ui = 0
Vi E I, h is an equilibrium solution for the network with the original cost functions. If
the multipliers Ui are> 0 for some links i E I, they represent the "road pricing" that
has to be imposed on these links in order to transform h into an equilibrium solution.
The solution of the problem (4.17) can be made arbitrarily close to the solution
of the problem (4.5) by choosing f sufficiently small. However high computational
difficulties can arise if we use a very small f value. For this reasons the popular
approach to the solution of problems that use penalty functions employs a sequence of
decreasing parameters f. With each new value of € a problem (4.17) is solved, starting
with the solution corresponding to the previously chosen parameter value.
5. A COMPUTATIONAL EXAMPLE
The method illustrated in the previous section has been applied to the small
network reported in fig. 2, which is travelled by two transport modes: car and transit.
Nodes 1. .. 5 are centroids: DID flows depart from them and arrive at them, but cannot
go through them. Every pair of centroids is joined on transit by only one link, without
need of passing through other nodes; every centroid is connected by a dummy link to
car links. Car and transit do not interfere with each other. Transport cost coincides
with time.
The journey time on transit is independent of flow, and is given by the sum of
line-haul time t~ reported in Tab. 1 for every pair w, and of access and waiting time ta
which is equal to 10 minutes for all pairs w. The time on dummy links is 0.1 minutes;
the time of every car link depends only on the flow Ii on it and is given by:
t; = ai + bi [ 10~O
r ]4 (5.1)
where parameters ai and bi are reported in Tab. 2 for every link i. Flows on both
car and transit links are measured in passengers per hour (Pph). Capacity constraints
express the condition that flow Ii on every car link has to be less than the EGi values
reported in Tab. 2.
97
(0:. 6- - - - - - - 7-------8~:Pl
1 2
I I I
I I I
I I I
I I I
o Centroids
I
I
I
I
I
L ______ -Q-
3
________
I
I
...J
o Nodes I 9 J J •••• 10 11
- Car links I
I
r------ JL - - - - - - - - - . . ,I
- - Transit links I I I
I I
........ Dummy links I I
I I
I12 ) 13 14 I
.I .~l
41~ - - - - - - - - - - - -
- - - ~'" 5
L _________________ J
~-----------------
Figure 2 Bimodal network considered in the computational example.
Tab. 1 - Transit line-haul times (min) between the OlD of the network in fig. 2
OlD 1 2 3 4 5
1 - 29 12 28 27
2 29 - 17 32 32
3 12 17 - 16 15
4 28 32 16 - 31
5 27 32 15 31 -
98
Tab. 2 - Parameters of cost functions and capacities ECi (Pph) on car links of the
network in fig. 2
The total transport demand d w on both modes for every pair w is fixed and is
given in Tab. 3. The demand for car travel d~ is obtained from Eq. (2.8), which in
the case under examination becomes:
de - d exp[-,Bt~]
(5.2)
w - Wexp[-,Bt:'u] + exp[-'Yta - ,Btt,]
where t:;, is the journey time by car for pair w, and coefficients ,B and 'Y are equal to
0.1 and 0.15 respectively.
Tab. 3 - Transport demand (Pph) between the OlD of the network in fig. 2
OlD 1 2 3 4 5
Tab. 4 - Flows Ii (Pph) and estimates uE;(min) of Lagrange multipliers for car links
of the network in fig. 2 corresponding to various value of parameter f
6-7 1286.75 0.867 1219.34 3.094 1201.19 6.071 1200.03 4.480 1200.00 4.748
6-9 1614.38 0.000 1569.93 0.000 1559.46 0.000 1559.02 0.000 1559.00 0.000
7-6 1278.77 0.788 1210.66 1.705 1200.37 1.916 1200.01 1.489 1200.00 1.904
7-8 1281.33 0.000 1276.26 0.000 1271.16 0.000 1270.46 0.000 1270.45 0.000
7-10 1128.86 0.000 1076.05 0.000 1041.43 0.000 1040.59 0.000 1040.57 0.000
8-7 1367.32 0.673 1331.28 5.004 1301.40 7.149 1300.05 7.487 1300.00 9.004
8-11 1743.15 4.431 1405.87 16.939 1304.12 21.118 1300.12 20.295 1300.00 21.221
9-6 1585.67 0.000 1527.32 0.000 1502.28 0.000 1501.63 0.000 1501.59 0.000
9-10 1090.09 0.000 1038.43 0.000 1020.03 0.000 1019.49 0.000 1019.49 0.000
9-12 1706.88 4.069 1392.62 14.819 1303.16 16.169 1300.10 16.542 1300.00 17.832
10-7 1034.89 0.000 1012.35 0.000 1010.38 0.000 1010.99 0.000 1011.02 0.000
10-9 1001.04 0.000 980.39 0.000 961.64 0.000 962.07 0.000 962.07 0.000
10-11 856.80 0.000 842.98 0.000 840.69 0.000 840.85 0.000 840.87 0.000
10-13 1136.87 0.000 1101.58 0.000 1082.63 0.000 1082.25 0.000 1082.23 0.000
11-8 1606.53 3.065 1360.34 9.654 1302.15 10.994 1300.07 11.613 1300.00 11.925
11-10 940.32 0.000 887.05 0.000 842.54 0.000 840.91 0.000 840.87 0.000
11-14 1639.52 2.395 1451.57 8.252 1402.03 10.372 1400.06 9.177 1400.00 9.934
12-9 1767.21 4.672 1408.05 17.288 1304.38 22.410 1300.13 21.654 1300.00 21.534
12-13 1356.61 1.566 1253.00 8.481 1202.14 10.962 1200.07 11.644 1200.00 12.883
13-10 1181.94 0.000 1114.61 0.000 1083.45 0.000 1082.29 0.000 1082.23 0.000
13-12 1276.94 0.769 1229.90 4.784 1201.28 6.539 1200.04 5.766 1200.00 6.263
13-14 1571.71 3.717 1280.95 12.952 1202.86 14.658 1200.08 13.803 1200.00 13.504
14-11 1586.41 1.864 1450.11 8.018 1401.91 9.756 1400.06 9.697 1400.00 9.415
14-13 1537.11 3.371 1270.88 11.340 1202.82 14.432 1200.09 14.907 1200.00 14.759
100
ti = 731[1n dwd~
_ dt, + fib
tw + ,ta ] . (5.3)
A cost function ii(fi) modified following (4.14) was attributed to every car link:
i- =
I
a'
I
+ b· [~]4
I 1000 if f,.I > EG·I
(5.4)
.
t·I = a'I + b·
I
h]4 + -(f,.
[-1000 1
f
- EG·)
I I ,
f > 0 if h ~ EG;.
REFERENCES
[1) S. Bazaraa, D. Sheraly and C.M. Shetty, "Nonlinear Programming: Theory and Algorithms",
John Wiley & Sons, New York, N.Y. ,1993.
[2) S. Dafermos, "The general multimodal network equilibrium problem with elastic demand", Net-
worL 12, 57-72, 1982.
[3) T.A. Domencich and D. McFadden, "Urban Travel Demand", North Holland, Amsterdam, 1975.
[4) P. Ferrari, "Road pricing and network equilibrium" , forthcoming in Transpn. Res. B 1994.
[5) D. Kinderleher and G. Stampacchia, "An Introduction to Variational Inequalities and Their
Applications" , Academic Press, New York, 1980.
[6) D. McFadden, "Conditionallogit analysis of qualitative choice behaviour", Frontiers in Econo-
metrics, Academic Press, New York, NY, 105-142, 1975.
[7) G. Ruberti, "I vincoli di capacita delle reti stradali urbane" "Internal Report of Istituto di Strade
e Trasporti dell'Universita di Pisa", 1994.
[8) Y. Sheffy, "Urban Transportation Networks: Equilibrium Analysis with Mathematical Program
ming Methods" , Prentice-Hall, Englewood Cliffs, N.J., 1985.
SEPARATION OF SETS AND GAP FUNCTIONS FOR
QUASI-VARIATIONAL INEQUALITIES
Giannessi Franco
Department of Mathematics
University of Pisa
Via F. Buonarroti 2, 56127 Pisa, Italy
e-mail: [email protected]
1. INTRODUCTION
101
102
where C is a closed and convex cone with apex at the origin. This form can be in-
terpreted as the search, among the fixed-points of the point-to-set map K, for one
which satisfies (2.1); it obviously includes a classic form of QVI, when Vy E X(y) the
condition g(y;x) E C is identically true (with respect to x) on X(y); the fact that the
format (2.1) embraces such kinds of 9 will be exploited later to recover known gap
functions.
When X(y) and g(y; x) are independent of y (in this case they will be denoted by
X and g(x), respectively), then (2.1) collapses to a Variational Inequality, which will
be denoted by VI. If, moreover, K(y) == X(y) and, Vy,K(y) is a closed and convex
cone with apex at the origin, then (2.1) collapses to a Quasi-Complementarity System
(in a Hilbert space):
where K*(y) denotes the (positive) polar of K(y). When K(y) is independent of y,
then (2.1)' becomes the Classical Complementarity System.
In [8J, Sect. 3, it has been proposed to associate an image space to a VI; this is
now done for a QVI. Such an approach will lead to define a wide class of gap functions,
containing the known ones. The gap functions we will obtain have a Lagrangian taste,
in the sense that they allow us to free ourselves from the constraints g(y; x) E C in
defining the gap function. The approach proposed in [8J starts with the obvious remark
that y E K(y) is a solution of (2.1) iff the system (in the unknown x):
is infeasible. The space where (u, v) runs is the image space associated to (2.1), and
the set:
is the image of (2.1). To system (2.2) we associate the set 1{ := {(u, v) E R x R m : u >
0, v E C}, which depends on the types of inequalities only; another obvious remark is
103
that the impossibility of (2.2) is equivalent to 'Ii n K(y) = 0. To show this disjunction
in [8] a separation scheme is proposed; this approach has been developed in [9]. It will
now be slightly modified and applied to (2.1) in order to define a general class of gap
functions; concepts and notations of [9] will be here understood.
In the remaining part of this section we will develop the particular case where
G(y;v;W):= (Gi(y;Vi;Wi), i E I) ,
m
Gi : :::: x R xQi -+ R, W = (Wi, i E I), Wi E Qi , Q = X Qi;
i=1
where the level sets are considered with respect to (u, v) only. Under these conditions
(2.3) is a weak separation function in the sense of [9].
Each G i may be considered as a transformation of gi; for this reason,
Vy E X(y), VWi E Qi, Gi must be such that
Remark 2.1. According to [9), (2.3) is a particular case of weak separation function
and then, in further investigation, might be useful to replace it with a more general
function:
w(y; u, v;w) := (Ju + ,(V; v;w), wE Q ,
will be shown to be a gap funtion for (2.1), and will be called weak gap function, since
it comes from (2.3).
In the definition of 'ljJw there is a maximization; this is performed on X(y), i.e. on
the relaxation of the domain of the QVI, and not on K(y), unlike what happens (in
the case of a VI) to the existing gap functions.
Note that, in the case (2.1)', we easily find:
Theorem 2.1. Let y E KO. Assume that the extrema in (2.6) exist and for each
y E K(y) there exists w(y) E Q, such that
(i) £(y;w(y» is convex;
(ii) U := {(u,v) E 1{ : v = O} <1 T(&(y;w(y))), where T denotes Bouligand
tangent cone at (u = 0, v = g(y; y».
Then y is a solution of (2.1) iff 'ljJw(y) = o.
Hence
max w(y;(F(y),y-x), g(Y;X);A,W(Y»~O, VAEC',
xEX(y)
105
Hence
max W(Yi (F(y), Y - x), 9(Yi x)i A,w(y» ~ 0, 'VA E C*,
"EX(y)
so that tPw(Y) ~ O. Ab absurdo, assume that tPw(Y) > O. Then 3 a > 0, such that
and hence
(2.7)
Since Y solves (2.1), i.e. (2.2) is impossible or 'H n K(YiW(Y» = 0 or, equivalently,
'H n t'(YiW(Y)) = 0, then (i) and the obvious convexity of'H imply the existence of at
least one hyperplane which separates 'H and t'(Yiw(y)). Because of Theorem 2.1 of
[9], (ii) is necessary and sufficient in order that at least one of the above hyperplanes,
say H, does not contain Ui and this implies that there exists disjunctive separation,
namely t'(YiW(Y)) can be included in one closed halfspace defined by Hand 'H in its
complement, or 3 XE C· such that:
which implies
(2.8)
and contradicts (2.7). Now assume that tPw(Y) = o. Then 3 AE c· such that
so that
(F(y),y-X) + (X,G(Yi9(YiX)iW(Y») :::;0, 'VxEX(y). (2.9)
(F(y), Y - x) > O.
Then, since 'VA E C· we have (A, 9(Yi x») ;::: 0, and, because of assumption (2.5), also
(A, G(Yi 9(Yi X)i w(y))) ~ 0 , we find
Remark 2.3. IfVy E X(y) the condition g(y;x) E C is identically true on X(y),
so that K(y) = X(y), then Theorem 2.1 holds without (i)-(ii), so that tPw(Y) is a
gap function whatever F(y) and X(y) may be. To see this the proof of Theorem 2.1
requires a few changes only. In fact, according to Remark 2.2, the sufficiency does not
require any change. The necessity holds without (i)-(ii) up to (2.7). At this point note
that now
Since JC(y) n 11 = 0, we have JC(y) ~ V\ 11, so that any hyperplane of R1+m, whose
equation is
p
u +L .xi Vi = 0 , if p > 0, or u = 0, if p = 0 ,
i=1
implies a disjunctive separation between JC(y) and 11. Then we again achieve (2.8) and
the absurd.
If the domain of the QVI is given as a set K(y) and not furtherly specified like in
(2.1), then - as said at the beginning of this section - suitable 9 and C can be found so
that, Vy E X(y), 9(Yi x) E C be identically true on X(y). Hence, as shown in Remark
2.3, (2.6) is gap function whatever G may be. For instance, we can set m = 1, C = R+,
and g(y;x) = 0, Vx E X(y), and choose G(y;v;w) = V Vv E R, Vw En:= R (the
definition n is of no importance); then (2.6) collapses to:
When X is independent of y, or QVI becomes a VI, then (2.10) becomes the gap
function introduced by Auslander [2].
In Theorem 2.1 we do not assume that K(y) is convex even when K(y) is indepen-
dent of y, unlike what happens in [2,7,16]. However it may happen that, K(y) being
not convex, (i) of Theorem 2.1 turns out to be satisfied. This is shown by the following
example.
Example. Consider (2.1) as VI (K(y) is independent of y; X(y) and K(y) are denoted
by X and K, respectively) in the particular case, where Xc R 2 , C = R+, m = 1, and
D;
set:
F(y) = G~: =: X = {(XI, X2) E R2 : ~ ~ Xi ~ 1, i = 1,2};
g(x)=x~+x~-1i
so that F is strictly isotone, but K is not convex. JC(y) is the set of (u, v) E R X R,
such that:
107
or, if Y2 =I- !,
V= (1-~Y2)2 {u 2 + 2[YI + Y2 - 2(yi + yn + (2YI - I)XI]U+
{
+2[2(Yi + y~ - YI - Y2) + l]xi+
+2[(YI + Y2 - 2y; - 2yD(2YI -l)]XI + (YI + Y2 - 2yi - 2y~)} - 1,
3V2 5 -3V2
v = (3 + 2V2)u 2 - (2 + V2)u - V2 - 4 , 2 - -2- ~ u ~ 4
and is art arc of a parabola; it does not intersect 'H., since such an Y solves (2.1), as it is
easy to see. If we select 'Y = v, namely if we choose a linear separation function (2.3),
then [(V; w) is not convex. While, if we consider (2.3)" and set 'Y = 1 - exp( -wv),
then simple calculations let us see that at w > 8 the above arcs of parabolas pass
from convexity to concavity. For instance, with w = 10 consider the arc of K(y; w) at
Y -_ (./2 ./2)
2'2'
X =
I 2'
./2.
3V2 5 - 3V2
v=1-exp{-10[(3+2v'2)u 2 -(V2+2)u]}, 2--2-~uE 4 ;
we find
.exp{-10[(3 + 2V2)u 2 -
3V2 , 5 - 43V2] '
(V2 + 2)u]) < 0, Vu E [ 2 - -2-
and hence the concavity of the arc follows. Now [(V; 10) is convex. (i) of Theorem 2.1
is now fulfilled, while before was not. The above arcs are the images of the segments
t
{(XI,X2) E R2: Xl =constartt; ~ X2 ~ I}, ~ Xl ~ 1. t
Since (F(y), Y - x) is linear in x, such arcs become concave because of the trartsforma-
tion of g; more precisely 'Y can be interpreted as the convexification of the restrictions
of 9 on the rays {(XI,X2) E R2: Xl =constant; X2 ~ 2~}' t ~ Xl ~ 1, and hence on
the above segments. In conclusion, since (ii) is obviously verified, Theorem 2.1 Cart be
applied, notwithstanding the fact that K is not convex.
Same results are obtained if the above operator F is replaced with an anti tone
operator, as
F(y)=(2-2 YI ).
2 - 2Y2
108
Now it will be shown that, correspondingly to what happens for the general situ-
ation of the Theory of Optimization, we have the strong case here too. This leads us
to define another class of gap functions. Consider again (2.1), and the function:
where h: B x R m x11 -+ R is, Vy E KO, such that lev>o S C 'H, and h and 11 must be
such that, Vy E K O , the following conditions are satisfied:
which will be called ~trong gap functionj the term strong is motivated by the fact that
(3.1) is a strong separation function in the sense of [9].
Theorem 3.1. Let y E K O• Assume that the maxima in (3.3) exist, and that
conditions (3.2) hold. Then y is a solution of (2.1) iff t/J.(y) = O.
so that 'H n K;(y) =F 0. Hence, because of (3.2b), 3w E 11 and 3x E K(y) such that:
S(Yj (F(y), Y - x), g(Yj x); w) > O.
Remark 3.1. Note that we have tf;.(y) 2: 0, Vy E KO, as shown by the proof of
Theorem 3.1. Moreover, the proof of Theorem 3.1 does not require the existence of
the two maxima in (3.3).
o"( ·v·w ) - { wv , if v 2: 0
y" - +00, if v < 0
we have that the function s obviously satisfies the inclusion lev>o s C H. At x = y
we find that (u = 0, v = 0) E K:(y) and s(y; 0, 0; w) = 0, Vy EX, Vw 2: 0; hence (3.2a)
is verified at any w 2: o. In order to discuss (3.2b), note that, Vy E X, we have:
K(y) = {(u,v) E R x R: u = (F(y),y - x) ,
(3.6)
v=(x-y,A(x-y)), xEX}C{(u,v)ERxR:v2:0};
x EX} { 3£ E X s.t.
(3.7)
(F(y), y - x) > 0 => (F(y), y - £) > (£ - y, A(£ - y)).
(3.6) is obvious. Now let us prove [15] (3.7). Suppose that there exists x E X such
that (F(y), y - x) > o. Let us set ii := (F(y), y - x) , /3:= (x - y , A(x - y)). (Note
that (ii,/3) E K(y)). If ii > /3, then £ := x is the required point. Assume that /3 2: ii.
We set x(t) = (1 - t)y + tx, t E [0,1]. Since y, x E X and X is convex, then x(t) EX.
Thus we have (a(t), f3(t)) E K(y), where
a(t) := (F(y), y - x(t)) , f3(t) := (x(t) - y, A(x(t) - y)) .
We have
a(t) = (F(y), y - x(t)) = (F(y), y - (1 - t)y - tx)
= t(F(y), y - x) = tii,
f3(t) = (x(t) - y, A(x(t) - y)) = (t(x - y), A(t(x - y))}
= t 2 (x - y, A(x - y)) = t 2 /3.
Since
f3(t) t 2 /3 /3
--
a( t)
= -tii
= t- --+ 0
ii
as t --+ 0,
¢.(y) = maxmax[(F(y),y
"'Ell ",EX
- x) - (x - y, wA(x - y»)],
which is the gap function considered by M. Fukushima [7]. Note that we could have
chosen other sets n. We have made the simplest choice: a singleton. This is not unique.
We might have chosen n = {w > O} finding the same result, or n = {O} finding again
(2.10) with X in place of X(y). The Auslander gap function is so particular that, with
it the two approaches become the same thing.
In a quite analogous way, by setting g(y;x) = A(x,y), A being non-negative,
continuously differentiable on X x X, strongly convex on X with respect to x, Vy E X,
and such that A(x,x) = 0, and
V.,A(y,y) =0 , Vy E X,
then we recover the gap function considered by D.L. Zhu and P. Marcotte [16].
In fact, we can prove the following [15].
f1(t) _ oCt) 0
at( )
-
ta
--+ as t --+ o.
Hence, 3i E ]0, 1[ such that ~~2 < 1, and x := xCi) = (1- i)y+tx is the required point.
This completes the proof. 0
111
Now consider again (2.1) as VI (X(y), K(y) and g(y; x) are replaced with X, K and
g(x), respectively) in the particular case where X = :=: = IRn , C = 1R+'; furthermore
let us set:
g(x)=Ax-b, S1=lR+" w='\, yT=(Yl, ... ,Yn), ( 4.1)
It is easily seen that now K(y) is affine so that fey), being the sum of an affine
variety and a closed convex cone (namely -cRH), is convex - so that (i) of Theorem
2.1 is fulfilled - and coincides with its Bouligand tangent cone at (u = 0, v = g(y)).
Now note that 1-{ n K(y) = 0 implies the existence of a hyperplane, say H, which
contains K(y) and such that 1-{ C H+, H+ being one open halfspace defined by H;
therefore, T( f(y)) equals f(y) and, being included in the complement of H+, does not
intersect 1-{ and hence U, so that (ii) of Theorem 2.1 is satisfied at those y which solve
the present VI. Since this is enough (as noted in Remark 2.2), then Theorem 2.1 can
be applied.
Now we have to analyse the minimization of (2.6). To this end note, first of all,
that, without any loss of generality, we can assume that the system ,\ 2': 0, ,\A = F(y)
be feasible. In fact, in the contrary case, Vy E K, a well known theorem of the
alternative for linear systems (see [12], page 33, Th. 10) would give the existence of
t E lR n (depending on y) such that:
(F(y), t} =1 At ~ O.
By setting x := y - t, we would get Ax 2': Ay 2': b and thus deduce the existence,
Vy E K, of x E K such that:
(F(y),x - y) < 0
Remark 4.1. The symbols adopted in the present Remark are generic and indepen-
dent of those used elsewhere. Assuming that the minima exist, note that we have:
where the I-st minimization is performed with respect to y, the 2-nd with respect to
A, and the the 3-rd with respect to (y,A). Denote by P',P",Q(y) the problems in
(4.4a) and that inside P', respectively; by M" and A(y) the sets of minimum points
of P" and Q(y), respectively. Let M' be the set of pairs (y, A(Y)), where A(y) E A(y)
and y is a minimum point of P'. Then we have
M'=M" . (4.4b)
The proof is quite obvious: ab absurdo let (y., A·) E M"\M' and (y,). := A(Y)) EM',
so that, A· being a feasible solution for Q(y.) and y. for P', we have:
Collecting (4.5) and (4.6) leads us to the absurdo. Hence (4.4) follow. o
where the minimization is performed with respect to (y, A), and where (w is considered
as fixed):
H(y,A):= - max [(A,G(y;g(y;x);w») - (F(y),x)].
xEX(,1)
In the case (4.1)-(4.2) we find H(y, A) = (A,b), so that, taking into account (4.3) and
the preceding analysis, (4.7a) becomes:
Remark 4.2 In (4.1) K is defined by inequality only. If there are both equalities,
say p, and inequalities, say m - p, we set C = RP X R+-P j then the preceding analysis
receives only formal changes.
113
Remark 4.3 The preceding development receives only formal changes, when g(x) =
Ax - b is replaced with g(y; x) = Ax + By - b, with B of order m x n, so that we have
now a QVI. Instead of (4. 7b), we find:
min
(A+B).~'
[(F(y), y) + (A, By - b)]. (4.7)'
F(.)-~A=O
,,~o
F(yf = C+ Cy (4.8)
where cT := (Cl,""C n ) and C:= (cij,i,j = 1, ... ,n). In this case (4.7b) becomes a
so~called quadratic programming problem, since C can be obviously replaced with
C := t(C + CT), if not symmetric. If moreover C is positive semidefinite, then (4.7b)
is a convex quadratic program. Note that in th<; case (4.8) the VI is a l~st order
optimality condition of problem min (y, C+ ~Cy) subject to y E K.
In some applications, like in the study of equilibrium of flows in a network, it is
useful to consider explicitly bounds on the variables. Hence X is replaced with
Now we will discuss two ways of solving the present VI by minimizing the gap function.
Because of Remark 4.1 the minimization of 'ljJw(y) now becomes:
mi~'ljJw(Y)
yEh
= min[(c,y)
.EK
+ (y,Cy) - (A,b) + cI>(y,A)], (4.11)
~~o
Note that cI>(y, A), being the maximum of a parametric linear problem, where the
gradient of the objective function depends linearly on the parameter (y, A), is convex
and piecewise linear; call Cl, ... ,Cr the cones where cI>(y,A) is linear.
Another way consists in setting J = J(y, A) := {j = 1, ... , n : {>.A - yTC}j 2: Cj};
thus we easily find:
where
2: + t f (l:
(4.13a)
!
subject to
n
~
L.J Aa"
I IJ
>
- Cj + L: YrCjr, j E J,
i=I r=I
m n
(4.13b)
L: Aiaij S Cj + L: YrCjr, j ~ J,
i=I r=I
A ~ O.
In fact, since now X is an interval of Rn, the maximization which appears in (2.6)
is easily solved in terms of the sign of the elements of vector AA - yTC - cT.
Note that i.f'w(Y), being the minimum of a parametric linear problem whose right-
hand side depends linearly on the parameter y, is convex and piecewise linear; call
DI, ... ,D. the cones where i.f'w(y) is linear.
If C is positive semidefinite, then (4.11) and (4.12) are convex and thus we can
use any method of convex quadratic optimization to minimize it on K.
If C is negative semidefinite, then the objective functions of (4.11) and (4.12) can
be viewed either as continuous and piecewise concave on the polyhedra
DI n K, ... , D. n K, or as a difference of convex functions, namely <I>(y, A) - (A, b)
or i.f'w(y) and -(c, y) - (y, Cy). For the former case some interesting analysis already
exists in [5]. In the latter case the minimization falls into a class of problems which
have received a lot of attention [11]; below some details will be discussed.
If C is indefinite, then we can use classic decomposition techniques to reduce
ourselves to subproblems of the above kinds.
Now assume that C be negative semidefinite and see (4.12) as a difference of two
convex functions (in short, DC-functions), so that the minimization of (4.12) can be
written as
min[1/!w(Y) = i.f'w(y) -u(y)], s.t. Y E K, (4.14)
y
where u(y) : -(c, y) - (y, Cy); i.f'w and u are obviously convex. We assume that
.((; := sup 1/!w(y) < +00 ; rp:= inf i.f'w(y) > -00 ;
yEK yEK
(4.15)
if := sup u(y) < +00.
yEK
where now R := [0, .((;]C It , and S := [rp, if + .((;]C R. In fact, by a well known device
(4.14) is obviously equivalent to
Since i.f'w(y) S u(y) + riff 3s E R such that i.f'w(y) S s S u(y) + r, problem (4.17)
turns out to be equivalent to
Now, to see the equivalence between (4.14) and (4.16) it is enough to note that, by
Theorem 2.1 and the subsequent remark, tPw(Y) ;:::: 0 Vy E K and thus on the feasible
region of (4.17) we have r ;:::: 0, and to take into account (4.15). Problem (4.18) is of
the kinds studied in [11], where some solving algorithms are proposed. Quite similar
remark can be done for (4.11).
This section contains some remarks, which aim to stimulate the investigation on
the duality theory for QVI. Among the several aspects, which concern the present
topic, only one is considered here: the connections between gap functions and duality
for extremum problems. Such connections can be analyzed in at least two ways. With
the former we perform suitable transformations directly on the gap function. With the
latter we replace the minimization of the gap function with a simplified minimization,
like (4.7), and perform the transformations on the simplified problem.
Now let us describe the former way. Consider again (2.6) and the transformed
operator Fo(y) := F(y) - F(O). Within square brackets of (2.6) substract and add
~(Fo(Y),Y); then the minimization of (2.6) becomes:
m~n
YEK(y)
{(F(Y) - ~Fo(Y),y)
2
+ ~(Fo(Y),y)+
2
(5.1)
min max [(A,G(y;g(y;x);w)) - (F(Y),x)]} ,
AEC· xEX(y)
where 8* denotes support function. Instead of (5.3) we can equivalently (in the sense
116
that the maxima appearing in them are equal) consider the following:
min!f1w(Y)
yEK
~ min(F(y)
yEK
- -21 Fo(y),y)-
1 (5.4)
- 'E~;~R+ [(A, b) - 2(Fo(y), y) - (fI, d)]
pOtAT'-F(.)
pERi-
where K* := {(A'fI) E Rf.' x Rf, : ATA - fI ~ e}, and where the equalities are moti-
vated by the fact that Th.2.1 can be applied and implies that the LHS of (5.5) is zero,
and by the so-called weak duality theorem for linear programs. As a consequence of
(5.5) we obtain the so-called strong duality theorem for linear programs (on polytopes).
In the particular case where F(y) = e + Cy (as in (4.1), (4.8), (4.9)), so that
Fo(Y) = Cy, (5.4) becomes:
and is obviously a (l-st order) necessary and sufficient optimality condition for y to
be minimum point of the I-st problem in the 2-nd side of (5.5). Hence, it is natural
to consider the analogous (l-st order) necessary and sufficient optimality condition for
the 2-nd problem in the 2-nd side of (5.5), namely
where K* is that of (5.5), and (X, jt) is the unknown; when d = +00, then in (5.8) we
must put fI = it = O. It is natural to call (5.8) dual variational inequality of (5.7).
Note that the operator of (5.7) is not injective. It is useful, even if quite obvious, to
note that, if we consider (5.8) in place of (2.1), and we repeat the development which
has led to (5.5) (note that now the domain of the VI is K*, so that now d = +00 and
A, b are respectively:
117
and F(y) is replaced with (-b, d), so that Fo == 0), then, in place of the problems which
appear in the 2-nd side of (5.5), we find respectively:
and
max
>'ERm ;JlERnjIER+
[-(c, y)] = min
Ay>b
(c, y).
O:s;i~d
I'ER+irER+;
_,T AT+IITS-bTill+r:S.d
Hence we find again the problems in the 2-nd side of (5.5), but in reverse order.
Consider again the case F(y) = c + Cy, where (2.1) becomes:
The 1-st order condition for the 2-nd problem in the 2-nd side of (5.6) is
where K* is now that of (5.6), and the unknown is (A,p,y). We call (5.10) dual
variational inequality of (5.9). We consider (5.10) in place of (2.1) and we repeat the
development which has led to (5.6). Now the domain of the VI is K*, so that d = +00
and A, b are respectively
and F(y) is replaced with (-b,d,Cy), so that Fo(A,p,y) = (O,O,Cy). Then, in place
of the problems which appear in the 2-nd side of (5.6), we find respectively:
and
max
.\ERm i}lER" iYER+.
[-(c, y) - -21 (Cy, y)] = min [(c, y)
Ay>b
+ -21 (Cy, y)] .
vER+ iTER+.
1;5
0;5 d
_liT AT+"T5-bTiy+r5d
.TC$(C.)T
Hence we find again the problems in the 2-nd side of (5.5), but in reverse order.
Now let us consider a second way for establishing relationships between QVI and
duality for extremum problems. (5.1) has been obtained through a special transforma-
tion of (2.6). Now we start with (4. 7a) instead of (2.6). Since in (2.6) the existence of
the maximum is not required (so that we can read supremum in place of maximum),
118
then the minimization of (2.6) is equivalent to (4. 7a). Let f : :::: -t R be a generic
function, and set
L(y; >.) := fey) - (F(y), t) + H(y, >.).
By exploiting (4.7a) we find:
min ~w(y) = min {fey) + [-L(y; >.)]} ~ min fey) - maxL(y; >.), (5.11)
YEK(y) .~:c(!) YEK(y) AEC·
where the 1-st and 3-rd minimizations are done with respect to y, and the 2-nd and
4-th with respect to (y, >.).
In the case (4.1)-(4.2), taking into account that H(y,>.) = (>.,b) and that we can
assume F(y) = >.A, so that L(y; >.) = fey) - (>.,Ay - b), (5.11) becomes:
Denote by P and P' respectively the problems in the RHS of (5.12). Let (y', >") be
an optimal solution of po. If y' is an optimal solution of P (so that Ay' ~ b), then
we easily deduce that (>", b) :::; (>", Ay*), and thus:
min fey)
Ay>O
= max
F(,)-AA=O
L(y; >.). (5.13)
- .\~o
When F(y) is the gradient of a convex function, which is chosen as fey), then P*
turns out to be the Lagrangian (Wolfe) dual [12J of P.
6. SOME EXTENSIONS
We shall now briefly discuss some questions that arise from the development car-
ried out in the previous sections.
The analysis of Sect.5 has been done by exploiting a weak gap function. It can be
done analogously in the case of a strong gap function of Sect.3.
In Theorem 2.1 we assume the existence of w. This is guaranteed if 9 is C-concave
and K convex. More general conditions, which ensure the linear separability of t: and
'H, are useful, as well as the comparison with the existing ones. Analogous question
arises in Theorem 3.1. An important class of problems, recently studied in terms
of VI and QVI, is the one that considers with the equilibrium of flows on a road
or on a computer network. Same questions for (ii) of Theorem 2.1; under convexity
assumptions the well known Slater condition (or its generalization) guarantees it.
119
between the separation scheme adopted in (14) for extremum problems and the one
exploited here. In the former case the image of an extremum problem receives a trans-
lation in the direction of u axis, if the candidate y for a minimum point is changed;
here, as (2.2) shows, the image receives any transformation, if y changes.
The uniqueness of solutions to a VI or to a QVI may be reduced to that for
the corresponding gap function, by using the following obvious proposition. Given a
function (not necessarily the gap function) t/J : K ~ JRn -+ JR, y is unique minimum
point iff, Va E R,n, the system (in the unknown z)
is impossible. In the particular case of VI with C = JR+ and t/J as the Auslander gap
function, it becomes: a solution y of the VI is unique iff, Va E R,n, the system (in the
unknown z):
is impossible. All this can be extended to QVI and to the infinite dimensional case.
The format (2.1) is not the most general one for VI. In [3) a more general scheme
is considered: to find y E K such that
F(y,x)~O, VxEK,
This differs from a general system, like that considered in [9), for the fact that
F( x, x) = 0 and for the dependence of the domain X on the unknown. The format
(2.1) contains also infinite dimensional QVI, like the case where g(y; x) E C represents
an integral constraint, but does not contain the case where g(y; x) E C is a geodesic
constraint or a differential equation or inclusion. Extension of the results of Sect.s 2-5
to these cases may be conceived by means of the multifunction approach outlined in
Sect.4 of [10).
ACKNOWLEDGMENT
The research has been partially supported by the Project "Trasporti" of Italian
National Research Council (CNR), which is here gratefully acknowledged.
121
REFERENCES
Gwinner Joachim
Institute of Mathematics
Department of Aerospace Engineering
University of the Federal Army Munich
D - 85577 Neubiberg, Germany
e-mail: [email protected]
Abstract: The convergence and stability theory of Mosco is extended by the more
general approach of monotone - convex functionals and by weakening uniform equi-
monotonicity and equicoervity assumptions to semicoervity conditions. The abstract
theory is firstly applied to a finite dimensional Variational Inequality that models dis-
tributed market equilibria with box constraints providing a stability result with respect
to the upper bounds. Then it is applied to obtain stability results with respect to co-
efficients and unilateral constraints in p-harmonic elliptic unilateral boundary value
problems that can be considered as scalar models of the unilateral contact problem
without, respectively with given friction.
1. INTRODUCTION
123
124
Daiermos and Nagurney [l1J give a stability result for perturbations in the cost func-
tions. More recently McLinden [33J comprehensively investigates stability understood
as nonemptiness of the solution set of locally perturbed Variational Inequalities.
In this paper we follow the stability concept of Mosco and use his notion of set
convergence, what is today called "Mosco convergence". Instead of considering single
valued monotone operators we more generally work with monotone-convex functionals
as initiated by Minty and Oettli [35,5J. This approach allows to consider also set-
valued monotone operators. However, the main difference between our contribution
and previous work is that we relax uniform monotonicity and hence coercivity to
semicoercivity conditions. Under these conditions solvability and even more stability
of the Variational Inequality depend in a subtile manner on the data of the problem,
namely on the monotonicity and growth behaviour of the nonlinear monotone-convex
functional, the constraints and the inhomogenous term at the right hand side. Thus
our analysis do not only cover the well-posed case [13, Chapter II, Section 5J, but also
treats the ill-posed case [14, section 7.3.3J, [26, Chapter VIIJ, where uniqueness may
be lost. Therefore we arrive at new results also in finite dimensions.
As the monograph of Glowinski [18J shows, Mosco convergence can be refined to
give an excellent tool for the analysis of the convergence of discretization methods, in
particular finite element methods applied to elliptic Variational Inequalities. Also the
abstract stability result presented in this paper can be understood as an approximation
result that applies to the convergence of finite element approximations for semicoercive
nonlinear elliptic unilateral boundary value problems; for details we refer to [22J. A
related discretization result in the case of bilinear forms, where uniform monotonicity
or equivalently positive definiteness is relaxed to a Garding inequality, is given in [20J
and applied in [21J to the convergence of boundary element approximations to contact
problems in linear elastostatics with Signorini boundary conditions. II-posed elliptic
Variational Inequalities in terms of weakly coercive bilinear forms and their application
to contact problems are treated by iterative multi-step prox-regularization methods in
the recent monograph [26, Chapter VIIJ. For earlier convergence analysis of Galerkin
discretization of monotone operator equations under relaxed coercivity conditions we
refer to the poetic paper [49J.
In the present paper convergence results for discretization methods are outside of
the scope. So to illustrate the power of our abstract stability result we derive sta-
bility results for perturbed data, like perturbed constraints and perturbed operators,
in concrete Variational Inequalities. These Variational Inequalities model some equi-
librium problems drawn from operations research, respectively model some unilateral
boundary value problems stemming from mathematical physics.
Let us formulate the Variational Inequality problem in the following way: Find ii.
in a real reflexive Banach space (V, II . II) such that
(P) ii.EK,cp(ii.,v)~'\(v-ii.), VvEK.
Here K ~ V is nonempty, closed, and convex, ,\ belongs to the dual space V·, and
cp : K x K -+ IR satisfies appropriate convexity and monotonicity assumptions. In
particular, following Minty-Oettli [35,5J, cp is termed monotone, if
(2.1)
125
Moreover, for any u E K, <pC u, .) is assumed to be convex and lower semi continuous on
K, and to satisfy <pC u, u) = O. Note that a set-valued monotone operator T : V ---> V·
gives rise to a monotone-convex functional by
Also <P(Vl,V2) = t/J(V2) - t/J(Vl) is monotone-convex for a given convex lower semi-
continuous functional t/J. Hence the above formulation includes so-called Variational
Inequalities of the second kind [18] without pain.
To introduce further assumptions, let I . I be a continuous seminorm on V. Thus
Y = {y E V : Iyl = O}
is a closed subspace. With ~ denoting weak convergence in contrast to strong conver-
gence denoted by -+ , we require:
(AI) For any sequence {Vn}nEN with Ivnl-+ 0, Vn ~ v and IIvnll ;::: ." for some.,,> 0
there exists a subsequence {v n • hEN such that Vn.-+ v.
Obviously in finite dimensions, (AI) trivially holds, since then weak and strong con-
vergence coincide. To state more concrete sufficient conditions for (AI) to hold, we
insert from [25, Sect. 5.3 (25)], [19, Sect. 3.4] the following
Lemma I The IL8sumption (A 1) is satisfied in any of the foliowing two cases:
(i) 11·11 is equivalent to 111·111 + 1·1, where 111·111 is another norm on V; the dimension
of Y is finite; there exists a > 0 such that
(ii) V is compactly embedded in another Banach space (X, II . IIx); II 'IIv == II· II is
equivalent to 1·1 + II . IIx .
Moreover the seminorm 1·1 is continuous.
Proof. The inequality 1·1 :$ 111·111 + 1·1, respectively 1·1 :$ 1I·lIx + 1·1 and the equivalence
of the sum to the original norm 1I·lIv implies the continuity of the seminorm 1·1.
To prove (i) let {Vn}nEN a sequence in V such that Vn ~ V, Ivnl-+ o. Then for any
n choose Yn E Y such that
and by the equivalence of the sum to the original norm 1I·lIv one obtains IIwnliv -+ O.
Thus Yn = Vn - Wn ~ v in the finite dimensional subspace (Y, II . II) whose unit ball
is compact. Hence for the bounded sequence {yn} there exists a subsequence {Yn.}
strongly converging to v. Altogether one concludes IIvn • - vllv = llYn. +W n • - vllv -+ O.
To prove (ii) let {Vn}nEN a sequence in V such that Vn ~ V, Ivnl -+ O. Since the
seminorm I· I is weakly lower semicontinuous, Ivi = 0 holds. By the boundedness of
the sequence {v n } and by the compact embedding of V in (X, II . IIx) there exists a
subsequence {v n .} strongly converging in II . IIx to some x EX. Since X* ~ V* and
126
the weak topology separates points, x = v holds. Finally the equivalence of the sum
to the original norm 1I·lIv again implies IIvnk - vllv --+ O. 0
As we shall see later, the assumptions of Lemma 1 can be satisfied in Sobolev spaces
in virtue of the Kondrachov-Rellich embedding theorem.
Now we introduce several assumptions concerning the functional cpo Having our
applications in mind we can take for granted that there exists some y E Y n K and
some constants ao ~ 0, al ~ 0 such that
Note that this condition is satisfied, if the affine minoranat, which is obtained by
subdifferentiability of cp(y, .) at y, is continuous with respect to I . I. Since we need
detailed information about the monotonicity, growth, and Lipschitz behaviour of cp,
we require:
where 10 ~ 0 ,,1 ~ 0 , and the functions 9 and G : [0, 00) --+ R are monotonously
increasing and positive for positive arguments. Moreover for t --+ 00 we require that
get) --+ 00 and G(t) --+ 00 and that
Before we address the existence of a solution u to (P), we insert the following useful
result for monotone-convex functionals which goes back to Minty [34] and was firstly
formulated in [35].
Proof. (P) => (2.3) is immediate from the monotonicity. To prove (2.3) => (P) fix
v E K. Then, for any s E [0,1), w. = v + s(u - v) E K, hence
cp(W., 11) ~ >.(u - w.) = (1 - s)>.(11 - v).
hence
SA(V - u) ~ Cf'(w.,v).
Thus by hemicontinuity, we arrive in the limit S --+ 1 at (P). o
Since we do not presuppose that K is bounded we use the concept of the asymptotic
cone of K, due to Fenchel [15], which is defined for some fixed Vo E K as follows:
In this section we present our abstract stability analysis. Thus in addition to the
real valued functional Cf' defined on K x K we are given a sequence {K" }"EN of nonvoid
convex closed subsets of V and a sequence {Cf'v}VEN of real valued functionals defined
on K" x K". We do not suppose that all the K" are contained in K. Instead we follow
the concept of Mosco [36] (see also [44,13]) of set convergence ("Mosco convergence")
introducing the subsequent two hypotheses:
(HI) If {V"}"EN weakly converges to v and for all". E IN, v" lies in K"" for a subsequence
{Kv"},,EN of the sequence {K,,}vEN, then the weak limit v belongs to K.
(H2) For any v E K there exists a sequence {VV}"EN strongly converging to v such
that v" lies in Kv for all large v.
These two hypotheses can be shortly phrased as
where in the sense of Kuratowski (see e.g. [2,29]) w-limsup, respectively s-liminf de-
notes the limes superior with respect to weak convergence, respectively the limes infe-
rior with respect to strong convergence. Note that only for simplicity we work with
sequences, all the analysis to follow carries through also with nets or filters.
Let us remark that Stummel [47] obviously independently developed this concept of
set convergence in his study of perturbations of linear elliptic boundary value problems
in Sobolev spaces. He notes in [47, p. 11] that these two hypotheses are equivalent to
w-limsupK" S;; K S;; w-liminfK", s-lim sup K" S;; K S;; s-lim inf Kv ,
"-+00 11-+00 11-+00
(H2') There exist a dense subset D of K and for all II E IN mappings r" : D -+ V with
the property that, for each v E D, r,,(v) -+ v (II -+ 00) and rAv) E K" for all
II ~ 1I0( v) for some 1I0( v) > o.
The emphasis of (H2') lies upon the fact that when applied to finite element approx-
imation, the mappings r ll can be explicitly stated by interpolation only on a dense
subset, given by smooth enough functions. Since we are here not interested in dis-
cretization analysis, we stick to the original Mosco convergence. On the other hand,
Stummel [47, p. 11,12] (see also [46, section 4.1(9) p.78]) proved:
The relation K ~ s-liminf K" holds, if and only if there exists a dense
subset D ~ K with the property D ~ s-liminf K".
In addition to (HI) and (H2) it is important for our stability analysis that the inter-
section K n n"EN K" should be nonvoid. In view of our applications it is not hard to
require that there is even some ii E K n Y such that ii E K., and (2.2) is uniformly
satisfied in II E IN, that is there exist some constants bo ~ 0, b1 ~ 0 such that
(3.1)
Also concerning the sequence {'>I',,}"EN, we now require that the assumptions (A2) and
(A3) hold uniformly in /I E IN. In order to approximate '>I' by {'>I',,} , we need the
following hypothesis:
This hypothesis corresponds to the assumption (Ass I) with Mosco; more precisely in
the special setting '>1'.,( v"' w,,) = (T,,( VII)' wlI - VII)' '>1'( v, w) = (T( v), W- v) for single val-
ued monotone operators Til, T (Ass I) requires the norm convergence of Til ( VII) to T( v)
and hence implies (H3). Moreover in the case Cf'1I(V." w,,) = ,p1l(W II ) - ,p1l(VII ), Cf'(v, w) =
,p(w) - ,p(v) for convex functionals ,pll',p Mosco convergence ,pll -+ ,p [13, Chapter IV,
Section 8], [48] implies (H3). We point out that our stability theory does not need
a uniform equimonotonicity assumption like (Ass III) with [36] or an equicoercivity
assumption like (Ass IV) with [36].
Altogether instead of the problem (P) we consider the perturbed problem of finding
(P,,)
By the existence theory discussed in the previous section, also solvability of the per-
turbed problems (P,,) can be guaranteed.
Now we can present our basic stability result extending Theorem A in the classical
paper of Mosco [36].
129
Theorem 3 Let the seminorm 1·1 and "P,>',K satisfy (AI) and (AS). Let the func-
tionals "Pv satisfy (A2) and (A9) uniformly with respect to v E IN together with (A4)·
Suppose that the sets K, Kv and the functionals "P, "Pv satisfy (HI)-(H9). Then there
exists a subsequence {u v.} weakly converging to u, where uv• is a solution of (Pv.)
and u is a solution of (P). If (A9) holds with 10 = II = 0, then in addition we have
lim lu v - ul = O. If moreover the solution u of (P) is unique, then v_oo
1(_00 K
lim lIu v - ull = 0
holds.
Proof. We divide the proof in five parts. We first show a priori estimates for {u v },
before we can establish the convergence results. In the following co, CI, • .• are generic
positive constants.
(1) An I· I-estimate for {u v }. Fix Vo E K, according to (H2) let {v v } be an
lim IIv v - voll = 0 and Vv E Kv for v > vo = vo(vo).
approximating sequence to Vo ,i.e. v_oo
Then with U v, a solution of (P v), we have
Since Vv converges to vo, >.(vv) and Ivv - iii are bounded. By (A3) we can estimate
Xv := U v - ii by
(3.4)
We claim that IZn I --+ O. Assume not. Then for a subsequence IZn.1 ;::: c > 0, hence
:= Ixn.1 = IIxn.lllzn.1 --+ 00. By the continuity of I . I
tk
This leads by (A4) to a contradiction, proving our claim. Thus we have IZnl--+ 0, Ilznll =
1 and Zn weakly convergent to £. By (AI), we can extract a subsequence, again denoted
by {zn}, that converges strongly to £. It follows that 11£11 = 1, in particular £ =I O.
Now we claim that £ E ac (Y n K) = Y n acK. Since by continuity of 1·1, £ E Y
and since ii E K, we only have to show that, for any fixed s > 0, ii + s£ belongs to K.
130
Because of JlunJi -+ +00, there exists an index nI such that Jlu n - YII > s for n ~ nI.
On the other hand, y E K"n and Un E K"n. By convexity for n ~ nI,
(4) Convergence with respect to 1·1. Here we modify the arguments of Glowinski
[18, p. 11]. Suppose that (A3) holds with /0 = /1 = O. Then as above, let Un solve
(P"n) and let Un converge weakly to u*, a solution of (P). By (A2),
131
Since 'P(u·,·) is convex and lower semicontinuous, hence [25, §4.3, p. 153] weakly lower
semicontinuous and 'P( u·, u*) = 0,
limsup [-'P(u*,u n )] ~
n_oo
o.
Now fix Vo E K, according to (H2) let {v~} be an approximating sequence to Vo. Then
Vn= V~n converges strongly to Vo and Un - Vn converges weakly to U* - Vo. Hence
lim sup
n--+oo
IVn - u*1 ~ lim
n--+oo
IVn - vol + Ivo - u*1 = Ivo - u*l·
Moreover, G(luni) is bounded above, say by c* > O. therefore (3.6) entails in the limit,
for any Vo E K,
o ~ lim sup
n_oo
IU n - u*1 g(lu n - u*i) ~ c*lvo - u*1 + A(U* - vo).
Moreover the transportation from i to j gives rise to unit costs 7rij. Further we associate
with each supply market i a supply price Pi and with each demand market j a demand
price qj. As in the recent model of Nagurney and Zhao, we assume there is given a
fixed minimum supply price Pi ~ 0 ('price floor') for each supply market i and also
a fixed maximum demand price qj > 0 ('price ceiling') for each demand market j.
These bounds can be absent and the standard traffic network equilibrium model due
to Dafermos ([9], [10]) results, where the markets are required to be cleared, i.e.
Si =0 for i = 1, ... , nj tj =0 for j = 1, ... , m
are required to hold. Since Si ~ 0 and tj ~ 0 are admitted, the model is also termed a
disequilibrium model. In addition to the model of [37] we also include upper bounds
Xij > 0 for the transportation fluxes Xij.
Assuming perfect equilibrium the economic market conditions take the following
form
Proof. Let u = (p,q,x) E M satisfy the market equilibrium conditions (4.1) - (4.5)
with
Si = ij;(p) - I>ij ~ 0, tj = /j(q) - I>ij ~ O.
i
If Pi= p., then Pi - Pi ~ 0 and the product in the first sum in (4.7) is nonnegative,
otherwi;~ by (4.3) Si = 0 holds and the product vanishs. By similar case distinctions
one obtains that each product in the second sum is nonpositive and each product in
the third sum is nonnegative. This proves (4.7).
Conversely let (4.7) hold. By the choices Pk = Pk, ifl = ql, Xkl = X kl for all indices k
except some fixed index i and for all indices 1 except some fixed index j one obtains
that in (4.7) all products in the first and in the third sum are nonnegative, whereas all
products in the second sum are nonpositive.
By the choice Pi> Pi (4.1) follows with Si ~ O. To verify (4.3) suppose for some i
that Si > 0 and Pi > p. both hold. Then the choice Pi = p. leads to a contradiction.
Now turn to the third sum. If Xij = 0, choose Xij = -iij > OJ if Xij = Xij, choose
Xij = OJ if Xij E (0, Xij) both choices are possible. This shows (4.5).
Let us consider the second sum. If qj E (0, qj]' choose ifj = O. Then obviously tj
given by (4.2) is nonnegative. Thus the case qj = 0 remains. Suppose tj < O. Then
n
LXij>f;=h(q)~O
i=l
by assumption. Therefore there exists some index i E {l, ... ,n}, such that Xij > O.
The already proved condition (4.5) entails
Pi + 7f'ij ~ qj =0.
Hence in virtue of Pi ~ p. ~ 0 we obtain 7f'ij = *ij(X) ~ 0, what contradicts that by
assumption *ij(X) > 0 h~lds because of Xij > O.
To verify finally (4.4) apply an analogous argument as in the proof of (4.3): tj > 0
and qj < qj cannot hold simultaneously. In view of the just proved bound tj ~ 0 this
proves (4.4). 0
The Variational Inequality (4.7) leads to the following functional
holds for all pI, p2 E R~. Also we require that 9 (but neither - / nor !) is Lipschitz *'
continuous with some constant 7]2 > 0 and an exponent f3 in the sense that
134
where we suppose ~v) > 0, xli) > o. Then jj = y = 0 E MCv) for all v E IN.
This leads to the perturbed variational problem: Find u Cv ) = (pCI') , qCv), xCv» E M(v)
such that
n m m n
L(Yi(pCv» - L xlj»(p. - plv» - L(/i(qCv» - L xli»(q; _ q~v»
.=1 ;=1 ;=1 .=1
n m
+L L(plv) + *i;(X Cv » - qJv»(x.j - xli» ~ 0, (4.8)
.=1 ;=1
Then we have the following result.
Corollary 5 Let cp, M and the sequence {MCv)}vEN be given as above in this section.
Suppose that the bounds q~v), xli) converge to qj, Xij respectively for all i = 1, ... ,n; j =
1, ... ,m. Then there exist.! a .!ub.!equence {uCvk)hEN that converge.! to a solution u of
(4.7). If the solution u of (4.7) is unique, then 1'_00
lim lIu Cv ) - ull = 0 holds.
Proof. By the assumed convergence of the bounds, the sets MCv) clearly Mosco-
converge to M. Therefore the claimed convergence of a subsequence is a direct con-
sequence of Theorem 3. In case of uniqueness, the convergence of the entire sequence
follows from an obvious contradiction argument. 0
To conclude this section let us remark that is not hard to extend this Corollary
5 to include perturbations of the mappings g, i, *.
One considers mappings gCv) that
converge to g in the sense that pCI') ~ p implies gCv)(pCv» ~ g(p) as v ~ 00; likewise for
i Cv ), *Cv). According to Theorem 3 it is enough to assume that the mappings gCv) satisfy
the same uniform monotonicity condition and the same Lipschitz continuity condition
as y, whereas the mappings -1<1'), *Cv) are allowed to be arbitarily monotone. - On
the other hand, if not all demand prices qj or fluxes Xi; are bounded above, then the
seminorm 1·1 has to be modified and also the mappings - i, *
have to satisfy appropriate
uniform monotonicity and Lipschitz continuity conditions. The details are omitted.
invoking Green's formula these problems can be seen to be the weak formulation of
a boundary value problem involving the p-harmonic differential operator as a typical
nonlinear monotone elliptic operator and the unilateral Signorini boundary condition,
respectively nonclassical boundary conditions describing given friction on the boundary
[14]. Thus both elliptic Variational Inequalities can be considered as scalar models of
the stationary unilateral contact problem [24] in solid Mechanics without and with
given friction, respectively, where a certain nonlinear behaviour of the material [39] is
admitted.
First we introduce the p-harmonic Signorini boundary value problem and collect
some technical tools. Then we apply our abstract stability theory to its variational
formulation, a semi coercive elliptic Variational Inequality, and investigate its stabil-
ity under perturbations of the coefficients of the differential operator and also of the
boundary obstacles. Finally we discuss a related Variational Inequality of the second
case and apply the abstract stability theory to perturbations in the coefficient in the
differential operator only.
u~x,
au
0liVullP-2an ~h,
au
(u-x)(0IiVuIlP-2an -h)=O onr,
In virtue of Sobolev's embedding [38, Theorem 3.8, p. 72] we have Wl,p(Q) C C(Q}.
Therefore the restriction of u on r is pointwise defined in the definition of K and
accordingly we assume X E C(r). To render -X a continuous linear form it suffices to
assume that f E L2(Q) and h E Ll(r). Finally to have the nonlinear form a well
defined we assume that the coefficient function 0 E LOO(Q) with 0 ~ 00 a.e. on Q for
some constant 00 > O.
As a continuous seminorm on V we have here
it follows
IIvlh,p :5 Ivh,p + IIvllo,p :5 2l1 vlh,p .
Hence and by the compact imbedding WI,p(n) c LP(n) ([38, §6.1], [1, Theorem 6.2])
Lemma 1 (ii) entails that the assumption (AI) is satisfied.
Referring to the proof of [8, (5.3.20), Theorem 5.3.3] we have a positive constant
CI such that for all u, v E V
-{'fl(u,v)+'fl(v,u)} = a(v,v-u)-a(u,v-u)
~ cI(ao)lv - ulr,p ,
and (A2) is satisfied with get) = cl t p - 1 • Since moreover with some positive constant C2
where TIo is the subspace of order zero polynomials, i.e. of constant functions on
n. Since TIo is isomorphic to R, hence finite dimensional, TIo is a closed subspace in
WI,p(n). Denoting the orthogonal subspace with respect to the duality by ..L. and using
the reflexivity of Wl,p we obtain
TIo = [X n T-I(O)].L.L = {cl [X.L + T-I(O).L]}.L j
since X.L = {OJ by density of X in Wl,p (e.g. [17], Theorem 7.9) finally
TIo = T-I(O).L.L = T-I(O) = Y
is obtained. o
In virtue of this lemma one can prove that the quotient norm of the quotient space
V/Y is equivalent to I . h,p, hence also Lemma 1 (i) guarantees that the assumption
(A 1) is satisfied.
Now we choose y as a large enough positive constant function. Then y E K n Y
with 'fl(Y,·) == o. Also by the preceding Lemma 6, A = Al satisfies (A5), if and only if
the following condition
Lemma 7 Suppose, (*) hold3. Then there exists (at most) one solution to the Varia-
tional Inequality (P), where <p, A, K are given by (5.1), (5.2), (5.9), respectively.
Proof. Suppose, there exist two distinct solutions '1£1 and '1£2 to (P). Then (A2) implies
lUI - '1£21 g(IUl - '1£21) ~ -<p( '1£1, '1£2) - <p( '1£2, ud
~ -A(U2 - '1£1) - A(UI - '1£2) = O.
Therefore '1£1 - '1£2 E Y, and by Lemma 6, U2(X) = Ul(X) + r (for a.e. x E n) with some
r E R\{O}. Since by monotonicity, the two solutions '1£1 and '1£2 fulfil
holds, we obtain
A(U2 - ud = r{fo fdx + 1r hds} = 0,
contradicting (*). o
For our announced stability result we do not only make use of uniqueness, but to
prove Mosco convergence in the unilateral constraints we need also a density result,
already given in the Hilbert space case by Glowinski [18, Lemma 4.2], now to be
presented in the general case 1 ~ s ~ 00. For that latter result we exploit the lattice
structure of W = w 1 .S(n). Namely, since n is supposed to be a Lipschitz domain,
w 1·"(n) is a Dirichlet space ([3, Theorem 5.23], [28, Corollary A.6]) in the following
sense: Let 8 : IR -+ R be a uniformly Lipschitz function such that the derivative 8'
exists except at finitely many points and that 8(0) = OJ then the induced map 8* on
W given by w E W -+ 80 w is a continuous map into W. In particular, the map
w E W -+ w+ = max(O,w) = Hw + Iwl) is a continuous map into W. Note that
w~'"(n) = C8"(n) is also a Dirichlet space without imposing a regularity assumption
on r.
Lemma 8 Let W = w 1·"(n) (1 ~ s ~ (0). Under the above assumptions on n,
k:={wEW:w~O a.e.onn}
satisfies
clw[k n C""(fi)] =k .
Proof. Since n is a Lipschitz domain, for any v E W there exists an extension ([38,
3.7, 3.8]) v E W 1 '"(RN ) = C8"(IRN) (here N = 2) such that vln = v. Since Wl'"(R N)
is a Dirichlet space, Ivl is also an extension of any v E k and coincides with v on n.
Hence for any v E k, we can assume the extension v ~ 0 on RN. Now using Friedrich's
regularization, we introduce
JR Nw(y)dy = 1, w ~ 0.
where Xv E C(r) and a v E LOO(11) with a v ~ 0'0 a.e. on 11. Thus we arrive at the
perturbed Variational Inequality problem (P v). For our convergence analysis we need
extensions X, Xv E V = W 1 ,P(11) of x, Xv' Then we can present the following
Theorem 9 Let a v -+ a in Loo(11) and Xv -+ X in V as v -+ 00. Suppose, the
condition (*) is satisfied. Then the Mlutions U v of the perturbed problems (P v) converge
strongly in Vas v -+ 00 to the solution u of the p-harmonic Signorini problem (P).
Proof. Looking at the preliminaries in the preceding subsection we only have to verify
the hypotheses (HI) - (H3) and apply Theorem 3.
Verification of (HI). Let a sequence {Vn}nEN be weakly convergent to v E V, where
VnE Kiln (Yn E IN). Since V is compactly imbedded in CO(IT) [38, Theorem 3.8J, it
follows for the trace operator, on r that after extracting a subsequence again denoted
by {v n }
(5.4)
Moreover by assumption, using X = ,X, XII = ,XII
(5.5)
lim IIbvn
n--+oo - Xlln ) - (,v - x)lIv"'(r) = O.
Thus ,VII - XII ~ 0 pointwise on r (Yv E IN) implies that for all t/J E CO(r) , t/J ~ 0
irb v - X) t/J ds ~ 0,
hence ,v - X ~ 0 on r or v E K. This proves (HI).
Verification of (H2). We claim that w- E W~'P(11) for any wE K - X. Indeed, since 11
is a Lipschitz domain, by density [38, §1.3J there exist Wj E Coo (IT) (j E IN) such that
W = limj--+oo Wj strongly in W. On the other hand the trace operator, : W -+ Ll(r) is
continuous [38, Theorem 4.6, Corollary 6.31 and by Egorov's theorem L1-convergence
implies pointwise convergence a.e. for some extracted subsequence which is again
denoted by {Wj}. Hence for any fixed c > 0, there exists jo E IN such that for all j ~ jo
we have
-,(Wj)(x) = -w;{x) :::; c (Yx E r\Nc:) ,
where meas(Nc:) < c. Hence
ICPv(vv,wv)-cp(v,w)1 :S lav(vv,wv-vv)-a(vv,wv-vv)1
+ la(vv,wv - vv) - a(v,wv - vv)1
+ la(v, (wv - vv) - (w - v))1
< in I [av - a]IIIVvvl p- Vvv . [Vwv - Vvv]ldx
2
By the estimate [8, (5.3.21),Theorem 5.3.3] the first summand above is bounded by
C2(2cO)Pllav - all oo and the secondsummand is bounded by Ilallooc2(2co)P-lllvv - vlh.p,
hence both convergent to zero. Since (w v - vv) ~ (w - v) and
gives a continuous linear form, the third summand also converges to zero. Therefore
(H3) is satisfied. 0
-div(adIVuIl P - 2Vu) f in n,
au
d(u) := aliVullp-2 an = h on rl ,
Id(u)l:::=;k,d(u)+klul = 0 onr 2 ,
where as above p > 2, IIVull2 = (al u)2 + (~u)2 and a,J,h,k are given data. Now
hE £l(rl ) and k E Loo(r 2) is given such that k(x) ~ ko > 0 a.e. on r 2 . Note that
the latter equation du + klul = 0 in the nonclassical boundary conditions involving the
conormal derivative d( u) is equivalent to
what reflects Coulomb's law of friction [14]. These nonclassical boundary conditions
give rise to the positively homogeneous, sublinear, hence convex functional
tft(v) = Jklvlds.
r2
whereas now simply K = V. Thus again the assumptions (AI) - (A4) are satisfied,
and A = Al satisfies (A5), if and only if the condition (*) is modified to
Corollary 10 Let a" -+ a in Loo(n). Suppose, the condition (*) is satisfied. Then
there exists a subsequence {U"k hEN of the solutions u" of the perturbed problems (P,,)
that converges weakly in V as v -+ 00 to a solution u of the p-harmonic boundary
value problem (P). If (P) has an unique solution, then norm convergence of the entire
sequence holds.
One can conjecture that the abstract stability theory also applies to perturbations
of the convex functional tft by studying the Mosco convergence of the epigraph sets by
means of arguments similar to the proof of Theorem 9.
141
REFERENCES
[I] R.A. Adams, "Sobolev Spaces," Academic Press, New York, 1975.
[2] H. Attouch, "Variational Convergence for Functions and Operators," Pitman, Boston, 1984.
[3] C. Baiocchi and A. Capelo, "Variational and Quasivariational Inequalities," Wiley, Chichester 1984.
[4] C. Baiocchi, F. Gastaldi and F. Tomarelli, "Some existence results on noncoercive variational
inequalities", Ann. Scuola Norm. Sup. Pisa Cl. Sci., Ser. IV 13:617-659, 1986.
[5] E.Blum and W. Oettli, "From optimization and variational inequalities to equilibrium problems",
The Mathematics Student 63:123-145,1994.
[6] L. Boccardo and F. Donati, "Existence and stability results for solutions of some strongly nonlinear
constrained problems", Nonlinear Analysis 5:975-988, 1981.
[7] L. Boccardo and F. Murat, "Nouveaux resultats de convergence dans des problemes unilateraux",
in: "Nonlinear Partial Differential Equations and Their Applications," College de France Seminar,
Vol. II, H. Brezis and J.-L. Lions, eds., Research Notes in Math. No. 60, Pitman, pp.64-85, 1982.
[8] P.G. Ciarlet, "The Finite Element Method for Elliptic Problems," North-Holland, Amsterdam,
1978.
[9] S. Dafermos, "Traffic equilibrium and variational inequalities", Transportation Sc. 14:42-54,1980.
[10] S. Dafermos, "Exchange price equilibria and variational inequalities", Math. Programming 46:391-
402,1990.
[11] S. Dafermos and A. Nagurney, "Sensitivity analysis for the asymmetric network equilibrium prob-
lem", Math. Programming 28:174-184,1984.
[12] P. Doktor and M. Kucera, "Perturbations of variational inequalities and rate of convergence of
solutions", Czech. Math. J. 30(105):426-437, 1980.
[13] A. Dontchev and T. Zolezzi, "Well Posed Optimization Problems," Lecture Notes in Mathematics
1543, Springer, 1993.
[14] G. Duvaut and J.-L. Lions, "Inequalities in Mechanics and Physics," Springer, Berlin, 1976.
[15] W. Fenchel, "Uber konvexe Funktionen mit vorgeschriebenen Niveaumannigfaltigkeiten", Math. Z.
63:496-506, 1956.
[16] G. Fichera: "Boundary value problems of elasticity with unilateral constraints", in: "Handbuch der
Physik - Encyclopedia of Physics, Band VI a/2 Festkorpermechanik II," S. Fliigge, ed., Springer,
Berlin pp. 391-424, 1972.
[17] D. Gilbarg and N.S. Trudinger, "Elliptic Partial Differential Equations of Second Order," Springer,
New York, 1977.
[18] R. Glowinski, "Numerical Methods for Nonlinear Variational Problems," Springer, New York, 1984.
[19] J. Gwinner, "Nichtlineare Variationsungleichungen mit Anwendungen," Ph.D. Thesis, Universitat
Mannheimj Haag & Herchen, Frankfurt, 1978.
[20] J. Gwinner, "Discretization of semicoercive variational inequalities", Aequationes Math. 42:72-79,
1991.
[21] J. Gwinner and E.P. Stephan, "A boundary element procedure for contact problems in plane linear
elastostatics", RAIRO Mathematical Modelling and Numerical Analysis 27:457-480, 1993.
[22] J. Gwinner, A discretization theory for monotone semicoercive problems and finite element con-
vergence for p-harmonic Signorini problems, Z. Angew. Math. Meeh. 74:417-427,1994.
[23] P. Hess, "On semi-coercive nonlinear problems", Indiana Univ. Mat". J. 23:645-654,1974.
[24] I. HlavaCek, J. Haslinger, J. Netas and I. Lovisek, "Solution of Variational Inequalities in Mechan-
ics," Springer, Berlin, 1988.
[25] H.G. Jeggle, "Nichtlineare Funktionalanalysis," Teubner, Stuttgart, 1978.
142
[26) A. Kaplan and R. Tichatschke, "Stable Methods for Ill-Posed Variational Problems," Akademie
Verlag, Berlin, 1994.
[27) N. Kenmochi, "Nonlinear operators of monotone type in reflexive Banach spaces and nonlinear
perturbations", Hiroshima Math. J. 4:229-263, 1974.
[28) D. Kinderlehrer and G. Stampacchia, "An Introduction to Variational Inequalities and Their Ap-
plications," Academic Press, New York, 1980.
[29) P. Kosmol,"Optimierung und Approximation," de Gruyter, Berlin, 1991.
[30) J. Lagnese, "Perturbations in variational inequalities", J. math. Analysis Applic. 55:302-328, 1976.
[31) P.L. Lions, "Two remarks on the convergence of convex functions and monotone operators", Non-
linear Analysis 2:553-562, 1978.
[32) A. Maugeri, "Stability results for variational inequalities and applications to traffic equilibrium
problems" , in: Atti del Convegno Celebrativo dell. Centenario del Circolo Matematico di Palermo,
Supplemento ai Rend. Cire. Mat. Palermo, Serie II, 8, 1985.
[33) L. McLinden, "Stable monotone variational inequalities", Math. Programming 48:303-338, 1990.
[34) G.J. Minty, "Monotone (non linear) operators in Hilbert space". Duke Mati•. J. 29:341-346, 1962.
(35) G.J. Minty and W. Oettli, "Variational inequalities for monotone-convex functionals", Universitat
Mannheim, 1974 (unpublished).
(36) U. Mosco, "Convergence of convex sets and of solutions of variational inequalities", Advances in
Math. 3:510-585, 1969.
(37) N. Nagurney and L. Zhao, "Disequilibrium and variational inequalities", J. Comput. Appl. Math.
33:181-198,1990.
(38) J.Neeas, "Les methodes directes en theorie des equations elliptiques," Academia/Masson,
Prague/Paris, 1967.
(39) P.D. Panagiotopoulos, "Inequality Problems in Mechanics and Applications," Birkhauser, Basel,
1985.
(40) A.A. Pankov, "Discrete approximation of convex sets and the solution of variational inequalities"
(in Russian), Math. Nachrichten 91:7-22, 1979.
(41) P.A. Samuelson, "A spatial price equilibrium and linear programming", Amer. Econom. Rev.,
42:283-303, 1952.
(42) F. Scarpini, "Some nonlinear complementarity systems algoritms and applications to unilateral
boundary-value problems". Ann. Mat. Pura Appl. IV 123:185-202, 1980.
(43) M.J. Smith, "The existence, uniqueness and stability of traffic equilibria", Transportation Res.
13B:295-304, 1979.
[44) Y. Sonntag, "Convergence au Sens de U. Mosco: theorie et applications a I'approximation des
solutions d'im;quations", These, Universite de Provence, 1982.
(45) G. Stampacchia, "Variational Inequalities", in: "Theory and Applications of Monotone Operators,"
A. Ghizetti, ed., Edizione Odersi, Gubbio, pp. 101-192, 1969.
[46) F. Stummel, "Diskrete Konvergenz linearer Operatoren I", Math. Ann. 190:45-92,1970.
(47) F. Stummel, "Perturbation theory for Sobolev spaces", Proc. Royal Soc. Edinburgl. 73A:5-49,
1974/75.
[48) R. Wets, "Convergence of convex functions, variational inequalities and convex optimization prob-
lems", in: "Variational Inequalities and Complementarity Problems," R.W. Cottle, F.Giannessi,
and J.-L. Lions, eds., Wiley, Chichester, pp. 375-403, 1980.
[49) F. Wille, "Galerkins Losungsnaherungen bei monotonen Abbildungen", Math. Z. 127:10-16,1972.
A PRIMAL-DUAL PROXIMAL POINT ALGORITHM
FOR VARIATIONAL INEQUALITY PROBLEMS
Abstract: The proximal point algorithm is an iterative method for finding a zero
of a maximal monotone set-valued mapping, and has served as a fundamental method
for solving Variational Inequality problems and, in particular, convex optimization
problems. In this paper, we propose a new variant of the primal-dual proximal point
algorithm for solving monotone Variational Inequality problems. The proposed method
adopts a Gauss-Seidel-like procedure to solve a subproblem in each iteration, which
updates primal variables and dual variables alternatively. We prove the convergence
of the proposed method and show that our method is particularly effective for Varia-
tional Inequality problems with certain separable structure, such as traffic assignment
problems.
1. INTRODUCTION
where ( " . ) denotes the inner product in R,N. We suppose that X is defined by
(1.1)
143
144
where g(x) = [gl(X),g2(X), ... ,gM(X)]T with gi: RN -+ R being continuously differen-
tiable and convex, hex) = [h 1 (x),h 2(x),···,h p (x)]T with hj : RN -+ JR being affine,
and S denotes a closed polyhedral convex subset of RN .
Over the years, various solution methods for problem VI have been developed [7].
In this paper, we propose a method for solving problem VI based on the proximal
point algorithm [2,14]. Our method turns out to be particularly effective for problem
VI with certain separable structure, such as traffic assignment problems.
The proximal point algorithm is a method for finding a zero of a maximal monotone
set-valued mapping T :
This method is an iterative method such that in iteration k one solves the equation
1
oE T(v) + C(k) (v - v(k»),
where v(k) is the current iterate and c(k) is a positive parameter. The proximal point
algorithm can be applied to problem VI in the following manner. Let ,\ E JRM and
1r E R P be Lagrange multipliers vectors associated with the inequality and equality
constraints in (1.1), respectively, and define the mapping T by
where R~ denotes the nonnegative orthant of RM, Ns(x) is the normal cone to S at
x, and NRM('\) is the normal cone to R~ at ,\. It is known [15,16] that the mapping
+
T is maximal monotone. Also, it is shown in [15,16] that any zero of the mapping T is
a solution of problem VI if the Slater's constraint qualification is satisfied, i.e., there
is an x E X satisfying g(x) < o. Then, the proximal point algorithm for the mapping
T is to solve at each iteration the following system of equations:
1
o E F(x)+Vg(x)'\ + Vh(x)1r + Ns(x) + c(k)(x-x(k»), (1.3)
where A(x,,\,c) and II(x,1r,c) are the vectors which have the following functions as
the components, respectively,
To solve problem VI, Rockafellar [15,16] has proposed an iterative procedure called the
proximal method of multipliers. The proximal method of multipliers is a modification of
the method of multipliers or the augmented Lagrangian method developed for nonlinear
145
Unfortunately the equation (1.10) is not separable even though F, g, hand S are sepa-
rable, since V' g(x) A(x, A(k), elk»~ and V'h(x) II(x, 7f'(k) , elk»~ usually involve cross terms.
This difficulty is common to the method of multipliers in nonlinear optimization, for
which a number of modifications have been proposed in order to restore the separabil-
ity [4,9,11,17). In this paper, bearing in mind the applicability to separable problems,
we propose an approach to solve subproblems (1.3)-(1.5) of the primal-dual proxi-
mal point algorithm in a way different from [15,16). This approach, roughly speaking,
adopts a Gauss-Seidel-like procedure to solve approximately (1.3)-(1.5). In other words,
one first updates x for fixed A and 7f', then one improves A and 7f' by using the new data
of x. This procedure thereby enables us to make the most of a separable structure the
given problem may have.
This paper is organized as follows. In Section 2, we state a new variant of the
primal-dual proximal point algorithm. In Section 3, we prove the convergence of the
proposed method. In Section 4, we apply the method to traffic assignment problems.
Section 5 concludes the paper.
2. ALGORITHM
In this section, we propose a variant of the primal-dual proximal point method for
solving problem VI with the constraints defined by (1.1). The method is in essence the
proximal point algorithm for the mapping T defined by (1.2). Let SP(x, A, 7f'; c) denote
the subproblem to be solved for the current iteration point (X,A,7f') (see (1.3)-(1.5».
Note that the solution (x, X, 1f) of subproblem SP(x, A, 7f'; c) is uniquely determined. In
fact, using the mapping T defined by (1.2), (x, X, 1f) can be expressed as (x, X, 1f) =
(I + eTtl(x, A, 7f'). We first state a procedure, called SSP(x, A, 7f'; e, E), to find an ap-
proximate solution (z, Ti, 17) of subproblem SP(x, A, 7f'; c).
Note that the functions A and II that appear in Step 4 are defined by (1.8) and (1.9)
in the previous section. Using this procedure, a primal-dual proximal point algorithm
for problem VI can be stated as follows. This algorithm uses two sequences of positive
parameters {elk)} and {elk)} such that
Algorithm PDPPA.
Step 1: Choose a starting point (x(O),>.(O),1T(O»;
k:=O.
Step 2: Obtain an approximate solution (x(k+ 1), >.(10+1), 1T(k+1» of subproblem
SP(x(k), >.0:), 1T(k) ; e(k» by executing procedure SSP(x(k), >.(10), 1T(k); elk) , elk) ).
Step 3: If (X(k+1), >.(k+1), 1T(k+1» = (x(k), >.(10), 1T(k», then stop. Else
k:= k + 1;
go to Step 2.
The termination condition (x(k+1), >.(10+1), 1T(k+1» = (x(k), >.(10), 1T(k» in Step 3 guarantees
that x(k+1) is a solution of problem VI. This fact will be proved in the next section.
Now let us consider a Variational Inequality problem with certain separable struc-
ture. Specifically, suppose that
where Fq: RNq --+ RNq are maximal monotone mappings, Yq: RNq --+ RM are mappings
whose components are continuously differentiable convex functions, and Aq and b are
P x N q matrices and an P-vector, respectively. Problem VI with the above structure
will be called a separable Variational Inequality problem (SVI). For problem SVI, the
equation to be solved in Step 2 of procedure SSP(x, >., 1T; e, e) can be decomposed into
(Q + 1) smaller equations as follows:
147
Note that these (Q+ 1) equations can be solved independently of each other. Therefore,
Algorithm PDPPA can take full advantage of the separable structure of problem SVI.
3. CONVERGENCE
In this section, we prove the convergence of the algorithm PDPPA proposed in
the previous section. First we establish a finite termination property of procedure
SSP(x,.x, 71"; c, f).
Let (x, X, 'if) denote the unique exact solution of subproblem SP(x,.x, 71"; c). Then, since
X = A(x,.x, c) and 'if = II(x, 71", c), we have for all z
E S
1
(F(x) +\7g(x)A(x,.x,c) + \7h(x) II(x, 71", c) +-(x-x), z -x)
c
~O.
Substituting z = x in the first inequality and z = z(l+1) in the second inequality, and
then adding the two inequalities, we have
( A(z(l), .x, c), g(x) - g(z(l+1») ) ~ ( A(z(l), .x, c), \7 g(z(l+1»)T (x - z(l+I») ) (3.3)
and
( A(x,.x, c), g(z(l+1») - g(x) ) ~ ( A(x, .x, c), \7 g(xf (z(l+1) - x) ). (3.4)
Since h( x) = Ax - b, we have
(3.5)
148
(3.8)
Since e2 (L; + p(AT A)) < 1 by assumption, the inequality (3.8) implies that {z{l)}
converges to x linearly. Therefore, the stopping criterion in Step 3 of procedure
SSP(x, A, 71'; e, E) must be satisfied in a finite number of iterations. 0
Remark. The inequality (3.8) indicates that the sequence {z(l)} generated by proce-
dure SSP(x, A, 71'; e, E) converges to an exact solution x at a linear rate.
The theorem below justifies the termination condition at Step 3 of algorithm PDPPA.
Theorem 1. Let the assumptions in Proposition 1 be satisfied. Suppose that e{k) are
chosen smaller than (L; + p(AT A)t~ for all k. If algorithm PDPPA terminates at
Step 3, then the last iterate x{k+I) is a solution of problem VI.
(3.9)
Suppose that procedure SSP( x{k), A{k), 71'{k); e{k), E{k») halts at iteration 1. That is to say,
we have
(3.10)
Also we have
(3.11)
149
from Step 1 of procedure SSP(X(k), >.(k) , 1l"(k); C(k), e(k»). It then follows from (3.9)-(3.11)
that
(3.12)
Now let (x, "X, 7f) denote the exact solution of subproblem SP(x(k), >.(k) , 1l"(k)j elk»). By
(3.8), if z(O) =/: x, then
(3.13)
(3.14)
Moreover, by (3.8), (3.13) implies x = z(1). In other words, procedure SSP(x(k), >.(k) , 1l"(k)j
elk), elk») necessarily terminates at the first iteration, i.e., j = O. Since >.(k+I) = fi and
fi = jl(~ = jl(O) , it follows from (3.14) that
(3.15)
This implies that >.(k+ 1) satisfies (1.4) with x = x(k+ 1). In a similar way, we obtain
(3.16)
which implies that 1l"(k+l) satisfies (1.5) with x = x(k+l). Moreover, from Step 2 of
procedure SSP(x(k), >.(k), 1l"(k)j e(k), elk»), we have
(3.17)
Since i = 0 implies z(1) = x(k+I),jl(O) = >.(k+l) and v(O) = 1l"(k+I), it follows from (3.17)
that (X(k+l), >.(k+ 1) , 1l"(k+ 1») satisfies (1.3). Consequently, if (X(k+l), >.(k+ 1) , 1l"(k+I») = (x(k),
>.(k),1l"(k») holds, it follows from (1.3)-(1.5) that (x(k+I), >.(k+I) , 1l"(k+l») is a zero of the
mapping T, i.e., x(k+ 1) is a solution of problem VI. 0
Now we show that algorithm PDPPA converges to a solution of problem VI, when
it generates an infinite sequence.
Theorem 2. Let the assumptions in Proposition 1 be satisfied. Suppose that elk) are
chosen smaller than (L; + p(AT A)t~ Vk. If problem VI has at least one solution,
then the sequence {(x(k),>.(k),1l"(k»)} generated by algorithm PDPPA with an arbitrary
starting point (x(O), >.(0), 1l"(0») converges to a particular solution (x*, >. *, 1l"*) of problem
VI.
(3.18)
150
For simplicity of notation, we shall write (X(k), A(k), 7r(k»), (X(k+l), A(k+l), 7r(k+ 1 »), elk)
and E(k) as (X,A,7r), (X+,A+,7r+), e and E, respectively. Moreover, let z+, z, /-l and v
stand for z(i+l), z(i), /-l(i) and v(i), respectively. Then, by the definition (1.2) of mapping
T, we have
namely,
(X+,A+,7r+)=J!j.(x,A-eg(z+)+eg(z),7r-eh(z+)+eh(z)), (3.23)
where Jf denotes the resolvent of T, i.e., Jf = (I + eT)-l. On the other hand, we have
II (X+,A+,7r+) - JT(X,A,7r) II
= II J!j. (x, A - eg(z+) + eg(z), 7r - eh(z+) + eh(z)) - J!j. (x, A, 7r) II
~e II (0, g(z) - g(z+), h(z) - h(z+)) II
= e( II g(z) - g(z+) 112 + II h(z) - h(z+) 112)t
~ e( L; + p(AT A))t II z+ - z II, (3.24)
where the first inequality follows from the nonexpansiveness of Jf and the last inequal-
ity follows from the Lipschitz continuity of 9 and the linearity of h. From (3.18) and
(3.24), we have
i.e.,
I (x(k+l), A(k+l), 7r(k+l») - Jf) (x(k), A(k), 7r(k») II < e(k)(L; + p(AT A))t f(k)
< f(k),
where the last inequality follows from the assumption on elk). Then, by Theorem 1 in
Rockafellar [14], the sequence {( x(k), A(k), 7r(k»)} converges to a particular solution of
problem VI. 0
151
Sq ={ Xq E RIAl'I I EX q = dq , Xq ~ O}
• Xo C RIAl: set of feasible total flows,
Xo = { Xo ERIAl/ Xo = t
q=l
X q , Xq E Sq, q = 1"" ,Q}
• C: travel cost function from RIAl into itself
We assume that the travel cost function C is maximal monotone. Then we can formu-
late the traffic assignment problem as the Variational Inequality problem:
A solution x~ of problem TA is called a user-optimal flow in the sense that no user can
decrease his/her travel cost by only changing his/her route. Clearly, problem TA can
be reduced to problem SVI with the following identifications:
oE !(Z(l+l) -
c q q . q
+
x ) - 1/(1) Ns (z(I+1» •
5. CONCLUSION
REFERENCES
[1) D.P. Bertsekas, "Constrained Optimization and Lagrange Multiplier Methods", Academic Press,
New York, 1982.
[2) D.P. Bertsekas and J.N. Tsitsiklis, "Parallel and Distributed Computation: Numerical Methods",
Prentice Hall, New Jersey, 1989.
[3) S. Dafermos, "Traffic equilibrium and Variational Inequalities," Transportation Science 14, pp.
42-54, 1980.
[4) M. Fukushima, "Application of the alternating direction method of multipliers to separable convex
programming problems," Computational Optimization and Applications 1, pp. 93-111, 1992.
[5) D. Gabay, "Application of the method of multipliers to Variational Inequalities, " in: M. Fortin
and R. Gorwinski eds., "Augmented Lagrangian Methods: Applications to the Numerical Solutions
of Boundary-Value Problems", North-Holland, Amsterdam, pp. 299-331, 1983.
[6) G.H. Golub and C.F. Van Loan, "Matrix Computations", The Johns Hopkins University Press,
Baltimore, 1983.
[7) P.T. Harker and J.S. Pang, "Finite-dimensional Variational Inequality and nonlinear complemen-
tarity problems: A survey of theory, algorithms and applications," Mathematical Programming 48,
pp. 161-220, 1990.
[8) M. Hestenes, "Multiplier and gradient methods," Journal of Optimization Theory and Applications
4, pp. 303-320, 1969.
[9) S. Ibaraki, M. Fukushima and T. Ibaraki, "Primal-dual proximal point algorithm for linearly con-
strained convex programming problems," Computational Optimization and Applications 1, pp.
207-226, 1992.
[10) T. Ito, M. Fukushima and T. Ibaraki, "An iterative method for Variational Inequalities with ap-
plication to traffic equilibrium problems," Journal of the Operations Research Society of Japan 3,
pp. 82-104, 1988.
[11) J.M. Mulvey and A. Ruszczyriski, "A diagonal quadratic approximation method for large scale
linear programs," Operations Research Leiters 12, pp. 205-215, 1992.
[12) M.J.D. Powell, "A method for nonlinear constraints in minimization problems," in: R. Fletcher
eds., "Optimization", Academic Press, London, pp. 283-298, 1969.
(13) R.T. Rockafellar, "Augmented Lagrange multiplier functions and duality in non-convex program-
ming," SIAM Journal on Control 12, pp. 268-285, 1974.
[14) R.T. Rockafellar, "Monotone operators and the proximal point algorithm," SIAM Journal on Con-
trol and Optimization 14, pp. 877-898, 1976.
[15) R.T. Rockafellar, "Mouotone operators and augmented Lagrangian methods in nonlinear program-
ming," in: O.L. Mangasarian, R.R. Meyer and S.M. Robinson, eds., "Nonlinear Programming 3",
Academic Press, New York, pp. 1-25, 1978.
[16) R.T. Rockafellar, "Lagrange multipliers and Variational Inequalities," in: R.W. Cottle, F. Gian-
nessi, and J .L. Lions eds., "Variational Inequality and Complementarity Problems: Theory and
Applications", Wiley, New York, pp. 330-322, 1980.
[17) P. Tseng, "Application of a splitting algorithm to decomposition in convex programming and
Variational Inequalities," SIAM Journal on Control and Optimization 29, pp. 119-138, 1991.
[18) C. Zhu, "Modified proximal point algorithm for extended linear-quadratic programming," Compu-
tational Optimization and Applications 1, pp. 185-205, 1992.
RELATIONS BETWEEN t-, S-, z-DOMAIN DESCRIPTIONS OF
PERIODICALLY-SWITCHED NETWORKS
Koksal Muhammet
Inonu University
Engineering Faculty
44100 Malatya, Turkiye
e-mail: [email protected]
1. INTRODUCTION
155
156
d
= Akxn,k(t) + L
J..
dt Xn,k(t) Bk,jU(')(t); tn,k-l < t < tn,k (Ua)
i=O
J.+J.+,-6(J.+tl
Xn,1+1(t~,k) = Fkxn,k(t;;,k) + L Gk,iU(i)(t;;,k)+
i=O
(Ub)
i=O
J.
Yn,k(t) = CkXn,k(t) + L Dk,iU(i)(t); t n,k-l < t < tn,k. (Uc)
i=O
These equations are known to be the state, switching, and output equations, respectively;
where n = 0,1,2, ... , k = 1,2, ... , K, and due to periodicity t n+1,k = tn,k + To. Note
that, although at any time the system reduces to a linear time-invariant system, the
whole system is periodically time-varying but still linear. Xn,k is the state vector at
the k-th interval; its dimension and content may vary from one interval to another, but
must be the same after each period To. Yn,k(t) represents the output function yet) in
the k-th interval; its dimension and entries are the same for all intervals. u(t) represents
the forcing (source) function or excitation of the system; it is assumed to be pice-wise
continuous in its Jk-th derivative during the intervals t n ,k-l < t < tn,k; and at the
switching instants tn,k to have left and right derivatives up to Jk + J1+1 - 6(h+d and
hH - 1 orders, respectively. h is the maximum order of the source derivative in the
k-th interval, and 6(J1+1) is zero if J1+1 = 0, if J1+1 ~ 1 it is l. Superscript (i) denotes
the i-th derivative. Finally, Ak, Bk,i, Fk, Gk,i, (h,iCk, Dk,i are constant matrices of
appropriate orders.
Although Eqs. (l.la), (l.lb), (l.lc) are originally derived for electrical networks
containing linear, lumped, bilateral components and periodically operated switches [2],
they may describe any physical system analogous to such networks; and the main concern
in this article is focused on the solution properties of these equations and the transfer
characteristics between the input u(t) and the output yet). Note that no nonlinearity
and time-delay as well as distributed parameters are allowed in the system.
When the input u(t) is an exponential, i.e., u(t) = U ept where U is a complex vector
and p is a complex scalar (frequency), the above equations easily simplify to the new
ones containing no source derivatives; for this case the explicit steady-state and transient
solutions for Xn,k(t) and yet) are already found [3]. With the assumption of piecewise
constant excitation u(t) during each interval t n,k-l < t < t",k, such as the case the
input is fed through a sample (at time instants tn,k-d and hold (for tn,k-l < t < tn,k)
circuit, similar simplifications result; further assumptions which are valid in most of
the switched-capacitor networks are the case of 00 time constants (zero natural system
dynamics Ak = 0), no direct coupling between state vectors and inputs (Bk~ = 0), and
between outputs and inputs (Dk,i = 0) during each interval; on the base of the previous
157
work [1,2,3], this case is considered by Nacaroglu [4] and the first general solution results
in z-domain (discrete frequency domain) are obtained starting from the time domain
equations in (1.1). In a later work, the assumptions Bk,i = 0, Dk,i = 0, and the picewise
constancy of excitation are removed and more general results are obtained in [5].
For the judgment of these assumptions and the material in the sequel, the following
basic knowledge about the periodically time-varying systems must be introduced.
1:
casual systems [6]. Time-varying system function H(s, t) is defined by [7,8]
where s denotes the complex frequency 0'+ jw. The forced response of the relaxed system
to any input u(t) with the Laplace transform U(s) is then given by
yet) 1 /00
= -2' H(s,t)U(s)e·tds = ~Residues of H(s,t)U(s)e 8t at its poles. (1.3)
7rJ -00
For a unit exponential input u(t) = ePt , i.e., U(s) = l/(s - p) where p denotes the real
frequency jw, the last expression yields the following steady-state response due to the
input, or the complete steady-state response if the system is asymptotically stable (all
natural modes die away with time, i.e., no steady-state natural modes exist); in other
words H(p, t) can be defined to be
where Wo = 27r/To, and H/(p)'s are the complex Fourier coefficients [9J. Substitution of
(1.6) into (1.4) yields that
00
yet) = L H/(p)ei(w+/wot). (1.7)
/=-00
158
This equation reveals the important property of the periodically time-varying linear
systems, which states that for a single input frequency w of the input, the output contains
infinitely many frequency components w + lwo of which only one is at the input frequency
appears in the output, i.e., H, = 0 if 1 =1= O. Therefore, Ho(p) is known to be the time-
invariant transfer function of a periodically time-varying linear system. Since the input
frequency component of the output is of concern in most of the applications, Ho(p),
being the average value of H(p, t), and hence through (1.5) the exponential input is very
important to study the behavior of periodically time-varying linear systems.
In many cases a switched-capacitor network is fed through a sample and hold circuit
[10]. This is equivalent to say the input u(t) is piecewise constant. Hence both of the
assumptions made in bi) and bii) are valid practically.
2. SOLUTION
for n = 0,1, .... Associated with the k-th interval (most frequently referred as phase),
these variables are referred to be the k-th input and output phases, respectively. Note
that the value of each variable at the switching instant tn,k-l is defined to be its value
after the switching at this instant; i.e., for e.g., y(t!,k_l) = y(tn,k-l) = Yk(tn,k-t};
however, y(t~,k_l) may not be equal to y(tn,k-t} and it is denoted as Yk-l(t~,k_l)'
With this assumption, the system in (1.1), without any other assumptions, has a unique
solution for the state and output vectors of which the only discontinuities may occur at
the switching instants, except the output may have additional discontinuities wherever
Jk-th derivative of the input is also discontinuous [6].
-+-:-::-:-:
t~,o
+--::-«:-:--+----~-+--~-~-:-))--II------I--
t~,l In,2
I
tn,k-l
I
tn,k
I
tn,k-l tn,k tn+1,l
nTO :
I I : tn~1.0
.: ..
I I
:.. ......_=--(_n~,..~l ) TO
!- 1
I I
~I ~I
I '/;'1 : 't'2 I I "k 't'k :
I I I
~1I2 I
I
I
I
I
I
~1'II3 I I
~I 'Ilk
~
~ ) t
Figure 1. The n-th switching period [tn,o, tn,k) and a) samples of variations of Uk(t) and
Yk(t), b) Uk(t) and iMt).
With the exponential input u(t) = ept where p = (J' + jw is the complex frequency
u(C k) = epr'u(t+ k-l)' and with the piecewise constant input u(t) = u(t; k-l) for
tn,kn~l < t < tn,k ;here u(t;;-,k) = U(t;,k_l)' Equations in (1.1) simplify to '
(2.2b)
J. J.+J.+ 1 -6(Jo+d J.
Bk = LBk,ipi, Gk = L Gk,ipi, Gk = 0, Dk = LDk,ipi, (2.3)
i=O i=O i=O
(2.4)
To derive the transfer characteristics between u and x, consider the solution for the i-th
phase due to j-th phase of the input.
For j = i, starting from x(tn,i-J), using Eq, (2.2a) for k = j, we obtain
160
This, together with the switching equation at tn,j (Eq. 2.2b with k = j) yields
since u(t;) = O. Note that although this equation is derived for the exponential input,
it is valid for the picewise constant input as well; infact if p __ 0, (e PTj - 1)/p __ Tj,
ePTj __ 1 by which the correct results are obtained for the picewise constant input case.
Up to time t;;+I,j_I' u(t) = uAt) is zero in (2.2a), hence x remains constant in each
interval of durations Tj+bTj+2, ... ,Tk,TI,T2, ... Tj-I. Therefore, the only variation in
the state vector x occurs at the switching instants; further, in the switching equations
(2.2b), due to zero input, only the first term becomes effective. Hence,
(2.7)
Finally, using the switching equation at tn+I,j-I and letting u(t;;-+I,j_I) = 0, we obtain
(2.9)
Using the shifted variables for the state vectors (Xn,k(t) = Xn,k(t + 1}k» as well as for
the input as defined before, and remembering that j = i, we can write (2.9) as
where
Faij = Fi-I" .FIFK .. .Fi, (2.11a)
Fcij = Gj - I . (2.11c)
For j > i, by following exactly similar steps, we can obtain (2.10) with
(2.12a)
Fbij = Fi-I .. .FIFK .. .Fj+I [FjBj ePTjp- 1 + Gje PTj + FjGj-I] , (2.12b)
Fcij = O. (2.12c)
(2.13a)
161
(2.13b)
(2.I3c)
In fact, Eq. (2.10) is a difference equation relating the discrete values of Xi and Uj.
Although Equation in (1.1) can be solved both for Xn,k and then Yn,k(t) starting
from the initial vector XO,l(O+) and using the state equation (lola) and the switching
equation (LIb) sequentially, and continuing this process along any desired number of
periods, this is time consuming if only the solution in the n-th period or the steady-state
solution is required.
To find the transient solution (complete solution is meant) in the n-th switching
period, the initial value of each phase is computed as follows: solving (2.2a) with n = 0
and for to,o < t < to,l we obtain
(2.I4a)
Using (2.2b)
(2.14b)
is obtained. Repeating this procedure further for i = 3,4, .. .K, we obtain the following
general formula:
(2.14c)
This formula can be used for i = 1 and 2 as well; care mut be paid that summations can
be taken to be zero and products be taken to be 1 when the upper limits are smaller
than the lower limits.
Using the headed phase variables, (2.14c) can be written as
(2.15)
162
After computing Xi(O), we now return to the difference equation (2.10). The solution of
this constant-coefficient first order difference equation can easily be obtained as
Using this equation, the state at t n,i-l can directly be reached without sequential
treatments of Eqs. (2.2a,b). The contribution of all input phases can be considered by
summing the result in (2.16) over the index j = 1,2, ... , K. This way Xn,k(tn,k-I) =
XA,(nTo) is computed; after each switching instant in the n-th switching period, hence
the solution in the intervals t n ,A,-1 < t < tn,A, can be found by using the integral solution
of (2.2a). Knowing solution for each state vector, the output vector yet) can be computed
from (2.2c) readily.
z-Domain Solution
To deal with the discrete frequency (z-) domain solution, we again return to the
difference equation in (2.10). Using the regular rules for the z-transform, we obtain
Xi(Z) = [I - z-1 Faij] -1 [z-l Fbij + Fcij] UAz) + [I - Z-l Faij] -1 [Xi(O) - FcijUj(O)] ,
(2.17)
where Xi(O) is computed from (2.15), and Uj(z) and Uj(O) are assumed to be known.
When all the phases are present, which is usually the case, Xj(z) is computed by sum-
ming (2.17) over j = 1,2, ... , K since the system is linear; i.e.,
K
Xi(Z) = L [I - Z-l Faij] -1 { [z-1 Fbii + Fcij] Uj(Z) - FcijUj(O) + Xi(O)} (2.18)
j=l
The discrete values of Xi(nTO) and hence Xi(nTO + Tfi) = Xi(t n,i-l) can be computed in
any interval specified by n, for e.g. by the formula
Tij(Z)
-
= [I - Z-1 Faij ] -1 [ Z-1 Fbij + Fcij ] , (2.20a)
(2.20b)
respectively. Transforming Eq. (2.2c) into discrete time domain (t = t n ,k-l), using the
headed (shifted) variables, noting that the j-th input phase is not effective on the i-th
163
output phase through the matrix Dk, and using the z-domain transformed variables, we
obtain
(2.21)
the last is the z-domain transfer function between the j-th input phase and the i-th
output phase.
Since superposition applies due to the linearity of the system, the state and output
responses due to all input phases can be computed from
K
Xi(z) = LTij(Z)Uj(z) (2.23a)
j=1
K
Y;(z) = LHij(Z)Uj(z) (2.23b)
j=1
for i = 1,2, ... , K. These equations can be written in matrix from as well. When this is
done for the output y we obtain
i!IK(Z)
H 2 K(Z)
1[ i!1(Z)(z) 1
U2
(2.24)
HKK(Z) UK(Z)
s-Domain Considerations
In this section, the time-varying system function defined by Eq. (1.2) (with s is
replaced by p) derived by using the z-domain analysis results and using the definition
in Eqs. (1.4,5).
Assuming an exponential input u(t) = ept , Uj(z) can be computed in the following
steps:
164
1'... (z )eP'Ij
Ii(O) = E Residues IJ
Z - eP
1',
0
(2.27)
We assume an asymptotically stable system so that all poles of T;j(z) are within the
unit circle and the residues at these poles will have no contribution on the steady-state
value of Ii(O). Hence, to find the time varying system function, the residues at the pole
z = epTo (p = jw and hence I ejwTo I = 1) is considered only. The result
(2.28)
is equal to X(to,i-1) = XO,i(1]i). Using the state equation (2.2a) for the i-th interval, the
solution in this interval can be expressed as
Using the output equation (2.2c), and the relation (2.22b) between TiiCz) and Hij(z),
Yo,i(t) can be obtained to be
(2.30)
which is valid for 1]i ::; t < 1]i+1. Comparison of this equation with (1.4) yields that for
this interval
- (pI',)
Hi,j (p, t ) = [Hij e 0 - Oij (CiBj
-p- + Dj )] ep('I·-
J
t)
+ Oij (Ci
-p-Bj + Dj ) . (2.31 )
When the effect of all input phases are considered, the time domain steady-state so-
lutions can be obtained by summing Eqs. (2.29,2.30) over the index j (superposition).
The results are
(2.32)
which are valid for 1]i ::; t < 1]i+1, i = 1,2, ... , K. With yet) = H(p, t)e pt and H(p, t +
nTo) = H(p,t), Eq. (2.33) defines yet) for all t 2:: O. Similarly, with x(t) = L(p,t)e pt
and L(p, t + nTo) = L(p, t), where
(2.34)
165
for '1/i ::; t < '1/i+l, i = 1,2, ... , K, x(t) can be computed for all t ~ o.
Since the transfer function is the average value of the time-varying system function,
averaging H(p, t) over (0, To), it is obtained from Eq. (2.33) to be
Ho(p) = _1_ LK [ K
e- P'l;(l- e- pr;) LHij(ePTO)eP'li+
pTo ;=1 j=1
(2.35)
+(e- pr; + pT; _ 1) (C:i + D) ].
We close this section by noting that the obtained results are valid for multi variable
systems having multiple input and/or output variables. In this case H(p, t) and hence
Ho(p) will be matrices which can be called, for e.g., time-varying system matrix and
transfer matrix, respectively. Similarly, each element of the generalized phase transfer
matrix in Eq. (2.24) will be a transfer matrix as well.
Another point following from the linearity is that when the input e pt is replaced by
U e Pt , where U is a complex scalar or vector, the responses for the state and output vectors
are still valid with ept is replaced by Ue pt , i.e., x(t) = L(p,t)Ue pt , yet) = H(p,t)Ue pt •
i) If the input u(t) is of a slow-varying type with respect to the periodicity of the
system, which is possible if the highest frequency component of the input is much smaller
than the switching frequency l/To , then it can be assumed that
The last sum is the row sum of the generalized phase transfer matrix and it is defined
as the transfer function between the input and the i-th phase of the output. In many
of the first order switched-capacitor filters [12,13] and second order SC biquads [14,15]
H;(z) is the same for all output phases, i.e., Hl(Z) = H2 (z) = ... = HJ(z), therefore it
is usual to define the z-domain transfer function of the filter as
J(
H(z) = H;(z) = LH;j(z) . (3.3)
j=l
ii) In many applications some output phases are effected by the corresponding input
phases only [16,17]' and/or the relation between the same input and output phases is of
concern; in this case H (z) is given by
(3.4)
iii) Sometime by using selector switches only, one input phase is made effective at
the output and only one of the output phases are selected; in this case
(3.5)
i.e., any entry of the GPTM is a transfer function of the whole circuit.
4. CONCLUSIONS
In this presentation a tutorial theory about the solution of piecewise constant but
periodically time-varying linear systems with zero natural dynamics is presented. In
addition to the state equation, switching equations are used to determine the discon-
tinuities between adjacent constant states. The theory finds its basic applications in
switched-capacitor networks which are widely used in integrated electronics in the last
decade due to their small chip area, environmental and constructional insensitivities,
and externally adjustable properties.
The solutions results are discussed in the time, complex frequency and discrete
frequency domains; the results especially relating the z-domain transfer functions to the
s-domain transfer functions, as far as the author's knowledge, are believed to be original
and carries much importance. When the system dynamics exist, which is inevitable when
the nonidealities are to be considered in physical systems, the presented analysis gets
much more difficult; although some exact solution results appear in the literature in
t- and s-domains [3], z-domain treatment similar to the one presented in this article
deserves to be theoretical importance.
Another particular interest would be the study of solution properties with zero
dynamics when the input is or can be assumed to be piecewise linear between switching
167
instants; which is a case that will naturally yield more accurate results than piecewise
constant approximation and better research pleasure.
It is believed that applications in the industry and real life would not be of much
interest for the mathematicians; therefore electronic circuits applying the dealt formulae
some of which can be found in the reference list not included in this material.
REFERENCES
[1] Koksal M. and Y. Tokad, "State space formulation of linear circuits containing periodically oper-
ated switches," Circuit Theory and Appl., vol. 5, no. 1, pp. 155-170, Jan. 1977.
[2] Koksal M., "Analysis and applications of linear circuits containing periodically operated switches" ,
Ph. D. dissertation, M.E.T.U., Elect. Eng. Dept., Ankara-Thrkey, May 1975.
[3] Koksal M., "On the solution of linear circuits containing periodically operated switches," Proc.
1976 European Conf. Circuit Theory and Design, Genoa, Italy, vol. 1, pp. 77-82, Sept. 7-10,
1970.
[4] Nacaroglu A., "Computer oriented analysis of linear circuits containing periodically operated
switches; Applications to switched capacitor networks", Ph. D. dissertation, M.E.T.U., Elect.-
Electron. Eng. Dept., Gaziantep-Thrkey, Dec. 1989.
[5] Kilci T., "Computer aided investigation of nonideal situations of linear circuits containing peri-
odically operated switches; Applications to switched-capacitor networks," Ph. D. dissertation,
Gaziantep University, Elect.-Electron. Eng. Dept., Gaziantep-Thrkey, Sept. 1993.
[6] Desoer C.A., "Notes for a Second Course on Linear Systems" , Van Nostrand Reinhold, New York,
1970.
[7] Zadeh L.A., "Frequency analysis of variable networks," Pmc. IRE, vol. 38, no. 3, pp. 291-299,
March 1950.
[8] Zadeh L.A., "Time-varying networks I," Pmc. IRE, vol. 49, no. 10, pp. 1488-1503, Oct. 1961.
[9] Pipes L.A. and L.R. Harvill, "Applied Mathematics for Engineers and Physicists", McGraw-Hill
Int. ed., Singapore, 1971.
[10] Farg S.C., Tsividis Y.P. and O. Wing, "SWITCAP: A switched-capacitor network analysis pro-
gram, part I: Basic features," IEEE Circuits and Systems Magazine, pp. 4-9, Sept. 1983.
[11] Lee M.S. "Switched-capacitor filters using floating-inductance simulation circuits," Electron.
Lett., vol. 15, no. 20, pp. 644-645, 27th Sept. 1979.
[12] Clement F.L. and W.K. Jenkins, "Computer-aided analysis of switched-capacitor filters," IEEE
Trans. Circuits and systems, vol. CAS-28, no. 7, pp. 681-691, July 1981.
[13] Martin K. and A.S. Sedra, "Effects of the op-amp finite gain and bandwidth on the performance of
switched-capacitor filters," lEE Trans. Circuits and Systems, vol. CAS-28, no. 9, pp. 822-829,
Aug. 1981.
[14] Bermudez J.C.M. and B.B. Bhattachanya, "A systematic procedure for generation and design of
parasitic insensitive SC biquads," lEE Trans. Circuits and Systems, vol. CAS-32, no. 8, pp.
767-783, Aug. 1985.
168
[15] Hegt J.A., "Contributions to switched-capacitor filter synthesis," Ph. D. Thesis, Eindhoven Univ.
of Technology, Jan. 1988.
[16] Rahim C.F., Copeland M.A. and C.H. Chan, "A functional MOS circuit for achieving the bilinear
transformation in switched-capacitor filters," IEEE J. Solid-State Circuits, vol. SC-13, no. 6,
pp. 906-911, Dec. 1978.
[17] Martin K., "Improved circuits for the realization of SC filters, lEE Trans. Circuits and Systems,
vol. CAS-27, no. 4, pp. 237-244, Apr. 1980.
ON SIDE CONSTRAINED MODELS OF TRAFFIC EQUILIBRIA
Division of Optimization
Department of Mathematics
Linkoping Institute of Technology
S-581 83 Linkoping, Sweden
1 e-mail: [email protected]
2 e-mail: [email protected]
Abstract: The basic model of traffic assignment does not capture complex traf-
fic flow relationships such as interactions among vehicles on different road links, joint
capacities on two-way streets, turning priorities in junctions, etc. Traditionally, such
relationships are introduced implicitly through generalizations of the separable travel
cost function of the basic model, and the refined models are usually stated as varia-
tional inequalities. However, this approach has found very little practical use, seem-
ingly mostly because the generalized travel cost functions are not easily chosen and
calibrated.
An alternative strategy for refining the basic model is to capture supplementary
traffic flow relationships explicitly through the introduction of side constraints. This
explicit approach has the advantage of always leading to optimization formulations,
and the immediate physical interpretations of the side constraints introduced may make
it easy to construct, calibrate and apply the resulting models compared to models with
non-separable cost functions.
We consider a traffic equilibrium model with general side constraints and show that
its optimality conditions correspond to a generalization of Wardrop's user equilibrium
principle. Further, we derive a close relationship between models of traffic equilibria
involving side constraints and generalized travel cost functions, respectively. These
results motivate a further study into the art of modelling traffic assignment problems
through the use of explicit side constraints.
169
170
are given, and each link a E A is associated with a positive and strictly increasing
travel cost function ta : ~~I >--+ ~++. The user equilibrium principle of Wardrop (1952)
states that for each origin-destination (O-D) pair (p, q), the routes utilized have equal
and minimal travel costs, that is, at an equilibrium flow, the conditions
[TAP)
[fa
mm T(f) := L Ju ta(s)ds, (1.4)
aEA 0
subject to
where
if route r E Rpq uses link a,
Va E A, Vr E Rpq, Yep, q) E C,
otherwise,
is the link-route incidence matrix, and fa denotes the total flow on link a.
The equilibrium model [TAP) is frequently applied in transportation analysis, and
many algorithms have been developed for its efficient solution. (See, e.g., Patriksson,
1995, for a thorough review of algorithms for [TAP).) The inherent simplicity of the
model, however, makes it inapplicable to more complex traffic problems (e.g., Sender
and Netter, 1970). For instance, it does not capture the interactions between the flows
on intersecting links, or between vehicles of different types. An illustrative example of
a deficiency of the model and its possible consequences is provided by Hearn (1980),
who comments on its property of allowing every road to carry arbitrarily large volumes
of traffic. This deficiency of the model causes that
"the predicted flow on some links will be far lower or far greater than the
traffic engineer knows they should be if all as.'lUmptions of the model are
correct. In practice, the result is that the model predictions are ignored, or,
more often, the user will perturb the components of the model (trip table,
volume delay formulas, etc.) in an attempt to bring the model output more
in line with the anticipated results."
171
In order to avoid such heuristic tampering with components of the model available,
traffic planners must be supplied with analysis tools whose underlying traffic models
are sufficiently general, reliable and accurate.
We distinguish two approaches for improving the model's ability to accurately de-
scribe, reproduce, or predict a real-world traffic situation.
The traditional approach is to capture additional flow relationships through the
introduction of non-separable, and typically also asymmetric, travel cost functions.
The solution of (1.2)-(1.3) can then, however, not be reformulated into an optimization
problem of the form [TAP], due to the non-integrability of the resulting travel cost
function c : lR~1 I-T lR~!. Instead, the Wardrop conditions are formulated as the
Variational Inequality problem of finding an f* E F such that
[VIP]
cu*fu - f*) ~ 0, "lIEF,
where F = {I E lR 1A1 I I satisfies (1.5)-(1. 7)}. The resulting class of models has
been extensively studied from a theoretical and algorithmical point of view (see, e.g.,
Nagurney, 1993, and Patriksson, 1995). Seemingly, the asymmetric models' popularity
is a consequence of their mathematical elegance and nice interpretations rather than
their applicability, since real-world applications seem to be lacking. A major reason
for this is probably the practical difficulty of choosing and calibrating the asymmetric
travel cost functions.
An alternative-but so far little studied-approach to improve the quality of the
basic traffic equilibrium model is to introduce a set of side constraints to model ad-
ditional restrictions on possible flow patterns. Such side constraints could be used to
describe, for instance, the interactions among vehicles in a junction, joint capacities
on two-way streets and links in intersections and roundabouts, requirements that ob-
served flows on some links should be reproduced in the calculated solution, a traffic
control policy, or dynamic aspects. We believe this approach to be appealing from a
practical point of view, since it is certainly easier for the traffic engineer to identify a
suitable set of side constraints-which may have immediate physical interpretations-
than to estimate proper values of parameters in complex travel cost functions. (In the
example provided by Hearn (1980), the proper improvement of the basic model is the
introduction of link capacity constraints corresponding to the engineer's anticipation of
reasonable levels of traffic flow.) The approach to improve a traffic equilibrium model
by introducing general side constraints was first discussed by Larsson and Patriksson
(1994).
Although this alternative approach is more straightforward than that based on
asymmetric cost functions, it has been given comparatively very limited attention.
We present a general side constrained assignment model and investigate its optimality
conditions; these may be interpreted as a generalization of Wardrop's equilibrium Prin-
ciple (1.2)-(1.3) in the sense that an equilibrium holds in terms of generalized travel
costs. Moreover, we show that the side constrained assignment problem may be equiv-
alently solved as a standard equilibrium model with a well-defined adjusted travel cost
function. This result leads to an interesting relationship between side constrained and
asymmetric models of traffic equilibria, which is one motivation for further studying
side constrained models.
172
Let gk : ~~I 1-+ ~, k E K, be convex and continuously differentiable, and define the
side constraints
VkEK.
Here, the index set K may, for instance, consist of the index set of the links, nodes,
routes, or O-D pairs, or any combination of subsets of them. (The constraints are,
without any loss of generality, given as inequalities.)
Consider the general side constrained traffic equilibrium problem
[TAP-SCl
subject to
We assume that the feasible set of [TAP-SC] is non-empty; in cases where some
functions gk are nonlinear, we also assume that a constraint qualification (e.g., Bazaraa
et al., 1993, Ch. 5) holds. The convexity of [TAP-SCl then ensures the existence of an
optimal solution, which is unique in the link flows and characterized by the first-order
optimality conditions. We next show that the optimality conditions of [TAP-SCl give
rise to a Wardrop equilibrium principle in terms of generalized route travel costs.
Proof. Stating the stationary point conditions for the Lagrangean function
°
and suppose that
agk(f) > Va E A, Vk E JC.
afa - ,
Consider any O-D pair (p, q). Let a route r E 'Rpq be defined to be unsaturated if for
all k E IC and all links a E A on route r,
Assume, without any loss of generality, that in a corresponding route flow solution to
[TAP-SC] the first I routes are actually used, and that among these the first mare
unsaturated. Then,
and the unused routes in the O-D pair have generalized route costs that are at least as
large as that of the used routes in the O-D pair.
174
Associated with the generalized route travel costs hence are the generalized link travel
costs
ta(f) ~ ta(fa) + E 13k a~~f) , Va E A. (2.18)
keIC Ja
The interpretations of the optimal Lagrange multipliers and the Conditions (2.6)-
(2.7) depend on the form of the constraint functions gk. For example, in the case of
simple upper bounds on the link flows (K = A and ga(f) = fa - U a, U a E [0, +ooJ,
for each a E A), (2.18) reduces to ta(fa) = ta(fa) + f3a, a E A, and the multipliers f3a
may be associated with the equilibrium queueing delays on the saturated links (i.e.,
those with flows on their respective upper bounds), and the multipliers 1f'pq with the
(minimal) sum of total travel cost and queueing delay in each O-D pair; see Miller et
al. (1975), Payne and Thompson (1975), Smith (1987), and Larsson and Patriksson
(1994). (In the case of capacitated traffic assignment, the constraint functions ga are
nondecreasing, and an unsaturated route contains no saturated links.)
The reader should note that the optimal multipliers 13 are not necessarily unique.
[TAP(f3)]
Theorem 3. (An Everett-type result for [TAP-SCD Let (3 E ~~I be a price vector.
Then the solution f«(3) to the priced-out problem [TAP«(3)] solves the side constrained
traffic assignment problem
[TAP-SC«(3)]
min T(I),
subject to
Proof. The strict convexity of T and the discussion following Theorem 6.5.1 of Bazaraa
et al. (1993) yield that r= f«(3). The result then follows since the link travel cost
mapping of [TAP«(3)] is \1L(·, (3) = t(·) + \1g(.)(3. 0
The side constrained assignment model [TAP-SC] may thus be solved as an equiv-
alent, convex, standard traffic equilibrium problem with an appropriately chosen ad-
justment of the travel costs ta; this problem will be referred to as [TAP-A]. Hence,
the link travel cost Mapping (3.1) provides a precise statement of the influence of the
side constraints on the travel cost perception of the users of the traffic network, and
therefore on their route-choice behaviour.
The Variational Inequality problem corresponding to the first-order optimality con-
ditions of the problem [TAP-A] is to find an rEF such that
VfE F.
176
[TAP-SCD]
max L«(3),
(3>O
where
L«(3) = min
IEF
L(f, (3).
Lagrangean dual problems are typically solved using simple iterative search meth-
ods for (essentially) unconstrained optimization. Within a dual solution procedure for
[TAP-SCD], the result of Theorem 3 may be utilized for monitoring the progress with
respect to the aim of finding a solution to [TAP-SC]. This result also facilitates the
finite termination of the dual algorithm when the solution is near-feasible with respect
to the side constraints. Clearly, near-feasible solutions are often satisfactory consid-
ering the uncertainties in the input data; near-feasibility is also satisfactory when the
side constraints are weak, in the sense that they do not need to be fulfilled exactly.
For a link capacity side constrained equilibrium model, Larsson and Patriksson
(1994) investigate and evaluate an augmented Lagrangean dualization (i.e., nonlinear
pricing) technique for finding optimal values of (3 and establish that it is more efficient
than traditional Lagrangean dualization; moreover, for certain instances of augmented
Lagrangean schemes, the sequence of dual iterates generated can be shown to converge
(at least linearly) although the set of dual solutions is not a singleton in general.
The foundation for the development of models of equilibria based on the inclusion of
explicit side constraints is the hypothesis that the additional flow relationships which
we want to capture when modelling a real-world traffic problem may be well repre-
sented by a set of side constraints. Under this hypothesis, we may establish a close
relationship between the improvements of the basic traffic equilibrium model through
the introduction of explicit side constraints and generalized travel cost functions, re-
spectively:
(i) In the generalized travel cost approach, the proper cost function to be used has
the form (3.1).
(ii) This generalized travel cost function involves gradients of the constraint functions
gk, which are unknown since they are not formulated explicitly in this approach.
177
(iii) The proper values of its travel cost parameters (3 are unknown, since the side
constrained problem is not solved in this approach.
Hence, the strategy of using generalized travel costs may be regarded as implicit,
which explains, at least partially, why equilibrium models with generalized travel cost
functions are difficult to calibrate, and therefore difficult to use in practice.
In contrast, the inclusion of side constraints in the equilibrium model constitutes an
explicit approach, in which the physical interpretations of these constraints facilitate
the calibration of the model. The strategy of extending the basic model with side
constraints also allows a large flexibility in the construction of the model, since the
side constraints may be nonlinear as well as non-separable. Thus, the use of side
constraints is a direct and general approach for improving the basic model.
To summarize, if explicit side constraints are utilized in a refinement of the basic
assignment model, the solution of the resulting model [TAP-SC] automatically pro-
duces the travel cost mapping of an equivalent standard traffic equilibrium model.
Hence, through a process in which one or more side constrained models are solved,
one may derive (i.e., determine the appropriate side constraints) and calibrate (i.e.,
find the proper coefficients (3) adjusted travel cost functions for use in existing trans-
portation analysis tools based on traditional equilibrium models, in order to (indirectly
through the cost functions) take into account the additional model components which
are described by the side constraints. The solution of an (augmented) Lagrangean dual
problem may then be viewed as a means for calibrating these travel cost functions.
The many possibilities for modelling traffic interactions with explicit side con-
straints, and the strong relationships to equilibrium models with generalized travel
costs, motivate the further exploration of this modelling strategy for traffic equilibrium
problems. The successful outcome of this exploration relies on cooperations between
operations researchers and users ot today's transportation planning systems.
ACKNOWLEDGEMENTS
The research leading to this report was sponsored in part by the Swedish Transport
and Communications Research Board (Grant TFB 92-128-63). The report was written
while the second author was on leave at the Department of Mathematics, University
of York, Heslington, York, England. The second author wishes to thank Professor
Michael J. Smith of the University of York for his hospitality and many interesting
discussions, and the Swedish Transport and Communications Research Board (Grant
TFB 93-131-63), the Swedish Institute (Grant 303 GH/MLH), and the Royal Swedish
Academy of Sciences for financing the visit.
REFERENCES
[1) M.S. Bazaraa, H.D. Sherali and C.M. Shetty, "Nonlinear Programming: Theory and Algorithms,"
second ed., John Wiley & Sons, New York, NY, 1993.
(2) M. Beckmann, C.B. McGuire and C.B. Winsten, "Studies in the Economics of Transportation,"
Yale University Press, New Haven, CT, 1956.
(3) S.C. Dafermos, "The traffic assignment problem for multicJass-user transportation networks",
Transportation Science 6:73-87, 1972.
178
[4] D.W. Hearn, "Bounding flows in traffic assignment models", Research Report 80-4, Department of
Industrial and Systems Engineering, University of Florida, Gainesville, FL, 1980.
[5] T. Larsson, and M. Patriksson, "An augmented Lagrangean dual algorithm for link capacity side
constrained traffic assignment problems", in: "Proceedings of the 2nd Meeting of the EURO
Working Group on Urban Traffic and Transportation, Paris, France, September 15-17, 1993,"
vol. 38 of Actes INRETS, F. Boillot, N. Bhouri, and F. Laurent, eds., Institut National de Recherche
sur les Transport et leur Securite (INRETS), Arcuei!, France, pp. 163-199. Also as Report LiTH-
MAT-R-93-22, Department of Mathematics, Linkiiping Institute of Technology, Linkiiping, Sweden,
1993.
[6] L.S. Lasdon, "Optimization Theory for Large Systems," Macmillan, New York, NY, 1970.
[7] S.D. Miller, H,J. Payne, and W.A. Thompson, "An algorithm for traffic assignment on capacity
constrained transportation networks with queues". Paper presented at the Johns Hopkins Confer-
ence on Information Sciences and Systems, The Johns Hopkins University, Baltimore, MD, April
2-4, 1975.
[8] A. Nagurney, "Network Economics: A Variational Inequality Approach," Kluwer Academic Pub-
lishers, Dordrecht, The Netherlands, 1993.
[9] M. Patriksson, "The Traffic Assignment Problem: Models and Methods," VSP, Utrecht, The
Netherlands, 1994.
[10] H,J. Payne, and W.A. Thompson, "Traffic assignment on transportation networks with capacity
constraints and queueing". Paper presented at the 47th National ORSA Meeting/TIMS 1975
North-American Meeting, Chicago, IL, April 30-May 2, 1975.
[11] J.G. Sender, and M. Netter, Equilibre offre-demande et tarification sur un reseau de transport,
Rapport de recherche 3, Departement Economie, Institut de Recherche des Transports, Arcueil,
France, 1970.
[12] M.J. Smith, "Traffic control and traffic assignment in a signal-controlled network with queueing".
Paper presented at the Tenth International Symposium on Transportation and Traffic Theory,
Boston, MA, 1987.
[13] J.G. Wardrop, "Some theoretical aspects of road traffic research", Proceedings of the Institute of
Civil Engineers, Part II, 325-378, 1952.
ADVANTAGES AND DRAWBACKS OF
VARIATIONAL INEQUALITIES FORMULATIONS
Marcotte Patrice
Abstract: In recent years, the Variational Inequality framework has been recognized
as a powerful modelling tool in operations research and economics, whenever equilib-
rium relationships among several agents occur. This interest has prompted researchers
to develop efficient algorithms for solving Variational Inequalities in finite dimensional
spaces. In this presentation we will try to assess the advantages and possible drawbacks
of Variational Inequality formulations, focusing on four problems: oligopoly models,
traffic assignment, bilevel programming, multicriterion equilibrium. Each topic will be
analyzed from the modelling and computational points of view. In particular we will
see whether the Variational Inequality formulations of these models has led to a better
understanding of their inner structure and to efficient solution algorithms.
1. INTRODUCTION
In this paper we consider four equilibrium situations inspired from the realm of
Operations Research and Economics, that can be formulated as Variational Inequal-
ities, and discuss whether this framework is adequate from either the modelling or
computational points of view. Our approach is not encyclopaedic but pragmatic, and
biased towards our own research interests. We basically wish to answer ("raise" would
perhaps be more adequate) two fundamental questions:
1. Have Operations Research (or Economics) models benefited from recent advances
in the field of Variational Inequalities?
179
180
The second question has been motivated by the fact that the Variational Inequality
formulation of practical situations frequently leads to nonmonotone mappings, and
that little research has been devoted to nonmonotone Variational Inequalities. One of
the goals of the present paper is to motivate research in this direction.
Let F be a mapping from the closed, convex set Xc R n into Rn. A vector x* is
a solution of the Variational Inequality V I(F, X) if x* E X and satisfies
where h .) denotes the inner product in Rn. We say that the mapping F is
- monotone on X if
- strictly monotone on X if
- pseudomonotone on X if
(F(y), x - y) ~ 0, VyEX.
1. g(x) ~ 0, Vx E X.
3. If g( x) > 0, then there exists y in X such that the directional derivative g'( Xj y-x)
is negative.
181
The simplest gap function is the primal gap function defined, for compact feasible
sets X, as
9(X) = max(F(x),x - y).
yEX
Based on the dual Variational Inequality, one can also define the dual gap function
g(x) = max(F{y),x
yEX
- y).
If the set X is not compact, one can use the projective gap function introduced by
Fukushima [I]:
1 IIx - YII~,
9B{X) = max(F{x), x - Y)B - -2
yEX a
where (" ')B and II . liB denote, respectively, the scalar product and norm associated
with a symmetric, positive definite matrix B, and a is a positive parameter. Several
other gap functions have been proposed in the literature.
The projective gap function 9B is differentiable but strongly nonconvex and difficult
to evaluate, unless the set X assumes a simple form. The primal gap function 9, on
the other hand, is simple to evaluate, weakly nonconvex (if F is affine and monotone,
9 is convex), but nondifferentiable. The dual gap function is convex, weakly nondif-
ferentiable (it is differentiable if F is affine and strongly monotone), but difficult to
evaluate.
Any gap function 9 can be used to transform V I{ F, X) into the mathematical
program
ming(x).
xEX
3. OLIGOPOLY MODELS
Consider a set of n firms competing to sell a product on a single market. Each firm
i E I = {I, ... ,n} is characterized by a production cost Ci{q;) that depends on its own
output qi, and the demand market by an inverse demand function p that relates the
market price p( Q) to the total output Q = L,iEI qi. A Nash-Cournot equilibrium, if
one exists, is a vector q* that satisfies the conditions
where the "max" operator should be interpreted in the global sense. Standard assump-
tions on the functions Ci and pare:
jel iel
Under assumptions AI-A5, the mapping F is not necessarily monotone. A sufficient
condition for F to be monotone is that the inverse demand function p be convex (Good-
man [2]). If this is the case, traditional algorithms can be implemented. Tiitonnement
processes, for instance the Jacobi scheme, have been suggested by Marcotte [3] and
Harker [4]. They require, at each major iteration k, the resolution of the n single-firm
mathematical programs:
q; E argmax qiP(qi + Lqj) - Ci(qi).
q.~O JEl
,tli
At iteration k + 1 we let qk+ 1 = tqk + (1- t)qk where t (0 < t ::::; 1) is an underrelaxation
parameter. The convergence proof given in [4] requires that the mapping F be strongly
monotone. This condition will be satisfied if p is strongly convex or if the c;'s are
strongly convex. Furthermore, the parameter t must be less than some threshold value
l that depends on the strong monotonicity modulus of F. While an overestimate of t
could lead to nonconvergence of the iterative process, an underestimation could affect
negatively the convergence rate of the method.
Clearly, this is a situation where the variational (or complementarity) formula-
tion hides the salient features of the problem. Computationally efficient algorithms
are more easily designed around the formulation of the oligopoly problem as a one-
dimensional equation. Indeed, if a total equilibrium output Q were known a priori, the
corresponding equilibrium outputs could be easily obtained from the conditions
qi= 0 if P(Q) - c:(O) ~ 0, (3.1)
p(Q) + qiP'(Q) - C:(qi) = 0, otherwise. (3.2)
These conditions are the first-order optimality conditions of the separable concave
program parameterized in Q
max
q>O
p(Q) ~ qi + -21p'(Q) ~ q; - ~Ci(qi) (3.3)
iel iel iel
subject to ~qi=Q.
iel
183
B(Q) = Lqi(Q),
iel
where qi(Q) is the set (interval) of optimal solutions of (3.1)-(3.2). It is shown in
Marcotte [6] that any solution Q is an optimal total output only if 0 E I}i( Q) ~
B(Q) - Q. Conversely, if 0 E I}i(Q), any set of outputs {qi(Q)}.eI that satisfies the
compatibility equation Eiel qi( Q) = Q yields an equilibrium. As in [5), a zero of I}i( Q)
can be located by dichotomic search. Contrary to the function >.( Q), the function B( Q)
is not necessarily monotone decreasing on the entire region R+. However, it has been
shown in (6) that one can restrict the binary search to an interval where the function
B( Q) is decreasing. This interval is [Qrnin, Qmax] where Qrnin is defined as
Qrnin = lI)ax{
.eI
q;run}
q;run = argm~qiP(qi)
q,-
- Ci(qi)
Qmax = ""'
L..Jqimax
iel
qffi&X = argmax
q~O
L0
EiEr qi ""'
p(t)dt - L..JCO(qi).
iel
Such results, intimately related to the problem's structure, cannot be derived from its
Variational Inequality formulation.
On a small five-firm example used in [5) our algorithm, coupled with the false
position method in its last iterations, could determine the equilibrium production
vector to 7 significant digits in 6 iterations. This compares favourably with Murphy's
method (7 iterations) and the Jacobi method (20 iterations). In the particular case
where all functions are affine, a finitely convergent greedy algorithm would be still
more efficient (see Marcotte [3)).
The situation is modified altogether when additional constraints are taken into
account. For instance, Marcotte [7) considered the case of firms shipping their products
to demand markets through a congested transportation network. Let us associate each
firm with an origin node i E I of the network, and each demand market k E K with
a destination node. Denoting by v~ the flow from origin (firm) i on arc a E A of the
network, dj. the quantity sold by firm i on market k, Sa the transportation cost function
on arc a and Pk the inverse demand function of market k, the ith firm payoff function,
to be maximized, is given by
Each firm's flow vector must satisfy its own flow conservation equations. The interac-
tion between firms takes place both on the network, through congestion, and at the de-
mand markets, through competition. Let us introduce the vectors Xi = (v~, di.)aEA,kEK,
x = (Xi)iEI and the set X of feasible x-vectors. Assuming that the payoff functions
are concave, a vector x is a Cournot-Nash equilibrium for the network oligopoly model
if and only if it satisfies the Variational Inequality VI(F,X) where the ith block F;
of the mapping F corresponds to the ith firm negative gradient with respect to its
decision variables: Fi(X) = -Vdi(X) Vi in I. A sufficient condition for VI(F,X) to
be monotone is that Goodman's condition be satisfied, i.e. that the demand functions
Pk be convex and that the functions Sa be concave, a rather unusual assumption for
congestion functions!
The structure of this model is characterized not so much by its underlying network
structure (see next section) than by the shape of the payoff functions. It is convenient
to linearize the functions Sa and Pk at the current iterate, since this yields a linear
Variational Inequality involving a gradient mapping, which can be transformed into a
concave quadratic maximization program. One must be aware that this linearization
scheme is not Newton's method, and that one should not expect the convergence
properties of the latter. A better approximation of the cost mapping F is obtained
through a partial linearization scheme that keeps, a la Jacobi, the nonlinear part of
F; associated with firm i, for all i in I. On the set of test problems solved in [7]
the linearization algorithm was preferable, in terms of accuracy and computer time,
to the nonlinear approximation, Jacobi and Gauss-Seidel algorithms. All algorithms
converged, whenever they did converge, to the same equilibrium solutions which, we
conjecture, might be unique.
Let us finally mention that, as the number of firms grows, the equilibrium solutions
converge to a Wardrop equilibrium on the transportation network (see Haurie and Mar-
cotte [8]). Although Wardrop equilibria are not necessarily Cournot-Nash equilibria,
this result shows that they can be obtained as limits of the above network oligopoly
equilibria.
In this section we only consider the fixed demand, static traffic assignment problem
(TAP in short), although most of the discussion could be extended to the VRl'iable
demand case. The static TAP has long served (still does!) as a benchmark for testing
Variational Inequality algorithms. This is due as well to the practical significance of
the problem and the availability of data, as to its specific decomposable structure.
The traffic assignment problem consists in determining flows on a congested trans-
portation network. Let Pod denote the set of paths joining the origin node E 0 °
to the destination node d E D, h;d the flow on a path P E Pod, god the demand for
°
transportation between nodes and d, and C,,( h) the generalized transportation cost
of traveling along path p, given a path-flow vector h = (h;d)(o,d)EOXD,"EPod ' By defini-
tion (Wardrop), an equilibrium h* is reached when flows are concentrated on shortest
paths with respect to the current path-cost vector C(h*), i.e., h* solves the Variational
Inequality V I(C, H), where H denotes the set of feasible path-flow vectors:
H = {h;dl E h;d = gOd, V(o, d) E 0 X D}.
"EPod
The cost of path P is the sum of the costs along the arcs making the path p. Therefore
185
we have v = Ah and
C(h) = AtF(Ah)
where A denotes the arc-path incidence matrix of the network and F is the arc-cost
mapping.
We will also consider the so-called arc-flow formulation of the TAP where one
searches for a total arc-flow vector v· compatible with a path-flow vector h· in Hand
such that
(F(v·),v· - v) :::; a
for all total arc-flow vector v compatible with some path-flow vector h in H. If the
Jacobian matrix V F( v) is symmetric for all v, then F is the gradient of some function
f, and VI(C,H) can be reduced to the problem of finding stationary points of the
mathematical program
mm f(v)
v,h
subject to v= Ah
hEH.
In this case, the Frank-Wolfe algorithm is the only algorithm able to exploit efficiently
the network structure of the TAP. At each iteration, it solves the linear program
obtained by substituting to f its first-order approximation at the current iterate vi:
mm f(v l ) + (Vf(vl),v-v l )
v,h
subject to v =Ah (4.1)
hEH.
This subproblem can efficiently be solved by computing shortest path trees rooted
at the origin nodes 0 E O. Let wi be the total arc-flow corresponding to one of its
solutions and set
(4.2)
The fact that the multi commodity character of the problem is "hidden" in the shortest
path subproblems makes this method especially attractive. However this algorithm is
difficult to extend to the case where F is not a gradient mapping (see Marcotte [9]
however) and its convergence rate is not good.
Researchers have tried to overcome the slow (typically sublinear) order of conver-
gence of the Frank-Wolfe algorithm by working directly in path-flow space. Since the
number of paths is in general prohibitive, it is convenient to solve, at a given iteration,
a Variational Inequality restricted to the convex hull HR of a restricted number of
extreme points of the set H:
HR = n::: Arhrl L Ar =
rER rER
1, Ar ~ a}.
rER
If the vector AR is not a solution of the original Variational Inequality, then the set R
can be enlarged by incorporating into it the path-flow solution obtained by solving the
linear program
186
Even if no extreme points are dropped, there is no guarantee that the sequence {g(hL)}
be decreasing, nor that the cardinality of R can be limited a priori, although it has
been observed that a very good approximation to an equilibrium solution could be
constructed from the convex combination of less than twenty extreme points. Indeed,
in disaggregated mode, three or four paths per origin-destination pair can be combined
to yield quasi-equilibrium path flows.
There remains to propose a procedure for solving the restricted subproblems. Define
G(>.) = CO::::rER>'rhr). If F is strongly monotone, the mapping C is co-coercive and
the sequence generated by the rule
>.+ = ProjsR(>' - aG(>.»
converges to a solution of the restricted Variational Inequality (see Marcotte and Zhu
[11]) provided that the parameter a is sufficiently small.
In practice, the cost mapping is not strongly monotone, but monotone at best.
Gap-decreasing methods can be adapted to this situation: Newton's direction is a
descent direction for the primal gap function (see Marcotte and Dussault [12]); the
projection
>: = ProjsR(>' - aG(>.»
induces a descent direction X - >. for Fukushima's projective gap, provided that the
parameter a be larger than some threshold value a that can be updated between
iterations (see Marcotte and Zhu [13]).
Since restriction-based algorithms work in path-flow space, a theoretical challenge
consists in devising a limited-memory method operating in arc-flow space. One could
for instance mimic the Frank-Wolfe algorithm and replace the linesearch (4.2) by the
one-dimensional Variational Inequality VI(F, [vl,w l ]), where wi is now the partial v-
solution of the linear program
min f(v l ) + (F(vl),v - vi)
v,h
v =Ah (4.3)
hEH.
If the Jacobian V F is "weakly" asymmetric, this heuristic scheme provides a reasonable
approximation. If V F is strongly asymmetric however, the method could zigzag, or
even cycle. This could be fixed by using a smaller stepsize, for instance:
1
+
VI 1 = Vi + _(wi - vi).
1
This constitutes a generalization of the well-known Brown-Robinson's fictitious play
algorithm (see Robinson [14]) for solving zero-sum matrix games. We conjecture that
it can be used to solve monotone Variational Inequalities as well.
187
Bilevel programming problems are two-stage decision problems where a leader op-
timizes her objective, taking into account the reaction to her course of action of a
follower, whose program (objective and/or constraint set) depend on the leader's de-
cisions. Players do not cooperate although, given two equivalent alternatives, the
follower will choose the one most favourable to the leader. Mathematically we have:
mill
r
F(x, y)
subject to (x,y)EX (5.1)
yEarg min J(x,z).
zEY(r)
Denote by S(x) the set of optimal solutions of the lower level program; then (5.1) can
be reformulated as the one-level program
mIn
r,Y
F(x, y)
subject to (x,y)EX
y E S(x).
This is a difficult problem. Even in the linear case, checking the local optimality of a
solution (x, y) is strongly NP-Hard (Hansen, Jaumard and Savard [16]).
It occurs frequently that the lower level variables are controlled by several, if not
infinitely many, agents or players. For instance, the x-variables might represent energy
taxes set by the government, and y the industrial sector's reaction to those tax levels.
The governments's objective function F could embody financial as well as economic,
social or environmental goals. At the lower level, the industries are price takers, and
seek to maximize their own profit functions. Assuming that they behave according to
the Cournot-Nash principle and that certain conditions are satisfied (see section on
oligopoly problems), their equilibrium state can be represented as a Variational In-
equality. We then obtain the generalized bilevel program where S( x) above is replaced
by the solution set S(x) of a Variational Inequality with cost mapping G, i.e., find y
in Y(x) such that:
(G(x,y),y - z) :S 0, Vz E Y(x).
There are several ways to reduce a generalized bilevel program to a single-level program.
One can for instance express S( x) as the set of points for which the primal gap function
is equal to zero:
mill
r,Y
F(x,y)
188
Obviously, any other gap function could have been used for that purpose. Even under
strong assumptions, the constraint g(x, y) = 0 in (5.2) is nonconvex, and cannot be
expected to satisfy any constraint qualification. By penalizing the nonconvex term,
one obtains the program
min F(x,y) + Kg(x,y). (5.3)
(X,II)eX
Under suitable assumptions (see Marcotte and Zhu [17]), this penalty is exact, i.e.,
there exists a threshold value K such that, for any K ~ K, a global solution of (5.3)
is a global solution of (5.2).
If F is monotone in y, for fixed x, and the set Y(x) is given explicitly as
where the functions h j are convex in y and satisfy some constraint qualification, then
the lower level Variational Inequality can be replaced by its Kuhn-Tucker necessary
and sufficient conditions, and the bilevel program (5.2) written as
mm F(x,y)
(x,y)eX
min
("',!I)eX
F(x,y) +KEujVhj(x,y) =0
jeJ
subject to Uj ~ 0 (5.5)
(Uj, hj(x, y») = 0, j E J.
Under certain conditions, for instance if all functions involved are affine, it can be
shown that a solution of (5.5) is globally optimal if and only if it is also globally
optimal for (5.4) (see Luo, Pang, Ralph and Wu [18]).
If the set Y(x) is a compact polyhedron Y = {By::; b} independent of x, one can
write the bilevel program as
mm F(x,y)
(x,y)ex
(G(x,y),y - yi) ::; 0, Vi E I
where {yihel is the (finite) set of extreme points of Y. Based on this formulation,
Marcotte [19] has suggested the following constraint generation scheme
189
Step 0 E +- 0; i +-1
Step 1 Solve
mm F(x,y) (5.6)
(r,II)EX
subject to (G(x,y),y_yi):50 VieE
Step 3 if (G(x, y), x - yi) :5 f then stop and record the approximate solution
(x, y(x» where y(x) is an optimal lower-level reaction to x.
else go to Step 4 below.
This method can only be efficient if the number of additional constraints in (5.6) is
small and if the nonconvex program (5.6) can be solved to global optimality. Otherwise,
one must resort to heuristic procedures. If the set X only involves the x-variables, one
such procedure, that works well for "weak" bilevel programs where the leader's and
follower's objectives are not too far apart, is to replace the original bilevel program by
the Variational Inequality: Find (x, y) «x, y) eX, ye Y(x» that satisfy
(5.7)
If F is convex in x and G( x, y) = 'V IIf( x, y) for some function f convex in the variable
y, then (x, y) is a Cournot-Nash equilibrium for a game involving two players having
objective functions F and f. Under suitable assumptions, the Variational Inequality
(5.7) is monotone and can be reduced to a convex optimization problem (see Marcotte
and Marquis [20]).
Bilevel programs are computationally intensive. Exact algorithms, based on the
implicit enumeration of the lower-level optimal faces, have only been tested on small
instances involving linear or quadratic objectives and constraints. Even local optimality
results are difficult to obtain (see Hansen, Jaumard and Savard [16] and Gauvin and
Savard [21]) and the theory of sensitivity analysis for Variational Inequalities is of no
great help, as local optima may be numerous.
In view of the intractability of the general problem, the main challenge will be
the development of efficient heuristic procedures able to solve to near-optimality large
instances of bilevel programs. These will have to be adapted to the specific structure
of the problem they intend to solve.
190
6. A BICRITERION PROBLEM
One shortcoming of the traffic assignment model presented in the next-to-Iast sec-
tion is that it yields nonunique equilibrium path flows, even when arc flows are unique.
This can be remedied by considering stochastic versions of the TAP (see Fisk [22]),
where all paths carry some amount of flow. The larger its cost, the lower will be the
amount of flow carried by a given path. These models have shortcomings of their own:
(i) they can only operate in path-flow space and require, a priori, the knowledge of
all paths, of which there are exponentially many (ii) they cannot discriminate well
between almost identical paths, e.g., paths sharing several arcs.
In this section, we consider a bicriterion extension of the TAP that possesses none
of these drawbacks and that can be implemented in arc-flow space, even when the
cost mapping is not a gradient mapping. This extension is formulated in an infinite-
dimensional space, but all steps of the proposed solution algorithm can be implemented
finitely, without resorting to discretization of any kind. Although it remains valid in a
wider context, we will restrict our analysis to the realm of transportation.
Let the generalized cost of traveling along an arc a of a network assume the form
1
Fa(V) = Ca(V) + -Go
a
where Ca(Vo) represents the traversal time of arc a, which depends on the total arc-
flow vector V, Go represents the out-of-pocket cost of travel along arc a, and a is a
parameter converting time units into cost units. The parameter a is referred to as the
"value-of-time" parameter. Following Leurent [23], we assume that the parameter a
varies continuously across the population of network users, and that its density function
h( a) is continuous and strictly positive on [0,00), with h(O) = O. Let us introduce the
notations:
dkh(a), if i = origin of k,
2;: v!(a) - 2;: v!(a) = { -dkh(a), if i =.destination of k,
oE.+ aE.- 0, otherwise.
191
Throughout the section, we will assume that the mapping C is monotone and con-
tinuously differentiable on n. In our assignment model, equilibrium is reached when,
for every possible value of time a, the flow density vector {v:(a)}aEA is assigned to a
shortest path from the origin to the destination node of k, with respect to the value
a, given the congestion levels Ca(V) on the network, e.g.,
(6.1)
for all flow vectors y in n compatible with the feasible total flow vector Y. If the
congestion functions are separable by arc, i.e., Ca(V) only depends on the flow Va, then
(6.1) is equivalent to finding a minimum point of the convex mathematical program
min L
vEO aEA
{l
0
Va
Ca(u)du 1
+ Ga 00 _a_da.
0
v; (a) }
a
This infinite dimensional problem can be solved efficiently using Frank and Wolfe's
linear approximation algorithm, without resorting to numerical discretization or path
enumeration. The convergence of the method follows from a result established by
Auslender and Brodeau [25].
For a current solution vector v and corresponding total flow vector V, the linear
subproblem resulting from the linearization of the objective with respect to the flow
variables is:
mm
tE[O,l)
L 1
aEA 0
Va +t (Ya- Va)
Ca(u)du+Ga
100 -[v,,(a)+t(Ya(a)-Va(a))]da
0
1
a
= rjJ(t) + (1 - t)Q + t R
192
and update
v.,+ = (1 - t*)V" + t*Y;..
It is remarkable that the above update can be realized without explicit reference to
the commodity flow variables, exactly as for the standard traffic assignment problem.
One must only be careful to keep track of the integrals I v" (0. ) / 0. and I Y;. ( 0. ) / 0.. This
is simple; indeed, by linearity:
One can obtain a closed form formula for the latter integrals only if the function h( 0.) / 0.
admits a closed form primitive. If this is not the case, an approximate but closed form
expression can be derived from polynomial quadratures for h(o.). It follows from the
previous discussion that the arc-separable bicriterion traffic assignment problem can
be solved in a similar fashion to the standard problem, with a parametric shortest path
problem being substituted to a standard shortest path problem.
If the delay mapping C is not arc-separable but its Jacobian 'VC is symmetric, the
Variational Inequality (6.1) can still be reduced to a convex optimization problem. If
the Jacobian 'VC is not symmetric, a slightly modified Frank-Wolfe algorithm can be
used to solve it, under the following "nondegeneracy" assumption:
Assumption A: For any two paths PI and Pi joining the origin node to the destina-
tion node of k, one has:
L G"f L G,..
,.ep~ ,.eP:
It is always possible to induce Assumption A through a suitable perturbation of
the cost vector G.
Let vI be the feasible flow vector at iteration 1 and V' the corresponding total flow
vector. Under Assumption A, the functional solution (yl, yl) of the linear program
subject to Y=LY"
"eK
is unique, except for possibly a finite number of values of the state variable 0.. This
remark implies that the primal gap function g( v) is continuously differentiable. More-
over, the extremal solution of (6.2) induces the feasible descent direction d' = yl-v l for
9 at vI. The next iterate V ' +1 can be determined by minimizing 9 along the direction
d' or by setting
where
0.1 ~ 0
lim
1-+00
0.1 =0
00
Lo.l = +00.
1=1
193
Once the solution yl is obtained, all the calculations can be performed with respect to
total flows, and one can forget the functional flows v. The convergence of the method
can be established by adapting a proof of convergence of a gap-decreasing algorithm
for solving Variational Inequalities defined over strongly convex sets (see Auslender
[26)). In both cases the solution of the linear approximation problem (6.2) is unique.
The uniqueness of the correspondence y( v) has other interesting consequences. In
particular, at any equilibrium solution v·, one has that y( v·) = v·. This result is
independent of any monotonicity assumption on the mapping C. Furthermore, given
only the equilibrium arc costs C(V*), a full equilibrium solution can be recovered as
the unique y-solution of (6.2).
The basic model presented in this section can be enhanced to involve flow de-
pendent functions Ga(V) and/or more than two criteria. Unfortunately, the resulting
Variational Inequality formulations are always nonmonotone. It would be interesting to
know whether they satisfy some generalized monotonicity condition, or can nonetheless
be solved efficiently.
REFERENCES
[1] M. Fukushima, "Equivalent differentiable optimization problems and descent methods for asym-
metric Variational Inequalities", Mathematical Programming 53,99-110, 1992.
[2] J. C. Goodman, "A note on existence and uniqueness of equilibrium points for concave N-person
games" ,Econometrica 48, 251, 1980.
[3] P. Marcotte, "Quelques notes et resultats nouveaux sur Ie probleme d'equilibre d'un oligopole",
RAIRO Recherche Opirationnelle 18,147-171,1984.
[4] P. T. Harker, "A Variational Inequality approach for the determination of oligopolistic market
equilibrium", Mathematical Programming 30, 105-111, 1984.
[5] F. H. Murphy, H. Sherali and A. L. Soyster, "A mathematical programming approach for deter-
mining oligopolistic market equilibrium", Mathematical Programming 24,92-106, 1982.
[6] P. Marcotte, "Un algorithme general de calcul de l'etat d'equilibre d'un oligopole", RAIRO
Recherche Opirationnelle 22,225-241, 1988.
[7] P. Marcotte, "Algorithms for the network oligopoly problem" , Journal of the Operational Researcl,
Society 38, 1051-1065,1987.
[8] A. Haurie and P. Marcotte, "A game-theoretic approach to network equilibrium" Mathematical
Programming Study 26, 252-255, 1986.
[9] P. Marcotte, "A new algorithm for solving Variational Inequalities, with application to the traffic
assignment problem", Mathematical Programming 33, 339-351, 1985.
[10] S. Lawphongpanich and D. W. Hearn, "Simplicial decomposition of the asymmetric traffic assign-
ment problem", Transportation Research B 18B, 123-133,1984.
[11] Zhu and Marcotte "Co-coercivity and its role in the convergence of iterative schemes for solving
Variational Inequalities", Publication CRT-926, Centre de recherche sur les transports, Universite
de Montreal, 1993.
[12] P. Marcotte and J.-P. Dussault, "A note on a globally convergent Newton method for solving
monotone Variational Inequalities", Operations Research Letters 6, 1,35--42, 1987.
[13] P. Marcotte and D. Zhu, "Modified descent methods for solving the monotone Variational Inequality
problem", Operations Research Letters 14, 111-120, 1993.
[14] J. Robinson, "An iterative method of solving a game", Annals of Mathematics 154, 296-301, 1951.
194
[15) J. H. Wu, "Dynamic network equilibrium problem formulated as an infinite dimensional Variational
Inequality problem", Publication CRT-952, Centre de recherche sur les transports, Universite de
Montreal, 1993.
[16) P. Hansen, B. Jaumard and G. Savard, "A variable elimination algorithm for bilevellinear pro-
gramming", SIAM Journal on Scientific and Statistical Computing 13, 1992.
[17) P. Marcotte and D. Zhu, "Exact and inexact penalty methods for the generalized bilevel pro-
gramming problem", Publication CRT-920, Centre de recherche sur les transports, Universite de
Montreal, 1993.
[18) Z.-Q. Luo, J .-S. Pang, D. Ralph and S.-Q. Wu, "Exact penalization and stationarity conditions
of mathematical programs with equilibrium constraints", Technical report 275, Communications
Research Laboratory, McMaster University, Hamilton, Canada, 1993.
[19) P. Marcotte, "Network design problem with congestion effects: A case of bilevel programming",
Mathematical Programming 34, 1, 142-162,1986.
[20) P. Marcotte and G. Marquis, "Efficient implementation of heuristics for the continuous network
design problem", Annals o/Operations Research 34,163-176,1992.
[21) J. Gauvin and G. Savard, "The steepest descent direction for the nonlinear bilevel programming
problem" , forthcoming in Operations Research Letters.
[22) C. Fisk, "Some developments in equilibrium traffic assignment", Transportation Researcl. B 14B,
243-255, 1980.
[23] F. Leurent, "Cost versus time equilibrium over a network", European Journal 0/ Operational Re-
search 71, 205-221, 1993.
[24] S. C. Dafermos, "A multicriterion route-mode choice traffic equilibrium model", Unpublished
manuscript, Lefschetz Center for Dynamical Systems, Brown University, 1981.
[25] A. Auslender, and F. Brodeau, "Convergence d'un algorithme de Frank et Wolfe applique 11 un
probleme de controle", RIRO 2, 3-12, 1968.
[26] A. Auslender, "Optimisation: methodes numeriques", Masson, Paris, 1976.
VARIATIONAL AND QUASI-VARIATIONAL INEQUALITIES IN
NETWORK FLOW MODELS. RECENT DEVELOPMENTS IN
THEORY AND ALGORITHMS
Maugeri Antonino
Dipartimento di Matematica
Universita di Catania,
Viale A. Doria, 6
95125 Catania, Italy
e-mail: [email protected]
1. INTRODUCTION
The theory of unilateral problems, which in the simplest case, deals with the
problem of finding a solution of the equation
~u = F in n, (Ll)
satisfying the boundary conditions
uau = 0 on an;
a" (1.2)
au
u;:: 0, a" ;:: 0 on an,
where ~ is the derivative of u along the inner normal direction; this problem was
solved in 1964 by Fichera [1], and this work marked the beginning of the theory, the
Variational Inequalities Theory, that received its best abstract setting for bilinear and
non-symmetric forms by Stampacchia ([2J, 1964), and by Lions and Stampacchia ([3],
1967). The same Stampacchia at the page 191 of the survey [4] writes:
195
196
2. MODELS
In this section I will present the usual model of traffic network, that in a natural
way leads to Quasi-Variational Inequalities. Subsequently, I will report the general
approach suggested by Giannessi ([7J, [15J, and this Volume). As usually, a set N of
nodes is given, say Nb ... , N p , a set A of n directed arcs, say Ab ... ,An and a set Wof
1 ordered points: Origin/Destination, say WI' ... ' Wi. Hence the network is identified
by the 3-triple (N, A, W).
The flow on the arc Ai is denoted by fi' f = (fl, ... , fn). We assume that the
nodes ofthe origin-destination (in short OlD) pair Wj are connected by rj ~ 1 paths,
whose set is denoted by Pj. PI! ... , p" with m = r1 + ... + rl, are all the considered
paths, and F1 , ••• , Fm are the corresponding flows; F:= (FI! ... , Fm).
We assume that the travel cost on arc Ai depends on t, and denote it by Ci(f),
c(f) = (C1(f), ... ,cn(f)). C.(F) will denote the travel cost on the path R. as a function
of the flows of all paths; C(F) = (Cl(f), ... ,Cm(F)). By means of Kronecker numbers:
I, if Ai E R., i = 1, ... , n,
6i • ={
0, if Ai ft R., s = 1, ... ,m,
the flows on arcs can be expressed in terms of flows on paths:
m
Also the cost C.(F) on the path R. can be expressed as a sum of all costs on the arcs
of R.:
n
C.(F) = L: 6i.Ci(~F), (2.2)
i=l
197
or
C(F) = ~TC(~F).
According to the classic Wardrop's definition, a vector H E R+ is said to be an
equilibrium pattern flow iff 'IRq, R. E Pi we have
= Cq(H), if H. > 0,
C.(H) { .
~ Cq(H), 1f H. = 0.
Then, when the equilibrium pattern flow is established, an equilibrium path cost Ci(H)
is associated to each O/D pair Wi:
Ci(H) = R.EPi
min C.(H). (2.4)
Now, if we denote by Pi the travel demands for the OlD pair Wj, in general each
demand depends on the costs for all O/D pairs:
be continuous and strongly monotone. Then the inverse function c) c)(p) exists;
moreover, taking into account the equilibrium conditions:
'IF E K(p*), Vp E R~ ,
where
K(p)={FER+: 'PF=p} PER~. (2.8)
In (9) the authors, taking into account that, as a matter offact, the demands depend
on H, proposed to embed the problem into a Quasi-Variational Scheme.
To this end, let me introduce the Kronecker numbers
1, if R. E Pi
'Pi. ={ and 'P = {'Pi.}, j = 1, ... ,1, s = 1, ... ,m
0, if R. ¢. Pi
198
by means of which the so-called flow conservation law can be written as:
cpF = p(H).
In (9] (see also (10], (11]) it is shown that HE K(H) = {F E R~: cpF = p(H)} is
an equilibrium pattern iff
The paper by De Luca of this Volume provides more information about solvability
of problem (2.9), but it is worth to report the formulation suggested by Giannessi
(see [7]), who remarks that the capacity of arcs must be taken into consideration; as
f;
concerns this question see the paper by Ferrari of this Volume. If we suppose that
every arc flow must satisfy the upper bounds
i = 1, ... ,n,
we have from (2.1)
m
L hirFr ::; 'Yi, i = 1, ... ,n
r=1
or
t:..F::; 'Y
and, therefore, the set of feasible path flows becomes
Hence the Problem (2.8) is solvable if compatibility conditions on K(H) are fulfilled.
(we use the usual notations) where S is a closed, convex and non-empty subset of R,n
and
f: S (or Rn) --+ R n
is a vector valued function with properties to be specified in what follows.
It is well known that, if f is differentiable and the Jacobian matrix
of (Of;)
ax = ax· ',J=l,
.. ...,n (3.2)
J
is symmetric, then the problem (3.1) can be considered as the first differentiable con-
dition for an optimization problem. Then, well-known theoretical and computational
results can be applied.
If (3.2) is not symmetric, i.e. we are out of the usual optimization framework,
it is possible to associate to the Variational Inequality (3.1) a non-negative function,
called gap function, that is zero iff the argument is a solution of (3.1) and such that,
if differentiable, its first optimality condition is verified by a solution of (3.1).
199
defined by
g(x) := max(f(x), x - y). (3.3)
liES
He showed that
1) g(x) ~ 0,
2) x* is a solution of the Variational Inequality (3.1) iff g(x*) = minxEs g(x) = O.
But g(x), in general, is not differentiable, and this has led some authors to construct
more refined gap functions.
Fukushima in [13] presents a new kind of gap function whose stationary point solves
the Variational Inequality (3.1). He considers a n X n symmetric matrix G and, for
any xES, the function
1
F(x) = -(f(x),H(x) - x) - 2(H(x) - x,G(H(x) - x») , (3.4)
with
H(x) := Pr S,G(x - G- 1 f(x»,
where PrjsG(x)
,
denotes the (unique) solution of problem min
liES
lIy - xliG' It is shown
that
F(x) ~ 0,
F(x*) = 0 iff x* is a solution of the Variational Inequality (3.1).
The result is achieved taking into account the fact that a solution of (3.1) verifies
the condition
x* = Pr SG(x* - G- 1 f(x*)) = H(x*), (3.5)
and that F(x) can be rewritten as
Then F(x) = 0 iff H(x) = x. The result is important since the following properties
hold.
Moreover, if ~ is positive definite for all xES, and x* is a stationary point of the
problem
minF(x) = F(x*), (3.7)
:rES
that is,
( a~~*) ,x - x* ) ~ 0, Vx E S, (3.8)
h: R n xS ---+ R,
it results
The gap function provides us also an iterative method to compute the solution of
(3.1). In fact, Fukushima showed the following:
k = 0,1, ...
where dk are given by dk = H(x k ) - xk, and tk E [0,1] are determined by Armijo-type
rule
F(x k + fidk)::; F(x k ) - a,Bllldk Il 2 •
If S is compact, then the mapping f( x) is differentiable and strongly monotone with a
modulus /-I > 0 on S, and f(x), 8~~) are Lipschitz continuous on S, then the generated
sequence {xk} lies in S and it converges to the unique solution of the Variational
Inequality (3.1) for any starting point XO E S, if the positive constant a of the Armijo-
type rule is chosen sufficiently small so that a < /-I.
(3.10)
(3.11)
-min{U(x),H(x) - x)
yES
+ !l(H(x),x)},
it is clear that H (x) is well-defined since h is a sum of a linear and a strongly convex
function.
The authors showed the following result:
where
K(y):= {x E X(y): g(y;x) E C} ( 4.1)
and where C is a closed, convex cone with apex at the origin, X(y) is a subset of a
Hilbert space 3, and
g: X(y) x X(x) __ Rm.
This form includes (3.1) and the approach proposed by Giannessi starts with the remark
that y is a solution of (4.1) iff the system (in the unknown x)
is not solvable. The space where (u, v) runs is the image space associated to (4.1) and
the set
is the image of (3.1). Setting 11.:= {(u,v) E R XRffl: u > 0, v E C}, the unsolvability
of system (4.2) is equivalent to 11. n K(y) = 0. Then, considering the particular case
let us set
where
= II n.,
ffl
Wi E n, n
.=1
and 'Y, n are such that Vy E X(y)
(4.3)
(weak separation).
Then the function
(strong separation)
will be proved to be a gap function and will be called strong gap function.
203
WV, if v ~ 0
~(y,v,w)= {
+00, if v < 0
and
g(y,x) = (x - y, G(x - y»)
(G is positive definite square matrix of order n), or
g(y,x) = Q(x,y)
HE K: r(H)(P - H) ~ 0, VP E K,
vPI, p 2 E K, (5.1 )
a being a positive constant. We remember that an equivalent condition for (5.1) holds
is that the Jacobian matrix ;~ is positive definite VP E K.
The construction of a sequence {Pi} c K that converges to the solution of the
Variational Inequality is developed in the following way:
- firstly, one fixes an arbitrary po E K, a symmetric positive definite matrix G, a
positive parameter 0 < p < ~, where
v = max
K {maxAi' l<i<m:
- - Ai is an eigenvalue of (u~pr )T G- 1(~pr)
u }., (5.2)
(5.3)
204
(5.4)
where
(5.5)
r
We observe that, being the Jacobian of i - 1 a positive definite symmetric matrix,
the problem (5.4) is equivalent to the problem of finding the minimum, over K, of the
quadratic, strictly convex function
(5.6)
Thus, one obtains a sequence {Fi} C K and it is possible to prove that this sequence
converges to the solution H.
Further, we have the following estimate
(2)1/2
A= ( 1 - - , (5.9)
pll
that represents, by (5.7), the best value for the rate of convergence.
The following theorem is proved by Zanni [18]:
Applying this theorem to (5.9) one has, in the affine case r(F) = J F + b:
2 ) 1/2
A ~ ( 1- 11 ;11 2 (5.10)
For the nonlinear case, denoting by a the minimum eigenvalue of a~<p, and by Ii
the maximum eigenvalue of the symmetric part [a~)f)l. of a~l{l, it results
a
YFEK: oX> 1-~.
- ~
.) 1/2
The graph of ( 1- ~ on ~/o. shows that, independently of G, the constant oX
takes values close to 1 also for values of ~ / 0. very close to 1.
Now, using the above results, we show a comparison between the projection method,
as developed by Dafermos, and the direct method.
We consider as an example the model of a circular highway exhibited by Bertsekas
and Gafni (see Fig. 1, and [19)).
In this case, the convex K is
I
F;+F;+s = 10' i = 1, ... ,5}
r is strongly monotone, and the components of r are quadratic functions of F:
r 1 (F) = 34Fl2 + 42FlF2 + 20Fl F3 + 20FI F4 + 42FlFs + 34Fl + 21F; + 20F2F3
+20F2Fs + 21F2 + 10F; + 10F3 + lOF; + 20F.Fs + lOF. + 21Fl + 21Fs + 34
r2(F) = 21Fl2 + 42FlF2 + 20Fl F3 + 20F1 Fs + 21Fl + 34F; + 42F2F3 + 20F2F4
+20F2Fs + 34F2 + 21F; + 20F3F4 + 21F3 + lOF; + lOF. + 10Fl + 10Fs + 34
r3(F) = lOFl2 + 20Fl F2 + 20Fl F3 + lOFl + 21F; + 42F2F3 + 20F2F. + 21F2 + 34F;
+42F3F. + 20F3FS + 34F3 + 21F; + 20F.Fs + 21F. + lOFl + 10Fs + 34
r.(F) = lOF; + 20Fl F. + 20FlFs + lOFl + lOF; + 20F2F3 + 20F2F. + lOF2 + 21F;
+20F3FS + 21F3 + 42F3F4 + 34F; + 42F4FS + 34F. + 21Fl + 21Fs + 34
rs(F) = 21F; + 20Fl F2 + 20F1 F. + 42FlFs + 21Fl + lOF; + 20F2Fs + lOF2 + lOF;
+20F3F. + 20F3FS + lOF3 + 21F; + 42F4FS + 21F. + 34Fl + 34Fs + 34
r6(F) = 23F; + 10F; + lOF;o + 20F6 F7 + 20F6FlO + 23F6 + 10FT + lOFlO + 23
r7(F) = 23F; + 10F; + lOF; + 20F Fs + 20F6Fr + 23F
T 7 + lOFs + lOF6 + 23
rs(F) = 23F; + 10F; + lOF; + 20FgFs + 20FsFr + 23Fs + lOFg + lOF7 + 23
rg(F) = 23F; + lOFl2o + lOF; + 20FgFlO + 20FsFg + 23Fg + lOFlO + lOFs + 23
rlO(F) = 23F;o + 10F; + lOF; + 20F6F lO + 20FlOFg + 23FlO + lOF6 + lOFs + 23.
Applying the direct method, proposed by [20] one finds, at the third step, the
unique solution
(%f ~ 0.003
250
200
150
100
50
Figure 1
The conclusion is that, except for some particular cases, the constant .A does not
assume those values that guarantee a rapid convergence of the method.
We conclude this presentation showing some problems where the analysis of sensi-
tivity of the equilibrium solution, with respect to changes in input data, leads to results
that seem counter-intuitive, and hence have been exhibited as paradoxes (see [21], [22]).
207
Even if several works are devoted to the sensitivity analysis, however these counter-
intuitive situations go out of the known analysis.
We present a generalization of Fisk's paradox [21] and, by means of the direct
method due to [20], we show that these situations may often occur.
p3.L-____________----------~P2
Figure 2
N {PI, P2,P3},
A W = {WI = (P3,P2), W2 = (P3,Pl ), W3 = (P2,Pl )},
R {RI,R 2,R3,R4}
with
Rl (P3, P2)
R2 (P3,Pd
R3 (P2 , PI)
R4 (P3,P2) U (P2,Pd.
Cl(f) 11
C2(f) h +a
C3(f) h
where a is a non-negative constant. Since
it F l +F4
12 F2
h = F3+ F4
the path costs are:
Cl(F) = Fl +F4
C2(F) = F2 +a
C3 (F) F3+ F4
C4 (F) Fl +F3 + 2F4 •
208
where
K = {F E R!: FI = Ph F2 + F4 = P2, F3 = P3},
Pi being constant positive travel demands.
Applying the method of [201, we find that the solution of the system:
CI(H) C2 (H)
HI = PI
H3 = P3
H2+H4 = p2
i.e.,
H_ (
- Ph
2P2 + PI + P3 -
3
a,P3, P2 - PI 3- P3 + a)
is the solution of (6.1) if:
Now, let us increase the travel demands related to the pair WI, from PI to PI + dO,
(d" > 0). The new solution is the vector H" with
H"I = HI +d"
d"
H; H2+-
3
HO3 = H3
d"
H"4 = H4 - -
3'
provided that
a - 2P2 ~ PI + P3 + d" ~ a + P2.
Then, if we calculate the total costs CO(HO)H" and C(H)H, we find:
~~----------------~~
"'-__________---------' Pl
Figure 3
Rl = (Pt,Pa)U(Pa,P,,)
R2 (PI, P2 ) U (P2 , P4 )
Ra (Pt , P2 ) U (P2 , Pa) U (Pa , P4 ).
CtC!) = 10ft
C2C!) c:t+/2
caC!) {3+ /3, 0<c:t<{3
C4C!) = c:t+h
csC!) 15
with the path flows F; given by:
ft = F2+Fa
h Ft
/3 F3
14 F2
15 Ft + Fa.
Then, the path costs are
with
K = {F e R 3+: Fl + F2 + F3 = p},
p being the travel demand for w.
Once again, using the above method, we obtain that the solution is given by the
system:
C1 (H) C2(H)
C2 (H) - C3(H)
HI +H2
~
°
P (6.4)
H3
HI
=
~
° H2 0, ~ 0.
Since it results
P
HI = H2 = 2' H3 = 0,
9
C2(H) - C3 (H) = -"2 P + a - {J < 0,
the vector H = (i, i,O) is the solution of (6.3) and the total cost is:
Now, we improve the total cost C3, considering a positive parameter {J* < a, instead
of {J, i.e.:
provided that
(6.5)
9 a* 31
"2 P +1-' < IS a . (6.6)
Thus, we find a paradoxical result for the class of problems where the parameters a,
{J, {J* and p are such that (6.5) and (6.6) are satisfied.
The case considered by Braess with a = 50, {J = 100, {J0 = 10, p = 6, is a particular
case of this class.
211
REFERENCES
[1) G. Fichera, "Problemi elastostatici con vincoli unilaterali: il problema di Signorini con ambigue
condizioni al contorno," Memorie dell'Academia Nazionale dei Lincei , pp. 91-140, 1964.
[2) G. Stampacchia, "Formes bilineares coercives sur les ensembles convexes," Compt. Rend. Acad.
Sci. Paris, pp. 4413-4416, 1964.
[3) J.-L. Lions and G. Stampacchia, "Variational inequalities," Comm. Pure Appl. Math. 22, pp. 493-
519, 1967.
[4) G. Stampacchia, "Variational inequalities," Proceedings of NATO Advanced Study Institute, Venice,
June 17-30, pp. 101-191,1968.
[5) O. G. Mancino and G. Stampacchia, "Convex programming and variational inequalities," J. Optim.
Theory Appl. 9, pp. 3-23, 1972.
[6) M. J. Smith, "The existence, uniqueness and stability of traffic equilibrium," Transportation Res.
138,pp. 295-304, 1979.
[7) F. Giannessi, "Theorems of the alternative, quadratic programs and complementarity problems" .
In "Variational inequalities and complementarity problems", R.W. Cottle et al. eds., J. Wiley, New
York, pp. 151-186, 1980.
[8) S. Dafermos, "The general multi modal network equilibrium problem with elastic demand," Network
12,pp. 57-72, 1982.
[9) M. De Luca and A. Maugeri, "Quasi-variational inequalities and application to equilibrium prob-
lems with elastic demand," in: F. H. Clarke et al. eds., "Nonsmooth optimization and related
topics", Plenum, New York, pp. 61-77, 1989.
[10) M. De Luca and A. Maugeri, "Discontinuous quasi-variational inequalities and application to equi-
librium problems," in: F. Giannessi ed., "Nonsmooth Optimization. Methods and Applications",
Gordon and Breach, U.K., pp. 70-75, 1992.
[11) M. De Luca and A. Maugeri, "Quasi-variational inequalities and application to the traffic equilib-
rium problem; discussion of a paradox," J. Compo and Appl. Math., 28, pp. 163-171,1989.
[12) A. Auslender, "Optimization: Methodes Numeriques", Masson, Paris, 1976.
[13] M. Fukushima, "Equivalent differentiable optimization problems and descent methods for asym-
metric variational inequality problems," Math. Progr. 53, pp. 99-110, 1992.
[14] D. L. Zhu and P. Marcotte, "A general descent framework for monotone variational inequalities,"
J. Optim. Theory Appl., 1993.
[15] F. Giannessi, "On some connections between combinational and continuous optimization," In "An-
nals of Operations Research", North-Holland. To appear.
[16] F. Giannessi, "Theorems of the alternative and optimality conditions." Jou. Optim. Th. Appl., Vol.
42, N. 3, pp. 331-365, 1984.
[17] S. Dafermos, "Traffic equilibrium and variational inequalities," Transport. Sci. 14, pp. 42-54,1980.
[18] L. Zanni, "On the convergence rate of two projection methods for variational inequalities in R n ,"
Calcolo, 1993.
[19] D. P. Bertsekas and E. M. Gafni, "Projection methods for variational inequalities with application
to the traffic assignment problem," Math. Progr. Study 17, pp. 139-159, 1982.
[20] A. Maugeri, "Convex programming, variational inequalities and applications to the traffic equilib-
rium problem," Appl. Math. Optim. 16, pp. 169-185, 1987.
[21] C. Fisk, "More paradoxes in the equilibrium assignment problem," Transport. Res. 13 B, pp. 305-
309, 1979.
[22] D. Braess, "Uber ein Paradoxon der Verkehrsplanung," Unternehmensforschung 12, pp. 256-268,
1968.
A VARIATIONAL INEQUALITY ASSOCIATED TO A SCALAR
CONSERVATION LAW WITH UNILATERAL CONSTRAINTS
Department of Mathematics
University of Pisa
Via F. Buonarroti 2, 56127 Pisa, Italy
e-mail: [email protected]
1. INTRODUCTION
213
214
tion closely follows the papers of Kruzkov [6] and Bardos-Leroux-Nedelec [1], a recent
note of Levi [9] and some results obtained independently by the author.
Starting from the fundamental work of E. Hopf [1] the central problem of the
weak or generalized solutions of conservation laws consists in describing the classes
of functions in which there is existence and eventually uniqueness for the solution of
the associated Cauchy problem or the mixed Cauchy-Dirichlet problem. The unique
solution, if it exists, should correspond to the physically relevent solution of the con-
servation law considered.
Ever since the fundamental paper [4] of Hopf the basic method used to study the
scalar conservation law equation
with bounded measurable initial values u(x,O) = uo(x) has been the so called van-
ishing viscosity method which consists in studying the family of solutions Up of the
corresponding parabolic equation
Remark 1. The function g(x, t, u) - E(oxj Ii )(x, t, u) is also referred to as the source
term.
In the one dimensional case, for feu) a convex function of u and for initial values
uo E Loo the problem is most studied. For a non convex feu), again in the one
dimensional case the problem was studied by Gelfand, Oleinik and others who also
determined a condition for the uniqueness among piecewise smooth solutions of the
Cauchy problem.
1, forr>O
sgn(r) = { 0, forr=O
-1, forr<O
and this can be approximated by the continuous function sgn>. defined, for A > 0, by
1, forr > A
sgn>.(T) = { 0, for Irl = A
-1, for r < A
In order to use suitable test functions starting from the weak solution of the equation
we introduce also the primitive of the function sgn>.:
riel
I>.(e) = 10 sgn>.(r') dr' for eE JRn
Then it is easily seen that
e--+ I>.(e) is a convex function on JRn and that
It is some times convenient to use the fact that a function w belongs to BV(JRn )
if, for every relatively compact open set wand for all h E JRn, we have
{ °
u E BV(JRnH )
E(u) = in JRnH
u(x,O) = uo(x)
We have the following global existence result:
Theorem 1 (Conway and Smoller) [2]. If Uo E BV(JRn ) then there exists a weak
solution of the Cauchy problem
{ u E BV(JRnH )
Otu+divf(u)=O 10 JRn+l
u(x,O) = uo(x)
such that, for all t = const, x 1-+ u(x, t) belongs to BV(JRn ).
217
The proof is based on the method of vanishing viscosity and the method is justified
only when the initial data uo is sufficiently smooth.
An existence result without uniqueness for the general quasilinear equation E( u) =
o was proved by Kuznecov in 1967.
In order to obtain existence and uniqueness in the Cauchy problem Kruzkov in-
troduced a some what different notion of generalized solutions in the class of functions
Loo in [6].
Assume
f : ZT X IR --+ IRR and g: ZT X IR --+ IR
JlZT[ [Iu(x, t) - klat ¢> + sgn( u(x, t) - k) l)h(x, t, u(x, t)) - hex, t, knox; ¢>-
Remark 2. For k = ±suplu(x, t)1 the inequality (*) implies the original weak formu-
lation.
u±(x, t) = p~±O
lim u«x, t) + pv(x, t))
we have, for all constants k,
lu+(x, t) - klcos(v(x, t), t) + sgn( u+(x, t) - k )[Jj(x, t, u+) - hex, t, k )]cos(v(x, t), x j)
218
::; lu-(x, t) - kjcos(v(x, t), t) + sgn( u-(x, t) - k )[/j(x, t, u-) - /j(x, t, k )]cos(v(x, t), Xj)
In order to state the principal results of Kruzkov we introduce the following no-
tation: Let
N = NM(R) = {x,t)E{lxl=:;R}
max [~)ou/j)2(x,t,uW/2
lul=:;M j
let /C denote the cone ({x,t); Ixl ::; R - NT, 0 ::; t::; To = min(T,~)} and
ST = /C n {t = T} for T E [0, To]
with the viscosity parameter E. Then one proves that {u.(x,t)}, {Otu.(x, t)} and
{ox;u.(x,t)} are bounded In ZT, and finally that {Otu.} and
{ox;u.} are uniformly Holder continuous. These imply that u. tends to a limit u
a.e in ZT which is a generalized solution of the problem.
219
In this section we follow closely the paper [1] of Bardos et al. We consider the
following Cauchy-Dirichlet problem for the quasilinear hyperbolic equation:
where, denoting the Lipschitz norm of a function with respect to the variable u by
II· II Lipu , we have
C = IIgllLipu + L 118"'i/jIlLipu
i
Next we multiply the equation E(u.) = €~u. by sgn(8t u.), integrate by parts over n
°
and using the initial condition u.(x,O) = we obtain an estimate of the form
Remark 4. It can be shown that if u = Ll(Qr) -lim U<m then u E BV(Qr) and
u(x,O) = uo(x) but u does not satisfy the homogeneous boundary condition on ~r.
This can be seen easily in the case of the well known Burger's equation
with u(x,O) = 1. An exact solution, given by the method of vanishing viscosity, for
t < lis
u(x, t) = {xlt
for x < t
1 forx>t
The difficulty consists in the fact that the speed of propagation depends on term
involving f(x, t, u) (which in the case of Burger's equation is ~~ = u.
These considerations suggest that the boundary condition should be reformulated
suitably. The following formulation of the problem was introduced in the paper of
Bardos, Leroux and Nedelec [1)
E(u) = 0 in Qr
{ u(x,O) = uo(x) in n
minkEJ('n::,u,O)(sgn('YE,u)[f(x,t,'n::,u)-f(x,t,k»),II(x,t» =0 on ~l
where J( 'YE, u, 0) denotes the closed interval [mine 'YE, u, 0), max( 'YE, u, 0»). We now have
the following definition of generalized solutions of the mixed problem.
JkT [lu(x, t) - klat ¢> + sgn(u(x, t) - k) L)Ji(x, t, u(x, t» -Ji(x, t, k)}axj ¢>
This definition generalizes that of Kruzkov for the Cauchy problem on the whole
space and hence without the boundary condition. The weak solution is the so called
entropy solution which characterizes the physically relevent solution among those ob-
tained by the vanishing viscosity method.
As we shall formulate the corresponding parabolic mixed problem (the associated
viscosity problem), taking into consideration of the above definition, for a more general
situation we shall postpone the details to the next section.
We have the following main result:
Theorem 4 (Bardos, Leroux and Nedelec) [1]. The Cauchy Dirichlet problem has a
unique weak solution given by the vanishing viscosity method.
The proof is again based on the vanishing viscosity method after an appropriate
weak formulation. We shall call a solution of the mixed problem in the sense introduced
by Bardos et al. a weak-entropy solution.
In this section we shall come to our main problem of the inequation associated to
the first order quasilinear hyperbolic equation. We consider a unilateral regular obsta-
cle condition which is compatible with the boundary condition. After a translation we
may assume this to be zero.
Let Kdenote the cone {u E LOO(QT) n BV(QT)i u(x, t) :::: 0 a.e.in QT} and let
Uo : n -+ lR with Uo :::: 0 be given.
We shall be concerned with the following
Problem. To find (u, E'), with u E K and E' a subset of the lateral boundary ET of
QT, such that
E(u) = atu + divf(x,t,u) + g(x,t,u) = 0
{
on Q~ = {(x, t) E QTi u(x, t) > O}
u(x,O)=uo(x) a.e.on n
u(x,t) = 0 on E'
where we assume the following hypothesis:
(i) f E C2(n X [0, t] X lR, lR n ), f = (II,'" ,In)
(ii) 9 E Cl(n X [0, t] X lR)
(iii) u 1-+ g(x, t, u) and u 1-+ ax; hex, t, u) are Lipschitz continuous in u uniformly
with respect to (x,t)
(iv)uo E LOO(n) n BV(n) with Uo :::: 0 a.e. in n.
We introduce the following definition, a slight modification of the definition of
weak solutions introduced in the previous section.
Here we give a very brief sketch of the idea of the proof. The proof is based on the
method of penalisation to reduce to a problem associated to a quasi linear equation of
the type studied by Bardos, Leroux and Nedelec in [1]:
To find, VTJ > 0 a solution u.,(x, t) and I;' C I;T satisfying
E(u ., ) - lu-.,.,
= 8tu., + div f(x, t, U.,) + g(x, t, U.,) - lu;;- =0
{ .
III
QT
U.,(x,O) = uo(x)
U.,(y, t) = 0 on
a.e. on n
I;'
.,
E(u'l .• ) - ~u;;-.•
= 8tu'l •• + divf(x,t,u'l .• ) + g(x,t,u'l .• ) - ~u;;-.• = f/~.U'l.f
{ in QT
u'l .• (x,O) = uo(x)
u'l .• (y, t) = 0 on
a.e. on
I;'
n
223
The weak form of this equation is formulated in the same way as above. By the result of
Bardos et al. (section 3) there exists a unique entropic solution u" E LOO( QT )nBV( QT)
obtained by the method of vanishing viscosity: let U",E be the unique viscous solution
of the parabolic problem (e > 0 is the viscosity parameter).
One proves that the family {U",E, e > 0, '7 > O} is bounded in the space Wl,l( QT)n
LOO(QT)' By using the monotonicity of the penalising operator (3(u) = - U - this
bounded set is seen to be independent of the penalisation parameter '7. It then follows
that, first passing to the limits as e -+ 0+ the family of solutions of the penalised
hyperbolic problems belong to a fixed bounded set in the space LOO(QT) n BV(QT).
Finally, using the compactness of the inclusion Ll(QT) n BV(QT) in L1(QT) we can
pass to the limits as '7 -+ 0+. By the result of Bardos et al. this limit is the required
unique weak-entropy solution.
REFERENCES
[I] C. Bardos, A.Y. Leroux and J.C. Nedelec, "First order quasilinear equations with boundary
conditions", Comm. Partial Differential Equations 4, 1017 - 1034, 1979.
[2] E. Conway and J. Smoller, "Global solutions of the Cauchy problem for quasi-linear first order
equations in several space variables", Comm. Pure Appl. Math. 19), 95 - 105, 1966.
[3] A. Friedman, "Partial differential equations of parabolic type", Prentice Hall, New York, 1969.
[4] E. Hopf, "The partial differential equation Ut + UU., = Jlu:u", Comm. Pure Appl. Math. 3,
201 -230, 1950.
[5] O.A Ladyzenskaja and N.N. Ural'ceva, "Boundary problems for linear and quasilinear parabolic
equations, I - II", Iw. Akad. Nauk SSSR, 26, 2 - 52, 1964'.
[6] S.N. Kruzkov, "First order quasilinear equations with several independent variables", Math.
USSR - Sh., 10,217 - 243,1970.
[7] S.N. Kruzkov, "Methods for constructing generalized solutions for the Cauchy problem for a
quasilinear equation of the first order", Uspehi Mat. Nauk 20, 112 - 118, 1965.
[8] P.O. Lax, "Shock waves, increase of entropy and loss of information", Publ. Math. Sci. Inat.,
2, 129 - 243, 1984.
[9] L.Levi, "Equations quasi lineaires du premier ordre avec contrainte unilaterale", C.R. Acad. Sci.
PaN, 317, 1133 - 1136, 1993.
[10] J .-L. Lions, "Quelques methodes de resolution des problemes aux limites non lineaires", Dunod,
Gauthier - Villars, Paris, 1969.
[11] F.Mignot and J.P. Puel, "Inequations variationelles et quasi variationelles hyperboliques du
premier ordre", J. Math. Pure. Appl. 55, 353 - 378, 1976.
[12] A.I. Vol'pert, "The spaces BV and quasilinear equations", Math. USSR Sb., 2, 225 - 267, 1967.
CONTINUITY OF THE SOLUTION SET
OF HOMOGENEOUS EQUILIBRIUM PROBLEMS
AND LINEAR COMPLEMENTARITY PROBLEMS
1 Universitat Mannheim
Fakultat fiir Mathematik und Informatik
68131 Mannheim, Germany
2 Hanoi Institute of Mathematics
P.O. Box 631 Bo Ho
Hanoi, Vietnam
Abstract: Denote by S(M, q) the solution set of the linear complementarity prob-
lem
where ME R,nxn and q E lRn. M is called an Ro-matrix iff S(M,O) = {O}. Jansen and
Tijs have proved that if M is an Ro-matrix, then the map S is upper semicontinuous at
(M, q) for every q E lRn. We prove that this property is characteristic for Ro-matrices.
Part of our results extends to homogeneous equilibrium problems of the type
z E K, J(z, y) + (q, y - z) ~ 0, Vy E K.
225
226
Here z ;::: a means that z belongs to the nonnegative orthant R+ of nn, and (.,.)
denotes the inner product in nn. The set of all z satisfying (1.1) is denoted by SCM, q).
Definition 1. M is said to be an Ro-matrix iff S(M, 0) = {a}. The set of all Ro-
matrices of order n is denoted by Ro. (In [4], [5], and [10], the same set is denoted by
E*(O). In [8] Ro-matrices are called also pseudo-regular matrices.)
First of all we consider a somewhat more general class of problems, namely ho-
mogeneous equilibrium problems.
K n (-K) = {O}. Then there exists a linear form (I,·) on R n such that
In fact, if K is pointed, then int K* "10, and every IE int K* has property (2.1). From
now on, some I which satisfies (2.1) is fixed. We set B := {x E K : (l,x) S; I}. Denote
by F the space of all continuous functions f: K x K - t R such that f(O,y) ~ 0 for
all y E K, and
f(>..x, >..y) = >..p+l f(x, y), 'Ix, y E K, V>" ~ 0, (2.2)
Let S(J, q) denote the solution set of (2.4). We consider the multi valued mapping S(·,·)
defined on F x R n with values in Rn. Using (2.2), (2.3), (2.4) it is easy to see that S
has closed graph. The latter means that if!k - t fin F, qk - t q in R n , Zk E S(Jk, qk),
and Zk - t Z in R n , then Z E S(J, q). Note that S(J,O) is a nonempty cone. We define
S is said to be upper semicontinuous at (J, q) E F x R n iff for each open set n :::> S(J, q)
there exist neighborhoods V of f in F and U of q in R n such that S(J', q') c n for
every (J', q') E V x U.
Proof: (a) Assume that f E Fo, q ERn, but S(f,q) is unbounded. Then there exist.s
a sequence {zd eRn such that IIZkll - t 00 and, for all k,
Let Vk := zk/llzkll. There exists a subsequence of {vd, denoted again by {vd, such
that Vk - t v "I O. Taking account of (2.2) we get from (2.6) that
y 1 y
Vk E K, f(vk, IIZkll) + IIZkll p (q, IIZkll - Vk) ~ 0, Vy E K,
228
and therefore
1
Vk E K, VTJ E IIzkll K = K.
Passing to the limit, we have
it follows, as from (2.6) in the proof of part (a), that v E S(f,O). Since f E F o , this
contradicts with v =I O.
( c) Assume that Fo is not open in F. Then there exists 1 E Fo and a sequence
Ud C F such that fk -+ 1 and !k f/. Fo, i.e., S(fk, 0) =I {O}. Since S(fk, 0) is a
nonempty cone there exists for every k some Zk E S(fk, 0) with IIzk II = 1. We may
assume that Zk -+ z =I o. Then z E Sa,O), since S has closed graph. Together with
Z =I 0 this gives 1 f/. F o, a contradiction.
(d) Ab absurdo, let us suppose that S(·,q) is u.s.c. at /, and / f/. Fo. Then there
exists z =I 0 such that
z E K, fez, y) "2 0, Vy E K. (2.7)
(l,X})P _
ft(x,y):=J(x,y)- ( t(I,Z) (q,y-x). (2.8)
We have ft()..x, )..y) = )..p+l ft(x, y) for all)" "2 0, so that ft E F. Moreover, the following
holds:
ft(Zt, y) = f(zt, y) - (q, y - Zt).
Thus for all y E K we have
1
ft(Zt, y) + (q, y - Zt)= f(zt, y) = tp+l fez, ty) "2 0,
from (2.7) and since ty E K. Therefore Zt E S(ft,q) for all t > O. Observe that ft -+ f
as t ! O. Since the set S(f, q) is bounded we can choose a bounded open set 0 :J S(f, q).
Since S(·,q) is u.s.c. at f we should have Zt EO for all t > 0 sufficiently small. But
this is not possible since II Zt II -+ 00 for t ! O. This completes the proof. 0
229
It can be shown (this is a particular case of [1], Theorem 2) that for (f,q) E Fx R,n
problem (2.4) has a solution if I is convex in the second argument, I(x,x) = 0 for all
x E K, and moreover the following conditions hold:
(i) I is copositive in the sense that l(x,O) :::; 0 for all x E K.
(ii) I is q-regular in the sense that {z E S(f,O) : (q, z) :::; O} = {OJ.
These same conditions (i), (ii) ensure also that S is u.s.c. at (f, q), provided we restrict
the first argument of S to copositive functions. To make this precise, let V denote the
set of all I E F such that I is copositive. Then we have the following extension of
Gowda's result quoted in Sect. 1.
Proof: (a) Assume, for contradiction, that S IVxRn is not u.s.c. at (f, q). Then
there exists an open set n :::> S(f,q) and sequences {fd c V, {qd c R n such that
Ik -+ I, qk -+ q, Zk E S(!k, qk) and Zk f/. n for every k. If there exists a subsequence of
{Zk} converging to some z, then we obtain simultaneously Z E S(f, q) and Z f/. n, which
is impossible. Thus we can assume that IIZkll -+ 00 and Vk := zk/llzkll -+ v =I- O. From
Zk E S(!k,qk) it follows, as in the proof of part (b) of Theorem 1, that v E S(f,O).
Moreover it follows from Zk E S(fk, qk) that
Since h, is copositive this implies (qk, -Zk) ~ 0, and in the limit (q,v) :::; O. Since I is
q-regular we must have v = 0, which contradicts v =I- O.
(b) Ab absurdo, suppose that S(·,q) Iv is u.s.c. at I, but I f/. Fo. Then there exists
Z =I- 0 such that z E S(f,O). For all t > 0 let Zt and It(x, y) be given by (2.8). As
in the proof of Theorem 1, part (d), it follows that It E F and Zt E S(ft,q) for all
t > O. Moreover, since I is copositive and q E -K* we have It(x, 0) :::; 0 for all x E K.
Hence It E V for all t > O. Since S(f, q) is bounded we can choose a bounded open
set n :::> S(f, q). Since It -+ I for t 1 0 and since S(" q) Iv is u.s.c. at I we should have
Zt En for all t > 0 sufficiently small. But this is impossible since IIZtll -+ 00 for t 1 O.
This completes the proof. 0
We observe that part (b) of Theorem 1 yields simultaneously the upper semiconti-
nuity of S(·,q) in an entire neighbourhood of the given element I E Fo, since Fo is
open by part (c) of Theorem 1. Likewise part (a) of Theorem 2 yields without further
assumptions the upper semicontinuity of S IVxRn in an entire neighbourhood of (f, q)
in V x R n. In fact, if I is q-regular, then there exists a neighbourhood U of (f, q) in
F x R n such that f' is q'-regular for all (f', q') E U. The proof of the latter fact is
like the proof of part (c) of Theorem 1.
230
Here p > 0 is fixed, and is the same as in the definition of:F. For I E F and q E RR
problem (2.4) becomes the homogeneous complementarity problem
since for every fixed z E J{ and ~ E RR one has the following equivalence:
For IE F and q E lRR let BU, q) denote the solution set of (2.4) or, which is the same,
of (2.11). Let Fo := {J E F: BU,O) = {O}}.
Let M be the space of all continuous mappings M(·) : J{ - t R n which satisfy
(2.10), and provided with the norm IIMII:= max
(x,y)EBxB
I(M(x), y)l. Clearly we may
identify F with M, and 11111 = IIMII if I and M are related via (2.9). For M E M
and q E RR we let S(M,q) denote the solution set of (2.11). Moreover we define
Mo := {M EM: S(M, 0) = {On. Clearly Mo can be identified with Fo via (2.9).
Proof: We identify M E M with I E j via (2.9), and take into account the equiva-
lence of (2.4) and (2.11) under this identification. Then the proof is a word-by-word
replica of the proof of Theorem 1, where we have to replace throughout:F by F,:Fo by
Fo, and S by S. Concerning the proof of part (d) we note that if f(x, y) = (M(x), y-x)
with M
Rearranging the content of parts (a), (b), (d) of Theorem 3 we obtain the following
characterization.
Corollary 1. Let ME M. Then MEMo if, and only if, for some (equivalently: for all)
qERR the set S(M,q) is bounded and the map S(',q) is upper semicontinuous at M.
231
There exists q E R n such that SCM, q) is bounded if, for some A > 0, the set
M(K \ AB) is convex and
In fact, it follows from (2.12) and the separation theorem ([16], p. 97) that there exists
a nonzero vector /-' ERn such that
(/-"M(x))~O, VxEK\)"B.
We turn now to the linear complementarity problem LCP(M, q), see (1.1). To this
end we have to assume that K := JR+, and that p = 1 in (2.10). R nxn can be con-
sidered as a subspace of M, by identifying the matrix M E R nxn with the linear
mapping M(·) E M. For consistency we have to provide R nxn with the seminorm
IIMII:= max I(Mx,y)l. If M E R nxn and K = R+, then problem (2.11) coin-
(x,y)EBxB
cides with the linear complement8.l·ity problem (1.1). So does problem (2.4), if we set
f(x,y) := (Mx,y - x). The solution mapping S(M,q) is now defined on R nxn x JR n ,
and 'Ro , as introduced in Sect. 1, coincides with M o n JR nxn .
Proof: The proof is completely analogous to the proof of Theorem 1, where we replace
F by R nxn and Fo by 'Ro. 0
232
Parts (a)-(c) of Theorem 4 have already been proved by Jansen and Tijs [10]. Gowda
[6] removed the hypothesis SCM, q) -10.
For LCP there is an additional fact (which probably fails to hold for the more general
cases considered in Sect. 2), namely:
Lemma 1. Let M E lR nxn and K = lR~. Then there exists q E lRn such that SCM, q)
is bounded.
Proof: Let q E lR n be such that S(M,q) is non empty and unbounded. Then there
exists a sequence {zk} C S(M,q) with IIzkll-+ 00. Let v k := Mzk + q. Then for all k
we have
Zk 2: 0, v k 2: 0, (zk, v k ) = O.
Let J(zk) := {j : zJ > OJ. Without any loss of generality we may assume that
J(zk) =: J is independent of k. Let I denote the complement of J. We have for
all k that
q=_MZk+V k .
where Mj denotes the j -th column of the matrix M, and Ij denotes the j -th column
of the (n x n) unit matrix I. Now II(zk,vk)ll-+ 00, and without any loss of generality
we may assume that (zk,vk)/lI(zk,vk)ll-+ (z,v) -I (0,0). Dividing (3.1) by II(z\vk)11
and passing to the limit we obtain
(3.2)
From (3.1) it follows that q is an element of the convex conical hull of the vectors
-Mj (j E J) and Ij (j E I). From (3.2) it follows that these n vectors are linearly
dependent. Thus, their convex conical hull has dimension less than n, and it is closed.
The union of finitely many closed convex sets in lR n, each having dimension less than
n, cannot equal all of lR n (as a consequence of Baire's Lemma, see [2], p. 15). In fact,
its complement is open and dense in lRn. Hence there must remain some q E lR n such
that S(M,q) is bounded (may be empty). 0
Combining Lemma 1 with parts (b) and (d) of Theorem 4 we obtain the following
characterization:
Theorem 5. Let M E lRnxn. Then M E Ro if, and only if, S(·,q) is upper semicon-
tinuous at M for all q E lR n.
Proof: Assume that M E lR n x nand M rfc 'Ru. Then there exists zf 0 such that
z 2 0, Mz 2 0, (z, Mz) = o.
Let v := Mz. Let J := {j : Zj > O} and J the complement of J. Since ZjVj = a for all
j, we have
REFERENCES
[1] E. Blum and W. Oettli, "From optimization and variational inequalities to equilibrium problems",
Math. Student 63, 123-145, 1994.
[2] H. Brezis, "Analyse fonctionnelle", Masson, Paris, 1983.
[3] R.W. Cottle, J.-S. Pang, and R.E. Stone, "The Linear Complementarity Problem", Academic
Press, New York, 1992.
[4] R.D. Doverspike, "Some perturbation result.s for the linear complementarity problem", Math. Pro-
gramming 23, 181-192, 1982.
[5] C.B. Garcia, "Some classes of matrices in linear complementarity theory", Math. Programming 5,
299-310, 1973.
(6) M.S. Gowda, "On the continuity of the solution map in linear complementarity problems" , SIAM
1. Optimization 2, 619-634, 1992.
[7] M.S. Gowda, "Applications of degree theory to linear complementarity problems", Math. Oper.
Res. 18, 868-879, 1993.
(8) M.S. Gowda and J .-S. Pang, "On solution stability of the linear complementarity problem", Math.
Oper. Res. 17,77-83, 1992.
[9) C.D. Ha, "Stability of the linear complementarity problem at a solution point", Math. Program-
ming 31, 327-338, 1985.
[10] M.J.M. Jansen and S.H. Tijs, "Robustness and non degenerateness for linear complementarity
problems", Math. Programming 37, 293-308, 1987.
234
[11] O.L. Mangasarian and T.-H. Shiau, "Lipschitz continuity of solutions of linear inequalities, pro-
grams and complementarity problems", SIAM J. Control Optim. 25, 583-595, 1987.
[12] K.G. Murty, "On the number of solutions to the complementarity problem and spanning properties
of complementarity cones", Linear Algebra Appl. 5, 65-108, 1972.
[13] K.G. Murty, "Linear Complement.arity, Linear and Nonlinear Programming", IIeldermann- Verlag,
Berlin, 1987.
[14] S.M. Robinson, "Generalized equations and their solutions, Part I: Basic Theory", Math. Pro-
gramming Study 10, 128-141, 1979.
[15] S.M. Robinson, "Some continuity properties of polyhedral multifunctions", Math. Programming
Study 14, 206-214, 1981.
[16] R.T. RockafeUar, "Convex Analysis", Princet.on University Press, Princeton, 1970.
TENSOR APPROXIMATIONS OF SMOOTH NONLINEAR
COMPLEMENTARITY SYSTEMS
Rapcs8.k Tamas
1. INTRODUCTION
235
236
coordinate representations and a new view concerning the structure of NLCS could
be shown. This approach develops further the results in papers (Rapcsak, 1989, 1991;
Rapcsak and Csendes, 1993; Rapcsak and Thang, 1994).
In Section 2 the problem will be set, in Section 3 a new notion, the tensor field
CS will be introduced defining a generalization of NLCS. Section 4 contains a tensor
formula with first-order and second-order tensor approximations of smooth functions
defined on a Riemannian manifold, based on Taylor expansions, by using covariant
derivatives and some consequences of this approach will be considered. In Section 5
the usefulness of this notion will be demonstrated by suggesting a class of penalty
methods based on globally convergent variable metric algorithms along geodesics of a
Riemannian manifold for solving NLCS.
h;(x) T
= = 1, ... ,n,
I
First, the notion of the tensor is recalled, thereafter the tensor and tensor field
CS will be defined. Let M be a C 2 n-manifold and m a point in M. The tangent
space TMm at m is an n-dimensional vector space. Let TM:;' be the dual space of
TMm, TM:;' endowed with its natural vector space structure. Now, the theory oflinear
algebra can be applied to define tensors (e.g., Spivak, 1979).
where T;, T;, i = 1, ... , n are p-covariant and q-contravariant tensor fields on A, respec-
tively. If all the tensor fields are considered at a fixed point mEA, a tensor CS is
obtained at m E M.
Since the inequality constraints of problems (2.1) and (2.3) are O-covariant tensor
fields on R n , thus problems (2.1) and (2.3) are special tensor field CS on Rn. Moreover,
the only equality constraint in (2.1) and (2.3) results in a manifold, thus the inequality
constraint functions can be considered on this manifold. Tensor field CS can be origi-
nated, e.g., by characterizing the equilibrium positions of a system of material points
on which active forces have an effect, mainly if the constraints depend on the veloc-
ity and the system is scleronomous and non-holonomous (e.g. Rapcsak and Szenthe,
1990). Tensor field optimization problems fitting better to the structure of a non-
linear optimization problem class whose equilibrium systems may be also formulated
as tensor field CS can be similarly defined (Rapcsak and Csendes, 1993). Since the
objective function of a general optimization problem defining a constraint set A ~ M
in a Riemannian manifold is a O-covariant tensor field on A, thus these optimization
problems are unconstrained tensor field optimization problems on A ~ M.
It turns out from Definition 3.2 that neither the values of the tensors nor the
solutions of problems (3.1) change by a nonlinear coordinate transformation, thus this
problem class can become an adequate tool to study the structure of problems (2.1)
and (2.3). Consider an arbitrary coordinate representation of the constraints manifold
M[hd or M[h2J in a neighbourhood of any point m in problem (2.1) or (2.3). Then,
a tensor CS assigned to a given point m and to a coordinate representation can be
considered an NLCS with a special structure, i.e., all the functions are multilinear.
This property may be useful for elaborating efficient algorithms.
To build and study tensor field CS the operations of tensor algebra (addition,
subtraction, multiplications and contraction) and of tensor analysis (covariant dif-
ferentiation) can be applied, as well as other operations which preserve the tensor
character.
In the paper, covariant differentiation (e.g., Spivak, 1989; Gabay, 1982) will be
the most important tool. In a system of local coordinates, the coefficient functions
of covariant differentiation (affine connection) r:~ '
2' 11, 12 , 13
n
= 1, ... , n define the co-
variant derivative for all the tensor fields. Let vr = E Vi3r:~'2' 11 ,12 = 1, ... , n,
1.=1
= 1, ... , n are the component functions of a covariant vector field, and let
where Vi., 13
n
vr = E V" r:~'2' 12 , h = 1, ... ,n, where V", 11 = 1, ... , n are the component func-
1,=1
tions of a contravariant vector field. The following result is well-known in differential
geometry (e.g., Spivak, 1979):
DV = Jv+vr, (3.3)
where JV denotes the Jacobian matrix of the corresponding vector field and vr is
the multiplication of the vector field and the 3-dimensional matrix r at each point
of an arbitrary coordinate neighbourhood. For an arbitrary tensor field the covariant
derivative forms a tensor field. If the tensor field is scalar (i.e., a smooth function on
M), then the covariant derivative is equal to the gradient.
Definition 3.3 If
Theorem 3.1 Let G be a symmetric matrix function defining the metric on a Rie-
mannian manifold M in any system of local coordinates. Then, there exists a unique
symmetric connection such that
g"'2)(G- 1 )'a o
ax, + ag'2
r'a1,'2 = ~1/2(ag"O ax, aaxo
0 _
~ (3.5)
0'=1 2 1
We have seen above that NLCS (2.1) and (2.3) can be considered 2n O-covariant
tensor field inequalities on M[hIJ and M[h2J, respectively. Let D f and D2 f denote the
first-order and second-order covariant derivatives of an arbitrary smooth function f on
240
a Riemannian C2 k-manifold M (with respect to the Riemannian metric), respectively.
In order to draw conclusions for the local structure of smooth NLCS, the function f
and the manifold M will be investigated in a neighbourhood of an arbitrary point in
the form of
f(x(u», (4.1)
Definition 4.1 A curve of M is a geodesic if its tangent is parallel along the curve. A
set A ~ M is geodesic convex if any two points of A are joined by a geodesic belonging
to A ~ M.
The definition of geodesics coincides with the classical one in the case of Rie-
mannian metrics (e.g., Spivak, 1979). Consider an arbitrary coordinate representation
of the manifold M in any geodesic convex neighbourhood of M. Then, every two
points of this neighbourhood can be joined by a unique geodesic given in the form of
x( u( s», s E [0, bl, where this function is twice continuously differentiable, s means the
arc length parameter and b is the length of the geodesic between the two points.
Theorem 4.1 In any geodesic convex coordinate neighbourhood of M and for every
geodesic x( u(s» joining two arbitrary points x( u(b» and x( uo) = x( u(O», s E [0, bl,
where e fulfills the equality lim.....o e(O; s)/ s2 = 0 and u~ is the tangent of the geodesic
at uo. IT x(uo) is a stationary point, then
1
f(x(u(s») = f(x(uo» + '2(u~fD2 f(x(uO»u~s2 + e(O;s). (4.3)
where the k X k x k matrix r( u) contains the second Christoffel symbols, u'( s), s E [0, bJ
are tangent vectors and (4.6) means that ui'(s) = -u'(s)Tri(u)u'(s), i = 1, ... , k.
Thus,
f(x( u(s))) = f(x( uo)) + ~(u~f D2 f(x( Uo ))u~s2 + c(O; s). (4.8)
o
where IT M[ hi] as index denotes restriction to the tangent spaces T M[ hi], i = 1,2.
Proof. The space R n should be considered at every point of M[h 1 ] or M[h2] the direct
sum of the tangent space and the normal space with respect to the Euclidean metric
spanned by the column vectors of J Xi(U), i = 1,2 and h:(Xi(U))T, i = 1,2, respectively.
Thus, it is possible to decompose any vector of R n at every point of M[hlJ or M[h2J
242
into a tangential and a normal component. By using these notations for the column
vectors of the n x (n - 1) x (n - 1) matrices x:' (uo), i = 1,2, the well-known Gauss
equation (e.g., Spivak, 1979) can be formulated as follows:
D2 f(x;(u)) =
Jx;(u)T f~(x;(U))Jxi(U) + f~(Xi(U))X:'(u) - f~(x;(u))Jx;(u)fi(U) =
= Jx;(u)T f~(x;(u))Jx;(u) + f~(x;(u))(x:'(u) - JXi(u)f;(u)) =
by Gauss equation,
(4.11)
= Jx;(uf f~(Xi( u))J x;(u) + f~(x;(u))h:(Xi(U))T Bi(U) =
= Jx;(u)T f~(x;(U))Jxi(U) + (f~(Xi(U))T + f~(X;(U))N)h:(x(u))TBi(U) =
= JXi(uf f~(Xi(U))Jxi(U) + f~(Xi( u))Nh:(Xi(U))T B;( u),
uEU~Rn-l, i=1,2.
As
h;(x;(u)) = 0, u E U ~ Rn - l , i = 1,2 (4.12)
i = 1,2. (4.13)
i = 1,2 (4.14)
is a positive definite tensor and the first order condition is fulfilled, then the strict
local optimality is guaranted in a neighbourhood.
(3) If we can extend continuously the positive second-order covariant derivative at a
stationary point (second-order optimality conditions) for a coordinate neighbour-
hood of the manifold as a second-order positive definite covariant tensor field, then
a Riemannian metric can be introduced there. In nonlinear optimization, by using
first-order and/or second-order information almost every method is based on an
approximation of such a Riemannian metric. The approximations of the Hessian
matrix of the Lagrangian function in a neighbourhood of a local optimum point
(e.g., Newton and Newton-like methods, continuous methods of second-order de-
gree, sequential quadratic programming methods, interior point methods, reduced
gradient methods, penalty function methods, etc.), in general, are not tensor ap-
proximations. The main idea of methods is to choose a decreasing direction v at
any iteration point such that the inner product of a projection of /' to the tangent
space of the constraint manifold and v should be minimized in the Riemannian
metric generated by this tensor field (Rapcsak and Thang, 1994).
Now, the second-order tensor approximations of functions g(x) and x in NLCS
(2.1) will be determined on the basis of formulas (4.2) and (4.9). These statements
are true in the case of any coordinate representation of the complementarity manifold
M[h l ], so we choose a convenient one. Let us fix a point Xo and since h~(xo) -I 0,
suppose that the nth component is not zero. Let
i.e., the first (n -1) coordinates give the parameters of the complementarity manifold
in a neigborhood of Xo. By applying the implicit function theorem, we obtain the
complementarity manifold M[hd = {x E Rnlhl(x) = O} in a neighbourhood of Xo as
follows:
x(U) = ( 4.16)
o
1 o
o ) ,
1 (4.17)
~i!!.!:l
- 8U n _l / aXn
5. AN OPTIMIZATION APPROACH
The history of algorithms for solving the finite-dimensional VI and NLCS is rela-
tively short. From the beginning it has been recognized that NLCS are special cases
of VI, still the research directions were different: VI have been considered in infinite-
dimensional metric spaces and NLCS in finite-dimensional Euclidean spaces. In the
early years of study on CS most effort was devoted to LCS due mainly to the applica-
bility of Lemke-type algorithms. Cottle (1966) developed the first method for NLCS
by extending the principal pivoting algorithms for LCS. This idea led to the class of
methods referred to as fixed-point and homotopy algorithms which are not employed
245
in the case of VI. Mangasarian (1976) transformed NLCS into a system of nonlinear
equations to put classical algorithms into forefront which seems to be a promising
approach. Nonsmooth equation formulations of NLCS led recently to a robust itera-
tive algorithm published by Pang and Gabriel (1993). The method NE/SQP solves
a sequence of nonnegatively constrained quadratic programs of the least-squares type
based on a robust SQP algoritm. By tensor approximations, a new smooth robust SQP
type approach may be introduced. The reformulation of NLCS (or VI) as optimization
problems has also the advantage to use classical methods for solving NLCS. In the case
of (2.1), the usual form is
where the solution point x· must meet the requirement of g(x*)T x* = O. In general, the
feasible region of (5.1) is not convex, but in the case of LCS, problem (5.1) is quadratic.
To solve the optimization problem, among other things, the direct use of software
MINOS, cutting plane based and continuation methods were suggested. Another idea
based on gap functions is to cast NLCS as an unconstrained minimization of a smooth
function derived from (5.1). Fukushima (1992) formulated asymmetric Variational
Inequality problems as differentiable optimization problems which are involved in an
implicit Lagrangian function approach published by Mangasarian and Solodov (1993).
The most general framework for gap functions seems to be introduced by Giannessi
(1994).
The fixed-point and optimization approaches have advantages and disadvantages,
namely, they lack either the computational efficiency necessary for solving large-scale
equilibrium problems or the generality. A detailed description of results on finite-
dimensional NLCS and VI can be read in a state-of-the-art paper by Harker and Pang
(1990).
A general approach for solving NLCS, VI, optimization problems and systems of
nonlinear equalities consists of linear or nonlinear approximations of problem func-
tions. Our approach emphasizes that tensor approximations invariant under nonlinear
coordinate transformations (representations) may be essential. In this part, NLCS will
be reformulated as optimization problems defined on a Riemannian manifold, which
ensures a convenient structure to study theoretical and numerical questions.
First, see NLCS (2.1) which can be considered 2n O-covariant tensor field in-
equalities on M[hlJ. A well-known step to reformulate NLCS (2.1) into optimization
problems different from (5.1) is as follows:
only one is formed in (5.2) by using the operation' max' for all the functions, then we
lose the property of smoothness, but efficient techniques of global optimization may be
applied. To ensure the geodesic convexity property of the feasible region of (5.2), some
additional requirements are needed (Rapcsak, 1994a,1994b,1994c), thus to preserve
a more general and flexible treatment of the nonconvex character, a penalty func-
tion approach is introduced. To solve unconstrained optimization problems a general
framework for globally convergent variable metric methods along geodesics can be in-
troduced, containing a generalization of the steepest descent, Newton, Quasi-Newton,
gradient projection, reduced gradient, projected gradient, SQP methods, etc., for the
optimization problem given in the form of
where f(x,w) means a suitable penalty function with 1 penalty parameters for opti-
mization problem (5.2). A general variable metric method along geodesics for solving
(5.3) in the case of a fix value wEn proceeds as follows:
Starting from an initial feasible solution Xo, let Xk be the feasible solution of the
kth iteration step, Wk the penalty parameters at the kth iteration step, Dk an n x n
matrix which is positive definite on TM[hlJ",. (the tangent space of M[hlJ at Xk)
defining a linear map Dk : TMx. -+ TMx., G 1 the Riemannian metric of R.n , G the
induced Riemannian metric of M[h l ]; furthermore, Dk and the metric G commute at
the iteration points.
Step 1. Compute the direction Pk as follows: Pk = Dkfb(Xk,Wk)T, where
fb(Xk,Wk)T is the projected gradient of f with respect to the Riemannian metric
G 1 to the tangent space.
Step 2. Let Xk+! = f",.(tk,Pk), where f",.(tk,Pk) is the arc ofthe geodesic starting
from Xk with tangent Pk, and the stepsize tk is determined by an exact geodesic search
(5.4)
It is also a weak separation function for every pair of parameters. In the case of
convex optimization problems with unbounded feasible region, nonempty and bounded
solution set, the algorithm converges to an optimal solution, and moreover, if the
interior of the feasible domain is nonempty, then a pair of parameters can be selected
such that an f.-approximate solution is reached after one global minimization (Rapcsak,
1975). It follows from the proof that if the global minimums of this penalty function
exist for every pair of parameters, then they converge to a solution of problem (5.2).
By extending the f.-exactness property to our case, the combination of this result and
Theorem 5.1 should result in a globally convergent algorithm for determining an f-
approximate solution of problem (5.2).
6. CONCLUDING REMARKS
In the paper, the structure of smooth NLCS and a class of penalty methods are
studied by using tensors. Some advantages of this approach are as follows:
(1) an attempt for a unified and global description of theoretical and algorithmical
results of NLCS,
(2) the independence of the theoretical results from the imbed dings of the comple-
mentarity manifold and the representations of the Riemannian metrics and
(3) a challange to use a new view in NLCS.
[1] D.P. Bertsekas, "Constrained optimization and Lagrange multiplier methods", Academic Press,
New York, London, 1982.
[2] R.W. Cottle, "Nonlinear programs with positively bounded Jacobians", SIAM Journal on Ap-
plied Mathematics 14, 147-158, 1966.
[3] F. Forgo, "A method for solving nonlinear programming problems approximately", Szigma 1,
67-75, 1969. (in Hungarian)
[4] M. Fukushima, "Equivalent differentiable optimization problems and descent methods for asym-
metric Variational Inequalities", Mathematical Programming 53, 99-110, 1992.
[5] D. Gabay, "Minimizing a differentiable function over a differentiable manifold", Journal of Op-
timization Th.eory and Applications 37,177-219,1982.
[6] F. Giannessi, "Theorems of the alternative and optimality conditions" , Journal of Optimization
Theory and Applications 42, 331-365, 1984.
[7] F. Giannessi, "Separation of sets and gap functions for Quasi-Variational Inequalities". This
Volume.
[8] P.T. Harker, and J.S. Pang, "Finite-dimensional Variational Inequality and nonlinear comple-
mentarity problems: a survey of theory, algorithms and applications", Mathematical Program-
ming 48, 161-220, 1990.
[9] O.L. Mangasarian, "Equivalence of the complementarity problem to a system of nonlinear equa-
tions", SIAM Journal on Applied Mathematics 31, 89-92, 1976.
[10] O.L. Mangasarian, and M.V. Solodov, "Nonlinear complementarity as unconstrained and con-
strained minimization", Mathematical Programming 62, 277-297, 1993.
[11] J.S. Pang, and S.A. Gabriel, NE/SQP: "A robust algorithm for the nonlinear complementarity
problem", Mathematical Programming 60,295-337, 1993.
[12] M. Pappalardo, "Image problem approach to penalty methods" , Journal of Optimization Theory
and Applications 64, 141-152, 1990.
[13] T. Rapcsak, "An exterior point algorithms for solving convex nonlinear programming problems" ,
Alkalmazott Matematikai Lapok 1, 357-364, 1975. (in Hungarian)
[14] T. Rapcsak, "Minimum problems on differentiable manifolds", Optimization 20, 3-13, 1989.
[15] T. Rapcsak, and J. Szenthe, "On the connection between mechanical force equilibrium and
nonlinear programming", ZAMM 70,557-564, 1990.
[16] T. Rapcsak, "Geodesic convexity in nonlinear programming", Journal of Optimization Theory
and Applications 69, 169-183, 1991.
[17] T. Rapcsak, and T. Csendes, "Nonlinear coordinate transformation for unconstrained optimiza-
tion, II. Theoretical background", Journal of Global Optimization 3, 359-375, 1993.
[18] T. Rapcsak, "On the connectedness of the solution set to linear complementarity systems",
Journal of Optimization Theory and Applications 80, 501-512, 1994a.
[19] T. Rapcsak, "On the connectedness of the solution set to nonlinear complementarity systems",
Journal of Optimization Theory and Applications 81, 619-631, 1994b.
[20] T. Rapcsak, "Geodesic convexity on Rn", in: "Generalized convexity", Lecture Notes in Eco-
nomics and Mathematical Systems 405, Koml6si, S., Rapcsak, T. and Schaible, S. (eds.),
Springer-Verlag, Berlin, Heidelberg, New York, 91-103, 1994c.
[21) T. Rapc:Ak, _d T.T. Thang, "On nonlineu coordinate representations of smooth optimization
problems", J_raal oj Optimization Tlt.tw'll and Appliccatioru. To appear.
[22) M. Spivak, A comprehensive introduction to differential geometry I-V., Publish or Perish,
IDe. Berkeley, 1979.
BASIC EXISTENCE THEOREMS FOR GENERALIZED VARIATIONAL
AND QUASI-VARIATIONAL INEQUALITIES
Ricceri Biagio
Department of Mathematics
University of Catania
Viale A.Doria 6
95125 Catania, Italy
e-mail:[email protected]
Abstract. We deal with the following problem: given a Hausdorff real topological
vector space E, a closed convex set X ~ E, a multifunction F : X --+ 2x and a
multifunction ~ : X --+ 2E ' , find (x, 'P) E X x E* in such a way that
Our aim is to report those results on the subject in which the basic assumption on ~
is the following: for each y E X - X, the set
is closed. Particular emphasis is put on the case where F is constant as well as on the
(difficult) open problems.
We deal with the following problem: given a real Hausdorff topological vector
space E, a closed convex set X ~ E, a multifunction F : X --+ 2x and a multifunction
~: X --+ 2E ', find (x,'P) E X x E* in such a way that
This problem, as one says, is the generalized Quasi- Variational Inequality associated
with X, F,~. We denote it by GQVI(X, F, ~). It was first introduced, in a finite-
dimensional setting, by Chan and Pang in [3]. When F(x) = X (resp. I}(x) is a
251
252
singleton) for all x E X, our problem reduces to the generalized Variational Inequality
associated with X, ~ (GVI(X, ~» (resp. Qu~i- Variational Inequality associated to
X,F,~ (QVI(X,F,~))). Finally, when, at the same time, F(x) = X and ~(x) is a
singleton for all x E X, we have the classical Variational Inequality associated with
(X,~) (VI(X,~».
Most papers devoted to our problem (in its various specifications) involve some
suitable notion of monotonicity for the operator~. The work done in that direction
is mainly related to contributions by Browder, Brezis, Hess. There are, however, some
writings where the conditions imposed to ~ concern merely the kind of its continuity.
Here we deal with this latter category of papers.
To focus better the essential things in the development of the theory as well as in
the problems still open, I start with the case of Variational Inequalities.
It is well known that one of the basic results is Hartman-Stampacchia's theorem,
stated below (in a more general form) as Theorem 1.
Theorem 1 ([61, Lemma 3.1).- Assume that X is compact and finite-dimensional, and
that ~ is single-valued and weakly-star continuous. Then, VI(X, ~) has some solution.
A question naturally arises: what happens when dim(X) = 00 ? At that point,
the situation becomes definitively more composite and delicate. Continue to suppose
that X is compact and that ~ is single-valued and weakly-star continuous. Denote by
U the family of all finite-dimensional linear subspaces of E meeting X. Consider U as
a directed set, with the set-theoretic inclusion. For each S E U, thanks to Theorem 1,
we get a point x sEX n S such that
Since X is compact, the net {XS}SEU admits some cluster point in X, say Xo. How
can we show that Xo is a solution of VI(X, ~) ? To this end, let {x .. } be a sub-net of
{XS}SEU converging to Xo. Fix y E X. Then, from the obvious equality
and
..
lim(~(x .. ),xo - y) = (~(xo),xo - y) ,
it is enough to know that
lim(~(x .. ),x .. - xo)
01
= O.
To get this latter information, there are just two natural ways. Namely, either we
suppose more on the kind of continuity of ~, requiring that it is continuous with
respect to the topology, on E*, of uniform convergence on the compact subsets of E
253
([1), Theorem 6), or we suppose more on the topology of E, requiring that the weakly-
star compact subsets of E* are equicontinuous ([2), Proposition 15). This latter fact
happens, for instance, if E is a Baire space.
When we are out of the two above-mentioned circumstances, Theorem 1, in gen-
eral, is no longer true without the condition dim(X) < 00. In this connection, we have
the following very interesting result by M.Frasca and A.Villani:
Theorem 2 ([5], Theorem 2). - Let (E, h,) be any infinite-dimensional Hilbert
space. Then, for each closed ball X in E, there exists some {strongly} continuous
affine operator ~ : E -+ E such that, for every x EX, one has
sup(~(x),x - y) > O.
,EX
It is clear how Theorem 2 serves to our purpose: we consider E endowed with the
weak topology and identify E* with E. Thus, X is weakly compact and, at the same
time, ~ is weakly continuous, being affine and strongly continuous.
On the basis of what we have now seen, the following question arises: is there some
general case in which VI(X, ~) has solutions, though ~ is only weak-star continuous
and, at the same time, no restriction is made on the topology of E? An answer to
this question is provided, for instance, by the following result of mine:
Theorem 3 ([8]) - Assume that the relative interior of X {that is the interior of X in
its affine hull} is non-empty, and that ~ is weakly-star continuous. Moreorer, let K, Kl
be two non-empty compact subsets of X, with Kl ~ K and Kl finite-dimensional, such
that, for each x EX \ K, one has
Actually, Theorem 3 is a very particular case of the following more general result
which is also the basis of our further considerations.
Theorem 4 ([8], Theoreme 1). - Let X,K,K1 be as in Theorem 9, and let ~(x) be
convex and wealcly-star compact for all x E X. Moreover, assume that:
(a) for each y E X -X, the set {x EX: inft;>E+(x){<P,y) ::; O} is compactly closed (that
is, its intersection with any compact subset of E is closed);
(b) for each x EX \ K, one has
REFERENCES
[1] F.E. Browder, "The fixed point theory of multi-valued mappings in topological vector spaces",
Math. Ann., 177, 283-301, 1968.
[2] F.E. Browder and P.HESS, "Nonlinear mappings of monotone type in Banach spaces", J. Fund.
Anal., 11, 251-294, 1972.
[3] D. Chan and l.S.PANG, "The generalized quasi-variational inequality problem", Math. Oper.
Res., 7, 211-222, 1982.
[4] P. Cubiotti, "An existence theorem for generalized quasi-variational inequalities", Set- Valued
Anal., 1, 81-87, 1993.
[5] M. Frasca and A.Villani, "A property of infinite-dimensional Hilbert spaces", J. Math. Anal.
Appl., 139, 352-361, 1989.
[6] P. Hartman and G. Stampacchia, "On some nonlinear elliptic differential equations" , Acta Math.,
115, 161-220, 1966.
[7] W.K. Kim, "Remark on a generalized quasi-variational inequality", Proc. Amer. Math. Soc.,
103, 667-668, 1988.
[8] B. Ricceri, "Un theoreme d'existence pour les inequations variationnelles", C.R. Acad. Sci.
Paris, Serie 1,301, 885-888, 1985.
[9] N.D. Yen, "On an existence theorem for generalized quasi-variational inequalities", Set- Valued
Anal., to appear.
SENSITIVITY ANALYSIS OF VARIATIONAL INEQUALITIES
BY NORMAL-MAP TECHNIQUES
Robinson Stephen M.
Abstract: This paper deals with the sensitivity analysis of solutions of variational
inequalities by an implicit-function approach that makes use of so-called normal maps.
Although some of the results can be extended to infinite-dimensional spaces, the treat-
ment here is restricted to variational inequalities over polyhedral convex sets in finite-
dimensional spaces, the case for which the strongest results can be established. Cov-
erage includes transformation of the variational inequality to the equivalent form of a
normal map, basic facts about normal maps, derivatives, conditions for nonsingularity,
the implicit-function theorem, and applications to sensitivity analysis. Some examples
are included to illustrate the techniques.
1. INTRODUCTION
257
258
which the functions are not differentiable, and this is an active research area. For a
sample of the kinds of results that can be obtained in that case we refer to [6].
The rest of this paper is organized in three main sections. The next section introduces
variational inequalities and normal maps, and shows how either one may be transformed
into the other. Then we examine the normal manifold associated with the projector
onto a polyhedral convex set C. This normal manifold plays an important part in the
analysis of normal maps, as well as in computations involving them. In that section we
also look at affine maps and ask when the normal maps generated from them will have
certain convenient nonsingularity prope!ties. In the following section, we exploit this
nonsingularity to obtain the basic implicit-function theorem, and we show how to apply
this theorem to do sensitivity analysis. Following these main sections we provide a brief
conclusion, acknowledgments of support, and references.
Throughout the paper we use without comment terms and constructs of convex anal-
ysis. An excellent reference for these is the fundamental book of Rockafellar [15].
In this section we introduce variational inequalities and normal maps, and we show that
these are essentially equivalent devices, in that either one may be transformed into the
other. Variational inequalities are more familiar from applications, whereas the normal
map formalism will be more useful for the analysis in this paper.
We begin with the parametric variational inequality
Here C is a nonempty polyhedral convex set in IR" and f is a function from a product
(C n fl) X e to IRn;e is an open subset of a normed linear space P and fl is an open
subset of IRn with C n fl f. 0. For many applications P = IRk, but in general it is
convenient to let it be a normed linear space because the arguments go through and
because one sometimes needs the extra generality.
Of course, y might be fixed, in which case one has an ordinary variational inequality.
However, the setup in (2.1) permits us to change the form of f and then to observe
the effect of such changes on the solution point. Sensitivity analysis of this kind is the
focus of the present paper. We point out that in general there is no reason for C to be
polyhedral, but that we make that assumption here because it covers the cases in which
we are interested, and it makes possible much stronger results than would otherwise be
available.
The problem (2.1) includes a great many special cases, including linear and nonlinear
programming (the latter through the device of the Kojima function; see the discussion in
[13]), and various kinds of equilibrium problems including those posed over networks. It is
appropriate to note that the work of Stampacchia was decisive in the initial development
of this subject: the first two existence theorems for variational inequalities seem to have
been proved in [18] and [5] respectively, and an exposition of the state of the theory at
an early point in its development is in [19]. We do not go into detail here regarding the
various problems to which the formalism (2.1) can be applied, but we refer instead to
[4], in which numerous references are given.
259
As mentioned above, another formulation of the problem will be more useful for
analysis. Suppose that we denote by TIc the Euclidean projector on C, and for the
moment we consider a function f : C n n --+ R n , where n is as above (that is, we
temporarily suppress the variable y). Define a new function fe from TIc1(n) to lR,n by
Jc(z) = f(IIc(z» + (z - IIe(z». (2.2)
Thus, to evaluate fe at a point z we first project z onto C to obtain a point x, then
evaluate f at x, and finally add to f(x) the remainder z - x. Given the set C, we
call this function fe the normal map induced by f. The definition implies that for a
function f and a point a E R n one has (f + a)c = fe + a, and that for functions f
and g and a scalar A E [0,1] one has [(1 - A)f + Ag]c = (1 - A)fe + Age.
To understand why this function fe is connected to the Variational Inequality
problem, suppose first that z and x are points of R n such that
fe(z) = 0, x = TIe(z). (2.3)
If we denote by Ne(s) the normal cone of Cat s, defined to be empty if s 1. C and
to be the set of all points c· with (c·, c - s) :S 0 for each c E C otherwise, then as
o = fe(z) we have - f(x) = z - x. The right-hand side is an element of the normal
cone Ne(x), so we see that
z =x - f(x), and 'Vc E C, (f(x),c - x) 2': 0, (2.4)
and therefore x = TIe(z) solves the Variational Inequality problem defined by f and
C.
Conversely, if z and x satisfy (2.4), we obtain z - x = -f(x) E Ne(x), or x =
(I + Ne)-l(z).But (I + Ne)-l is the projector TIc, so we have x = TIe(z). This
implies
0= f(x) + (z - x) = f(TIe(z» + (z - TIe(z» = fe(z),
and therefore z and x satisfy (2.3). Accordingly, we can move back and forth between
(2.4) and (2.3) because they are completely equivalent.
°
Returning to (2.1), we see that to analyze its solutions x(y) we can solve instead
the normal-map equation f(', y)e(z) = to obtain a point z(y), then project z(y) onto
C with TIc to obtain x(y) satisfying (2.1). That is what we shall do in the remainder
of this paper.
An example might help to clarify these ideas. Suppose that we want to minimize
the Frechet differentiable function 1> over the polyhedral convex set C. The first-order
necessary optimality condition is then just (2.4), with f = d1>, the derivative of 1>.
Accordingly, the corresponding normal-map equation is
d1>e(z) = 0,
and if z is a solution of this equation then the corresponding first-order point is x =
TIe(z).
In this section we introduced the Variational Inequality (2.1) and the formalism
of normal maps. We further showed that any Variational Inequality is completely
equivalent to the corresponding normal map, so that we can use the more convenient
single-valued normal map for analysis. In the next section we discuss a geometric
construction associated with normal maps, whose properties are important for the
analysis we shall do later.
260
In this section we study a subdivision of the space R n that is closely associated with
the normal map construction. This subdivision has numerous interesting geometric
properties, but we shall only note a few of these because most are not required for our
purposes. Additional detail, and proofs, can be found in [13].
To begin, recall that the set C is polyhedral, so it has a finite number of faces.
Further, it is easy to show that on the relative interior of each such face, say F, the
normal cone Ne(x) takes a constant value, which we denote by NF. Let us write
(J'F = F + NFj this set (J'F is a polyhedral convex subset of RR having dimension n.
Note that on (J'F the projector lIe coincides with the affine map aF that projects points
onto the affine hull of F, so that on (J'F fe coincides with (f 0 aF) + I - aF, a map
that is smooth if f is smooth, and even affine if f is affine. Therefore fe is piecewise
smooth or piecewise affine respectively, and the (J'F are in general the maximal sets
with respect to which the smoothness or the affine property will hold for fe.
It is easy to show that the union of the (J'F over all faces F of C covers Rnj in fact,
if we consider the slightly smaller sets u' = (riF) + N F , these actually partition the
space Rn. We shall work with the (J'F because they are more convenient to deal with
(being closed)j further, the collection
forms a .nibdivided piecewi&e linear (PLJ manifold in the sense of [2]j this is the normal
manifold of C. We call the sets (J'F the n-cells of .Ne . Note the difference between the
normal manifold.Ne and the normal-cone operator Ne.
If we wish to apply the ordinary implicit-function theorem to a sufficiently differ-
entiable function, an important hypothesis is that an appropriate partial derivative be
nonsingular: this is equivalent to saying that it is a local homeomorphism. It will turn
out later that a precisely analogous condition holds for our piecewise smooth function
fej this condition requires that a certain normal map Ae induced by an affine trans-
formation A be a local homeomorphism. Therefore it is of great importance to have a
suitable condition to ensure this property.
It turns out that the required condition is conceptually simple and easy to state
(these adjectives do not, however, also apply to its proof!). In order to state it, we need
some terminology. First, we say that a piecewise affine map a from an n-dimensional
PL manifold M c R n to R n is coherently oriented if the determinants of the affine
transformations with which a coincides in the n-cells of M all have the same nonzero
sign. Second, for any point z E Rn we define the critical cone of C with respect to z
to be the polyhedral convex cone defined by
the common (n - I)-face separating them. Let al and a2 be the affine representatives
of a in these two cells. Choose a point Xo in the relative interior of H; by making
translations if necessary we can always assume that Xo = 0, so that H coincides near
the origin with a subspace L, and that al(O) = 0 = a2(0), so that al and a2 are
linear. Let v be any point in T2 but not in L, and let h2 , ••• , h n be a basis for L. By
nonsingularity ale v), al (h 2), . .. ,al(hn ) are linearly independent. Take the following
bases for R n , regarded as, respectively, the domain space and the image space of the
al:
With respect to these bases the matrix of al is the identity. If we define 0', f32, . .. ,f3n
by
n
Accordingly, the determinants of the matrices representing al and a2 will fail to have
the same sign exactly when 0' < O. In that case, the point Vi = O'V + Li=2 f3ihi belongs
to rl. But then (3.1) shows that
is irrelevant, because with finitely many cells in Nc any coherently oriented piecewise
affine map must be proper. Further, Ralph (10) has shown that for any polyhedral
set C, the branching number of Nc does not exceed 4. Therefore our map Ac will be
bijective if and only if it is coherently oriented. But any piecewise affine map on a PL
manifold with finitely many cells is Lipschitzian, so in this case bijectivity is equivalent
to the homeomorphism property that we are after. This proves our first assertion. A
different proof is given in (13). The second assertion is Theorem 5.2 of (13). 0
Although this theorem gives very precise conditions for the homeomorphism prop-
erties to hold, it is not necessarily the case that those conditions are easy to check. In
general, one has to evaluate the affine representatives in all of the cells of the normal
manifold and to check the signs of their determinants, which can be a lot of work. Of
course, sometimes it is possible to use structural information to make this job easier,
but that depends on the particular case at hand.
However, if the matrix of the transformation in question has a certain symmetry
property then the conditions of the theorem can be made much simpler. We now
describe this situation, giving results only for the global homeomorphism property
since by part (b) of Theorem 1 the local results for the original normal map are the
same as global results for a map based on the critical cone.
To introduce the symmetry property we need to make a certain reduction. We
describe this in the following proposition, which for simplicity we phrase in terms
of linear transformations; as noted above, if a( x) = Ax + a then we always have
ac = Ac + a, so there is no loss of generality in considering only the linear case.
The proposition uses the concept of Schur complement, defined as follows: if Mis
a matrix partitioned as
(1~~ 1~:)·
Then Ac is a homeomorphism of R n if and only if (i) ALL is nonsinguiar, and (ii)
(A/ALdD is a homeomorphism of Rn-j.
b. Suppose H is the affine hull of C, and define an affine transformation T :
H --+ H by T(h) = IIHA(h) for h E H. Then Ac is a homeomorphism of R n if and
only if Tc is a homeomorphism of H.
Proof. Part (a) is Proposition 4.1 of (13) and part (b) is a specialization of Propo-
sition 2.3 of [14]. 0
We use Proposition 1 by starting with an arbitrary polyhedral convex set C, then
using part (a) to reduce the homeomorphism question to one for a normal map over the
263
over the nonnegative orthant R!. Here I: can be any real parameter. We can express
this problem by writing the first-order necessary condition in the form
( -1 2)
where
A = 2 1 ' C=R!.
It is not hard to see that Zo = (-1,0). The critical cone of Cat Zo is then J{ = {O} xR+,
and to check whether Ac is a local homeomorphism at Zo we should see if AK is a
homeomorphism. The set K contains no lines, but it is not offull dimension. Therefore
we reduce the problem using part (b) of Proposition 1, using R 1 as a representation for
the affine hull {O} x R of K, and obtaining the trivial normal map (l)R+' The matrix
here is certainly positive definite, so the original Ac was a local homeomorphism.
In fact, for I: near zero the solution is
_ { Zo + (-21:, 1:) if I: ~ 0,
Z, - Zo + (0, 1:) otherwise,
and the stationary points are the projections of these z, onto C, namely
_ { (0, 1:) if I: ~ 0,
x, - (0, 0) otherwise,
so that both z, and x, are single-valued, continuous (but nonsmooth) functions of 1:.
The theory of this section showed us how to determine if solutions like the z, and
x, of the last example existed, were unique, and were continuous functions of f. In
the next section we see that this theory also helps us to establish a general implicit-
function theorem that will predict the existence, uniqueness, and (Lipschitz) continuity
of solutions for normal maps induced by differentiable nonlinear functions, rather than
just the affine functions we have been considering here. In addition, we shall see that
the affine theory provides a powerful computational tool for locally approximating the
solutions of the nonlinear problems.
In this section we return to the Variational Inequality (2.1), and using the tools
we have already developed we rewrite it in the form
Because of the equivalence of (2.3) and (2.4), a solution z(y) of (4.1) yields a solution
x(y) of (2.1) via the projection operator: x(y) = IIc(z(y)). Conversely, any solution
x(y) of (2.1) is derived via projection from the solution z(y) = x(y) - f(x(y), y) of
(4.1). So we need only be concerned with solving (4.1). We shall give a version of the
implicit-function theorem that applies to that equation, then a reformulated corollary
that applies directly to (2.1).
265
In stating the first result we need the concept of B-derivativej a function 9 from
an open subset r of R m to Rq is said to be B-differentiable (Bouligand differentiable)
at Xo E r if there is a positively homogeneous function dg(xo) : R m -+ Rq (the B-
derivative of 9 at xo) having the property that
One expects implicit-function theorems to involve derivatives, and this map is going
to play the part of a derivative for our purposes.
Theorem 3 Let C, fl, e, and f be as in Section 2. Let CIi = IIc1 (fl) and let Zo E CIi
and Yo E e with fh yo)c(zo) = 0. Suppose that for some 8 and each x E fl, f(x,.) is
Lipschitzian on e with modulus 8. Write K = K(zo) and Xo = IIc(zo), and define L
by (./.2). Assume the following hypotheses:
a. The partial derivative dxf(xo, Yo) is strong.
b. LK is a homeomorphism.
Then 6(LK' RR) > 0, and for each ,\ > 6(LK' RRt 18 there exist neighborhoods Z
of Zo, X of Xo, and Y of Yo, and a function z : Y -+ RR, such that:
a. z(Yo)= zoo
b. z is Lipschitzian with modulus ,\.
c. For each y E Y, z(y) is the unique point in Z satisfying (4.1), and IIc(z(y))
is the unique point in en X satisfying (2.1).
place of the function f appearing in Theorem 3.2 of (12) (not to be confused with
our present I). Second, we must prove the claim that 6(LK' Rn) > o. In fact, the
second point is easy, because we have assumed that LK is a homeomorphism, so that
its inverse is well defined on all of RR. As the inverse, like L K , is affine in each of the
finitely many cells of JVK, it must be Lipschitzian. Then the reciprocal of the Lipschitz
constant provides a positive lower bound for 6(LK,RR).
To prove the first assertion, we note that Proposition 4.1 of (12) shows that a strong
approximation of fO(·,y)(z) in z at (zo, Yo) is given by
d",f(xo, Yo)[I1c(z) - xo] + [(z - zo) - (I1c(z) - xo)]. (4.3)
All we need to do is to show that in our case (4.3) is identical to LK(Z - zo) for z
near zoo We first note that as C is polyhedral, by Lemma 5(i) of [9] we have for all
sufficiently small hERR the equation
I1c{zo + h) = I1c(zo) + I1K{h). (4.4)
Therefore we can rewrite (4.3) as
d,.f(xo, Yo)[IIK(z - zo») + [(z - zo) - IIK(z - zo»),
which is exactly LK(Z - zo). This completes the proof. 0
At this point, we know that under the hypotheses of Theorem 3 there is a locally
unique, Lipschitzian function giving a solution of (4.1), or equivalently of (2.1), for
each y near Yo. That is, we have identified conditions under which slightly perturbed
problems will have solutions that behave reasonably nicely as functions of the per-
turbations. Our final result will give a method of approximating those solutions by
solving linearized Variational Inequalities or normal maps that are, in general, much
simpler to deal with than are (2.1) or (4.1) respectively. The idea is simply to solve an
approximate problem to obtain an approximate solution, and the approximate problem
is, just as in the classical implicit-function theorem, constructed by using derivatives
of the nonlinear function of the original problem.
Theorem 4 Assume the notation and hypotheses of Theorem 9, and suppose in ad-
dition that f has a partial B-derivative dyf(xo, yo) with respect to y at (xo, yo). Then
the functions z(y) and xCV) of Theorem 9 are B-differentiable at Yo with
(4.5)
and
= IlK 0 dz{yo).
dx{yo) (4.6)
Proof. We know from Theorem 3 that xCV) = I1c(z{y». Therefore if we prove (4.5)
then (4.6) will follow from the chain rule for B-derivatives (e.g., Corollary A.4 of [11]),
since (4.4) implies in particular that dIlc(zo) = I1K(zo) = IlK.
To prove (4.5) we can use Part (a) of Theorem 3.3 in [12], provided that we show
that d,J(xo,yo)(·) is the partial B-derivative in y of fc(·,Y)(z) at the point (zo,Yo) (in
fact, it is then an F-derivative because it is linear). To do this we just write
fc(-,yo + k)(zo) = f(xo, Yo + k) + (zo - xo)
= f(xo, Yo) + dyf(xo, yo)(k) + o(k) + (zo - xo)
= fO(·, Yo){zo) + dyf{xo, Yo)(k) + o(k).
267
and that dX(Yo)(k) is the projection on K of dz(Yo)(k), then we see that w = dx(Yo)(k)
must solve the linear Variational Inequality problem of finding w E K such that
further, under the assumptions of Theorems 3 and 4, (4.7) will have a unique solution
for each kERn. So we have reduced the question of sensitivity analysis of the nonlin-
ear Variational Inequality (2.1) to that of computing solutions of the linear problem
(4.7). At least in the finite-dimensional case, (4.7) is frequently easy to solve. Further,
the possibility of computing dz(yo) and dx(yo) leads applications other than sensitiv-
ity analysis; for example, Sellami [16] has recently used it in developing a homotopy
method of predictor-corrector type for solving nonlinear Variational Inequalities when
accurate starting points are not available.
Theorem 4 completes our mathematical coverage of sensitivity analysis for Varia-
tional Inequalities. We conclude this section with an example illustrating the use of
Theorems 3 and 4.
Consider finding stationary points for the problem of minimizing (in x)
°
We can see that for Xo = and EO = 0, d",if>(xo, Eo) = (-1,0,1), so that the origin is
a stationary point (in fact it must be a minimizer, because for E = the objective °
function is convex). We write f for d",if> and take
= Xo - f(xo, Eo) = (1
zo -1), °
and then we have fc(zo, Eo) = o. The tangent cone of C at the origin is C itself, so the
critical cone K is the intersection of C with the hyperplane consisting of all X E R3
with Xl = X3. Therefore
We have
f(x, E) = (-1 X2 - E -EX~ + X3 + 1) ,
so that
to check that d.J(xo,f{J) is positive definite on the subspace M comprising the affine
hull of K. The two vectors
U~~)o~nun~on.
which is certainly positive definite. Therefore we know that our theoretical results
apply, so for E near zero the locally unique stationary points comprise a Lipschitzian
function of E.
Moreover, we can obtain a first-order approximation to the curve of stationary
points by solving the linear problem (4.7) over our critical cone K. Here we have
5. CONCLUSION
REFERENCES
[1] R. W. Cottle, "Manifestations of the Schur complement" , Linear Algebra and Its ApplicatioIls 8,
189-211, 1974.
269
[2] B. C. Eaves, "A short course in solving equations with PL homotopies" ,in: "Nonlinear Program-
ming, SIAM-AMS Proceedings Vol 9," R. W. Cottle and C. E. Lemke, eds., American Mathemat-
ical Society, Providence, RI 1976.
[3] A. V. Fiacco, "Sensitivity analysis for nonlinear programming using penalty methods", Mathe-
matical Programming 10,287-311, 1976.
[4] P. T. Harker and J. S. Pang, "Finite-dimensional Variational Inequality and nonlinear complemen-
tarity problems: A survey of theory, algorithms and applications", Mathematical Programming
48, 161-220, 1990.
[5] P. Hartman and G. Stampacchia, "On some nonlinear elliptic differential functional equations",
Acta Mathematica 115, 153-188, 1966.
[6] A. J. King and R. T. RockafelIar, "Sensitivity analysis for nonsmooth generalized equations",
Mathematical Programming 55, 193-212, 1992.
[7] M. Kojima and R. Saigal, "A study of PC I homeomorphisms on subdivided polyhedrons" , SIAM
Journal on Mathematical Analysis 10,1299-1312,1979.
[8] D. Kuhn and R. Liiwen, "Piecewise affine bijections of R n , and the equation Sx+ - Tx- = y",
Linear Algebra and Its Applications 96, 109-129,1987.
[9] J. S. Pang, "Newton's method for B-differentiable equations", Mathematics of Operations Research
15,311-341,1990.
[10] D. Ralph, "On branching numbers of normal manifolds", Nonlinear Analysis: Theory, Methods
and Applications 22, 1041-1050,1994.
[11] S. M. Robinson, "Local structure offeasible sets in nonlinear programming, Part III: Stability and
sensitivity", Mathematical Programming Study 30,45-66, 1987.
[12] S. M. Robinson, "An implicit-function theorem for a class of nonsmooth functions", Mathematics
of Operations Research 16, 292-309, 1991.
[13] S. M. Robinson, "Normal maps induced by linear transformations", Mathematics of Operations
Research 17691-714, 1992.
[14] S. M. Robinson, "Nonsingularity and symmetry for linear normal maps", Mathematical Program-
ming 62,415-425, 1993.
[15] R. T. RockafelIar, "Convex Analysis," Princeton University Press, Princeton, NJ, 1970.
[16] H. Sellami, "A Continuation Method for Normal Maps," Ph.D. Dissertation, Departments of
Mathematics and Industrial Engineering, University of Wisconsin-Madison, Madison, WI, 1994.
[17] G. Stampacchia, "Variational inequalities", in: "Theory and Applications of Monotone Operators,
Proceedings of the NATO Advanced Study Institute, Venice, Italy," Edizioni Oderisi, Gubbio,
Italy, 1968.
MONOTONE RELATIONS AND NETWORK EQUILIBRIUM
Rockafellar R. Tyrrell
1. INTRODUCTION
271
272
An alternative form, which is preferable for many reasons and will be especially fruitful
in what follows, is obtained by utilizing the notion of the normal cone Nz(z) to Z at
z, which in convex analysis consists of all vectors w such that (w, z - z) :::; 0 for all
z E Z (see [1]). The problem is then:
In either form the inspiration comes from the case where F is the gradient mapping
V' f associated with a continuously differentiable function f defined on a neighborhood
of Z. The Variational Inequality then expresses the first-order necessary condition for
optimality in the minimization of f over Z, this being not just necessary but sufficient
when f happens to be convex.
Variational inequalities in which F = V' f are usually called $ymmetric, whereas
all others are a$ymmetric. Really, this terminology is appropriate only under the
additional assumption that F is continuously differentiable, since that allows the
existence of a function f with F = V' f to be identified with the property of F that
the Jacobian matrix V' F( z) is symmetric everywhere. It is possible, of course, to have
F = V' f without F being differentiable at all, so that the Jacobian does not even exist.
On the other hand, in circumstances where F is differentiable but not continuously
differentiable the Jacobian might exist but not be symmetric, as indicated by classical
examples of twice differentiable functions f for which the matrix of second partial
derivatives is not symmetric.
The distinction between the symmetric and asymmetric cases is sometimes inter-
preted as marking the division between the Variational Inequality problems directly
reducible to optimization and the ones not so reducible. But this view is inaccurate
and potentially misleading. Variational inequalities can correspond to optimization
despite asymmetry, and whenever that is true they can be solved by optimization
techniques just as well as if they were symmetric, and without resorting to the intro-
duction of an artificial "gap" function.
For example, the Kuhn-Tucker conditions for a minimization problem with func-
tional constraints express first-order optimality in terms of an asymmetric Variational
Inequality in the primal and dual variables jointly. Suppose the problem consists of
minimizing fo(x) over all x E X satisfying (!I(x), ... ,fm(x)) E K, where K is a
273
closed, convex cone in JR m , X is a nonempty, closed, convex set in JR n , and the func-
tions Ii are continuously differentiable. Let lex, y) stand for the Lagrangian expression
Io(x) + L~l Ydi(X). The generalized Kuhn-Tucker conditions that apply to this set-
ting, as established in [2] (Thms. 4.2, 10.6) under a basic constraint qualification, say
that for x E X to give a local minimum there must be a vector y E Y, where Y is the
cone polar to K, such that
The search for a pair (x, y) satisfying this double relation is the Variational Inequality
problem (VI') in the case of
The Variational Inequality is asymmetric because F is not actually the gradient map-
ping of any function. Indeed, when F is continuously differentiable (through the
functions J; being twice continuously differentiable) its Jacobian matrix has an obvi-
ous lack of symmetry:
(We denote by (.,.) the canonical inner product.) When F is continuously differ-
entiable this property is equivalent to the positive semidefiniteness of the (possibly
asymmetric) Jacobian matrix V' F( z) at every point z E Z. In the symmetric case
with F = V' f it corresponds to f being convex. The Variational Inequality then
describes the solution(s) to a problem of minimizing a convex function over a convex
set.
An important asymmetric example of monotonicity is encountered when F has the
form (2.2) relative to a product of convex sets X and Y, and the function I(x, y) is
convex in x E X and concave in y E Y. Such a Variational Inequality corresponds to
convex optimization as well. It characterizes solutions x to a certain primal problem of
minimization by means of a saddle point (x, jj), where jj solves a certain dual problem
of maximization.
274
As valuable as the notion of a Variational Inequality has turned out to be, it has
definite limitations which need to be appreciated if connections with optimization are
fully to be understood. One limitation, which fortunately is easy to get around, is the
single-valuedness of the mapping F. On the surface, this excludes applications to areas
like nonsmooth optimization. A more serious limitation, however, is the requirement
that the set Z be convex. When a Variational Inequality problem is stated in the
form (VI), the convexity of Z is essential for it to make good sense, but in form (VI ')
the way is open to assigning to the normal cone Nz(z) a definition appropriate not
only for convex sets Z but nonconvex sets as well. For instance, N z( z) can be taken
to be the Clarke normal cone or the smaller cone that has received special emphasis
in the work of Mordukhovich (see [2], Sect. 10).
The point is that although Variational Inequalities in which Z is convex do cover
some problems of nonconvex optimization through extended Kuhn-Tucker conditions,
as already discussed, there is something rather strained about the formulation. The
case where Z is convex and F is monotone is natural in providing a platform for a
theory of Variational Inequalities that mirrors convex optimization. The case where
Z is potentially not convex and F not monotone is well motivated too, if interpreted
in the manner just described. But the hybrid case where Z is convex, yet F is not
monotone, draws boundaries rather artificially.
Still another way of stating the basic Variational Inequality problem, which will
serve as a guide in our discussion of network equilibrium, is:
At first this format may seem unappealing because it requires working with a set-
valued mapping T. As support for an alternative point of view, however, it is rich in
theoretical implications.
A general mapping T that assigns to each z E R,N a subset T(z) C JRN can be
regarded as an ordinary single-valued mapping from JRN to the space 2RN. For most
purposes, though, there is much more to be gained by identifying T with the set
as its "graph" and thinking of it thus as expressing a relation between vectors z and
w. The "effective domain" dom T and "effective range" rge T of T are defined then
by
domT:= {z 1 T(z) I- 0}, rgeT:= {w 13 z, wE T(z)}. (2.5)
In this framework, which we signal this framework by writing T : JRN =t R,N in place
of T : nN -+ n N, T is regarded as single-valued, empty-valued or multivalued at z
according to whether T(z) is a singleton, the empty set, or a set with more than one
element. The "inverse" of T is the mapping T- 1 : JRN =t JRN defined by
3. NETWORK EQUILIBRIUM
Classical network theory is concerned with scalar-valued flows, but here we will
be occupied with vector-valued flows. A d-dimensional flow x is a "supervector"
(Xl, ... , xn) where each component Xj is a vector (Xjl, ... , Xjd) E JRd. In applications,
X jk will represent the amount of scalar flow of type k in arc j. Constraints on the
magnitude and direction of such flow amounts may be imposed later, but for now we
note merely that a quantity Xjk > 0 is to be interpreted as flowing from the initial
node of arc j to its terminal node, whereas a quantity Xjk < 0 refers to physical flow
in the opposite direction.
The divergence of the flow X at node i is the vector Yi = (Yil, .. . ,Yid) E JRd in
which Yik gives the net amount of flow type k that originates at node i. This is
expressed by
where y is the supervector (Yl, ... ,Ym). Node i is a source for flow type k under x if
Yik > 0 and a sink if Yik < o. Flow type k is conserved at node i if Yki = o.
Dual to the concept of flow is that of potential. Ad-dimensional potent'jul U
is a supervector (Ub ... , urn), each component of which designates a vector Ui =
(Uil, ... , Uid). The quantity Uik refers to the potential of type k at node i, and abstract
quantity which in economic applications may have a price interpretation. Relative to
such a vector-valued potential u, the tension Vj in arc j is the difference Ui' - Ui, where
276
i is the initial node of arc j and i' is the terminal node. In terms of the incidence
matrix E this comes out as
where v = (VI, ... ,vn ). Each tension vector Vj = (Vjl,' .. ,Vjd) E IRd has components
Vjk giving the difference in potential type k in arc j.
Equilibrium problems in this context can usefully be set up on several levels.
To begin with, we consider the case of fixed supplies and demands. By a supply
b = (bI, ... ,bm ) in the network we will mean the assignment to each node i of a vector
bi = (bi1 , ... , bid), where bik designates the supply of flow type k at node i, this being
the value that the divergence Yki will be required to have. Negative supply values bik
correspond of course to demand. A value bik = 0 indicates that flow type k is required
to be conserved at node i.
By a flow-tension relation in arc j we will mean a subset of IRd x IRd specifying
the flow-tension pairs (xj,Vj) permitted to coexist in arc j. We interpret this subset
as the graph gphRj of a mapping R j : IRd ~ IRdj thus, Xj and Vj are related in the
required manner if and only if Vj E Rj(xj), or equivalently Xj E R j l(Vj).
The classical analogy for scalar-valued flows in an electrical network lies with
resistance and conductance. In such a network each arc j represents an electrical
component with a certain "characteristic curve" which describes how the flow (elec-
trical current) through j corresponds to the tension (voltage difference) across j. This
characteristic curve is the graph of R j , and the R j is "resistance mapping" for the arc
j j the inverse Rjl is the "conductance mapping" for the arc.
If arc j represents an ideal resistor, behaving in accordance with Ohm's Law with
resistance value r j > 0, its characteristic curve is a line in IR x IR with slope r j.
Then both R j and Rt are single-valued and linear. Nonlinear resistors correspond
to more complicated curves in IR x IR. Sometimes R j or Rjl, or both, can fail to be
single-valued in such a context. For instance, in the case of an ideal diode, the graph
of R j is the subset of IR x IR formed by the union of the nonnegative xj-axis and the
nonpositive vraxis.
Equilibrium Problem 1 Given for each arc j a mapping R j : IRd ~ IRd and
for each node i a supply vector bi E IR d, find a flow x and a potential U for which the
corresponding divergence Y and tension v satisfy
A more general formulation of equilibrium dispenses with fixed supplies and de-
mands and instead allows a divergence-potential relation to be assigned to each node.
Again, we think of such a relation in terms of a subset of IRd x IRd viewed as the
graph G(Si) of a mapping Si : IRd ~ IRd. The divergence Yi and potential Ui at node
i are related in the required manner when Ui E Si(Yi), or equivalently, Yi E Si- 1( Ui).
In Equilibrium Problem 1, the graph of Si is the set {bi} X IRd for every node ij in
other words, we have
But instead now, Si- 1 might for instance be a nonconstant, single-valued mapping.
An economic interpretation in some models where the kinds of flow represent different
277
In applications to traffic equilibrium, the following model is basic. For each node i
consider a supply vector bi = (bib ... , bid). For each arc j let Xj = [0, eilJ x··· X [0, eNJ,
where eik is the upper bound for flow of type k in the arc in question, ejk ~ O. (When
ejk = 0, flow of type k is forbidden in this arc.) For Xj E X j let
for a continuous, nondecreasing function Cj : [0,00) -+ [0,00) and fixed weights Wjk ~
o. The conditions for traffic equilibrium are taken then to be those of Equilibrium
Problem 1 with
In models of this kind it is common to have only one source and one sink for each
type of traffic. Then for each k there is exactly one node i with bik > 0 and exactly
one other node i' with bilk < O. Often the models are set up in terms of flows along
particular paths instead of just flow amounts in each arc. Such models are much more
complicated to work with, yet they seem not to offer any serious advantages, because
the flow of traffic of type k can readily be represented, at any stage of computation
or analysis where desired, as a sum of flows along paths from source to sink. See
Rockafellar [5], Secs. 4A and 4B, for the elementary algorithm that is involved.
Generalized traffic models in the format of Equilibrium Problem 2 instead of Equi-
librium Problem 1 might arise from situations in which the supply and demand for the
278
different kinds of flow could be affected by the state of congestion. Models attempting
to treat the difficulties of passing through various nodes would not necessarily require
passage to Equilibrium Problem 2. Instead one might use the device of introducing
"internal arcs" in such nodes; the needed equilibrium conditions could then be cen-
tered on such arcs, see [5], Sec. 3L. In either approach, dynamical networks could be
formulated in a space-time framework to handle traffic equilibrium in the sense of a
day-to-day cycle; see [5], Secs. 1H and 3L.
(4.2)
(4.3)
This Variational Inequality is monotone when every Fj and Wi is monotone; then one
has a maximal monotone version of Equilibrium Problem 2.
Yo are polyhedral. Since ri L = L we see that the assumption in (a) about a certain
How x corresponds to the hypothesis of Theorem 3 when applied to the intersection
Z = L n X~ n ... n X~ n Y{ n ... n Y':'. We deduce thereby that
Nz(z) = NL(Z) + NXj(z) + ... + Nx:.(z) + Ny;(z) + ... + Ny,:. (z).
Here NL{Z) = LJ. = {(v,u) I v = -ETu}, whereas
NXj{z) + ... + Nx:.{z) + Ny;(z) + ... + Ny,:.{z)
= Nx,(xd X ••• x NXn(xn) X Ny,(Yl) x··· X NYn(Yn) = Nzo(z).
Thus, Nz(z) = LJ. + Nzo(z).
The Variational Inequality for Z and F, expressed in form (V!'), refers therefore
to the existence of z E L n Zo such that there exists 'Iii E LJ. with 'Iii - F(z) E Nzo{z).
To say that z E L n Zo is to say that z = (x,y) with y = Ex, Xj E Xj and y E Yo. To
say that 'Iii E LJ. with 'Iii - F{z) E Nzo(z) is to say that w = (ii, u) with ii = -ETu,
Vj - Fj(xj) E NxJ(xj) and Ui - Gi(Yi) E Ny,(Yi) for all arcs j and nodes i. From (4.1)
and (4.3) we conclude that the Variational Inequality comes down to the equilibrium
conditions in Equilibrium Problem 2.
When all the mappings F j and Gi in (a) are monotone, F is obviously monotone
as well. Then too, every relation Rj and Si is maximal monotone by Theorem 1, so
we have a maximal monotone version of Equilibrium Problem 2.
In case (b) the argument is closely parallel. We have Z = L n Zo for the subspace
L = {{v,u) I v = -ETu} and set Zo = 'Vi x ... x Vn X Ul X ••• X Urn. Again
through Theorem 3, the assumption about a potential u guarantees that Nz(z) =
NL(z) + Nzo{z) with NL(z) = LJ. = {(x,y) I Y = Ex} and Nzo(z) = NV,{Vl) x ... x
Nvn{vn ) x Nu,{ut} x··· X Num(u m ). The specified Variational Inequality in form (VI')
reduces then to the existence of z = (v,u) ELand w = (x,y) E LJ. such that iij E Vi
and Xj - CIi;(vj) E NvJ(vj) for all arcs j, while Ui E Ui and Yi -Wi(iii) E Nu;(iii) for all
nodes i. Because of (4.2) and (4.4), these conditions are identical to Xj E Rjl(Vj) and
Yi E Si- l (iii)' which are just another way of writing the ones in Equilibrium Problem 2.
When all the mappings CIi j and Wi in (b) are monotone, F is monotone too, and
through Theorem 1 the relations Rjl and Si- 1 are maximal monotone. Then R j and
Si are maximal monotone and we have a maximal monotone version of Equilibrium
Problem 2.
Case (c) is simpler and does not require Theorem 3. Without having to invoke
any constraint qualification we know that
Nz(z) = Nx,(xt} X .•• X NXn(x n) x NU,(Ul) X ••• X Num(u m ).
Consider z = (x, u) and let y = Ex and v = -ETu. To say that z E Z is to say
that Xj E Xj for all j and Ui E Ui for all i. The condition -F(z) E Nz(z) takes the
form then that -Fj(xj)+Vj E NxJ(xj) and -Wi(Ui)+Yi E Nu;(iii) for all i andj. By
(4.1) and (4.4) these properties are equivalent to having iij E Rj(xj) and Yi E S-I{Ui),
which are the same as the equilibrium conditions in Equilibrium Problem 2.
When Fj and Wi are monotone, the mapping
(x,u) 1-+ (Fl(Xl),'" ,Fn(xn), Wl(Ul),"" wm(um »
in (c) is monotone. The linear mapping (x,u) 1-+ (ETu,-Ex) is always monotone
(because its matrix is antisymmetric). Then F, as the sum of two monotone mappings,
is itself monotone. In this case the Variational Inequality in (c) is monotone. At the
same time the mappings Rj in (4.1) and Si- l in (4.4) are maximal monotone by virtue
of Theorem 1, so we have a maximal monotone version of Equilibrium Problem 2. 0
281
This Variational Inequality is monotone when every <flj is monotone; then one has a
maximal monotone version of Equilibrium Problem 1.
(c) (primal-dual case). In Equilibrium Problem 1, suppose all the flow-tension
relations are of primal VI-type: (4.1) holds. The problem is equivalent then to sol'ving
the Variational Inequality for
This Variational Inequality is monotone when every Fj is monotone; then one has a
maximal monotone version of Equilibrium Problem 1.
set. The "effective domain" domf := {z I fez) < oo} of such a function is then a
convex set in particular.
A convex function f on R,N is proper if fez) > -00 for all z, and fez) < 00 for
at least one z. It is lower Ilemicontinuous (lsc) if the set epif is closed in JRN x JR.
The function j* conjugate to f is defined by
When f is convex, proper and lsc, the conjugate function j* likewise is convex, proper
and lsc, and the function j** conjugate to j* is in turn f:
The set of these is denoted by af(z). Thus af denotes a mapping JRN =f JRN in the
general sense adopted in Section 2. Furthermore, the inverse mapping (al)-1 is the
subgradient mapping associated with the conjugate function j*:
Primal Problem 1 Given for each arc j a convex, proper, lsc function fi on JRd
and for each node i a supply vector bi E JRd, minimize 1:jeJ fi(Xj) over all flows x
with divergence Y satisfying Yi = bi for all i E I.
Primal Problem 2 Given for each arc j a convex, proper, lsc function fi on JRd
and for each node i a convex, proper, lsc function gi on JRd, minimize 1:jEJ fi(Xj) +
1:iEI gi(Yi) over all flows x, where Y is the divergence of x.
It should be kept in mind that these problems have implicit constraints represented
through 00. In Primal Problem 1, a flow x with divergence y = b is not regarded as
feasible unless I;(Xj) < 00 for all j, i.e., Xj belongs to the convex set dom I; C JRd for
all j. In Primal Problem 2, there is the further requirement that Yi should belong to
the convex set domgi for all i. Primal Problem 2 reduces to Primal Problem 1 when
With these primal problems we associate the following dual problems in terms of
the conjugate convex functions.
283
Dual Problem 1 Given for each arc j a convex, proper, lsc function Ii on JRd
and for each node i a supply vector bi E JRd, maximize - LjEJ f;(vj) - LiEI -(bi, Ui)
over all potentials U and their tensions v.
Dual Problem 2 Given for each arc j a convex, proper, lse function fj on JRd
and for each node i a convex, proper, lsc function gi on JRd, maximize - LjEJ f;( Vj)-
LiEf g:( Ui) over all potentials U and their tensions v.
By the optimal values in these primal and dual problems we will mean the values
(in JR) giving the infimum or supremum in each case. The optimal solutions are the
elements (flows or potentials) for which these values are achieved, if any.
Finally, we formulate saddle point problems of Lagrangian type corresponding to
the optimization problems on both levels.
Saddle Problem 1 Given for each arc j a convex, proper, lsc function fj on JRd
and for each node i a supply vector bi E JRd, find a saddle point of the Lagrangian
function
L(x, u) = 'Lfi(Xj) - 'L(bi, Ui) + 'L
eij(ui, Xj)
jEJ iEI jEJ, iEI
with respect to minimizing over flows x but maximizing over potentials u.
Saddle Problem 2 Given for each arc j a convex, proper, lsc function Ii on JRd
and for each node i a convex, proper, lsc function gi on JRd, find a saddle point of the
Lagrangian function
Proof. The condition for (x,u) to be a saddle point in (b) is that the expression
L(x, ii) should achieve its minimum over all flows x at x, whereas L(x, u) should
achieve its maximum over all potentials u at ii. The minimization part is equivalent
to having f;(xj) - (Vj,Xj) achieve its minimum over Xj E lR,d at Xj for each j E J,
where Vj = - EiEleijiii. This means that Vj E 8fj(xj) for all j. The maximization
part is equivalent to having (Yi,Ui) - g;(Ui) achieve its maximum over Ui E lR,d at
Ui for each i E I, where iii = EjEJ eijXj. This means that iii E 8g;(iii) for all i, or
since 8g: = (8git\ that iii E 8gi(Yi). These subgradient conditions are the same as
Vj E Rj(xj) and Ui E Si(Yi) under our hypothesis. Thus, (a) and (b) are equivalent.
To establish the equivalence of (b) with (c), let r( x) = supu L( X, '11.) and s( '11.) =
inf",L(x,u). It is elementary and well known in general minimax theory (cf. [1],
Sec. 36) that (x,u) furnishes a saddle point of L if and only if x minimizes rex),
ii maximizes s(u), and the minimum value rex) agrees with the maximum value
s(u). We merely need to observe now that rex) = EjEJ I;(Xj) + EiEI gi(Yi) (with Yi
standing for EjEJ eijXj), whereas s(u) = - EjEJ fi(Vj) - EiEI g;(Ui) (with Vj standing
for - EiEI eijUi). 0
Theorem 9 Either one of the following assumptions suffices to guarantee that the
optimal values in Primal Problem 2 and Dual Problem 2 are equal:
(a) There is a flow x with divergence y such that xjEri ( dom 1;) for all arcs
j E J, and Yi E riC domgi) for all nodes i E f.
(b) There is a potential u with tension v such that Vj E ri ( dom fn for all arcs
j E J, and Ui E riC domgt) for all nodes i E I.
Proof. This follows from the Fenchel duality theorem in convex optimization,
specifically as a case of [1], Cor. 31.2.1. 0
The constraint qualifications in Theorem 9 could be refined along the lines of the
one in Theorem 3. Recall that a convex function is polyhedral when its epigraph is a
polyhedral set; then the conjugate function is polyhedral as well; see [1], Sees. 19 and
20. Theorem 10 remains valid if in (a) the condition Xj E ri (domJ;) is weakened to
ifj E dom J; when J; is polyhedral, while the condition iii E ri ( dom gi) is weakened to
iii E domgi when gi is polyhedral; similarly in (b) and Theorem 10. This refinement
can be verified on the basis of the polyhedral results in [1], Theorem 31.1, by splitting
up the polyhedral and nonpolyhedral parts of the problem in a suitable way.
For scalar-valued flows and potentials, "ri" can be dropped entirely from Theo-
rems 9 and 10-see [5], Sec. 8H.
An example of particular importance is the one where the functions J; have the
form
r(X') _ {<pj(Xj), if Xj E Xj, (5.4)
J J - 00
, ifx·dX·
J y:. J'
with X j a nonempty, closed, convex subset of JRd and <Pj a differentiable convex
function defined on a open set containing Xj. Then
In other words, the relation Rj = of; is of primal VI-type (4.1) with Fj = V<pj (this
mapping being continuous, because any differentiable convex function is continuously
differentiable). Then Primal Problem 1 corresponds to minimizing LjEJ <PiC x j) subject
to Xj E Xj and Yi = bi.
Other cases, where Rj is of dual VI-type, or where Si is of primal or dual VI-type,
can be identified in like manner. This reveals the extent to which the Variational
Inequalities developed for Equilibrium Problems 1 and 2 in Theorems 4 and 5 can
be viewed as coming from optimization with convexity. It should be noted that the
primal-dual Variational Inequalities correspond to the saddle point problems.
Because of this property of dom T and rge T we can speak of the relative interiors
ri ( dom T) and ri ( rge T), these being the same as the relative interiors of the convex
sets C and D in Theorem 11. When T = af for a convex, proper, lsc function
f : R,N -+ R, (cf. Theorem 6), one has ri ( dom T) = ri ( dom f) and ri ( rge T) =
riC domj*), see [1], Thm. 23.4.
This result holds not only for the sum of two mappings, but any number. A proof
by induction is immediate from the fact that for convex sets CI, ... ,CT one has
if Z E L,
if Z rJ. L,
Proof. In all cases the equivalence of (ZP) with Equilibrium Problem 2 is elemen-
tary. What we have to verify are the maximal monotonicity assertions. For these we
will rely on Theorem 12. In (a), let
whereas dom Tl is the product of the sets dom R j and dom Sj. The assumption in (a)
about x is equivalent therefore to the assumption that ri (dom T1 ) n ri (dom T2 ) "# 0,
and Theorem 12 then gives us the maximal monotonicity of T.
Similarly in (b), let
if Ex = b,
if Ex"# b.
In the maximal monotone version of Equilibrium Problem 1, T is maximal monotone
as long as there exists a flow x with Ex = b and x jEri ( dom Rj) for all j E J.
(b) (dual case). Equilibrium Problem 1 can be identified with (ZP) in the case of
z = (v, u) and the mapping
if z E L,
if z ~ L,
288
REFERENCES
[1] R. T. Rockafellar, "Convex Analysis", Princeton University Press, Princeton, NJ, 1970.
[2] R. T. Rockafellar, "Lagrange multipliers and optimality," SIAM Review 35, 183-238, 1993.
[3] R. T. Rockafellar, "On the maximality of sums of nonlinear monotone operators," Transactions
Amer. Math. Soc. 149, 75-88, 1970.
[4] G. J. Minty, "Monotone networks," Proc. Roy. Soc. London A 257,194-212,1960.
[5] R. T. Rockafellar, "Network Flows and Monotropic Optimization", Wiley, New York, 1984.
[6] J. J. Moreau, "Proximite et dualite dans un espace hilbertien," Bull. Soc. Math. France 93,
273-299, 1965.
[7] G. J. Minty, "On the maximal domain of a 'monotone' function," Michigan Math. J. 8,135-137,
1961.
GENERALIZED MONOTONICITY - CONCEPTS AND USES
Schaible Siegfried
1. INTRODUCTION
289
290
We begin with a brief description of three models in order to provide the necessary
context for generalized monotone maps. These models have been extensively analyzed
and applied in the second part of this century, e.g. [6], [7], [8], [15], [20], [38], [46].
applications in the natural sciences, often in infinite dimensional spaces, and economic
equilibrium problems as well as problems in a management science give rise to VIP.
(2.4)
Under certain regularity assumptions theoretical as well as algorithmic results for these
models can be established. Classical assumptions are, in addition to closedness and
convexity of C:
for MP: f is convex
for CP/VIP: F is monotone.
Under these assumptions the models above have many nice properties, and algorithms
converge to a solution. Indeed, in many applications these convexity /monotonicity con-
ditions hold.
291
However there are also many applications were these classical assumptions fail to be true.
We realize that they are only sufficient conditions which are not necessary in the theory
of MP, CP and VIP. In case of MP, much progress has been made to develop various
kinds of generalized convex /unctions and demonstrate their usefulness in application,
see [1), [3), [33), [49), [50). Recently various kinds of generalized monotone maps have
been introduced and first uses in CP /VIP have been shown. Most of the work so far has
been conceptual [2), [4), [5), [10), [11), [12), [16), [17), [18), [22), [23), [24), [25), [26), [27),
[28), [29), [30), [31), [32), [33), [35), [36), [39), [40), [46], [47], [54].
However first results demonstrating the usefulness of generalized monotonicity in CP /
VIP have already appeared [2), [9), [19), [20), [23), [24), [26), [48), [52), [53). Many of
these results are related to the existence and uniqueness of solutions.
In [18), [27) the following types of (generalized) monotone maps have been intro-
duced:
quasimonotone
i
monotone --+ pseudomonotone --+ semistrictly quasimonotone
i i i
strictly monotone --+ strictly pseudomonotone --+ strictly quasimonotone
i i
strongly monotone --+ strongly pseudomonotone
(2.5)
In the special case of a gradientmap, these properties correspond to the following (gen-
eralized) convexity properties of the underlying function [1) as shown in [18), [27):
quasi convex
i
convex --+ pseudoconvex --+ semistrictly quasi convex
i i i
strictly convex --+ strictly pseudo convex --+ strictly quasi convex
i i
strongly convex --+ strongly pseudoconvex
(2.6)
We remark that equivalence between strong pseudoconvexity and strong pseu-
domonotonicity of the gradient only holds if the new definition in [18) is used. For
the more restrictive definition in [27) strong pseudomonotonicity of the gradient implies
strong pseudoconvexity of the underlying function, but the reverse is not true, as shown
in [18).
We now give the definitions of (generalized) monotone maps [18), [26), [27).
For the definition of (generalized) convex functions and their characteristic geometrical
properties see [1). We assume F: C -+ R n where C ~ Rn.
F is strongly monotone (str.m) on C if there exists (3 > 0 such that for all x, y E C
(2.9)
which is equivalent to
(2.13)
(2.14)
A: vTF(x) = 0 (2.16)
Let B be a real Banach space, B* its dual and (x,u) the primal-dual pairing
for x E B, u E B*. Furthermore, let K ~ B be a nonempty closed convex set and
F : K -+ B*. We will study the Variational Inequality Problem:
According to the classical result by Hartman and Stampacchia [21] in 1966 there exists a
solution of the VIP if B = R,n , K is compact and F is continuous. This existence result
has been extended in many ways, often assuming monotonicity of F. For monotone VIP
the following properties hold:
- the solution set is convex (it may be empty)
- a solution is unique if F is strictly monotone
- there exists a unique solution if F is strongly monotone.
Also, most algorithms for VIP are derived under the monotonicity assumption.
Turning now to nonmonotone VIP, Karamardian [26] proved the following existence
result for Complementarity Problems:
- the VIP (3.1) has a solution if B = R,n , K ~ R,n is a pointed solid closed convex
cone, there exists x E K such that F( x) E int K*, where K* is the dual of K, and
F is continuous and pseudomonotone.
This result, overlooked for some time, has recently been extended in several ways,
always assuming pseudomonotonicity of F. According to [20], K does not have to be
a cone, i.e. the existence result holds not only for Complementarity Problems, but for
Variational Inequality problems. Cottle and Yao [9] extend Karamardian's result to
Complementarity Problems in Hilbert space, at the same time relaxing the continuity
assumption to continuity on finite dimensional subspaces. Yao [53] extends these results
further to reflexive Banach spaces. In [52] F is allowed to be multivalued, i.e. existence
of solutions is established for generalized VIP.
In [19] Hadjisavvas and Schaible extend Karamardian's result [26] in three direc-
tions: a) B is a reflexive Banach space, b) F is quasimonotone, but not necessarily
pseudomonotone, and c) K has an "inner" point, but not necessarily an algebraic inte-
rior point. The proofs of the existence results in [19] differ significantly from those in
295
previous extensions of Karamardian's existence result. The main reason for this is that
quasimonotone maps form a much larger class than pseudomonotone maps so that new
proof techniques are needed.
In [19) the following four types of generalized monotone maps are used:
pseudomonotone ---+ quasimonotone
1 1
strictly pseudomonotone ---+ striclty quasimonotone.
In the result in (19) the existence of a socalled inner point is assumed, replacing the
more restrictive assumption of the existence of a relative algebraic interior point. A
point Xo E B is called inner point of a nonempty set D ~ B iffor all u E B* \ {O}
(x,u):::; (xo,u) for all xED implies (x,u) = (xo,u) for all xED. (3.2)
A relative algebraic interior point is an inner point, but the converse is not true. It can
be shown that all nonempty separable closed convex sets have inner points. They may
however not have algebraic interior points; e.g. important subsets in CP(1 :::; p < 00).
Theorem 2. Let B a real, reflexive Banach space. Consider a nonempty closed convex
set K ~ B which has an inner point and which is either bounded or there exists p > 0
such that for all x E K with IIxll ~ p there exists Y E K satisfying IIYII < p and
(x - Y , F( x» ~ O. Let F : K -+ B* be hemicontinuous and quasimonotone. Then the
VIP (3.1) has a solution.
This extends not only Karamardian's result in (26), but also the extensions of it in (9), (20)
and [53). The Hilbert space is replaced by a reflexive Banach space, pseudomonotonicity
by quasimonotonicity, and the continuity and interior point assumptions are weakened.
In (19) additional existence results for quasimonotone VIP are obtained. Further-
more, properties of the set of solutions of (3.1) and sufficient conditions for the unique-
ness of a solution are established. An application to a problem in equilibrium analysis
is presented as well.
As above, let B be a real Banach space. Consider a nonempty closed convex cone
K ~ B which induces the partial order
For strictly monotone operators T the equivalence of the above five problems has
been established by ruddell in [42] under certain regularity assumptions. His work
extends earlier results by Stampacchia [51], Lions and Stampacchia [34] and Cryer and
Dempster [13]. In [48] Schaible and Yao extend ruddell's results further from strictly
monotone to strictly pseudomonotone operators.
The map T is called Z-map if
pseudomonotone
T
strictly pseudomonotone
T
strongly pseudomonotone (as in [27]).
Using an existence result by Yao in [52] for certain Variational Inequalities, the following
extension of ruddell's result to strictly pseudomonotone operators is derived in [48]:
In [48] additional results on the equivalence of the above five problems under generalized
monotonicity of the operator are derived.
REFERENCES
[1) M. Avriel, W.E. Diewert , S. Schaible and I. Zang, "Generalized concavity", Plenum Publishing
Corporation, New York, 1988.
(2) M. Bianchi, "Pseudo P-monotone operators and Variational Inequalities", Working Paper No.
6, Istituto di Matematica Generale, Finanziaria ed Econometrica, Universitit Cattolica, Milano,
1993.
(3) A. Cambini, E. Castagnoli, L. Martein, P. Mazzoleni and S. Schaible (eds), "Generalized convexity
and fractional programming with economic applications", Springer-Verlag, Berlin-Heidelberg-
New York, 1990.
(4) E. Castagnoli and P. Mazzoleni, "Order-preserving functions and generalized convexity", Rivista
di Matematica per Ie Scienze Economiche e Sociali 14, 33-45, 1991.
[5) E. Castagnoli and P. Mazzoleni, "Orderings, generalized convexity and monotonicity", in: S.
Komlosi, T. Rapcsak and S. Schaible (eds.) "Generalized convexity", Proceedings, Pees. Hungary
1992, Springer-Verlag, Berlin-Heidelberg-New York, 250-262, 1994.
[6) R.W. Cottle, F. Giannessi and J.L. Lions (eds.) "Variational Inequalities and Complementarity
Problems: theory and applications" , J. Wiley, New York, 1980.
[7) R.W. Cottle, J. Kyparisis and J.S.Pang (eds.), "Variational inequality problems", Mathematical
Programming, Series B, 48, No.2, 1990.
[8] R.W. Cottle, J. Kyparisis and J.S.Pang (eds.), "Complementarity problems", Mathematical Pro-
gramming, Series B, 48, No.3, 1990.
[9) R.W. Cottle and J.S. Yao, "Pseudomonotone complementarity problems in Hilbert space", J. of
Optimization Theory and Applications 75, 281-295, 1992.
[10) J.P. Crouzeix and J .A. Ferland, "Criteria for differentiable generalized monotone maps", M athe-
matical Programming (to appear), 1993.
[11] J.P. Crouzeix and A. Hassouni, "Quasimonotonicity of separable operators and monotonicity in-
dices", Working Paper, Mathematiques Appliquees, Universite Blaise Pascal, Clermont-Ferrand,
1992.
[12) J.P. Crouzeix and A. Hassouni, " Generalized monotonicity of a separable product of opera-
tors: the multivalued case", Working Paper, Mathematiques Appliquees, Universite Blaise Pascal,
Clermont-Ferrand, 1993.
[13] C.W. Cryer and M.A.H. Dempster, "Equivalence of linear Complementarity Problems and linear
programs in vector lattice Hilbert spaces", SIAM J. Control Optimization 18, 76-90, 1980.
[14] J .A. Ferland, "Quasi-convex and pseudo-convex functions on solid convex sets" , Technical Report
71-4, Department of Operations Research, Stanford University, 1971.
298
[15] F. Giannessi and F. Niccolucci, "Connections between nonlinear and integer programming prob-
lems". Symposia Mathematica, Vol. XIX, Academic Press, 160-175,1976.
[16] M.S. Gowda, "Pseudomonotone and copositive star matrices, Linear Algebra and Its Applications
113, 107-118, 1989.
[17] M.S. Gowda, "Affine pseudomonotone mappings and the linear Complementarity Problem", SIAM
J. of Matrix Analysis and Applications 11,373-380, 1990.
[18] N. Hadjisavvas and S. Schaible, "On strong pseudomonotonicity and (semi) strict quasimonotonic-
ity", J. of Optimization Theory and Applications 79, 139-155, 1993.
[19] N. Hadjisavvas and S. Schaible, "Quasimonotone Variational Inequalities in Banach spaces" , Work-
ing Paper 94-01, Graduate School of Management, University of California, Riverside, 1994.
[20] P.T. Harker and J.S. Pang, "Finite dimensional Variational Inequality and nonlinear Complemen-
tarity Problems: a survey of theory, algorithms and applications", in: R.W. Cottle, J. Kyparisis
and J.S. Pang (eds.), "Variational inequality problems", Mathematical Programming 48, Series E,
161-220, 1990.
[21] P. Hartman and G. Stampacchia, "On some non-linear elliptic differential-functional equations",
Acta Math. 115,271-310, 1966.
[22] A. Hassouni, "Sous-differentiels des fonctions quasi-convexes" , These de 3 erne Cycle, Mathemati-
ques Appliquees, Universite Paul Sabatier, Toulouse, 1983.
[23] A. Hassouni, "'Quasimonotone multifunctions, applications to optimality conditions in quasiconvex
programming", Numerical Functional Analysis and Optimization 13,267-275, 1992.
[24] A. Hassouni, "Operateurs quasimonotone, applications a certain problemes variationnels" , These,
Mathematiques Appliquees, Universite Paul Sabatier, Toulouse, 1993.
[25] A. Hassouni and R. Ellaia, "Characterizations of nonsmooth functions through their generalized
gradients", Optimization 22, 401-416, 1991.
[26] S. Karamardian, "Complementarity over cones with monotone and pseudomonotone maps", J. of
Optimization Theory and Applications 18,445-454, 1976.
[27] S. Karamardian and S. Schaible, "Seven kinds of monotone maps", J. of Optimization Theory and
Applications 66, 37-46, 1990.
[28] S. Karamardian and S. Schaible, "First-order characterizations of generalized monotone maps" ,
Working-Paper 90-5, Graduate School of Management, University of California, Riverside, 1989.
[29] S. Karamardian, Schaible S. and J.P. Crouzeix , "Characterizations of generalized monotone
maps", J. of Optimization Theory and Applications 76, 399-413, 1993.
[30] S. Komlosi, "On generalized upper quasidifferentiability", in: F. Giannessi (ed.), "Nonsmooth
optimization methods and applications". Gordon and Breach, Amsterdam 189-200, 1992.
[31] S. Komlosi, "Generalized monotonicity in non-smooth analysis", in: S. Komlosi, T. Rapcsak and
S. Schaible (eds.), "Generalized convexity", Proceedings, Pees. Hungary 1992, Springer-Verlag,
Berlin-Heidelberg-New York, 263-275, 1994.
[32] S. Komlosi, "Generalized monotonicity and generalized convexity", J. of Optimization Theory and
Applications (to appear), 1993.
[33J S. Komlosi, T. Rapcsak and S. Schaible (eds.), "Generalized convexity", Proceedings, Pees. Hun-
gary 1992, Springer-Verlag, Berlin-Heidelberg-New York 1994.
[34] J.L. Lions and G. Stampacchia, "Variational inequalities", Comm. Pure Appl. Math. 20,493-519,
1967.
299
[35] D.T. Luc, Characterizations of quasiconvex functions", Bulletin of the Australian Mathematical
Society 48, 393-405, 1993.
[36] D.T. Luc, "On generalized convex nonsmooth functions", Bulletin of the Australian Mathematical
Society 49, 139-149, 1994.
[37] D.T. Luc and S. Schaible, "Generalized monotone nonsmooth maps", Working Paper, Graduate
School of Management, University of California, Riverside, 1994.
[38] A. Maugeri, "Convex programming, Variational Inequalities and applications to the traffic equi-
librium problem", Appl. Math. Optim. 16, 169-185, 1987.
[39] J.P. Penot and P.H. Quang, "On generalized convex functions and generalized monotonicity of
set-valued maps", J. of Optimization Theory and Applications, (to appear), 1993.
[40] R. Pini and S. Schaible, ''Invariance properties of generalized monotonicity", Optimization, 28,
211-222, 1994.
[41] W.C. Rheinboldt, "On M-functions and their application to nonlinear Gauss-Seidel iterations and
network flows", J. Math. Anal. Appl. 32, 274-307, 1970.
[42] R.C. Riddel, "Equivalence of nonlinear complementarity problems and least element problems in
Banach lattices", Mathematics of Operations Research 6, 462-474, 1981.
[43] S. Schaible, "Beitraege zur quasikonvexen Programmierung", Doctoral Dissertation, Universitaet
Koeln, 1971.
[44] Schaible S., "Quasiconvex, pseudoconvex and strictly pseudoconvex quadratic functions", J. of
Optimization Theory and Applications 35, 303-338, 1981.
[45] S. Schaible, "Generalized convexity of quadratic functions", in: S. Schaible and W.T. Ziemba
(eds.) , "Generalized concavity in optimization and economics". Academic Press, New York, 183-
197, 1981.
[46] S. Schaible, "Generalized monotone maps", in: F. Giannessi (ed.), "Nonsmooth optimization
methods and applications", Proceedings, Erice 1991, Gordon and Breach, Amsterdam, 392-408,
1992.
[47] S. Schaible, " Generalized monotonicity - a survey", in: Komlosi S., Rapcsak T. and S. Schaible
(eds.), "Generalized convexity", Proceedings, Pees. Hungary 1992, Springer-Verlag, Berlin-Heidel-
berg-New York, 229-249, 1994.
[48] S. Schaible and J .C. Yao, "On the equivalence of nonlinear Complementarity Problems and least
element problems", Mathematical Programming (to appear), 1993.
[49] S. Schaible and W.T. Ziemba (eds.), "Generalized concavity in optimization and economics",
Academic Press, New York, 1981.
[50] C. Singh and B.K. Dass (eds.), "Continuous-time, fractional and multiobjective programming",
Analytic Publishing Company, New Delhi, 1989.
[51] G. Stampacchia, "Le probleme de Dirichlet pour les equations elliptiques du second ordre a coef-
ficients discontinus", Ann. Inat. Fourier (Grenoble) 15, 189-258, 1965.
[52] J .C. Yao, "Multi-valued Variational Inequalities with K-pseudomonotone operators", J. of Opti-
mization Theory and Applications 83, 1994 (to appear).
[53] J .C. Yao, "Variational inequalities with generalized monotone operators", Mathematics of Opera-
tions Research (to appear), 1993.
[54] D.L. Zhu and P. Marcotte, "New classes of generalized monotonicity", J. of Optimization Theory
and Applications (to appear), 1994.
CONTRIBUTORS
301
302
303
304