Computable Models of Dynamic Spatial

Oligopoly from the Perspective of

Differential Variational Inequalities

Terry L. Friesz and Amir H. Meimand

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
2 The Notion of a Nash Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
3 Competitive Oligopoly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
3.1 Competitive Spatial Oligopoly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
3.2 Variational Inequality (VI) Formulations of Spatial Oligopoly . . . . . . . . . . . . . . . . . . . . . . 306
3.3 Diagonalization Algorithm for Variational Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
4 Static Competitive Network Oligopoly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
5 Dynamic Network Oligopoly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
5.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
5.2 The Firm’s Objective Functional, Dynamics, and Constraints . . . . . . . . . . . . . . . . . . . . . . 310
5.3 The Differential Variational Inequality Formulation of Dynamic Network
Oligopoly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
5.4 Discrete-Time Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
5.5 A Comment About Path Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
5.6 Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
5.7 Interpretation of Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
7 Cross-References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323

Abstract
We begin this chapter with the basic definition of Nash equilibrium and formulation
of static spatial and network oligopoly models as variational inequalities (VIs),

T. L. Friesz (*)
Department of Industrial and Manufacturing Engineering, Pennsylvania State University,
University Park, PA, USA
e-mail: tlf13@psu.edu; tfriesz@psu.edu
A. H. Meimand
Zilliant, Austin, TX, USA
e-mail: amirhosein.meimand@gmail.com

© Springer-Verlag GmbH Germany, part of Springer Nature 2021 301

M. M. Fischer, P. Nijkamp (eds.), Handbook of Regional Science,
https://doi.org/10.1007/978-3-662-60723-7_105
302 T. L. Friesz and A. H. Meimand

which can be solved by well-known numerical methods presented in the literature.

We then move on to dynamic network oligopoly models and show the differential
Nash game describing competitive network oligopoly may be articulated as a
differential variational inequality (DVI) involving both control and state variables.
Finite-dimensional time discretization is employed to approximate the model as a
mathematical program, which may be solved by the multistart global optimization
scheme found in the off-the-shelf software package GAMS when used in conjunc-
tion with the commercial solver MINOS. We also present a small-scale numerical
example of differential network oligopoly approached from the DVI perspective.

Keywords
Spatial Oligopoly · Dynamic Network Oligopoly · Differential Variational
Inequalities

1 Introduction

The theory of competitive oligopoly is reviewed in Greenhut et al. (1987) and

Greenhut and Lane (1989), and some basic models of spatial oligopoly are presented
by Raa (1984), Dafermos and Nagurney (1987), Henderson and Quandt (1980),
Novshek (1980), Matsushima and Matsumura (2003), and Taheri et al. (2014).
Moreover Dafermos and Nagurney (1987), Harker (1984), Nagurney (1999), and
Nagurney and Dong (2016a) studied the variational inequality (VI) approach to
determine market equilibrium for a general static oligopoly model.
The intention of this chapter is to construct computable general equilibrium
models for static and dynamic spatial oligopoly with a computational rather than
theoretical orientation. That said, it is nonetheless necessary to give a mathematical
form to oligopoly models and explore some of their qualitative properties. Accord-
ingly, we begin this chapter with the definition of Nash equilibrium and formulation
of static spatial and network oligopoly as variational inequality in order to set the
stage for the dynamic models we study subsequentially. In the network case, a few
firms compete as Nash agents in the output market for a single homogeneous
commodity. The firms are located at nodes of a transportation network in which
common freight tariffs expressed as a fee per unit of flow on each network arc are
known and faced by each oligopolist. We then move on to modeling and computing
dynamic network oligopoly described by the notion of a differential Nash equilib-
rium and mathematically expressed as a differential variational inequality (DVI).
Models like those presented in this chapter arise when constructing spatial
computable general equilibrium models (Tobin and Friesz 1983; Friesz and Harker
1984; Dafermos and Nagurney 1987; Beckmann and Puu 1990; Friesz 1993), as well
as partial equilibrium models, when detailed freight flows are emphasized for a
specific application. Throughout, our emphasis is on computation rather than theory,
and our style is that of a simple primer in order to make the material accessible by the
widest possible audience. Although some of the network models considered herein
Computable Models of Dynamic Spatial Oligopoly from the Perspective of. . . 303

are notationally complex, our presentation relies only on very basic notions from
microeconomic theory and elementary optimization.

2 The Notion of a Nash Equilibrium

Suppose there are N players in a game and each player i  [1, N]
I++ chooses a
feasible strategy tuple xi from strategy set Ωi to maximize the utility function
Θi : Ωi ! ℜ1, which will generally depend on other players’ strategies. So every
player i  [1, N] is trying to solve his/her best response problem:


max Θðxi ; xi Þ
: ð1Þ
s:t: x i  Ωi

Note that, in Eq. (1), we use the notation:

 
xi ¼ x j, j 6¼ i i  ½1, N 

to refer to the tuple of strategies of players other than i. It is assumed that the
non-own tuple xi is known to player i. A Nash equilibrium is the vector of strategies
 
x ¼ xi : i  ½1, N 

formed from player-specific tuples such that each xi solves the mathematical pro-
gram (1). We denote a Nash equilibrium by NE(Θ, Ω) (Friesz 2010).

3 Competitive Oligopoly

Consider there are N firms i ¼ 2, . . . , N, which supply a homogeneous commodity in

a noncooperative fashion. Let π ðQÞ: ℜ1þþ ! ℜ1þþ denote the inverse demand
function, where Q  0 is the total supply in the market and qi  0 is the supply
for firm i:

X
N
Q¼ qi : ð2Þ
i¼1

Also, let V i ðqi Þ: ℜ1þþ ! ℜ1þþ denote the total production cost of firm i to supply qi
units. Given the other firms strategy, qi, firm i is trying to maximize its total profit
jNj
Θi ðqi ; qi Þ: ℜþþ ! ℜ1 by solving its best response problem:
   
Max Θi qi ; qi ¼ qi π ðQÞ  V i qi ð3Þ
qi

subject to
304 T. L. Friesz and A. H. Meimand

X
N
Q¼ qi
i¼1
qi  0:

It is obvious that there is Nash game between the players (firms) choosing a feasible
strategy qi  0 and every player’s objective function depends on all other players’
strategies. This game can be defined as a Nash equilibrium NE(Θ, Ω) (Harker 1984).
If V i(.) is convex and continuously differentiable for i ¼ 1, 2, . . . , N, the inverse
demand function π(.) is strictly decreasing and continuously differentiable, and the

industry revenue function Qπ(Q) is concave, and then q ¼
 1 2   jNj
q , q , . . . , q N  ℜþþ is a Nash equilibrium solution if and only if:
         
ðiÞ π Q þ q i ∇qi π Q  ∇qi V i q i q i ¼ 0 8i ¼ 1, 2, . . . , N
      
ðiiÞ π Q þ q i ∇qi π Q  ∇qi V i q i  0 8i ¼ 1, 2, . . . , N

where


X
N

Q ¼ qi
i¼1

q i  0:

Conditions (i) and (ii) are the first-order necessary (and sufficient) conditions for the
set of problems defined by Eq. (3) for each i ¼ 1, 2, . . . , N (Murphy et al. 1982).

3.1 Competitive Spatial Oligopoly

Now we can generalize the classical oligopoly model to the spatial case. Assume that
there are N firms and M demand markets that are generally spatially separated.
Assume the homogenous commodity is produced by N firms and consumed in
M markets. Let qi denote the output of firm i  [1, N]
I++ and c j denote the
demand for the commodity at market j  [1, M]
I++. Let sij denote the commodity
shipment form firm i to the market j. The following conservation of flow equations
must hold:

X
M
qi  sij ¼ 0 8i ¼ 1, 2, . . . , N ð4Þ
j¼1

X
N
cj  sij ¼ 0 8j ¼ 1, 2, . . . , M ð5Þ
i¼1
Computable Models of Dynamic Spatial Oligopoly from the Perspective of. . . 305

where

0  c  cmax , 0  q  qmax , 0  s  smax : ð6Þ

Note that in these constraints, the column vectors

jNj
q  ℜþþ
jMj
c  ℜþþ
jNjjMj
s  ℜþþ

denote the production outputs, demands, and the commodity shipments. As previously,
V i(qi) is the total production cost for firm i to supply qi units, r ij is the freight rate (tariff)
charged per unit of flow sij , and π j(c j) is the inverse demand function at market j. To
consider the general situation, we assume the production cost may depend on the entire
jNj
production pattern V i ¼ V i ðqÞ: ℜþþ ! ℜ1þþ , the freight rate (tariff) may depend on
jNjjMj
the entire shipment pattern, r ij ¼ r ij ðsÞ: ℜþþ ! ℜ1þþ , and the inverse demand
function at market j may depend on the entire consumption pattern, π j ¼ π j ðcÞ:
jMj
ℜþþ ! ℜ1þþ . Note that the profit of firm i, as in Dafermos and Nagurney (1987), is

X
M X
M
Θi ðc, q, sÞ ¼ π j ðcÞsij  V i ðqÞ  r ij ðsÞsij : ð7Þ
j¼0 j¼0

In light of constraints (Eqs. (4) and (5)), one my write

Θ ¼ ΘðsÞ:

Next we consider the usual oligopolistic market mechanism, in which the N firms
behave as noncooperative agents and satisfy demand while maximizing their own
profit. We seek to determine a commodity shipment pattern s for which the N firms will
be in a state of equilibrium a fundamental concept explained in Nagurney (1999).

Definition 1 (Spatial Cournot-Nash Equilibrium).

A commodity shipment pattern s is said to constitute a Cournot-Nash equilib-
rium if for each for i ¼ 1, 2, . . . , N:
  
  

Θi sji , sji  Θi sji , sji 8sji  Ω

where Ω ¼ {Eqs. (4), (5), (6)}.

306 T. L. Friesz and A. H. Meimand

3.2 Variational Inequality (VI) Formulations of Spatial Oligopoly

In this section we present the variational inequality (VI) formulation of the Cournot-
Nash equilibrium. Letting
x ¼ ðc, q, sÞ

we make use of the following theorem:

Theorem 1 (Nash equilibrium equivalent to a variational inequality).

The general Nash equilibrium NE(Θ, Ω) is equivalent to the following varia-
tional inequality VI(∇Θ, Ω):

find x  Ω ð8Þ

such that
   T  
∇Θ x xx 0 8x  Ω

the following regularity conditions hold: (i) each Θi(x) : Ωi ! ℜ1 is convex

and continuously differentiable in xi; and (ii) each Ωi is a closed convex subset
of ℜni .

Proof See (Friesz 2010) ■

Applying Theorem 1, the competitive oligopoly problem may be formulated as the
following VI (Nagurney 1999):

find x  0 ð9Þ
such that
X
N           
∇qi V i qi  π Q  ∇qi π Q qi qi  qi  0
i¼1


X
N

where Q ¼ qi and qi  0:
i¼1

3.3 Diagonalization Algorithm for Variational Inequality

In this section, the diagonalization algorithm (or diagonalization for short) discussed
is one of several algorithms that can be used to solve a variational inequality. It is an
algorithmic philosophy very similar to the Gauss-Seidel method familiar from the
Computable Models of Dynamic Spatial Oligopoly from the Perspective of. . . 307

numerical analysis literature (Friesz 2010). Diagonalization is appealing for solving

finite-dimensional variational inequalities because the resulting subproblems are all
nonlinear programs that can be efficiently solved with well-understood nonlinear
programming algorithms, which are often available in the form of commercial
software. This fact notwithstanding diagonalization may fail to converge, and its
use on large-scale problems can be frustrating.
We next describe the algorithm, again making use of the shorthand

x ¼ ðc, q, sÞ

Step 0. Initialization. Determine a feasible initial solution x0  Ω, and set k ¼ 0.

k
Step
 1. Diagonalize  and solve the subproblem. We create separable functions Fi ðxi Þ
Fi xi ; xkj 8j 6¼ i and then proceed to solve the variational inequality

X
N   
Fki xki xi  xki  0 x, xk  Ω
i¼1

by solving the mathematical program

N ð xi
X
max J ðxÞ ¼ Fki ðxi Þ dxi xΩ
i¼1 0

and call the solution xk+1.

Step 2. For a suitably small positive constant ξ  ℜ1þþ , if

max j xkþ1
i  xki j ξ
i  ½1, N 

stop; otherwise set k ¼ k + 1 and go to Step 1.

The other numerical methods for VI such as gap function method, fixed-point
method, and successive linearization method and Lemke’s algorithm are presented
in Friesz (2010) with some numerical examples.

4 Static Competitive Network Oligopoly

As background for static competitive network oligopoly, we imagine a transportation

network that allows firms to ship their commodity to multiple markets through
multiple paths. In this section, we consider the impact of spatial transportation
networks on the firms’ economic decisions. For the time being, we restrict our
consideration to the static case, while in a later section, we will extend this model
308 T. L. Friesz and A. H. Meimand

to the dynamic case. Several papers exist in the literature about modeling spatial
network oligopoly. For example, Hakimi (1983) chooses locations for new facilities
in an oligopolistic environment. Tobin and Friesz (1983) show that an equivalent
optimization problem for spatial price network equilibrium may be formulated
without path variables. Moreover, Rovinskey et al. (1980) developed a model for
Cournot-Nash equilibrium of firms who are spatially separated but supply a com-
modity to a single market. Hashimoto (1985) and Harker (1986) formulated
more general Cournot-Nash models in which consumers are also spatially dispersed.
More recently Nagurney (2016b) presented an alternative variational inequality
formulation and studied its application in supply chain network with product
differentiation.
In this section we are going to present the static network oligopoly model. In
this model there is Nash game among firms whose facilities and final demand
markets are fixed at distinct nodes of a distribution network and connected by
paths involving chains of arcs of that network. The network is represented by G(N,
W ), in which N is set of nodes and W is set of origin-destination (OD) pairs (i, j).
Also F is a set of firms competing on the network. Nf is the set of nodes at which
firm f has economic presence, and Wf is the set of OD pair used by firm f to
transport commodity. Moreover, sijf is the shipment of firm f for OD pair (i, j), rij is
the freight rate (tariff) charged per unit of flow sijf for OD pair (i, j), cif is the
f
  of firm f  F at node i  N f , qi is output of firm f  F at
allocation of the output
node i  N f , V i qi is the variable cost of production for qif units. The best
f f

response problem for firm f is:

!
 f f f f f f  X X g f X f f X
max Θ f c , q , h ; c , q , h ¼ πi ci ci  V i qi  rij sijf
iN f gF iN f ði, jÞ  W f

ð10Þ

such that
X X
cif ¼ qif þ sjif  sijf 8i  N f ð11Þ
ði, jÞ  W ð j, iÞ  W

0  cif  cmax f f
i , 0  qi  qi , 0  sij  sij :
max max
ð12Þ

Each firm is seeking to maximize its total profit (Eq. (10)) expressed as revenue less
production and transportation cost. Constraint (11) guarantees that flow is conserved
at each node. Moreover, per constraints (12), all consumption, production, and
shipping variables are nonnegative and bounded from above.
The static network oligopoly can be formulated as the following VI:
   
find c f , q f , s f  Ω ð13Þ
Computable Models of Dynamic Spatial Oligopoly from the Perspective of. . . 309

such that
2 3
X X  
 X  
 X  

4 f
∇q f Θ f q i  q i f
þ f
∇ c f Θ f ci  ci f
þ ∇s f Θ f sij  sij 5  0
f f
i i ij
f F iN f iN f ði, jÞ  W f

   
for all (c, q, s)  Ω where Ω ¼ c , q , s : Eqs: ð11Þ, ð12Þ hold

5 Dynamic Network Oligopoly

We now turn to the problem of modeling and computing differential Nash equilibria
among the oligopolistic firms. The oligopolistic firms of interest, embedded in a
network economy, form a competitive oligopoly under the influence of dynamics
that describe the trajectories of inventories/backorders and correspond to flow
conservation for each firm at each node of the network. The oligopolistic firms,
acting as shippers, perfectly compete as price takers in the market for physical
distribution services. Perfect competition in shipping arises because numerous
shipping companies serve numerous customers due to the involvement of shippers
in the numerous output markets of the network economy. The time scale we consider
is neither short nor long but rather of sufficient length to allow output and shipping
pattern adjustments, while not long enough for firms to relocate, enter, or leave the
network economy. This model, which takes the form of a differential variational
inequality, is presented in this section. Dynamic network oligopoly models were
previously studied by Brander and Zhang (1993), Nagurney et al. (1994, 2002,
2016), Wie and Tobin (1997), Friesz et al. (2006), Markovich (2008), Mozafari et al.
(2015), and Chan et al. (2018). For the most part, these prior models assumed
inventory was cleared during each time period of interest. To develop a more general
mathematical formulation of network oligopoly, we follow the exposition in
Friesz (2010).

5.1 Notation

Fortunately, much of the notation introduced in previous sections of this chapter is

still relevant. Yet, because there are some subtle differences between the dynamic
oligopoly model that we now study and problems explored previously in this
chapter, we choose to give an exhaustive list of the notation to be employed
below, even though that will involve some duplication. In particular, we again let
continuous time be denoted by the scalar t ℜ1þ , initial
 time by t0  ℜ1þ, and final
1
time by t f  ℜþþ , with t0 < tf so that t  t0 , t f ℜ1þ . There are several sets
important to articulating a model of oligopolistic competition on a network; these are
as follows: F for firms, A for directed arcs, N for nodes, and W for origin-destination
(OD) pairs. Subsets of these sets are formed as is meaningful by using the subscripts
f for a specific firm, i for a specific node, and ij for a specific OD pair (i, j).
310 T. L. Friesz and A. H. Meimand

Each firm f  F controls production (output) rates q f, allocation of output to

meet demand c f, and shipping pattern s f. Inventories I f are state variables determined
by the controls. In particular, concatenations of the firm-specific vectors c f, q f
  jN jjF j   jN jjF j   jW jjF j
and s f give the c  L2 t0 , t f , q  L2 t0 , t f , s  L2 t0 , t f .
Furthermore, the state operator, once again, will be
  jN jjF j  2  jN jjF j  2  jW jjF j
I ðc, q, sÞ: L2 t0 , t f  L t0 , t f  L t0 , t f
  jN jjF j
! H 1 t0 , t f
1
 
where L2[t0, tf] is the space of square-integrable
  functions and H t 0 , t f is a
1
Sobolev space for the real interval t0 , t f  ℜþ .

5.2 The Firm’s Objective Functional, Dynamics, and Constraints

Each firm has the objective of maximizing net profit expressed as revenue less cost
and taking the form of an operator acting on allocations of output to meet demands,
production rates, and shipment patterns. For each firm f  F , net profit is given by
the following functional:
     
Φf c f , q f , s f ; cf , qf ¼ eρtf Z f I tf , t f
8 !
ð tf <X X g X f
ρt

þ e πi c i , t c fi V i qf , t
t0 : gF
iN iNf

X X   X
 rij ðtÞ s fij  V fi q f , t  rij ðtÞ s fij
ði , j Þ  W f iNf ði , j Þ  W f

X  
 ψ fi I fi ,t dt (14)
iN

where ρ  ℜ1þþ is a constant nominal rate of discount, r ij  ℜ1þþ is the freight

rate (tariff) charged per unit of flow sij for OD pair ði, jÞ  W f , ψ if is firm f ’s
inventory cost at node i, and I if is the inventory/backorder of firm f at node i.
In expression (14), cif is the allocation of the output of firm f  F at node
i  N to consumption at that node. Also Z f [I f(tf ), tf ] is the liquidation value of
inventory remaining at the terminal time, where I f ¼ ðI if: i  N f Þ. Because our
formulation is in terms of flows, P it is convenient to employ the inverse demand
functions π i(ci, t) where ci ¼ g  F cgi is the total allocation of output to con-
sumption for node i. Furthermore, qif is the output of firm f  F at node i  N .
Again V if ðq, tÞ is the variable production cost for firm f  F at node i  N . The
reader should note that θ f (c f, q f, s f; cf, qf) is a functional that is determined by
the controls c f, q f and s f when non-own allocations to consumption and non-own
production rates
Computable Models of Dynamic Spatial Oligopoly from the Perspective of. . . 311

0 0
cf ðc f : f 0 6¼ f Þ, qf ðq f : f 0 6¼ f Þ

are taken to be exogenous data by firm f. The first term of the objective functional
θ f (c f, q f, s f; cf, qf) in expression (14) is the firm’s revenue; the second term is the
firm’s cost of production; the third term is the firm’s shipping costs; and the last term
is the firm’s inventory or holding cost.
We next impose the terminal time inventory constraints

  f
I if t f  K~i 8f  F , i  N f ð15Þ

f
where K~i  ℜ1þþ are exogenous. Again all consumption, production, and shipping
variables are nonnegative and bounded from above; that is,

Cf  cf  0 ð16Þ

Qf  qf  0 ð17Þ

Sf  sf  0 ð18Þ

where

jF j jF j jW j
C f  ℜþþ , Q f  ℜþþ , S f  ℜþþf :

As for the monopoly, constraints (Eqs. (16), (17), and (18)) are recognized as pure
control constraints, while inequalities (Eq. 15) are terminal conditions for the state
space variables. Naturally
 f f f
Ωf ¼ c , q , s : Eqs: ð16Þ, ð17Þ, ð18Þ

is the set of feasible controls.

Firm f solves an optimal control problem to determine its production q f, alloca-
tion of production to meet demand c f, and shipping pattern s f by maximizing its
profit functional Φf (c f, q f, s f; cf, qf) subject to inventory dynamics expressed as
flow balance equations and pertinent production and inventory constraints. The
inventory dynamics for firm f  F , expressing simple flow conservation, obey

dI if X f X f
¼ qif þ sji  sij  cif 8i  N f ð19Þ
dt
ðj, iÞ  W ði, jÞ  W

I if ðt0 Þ ¼ K if 8i  N f ð20Þ
  f
I if t f ¼ K~i 8i  N f ð21Þ
f
where K if  ℜ1þþ and K~i  ℜ1þ are exogenous. In addition to the terminal time
inventory (state) constraints (21), the model is general enough to handle inventory
312 T. L. Friesz and A. H. Meimand

constraints over the entire planning horizon [t0, tf]. For instance, nonnegativity of the
inventory (state) variables could be imposed to restrict firms from taking backorders.
In light of the above development, we write
8
< dI f X X
I ðc, q, sÞ ¼ arg i
¼ q fi þ s fji  s fij  c fi
: dt ðj,iÞ  W ði,jÞ  W
o
f 
¼ K~ i
f f
I i ðt 0 Þ ¼ K fi , I i tf 8f  F i  N f

where we have assumed that the dynamics have solutions for all feasible controls.
We next note that firm f ’s problem is with the cf and qf as exogenous inputs,
compute c f, q f and s f (thereby finding I f) in order to solve the following extremal
problem:
 )
max Φ f c f , q f , s f ; cf , qf
 f f f 8f  F ð22Þ
s:t: c , q , s Ωf
where
 
Ωf ¼ c f , q f , s f : ð15Þ, ð16Þ, ð17Þ, ð18Þ hold

also for all f  F . That is, each firm is a Nash agent that knows and employs the
current instantaneous values of the decision variables of other firms to make its own
noncooperative decisions. As such, Eq. (22) is a differential Nash game. Moreover,
each firm’s best response problem (22) is a continuous time optimal control problem.

5.3 The Differential Variational Inequality Formulation

of Dynamic Network Oligopoly

To continue our study of oligopoly, we assume the Nash game expressed above is
regular in the sense of the following definition:

Definition 2 The dynamic oligopolistic network competition problem intro-

duced above will be considered regular if (i) the state operator I(c, q, s) exists
and is unique, while each of its components is continuous and G-differentia-
ble; (ii) the inverse demand, production cost, and inventory cost functions are
continuously differentiable with respect to controls and states; and (iii) for
each f  F , the composite terminal cost function
    X fh f  i
Zf If tf ,tf þ γ i K~i  I if t f
iN f
 
is continuously differentiable with respect to I if t f for all i  N f .
Computable Models of Dynamic Spatial Oligopoly from the Perspective of. . . 313

In the above definition, each γ if is a constant dual variable that prices out the terminal
constraint on inventory. We also note that the best response problem (22) is an
optimal control problem with fixed terminal time. Its Hamiltonian is

 
H f c f , q f , s f , I f , α f , β f , λ f ; cf ; qf ; t
   
Φ f c f , q f , s f , I f ; cf , qf ; t þ Ψ f c f , q f , s f , I f , α f , β f , λ f

where
8 !
  <X X
f f f
Φf c , q , s , I ; c , q ; t ¼ e f f f ρt
πi cgi , t cif
:i  N gF
f

9
X X X  =
 V if ðq, tÞ  r ij ðtÞsijf  ψ if I if , t ð23Þ
iN ði, jÞ  W iN
;
f f f

and
0 1
  X X X
f f f
Ψf c , q , s , I , α , β , λ f f f f
¼ λif @qif þ sjif  sijf  cif A
iN f ð j, iÞ  W ði, jÞ  W

ð24Þ

where αif  ℜ1þ and βif  ℜ1þ are dual variables for the inventory-bounding con-
straints (Eq. 15), while α f  ℜ j N f j and β f  ℜjN f j ; also λif  ℜ1þ is the adjoint
 jN j
1
variable for the dynamics of firm f at node i, while λ f  ℋ t0 , t f . Clearly Φf
is the instantaneous profit. To interpret Ψf we need to understand the relevant
dynamic shadow benefits and shadow costs of this model. To that end, recall that,
along an optimal trajectory, the adjoint variables obey

@J f
λif ¼ :
@I if
Consequently,

X @J f dI f
i
Ψf ¼ f
i  N @I
dt
f i

which is recognized as the shadow value of dynamic benefits arising from current
inventory held; it can be either a cost or a benefit, depending on its sign.
314 T. L. Friesz and A. H. Meimand

Due to regularity, the Pontryagin’s maximum principle (Sethi and Thompson

2000) takes the form of requiring that the nonlinear program

   
max H f s:t: C f , Q f , S f  c f , q f , s f  0

be solved by every firm f  F for every instant of time t  [t0, tf]. Consequently, since
the feasible set is convex, the finite-dimensional variational inequality principle from
the necessary conditions requires any optimal solution to satisfy (Friesz et al. 2006)

@H f  


cif  ci f 0 ð25Þ
@cif

@H f  


qif  qi f 0 ð26Þ
@qif

@H f  


sijf  sijf 0 ð27Þ
@sijf

for every f  F at every time, t  [t0, tf]. Familiarity with variational inequalities
suggests that the following variational inequality has solutions that are differential
Nash equilibria for a noncooperative game in which individual firms maximize net
profits in light of current information about their competitors:
   
find c f , q f , s f  Ω such that
"
X ðt f X @Φ f   X @Φ f  f 
 
f 
0 f
cif  ci þ f
qi  qi f
f  F t0 i  N f @ci i  N f @qi
#
X @Φ f  f 

f
þ f
sij  sij dt for all ðc, q, sÞ  Ω ð28Þ
ði, jÞ  W f @sij

where

      
Φ f ¼ Φ f c f , q f , s f , I f ; cf , qf ; t ð29Þ
Y
Ω¼ Ωf: ð30Þ
f F

We note that Eq. (28) is a differential variational inequality expressing the

differential Nash game that is our present interest. This formulation also provides
guidance in devising a computational strategy, as we show in Sect. 5.4. The issue of
immediate concern is to formally demonstrate that solutions of Eq. (28) are differ-
ential Nash equilibria. In fact, we state and prove the following result:
Computable Models of Dynamic Spatial Oligopoly from the Perspective of. . . 315

Theorem 2 Differential variational inequality formulation of dynamic oli-

gopolistic network competition. Any solution of Eq. (28) is a solution of the
dynamic oligopolistic network competition problem when regularity in the
sense of Definition 2 holds.

Proof We follow the presentation in Friesz (2010) and begin by noting that Eq. (28)
is equivalent to the following optimal control problem:
2 3
X ð t f X @Φ f f
  
X @Φ f f X @Φ f f
max Gðc, q, sÞ ¼ 4 c þ q þ s 5dt
f i f i f ij
f  F t0 i  N f @ci i  N f @qi ði, jÞ  W f @sij

s:t: Eqs: ð15Þ, ð16Þ, ð17Þ, ð18Þ, and ð19Þ

where it is essential to recognize that G(c, q, s) is a linear functional that assumes
knowledge of the solution to our oligopolistic game; as such, G(c, q, s) is a mathe-
matical construct for use in analysis and has no meaning as a computational device.
The augmented Hamiltonian for this artificial optimal control problem is
2 3
  
X X @Φ f f X @Φ f f X @Φ f f X
H0 ¼ 4 c þ q þ s 5þ Ψf:
f i f i f ij
f  F i  N f @ci i  N f @qi ði, jÞ  W f @sij f F

The associated maximum principal requires

   
max H 0 s:t: C f , Q f , S f  c f , q f , s f  0:

The corresponding necessary conditions for this mathematical program are identical
to Eq. (25) through Eq. (27), since

   
@H 0 @Φ f @Ψ f @H f
¼ þ ¼
@cif @cif @cif @cif
   
@H 0 @Φ f @Ψ f @H f
¼ þ ¼
@qif @qif @qif @qif
   
@H 0 @Φ f @Ψ f @H f
¼ þ ¼
@sijf @sijf @sijf @sijf
where
2 3
  
X X @Φ f  X @Φ f  X @Φ f  X

H0 ¼ 4 ci f þ qi f þ sijf 5 þ

Ψf
f F iN f
@cif iN f
@qif ði, jÞ  W f
@sijf f F

and
316 T. L. Friesz and A. H. Meimand

      


f f
Ψf ¼ Ψf c ,qf ,sf ,If ,αf ,βf ,λ :

We next note that the following existence result holds:

Theorem 3 Existence of dynamic oligopolistic network equilibrium. When the

variational inequality of Theorem 2 is regular in the sense of Definition 2,
there exists a solution of the dynamic oligopolistic network competition
problem.

Proof Note that each set of admissible controls Ωf is convex and compact by the
virtue of the given and the explicit lower and upper bounds of the formulation. Note
also that continuity of the principal operator of DVI (Eq. (28)) is assured by
regularity. Existence is then an immediate consequence of the Stampacchia existence
theorem, as presented in Friesz (2010). ■

5.4 Discrete-Time Approximation

Let us define the discrete instant of time tk ¼ t0 + kΔ, where Δ is the time step
employed, while

t f  t0
N¼
Δ
is the number of discretizations and tN ¼ tf. Then, the extremal problem (22) for all
firms f  F becomes the following:
8 !
  XN <X X
max Φ f c f , q f , s f ; cf , qf Δ τðtk Þe ρtk
πi cgi ðtk Þ, tk cif ðtk Þ
k¼0 :i  N gF
f

X f  X
 V i q f ðtk Þ, tk  r ij ðtk Þ sijf ðtk Þ
iN f ði, jÞ  W f

X  o
 ψ if I if ðtk Þ, tk Þ
iN f

subject to
Computable Models of Dynamic Spatial Oligopoly from the Perspective of. . . 317

2 3
X X
I if ðtk Þ ¼ I if ðtk1 Þ þ Δ4qif ðtk Þ þ sjif ðtk Þ  sijf ðtk Þ  cif ðtk Þ5
ð j, iÞ  W ði, jÞ  W

8k ¼ 1, . . . , N and i  N f

I if ðt0 Þ¼ K if
8i  N f
  f
I if t f ¼ K~i 8i  N f
0  c f ðt k Þ  C f 8k  ½1, N 
f f
0  q ðtk Þ  Q 8k  ½1, N 
f f
0  h ðtk Þ  H 8k  ½1, N 

where the vectors c f, q f, and h f have the obvious definitions; moreover, τ(t) is
presently the coefficient which arises from a trapezoidal approximation of the
present value integral; that is,
8
< 0:5 if
> t ¼ t0
τðtÞ ¼ 0:5 if t ¼ tf
>
:
1 if t0 < t < t f :

One advantage of time discretization is that we can now completely eliminate

state variables (inventories) from the problem by noting that
2 3
X
t X X
I if ðtkþ1 Þ ¼ K if þΔ 4q f ðtk Þ þ sjif ðtk Þ  sijf ðtk Þ  cif ðtk Þ5 ð31Þ
i
k¼0 ð j, iÞ  W ði, jÞ  W

  f
I if t f ¼ K~i ð32Þ

for t ¼ 0, . . . , N  1 and all i  N f . As a consequence, one obtains a finite-

dimensional variational inequality involving only upper and lower bound constraints
on the remaining control variables. This finite-dimensional variational inequality
may be solved by conventional algorithms developed for such problems or a finite-
dimensional nonlinear complementarity formulation may be created and used in
combination with a successive linearization scheme and a linear complementarity
solver.

5.5 A Comment About Path Variables

It should be noted that one may introduce path flows in the above formulation by
re-expressing the state dynamics as
318 T. L. Friesz and A. H. Meimand

dI if X X X X
¼ qif þ hpf  hpf  cif
dt jN pP jN pP
f ji f ij

for every firm f  F and node i  N f , where Pji is the set of paths from node j  N f
to node i  N f ; furthermore, hp is the flow on path p  Pji. There are corresponding,
but quite obvious, changes in the firm’s objective function and the upper and lower
bound constraints on its controls. We omit a complete statement of such details for
the sake of brevity.

5.6 Numerical Example

Let us consider a network of five arcs, four nodes, and four firms, where a single firm
f is located at each node i ¼ 1, 2, 3, 4. Consumption of each firm’s output
potentially occurs at every node; this consumption may be of local or of imported
output as the network topology permits. Figure 1 illustrates the network.
The time interval of interest is [0, 20]; that is, t0 ¼ 0 and tf ¼ 20. In this example,
firm 1 has an economic presence at all nodes; firm 2 at nodes 2, 3, and 4; firm 3 at
nodes 3 and 4; and finally firm 4 at node 4 only. Therefore F ¼ {1, 2, 3, 4} and
N 1 ¼ f1, 2, 3, 4g, N 2 ¼ f2, 3, 4g, N 3 ¼ f3, 4g, N 4 ¼ f4g. Before time discreti-
zation, there are 29 controls and ten state variables associated with this example;
these are enumerated in Table 1.
At time t0 ¼ 0, every firm has an inventory of 100 units at their respective
locations. That is I if ð0Þ ¼ 100 for f  F and i  N f . In addition, we impose the
condition that no backordering is allowed by any firm at any node at the terminal
time tf ¼ 20. That is,

I if ð20Þ  0 for f  F and i  N f : ð33Þ

Market 2

a1 a4

Market 1 1 a3 4 Market 3

a2 a5

Market 4

Fig. 1 Network of five arcs, four nodes, and four firms

Computable Models of Dynamic Spatial Oligopoly from the Perspective of. . . 319

Table 1 Controls by node or path

Firm Controls by node or path States
1 c11 c12 c13 c14 I 11 I 12 I 13 I 14
2 c22 c23 c24 I 22 I 23 I 24
3 c33 c34 I 33 I 34
4 c44 I 44
All q11 q22 q33 q44
1 h11 h12 h13 h14 h15 h16 h17 h18 h19 h110
2 h27 h28 h29 h210
3 h310

The inventory dynamics are the flow balance equations:

dI 11
¼ q11  h11  h12  h13  h14  h15  h16  c11
dt
dI 12
¼ h11  h17  h18  h19  c12
dt
dI 13 ð34Þ
¼ h12 þ h13 þ h17  h110  c13
dt
⋮
dI 44
¼ q44  c44
dt
which we only partially enumerated in the interest of saving space. We assume the
inverse demands at each node i take the following form:

π i ðci , tÞ ¼ αi  βi ðci Þm ð35Þ

where m  ℜ1þþ is a constant. Also αi  ℜ1þþ and βi  ℜ1þþ for all i are constants.
The production cost functions for each firm f have the form

1  2 1  3
V ii ¼ ρii qii þ σ ii qii for i ¼ 1, 2, 3, 4 ð36Þ
2 3
where ρif and σ if  ℜ1þþ are also constants for all allowed i and f. In expression (36),
we consider nonconvex production cost functions in order to capture both increasing
and decreasing economies of scale for different production rate regimes (Fig. 2). We
assume the holding costs are quadratic and of the form
 2
1
ψ if ¼ ηif I if for f  F and i  N f ð37Þ
2
where ηij  ℜ1þþ are constants, again for allowed i and f. The relationships between
arc and path variables are summarized in Table 2.
320 T. L. Friesz and A. H. Meimand

80
Q11 Q33
70
Q22 Q44
60
Production rate

50

40

30

20

10

0
0 2 4 6 8 10 12 14 16 18 20
Time

Fig. 2 Production rates

Table 2 Path definitions Path Arc sequence

p1 a1
p2 a2
p3 a1 , a3
p4 a1 , a4
p5 a1 , a3 , a5
p6 a2 , a5
p7 a3
p8 a4
p9 a3 , a5
p10 a5

Furthermore, the relevant arc-path incidence matrix is

2 3
1 0 1 1 1 0 0 0 0 0
6 7
60 1 0 0 0 1 0 0 0 07
  6 7
Δp ¼ δap ¼6
60 0 1 0 1 0 1 0 1 077
6 7
40 0 0 1 0 0 0 1 0 05
0 0 0 0 1 1 0 0 1 1

The associated path costs are

R ¼ ΔT r ð38Þ
Computable Models of Dynamic Spatial Oligopoly from the Perspective of. . . 321

where

r ¼ ðr ai : i ¼ 1, 2, 3, 4, 5Þ

and the

rai ¼ Ai þ Bai ð f i Þn i ¼ 1, 2, 3, 4, 5

are unit freight rates for individual arcs and the Ai  ℜ1þþ and Bi  ℜ1þþ are known
constants. We impose the following vectors of bounds on control variables:

C f ¼ Q f ¼ H f ¼ 75:

Each firm’s instantaneous profit function is found by substituting the functional

forms (35), (36), (37), and (38) into Eq. (23), where ρ  ℜ1þþ is again the fixed
nominal interest rate. A discrete-time approximation of the corresponding differen-
tial variational inequality is created using N ¼ 21 equal time steps. The resulting
finite-dimensional variational inequality is restated as a nonlinear complementarity
problem and solved using GAMS with the PATH solver. The numerical values of the
model’s parameters are presented in below Table.

Parameter Value Parameter Value Parameter Value

ρ 0.05 A1 2 A2 2
A3 2 A4 2 A5 2
B1 0.9 B2 0.9 B3 0.9
B4 0.9 B5 0.9 α1 2000
β1 12 α2 2200 β2 16
α3 2400 β3 14 α4 2500
β4 18 ρ11 0.3 ρ22 , ρ44 0.1
ρ33 0.2 σ ii , i ¼ 1, . . . , 4 1 η12 , η14 , η34 , η44 1
η11 4 η13 , η22 , η23 2 η24 , η33 3
t0 0 tf 20 N 20
Δ 1 n 1 m 1

5.7 Interpretation of Numerical Results

The production rates for the four firms and prices of finished goods in four spatially
separated markets are plotted in Fig. 3. Each firm seems to follow a different
production plan; firm 2 operates at its full capacity for the first ten time units,
abruptly halts production, and then returns to full production for the last time period
to meet the final inventory constraints, whereas firm 1 slowly increases production
until near the end of the planning horizon where production begins to decline. There
322 T. L. Friesz and A. H. Meimand

Market Price
3000
Market 1 Market 2 Market 3 Market 4
2500

2000
Price ($)

1500

1000

500

0
0 2 4 6 8 10 12 14 16 18 20
Time

Fig. 3 Market price

100%

80%
% of Total costs (NPV)

60%

40%

20%

0%
Firm 1 Firm 2 Firm 3 Firm 4
Firms

Production cost Inventory holding cost Freight cost

Fig. 4 Cost by firms

is a significant increase in price in market 4 where all the firms can compete. By
contrast, only firm 1 can sell in market 1 which shows relatively small change in
price.
In Fig. 4 we compare the net present of cumulative production, inventory holding,
and transportation costs incurred by the four firms.
Computable Models of Dynamic Spatial Oligopoly from the Perspective of. . . 323

The present values of profits for firm 1 is –$ 185, 592, firm 2 is –$926, 070, firm
3 is $ 248, 179, and firm 4 is –$314, 978. It is evident from the above that the only
firm to realize positive profits is firm 3; all other firms experience losses. The
numerical results for dynamic network oligopoly show a complicated temporal
behavior from firms that cannot be deduced prior to numerical analysis. This is
significant, suggesting that firms must be extremely capable of dramatically altering
production and distribution schedules if they are to compete in the final goods
market successfully.

6 Conclusions

In this chapter we have presented computable models for both static and dynamic
oligopoly networks. We started from a simple static model and move on to the
complex dynamic network. Variational inequality is also presented as an alternative
formulation for static oligopoly network. And diagonalization algorithm is discussed
as one of the numerical methods that can be employed to solve VI. We have also
shown that the Nash game that is dynamic oligopolistic network competition is
easily and naturally formulated as a differential variational inequality (DVI). In the
numerical example, section finite-dimensional and discrete-time approximation are
used to solve DVI.

7 Cross-References

▶ Supply Chains and Transportation Networks

References
Beckmann MJ, Puu T (1990) Spatial structures. Springer, Berlin/Heidelberg/New York
Brander JA, Zhang A (1993) Dynamic oligopoly behavior in the airline industry. Int J Ind Organ
11(3):407–435
Chan CK, Zhou Y, Wong KH (2018) A dynamic equilibrium model of the oligopolistic closed-loop
supply chain network under uncertain and time-dependent demands. Transp Res E Logist
Transp Rev 118(12):325–354
Dafermos S, Nagurney A (1987) Oligopolistic and competitive behavior of spatially separated
markets. Reg Sci Urban Econ 17(2):245–254
Friesz TL (1993) A spatial computable general equilibrium model. In: Proceedings of the workshop
on transportation and computable general equilibrium models, Venice, 19–21 May 1993
Friesz TL (2010) Dynamic optimization and differential games. Springer, New York
Friesz TL, Harker PT (1984) Multicriteria spatial price equilibrium network design: theory and
computational results. Transp Res B 17(5):411–426
Friesz TL, Rigdon MA, Mookherjee R (2006) Differential variational inequalities and shipper
dynamic oligopolistic network competition. Transp Res B 40(6):480–503
Greenhut M, Lane WJ (1989) Theory of oligopolistic competition. Manch Sch 57(3):248–261
Greenhut M, Norman G, Hung CS (1987) The economics of imperfect competition: a spatial
approach. Cambridge University Press, Cambridge
324 T. L. Friesz and A. H. Meimand

Hakimi SL (1983) On locating new facilities in a competitive environment. Eur J Oper Res
12(1):29–35
Harker PT (1984) Variational inequality approach for the determination of oligopolistic market
equilibrium. Math Program 30(1):105–111
Harker PT (1986) Alternative models of spatial competition. Oper Res 34(13):410–425
Hashimoto H (1985) A spatial Nash equilibrium model. In: Harker PT (ed) Spatial price equilibria:
advances in theory, computation, and application. Springer, Berlin/Heidelberg/New York,
pp 20–41
Henderson JM, Quandt RE (1980) Microeconomic theory: a mathematical approach, 3rd edn.
McGraw-Hill, New York
Markovich S (2008) Snowball: a dynamic oligopoly model with indirect network effects. J Econ
Dyn Control 132(3):909–938
Matsushima N, Matsumura T (2003) Mixed oligopoly and spatial agglomeration. Can J Econ
36(1):62–87
Mozafari M, Karimi B, Mahootchi M (2015) Dynamic pricing and freight transportation planning in
oligopolistic freight networks: a game theoretic approach. Transportmetrica A Transp Sci
11(10):918–938
Murphy FH, Sherali HD, Soyster AL (1982) A mathematical programming approach for determin-
istic oligopolistic market equilibriums. Math Program 24(1):92–106
Nagurney A (1999) Network economics: a variational inequality approach, Revised 2nd edn.
Kluwer, Boston
Nagurney A, Dong J (2016a) A dynamic network oligopoly model with transportation costs,
product differentiation, and quality competition. Comput Econ 44(2):201–229
Nagurney A, Dong J (2016b) Supply chain network oligopolies with product differentiation. In:
Competing on supply chain quality. Springer, Cham, pp 119–147
Nagurney A, Dupuis P, Zhang D (1994) A dynamical systems approach for network oligopolies and
variational inequalities. Ann Oper Res 28(3):263–293
Nagurney A, Dong J, Zhang D (2002) A supply chain network equilibrium model. Transp Res
E Logist Transp Rev 38(5):281–303
Novshek W (1980) Equilibrium in simple spatial (or differentiated product) models. J Econ Theory
22(2):313–326
Raa T (1984) The distribution approach to spatial economics. J Reg Sci 24(1):105–117
Rovinskey RB, Shoemaker CA, Todd MJ (1980) Determining optimal use of resources among
regional producers under differing levels of cooperation. Oper Res 28(4):859–866
Taheri I, Rashidnejad M, Badri A, Rahimi-Kian A (2014) IEEE System 9(4):1452–1462
Tobin RL, Friesz TL (1983) Formulating and solving the spatial price equilibrium problem with
transshipment in terms of arc variables. J Reg Sci Res 28(4):859–866
Sethi SP, Thompson GL (2000) Optimal control theory: Applications to management science and
economics (second edition), Kluwer Academic Publishers, Boston, MA
Wie BW, Tobin RL (1997) A dynamic spatial Cournot-Nash equilibrium model and an algorithm.
Comput Econ 10(1):15–45